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
|
|---|---|---|---|---|---|---|---|---|---|
1312.7581 | 4 | 1312 | 2015-04-20T07:32:28 | On the Learning Behavior of Adaptive Networks - Part I: Transient Analysis | [
"cs.MA",
"eess.SY",
"math.OC"
] | This work carries out a detailed transient analysis of the learning behavior of multi-agent networks, and reveals interesting results about the learning abilities of distributed strategies. Among other results, the analysis reveals how combination policies influence the learning process of networked agents, and how these policies can steer the convergence point towards any of many possible Pareto optimal solutions. The results also establish that the learning process of an adaptive network undergoes three (rather than two) well-defined stages of evolution with distinctive convergence rates during the first two stages, while attaining a finite mean-square-error (MSE) level in the last stage. The analysis reveals what aspects of the network topology influence performance directly and suggests design procedures that can optimize performance by adjusting the relevant topology parameters. Interestingly, it is further shown that, in the adaptation regime, each agent in a sparsely connected network is able to achieve the same performance level as that of a centralized stochastic-gradient strategy even for left-stochastic combination strategies. These results lead to a deeper understanding and useful insights on the convergence behavior of coupled distributed learners. The results also lead to effective design mechanisms to help diffuse information more thoroughly over networks. | cs.MA | cs | IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
1
On the Learning Behavior of Adaptive Networks --
Part I: Transient Analysis
Jianshu Chen, Member, IEEE, and Ali H. Sayed, Fellow, IEEE
5
1
0
2
r
p
A
0
2
]
A
M
.
s
c
[
4
v
1
8
5
7
.
2
1
3
1
:
v
i
X
r
a
Abstract -- This work carries out a detailed transient analysis
of the learning behavior of multi-agent networks, and reveals
interesting results about the learning abilities of distributed
strategies. Among other results, the analysis reveals how com-
bination policies influence the learning process of networked
agents, and how these policies can steer the convergence point
towards any of many possible Pareto optimal solutions. The
results also establish that the learning process of an adaptive
network undergoes three (rather than two) well-defined stages
of evolution with distinctive convergence rates during the first
two stages, while attaining a finite mean-square-error (MSE)
level in the last stage. The analysis reveals what aspects of the
network topology influence performance directly and suggests
design procedures that can optimize performance by adjusting
the relevant topology parameters. Interestingly,
it is further
shown that, in the adaptation regime, each agent in a sparsely
connected network is able to achieve the same performance level
as that of a centralized stochastic-gradient strategy even for left-
stochastic combination strategies. These results lead to a deeper
understanding and useful insights on the convergence behavior
of coupled distributed learners. The results also lead to effective
design mechanisms to help diffuse information more thoroughly
over networks.
Index Terms -- Multi-agent learning, multi-agent adaptation,
distributed strategies, diffusion of information, Pareto solutions.
I. INTRODUCTION
In multi-agent systems, agents interact with each other to
solve a problem of common interest, such as an optimization
problem in a distributed manner. Such networks of interacting
agents are useful in solving distributed estimation, learning
and decision making problems [2] -- [39]. They are also useful
in modeling biological networks [40] -- [42], collective rational
behavior [12], [13], and in developing biologically-inspired
designs [2], [43]. Two useful strategies that can be used to
guide the interactions of agents over a network are consensus
strategies [5] -- [11] and diffusion strategies [16] -- [27]. Both
Manuscript received December 28, 2013; revised November 21, 2014;
accepted April 08, 2015. This work was supported in part by NSF grants CCF-
1011918 and ECCS-1407712. A short and limited version of this work appears
in the conference publication [1] without proofs and under more restrictive
conditions than considered in this broader and expanded work.
J. Chen was with Department of Electrical Engineering, University of
California, Los Angeles, and is currently with Microsoft Research, Redmond,
WA 98052. This work was performed while he was a PhD student at UCLA.
Email: [email protected].
A. H. Sayed is with Department of Electrical Engineering, University of
California, Los Angeles, CA 90095. Email: [email protected].
Communicated by Prof. Nicol`o Cesa-Bianchi, Associate Editor for Pattern
Recognition, Statistical Learning, and Inference.
Copyright (c) 2014 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].
classes of algorithms involve self-learning and social-learning
steps. During self-learning, each agent updates its state using
its local data. During social learning, each agent aggregates
information from its neighbors. A useful feature that results
from these localized interactions is that the network ends
up exhibiting global patterns of behavior. For example, in
distributed estimation and learning, each agent is able to attain
the performance of centralized solutions by relying solely on
local cooperation [6], [19], [22], [23].
In this article, and the accompanying Part II [44], we con-
sider a general class of distributed strategies, which includes
diffusion and consensus updates as special cases, and study the
resulting global learning behavior by addressing four important
questions: (i) where does the distributed algorithm converge
to? (ii) when does it converge? (iii) how fast does it converge?
and (iv) how close does it converge to the intended point? We
answer questions (i) -- (iii) in Part I and question (iv) in Part
II [44]. We study these four questions by characterizing the
learning dynamics of the network in some great detail. An
interesting conclusion that follows from our analysis is that the
learning curve of a multi-agent system will be shown to exhibit
three different phases. In the first phase (Transient Phase I), the
convergence rate of the network is determined by the second
largest eigenvalue of the combination matrix in magnitude,
which is related to the degree of network connectivity. In
the second phase (Transient Phase II), the convergence rate is
determined by the entries of the right-eigenvector of the com-
bination matrix corresponding to the eigenvalue at one. And,
in the third phase (the steady-state phase) the mean-square
performance of the algorithm turns out to depend on this same
right-eigenvector in a revealing way. Even more surprisingly,
we shall discover that
the agents have the same learning
behavior starting at Transient Phase II, and are able to achieve
a performance level that matches that of a fully connected
network or a centralized stochastic-gradient strategy. Actually,
we shall show that the consensus and diffusion strategies can
be represented as perturbed versions of a centralized reference
recursion in a certain transform domain. We quantify the
effect of the perturbations and establish the aforementioned
properties for the various phases of the learning behavior of
the networks. The results will reveal the manner by which the
network topology influences performance in some unexpected
ways.
There have been of course many insightful works in the
literature on distributed strategies and their convergence be-
havior. In Sections II-B and IV-A further ahead, we explain
in what ways the current manuscript extends these earlier
investigations and what novel contributions this work leads
2
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
to. In particular, it will be seen that several new insights are
discovered that clarify how distributed networks learn. For the
time being, in these introductory remarks, we limit ourselves
to mentioning one important aspect of our development. Most
prior studies on distributed optimization and estimation tend to
focus on the performance and convergence of the algorithms
under diminishing step-size conditions [5] -- [10], [28] -- [30],
[45], or on convergence under deterministic conditions on the
data [10]. This is perfectly fine for applications involving static
optimization problems where the objective is to locate the
fixed optimizer of some aggregate cost function of interest.
In this paper, however, we examine the learning behavior
of distributed strategies under constant step-size conditions.
This is because constant step-sizes are necessary to enable
continuous adaptation, learning, and tracking in the presence
of streaming data and drifting conditions. These features
would enable the algorithms to perform well even when the
location of the optimizer drifts with time. Nevertheless, the use
of constant step-sizes enriches the dynamics of (stochastic-
gradient) distributed algorithms in that the gradient update
term does not die out with time anymore, in clear contrast
to the diminishing step-size case where the influence of the
gradient term is annihilated over time due to the decaying
value of the step-size parameter. For this reason, more care
is needed to examine the learning behavior of distributed
strategies in the constant step-size regime since their updates
remain continually active and the effect of gradient noise is
always present. This work also generalizes and extends in
non-trivial ways the studies in [18], [20]. For example, while
reference [18] assumed that the individual costs of all agents
have the same minimizer, and reference [20] assumed that each
of these individual costs is strongly convex, these requirements
are not needed in the current study: individual costs can have
distinct minimizers and they do not even need to be convex
(see the discussion after expression (32)). This fact widens
significantly the class of distributed learning problems that are
covered by our framework. Moreover, the network behavior
is studied under less restrictive assumptions and for broader
scenarios, including a close study of the various phases of
evolution during the transient phase of the learning process.
We also study a larger class of distributed strategies that
includes diffusion and consensus strategies as special cases.
To examine the learning behavior of adaptive networks
under broader and more relaxed conditions than usual, we
pursue a new analysis route by introducing a reference
centralized recursion and by studying the perturbation
of the diffusion and consensus strategies relative to this
centralized solution over time. Insightful new results are
obtained through this perturbation analysis. For example,
we are now able to examine closely both the transient
phase behavior and the steady-state phase behavior of the
learning process and to explain how behavior in these two
stages relate to the behavior of
the centralized solution
(see Fig. 2 further ahead). Among several other results,
the mean-square-error expression (52) derived later in Part
II [44] following some careful analysis, which builds on
the results of this Part I, is one of the new (compact and
powerful) insights; it reveals how the performance of each
agent is closely related to that of the centralized stochastic
approximation strategy -- see the discussion right after
(52). As the reader will ascertain from the derivations in the
appendices, arriving at these conclusions for a broad class of
distributed strategies and under weaker conditions than usual
is demanding and necessitates a careful study of the evolution
of the error dynamics over the network and its stability.
When all is said and done, Parts I and II [44] lead to several
novel insights into the learning behavior of adaptive networks.
Notation. 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) diago-
nal 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, and the notation X (cid:22) Y
means each entry of the matrix X is less than or equal to the
corresponding entry of the matrix Y . The notation x = vec(X)
denotes the vectorization operation that stacks the columns
of a matrix X on top of each other to form a vector x, and
X = vec−1(x) is the inverse operation. The operators ∇w and
∇wT denote the column and row gradient vectors with respect
to w. When ∇wT is applied to a column vector s, it generates
a matrix. The notation a(µ) = O(b(µ)) means that there exists
a constant C > 0 such that
for all µ, a(µ) ≤ C · b(µ).
II. PROBLEM FORMULATION
A. Distributed Strategies: Consensus and Diffusion
We consider a connected network of N agents that are
linked together through a topology -- see Fig. 1. Each agent
k implements a distributed algorithm of the following form to
update its state vector from wk,i−1 to wk,i:
l=1
N(cid:88)
N(cid:88)
N(cid:88)
l=1
φk,i−1 =
ψk,i =
wk,i =
a1,lkwl,i−1
a0,lkφl,i−1 − µk sk,i(φk,i−1)
a2,lkψl,i
(1)
(2)
(3)
l=1
where wk,i ∈ RM is the state of agent k at time i, usually
an estimate for the solution of some optimization problem,
φk,i−1 ∈ RM and ψk,i ∈ RM are intermediate variables gen-
erated at node k before updating to wk,i, µk is a non-negative
constant step-size parameter used by node k, and sk,i(·) is
an M × 1 update vector function at node k. In deterministic
optimization problems, the update vectors sk,i(·) can be the
gradient or Newton steps associated with the cost functions
[10]. On the other hand, in stocastic approximation problems,
such as adaptation, learning and estimation problems [5] -- [9],
[14] -- [23], [25] -- [29], the update vectors are usually computed
from realizations of data samples that arrive sequentially at
the nodes. In the stochastic setting, the quantities appearing
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
3
Fig. 1. A network representing a multi-agent system. The set of all agents
that can communicate with node k is denoted by Nk.The edge linking any two
agents is represented by two directed arrows to emphasize that information
can flow in both directions.
in (1) -- (3) become random and we use boldface letters to
highlight their stochastic nature. In Example 1 below, we
illustrate choices for sk,i(w) in different contexts.
The combination coefficients a1,lk, a0,lk and a2,lk in (1) --
(3) are nonnegative weights that each node k assigns to
the information arriving from node l; these coefficients are
required to satisfy:
N(cid:88)
N(cid:88)
N(cid:88)
a1,lk = 1,
a0,lk = 1,
a2,lk = 1
l=1
l=1
a0,lk ≥ 0,
a1,lk ≥ 0,
a1,lk = a2,lk = a0,lk = 0,
l=1
a2,lk ≥ 0
if l /∈ Nk
(4)
(5)
(6)
Observe from (6) that the combination coefficients are zero
if l /∈ Nk, where Nk denotes the set of neighbors of node
k. Therefore, each summation in (1) -- (3) is actually confined
to the neighborhood of node k. In algorithm (1) -- (3), each
node k first combines the states {wl,i−1} from its neighbors
and updates wk,i−1 to the intermediate variable φk,i−1. Then,
the {φl,i−1} from the neighbors are aggregated and updated
to ψk,i along the opposite direction of sk,i(φk,i−1). Finally,
the intermediate estimators {ψl,i} from the neighbors are
combined to generate the new state wk,i at node k.
Example 1: The distributed algorithm (1) -- (3) can be ap-
plied to optimize aggregate costs of the following form:
N(cid:88)
k=1
• In distributed estimation and learning,
node k is chosen as the deterministic gradient (column)
vector ∇wJk(·).
the individual
cost function at each node k is usually selected as
the expected value of some loss function Qk(·,·), i.e.,
Jk(w) = E{Qk(w, xk,i)} [18], where the expectation
is with respect to the randomness in the data samples
{xk,i} collected at node k at time i. The exact expression
for ∇wJk(w) is usually unknown since the probability
distribution of the data is not known beforehand. In
these situations, the update vector sk,i(·) is chosen as an
instantaneous approximation for the true gradient vector,
such as, sk,i(·) = (cid:92)
∇wJk(·) = ∇wQk(·, xk,i), which is
known as stochastic gradient. Note that the update vector
sk,i(w) is now evaluated from the random data sample
xk,i. Therefore, it is also random and time dependent.
The update vectors {sk,i(·)} may not necessarily be the
gradients of cost functions or their stochastic approximations.
They may take other forms for different reasons. For example,
in [6], a certain gain matrix K is multiplied to the left of
the stochastic gradient vector (cid:92)
∇wJk(·) to make the estimator
asymptotically efficient for a linear observation model.
Returning to the general distributed strategy (1) -- (3), we note
that it can be specialized into various useful algorithms. We
let A1, A0 and A2 denote the N × N matrices that collect the
coefficients {a1,lk}, {a0,lk} and {a2,lk}. Then, condition (4)
is equivalent to
AT
1
1 = 1
1 = 1, AT
0
1 = 1, AT
2
(9)
where 1 is the N × 1 vector with all its entries equal to
one. Condition (9) means that the matrices {A0, A1, A2} are
left-stochastic (i.e., the entries on each of their columns add
up to one). Different choices for A1, A0 and A2 correspond
to different distributed strategies, as summarized in Table I.
Specifically, the traditional consensus [5] -- [11] and diffusion
(ATC and CTA) [16] -- [23] algorithms with constant step-sizes
are given by the following iterations:
Consensus :
l∈Nk
(cid:88)
(cid:88)
a0,lkwl,i−1
wk,i = φk,i−1 − µk sk,i(wk,i−1)
φk,i−1 =
φk,i−1 =
ψk,i = wk,i−1−µk sk,i(wk,i−1)
(cid:88)
wk,i = φk,i−1 − µk sk,i(φk,i−1)
a1,lkwl,i−1
a2,lkψl,i
wk,i =
l∈Nk
l∈Nk
(10)
(11)
(12)
J glob(w) =
Jk(w)
(7)
CTA diffusion :
or to find Pareto-optimal solutions to multi-objective optimiza-
tion problems, such as:
w {J1(w), . . . , JN (w)}
min
(8)
ATC diffusion :
where Jk(w) is an individual convex cost associated with
each agent k. Optimization problems of the form (7) -- (8) arise
in various applications -- see [3] -- [31]. Depending on the
context, the update vector sk,i(·) may be chosen in different
ways:
• In deterministic optimization problems, the expressions
for {Jk(w)} are known and the update vector sk,i(·) at
Therefore, the convex combination steps appear in different
locations in the consensus and diffusion implementations. For
instance, observe that the consensus strategy (10) evaluates the
update direction sk,i(·) at wk,i−1, which is the estimator prior
to the aggregation, while the diffusion strategy (11) evaluates
the update direction at φk,i−1, which is the estimator after the
Nkk123456789a1kak110 4
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
DIFFERENT CHOICES FOR A1, A0 AND A2 CORRESPOND TO DIFFERENT
DISTRIBUTED STRATEGIES.
TABLE I
Consensus
Distributed Strategeis A1 A0 A2
I
A
I
ATC diffusion
CTA diffusion
I
I
A
A
I
I
A1A0A2
A
A
A
aggregation. In our analysis, we will proceed with the general
form (1) -- (3) to study all three schemes, and other possibilities,
within a unifying framework.
We observe that there are two types of learning processes
involved in the dynamics of each agent k: (i) self-learning
in (2) from locally sensed data and (ii) social learning in
(1) and (3) from neighbors. All nodes implement the same
self- and social learning structure. As a result, the learning
dynamics of all nodes in the network are coupled; knowledge
exploited from local data at node k will be propagated to
its neighbors and from there to their neighbors in a diffusive
learning process. It is expected that some global performance
pattern will emerge from these localized interactions in the
multi-agent system. In this work and the accompanying Part
II [44], we address the following questions:
• Limit point: where does each state wk,i converge to?
• Stability: under which condition does convergence occur?
• Learning rate: how fast does convergence occur?
• Performance: how close is wk,i to the limit point?
• Generalization: can wk,i match the performance of a
centralized solution?
We address the first three questions in this part, and examine
the last two questions pertaining to performance in Part II [44].
We address the five questions by characterizing analytically the
learning dynamics of the network to reveal the global behavior
that emerges in the small step-size regime. The answers to
these questions will provide useful and novel insights about
how to tune the algorithm parameters in order to reach desired
performance levels -- see Sec. VI in Part II [44].
B. Relation to Prior Work
In comparison with the existing literature [5] -- [10], [28] --
[30], [45] -- [48], it is worth noting that most prior studies on
distributed optimization algorithms focus on studying their
performance and convergence under diminishing step-size con-
ditions and for doubly-stochastic combination policies (i.e.,
matrices for which the entries on each of their columns and
on each of their rows add up to one). These are of course
useful conditions, especially when the emphasis is on solving
static optimization problems. We focus instead on the case of
constant step-sizes because, as explained earlier, they enable
continuous adaptation and learning under drifting conditions;
in contrast, diminishing step-sizes turn off learning once they
approach zero. By using constant step-sizes,
the resulting
algorithms are able to track dynamic solutions that may slowly
drift as the underlying problem conditions change.
Moreover, constant step-size implementations have merits
even for stationary environments where the solutions remain
static. This is because, as we are going to show later in this
work and its accompanying Part II [44], constant step-size
learning converges at a geometric rate, in the order of O(γi)
for some 0 < γ < 1, towards a small mean-square error in
the order of the step-size parameter. This means that these
solutions can attain satisfactory performance even after short
intervals of time. In comparison, implementations that rely on
a diminishing step-size of the form µ(i) = µo/i, for some
constant µo, converge almost surely to the solution albeit at
the slower rate of O(1/i). Furthermore, the choice of the
parameter µo is critical to guarantee the O(1/i) rate [49, p.54];
if µo is not large enough, the resulting convergence rate can
be considerably slower than O(1/i). To avoid this slowdown
in convergence, a large initial value µo is usually chosen in
practice, which ends up leading to an overshoot in the learning
curve; the curve grows up initially before starting its decay at
the asymptotic rate, O(1/i).
We remark that we also do not limit the choice of com-
bination policies to being doubly-stochastic; we only require
condition (9). It turns out that left-stochastic matrices lead to
superior mean-square error performance (see, e.g., expression
(63) in Part II [44] and also [17]). The use of both constant
step-sizes and left-stochastic combination policies enrich the
learning dynamics of the network in interesting ways, as we
are going to discover. In particular, under these conditions, we
will derive an interesting result that reveals how the topology
of the network determines the limit point of the distributed
strategies. We will show that the combination weights steer
the convergence point away from the expected solution and
towards any of many possible Pareto optimal solutions. This is
in contrast to commonly-used doubly-stochastic combination
policies where the limit point of the network is fixed and
cannot be changed regardless of the topology. We will show
that the limit point is determined by the right eigenvector that
is associated with the eigenvalue at one for the matrix product
A1A0A2. We will also be able to characterize in Part II [44]
how close each agent in the network gets to this limit point
and to explain how the limit point plays the role of a Pareto
optimal solution for a suitably defined aggregate cost function.
We note that the concept of a limit point in this work is
different from earlier studies on the limit point of consen-
sus implementations that deal exclusively with the problem
of evaluating the weighted average of initial state values
there
at
are no adaptation steps and no streaming data;
the step-
size parameters {µk} are set to zero in (2), (10), (11) and
(12). In contrast, the general distributed strategy (1) -- (3) is
meant to solve continuous adaptation and learning problems
from streaming data arriving at the agents. In this case, the
adaptation term sk,i(·) (self-learning) is necessary, in addition
to the combination step (social-learning). There is a non-trivial
coupling between both steps and across the agents. For this
reason, identifying the actual limit point of the distributed
strategy is rather challenging and requires a close examination
of the evolution of the network dynamics, as demonstrated by
the technical tools used in this work. In comparison, while
the evolution of traditional average-consensus implementations
can be described by linear first-order recursions, the same is
not true for adaptive networks where the dynamics evolves
the agents (e.g., [50]). In these implementations,
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
5
according to nonlinear stochastic difference recursions.
where πk is the kth entry of the vector π.
III. MODELING ASSUMPTIONS
In this section, we collect the assumptions and definitions
that are used in the analysis and explain why they are justified
and how they relate to similar assumptions used in several
prior studies in the literature. As the discussion will reveal, in
most cases, the assumptions that we adopt here are relaxed
(i.e., weaker) versions than conditions used before in the
literature such as in [5] -- [7], [10], [12], [18] -- [20], [28] -- [30],
[49], [51]. We do so in order to analyze the learning behavior
of networks under conditions that are similar to what
is
normally assumed in the prior art, albeit ones that are generally
less restrictive.
Assumption 1 (Strongly-connected network): The N × N
matrix product A (cid:44) A1A0A2 is assumed to be a primitive
left-stochastic matrix, i.e., AT 1 = 1 and there exists a finite
integer jo such that all entries of Ajo are strictly positive.
This condition is satisfied for most networks and is not
restrictive. Let A = [alk] denote the entries of A. Assumption
1 is automatically satisfied if the product A corresponds to a
connected network and there exists at least one akk > 0 for
some node k (i.e., at least one node with a nontrivial self-loop)
[21], [23]. It then follows from the Perron-Frobenius Theorem
[52] that the matrix A1A0A2 has a single eigenvalue at one
of multiplicity one and all other eigenvalues are strictly less
than one in magnitude, i.e.,
1 = λ1(A) > λ2(A) ≥ ··· ≥ λN (A)
(13)
Obviously, 1T is a left eigenvector for A1A0A2 corresponding
to the eigenvalue at one. Let θ denote the right eigenvector
corresponding to the eigenvalue at one (the Perron vector) and
whose entries are normalized to add up to one, i.e.,
Aθ = θ,
1T θ = 1
(14)
Then, the Perron-Frobenius Theorem further ensures that all
entries of θ satisfy 0 < θk < 1. Note that, unlike [5] -- [11],
[28], [29], we do not require the matrix A1A0A2 to be doubly-
stochastic (in which case θ would be 1/N and, therefore, all its
entries will be identical to each other). Instead, we will study
the performance of the algorithms in the context of general
left-stochastic matrices {A1, A0, A2} and we will examine the
influence of (the generally non-equal entries of) θ on both the
limit point and performance of the network.
Definition 1 (Step-sizes): Without loss of generality, we ex-
press the step-size at each node k as µk = µmaxβk, where
µmax (cid:44) max{µk} is the largest step-size, and 0 ≤ βk ≤ 1.
We assume βk > 0 for at least one k. Thus, observe that we
are allowing the possibility of zero step-sizes by some of the
agents.
Definition 2 (Useful vectors): Let π and p be the following
N × 1 vectors:
π (cid:44) A2θ
p (cid:44) col{π1β1, . . . , πN βN}
(15)
(16)
The vector p will play a critical role in the performance of
the distributed strategy (1) -- (3). Furthermore, we introduce the
following assumptions on the update vectors sk,i(·) in (1) -- (3).
Assumption 2 (Update vector: Randomness): There exists
an M × 1 deterministic vector function sk(w) such that, for
all M × 1 vectors w in the filtration Fi−1 generated by the
past history of iterates {wk,j} for j ≤ i− 1 and all k, it holds
that
E{sk,i(w)Fi−1} = sk(w)
for all i, k. Furthermore, there exist α ≥ 0 and σ2
that for all i, k and w ∈ Fi−1:
E(cid:110)
(cid:107)sk,i(w)−sk(w)(cid:107)2(cid:12)(cid:12)Fi−1
(cid:111)
≤ α·(cid:107)w(cid:107)2 +σ2
v
(18)
(17)
v ≥ 0 such
Condition (18) requires the conditional variance of the
random update direction sk,i(w) to be bounded by the square-
norm of w. Condition (18) is a generalized version of As-
sumption 2 from [18], [20]; it is also a generalization of the
assumptions from [28], [49], [51], where sk,i(w) was instead
modeled as the following perturbed version of the true gradient
vector:
sk,i(w) = (cid:92)
∇wJk(w) = ∇wJk(w) + vk,i(w)
(19)
with sk(w) = ∇wJk(w), in which case conditions (17) -- (18)
translate into the following requirements on the gradient noise
vk,i(w):
(cid:111)
E(cid:110)
(cid:107)vk,i(w)(cid:107)2(cid:12)(cid:12)Fi−1
E{vk,i(w)Fi−1} = 0
(zero mean)
≤ α·(cid:107)w(cid:107)2 +σ2
v
(20)
(21)
In Example 2 of [18], we explained how these conditions
are satisfied automatically in the context of mean-square-error
adaptation over networks. Assumption 2 given by (17) -- (18)
is more general than (20) -- (21) because we are allowing the
update vector sk,i(·) to be constructed in forms other than
(19). Furthermore, Assumption (21) is also more relaxed than
the following variant used in [49], [51]:
≤ α·(cid:107)∇wJk(w)(cid:107)2 +σ2
v
(22)
This is because (22) implies a condition of the form (21).
Indeed, note that
(cid:111)
E(cid:110)
(cid:107)vk,i(w)(cid:107)2(cid:12)(cid:12)Fi−1
(cid:111)
E(cid:110)
(cid:107)vk,i(w)(cid:107)2(cid:12)(cid:12)Fi−1
v
(b)
(a)
= α·(cid:107)∇wJk(w) − ∇wJk(0) + ∇wJk(0)(cid:107)2 +σ2
≤ 2α·(cid:107)∇wJk(w) − ∇wJk(0)(cid:107)2 + 2α(cid:107)∇wJk(0)(cid:107)2 +σ2
≤ 2αλ2
(cid:44) α
U · (cid:107)w(cid:107)2 + 2α(cid:107)∇wJk(0)(cid:107)2 +σ2
· (cid:107)w(cid:107)2 + σ2
v(cid:48)
(23)
where step (a) uses the relation (cid:107)x + y(cid:107)2 ≤ 2(cid:107)x(cid:107)2 + 2(cid:107)y(cid:107)2,
and step (b) used (24) to be assumed next.
v
v
(cid:48)
6
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
N(cid:88)
(cid:104)
(cid:105)
Assumption 3 (Update vector: Lipschitz): There exists a
nonnegative λU such that for all x, y ∈ RM and all k:
(cid:107)sk(x) − sk(y)(cid:107) ≤ λU · (cid:107)x − y(cid:107)
(24)
where the subscript "U" in λU means "upper bound".
A similar assumption to (24) was used before in the literature
for the model (19) by requiring the gradient vector of the
individual cost functions Jk(w) to be Lipschitz [5], [12], [30],
[49], [51]. Again, condition (24) is more general because we
are not limiting the construction of the update direction to
(19).
Assumption 4 (Update vector: Strong monotonicity): Let
pk denote the kth entry of the vector p defined in (16). There
exists λL > 0 such that for all x, y ∈ RM :
pk
(x − y)T ·
≥ λL · (cid:107)x − y(cid:107)2
where the subscript "L" in λL means "lower bound".
sk(x) − sk(y)
k=1
(25)
Remark 1: Applying the Cauchy-Schwartz inequality [52,
p.15] to the left-hand side of (25) and using (24), we deduce
the following relation between λL and λU :
λU · (cid:107)p(cid:107)1 ≥ λL
(26)
The following lemma gives the equivalent forms of Assump-
where (cid:107) · (cid:107)1 denotes the 1−norm of the vector argument.
tions 3 -- 4 when the {sk(w)} happen to be differentiable.
Lemma 1 (Equivalent conditions on update vectors):
Suppose {sk(w)} are differentiable in an open set S ⊆ RM .
Then, having conditions (24) and (25) hold on S is equivalent
to the following conditions, respectively,
(cid:107)∇wT sk(w)(cid:107) ≤ λU
[Hc(w) + H T
c (w)] ≥ λL · IM
1
2
(27)
(28)
for any w ∈ S, where (cid:107)·(cid:107) denotes the 2-induced norm (largest
singular value) of its matrix argument and
Hc(w) (cid:44) n(cid:88)
Proof: See Appendix B.
k=1
Since in Assumptions 3 -- 4 we require conditions (24)
and (25) to hold over the entire RM , then the equivalent
conditions (27) -- (28) will need to hold over the entire RM
when the {sk(w)} are differentiable. In the context of dis-
tributed optimization problems of the form (7) -- (8) with twice-
differentiable Jk(w), where the stochastic gradient vectors
are constructed as in (19), Lemma 1 implies that the above
Assumptions 3 -- 4 are equivalent to the following conditions
on the Hessian matrix of each Jk(w) [49, p.10]:
wJk(w)(cid:13)(cid:13) ≤ λU
(cid:13)(cid:13)∇2
N(cid:88)
pk∇2
wJk(w) ≥ λLIM > 0
k=1
pk∇wT sk(w)
(29)
i.e.,
J glob,(cid:63)(w) (cid:44) N(cid:88)
Condition (31) is in turn equivalent to requiring the following
weighted sum of the individual cost functions {Jk(w)} to be
strongly convex:
pkJk(w)
(32)
k=1
We note that strong convexity conditions are prevalent
in
many studies on optimization techniques in the literature. For
example, each of the individual costs Jk(w) is assumed to be
stronlgy convex in [29] in order to derive upper bounds on
the limit superior ("lim sup") of the mean-square-error of the
estimates wk,i or the expected value of the cost function at
wk,i. In comparison, the framework in this work does not
require the individual costs to be strongly convex or even
convex. Actually, some of the costs {Jk(w)} can be non-
convex as long as the aggregate cost (32) remains strongly
convex. Such relaxed assumptions on the individual costs
introduce challenges into the analysis, and we need to develop
a systematic approach to characterize the limiting behavior of
adaptive networks under such less restrictive conditions.
Example 2: The strong-convexity condition (31) on the
aggregate cost (32) can be related to a global observability
condition similar to [6] -- [8]. To illustrate this point, we con-
sider an example dealing with quadratic costs. Thus, consider
a network of N agents that are connected according to a
certain topology. The data samples received at each agent k
at time i consist of the observation signal dk(i) ∈ R and
the regressor vector uk,i ∈ R1×M , which are assumed to be
related according to the following linear model:
(33)
where vk(i) ∈ R is a zero-mean additive white noise that
is uncorrelated with the regressor vector u(cid:96),j for all k, (cid:96), i, j.
Each agent in the network would like to estimate wo ∈ RM
by learning from the local data stream {dk(i), uk,i} and by
collaborating with its intermediate neighbors. The problem can
be formulated as minimizing the aggregate cost (7) with Jk(w)
chosen to be
dk(i) = uk,iwo + vk(i)
Jk(w) =
J glob(w) =
1
2
Edk(i) − uk,iw2
N(cid:88)
1
2
Edk(i) − uk,iw2
k=1
(34)
(35)
This is a distributed least-mean-squares (LMS) estimation
problem studied in [16], [17], [19]. We would like to explain
that condition (31) amounts to a global observability condition.
First, note that the Hessian matrix of Jk(w) in this case is the
covariance matrix of the regressor uk,i:
k,iuk,i}
(36)
Therefore, condition (31) becomes that there exists a λL > 0
such that
Ru,k (cid:44) E{uT
N(cid:88)
(37)
pkRu,k ≥ λLIM > 0
k=1
(30)
(31)
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
7
it can be verified that
Furthermore,
the above inequality
holds for any positive {pk} as long as the following global
observability condition holds:
Ru,k > 0
(38)
N(cid:88)
k=1
N(cid:88)
N(cid:88)
k=1
To see this, let pmin = mink pk, and write the left-hand side
of (37) as
N(cid:88)
k=1
pkRu,k = pmin
Ru,k +
N(cid:88)
(cid:124)
k=1
(pk − pmin)Ru,k
(cid:123)(cid:122)
(cid:125)
·IM
(39)
≥ pmin
Ru,k > pminλu,min
k=1
(cid:44)λL
where λu,min denotes the minimum eigenvalue of(cid:80)N
k=1 Ru,k.
Note that the left-hand side of (38) is the Hessian of J glob(w)
in (35). Therefore, condition (38) means that the aggregate
cost function (35) is strongly convex so that the information
provided by the linear observation model (33) over the entire
network is sufficient to uniquely identify the minimizer of (35).
Similar global observability conditions were used in [6] -- [8]
to study the performance of distributed parameter estimation
problems. Such conditions are useful because it implies that
even if wo is not locally observable to any agent in the network
but is globally observable, i.e., Ru,k > 0 does not hold for any
k = 1, . . . , N but (38) holds, the distributed strategies (10) --
(12) will still enable each agent to estimate the correct wo
through local cooperation. In Part II [44], we provide more
insights into how cooperation benefits the learning at each
agent.
Assumption 5 (Jacobian matrix: Lipschitz): Let wo denote
the limit point of the distributed strategy (1) -- (3), which is
defined further ahead as the unique solution to (42). Then, in
a small neighborhood around wo, we assume that sk(w) is
differentiable with respect to w and satisfies
(cid:107)∇wT sk(wo + δw) − ∇wT sk(wo)(cid:107) ≤ λH · (cid:107)δw(cid:107)
(40)
for all (cid:107)δw(cid:107) ≤ rH for some small rH, and where λH is a
nonnegative number independent of δw.
In the context of distributed optimization problems of the form
(7) -- (8) with twice-differentiable Jk(w), where the stochastic
gradient vectors are constructed as in (19), the above Assump-
tion translates into the following Lipschitz Hessian condition:
(cid:107)∇2
wJk(wo + δw) − ∇2
wJk(wo)(cid:107) ≤ λH · (cid:107)δw(cid:107)
(41)
Condition (40) is useful when we examine the convergence
rate of the algorithm later in this article. It is also useful in
deriving the steady-state mean-square-error expression (52) in
Part II [44].
IV. LEARNING BEHAVIOR
A. Overview of Main Results
Before we proceed to the formal analysis, we first give
a brief overview of the main results that we are going to
establish in this part on the learning behavior of the distributed
strategies (1) -- (3) for sufficiently small step-sizes. The first
major conclusion is that for general left-stochastic matrices
{A1, A0, A2}, the agents in the network will have their esti-
mators wk,i converge, in the mean-square-error sense, to the
same vector wo that corresponds to the unique solution of the
following algebraic equation:
N(cid:88)
k=1
N(cid:88)
k=1
N(cid:88)
k=1
pksk(w) = 0
(42)
For example, in the context of distributed optimization prob-
lems of the form (7), this result implies that for left-stochastic
matrices {A1, A0, A2}, the distributed strategies represented
by (1) -- (3) will not converge to the global minimizer of the
original aggregate cost (7), which is the unique solution to the
alternative algebraic equation
∇wJk(w) = 0
(43)
Instead, these distributed solutions will converge to the global
minimizer of the weighted aggregate cost J glob,(cid:63)(w) defined
by (32) in terms of the entries pk, i.e., to the unique solution
of
pk∇wJk(w) = 0
(44)
Result (42) also means that the distributed strategies (1) -- (3)
converge to a Pareto optimal solution of the multi-objective
problem (8); one Pareto solution for each selection of the
topology parameters {pk}. The distinction between the aggre-
gate costs J glob(w) and J glob,(cid:63)(w) does not appear in earlier
studies on distributed optimization [5] -- [11], [28] -- [30] mainly
because these studies focus on doubly-stochastic combination
matrices, for which the entries {pk} will all become equal to
each other for uniform step-sizes µk ≡ µ or µk(i) ≡ µ(i). In
that case, the minimizations of (7) and (32) become equivalent
and the solution of (43) and (44) would then coincide. In
other words, regardless of the choice of the doubly stochastic
combination weights, when the {pk} are identical, the limit
point will be unique and correspond to the solution of
sk(w) = 0
(45)
k=1
In contrast, result (42) shows that left-stochastic combination
policies add more flexibility into the behavior of the network.
By selecting different combination weights, or even different
topologies, the entries {pk} can be made to change and the
limit point can be steered towards other desired Pareto optimal
solutions. Even in the traditional case of consensus-type im-
plementations for computing averages, as opposed to learning
from streaming data, it also holds that it is beneficial to relax
the requirement of a doubly-stochastic combination policy in
order to enable broadcast algorithms without feedback [53].
The second major conclusion of the paper is that we will
show in (129) further ahead that there always exist sufficiently
small step-sizes such that the learning process over the network
N(cid:88)
8
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
at the same value, say, e.g., wk,0 = 0, then the learning
curves start at Transient Phase II directly.
• Transient Phase II:
In this phase, the trajectories of all agents are uniformly
close to the trajectory of the reference recursion; they
converge in a coordinated manner to steady-state. The
learning curves at this phase are well modeled by the
same reference recursion (47) since we will show in (145)
that:
E(cid:107) wk,i(cid:107)2 = (cid:107) wc,i(cid:107)2 + O(µ1/2
Furthermore, for small step-sizes and during the later
stages of this phase, ¯wc,i will be close enough to wo
and the convergence rate r will be shown to satisfy:
max) · γi
c + O(µmax)
(49)
r =(cid:2)ρ(IM − µmaxHc)(cid:3)2
+ O(cid:0)(µmax)
2(M−1)(cid:1)
1
(50)
where ρ(·) denotes the spectral radius of its matrix
argument, is an arbitrarily small positive number, and
Hc is the same matrix that results from evaluating (29)
at w = wo, i.e.,
pkHk = Hc(wo)
(51)
Hc (cid:44) N(cid:88)
where Hk (cid:44) ∇wT sk(wo).
• Steady-State Phase:
k=1
The reference recursion (47) continues converging to-
wards wo so that (cid:107) wc,i(cid:107)2 = (cid:107)wo − ¯wc,i(cid:107)2 will converge
to zero (−∞ dB in Fig. 2). However, for the distributed
the mean-square-error E(cid:107) wk,i(cid:107)2 =
strategy (1) -- (3),
E(cid:107)wo − wk,i(cid:107)2 at each agent k will converge to a finite
steady-state value. We will be able to characterize this
value in terms of the vector p in Part II [44] as follows:1
E(cid:107) wk,i(cid:107)2 = µmax·Tr(cid:8)X(pT⊗IM )Rv(p⊗IM )(cid:9)
lim
i→∞
+ o(µmax)
(52)
where X is the solution to the Lyapunov equation de-
scribed later in (41) of Part II [44] (when Σ = I),
and o(µmax) denotes a strictly higher order term of
µmax. Expression (52) is a revealing result. It is a non-
trivial extension of a classical result pertaining to the
mean-square-error performance of stand-alone adaptive
filters [54] -- [57] to the more demanding context when
a multitude of adaptive agents are coupled together in
a cooperative manner through a network topology. This
result has an important ramification, which we pursue
in Part II [44]. We will show there that no matter how
the agents are connected to each other, there is always
a way to select the combination weights such that the
performance of the network is invariant to the topology.
This will also imply that, for any connected topology,
there is always a way to select the combination weights
such that the performance of the network matches that of
the centralized solution.
1The interpretation of the limit in (52) is explained in more detail in Sec.
IV of Part II [44].
Fig. 2. A typical mean-square-error (MSE) learning curve includes a transient
stage that consists of two phases and a steady-state phase. The plot shows
how the learning curve of a network of agents compares to the learning curve
of a centralized reference solution. The analysis in this work, and in the
accompanying Part II [44] characterizes in detail the parameters that determine
the behavior of the network (rate, stability, and performance) during each
phase of the learning process.
is mean-square stable. This means that the weight error vectors
relative to wo will satisfy
lim sup
i→∞
E(cid:107) wk,i(cid:107)2 ≤ O(µmax)
(46)
so that the steady-state mean-square-error at each agent will
be of the order of O(µmax).
The third major conclusion of our analysis is that we will
show that, during the convergence process towards the limit
the learning curve at each agent exhibits three
point wo,
distinct phases: Transient Phase I, Transient Phase II, and
Steady-State Phase. These phases are illustrated in Fig. 2
and they are interpreted as follows. Let us first introduce
a reference (centralized) procedure that is described by the
following centralized-type recursion:
which is initialized at
¯wc,i = ¯wc,i−1 − µmax
N(cid:88)
¯wc,0 =
pksk( ¯wc,i−1)
(47)
k=1
θkwk,0
(48)
k=1
where θk is the kth entry of the eigenvector θ, µmax, and {pk}
are defined in Definitions 1 -- 2, wk,0 is the initial value of the
distributed strategy at agent k, and ¯wc,i is an M × 1 vector
generated by the reference recursion (47). The three phases
of the learning curve will be shown to have the following
features:
• Transient Phase I:
If agents are initialized at different values,
then the
estimates of the various agents will initially evolve in
such a way to make each wk,i get closer to the reference
recursion ¯wc,i. The rate at which the agents approach
¯wc,i will be determined by λ2(A), the second largest
eigenvalue of A in magnitude. If the agents are initialized
N(cid:88)
Phase&I&Phase&II&Steady,State&MSE&(dB)&Number&of&Itera9ons&&Reference&(centralized)&strategy&&&&&Steady,state&MSE&&&&&&Distributed&strategies&(49)&and&(50)&(52)& CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
9
Note that the above results are obtained for the general
distributed strategy (1) -- (3). Therefore, the results can be spe-
cialized to the consensus, CTA diffusion, and ATC diffusion
strategies in (10) -- (12) by choosing the matrices A1, A2, and
A0 according to Tab. I. The results in this paper and its
accompanying Part II [44] not only generalize the analysis
from earlier works [16] -- [20] but, more importantly, they also
provide deeper insights into the learning behavior of these
adaptation and learning strategies.
V. STUDY OF ERROR DYNAMICS
A. Error Quantities
We shall examine the learning behavior of the distributed
strategy (1) -- (3) by examining how the perturbation between
the distributed solution (1) -- (3) and the reference solution (47)
evolves over time -- see Fig. 3. Specifically, let wk,i denote
the discrepancy between wk,i and ¯wc,i, i.e.,
wk,i (cid:44) wk,i − ¯wc,i
(53)
and let wi and wi denote the global vectors that collect the
wk,i and wk,i from across the network, respectively:
wi (cid:44) col{w1,i, . . . , wN,i}
wi (cid:44) col{ w1,i, . . . , wN,i} = wi − 1 ⊗ ¯wc,i
(54)
(55)
It turns out that it is insightful to study the evolution of wi
in a transformed domain where it is possible to express the
distributed recursion (1) -- (3) as a perturbed version of the
reference recursion (47).
Definition 3 (Network basis transformation): We
define
the transformation by introducing the Jordan canonical
decomposition of the matrix A = A1A0A2. Let
AT = U DU
−1
(56)
where U is an invertible matrix whose columns correspond to
the right-eigenvectors of AT , and D is a block Jordan matrix
with a single eigenvalue at one with multiplicity one while
all other eigenvalues are strictly less than one. The Kronecker
form of A then admits the decomposition:
AT (cid:44) AT ⊗ IM = UDU
(57)
−1
where
U (cid:44) U ⊗ IM ,
D (cid:44) D ⊗ IM
We use U to define the following basis transformation:
w(cid:48)
w(cid:48)
i
i
(cid:44) U
(cid:44) U
−1wi = (U
−1 wi = (U
−1 ⊗ IM )wi
−1 ⊗ IM ) wi
(58)
(59)
(60)
The relations between the quantities in transformations (59) --
(60) are illustrated in Fig. 3(a).
We can gain useful insight into the nature of this transforma-
tion by exploiting more directly the structure of the matrices
−1. By Assumption 1, the matrix AT has an
U, D, and U
eigenvalue one of multiplicity one, with the corresponding
left- and right-eigenvectors being θT and 1, respectively. All
UR
(62)
(63)
(64)
(65)
UL (cid:44) UL ⊗ IM
UR (cid:44) UR ⊗ IM
DN−1 (cid:44) DN−1 ⊗ IM
other eigenvalues of D are strictly less than one in magnitude.
Therefore, the matrices D, U, and U−1 can be partitioned as
(cid:35)
U =(cid:2) 1 UL
(cid:3) U
−1 =
(cid:35)
(cid:34)
θT
UR
(cid:34)
D =
1
DN−1
(61)
where DN−1 is an (N − 1) × (N − 1) Jordan matrix with all
diagonal entries strictly less than one in magnitude, UL is an
(cid:35)
N × (N − 1) matrix, and UR is an (N − 1)× N matrix. Then,
the Kronecker forms D, U, and U
D =
−1 can be expressed as
θT ⊗IM
, U =(cid:2) 1⊗IM UL
(cid:3) , U
−1 =
(cid:34)
(cid:34)
(cid:35)
IM
DN−1
where
N(cid:88)
k=1
It is important to note that U−1U = IN and that
θT 1 = 1,
We first study the structure of w(cid:48)
(cid:125)
w(cid:48)
i = col{(θT ⊗ IM )wi
θT UL = 0, UR1 = 0, URUL = IN−1 (66)
i defined in (59) using (61):
(67)
, (UR ⊗ IM )wi
(cid:123)(cid:122)
(cid:123)(cid:122)
(cid:124)
(cid:124)
(cid:125)
}
(cid:44)wc,i
(cid:44)we,i
The two components wc,i and we,i have useful interpretations.
Recalling that θk denotes the kth entry of the vector θ, then
wc,i can be expressed as
wc,i =
θkwk,i
(68)
As we indicated after Assumption 1, the entries {θk} are
positive and add up to one. Therefore, wc,i is a weighted
average (i.e., the centroid) of the estimates {wk,i} across all
agents. To interpret we,i, we examine the inverse mapping of
(59) from w(cid:48)
i to wi using the block structure of U in (61):
wi = (U ⊗ IM )w(cid:48)
i
= (1 ⊗ IM )wc,i + (UL ⊗ IM )we,i
= 1 ⊗ wc,i + (UL ⊗ IM )we,i
the individual estimates at
(69)
the various
which implies that
agents satisfy:
wk,i = wc,i + (uL,k ⊗ IM )we,i
(70)
where uL,k denotes the kth row of the matrix UL. The network
basis transformation defined by (59) represents the cluster of
iterates {wk,i} by its centroid wc,i and their positions {uL,k⊗
IM )we,i} relative to the centroid as shown in Fig. 3. The two
parts, wc,i and we,i, of w(cid:48)
i in (67) are the coordinates in this
new transformed representation. Then, the actual error quantity
wk,i relative to wo can be represented as
wk,i = wo − ¯wc,i − (wk,i − ¯wc,i)
= wo − ¯wc,i − (wc,i + (uL,k ⊗ IM )we,i − ¯wc,i) (71)
10
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
(a)
(b)
Fig. 3.
centroid, wc,i, of the N iterates across the network.
(a) Network basis transformation. (b) The diagrams show how the iterate wk,i is decomposed relative to the reference ¯wc,i and relative to the
Introduce
(72)
(73)
Then, from (71) we arrive at the following critical relation for
our analysis in the sequel:
wc,i (cid:44) wo − ¯wc,i
wc,i (cid:44) wc,i − ¯wc,i
wk,i = wc,i − wc,i − (uL,k ⊗ IM )we,i
(74)
This relation is also illustrated in Fig. 3. Then, the behavior
of the error quantities { wk,i} can be studied by examining
wc,i, wc,i and we,i, respectively, which is pursued in Sec.
VI further ahead. The first term is the error between the
reference recursion and wo, which is studied in Theorems 1 --
3. The second quantity is the difference between the weighted
centroid wc,i of the cluster and the reference vector ¯wc,i, and
the third quantity characterizes the positions of the individual
iterates {wk,i} relative to the centroid wc,i. As long as the
second and the third terms in (74), or equivalently, wc,i
and we,i, are small (which will be shown in Theorem 4),
the behavior of each wk,i can be well approximated by the
behavior of the reference vector ¯wc,i. Indeed, wc,i and we,i
are the coordinates of the transformed vector w(cid:48)
i defined by
(60). To see this, we substitute (61) and (55) into (60) to get
−1 ⊗ IM )(wi − 1 ⊗ ¯wc,i)
i − (U
−11 = col{θT 1, UR1}
= col{1, 0N−1}
(75)
Recalling (66) and the expression for U−1 in (61), we obtain
(76)
where 0N−1 denotes an (N−1)×1 vector with all zero entries.
Substituting (76) and (67) into (75), we get
(77)
w(cid:48)
i = col{wc,i − ¯wc,i, we,i} = col{ wc,i, we,i}
w(cid:48)
i = (U
= w(cid:48)
U
−11) ⊗ ¯wc,i
Therefore, it suffices to study the dynamics of w(cid:48)
i and its mean-
square performance. We will establish joint recursions for wc,i
and we,i in Sec. V-B, and joint recursions for wc,i and we,i
in Sec. V-C. Table II summarizes the definitions of the various
quantities, the recursions that they follow, and their relations.
B. Signal Recursions
We now derive the joint recursion that describes the evolu-
tion of the quantities wc,i = wc,i − ¯wc,i and we,i. Since ¯wc,i
follows the reference recursion (47), it suffices to derive the
joint recursion for wc,i and we,i. To begin with, we introduce
the following global quantities:
A = A ⊗ IM
A0 = A0 ⊗ IM
A1 = A1 ⊗ IM
A2 = A2 ⊗ IM
M = Ω ⊗ IM
Ω = diag{µ1, . . . , µN}
(78)
(79)
(80)
(81)
(82)
(83)
We also let the notation x = col{x1, . . . , xN} denote an
arbitrary N×1 block column vector that is formed by stacking
M×1 sub-vectors x1, . . . , xN on top of each other. We further
define the following global update vectors:
si(x) (cid:44) col{s1,i(x1), . . . , sN,i(xN )}
s(x) (cid:44) col{s1(x1), . . . , sN (xN )}
(84)
(85)
Then, the general recursion for the distributed strategy (1) --
(3) can be rewritten in terms of these extended quantities as
follows:
wi = AT wi−1 − AT
2 Msi(φi−1)
(86)
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
11
SUMMARY OF VARIOUS ITERATES, ERROR QUANTITIES, AND THEIR RELATIONS.
TABLE II
Quantity
wk,i
wk,i
wk,i
Original system
Definition
Iterate at agent k
wo−wk,i wk,i− ¯wc,i
a The transformation is defined by (59) -- (60).
Eqs. (1) -- (3)
Recursion
where
φi (cid:44) col{φ1,i, . . . , φN,i}
and is related to wi and w(cid:48)
φi = AT
i via the following relation
1 wi = AT
1 Uw(cid:48)
i
(87)
(88)
Applying the transformation (59) to both sides of (86), we
obtain the transformed global recursion:
i = Dw(cid:48)
w(cid:48)
i−1 − U
−1AT
2 Msi (φi−1)
(89)
We can now use the block structures in (62) and (67) to derive
recursions for wc,i and we,i from (89). Substituting (62) and
(67) into (89), and using properties of Kronecker products [58,
p.147], we obtain
wc,i = wc,i−1 − (θT ⊗ IM )AT
2 Msi (φi−1)
= wc,i−1 − (θT AT
2 Ω ⊗ IM )si (φi−1)
= wc,i−1 − µmax · (pT ⊗ IM )si (φi−1)
and
we,i = DN−1we,i−1 − URAT
2 Msi (φi−1)
where in the last step of (90) we used the relation
µmax · p = ΩA2θ
(90)
(91)
(92)
which follows from Definitions 1 and 2. Furthermore, by
adding and subtracting identical factors, the term si (φi−1)
that appears in (90) and (91) can be expressed as
(cid:125)
si (φi−1) = s(1 ⊗ wc,i−1) + si (φi−1)−s (φi−1)
(cid:123)(cid:122)
(cid:125)
(cid:124)
+ s (φi−1)−s(1 ⊗ wc,i−1)
(cid:44)vi(φi−1)
(cid:123)(cid:122)
(cid:124)
(cid:44)zi−1
(93)
where the first perturbation term vi(φi−1) consists of the dif-
ference between the true update vectors {sk(φk,i−1)} and their
stochastic approximations {sk,i(φk,i−1)}, while the second
perturbation term zi−1 represents the difference between the
same {sk(φk,i−1)} and {sk(wc,i−1)}. The subscript i − 1 in
zi−1 implies that this variable depends on data up to time i−1
and the subscript i in vi(φi−1) implies that its value depends
on data up to time i (since, in general, si(·) can depend on
data from time i -- see Eq. (33) in Part II for an example).
Then, si (φi−1) can be expressed as
si (φi−1) = s(1 ⊗ wc,i−1) + vi + zi−1
(94)
Transformed systema
N(cid:88)
wc,i
wc,i
θkwk,i wc,i− ¯wc,i
we,i
URwi
k=1
Eq. (96)
Eq. (103)
Eq. (104)
Eq. (47)
Reference system
wc,i
¯wc,i
Ref. Iterate
wo− ¯wc,i
Lemma 2 (Signal dynamics): In summary,
the previous
derivation shows that the weight iterates at each agent evolve
according to the following dynamics:
wk,i = wc,i + (uL,k ⊗ IM )we,i
wc,i = wc,i−1 − µmax · (pT ⊗ IM )si (φi−1)
we,i = DN−1we,i−1 − URAT
2 Msi (φi−1)
si (φi−1) = s(1 ⊗ wc,i−1) + vi + zi−1
(95)
(96)
(97)
(98)
C. Error Dynamics
To simplify the notation, we introduce the centralized op-
erator Tc : RM → RM as the following mapping for any
x ∈ RM :
Tc(x) (cid:44) x − µmax · (pT ⊗ IM ) s(1 ⊗ x)
= x − µmax
pksk(x)
(99)
N(cid:88)
k=1
Substituting (94) into (96) -- (97) and using (99), we find that
we can rewrite (96) and (97) in the alternative form:
wc,i = Tc(wc,i−1) − µmax · (pT ⊗ IM ) [zi−1 + vi]
we,i = DN−1we,i−1 − URAT
(100)
2 M [s(1 ⊗ wc,i−1) + zi−1 +vi]
(101)
Likewise, we can write the reference recursion (47) in the
following compact form:
¯wc,i = Tc( ¯wc,i−1)
(102)
Comparing (100) with (102), we notice that the recursion for
the centroid vector, wc,i, follows the same update rule as the
reference recursion except for the two driving perturbation
terms zi−1 and vi. Therefore, we would expect the trajectory
of wc,i to be a perturbed version of that of ¯wc,i. Recall from
(73) that
wc,i (cid:44) wc,i − ¯wc,i
To obtain the dynamics of wc,i, we subtract (102) from (100).
Lemma 3 (Error dynamics): The error quantities that ap-
pear on the right-hand side of (77) evolve according to the
following dynamics:
wc,i = Tc(wc,i−1) − Tc( ¯wc,i−1)
− µmax·(pT ⊗IM ) [zi−1 + vi]
we,i = DN−1we,i−1
− URAT
2 M [s(1 ⊗ wc,i−1)+zi−1 +vi]
(103)
(104)
12
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
The analysis in sequel will study the dynamics of the variances
of the error quantities wc,i and we,i based on (103) -- (104). The
main challenge is that these two recursions are coupled with
each other through zi−1 and vi. To address the difficulty, we
will extend the energy operator approach developed in [20] to
the general scenario under consideration.
D. Energy Operators
To carry out the analysis, we need to introduce the following
operators.
Definition 4 (Energy vector operator): Suppose
x = col{x1, . . . , xN} is an arbitrary N × 1 block
column vector that is formed by stacking M0 × 1 vectors
x1, . . . , xN on top of each other. The energy vector operator
PM0 : CM0N → RN is defined as the mapping:
PM0 [x] (cid:44) col{(cid:107)x1(cid:107)2, . . . ,(cid:107)xN(cid:107)2}
(105)
where (cid:107) · (cid:107) denotes the Euclidean norm of a vector.
Definition 5 (Norm matrix operator): Suppose X is an ar-
bitrary K × N block matrix consisting of blocks {Xkn} of
size M0 × M0:
X =
(106)
The norm matrix operator ¯PM0 : CM0K×M0N → RK×N is
defined as the mapping:
··· X1N
...
··· XKN
...
XK1
X11
(cid:107)X11(cid:107)
...
the individual agents, {wk,i}, get to the reference recursion.
Recalling from (77) that wc,i and we,i are the two blocks of
the transformed vector w(cid:48)
i defined by (60), we can examine
instead the evolution of
(cid:44) EP [ w(cid:48)
i] = col{EP [ wc,i], EP [we,i]}
= col(cid:8)E(cid:107) wc,i(cid:107)2, EP [we,i](cid:9)
(108)
W
Specifically, we will study the convergence of wc,i in Sec.
(cid:48)
VI-A, the stability of W
i in Sec. VI-B, and the two transient
phases of wk,i in Sec. VI-C.
(cid:48)
i
A. Limit Point
(cid:48)
Before we proceed to study W
i, we state the following the-
orems on the existence of a limit point and on the convergence
of the reference recursion (102).
Theorem 1 (Limit point): Given Assumptions 3 -- 4,
there
exists a unique M × 1 vector wo that solves
pksk(wo) = 0
(109)
N(cid:88)
k=1
where pk is the kth entry of the vector p defined in (16).
Proof: See Appendix E.
Theorem 2 (Convergence of the reference recursion): Let
wc,i (cid:44) wo − ¯wc,i denote the error vector of the reference
recursion (102). Then, the following non-asymptotic bound
on the squared error holds for all i ≥ 0:
(1−2µmax(cid:107)p(cid:107)1λU )i·(cid:107) wc,0(cid:107)2 ≤ (cid:107) wc,i(cid:107)2 ≤ γ2i
c ·(cid:107) wc,0(cid:107)2
(110)
¯PM0[x] (cid:44)
···
(cid:107)X1N(cid:107)
...
(cid:107)XK1(cid:107)
··· (cid:107)XKN(cid:107)
where
(107)
γc (cid:44) 1 − µmaxλL +
1
2
µ2
max(cid:107)p(cid:107)2
1λ2
U
(111)
Furthermore, if the following condition on the step-size holds
where (cid:107) · (cid:107) denotes the 2−induced norm of a matrix.
By default, we choose M0 to be M, the size of the vector
wk,i. In this case, we will drop the subscript inPM0[·] and use
P [·] for convenience. However, in other cases, we will keep the
subscript to avoid confusion. Likewise, ¯PM0 [·] characterizes
the norms of different parts of a matrix it operates on. We
will also drop the subscript if M0 = M.
In Appendix A,
we collect several properties of the above energy operators,
and which will be used in the sequel to characterize how the
energy of the error quantities propagates through the dynamics
(103) -- (104).
0 < µmax <
2λL
(cid:107)p(cid:107)2
1λ2
U
then, the iterate wc,i converges to zero.
Proof: See Appendix F.
(112)
Note from (110) that, when the step-size is sufficiently
small, the reference recursion (47) converges at a geometric
rate between 1 − 2µmax(cid:107)p(cid:107)1λU and γ2
c = 1 − 2µmaxλL +
o(µmax). Note that this is a non-asymptotic result. That is,
the convergence rate rRef of the reference recursion (102) is
always lower and upper bounded by these two rates:
VI. TRANSIENT ANALYSIS
Using the energy operators and the various properties, we
can now examine the transient behavior of the learning curve
more closely. Recall from (74) that wk,i consists of three
parts: the error of the reference recursion, wc,i, the difference
between the centroid and the reference, wc,i, and the position
of individual iterates relative to the centroid, (uL,k⊗ IM )we,i.
The main objective in the sequel is to study the convergence of
the reference error, wc,i, and establish non-asymptotic bounds
for the mean-square values of wc,i and we,i, which will
allow us to understand how fast and how close the iterates at
1 − 2µmax(cid:107)p(cid:107)1λU ≤ rRef ≤ γ2
c ,
∀i ≥ 0
(113)
We can obtain a more precise characterization of the conver-
gence rate of the reference recursion in the asymptotic regime
(for large enough i), as follows.
Theorem 3 (Convergence rate of the reference recursion):
Specifically, for any small > 0, there exists a time instant
i0 such that, for i ≥ i0, the error vector wc,i converges to
zero at the following rate:
rRef =(cid:2)ρ(IM − µmaxHc)(cid:3)2
+ O(cid:0)(µmax)
2(M−1)(cid:1)
1
(114)
Proof: See Appendix G.
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
13
(cid:48)
i): Suppose
Theorem 4 (Non-asymptotic bound for W
which is O(µmax). This insight is exploited to establish a non-
(cid:48)
i = col{E(cid:107) wc,i(cid:107)2, EP [we,i]} in the
asymptotic bound on W
following theorem.
the
matrix Γ defined in (116) is stable, i.e., ρ(Γ) < 1. Then, the
following non-asymptotic bound holds for all i ≥ 0:
EP [ wc,i] (cid:22) µmaxhc(µmax)·1T (γcI−Γe)
cI−Γi
(cid:48)
+ W ub
c,∞
(cid:48)
EP [we,i] (cid:22) Γi
eWe,0 + W ub
e,∞
where We,0 (cid:44) EP [we,0], W ub
bounds of EP [ wc,i] and EP [we,i], respectively:
e,∞ are the lim sup
c,∞ and W ub
−1(cid:0)γi
We,0
(124)
(125)
(cid:1)
e
(cid:48)
(cid:48)
(cid:0)λL + hc(0)(cid:1)1T (I − Γe)−1We,0 + bv,cλL
(cid:0)λL + hc(0)(cid:1)1T (I − Γe)−1We,0 + bv,eλL
(126)
λ2
L
(cid:48)
W ub
c,∞ = µmax ·
ψ0
(cid:48)
e,∞ = µ2
W ub
−11 + o(µ2
max)
(127)
where o(·) denotes strictly higher order terms, and hc(0) is
the value of hc(µmax) (see (121)) evaluated at µmax = 0. An
important implication of (124) and (126) is that
∀i ≥ 0
EP [ wc,i] ≤ O(µmax),
Furthermore, a sufficient condition that guarantees the stability
of the matrix Γ is that
(128)
+ o(µmax)
ψ0
max·
× (I−Γe)
λL
(cid:26)
(cid:115)
1
λL
(cid:16)1−λ2(A)
1λ2
3 ψ0
U + 1
2(cid:107)p(cid:107)2
3(1−λ2(A))2N+1
2
,
22N +2ψ0
λL
(cid:0)
(cid:17)−2N ,
(cid:1)(cid:27)
(129)
(130)
(cid:107) ¯P1[AT UL](cid:107)2∞ + 1
Corollary 1 (Asymptotic bounds): It holds that
Proof: See Appendix J.
(cid:107)p(cid:107)2
1λ2
U
2
lim sup
i→∞
lim sup
i→∞
E(cid:107) wc,i(cid:107)2 ≤ O(µmax)
E(cid:107)we,i(cid:107)2 ≤ O(µ2
max)
(131)
Proof: The bound (130) holds since E(cid:107) wc,i(cid:107)2 = EP [ wc,i] ≤
O(µmax) for all i ≥ 0 according to (128). Furthermore,
inequality (131) holds because
E(cid:107)we,i(cid:107)2 (a)
1T EP [we,i]
lim sup
i→∞
Note that since (114) holds for arbitrary > 0, we can choose
to be an arbitrarily small positive number. Therefore, the
convergence rate of the reference recursion is arbitrarily close
to [ρ(IM − µmaxHc)]2.
B. Mean-Square Stability
inequality recursion for the transformed energy vector W
EP [ w(cid:48)
Now we apply the properties from Lemmas 5 -- 6 to derive an
(cid:48)
i =
i]. The results are summarized in the following lemma.
(cid:48)
Lemma 4 (Inequality recursion for W
i): The N × 1 vector
(cid:48)
i defined by (108) satisfies the following relation for all
W
time instants:
(cid:48)
i (cid:22) Γ W
W
(cid:48)
i−1 + µ2
maxbv
(115)
(116)
(117)
(cid:26)
(cid:34)
(cid:35)
maxψ0 · 11T ∈ RN×N
where
Γ (cid:44) Γ0 + µ2
γc µmaxhc(µmax) · 1T
Γ0 (cid:44)
0
bv (cid:44) col{bv,c, bv,e · 1} ∈ RN
...
...
λ2(A)
1−λ2(A)
Γe (cid:44)
...
Γe
2
∈ RN×N
2
(118)
1−λ2(A)
λ2(A)
∈ R(N−1)×(N−1)
1 UL](cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [AT
(cid:18)
1, 4α(cid:107)p(cid:107)2
1 ·
1 UL](cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [AT
∞ λ2
(cid:19)(cid:27)
(cid:104)
1 − λ2(A)
∞ λ2
α
λ2
U
(cid:19)
(119)
(120)
(cid:105)
∞ ,
+
U
U
1
,
1
3
(cid:18)
ψ0 (cid:44) max
4α(cid:107)p(cid:107)2
4N ·
4N ·
·
1 − λ2(A)
hc(µmax) (cid:44) (cid:107)p(cid:107)2
1·
1 UL]
bv,c (cid:44) (cid:107)p(cid:107)2
1 ·
2 ](cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [URAT
2 ](cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [AT
(cid:2)4α((cid:107) wc,0(cid:107)2 + (cid:107)wo(cid:107)2) + σ2
(cid:18)
bv,e (cid:44) N(cid:13)(cid:13) ¯P [URAT
2 ](cid:13)(cid:13)2
(cid:19)
∞ ·
α
+
λ2
U
∞ λ2
U ·
λL− 1
12
∞
v
+ 4α((cid:107) wc,0(cid:107)2 + (cid:107)wo(cid:107)2) + σ2
v
2 µmax(cid:107)p(cid:107)2
(cid:3)
1λ2
U
(121)
(122)
U(cid:107) wc,0(cid:107)2 + (cid:107)go(cid:107)∞
λ2
1 − λ2(A)
(123)
where go (cid:44) P [s(1 ⊗ wo)].
Proof: See Appendix H.
From (116) -- (117), we see that as the step-size µmax be-
comes small, we have Γ ≈ Γ0, since the second term in
the expression for Γ depends on the square of the step-
size. Moreover, note that Γ0 is an upper triangular matrix.
Therefore, wc,i and we,i are weakly coupled for small step-
sizes; EP [we,i] evolves on its own, but
it will seep into
the evolution of EP [ wc,i] via the off-diagonal term in Γ0,
The scalars ψ0, hc(µ), bv,c and bv,e are defined as
0 < µmax < min
(b)
= lim sup
i→∞
(cid:48)
(cid:22) 1T W ub
e,∞
(c)
= O(µ2
max)
(132)
where step (a) uses property (157) of the energy operator P [·],
step (b) uses (125), and step (c) uses (127).
Finally, we present following main theorem that character-
izes the difference between the learning curve of wk,i at each
agent k and that of wc,i generated by the reference recursion
(102).
14
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
Theorem 5 (Learning behavior of E(cid:107) wk,i(cid:107)2): Suppose the
stability condition (129) holds. Then, the difference between
the learning curve of the mean-square-error E(cid:107) wk,i(cid:107)2 at each
agent k and the learning curve of (cid:107) wc,i(cid:107)2 is bounded non-
asymptotically as
(cid:12)(cid:12)E(cid:107) wk,i(cid:107)2 − (cid:107) wc,i(cid:107)2(cid:12)(cid:12)
(cid:113)
≤ 2(cid:107)uL,k ⊗ IM(cid:107)2 · 1T Γi
eWe,0
+ 2(cid:107) wc,0(cid:107) · (cid:107)uL,k ⊗ IM(cid:107) ·
+ γi
max) + O(µmax)
1
2
c · O(µ
where γc was defined earlier in (111).
Proof: See Appendix L.
C. Interpretation of Results
1T Γi
eWe,0
for all i ≥ 0 (133)
The result established in Theorem 5 is significant because
it allows us to examine the learning behavior of E(cid:107) wk,i(cid:107)2.
First, note that the bound (133) established in Theorem 5 is
non-asymptotic; that is, it holds for any i ≥ 0. Let e1(i), e2(i),
e3(i) and e4 denote the four terms in (133):
(cid:113)
eWe,0
e1(i) (cid:44) 2(cid:107)uL,k ⊗ IM(cid:107)2 · 1T Γi
e2(i) (cid:44) 2(cid:107) wc,0(cid:107) · (cid:107)uL,k ⊗ IM(cid:107) ·
e3(i) (cid:44) γi
c · O(µ
e4 (cid:44) O(µmax)
max)
1
2
1T Γi
eWe,0
(134)
(135)
(136)
(137)
is,
Then, inequality (133) can be rewritten as
(cid:107) wc,i(cid:107)2−[e1(i) + e2(i) + e3(i) + e4]
≤ E(cid:107) wk,i(cid:107)2 ≤ (cid:107) wc,i(cid:107)2 +[e1(i)+e2(i)+e3(i)+e4]
(138)
the learning curve of E(cid:107) wk,i(cid:107)2
for all i ≥ 0. That
is a perturbed version of the learning curve of reference
recursion (cid:107) wc,i(cid:107)2, where the perturbation consists of four
parts: e1(i), e2(i), e3(i) and e4. We now examine the non-
asymptotic convergence rates of these four perturbation terms
relative to that of the reference term (cid:107) wc,i(cid:107)2 to show how
the learning behavior of E(cid:107) wk,i(cid:107)2 can be approximated by
(cid:107) wc,i(cid:107)2. From their definitions in (134) -- (135), we note that
e1(i) and e2(i) converge to zero at the non-asymptotic rates
of ρ(Γe) and [ρ(Γe)]
2 over i ≥ 0. According to (119), we
have ρ(Γe) = λ2(A), the magnitude of the second largest
eigenvalue of A. Let re1 and re2 denote the convergence rates
of e1(i) and e2(i). We have
1
1
2
re2 = λ2(A)
re1 = λ2(A),
(139)
By Assumption 1 and (13), the value of λ2(A) is strictly less
than one, and is determined by the network connectivity and
the design of the combination matrix; it is also independent
of the step-size parameter µmax. For this reason, the terms
e1(i) and e2(i) converge to zero at rates that are determined
by the network and are independent of µmax. Furthermore, the
term e3(i) is always small, i.e., O(µ
max), for all i ≥ 0, and
converges to zero as i → ∞. The last term e4 is also small,
namely, O(µmax). On the other hand, as revealed by Theorem
1
2
(141)
(142)
(143)
2 and (113), the non-asymptotic convergence rate of the term
(cid:107) wc,i(cid:107)2 in (138) is bounded by
1 − 2µmax(cid:107)p(cid:107)1λU ≤ rRef ≤ γ2
c = 1 − 2µmaxλL + o(µmax)
(140)
Therefore, as long as µmax is small enough so that2
λ2(A) < λ2(A)
2 < 1 − 2µmax(cid:107)p(cid:107)1λU
1
which is equivalent to requiring
µmax <
1
2
1 − λ2(A)
2(cid:107)p(cid:107)1λU
we have the following relation regarding the non-asymptotic
rates of e1(i), e2(i) and (cid:107) wc,i(cid:107)2:
re1 < re2 < rRef
This means that, for sufficiently small µmax satisfying (142),
e1(i) and e2(i) converge faster than (cid:107) wc,i(cid:107)2. For this reason,
the perturbation terms e1(i) and e2(i) in (138) die out earlier
than (cid:107) wc,i(cid:107)2. When they are negligible compared to (cid:107) wc,i(cid:107)2,
we reach the end of Transient Phase I. Then, in Transient
Phase II, we have
(cid:107) wc,i(cid:107)2−[e3(i) + e4] ≤ E(cid:107) wk,i(cid:107)2 ≤ (cid:107) wc,i(cid:107)2 + [e3(i) + e4]
(144)
By (136) and (137), the above inequality (144) is equivalent
to the following relation:
1
2
c + O(µmax)
max) · γi
E(cid:107) wk,i(cid:107)2 = (cid:107) wc,i(cid:107)2 + O(µ
(145)
This means that E(cid:107) wk,i(cid:107)2 is close to (cid:107) wc,i(cid:107)2 in Transient
Phase II, and the convergence rate of E(cid:107) wk,i(cid:107)2 is the same
as that of (cid:107) wc,i(cid:107)2. Furthermore, in the later stage of Transient
Phase II, the convergence rate of E(cid:107) wk,i(cid:107)2 would be close
to the asymptotic rate of (cid:107) wc,i(cid:107)2 given by (114). Afterwards,
c converge to zero
as i → ∞, both (cid:107) wc,i(cid:107)2 and O(µ
and the only term remaining will be e4 = O(µmax), which
contributes to the steady-state MSE. More specifically, taking
the lim sup of both sides of (133) leads to
max) · γi
1
2
(cid:12)(cid:12)E(cid:107) wk,i(cid:107)2 − (cid:107) wc,i(cid:107)2(cid:12)(cid:12)
lim sup
i→∞
E(cid:107) wk,i(cid:107)2 = lim sup
i→∞
≤ O(µmax)
(146)
We will go a step further and evaluate this steady-state MSE
in closed-form for small step-sizes in Part II [44]. Therefore,
wk,i converges to wo with a small steady-state MSE that is
on the order of O(µmax). And the steady-state MSE can be
made arbitrarily small for small step-sizes.
In view of this discussion, we now recognize that the results
established in Theorems 1 -- 4 reveal the evolution of the three
components, wc,i, wc,i and we,i in (74) during three distinct
phases of the learning process. From (124), the centroid wc,i
of the distributed algorithm (1) -- (3) stays close to ¯wc,i over the
entire time for sufficiently small step-sizes since the mean-
square error E(cid:107)wc,i − ¯wc,i(cid:107)2 = EP [ wc,i] is always of the
order of O(µmax). However, We,0 = EP [we,i] in (125) is
2λ2(A) < λ2(A) 1
2 always holds since λ2(A) is strictly less than one.
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
15
BEHAVIOR OF ERROR QUANTITIES IN DIFFERENT PHASES.
TABLE III
Transient Phase I
Convergence rate r a
1 − 2µmax(cid:107)p(cid:107)1λU ≤ r ≤ γ2
Error quantity
(cid:107) wc,i(cid:107)2
E(cid:107) wc,i(cid:107)2
EP [ we,i]
E(cid:107) wk,i(cid:107)2
a γc is defined in (111), and γ2
b We only show the leading term of the convergence rate for r. More precise expression can be found in (114).
c Closer studies of the steady-state performance can be found in Part II [44].
Value
c (cid:29) O(µmax)
O(µmax)
(cid:29) O(µmax)
(cid:29) O(µmax)
converged
r ≤ λ2(A)
Multiple modes
c = 1 − 2µmaxλL + o(µmax).
Transient Phase II
Value
Convergence rate r
b
r = [ρ(IM − µmaxHc)]2 (cid:29) O(µmax)
O(µmax)
O(µ2
max)
r = [ρ(IM − µmaxHc)]2 (cid:29) O(µmax)
converged
converged
Steady-State c
Value
0
O(µmax)
O(µ2
max)
O(µmax)
(a)
(b)
Fig. 4. The evolution and learning curves of various quantities in a diffusion LMS adaptive network (best viewed in color), where M = 2, and the regressors
are spatially and temporally white, and isotropic across agents. (a) The evolution of the iterates {wk,i} at all agents, the centroid wc,i, and the reference
recursion ¯wc,i on the two-dimensional solution space; the horizontal axis and vertical axis are the first and second elements of wk,i, respectively. The clusters
of {wk,i} are plotted every 50 iterations. (b) The MSE learning curves, averaged over 1000 trials, for the iterates {wk,i} at all agents, and the reference
recursion ¯wc,i. The zoom-in region shows the learning curves for different agents, which quick shrink together in Phase I.
not necessarily small at the beginning. This is because, as we
pointed out in (70) and Fig. 3, we,i characterizes the deviation
of the agents from their centroid. If the agents are initialized
at different values, then EP [we,0] (cid:54)= 0, and it takes some time
for EP [we,i] to decay to a small value of O(µ2
max). By (125),
the rate at which EP [we,i] decays is ρ(Γe) = λ2(A). On
the other hand, recall from Theorems 2 -- 3 that the error of
the reference recursion, wc,i converges at a rate between 1 −
2µmax(cid:107)p(cid:107)1λU and r2
c = 1− 2µmaxλL + o(λmax) at beginning
and then [ρ(IM −µmaxHc)]2 later on, which is slower than the
convergence rate of EP [we,i] for small step-size µmax. Now,
returning to relation (74):
wk,i = wc,i − wc,i − (uL,k ⊗ IM )we,i
(147)
this means that during the initial stage of adaptation, the third
max) at a faster rate than the first
term in (147) decays to O(µ2
term, although wc,i will eventually converge to zero. Recalling
from (70) and Fig. 3 that we,i characterizes the deviation of the
agents from their centroid, the decay of we,i implies that the
agents are coordinating with each other so that their estimates
wk,i are close to the same wc,i -- we call this stage Transient
Phase I. Moreover, as we just pointed out, the term EP [ wc,i] is
O(µmax) over the entire time domain so that the second term
in (147) is always small. This also means that the centroid of
the cluster in Fig. 3, i.e., wc,i, is always close to the reference
recursion ¯wc,i since wc,i = wc,i − ¯wc,i
is always small.
Now that E(cid:107) wc,i(cid:107)2 is O(µmax) and EP [we,i] is O(µ2
max),
the error wk,i at each agent k is mainly dominated by the
first term, wc,i, in (147), and the estimates {wk,i} at different
agents converge together at the same rate as the reference
recursion, given by (114), to steady-state -- we call this stage
Transient Phase II. Furthermore, if We,0 = 0, i.e., the iterates
wk,i are initialized at the same value (e.g., zero vector), then
(125) shows that EP [we,i] is O(µ2
max) over the entire time
domain so that the learning dynamics start at Transient Phase
II directly. Finally, all agents reach the third phase, steady-
state, where wc,i → 0 and wk,i is dominated by the second
and third terms in (147) so that E(cid:107) wk,i(cid:107)2 becomes O(µmax).
We summarize the above results in Table III and illustrate
the evolution of the quantities in the simulated example in
Fig. 4. We observe from Fig. 4 that the radius of the cluster
shrinks quickly at the early stage of the transient phase, and
then converges towards the optimal solution.
VII. CONCLUSION
In this work, we studied the learning behavior of adaptive
networks under fairly general conditions. We showed that, in
the constant and small step-size regime, a typical learning
curve of each agent exhibits three phases: Transient Phase I,
Transient Phase II, and Steady-state Phase. A key observation
−2024681012−20246810121416The first element of the estimated vectorThe second element of the estimated vector Optimal valueDiffusion initial iteratesDiffusion iteratesReference recursionCentroidi=0i=50i=100...i=15002004006008001000-20-1001020Number of iterationsMean square error (dB) Reference recursion05010015020010152025Number of iterationsMean-square-error (dB) 16
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
is that, the second and third phases approach the performance
of a centralized strategy. Furthermore, we showed that the right
eigenvector of the combination matrix corresponding to the
eigenvalue at one influences the limit point, the convergence
rate, and the steady-state mean-square-error (MSE) perfor-
mance of the distributed optimization and learning strategies
in a critical way. Analytical expressions that illustrate these
effects were derived. Various implications were discussed and
illustrative examples were also considered.
APPENDIX A
PROPERTIES OF THE ENERGY OPERATORS
In this appendix, we state lemmas on properties of the
operators PM0[·] and ¯PM0 [·]. We begin with some basic
properties.
Lemma 5 (Basic properties): Consider N ×1 block vectors
x = col{x1, . . . , xN} and y = col{y1, . . . , yN} with M × 1
entries {xk, yk}. Consider also the K × N block matrix X
with blocks of size M × M. Then, the operators P [·] and ¯P [·]
satisfy the following properties:
1) (Nonnegativity): P [x] (cid:23) 0, ¯P [X] (cid:23) 0.
2) (Scaling): For any scalar a ∈ C, we have P [ax] =
a2P [x] and ¯P [aX] = a · ¯P [X].
3) (Convexity):
are N × 1
block vectors formed in the same manner as x,
X (1), . . . , X (K) are K×N block matrices formed in the
same manner as X, and let a1, . . . , aK be non-negative
real scalars that add up to one. Then,
suppose x(1), . . . , x(K)
P [a1x(1) + ··· + aKx(K)]
¯P [a1X (1) + ··· + aKX (K)]
(cid:22) a1P [x(1)] + ··· + aKP [x(K)]
(148)
(cid:22) a1 ¯P [X (1)] + ··· + aK ¯P [X (K)]
(149)
4) (Additivity): Suppose x = col{x1, . . . , xN} and y =
col{y1, . . . , yN} are N × 1 block random vectors that
satisfy Ex∗
kyk = 0 for k = 1, . . . , N, where ∗ denotes
complex conjugate transposition. Then,
EP [x + y] = EP [x] + EP [y]
(150)
5) (Triangular inequality): Suppose X and Y are two K×
N block matrices of same block size M. Then,
¯P [X + Y ] (cid:22) ¯P [X] + ¯P [Y ]
(151)
6) (Submultiplicity): Suppose X and Z are K×N and N×
L block matrices of the same block size M, respectively.
Then,
¯P [XZ] (cid:22) ¯P [X] ¯P [Z]
and b ∈ CM . Then,
(152)
7) (Kronecker structure): Suppose X ∈ CK×N , a ∈ CN
(153)
(154)
where by definition, ¯P1[·] and P1[·] denote the operators
that work on the scalar entries of their arguments.
P [a ⊗ b] = (cid:107)b(cid:107)2 · P1[a]
¯P [X ⊗ IM ] = ¯P1[X]
When X consists of nonnegative entries, relation (153)
becomes
(155)
is the
(156)
¯P [X ⊗ IM ] = X
b,∞ (cid:44)(cid:0) max
N(cid:88)
8) (Relation to norms): The ∞−norm of P [x]
(cid:1)2
squared block maximum norm of x:
(cid:107)P [x](cid:107)∞ = (cid:107)x(cid:107)2
1≤k≤N (cid:107)xk(cid:107)
Moreover, the sum of the entries in P [x] is the squared
Euclidean norm of x:
1T P [x] = (cid:107)x(cid:107)2 =
(cid:107)xk(cid:107)2
(157)
k=1
9) (Inequality preservation): Suppose vectors x, y and
matrices F , G have nonnegative entries, then x (cid:22) y
implies F x (cid:22) F y, and F (cid:22) G implies F x (cid:22) Gx.
10) (Upper bounds): It holds that
¯P [X] (cid:22) (cid:107) ¯P [X](cid:107)1 · 11T
¯P [X] (cid:22) (cid:107) ¯P [X](cid:107)∞ · 11T
(158)
(159)
where (cid:107) · (cid:107)∞ denotes the ∞−induced norm of a matrix
(maximum absolute row sum).
Proof: See Appendix C.
More importantly, the following variance relations hold for
the energy and norm operators. These relations show how error
variances propagate after a certain operator is applied to a
random vector.
Lemma 6 (Variance relations): Consider N × 1 block vec-
tors x = col{x1, . . . , xN} and y = col{y1, . . . , yN} with
M × 1 entries {xk, yk}. The following variance relations are
satisfied by the energy vector operator P [·]:
1) (Linear transformation): Given a K × N block matrix
Q with the size of each block being M ×M, Qx defines
a linear operator on x and its energy satisfies
P [Qx] (cid:22) (cid:107) ¯P [Q](cid:107)∞ · ¯P [Q] P [x]
(cid:22) (cid:107) ¯P [Q](cid:107)2∞ · 11T · P [x]
(160)
(161)
As a special case, for a left-stochastic N × N matrix A,
we have
(162)
2) (Update operation): The global update vector defined
P [(AT ⊗ IM )x] (cid:22) AT P [x]
by (85) satisfies the following variance relation:
U P [x − y]
P [s(x) − s(y)] (cid:22) λ2
(163)
3) (Centralized operation): The centralized operator Tc(x)
defined by (99) satisfies the following variance relations:
(164)
P [Tc(x) − Tc(y)] (cid:22) γ2
P [Tc(x) − Tc(y)] (cid:23) (1 − 2µmax(cid:107)p(cid:107)1λU ) · P [x − y]
(165)
c · P [x − y]
where
γc (cid:44) 1 − µmaxλL +
1
2
µ2
max(cid:107)p(cid:107)2
1λ2
U
(166)
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
17
...
...
2
1−d2
d2
(171)
t · δwT ·
N(cid:88)
k=1
5) (Stable Kronecker Jordan operator): Suppose DL =
DL ⊗ IM , where DL is the L× L Jordan matrix defined
in (168) -- (169). Then, for any LM × 1 vectors xe and
ye, we have
P [DLxe + ye] (cid:22) Γe · P [xe] +
Proof: See Appendix D.
2
1 − d2 · P [ye]
(172)
APPENDIX B
PROOF OF LEMMA 1
First, we establish that conditions (27) and (28) imply (24)
and (25), respectively. Using the mean-value theorem [49, p.6],
(cid:13)(cid:13)(cid:13)(cid:13)(cid:90) 1
we have for any x, y ∈ S:
(cid:90) 1
(cid:107)sk(x) − sk(y)(cid:107) =
0 ∇wT sk(y + t(x − y))dt · (x − y)
(cid:13)(cid:13)(cid:13)(cid:13)
≤
≤ λU · (cid:107)x − y(cid:107)
0 (cid:107)∇wT sk(y + t(x − y))(cid:107) dt · (cid:107)x − y(cid:107)
(173)
where we used the fact that y + t(x − y) = tx + (1 − t)y ∈ S
given x, y ∈ S and 0 ≤ t ≤ 1. Likewise, we have
(x − y)T ·
pk[sk(x) − sk(y)]
N(cid:88)
k=1
Moreover, it follows from (26) that
γc ≥ 1 − µmaxλL +
1
2
µ2
maxλ2
L
= (1 −
µmaxλL)2 +
(167)
4) (Stable Jordan operation): Suppose DL is an L × L
Jordan matrix of the following block form:
µ2
maxλ2
L > 0
1
4
1
2
DL (cid:44) diag{DL,2, . . . , DL,n0}
(168)
where the nth Ln × Ln Jordan block is defined as (note
that L = L2 + ··· + Ln0)
dn
DL,n (cid:44)
(169)
1
...
...
...
1
dn
We further assume DL to be stable with 0 ≤ dn0 ≤
··· ≤ d2 < 1. Then, for any L × 1 vectors x(cid:48) and y(cid:48),
we have
P1[DLx
(cid:48)
(cid:48)
+ y
] (cid:22) Γe · P1[x
(cid:48)
] +
2
1 − d2 · P1[y
(cid:48)
] (170)
where Γe is the L × L matrix defined as
d2
2
1−d2
...
Γe (cid:44)
N(cid:88)
(cid:90) 1
k=1
(cid:90) 1
N(cid:88)
pk
= (x − y)T ·
0 ∇wT sk(y + t(x − y))dt · (x − y)
0
(29)
k=1
Hc(y + t(x − y)) + H T
pk∇wT sk(y + t(x − y))dt · (x − y)
= (x − y)T ·
= (x − y)T · Hc(y + t(x − y)) · (x − y)
= (x−y)T ·
≥ λL · (cid:107)x − y(cid:107)2
Next, we establish the reverse direction that conditions (24)
and (25) imply (27) and (28). Choosing x = w + t · δw and
y = w in (24) for any δw (cid:54)= 0 and any small positive t, we
get
c (y + t(x − y))
·(x−y)
(174)
2
(cid:107)sk(w + t · δw) − sk(w)(cid:107) ≤ t · λU · (cid:107)δw(cid:107)
⇒ lim
t→0+
(cid:13)(cid:13)(cid:13)(cid:13) sk(w + t · δw) − sk(w)
(cid:13)(cid:13)(cid:13)(cid:13) lim
sk(w + t · δw) − sk(w)
⇒
⇒ (cid:107)∇wT sk(w)δw(cid:107) ≤ λU · (cid:107)δw(cid:107)
⇒ (cid:107)∇wT sk(w)(cid:107) (cid:44) sup
δw(cid:54)=0
t→0+
t
t
(cid:13)(cid:13)(cid:13)(cid:13) ≤ λU · (cid:107)δw(cid:107)
(cid:13)(cid:13)(cid:13)(cid:13) ≤ λU · (cid:107)δw(cid:107)
(cid:107)∇wT sk(w)δw(cid:107)
≤ λU (175)
Likewise, choosing x = w + t · δw and y = w in (25) for any
δw (cid:54)= 0 and any small positive t, we obtain
(cid:107)δw(cid:107)
(cid:19)
pk[sk(w + t · δw) − sk(w)] ≥ t2 · λL · (cid:107)δw(cid:107)2
(cid:18)
sk(w + t · δw) − sk(w)
t
⇒ δwT ·
pk
N(cid:88)
lim
t→0+
N(cid:88)
≥ λL · (cid:107)δw(cid:107)2
k=1
pk∇wT sk(w) · δw ≥ λL · (cid:107)δw(cid:107)2
k=1
⇒ δwT ·
⇒ δwT Hc(w)δw ≥ λL · (cid:107)δw(cid:107)2
⇒ δwT Hc(w) + H T
⇒
2
Hc(w) + H T
c (w)
≥ λL · IM
c (w)
2
δw ≥ λL · (cid:107)δw(cid:107)2
(176)
APPENDIX C
PROOF OF LEMMA 5
Properties 1-2 are straightforward from the definitions of
P [·] and ¯P [·]. Property 4 was proved in [20]. We establish the
remaining properties.
(Property 3: Convexity) The convexity of P [·] has already
been proven in [20]. We now establish the convexity of the
operator ¯P [·]. Let X (k)
qn denote the (q, n)−th M × M block
of the matrix X (k), where q = 1, . . . , K and n = 1, . . . , N.
(cid:35)
Then,
(cid:34) K(cid:88)
k=1
¯P
akX (k)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
(G− F )x (cid:23) 0, which is true because all entries of G− F and
x are nonnegative due to F (cid:22) G and x (cid:23) 0.
(Property 10: Upper bounds) By the definition of ¯P [X] in
(107), we get
(177)
¯P [X] (cid:22)
(cid:18)
(cid:19)
l,k (cid:107)Xlk(cid:107)
max
(cid:32) N(cid:88)
(cid:33)
· 11T
l
(cid:22) max
(cid:107)Xlk(cid:107)
= (cid:107) ¯P [X](cid:107)∞ · 11T
k=1
APPENDIX D
PROOF OF LEMMA 6
(cid:13)(cid:13)(cid:13)X (k)
(cid:13)(cid:13)(cid:13)X (k)
K1
11
(cid:13)(cid:13)(cid:13) ···
(cid:13)(cid:13)(cid:13) ···
(cid:80)K
(cid:80)K
k=1 ak
...
k=1 ak
(cid:13)(cid:13)(cid:13)X (k)
(cid:13)(cid:13)(cid:13)X (k)
1N
KN
(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)
18
(cid:22)
=
(cid:80)K
(cid:80)K
K(cid:88)
k=1
k=1 ak
...
k=1 ak
ak ¯P [X (k)]
(Property 5: Triangular inequality) Let Xqn and Yqn
denote the (q, n)−th M × M blocks of the matrices X and
Y , respectively, where q = 1, . . . , K and n = 1, . . . , N. Then,
by the triangular inequality of the matrix norm (cid:107) · (cid:107), we have
(cid:107)X11(cid:107) + (cid:107)Y11(cid:107)
...
¯P [X + Y ] (cid:22)
···
(cid:107)X1N(cid:107) + (cid:107)Y1N(cid:107)
...
= ¯P [X] + ¯P [Y ]
(cid:107)XK1(cid:107) + (cid:107)YK1(cid:107)
··· (cid:107)XKN(cid:107) + (cid:107)YKN(cid:107)
(178)
(Property 6: Submultiplicity) Let Xkn and Znl be the
(k, n)−th and (n, l)−th M × M blocks of X and Z, respec-
tively. Then, the (k, l)−th M ×M block of the matrix product
XZ, denoted by [XZ]k,l, is
n=1
N(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤
N(cid:88)
[XZ]k,l =
XknZnl
(179)
(cid:2) ¯P [XZ](cid:3)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
Therefore, the (k, l)−th entry of the matrix ¯P [XZ] can be
bounded as
n=1
k,l =
XknZnl
(cid:107)Xkn(cid:107) · (cid:107)Znl(cid:107) (180)
Note that (cid:107)Xkn(cid:107) and (cid:107)Znl(cid:107) are the (k, n)−th and (n, l)−th
entries of the matrices ¯P [X] and ¯P [Z], respectively. The right-
hand side of the above inequality is therefore the (k, l)−th
entry of the matrix product ¯P [X] ¯P [Z]. Therefore, we obtain
n=1
¯P [XZ] (cid:22) ¯P [X] ¯P [Z]
(181)
(Property 7: Kronecker structure) For (153), we note that
the (k, n)−th M × M block of X ⊗ IM is xknIM . Therefore,
by the definition of ¯P [·], we have
···
x1N
= ¯P1[X]
x11
...
¯P [X ⊗ IM ] =
(182)
xK1
···
xKN
In the special case when X consists of nonnegative entries,
¯P1[X] = X, and we recover (155). To prove (154), we let
a = col{a1, . . . , aN} and b = col{b1, . . . , bM}. Then, by the
definitin of P [·], we have
P [a ⊗ b] = col{a12 · (cid:107)b(cid:107)2, . . . ,aN2 · (cid:107)b(cid:107)2} = (cid:107)b(cid:107)2 · P1[a]
(183)
(Property 8: Relation to norms) Relations (156) and (157)
are straightforward and follow from the definition.
(Property 9: Inequality preservation) The proof that x (cid:22) y
implies F x (cid:22) F y can be found in [20]. We now prove that
F (cid:22) G implies F x (cid:22) Gx. This can be proved by showig that
· 11T
(184)
Likewise, we can establish that ¯P [X] (cid:22) (cid:107) ¯P [X](cid:107)1 · 11T .
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(Property 1: Linear transformation) Let Qkn be the (k, n)-
th M × M block of Q. Then
P [Qx] = col
Q1nxn
, . . . ,
QKnxn
(185)
Using the convexity of (cid:107)·(cid:107)2, we have the following bound on
each n-th entry:
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
n=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
n=1
(cid:13)(cid:13)(cid:13)(cid:13)2
n=1
(a)
=
Qknxn
(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:34) N(cid:88)
(cid:34) N(cid:88)
(cid:34) N(cid:88)
(cid:34) N(cid:88)
≤
n=1
n=1
n=1
=
(b)
≤ max
k
n=1
(cid:107)Qkn(cid:107)
(cid:107)Qkn(cid:107)
(cid:107)Qkn(cid:107)
·
(cid:35)2
(cid:35)2
(cid:35)
·
n=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:80)N
(cid:107)Qkn(cid:107)
N(cid:88)
l=1 (cid:107)Qkl(cid:107)
(cid:80)N
(cid:107)Qkn(cid:107)
N(cid:88)
l=1 (cid:107)Qkl(cid:107)
(cid:35)
(cid:107)Qkn(cid:107) · (cid:107)xn(cid:107)2
N(cid:88)
n=1
n=1
·
·
n=1
(cid:107)Qkn(cid:107) · (cid:107)xn(cid:107)2
xn
Qkn
·
(cid:107)Qkn(cid:107)
· (cid:107)Qkn(cid:107)2
(cid:107)Qkn(cid:107)2(cid:107)xn(cid:107)2
= (cid:107) ¯P [Q](cid:107)∞ ·
(cid:107)Qkn(cid:107) · (cid:107)xn(cid:107)2
(186)
where in step (b) we applied Jensen's inequlity to (cid:107) · (cid:107)2.
Note that if some (cid:107)Qkn(cid:107) in step (a) is zero, we eliminate
the corresponding term from the sum and it can be verified
that the final result still holds. Substituting into (185), we
establish (160). The special case (162) can be obtained by
using ¯P [AT⊗IM ] = AT and that (cid:107)AT(cid:107)∞ = 1 (left-stochastic)
in (160). Finally, the upper bound (161) can be proved by
applying (159) to ¯P [Q].
(Property 2: Update operation) By the definition of P [·]
and the Lipschitz Assumption 3, we have
P [s(x) − s(y)]
= col{(cid:107)s1(x1) − s1(y1)(cid:107)2, . . .(cid:107)sN (xN ) − sN (yN )(cid:107)2}
(cid:22) col{λ2
= λ2
U · (cid:107)x1 − y1(cid:107)2, . . . λ2
U · (cid:107)xN − yN(cid:107)2}
(187)
U · P [x − y]
(cid:107)Qkn(cid:107)
N(cid:88)
n=1
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
(a)
(cid:13)(cid:13) N(cid:88)
(cid:23) (cid:107)x − y(cid:107)2 − 2µmax · (cid:107)x − y(cid:107) ·
N(cid:88)
N(cid:88)
(cid:23) (cid:107)x − y(cid:107)2 − 2µmax · (cid:107)x − y(cid:107) ·
(cid:23) (cid:107)x − y(cid:107)2 − 2µmax · (cid:107)x − y(cid:107) ·
= (1 − 2µmaxλU(cid:107)p(cid:107)1) · (cid:107)x − y(cid:107)2
k=1
k=1
k=1
(b)
19
pk [sk(x) − sk(y)](cid:13)(cid:13)
pk (cid:107)sk(x) − sk(y)(cid:107)
pkλU(cid:107)x − y(cid:107)
(190)
where in step (a), we used the Cauchy-Schwartz inequality
xT y ≤ xT y ≤ (cid:107)x(cid:107) · (cid:107)y(cid:107), and in step (b) we used (24).
matrix DL,n can be written as
(Property 4: Stable Jordan operator) First, we notice that
pk [sk(x) − sk(y)]
(Property 3: Centralized opertion) Since Tc : RM → RM ,
the output of P [Tc(x) − Tc(y)] becomes a scalar. From the
definition, we get
P [Tc(x) − Tc(y)]
=(cid:13)(cid:13)x − y − µmax · (pT ⊗ IM ) [s(1 ⊗ x) − s(1 ⊗ y)](cid:13)(cid:13)2
=(cid:13)(cid:13)x − y − µmax ·
N(cid:88)
We first prove the upper bound (164) as follows:
P [Tc(x) − Tc(y)]
=
k=1
k=1
k=1
max
N(cid:88)
N(cid:88)
N(cid:88)
= (cid:107)x − y(cid:107)2 − 2µmax · (x − y)T
+ µ2
= (cid:107)x − y(cid:107)2 − 2µmax · (x − y)T
+ µ2
pk [sk(x) − sk(y)](cid:13)(cid:13)2
(cid:13)(cid:13)2
(cid:13)(cid:13)(pT ⊗ IM ) [s(1 ⊗ x) − s(1 ⊗ y)]
maxP(cid:2)(pT ⊗ IM ) [s(1 ⊗ x) − s(1 ⊗ y)](cid:3)
(cid:13)(cid:13)2
(cid:13)(cid:13)x − y − µmax ·
(cid:13)(cid:13) N(cid:88)
(cid:13)(cid:13) N(cid:88)
(cid:2) N(cid:88)
(cid:34) N(cid:88)
N(cid:88)
(cid:13)(cid:13)2
pk[sk(x) − sk(y)](cid:13)(cid:13)2
(cid:3)2
≤ (cid:107)x − y(cid:107)2 − 2µmax · λL · (cid:107)x − y(cid:107)2
+ µ2
pk(cid:107)sk(x) − sk(y)(cid:107)
(cid:35)2
≤ (cid:107)x − y(cid:107)2 − 2µmax · λL · (cid:107)x − y(cid:107)2
+ µ2
≤ (cid:107)x − y(cid:107)2 − 2µmax · λL · (cid:107)x − y(cid:107)2
+ µ2
= (cid:107)x − y(cid:107)2 − 2µmax · (x − y)T
pk[sk(x) − sk(y)]
pk[sk(x) − sk(y)]
max ·
max ·
max ·
+ µ2
k=1
k=1
k=1
k=1
k=1
(25)
(24)
pk · λU · (cid:107)x − y(cid:107)
max ·
k=1
1λ2
max · (cid:107)p(cid:107)2
= (cid:107)x − y(cid:107)2 − 2µmax · λL · (cid:107)x − y(cid:107)2
+ µ2
(cid:19)2
(cid:18)
= (1 − 2µmaxλL + µ2
1) · (cid:107)x − y(cid:107)2
1
maxλ2
µ2
≤
2
U(cid:107)p(cid:107)2
U(cid:107)p(cid:107)2
1 − µmaxλL +
U · (cid:107)x − y(cid:107)2
maxλ2
1
· (cid:107)x − y(cid:107)2
(189)
2 x)2.
where in the last step we used the relation (1−x) ≤ (1− 1
Next, we prove the lower bound (165). From (188), we
notice that the last term in (188) is always nonnegative so that
P [Tc(x) − Tc(y)]
(cid:23) (cid:107)x − y(cid:107)2 − 2µmax · (x − y)T
N(cid:88)
k=1
pk [sk(x) − sk(y)]
pk [sk(x) − sk(y)]
(191)
where ΘLn is an Ln × Ln strictly upper triangular matrix of
the following form:
DL,n = dn · ILn + ΘLn
(188)
0
1
...
...
...
1
0
(cid:44)
ΘLn
(192)
(193)
(194)
pk[sk(x) − sk(y)]
Define the following matrices:
ΛL (cid:44) diag{d2IL2, . . . , dn0 ILn0}
(cid:48)
L
(cid:44) diag{ΘL2 , . . . , ΘLn0}
Θ
Then, the original Jordan matrix DL can be expressed as
DL = ΛL + Θ
so that
(cid:48)
L
(cid:48)
P1[DLx
]
+ y
(cid:48)
= P1 [ΛLx
+ Θ
(cid:48)
Lx
(cid:48)
(cid:48)
(195)
(cid:48)
(cid:48)
Lx
Θ
(cid:48)(cid:21)
2
1 − d2
(cid:20)
ΛLx
1 − d2
+
ΛLx
·
y
2
2
+
+ y
(cid:48)
]
1 − d2
(cid:48)(cid:21)
(cid:48)(cid:21)
(cid:48)(cid:21)
1 − d2
(cid:48)
1 − d2 ·P1 [Θ
Lx
1 − d2
2
2
2
y
(cid:20)
]+
· ¯P1[ΛL] · P1 [x
(cid:48)
]
· P1
(cid:48)
Lx
Θ
2
1 − d2
(cid:48)
]+
2
1 − d2 ·P1 [y
(cid:48)
]
]+
(cid:48)
] +
(cid:48)
· ¯P1[Θ
L] · P1 [x
(cid:48)
2
L · P1 [x
1−d2 ·Θ
(cid:48)
1−d2 ·ΘL · P1 [x
1 − d2 · P1 [y
2
2
(cid:48)
]
]+
(cid:48)
2
]
2
]+
1 − d2 · P1 [y
(cid:48)
1−d2 ·P1 [y
1−d2 ·P1 [y
]+
2
(cid:48)
]
(cid:48)
]
(196)
(cid:48)
(cid:20)
+
= P1
1
d2 ·
d2
1 − d2
(cid:20) 1
2
·
(a)
(c)
(b)
=
2
+
(cid:22) d2 · P1
1 − d2
1
d2
· P1
(cid:48)
d2 ·P1 [ΛLx
(cid:22) (cid:107) ¯P1[ΛL](cid:107)∞
d2
2(cid:107) ¯P1[Θ(cid:48)
L](cid:107)∞
1 − d2
(cid:22) ¯P1[ΛL]·P1 [x
(cid:48)
(cid:22) d2·IL·P1 [x
= Γe · P1 [x
] +
+
(cid:48)
(cid:48)
(f )
(d)
(e)
20
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
L] = Θ(cid:48)
L and (cid:107) ¯P1[Θ(cid:48)
L(cid:107)∞ = 1, step (e) uses ¯P1[ΛL] (cid:22) d2 · IL and Θ(cid:48)
where step (a) uses the convexity property (148), step (b) uses
the scaling property, step (c) uses variance relation (160), step
(d) uses (cid:107) ¯P1[ΛL](cid:107)∞ = d2, ¯P1[Θ(cid:48)
L](cid:107)∞ =
(cid:107)Θ(cid:48)
L (cid:22) ΘL,
where ΘL denotes a matrix of the same form as (192) but of
size L×L, step (f) uses the definition of the matrix Γe in (171).
The above derivation assumes d2 (cid:54)= 0. When d2 = 0, we
can verify that the above inequality still holds. To see this, we
first notice that when d2 = 0, the relation 0 ≤ dn0 ≤ ··· ≤
d2 implies that dn0 = ··· = d2 = 0 so that ΛL = 0 and
DL = Θ(cid:48)
L -- see (193) and (195). Therefore, similar to the
steps (a) -- (f) in (196), we get
(cid:48)
(cid:48)
(cid:48)
(cid:48)
P1[DLx
+ y
(cid:48)(cid:3)
(cid:2) 1
]
(cid:48)
+
]
(cid:48)
= P1
(cid:22)
=
] +
(cid:48)
(cid:48)
+ y
Lx
] = P1[Θ
(cid:48)
2 · 2Θ
Lx
(cid:48)
(cid:48)
1
2 · P1[2Θ
Lx
(cid:48)
1
2 · 22 · P1[Θ
Lx
1
2 · 2y
1
2 · P1[2y
1
2 · 22 · P1[y
] +
(cid:48)
(cid:48)
(cid:48)
]
Lx
= 2P1[Θ
] + 2P1[y
(cid:48)
(cid:48)
(cid:22) 2(cid:107) ¯P1[Θ
L](cid:107)∞ · ¯P1[Θ
L]P1[x
(cid:48)
(cid:48)
(cid:48)
] + 2P1[y
= 2Θ
LP1[x
]
(cid:48)
(cid:48)
(cid:22) 2ΘLP1[x
] + 2P1[y
(cid:48)
]
(cid:48)
]
] + 2P1[y
(cid:48)
]
(197)
By (171), we have Γe = 2ΘL when d2 = 0. Therefore, the
above expression is the same as the one on the right-hand side
of (196).
(Property 5: Stable Kronecker Jordan operator) Using
(195) we have
P [DLxe + ye]
(cid:104)
(cid:105)
(cid:48)
L ⊗ IM )xe + ye]
·
2
2
2
2
1
2
(a)
+
+
1−d2
1−d2
(cid:22) d2·P
= P [(ΛL ⊗ IM )xe + (Θ
= P
d2·
1 − d2
(cid:105)
d2 ·(ΛL⊗IM )xe +
(cid:104) 1
(cid:105)
1 − d2 · ye
(cid:104)
d2 ·(ΛL⊗IM )xe
(cid:105)
(cid:104)
(cid:48)
·P
1−d2 ·(Θ
L ⊗ IM )xe
1−d2 ·ye
d2 · P [(ΛL ⊗ IM )xe] +
1 − d2 · P [ye]
(cid:22) (cid:107) ¯P [(ΛL⊗IM )](cid:107)∞
d2
2(cid:107) ¯P [Θ(cid:48)
L⊗IM ](cid:107)∞
1 − d2
1−d2
1
· P
(b)
=
+
+
(c)
2
2
2
+
· ¯P [(ΛL ⊗ IM )]·P [xe]
(cid:48)
· ¯P [Θ
L⊗IM ]·P [xe] +
(cid:48)
(cid:22) ¯P [(ΛL ⊗ IM )] · P [xe] +
1 − d2 · Θ
L · P [xe]
+
2
2
(d)
1 − d2 · P [ye]
2
(cid:48)
1 − d2 · P [(Θ
L ⊗ IM )xe]
·
2
(cid:48)
1−d2 ·(Θ
L⊗IM )xe
2
1−d2 ·P [ye]
(e)
2
(cid:22) d2·IL·P [xe]+
= Γe · P [xe] +
(f )
1−d2 ·ΘL · P [xe]+
1 − d2 · P [ye]
2
2
1−d2 ·P [ye]
(198)
where step (a) uses the convexity property (148), step (b) uses
the scaling property, step (c) uses variance relation (160), step
L ⊗ IM ] = Θ(cid:48)
(d) uses (cid:107) ¯P [ΛL ⊗ IM ](cid:107)∞ = d2 and ¯P [Θ(cid:48)
L,
step (e) uses ¯P [ΛL ⊗ IM ] (cid:22) d2 · IL and Θ(cid:48)
L (cid:22) ΘL, and step
(f) uses the definition of the matrix Γe in (171). Likewise,
we can also verify that the above inequalty holds for the case
d2 = 0.
APPENDIX E
PROOF OF THEOREM 1
Consider the following operator:
T0(w) (cid:44) w −
λL
(cid:107)p(cid:107)2
1λ2
U
N(cid:88)
k=1
pksk(w)
(199)
As long as we are able to show that T0(w) is a strict
contraction mapping, i.e., ∀ x, y, (cid:107)T0(x)−T0(y)(cid:107) ≤ γ0(cid:107)x−y(cid:107)
with γo < 1, then we can invoke the Banach fixed point
theorem [59, pp.299-300] to conclude that there exists a unique
wo such that wo = T0(wo), i.e.,
N(cid:88)
k=1
N(cid:88)
k=1
wo = wo −
λL
1λ2
(cid:107)p(cid:107)2
U
pksk(wo) ⇔
pksk(wo) = 0
(200)
as desired. Now, to show that T0(·) defined in (199) is indeed a
contraction, we compare T0(·) with Tc(·) in (99) and observe
that T0(w) has the same form as Tc(·) if we set µmax =
in (99). Therefore, calling upon property (164) and
λL(cid:107)p(cid:107)2
1λ2
U
using µmax = λL(cid:107)p(cid:107)2
in the expression for γc in (166), we
1λ2
U
obtain
(cid:32)
(cid:18)
(cid:22)
1−
P [T0(x) − T0(y)]
λL
1λ2
(cid:107)p(cid:107)2
U
λ2
1
L
1λ2
(cid:107)p(cid:107)2
2
U
1 −
=
λL +
1
2
(cid:19)2
(cid:18) λL
(cid:19)2
1λ2
U
(cid:107)p(cid:107)2
· P [x − y]
(cid:33)2
1λ2
U
(cid:107)p(cid:107)2
·P [x − y]
(201)
By the definition of P [·] in (105), the above inequality is
equivalent to
(cid:18)
(cid:19)2
(cid:107)T0(x) − T0(y)(cid:107)2 ≤
1
1 −
2
It remains to show that 1− λ2
1 and λ2
the fact that λL, (cid:107)p(cid:107)2
1
2
< 1 −
1
2
· (cid:107)x − y(cid:107)2
λ2
L
1λ2
(cid:107)p(cid:107)2
U
L/(2(cid:107)p(cid:107)2
U are positive, we have
1λ2
U ) < 1. By (26) and
(202)
λ2
L
1λ2
(cid:107)p(cid:107)2
U
< 1
(203)
Therefore, T0(w) is a strict contraction mapping.
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
21
APPENDIX F
PROOF OF THEOREM 2
N(cid:88)
k=1
By Theorem 1, wo is the unique solution to equation (109).
Subtracting both sides of (109) from wo, we recognize that
wo is also the unique solution to the following equation:
wo = wo − µmax
pksk(wo)
(204)
so that wo = Tc(wo). Applying property (164), we obtain
(cid:107) wc,i(cid:107)2 = P [wo − ¯wc,i]
= P [Tc(wo) − Tc( ¯wc,i−1)]
(cid:22) γ2
(cid:22) γ2i
= γ2i
c · P [wo − ¯wc,i−1]
c · P [wo − ¯wc,0]
c · (cid:107) wc,0(cid:107)2
(205)
Since γc > 0, the upper bound on the right-hand side will
converge to zero if γc < 1. From its definition (166), this
condition is equivalent to requiring
1 − µmaxλL +
1
2
µ2
max(cid:107)p(cid:107)2
1λ2
U < 1
(206)
Solving the above quadratic inequality in µmax, we obtain
(112). On the other hand, to prove the lower bound in (112),
we apply (165) and obtain
(cid:107) wc,i(cid:107)2 = P [wo − ¯wc,i]
= P [Tc(wo) − Tc( ¯wc,i−1)]
(cid:23) (1 − 2µmax(cid:107)p(cid:107)1λU ) · P [wo − ¯wc,i−1]
(cid:23) (1 − 2µmax(cid:107)p(cid:107)1λU )i · P [wo − ¯wc,0]
= (1 − 2µmax(cid:107)p(cid:107)1λU )i · (cid:107) wc,0(cid:107)2
(207)
APPENDIX G
PROOF OF THEOREM 3
Since (110) already establishes that ¯wc,i approaches wo
asymptotically (so that wc,i → 0), and since from Assumption
5 we know that sk(w) is differentiable when (cid:107) wc,i(cid:107) ≤ rH for
large enough i, we are justified to use the mean-value theorem
[49, p.24] to obtain the following useful relation:
sk( ¯wc,i−1) − sk(wo)
(cid:20)(cid:90) 1
(cid:21)
0 ∇wT sk(wo − t wc,i−1)dt
(cid:90) 1
∇wT sk(wo − t wc,i−1) − ∇wT sk(wo)(cid:3)dt · wc,i−1
(cid:2)
= −
= −∇wT sk(wo) · wc,i−1
−
wc,i−1
0
(208)
Therefore, subtracting wo from both sides of (47) and using
(109) we get,
wc,i = wc,i−1 + µmax
pk(sk( ¯wc,i−1) − sk(wo))
= [I − µmaxHc] wc,i−1 − µmax · ei−1
(209)
N(cid:88)
k=1
where
Hc (cid:44) N(cid:88)
ei−1 (cid:44) N(cid:88)
k=1
0
pk
k=1
(210)
(211)
Furthermore, the perturbation term ei−1 satisfies the following
bound:
(cid:90) 1
pk∇wT sk(wo)
(cid:2)
∇wT sk(wo−t wc,i−1)−∇wT sk(wo)(cid:3)dt· wc,i−1
(cid:90) 1
N(cid:88)
(cid:90) 1
N(cid:88)
· (cid:107) wc,i−1(cid:107)
(cid:13)(cid:13)∇wT sk(wo−t wc,i−1)−∇wT sk(wo)(cid:13)(cid:13)dt
k=1
pk
0
pk
λH · t · (cid:107) wc,i−1(cid:107)dt · (cid:107) wc,i−1(cid:107)
(cid:107)ei−1(cid:107) ≤
≤
=
0
k=1
1
2(cid:107)p(cid:107)1λH · (cid:107) wc,i−1(cid:107)2
Evaluating the weighted Euclidean norm of both sides of
(209), we get
(cid:107) wc,i(cid:107)2
Σ = (cid:107) wc,i−1(cid:107)2
BT
c ΣBc − 2µmax · wT
c,i−1BT
c Σei−1
+ µ2
max · (cid:107)ei−1(cid:107)2
Σ
(212)
(213)
where
(cid:12)(cid:12) wT
(cid:12)(cid:12)
Bc = I − µmaxHc
(214)
Σ = xT Σx, and Σ is an arbitrary positive semi-
Moreover, (cid:107)x(cid:107)2
definite weighting matrix. The second and third terms on the
right-hand side of (213) satisfy the following bounds:
(a)
c,i−1BT
c Σei−1
≤ (cid:107) wc,i−1(cid:107) · (cid:107)BT
≤ (cid:107) wc,i−1(cid:107) · (cid:107)BT
≤ (cid:107) wc,i−1(cid:107) · (cid:107)BT
c (cid:107) · (cid:107)Σ(cid:107) · (cid:107)ei−1(cid:107)
c (cid:107) · Tr(Σ) · (cid:107)ei−1(cid:107)
λH(cid:107)p(cid:107)1
c (cid:107) · Tr(Σ) ·
2
· (cid:107) wc,i−1(cid:107)2
(215)
and
(cid:107)ei−1(cid:107)2
Σ ≤ (cid:107)Σ(cid:107) · (cid:107)ei−1(cid:107)2
(b)
≤ Tr(Σ) · (cid:107)ei−1(cid:107)2
H(cid:107)p(cid:107)2
λ2
≤ Tr(Σ) ·
4
1
· (cid:107) wc,i−1(cid:107)4
(216)
where for steps (a) and (b) of the above inequalities we used
the property (cid:107)Σ(cid:107) ≤ Tr(Σ). This is because we consider here
the spectral norm, (cid:107)Σ(cid:107) = σmax(Σ), where σmax(·) denotes the
maximum singular value. Since Σ is symmetric and positive
semidefinite, its singular values are the same as its eigenvalues
so that (cid:107)Σ(cid:107) = λmax(Σ) ≤
m λm(Σ) = Tr(Σ). Now, for any
given small > 0, there exists i0 such that, for i ≥ i0, we
have (cid:107) wc,i−1(cid:107) ≤ so that
(cid:80)
(cid:12)(cid:12) ≤ · (cid:107)BT
c (cid:107)·Tr(Σ)·
λH(cid:107)p(cid:107)1
2
(cid:12)(cid:12) wT
c,i−1BT
c Σei−1
(cid:107)ei−1(cid:107)2
Σ ≤ 2 · Tr(Σ) ·
H(cid:107)p(cid:107)2
λ2
4
1
·(cid:107) wc,i−1(cid:107)2
(217)
· (cid:107) wc,i−1(cid:107)2 (218)
22
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
(cid:3)qqT
(cid:3)qqT
(221)
and that
where
c (cid:107)+µmax
λH(cid:107)p(cid:107)1
4
(cid:2)
(cid:107)BT
(cid:2)
(cid:107)BT
(cid:3)
λH(cid:107)p(cid:107)1
c ΣBc+∆ (219)
Substituting (217) -- (218) into (213), we obtain
Σ ≤ (cid:107) wc,i−1(cid:107)2
(cid:107) wc,i−1(cid:107)2
where
c ΣBc−∆ ≤ (cid:107) wc,i(cid:107)2
BT
BT
(cid:2)
∆ (cid:44) µmax · λH(cid:107)p(cid:107)1 ·
(cid:107)BT
= O(µmax) · Tr(Σ) · IM
4
c (cid:107) + µmax
· Tr(Σ) · IM
(220)
Let σ = vec(Σ) denote the vectorization operation that stacks
the columns of a matrix Σ on top of each other. We shall
Σ interchangeably to denote
use the notation (cid:107)x(cid:107)2
the weighted squared Euclidean norm of a vector. Using the
Kronecker product property [58, p.147]: vec(U ΣV ) = (V T ⊗
U )vec(Σ), we can vectorize the matrices BT
c ΣBc + ∆ and
c ΣBc − ∆ in (219) as F+σ and F−σ, respectively, where
BT
F+ (cid:44) BT
σ and (cid:107)x(cid:107)2
c (cid:107)+µmax
λH(cid:107)p(cid:107)1
4
= BT
F− (cid:44) BT
c ⊗BT
c ⊗ BT
c ⊗BT
c ⊗ BT
c +µmax·λH(cid:107)p(cid:107)1·
c + O(µmax)
c −µmax·λH(cid:107)p(cid:107)1·
c − O(µmax)
= BT
(222)
where q (cid:44) vec(IM ), and we have used the fact that Tr(Σ) =
Tr(ΣIM ) = vec(IM )T vec(Σ) = qT σ. In this way, we can
write relation (219) as
σ ≤ (cid:107) wc,i−1(cid:107)2F+σ
(cid:107) wc,i−1(cid:107)2F−σ ≤ (cid:107) wc,i(cid:107)2
(223)
Using a state-space technique from [60, pp.344-346], we con-
Σ converges at a rate that is between ρ(F−)
clude that (cid:107) wc,i(cid:107)2
and ρ(F+). Recalling from (221) -- (222) that F+ and F− are
c , and since the perturbation
perturbed matrices of BT
term is O(µmax) which is small for small , we would expect
the spectral radii of F+ and F− to be small perturbations of
c ⊗ BT
ρ(BT
Lemma 7 (Perturbation of spectral radius): Let (cid:28) 1 be
a sufficiently small positive number. For any M × M matrix
X, the spectral radius of the perturbed matrix X + E for
E = O() is
c ). This claim is justified below.
c ⊗ BT
ρ(X + E) = ρ(X) + O(cid:0)
2(M−1)(cid:1)
1
(224)
Proof: Let X = T JT −1 be the Jordan canonical form of the
matrix X. Without loss of generality, we consider the case
where there are two Jordan blocks:
(225)
where J1 ∈ RL×L and J2 ∈ R(M−L)×(M−L) are Jordan
blocks of the form
J = diag{J1, J2}
λk
Jk =
1
...
...
...
1
λk
(226)
with λ1 > λ2. Since X + E is similar to T −1(X + E)T ,
the matrix X + E has the same set of eigenvalues as J + E0
where
E0 (cid:44) T
−1ET = O()
(227)
Let
(cid:44) diag(cid:8)1, 0, . . . , M−1
0 (cid:44)
D0
2(M−1)
1
0
(cid:9)
(228)
(229)
Then, by similarity again, the matrix J + E0 has the same set
of eigenvalues as
−1
(230)
0 (J + E0)D0 = D
where E1 (cid:44) D−1
0 E0D0. Note that the ∞-induced norm (the
maximum absolute row sum) of E1 is bounded by
−1
0 JD0 + E1
D
−1
0 (cid:107)∞ · (cid:107)E0(cid:107)∞ · (cid:107)D0(cid:107)∞
(cid:107)E1(cid:107)∞ ≤ (cid:107)D
1
M−1
· O() · 1 =
1
2 · O() = O(
1
=
0
1
2 )
(231)
(232)
(233)
−1
0 JD0 = diag{J
D
(cid:48)
1, J
(cid:48)
2}
λk
(cid:48)
J
k =
0
...
...
...
0
λk
M(cid:91)
Gm
Then, by appealing to Gersgorin Theorem [52, p.344], we
conclude that the eigenvalues of the matrix D−1
0 JD0 + E1,
which are also the eigenvalues of the matrices J + E0 and
X + E, lie inside the union of the Gersgorin discs, namely,
(234)
m=1
Gm (cid:44)
(cid:41)
where Gm is the mth Gersgorin disc defined as
(cid:41)
λ : λ − λ1 ≤ 0 +
E1,m(cid:96)
(cid:96)=1
(cid:40)
(cid:40)
(cid:110)
λ : λ−λ1 ≤ O(cid:0)
(cid:110)
λ : λ−λ2 ≤ O(cid:0)
λ : λ − λ2 ≤ 0 +
M(cid:88)
M(cid:88)
2(M−1)(cid:1)(cid:111)
2(M−1)(cid:1)(cid:111)
(cid:96)=1
1
E1,m(cid:96)
=
1
, 1 ≤ m ≤ L
, L < m ≤ M
, 1≤ m ≤ L
, L < m≤ M
(235)
1
and where E1,m(cid:96) denotes the (m, (cid:96))-th entry of the matrix
E1. In the last step we used (228) and (231). Observe from
(235) that there are two clusters of Gersgorin discs that are
centered around λ1 and λ2, respectively, and have radii on
2(M−1) ). A further statement from Gersgorin
the order of O(
theorem shows that if the these two clusters of discs happen
to be disjoint, which is true in our case since λ1 > λ2 and
we can select to be sufficiently small to ensure this property.
Then there are exactly L eigenvalues of X + E in ∪L
m=1Gm
while the remaining M − L eigenvalues are in ∪M
m=M−LGm.
From λ1 > λ2, we conclude that the largest eigenvalue of
D−1
2(M−1) ), which establishes (224).
0 JD0 + E1 is λ1 + O(
1
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
23
(d)
+
k=1
k=1
(e)
+
pl
pl
max
max
vk,i
pl
l=1
l=1
k=1
l=1
k=1
pl
·
E
max
l=1
k=1
max
E
+ µ2
+ µ2
µ2
max
µ2
max
l=1 pl
l=1 pl
l=1 pl
l=1 pl
pkvk,i
max·
max·
zk,i−1
k=1
µ2
(cid:13)(cid:13)2
pkzk,i−1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
µ2
1 − γc
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
+ µ2
= γc · EP [ wc,i−1] +
E(cid:13)(cid:13)(pT ⊗ IM )vi
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
+ µ2
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:32) N(cid:88)
(cid:33)2
= γc · EP [ wc,i−1]
pk(cid:80)N
·E
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) N(cid:88)
(cid:32) N(cid:88)
(cid:33)2
1 − γc·
pk(cid:80)N
·E
(cid:32) N(cid:88)
(cid:33)2
N(cid:88)
≤ γc · EP [ wc,i−1]
pk(cid:80)N
E(cid:107)zk,i−1(cid:107)2
(cid:33)2
(cid:32) N(cid:88)
1 − γc·
N(cid:88)
pk(cid:80)N
(cid:32) N(cid:88)
(cid:33)
E(cid:107)vk,i(cid:107)2
N(cid:88)
·
(cid:32) N(cid:88)
(cid:33)
= γc · EP [ wc,i−1] +
= γc · EP [ wc,i−1] +
+ µ2
= γc · EP [ wc,i−1] +
+ µ2
(cid:22) γc · EP [ wc,i−1]
µmax(cid:107)p(cid:107)1
2(cid:107)p(cid:107)2
1λ2
U
1 UL]
max · (cid:107)p(cid:107)1 · pT EP [vi]
λL − 1
max · (cid:107)p(cid:107)1 · pT EP [vi]
(cid:13)(cid:13)2
∞ · 11T · EP [we,i−1]
4α · 1 · EP [ wc,i−1]
(cid:111)
(cid:3)
1 UL](cid:107)2∞ · 11T EP [we,i−1]
(cid:3)
· 1
(cid:13)(cid:13) ¯P [AT
1 UL](cid:13)(cid:13)2
· EP [ wc,i−1]
(cid:105)
∞ · λ2
(cid:107) wc,0(cid:107)2 + (cid:107)wo(cid:107)2(cid:1) + σ2
(cid:2)4α(cid:0)
1λ2
U
· pT(cid:110)
(cid:13)(cid:13) ¯P [AT
max · (cid:107)p(cid:107)1 · pT(cid:110)
+(cid:2)4α · ((cid:107) wc,0(cid:107)2 + (cid:107)wo(cid:107)2) + σ2
= (cid:2)γc + µ2
(cid:104)
+ (cid:107)p(cid:107)2
1 ·
µmax
λL − 1
2 µmax(cid:107)p(cid:107)2
max · (cid:107)p(cid:107)2
1 ·
1 − γc ·
N(cid:88)
·
k=1
µ2
1 − γc · (cid:107)p(cid:107)1 · pT EP [zi−1]
µmax · (cid:107)p(cid:107)1
2 µmax(cid:107)p(cid:107)2
+ 4α · (cid:107) ¯P [AT
pkE(cid:107)vk,i(cid:107)2
max · 4α(cid:107)p(cid:107)2
λL − µmax
·
+ µ2
U
+ 4µ2
max ·
λ2
U ·
α
λ2
U
(cid:111)
+ µ2
l=1
k=1
pl
·
pl
l=1
(f )
(g)
+ µ2
max
max
+
1
(h)
1
v
pkE(cid:107)zk,i−1(cid:107)2
U · pT EP [zi−1]
1λ2
that
Using (224) for F+ and F− in (221) -- (222), we conclude
ρ(F+) =(cid:2)ρ(IM − µmaxHc)]2 + O(cid:0)(µmax)
ρ(F−) =(cid:2)ρ(IM − µmaxHc)]2 + O(cid:0)(µmax)
2(M−1)(cid:1) (236)
2(M−1)(cid:1) (237)
1
1
which holds for arbitrarily small . Since the convergence rate
of (cid:107) wc,i(cid:107)2 is between ρ(F+) and ρ(F−), we arrive at (114).
APPENDIX H
PROOF OF LEMMA 4
From the definition in (108), it suffices to establish a joint
inequality recursion for both EP [ wc,i] and EP [we,i]. To begin
with, we state the following bounds on the perturbation terms
in (93).
Lemma 8 (Bounds on the perturbation terms): The
following bounds hold for any i ≥ 0.
(cid:13)(cid:13) ¯P1[AT
1 UL](cid:13)(cid:13)2
P [zi−1] (cid:22) λ2
U ·
P [s(1⊗wc,i−1)] (cid:22) 3λ2
U ·P [ wc,i−1]·1+3λ2
E{P [vi]Fi−1} (cid:22) 4α · 1 · P [ wc,i−1]
∞·11T ·P [we,i−1]
(238)
U(cid:107) wc,0(cid:107)2·1+3go
(239)
+ 4α · (cid:107) ¯P [AT
1 UL](cid:107)2∞ · 11T P [we,i−1]
EP [vi] (cid:22) 4α · 1 · EP [ wc,i−1]
(cid:3)
+(cid:2)4α · ((cid:107) wc,0(cid:107)2 +(cid:107)wo(cid:107)2)+σ2
(cid:3)
+(cid:2)4α · ((cid:107) wc,0(cid:107)2 +(cid:107)wo(cid:107)2)+σ2
where P [ wc,i−1] = (cid:107) wc,i−1(cid:107)2, and go (cid:44) P [s(1 ⊗ wo)].
Proof: See Appendix I.
Now, we derive an inequality recursion for EP [ wc,i] from
· 1 (240)
1 UL](cid:107)2∞ · 11T EP [we,i−1]
·1 (241)
+ 4α · (cid:107) ¯P [AT
v
v
(103). Note that
EP [ wc,i] = E(cid:107) wc,i(cid:107)2
(a)
1
γc
+ µ2
= EP
γc ·
+ (1 − γc) · −µmax
1 − γc
(cid:21)
(Tc(wc,i−1) − Tc( ¯wc,i−1))
(pT ⊗ IM )zi−1
= EP(cid:2)Tc(wc,i−1) − Tc( ¯wc,i−1) − µmax · (pT ⊗ IM )zi−1
− µmax · (pT ⊗ IM )vi)(cid:3)
= EP(cid:2)Tc(wc,i−1) − Tc( ¯wc,i−1) − µmax · (pT ⊗ IM )zi−1
max · EP(cid:2)(pT ⊗ IM )vi)(cid:3)
(cid:20)
max · EP(cid:2)(pT ⊗ IM )vi)(cid:3)
(cid:3)
EP(cid:2)(pT ⊗ IM )zi−1
EP(cid:2)(pT ⊗ IM )vi
(cid:3)
EP(cid:2)(pT ⊗ IM )zi−1
(cid:3)
(cid:3)
EP(cid:2)(pT ⊗ IM )vi
(cid:13)(cid:13)2
E(cid:13)(cid:13)(pT ⊗ IM )zi−1
+ (1 − γc) ·
+ µ2
(cid:22) γc · EP [ wc,i−1] +
+ µ2
= γc · EP [ wc,i−1] +
EP [Tc(wc,i−1) − Tc( ¯wc,i−1)]
µ2
(1 − γc)2
µ2
1 − γc
(cid:22) γc ·
1
γ2
c
+ µ2
max
max
max
max
(c)
(b)
(cid:3)
max
µ2
1 − γc
(cid:3)
· 1T EP [we,i−1]
(242)
v
where step (a) uses the additivity property in Lemma 5 since
the definition of zi−1 and vi in (93) and the definition of
wc,i−1 in (67) imply that zi−1 and wc,i−1 depend on all {wj}
(cid:18)
(cid:19)
(cid:13)(cid:13)2
(cid:13)(cid:13)2
U N
1
11T
α
λ2
U
+
2 ]
∞ λ2
(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13) ¯P [URAT
×
1−λ2(A)
+ 4µ2
max ·
· 1 · E(cid:107) wc,i−1(cid:107)2
+ µ2
max·N
+4α((cid:107) wc,0(cid:107)2 +(cid:107)wo(cid:107)2)+σ2
∞·
(cid:20)
2 ]
12
v
(cid:21)
(cid:21)
(cid:18)
·EP [we,i−1]
3
1 − λ2(A)
(cid:19)
+
α
λ2
U
U(cid:107) wc,0(cid:107)2 +(cid:107)go(cid:107)∞
λ2
1 − λ2(A)
·1
(243)
where step (a) uses the additivity property in Lemma 5 since
the definition of zi−1 and vi in (93) and the definitions of
wc,i−1 and we,i−1 in (67) imply that zi−1, wc,i−1 and we,i−1
depend on all {wj} for j ≤ i−1, meaning that the cross terms
between vi and all other terms are zero, just as in step (a) of
(242), step (b) uses the variance relation of stable Kronecker
Jordan operators from (172) with d2 = λ2(A), step (c) uses
the variance relation of linear operator (161), step (d) uses the
submultiplictive property (152) and P [x + y] (cid:22) 2P [x] + 2P [y]
derived from the convexity property (148) and the scaling
property in (238), (239), and (241), step (e) substitutes the
bounds on the perturbation terms from (238) -- (241), and step
(f) uses the inequality 1T go ≤ N(cid:107)go(cid:107)∞.
rewrite recursions (242) and (243) as
EP [ wc,i] (cid:22) (γc +µ2
Finally, using the quantities defined in (120) -- (123), we can
maxψ0)·EP [ wc,i−1]
maxψ0) · 1T EP [we,i−1]
(244)
+(µmaxhc(µmax)+µ2
+µ2
maxbv,c
EP [we,i] (cid:22) µ2
maxψ01 · EP [ wc,i−1]
+ (Γe + µ2
+ µ2
maxψ011T ) · EP [we,i−1]
maxbv,e · 1
(245)
where EP [ wc,i] = E(cid:107) wc,i(cid:107)2. Using the matrices and vectors
defined in (116) -- (118), we can write the above two recursions
in a joint form as in (115).
APPENDIX I
PROOF OF LEMMA 8
First, we establish the bound for P [zi−1] in (238). Substi-
tuting (69) and (88) into the definition of zi−1 in (93) we
get:
for j ≤ i − 1, meaning that the cross terms are zero:
E[vizT
i−1] = E(cid:8)E [viFi−1] zT
(cid:9) = 0
E(cid:8)vi[Tc(wc,i−1) − Tc( ¯wc,i−1)]T(cid:9)
= E(cid:8)E [viFi−1] [Tc(wc,i−1)−Tc( ¯wc,i−1)]T(cid:9) = 0
i−1
(cid:3)
+
+
2
2
(a)
(b)
2 M
URAT
URAT
− URAT
Step (b) uses the convexity property in Lemma 5, step (c) uses
the variance property (164), step (d) uses the notation zk,i−1
and vk,i to denote the kth M × 1 block of the N M × 1 vector
zi−1 and vi, respectively, step (e) applies Jensen's inequality
to the convex function (cid:107) · (cid:107)2, step (f) substitutes expression
(166) for γc, step (g) substitutes the bounds for the perturbation
terms from (238), (239), and (241), step (h) uses the fact that
pT 1 = (cid:107)p(cid:107)1.
for we,i in (101):
EP [we,i]
Next, we derive the bound for EP [we,i] from the recursion
2 M s(1 ⊗ wc,i−1)
2 Mvi
2 M (s(1 ⊗ wc,i−1) + zi−1)(cid:3)
(cid:1)(cid:3)
(cid:0)s(1 ⊗ wc,i−1)+zi−1
2 M](cid:13)(cid:13)2
DN−1we,i−1 − URAT
2 Mzi−1 − URAT
DN−1we,i−1 − URAT
URAT
= EP(cid:2)
= EP(cid:2)
+ EP(cid:2)
(cid:3)
2 Mvi
1 − λ2(A) ·EP(cid:2)
(cid:22) Γe · EP [we,i−1]
+EP(cid:2)
(cid:3)
2 Mvi
(cid:13)(cid:13) ¯P [URAT
(cid:22) Γe · EP [we,i−1]
1 − λ2(A) ·
2 M](cid:13)(cid:13)2
· 11T · EP [s(1 ⊗ wc,i−1) + zi−1]
∞ · 11T · EP [vi]
2 ](cid:13)(cid:13)2
4(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13)2
∞ · 11T · EP [vi]
(cid:19)
1 UL]
2 ]
(cid:13)(cid:13) ¯P [URAT
2 ](cid:13)(cid:13)2
(cid:20)
(cid:13)(cid:13) ¯P [URAT
(cid:13)(cid:13)2
(cid:13)(cid:13) ¯P [URAT
2 ](cid:13)(cid:13)2
+(cid:13)(cid:13) ¯P [URAT
(cid:8)EP [s(1 ⊗ wc,i−1)]+EP [zi−1](cid:9)
(cid:13)(cid:13) ¯P [AT
(cid:13)(cid:13)2
(cid:21)
(cid:18)
∞ λ2
(cid:18)
·EP [we,i−1]
×
1−λ2(A)
+ 4µ2
max ·
· 1 · E(cid:107) wc,i−1(cid:107)2
+ µ2
max·
2 ]
(cid:21)
(cid:13)(cid:13) ¯P [AT
1 UL](cid:13)(cid:13)2
(cid:22) Γe · EP [we,i−1]
+µ2
1 − λ2(A)
·1
∞ λ2
+N [4α((cid:107) wc,0(cid:107)2 +(cid:107)wo(cid:107)2)+σ2
v]
U(cid:107) wc,0(cid:107)2N +1T go
λ2
(cid:19)
α
λ2
U
max ·
Γe +4µ2
1 − λ2(A)
1−λ2(A)
∞·
11T
max·
1
·
+ µ2
Γe +4µ2
·11T
max ·
max·
∞ λ2
U N
α
λ2
U
(f )
(cid:22)
(e)
(cid:22)
∞·
∞·
(cid:20)
(cid:20)
U N
U N
2 ]
+
3
+
12
(c)
(d)
∞
∞
24
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
P [zi−1]
(cid:22) P(cid:2)s(cid:0)1⊗wc,i−1 +(AT
U · P(cid:2)(AT
(cid:13)(cid:13) ¯P [AT
(cid:13)(cid:13)2
(cid:22) λ2
(cid:22) λ2
U ·
(a)
(b)
1 UL⊗IM )we,i−1
(cid:3)
1 UL ⊗ IM )we,i−1
1 UL]
∞ · 11T · P [we,i−1]
(cid:1)
−s(1⊗wc,i−1)(cid:3)
(246)
where step (a) uses the variance relation (163), and step (b)
uses property (161).
Next, we prove the bound on P [s(1 ⊗ wc,i−1)]. It holds
that
P [s(1 ⊗ wc,i−1)]
(cid:104) 1
3 · 3(cid:0)s(1 ⊗ wc,i−1) − s(1 ⊗ ¯wc,i−1)(cid:1)
= P
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
25
(a)
1
1
+
+
(cid:22)
(cid:105)
1
3 · 3 · s(1 ⊗ wo)
3 · 3(cid:0)s(1 ⊗ ¯wc,i−1) − s(1 ⊗ wo)(cid:1)
3 · P(cid:2)3(cid:0)s(1 ⊗ wc,i−1) − s(1 ⊗ ¯wc,i−1)(cid:1)(cid:3)
3 · P(cid:2)3(cid:0)s(1 ⊗ ¯wc,i−1) − s(1 ⊗ wo)(cid:1)(cid:3)
3 · P(cid:2)3 · s(1 ⊗ wo)(cid:3)
= 3P(cid:2)s(1 ⊗ wc,i−1)−s(1 ⊗ ¯wc,i−1)(cid:3)
+ 3P(cid:2)s(1 ⊗ ¯wc,i−1)−s(1 ⊗ wo)(cid:3) + 3P(cid:2)s(1 ⊗ wo)(cid:3)
U · P(cid:2)1 ⊗ (wc,i−1 − ¯wc,i−1)(cid:3)
+
+
1
1
(b)
U · P [1 ⊗ ( ¯wc,i−1 − wo)] + 3P [s(1 ⊗ wo)]
U · (cid:107) wc,i−1(cid:107)2 · 1 + 3λ2
U · (cid:107) ¯wc,i−1 − wo(cid:107)2 · 1
(c)
(cid:22) 3λ2
+ 3λ2
(d)
= 3λ2
+ 3P [s(1 ⊗ wo)]
(cid:22) 3λ2
(e)
U · (cid:107) wc,i−1(cid:107)2·1+3λ2
U(cid:107) wc,0(cid:107)2·1+3P [s(1 ⊗ wo)]
(247)
where step (a) uses the convexity property (148), step (b) uses
the scaling property in Lemma 5, step (c) uses the variance
relation (163), step (d) uses property (154), and step (e) uses
the bound (110) and the fact that γc < 1.
Finally, we establish the bounds on P [vi] in (240) -- (241).
Introduce the M N × 1 vector x:
(a)
x (cid:44) 1 ⊗ wc,i−1 + AT
1 ULwe,i−1 ≡ φi−1
(cid:107)s1,i(x1) − s1(x1)(cid:107)2(cid:12)(cid:12)Fi−1
= col(cid:8)E(cid:2)
(cid:3) ,
(cid:107)sN,i(xN ) − sN (xN )(cid:107)2(cid:12)(cid:12)Fi−1
. . . , E(cid:2)
(cid:22) col(cid:8)α · (cid:107)x1(cid:107)2 +σ2
(248)
We partition x in block form as x = col{x1, . . . , xN}, where
each xk is M × 1. Then, by the definition of vi from (93),
we have
E{P [vi]Fi−1} = E{P [si(x) − s(x)]Fi−1}
(cid:3)(cid:9)
(cid:9)
v, . . . , α · (cid:107)xN(cid:107)2 +σ2
(249)
(cid:3)
1 ULwe,i−1
4 ·4·1 ⊗ wo(cid:105)
4 ·4·1 ⊗ (wc,i−1− ¯wc,i−1)+
1
4 ·4·AT
+
4 ·4·1 ⊗ wo(cid:105)
4 ·4·1 ⊗ wc,i−1 +
1
4 ·4·AT
+
P [x] = P(cid:2)1 ⊗ wc,i−1 + AT
1 ULwe,i−1 +
1
4 ·4·1 ⊗ wc,i−1
where step (a) uses Assumption (18). Now we bound P [x]:
= α · P [x] + σ2
(cid:104) 1
(cid:104) 1
1
4 ·4·1 ⊗ ( ¯wc,i−1−wo)
= P
= P
1
1
1
v
v
(a)
(cid:22)
1 ULwe,i−1 +
1
4 · 42·P [1 ⊗ wc,i−1] +
1
4 · 42·P [AT
+
1 ULwe,i−1] +
1
4 · 42·P [1 ⊗ wc,i−1]
1
4 · 42·P [1 ⊗ wo]
(b)
= 4 · (cid:107) wc,i−1(cid:107)2 · 1 + 4 · (cid:107) wc,i−1(cid:107)2 · 1
+ 4 · P [AT
1 ULwe,i−1] + 4 · (cid:107)wo(cid:107)2 · 1
(c)
(cid:22) 4 · (cid:107) wc,i−1(cid:107)2·1+4·(cid:107) ¯P [AT
+ 4 · (cid:107) wc,0(cid:107)2 · 1 + 4 · (cid:107)wo(cid:107)2 · 1
1 UL](cid:107)2∞·11T ·P [we,i−1]
(250)
where step (a) uses the convexity property (148) and the
scaling property in Lemma 5, step (b) uses the Kronecker
property (154), step (c) uses the variance relation (160) and the
bound (110). Substituting (250) into (249), we obtain (240),
and taking expectation of (240) with respect to Fi−1 leads to
(241).
APPENDIX J
PROOF OF THEOREM 4
Assume initially that the matrix Γ is stable (we show further
ahead how the step-size parameter µmax can be selected to
ensure this property). Then, we can iterate the inequality
recursion (115) and obtain
i−1(cid:88)
j=0
max
(cid:48)
0 + µ2
max
W
(cid:48)
i (cid:22) Γi W
∞(cid:88)
j=0
Γj W
(cid:22)
(cid:22) (I − Γ)
(cid:48)
0 + µ2
−1( W
Γjbv
∞(cid:88)
Γjbv
(cid:48)
0 + µ2
j=0
maxbv)
(251)
where the first two inequalities use the fact that all entries of
Γ are nonnegative. Moreover, substituting (116) into (115), we
get
(cid:48)
i (cid:22) Γ0 W
W
(cid:48)
i−1 + µ2
maxψ011T W
(cid:48)
i−1 + µ2
maxbv
(252)
Substituting (251) into the second term on the right-hand side
of (252) leads to
(cid:48)
W
i (cid:22) Γ0 W
max · cv(µmax)
(cid:48)
i−1 + µ2
(253)
where
cv(µmax) (cid:44) ψ0 · 1T (I − Γ)
−1( W
(cid:48)
0 + µ2
maxbv) · 1 + bv
(254)
Now iterating (253) leads to the following non-asymptotic
bound:
maxΓj
µ2
0 · cv(µmax) (cid:22) Γi
0
W
(cid:48)
(cid:48)
0 + W ub
∞
i−1(cid:88)
j=0
(cid:48)
Wi
(cid:22) Γi
0
(cid:48)
0 +
W
where
(255)
(256)
(cid:48)
∞ (cid:44) µ2
W ub
max(I − Γ0)
−1 · cv(µmax)
W
We now derive the non-asymptotic bounds (124) -- (125) from
(255). To this end, we need to study the structure of the
(cid:48)
term Γi
0. Our approach relies on applying the unilateral
0
0, i ≥ 0} to get
z−transform to the causal matrix sequence {Γi
−1
(257)
since Γ0 is a stable matrix. Note from (117) that Γ0 is a 2 ×
2 block upper triangular matrix. Substituting (117) into the
above expression and using the formula for inverting 2 × 2
(cid:9) = z(zI − Γ0)
Γ0(z) (cid:44) Z
(cid:8)Γi
0
26
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
block upper triangular matrices (see formula (4) in [58, p.48]),
we obtain
where
(cid:35)
(cid:34) z
z−γc
0
Γ0(z) =
µmaxhc(µmax) ·
z
z−γc · 1T (zI − Γe)−1
z(zI − Γe)−1
(258)
Next we compute the inverse z−transform to obtain Γi
0. Thus,
observe that the inverse z−transform of the (1, 1) entry, the
(2, 1) block, and the (2, 2) block are the causal sequences γi
c,
0, and Γi
e, respectively. For the (1, 2) block, it can be expressed
in partial fractions as
z
(cid:18) z
z − γc · 1T (zI − Γe)
−1(cid:0)γi
µmaxhc(µmax) ·
= µmaxhc(µmax)·1T (γcI−Γe)
from which we conclude that the inverse z−transform of the
(1, 2) block is
i ≥ 0 (259)
µmaxhc(µmax)·1T (γcI−Γe)
(cid:35)
(cid:1)
It follows that
(cid:1) ,
c µmaxhc(µmax) · 1T (γcI − Γe)−1(cid:0)γi
I−z(zI−Γe)
cI − Γi
cI−Γi
z−γc
(cid:19)
(cid:34)
−1
−1
−1
e
e
Γi
0 =
γi
0
Γi
e
¯wc,0 = (cid:80)N
(260)
Furthermore, as indicated by (48) in Sec. IV-A, the reference
recursion (47) is initialized at the centroid of the network, i.e.,
k=1 θkwk,0. This fact, together with (68) leads to
¯wc,0 = wc,0, which means that wc,0 = 0. As a result, we get
(cid:48)
0:
the following form for W
(cid:48)
0 = col{0, EP [we,0]}
W
Multiplying (260) to the left of (261) gives
µmaxhc(µmax) · 1T (γcI−Γe)−1(cid:0)γi
(cid:34)
Γi
0
(cid:48)
0 =
W
Γi
eWe,0
(261)
(cid:35)
(cid:1)
cI−Γi
e
We,0
(262)
where We,0 = EP [we,0]. Substituting (262) into (255), we
(cid:35)
(cid:1) EP [we,0]
obtain
(cid:34)
µmaxhc(µmax)·1T (γcI−Γe)−1(cid:0)γi
cI−Γi
e
(cid:48)
i (cid:22)
W
EP [we,0]
Γi
e
(cid:48)
+ W ub
∞
(cid:26)
(cid:104)
1 − ψ0 ·
(cid:34)
(cid:34)
·
+
µmax
ζ(µmax) =
(cid:18)
+µ2
1+
(cid:19)
µmax
λL− µmax
2 (cid:107)p(cid:107)2
1T (I−Γe)
1λ2
U
hc(µmax)
2 (cid:107)p(cid:107)2
(cid:105)(cid:27)−1
−11
(266)
max
1λ2
U
λL− µmax
Proof: See Appendix K.
Substituting (254), (261), (264) and (265) into (256) and after
some algebra, we obtain
W ub
∞ = ψ0·ζ(µmax)f (µmax)
λL− µmax
µmax
(cid:35)
−11
(cid:48)
(cid:35)
−11bv,e
1λ2
U
λL− µmax
1+µmaxhc(µmax)1T (I−Γe)
2 (cid:107)p(cid:107)2
max · (I − Γe)−11
µ2
bv,c+µmaxhc(µmax)1T (I−Γe)
2 (cid:107)p(cid:107)2
maxbv,e · (I − Γe)−11
(cid:19)
1λ2
U
1λ2
U
µ2
(cid:1)
·1T (I−Γe)
1λ2
U
1bv,e
max
(267)
−1
(268)
where
f (µmax) (cid:44)
(cid:18)
λL− µmax
+
µmaxbv,c
2 (cid:107)p(cid:107)2
λL− µmax
1+
hc(µmax)
2 (cid:107)p(cid:107)2
(cid:0)EP [we,0]+µ2
Introduce
c,∞ (cid:44) ψ0·ζ(µmax)f (µmax)
W ub
×
(cid:48)
· µmax
+ µmax
1 + µmaxhc(µmax)1T (I − Γe)−11
bv,c + µmaxhc(µmax)1T (I − Γe)−11bv,e
(269)
λL − µmax
λL − µmax
2 (cid:107)p(cid:107)2
2 (cid:107)p(cid:107)2
1λ2
U
1λ2
U
(270)
(271)
(cid:48)
e,∞ (cid:44) ψ0·ζ(µmax)f (µmax) · µ2
W ub
−11
maxbv,e · (I − Γe)
+ µ2
max · (I − Γe)
−11
Then, we have
Substituting (271) into (263), we conclude (124) -- (125). Now,
to prove (126) -- (127), it suffices to prove
(cid:48)
W ub
∞ = col{ W ub
c,∞, W ub
e,∞}
(cid:48)
(cid:48)
(cid:0)λL +hc(0)(cid:1)1T (I−Γe)−1We,0 +bv,cλL
(cid:0)λL +hc(0)(cid:1)1T (I − Γe)−1We,0 +bv,eλL
(272)
λ2
L
λL
(273)
−11
· (I−Γe)
(263)
(cid:48)
W ub
c,∞
µmax
(cid:48)
W ub
e,∞
µ2
max
ψ0
=
ψ0
=
lim
µmax→0
lim
µmax→0
Substituting (269) and (270) into the left-hand side of (272)
and (273), respectively, we get
(cid:48)
Finally, we study the behavior of the asymptotic bound
W ub
∞ by calling upon the following lemma.
Lemma 9 (Useful matrix expressions): It holds that
1T (I − Γ)
(cid:104)
= ζ(µmax)
−1
·
max
λL− µmax
2 (cid:107)p(cid:107)2
1T (I − Γe)−1(cid:105)
hc(µmax)
2 (cid:107)p(cid:107)2
λL− µmax
(cid:16)
(cid:17)
1 +
1λ2
U
1λ2
U
−1
µ
(cid:34)
−1
(I − Γ0)
=
µ
−1
max
λL− µmax
2 (cid:107)p(cid:107)2
0
1λ2
U
lim
µmax→0
(cid:48)
(cid:26)
W ub
c,∞
µmax
= lim
µmax→0
(264)
(cid:35)
λL−µmax
hc(µmax)
2 (cid:107)p(cid:107)2
1T (I−Γe)−1
1λ2
U
(I − Γe)−1
(265)
ψ0·ζ(µmax)f (µmax)·
1+µmaxhc(µmax)1T (I−Γe)−11
λL − µmax
2 (cid:107)p(cid:107)2
1λ2
U
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
27
·
(a)
+
λL
1+
λ2
L
1λ2
U
ψ0
=
2 (cid:107)p(cid:107)2
1
λL
hc(0)
−1EP [we,0]
= ψ0·ζ(0)f (0) ·
= ψ0·1 ·
λL − µmax
(cid:19)
+
bv,c
λL
·1T (I−Γe)
bv,c + µmaxhc(µmax)1T (I − Γe)−11bv,e
(cid:20)(cid:18)
(cid:21)
(cid:0)λL + hc(0)(cid:1)1T (I − Γe)−1We,0 + bv,cλL
(cid:8)ψ0·ζ(µmax)f (µmax) · (I − Γe)
−11(cid:9)
(cid:20)(cid:18)
(cid:21)
−11 + bv,e · (I − Γe)
−1EP [we,0]
(cid:0)λL + hc(0)(cid:1)1T (I − Γe)−1We,0 + bv,eλL
+ bv,e · (I − Γe)
= ψ0·ζ(0)f (0) · (I − Γe)
= ψ0·
λL
+ bv,e · (I − Γe)
ψ0
·1T (I−Γe)
−11
(cid:48)
W ub
e,∞
µ2
µmax→0
= lim
µmax→0
(cid:19)
−11
hc(0)
lim
1+
max
(b)
−11
=
λL
−11
(275)
where steps (a) and (b) use the expressions for ζ(µmax) and
f (µmax) from (266) and (268).
·(I−Γe)
Now we proceed to prove (128). We already know that
the second term on the right-hand side of (124), W ub
c,∞, is
O(µmax) because of (126). Therefore, we only need to show
that the first term on the right-hand side of (124) is O(µmax)
(cid:1)
for all i ≥ 0. To this end, it suffices to prove that
cI−Γi
(cid:13)(cid:13)µmaxhc(µmax)·1T (γcI−Γe)−1(cid:0)γi
We,0
(cid:13)(cid:13)
lim
e
(cid:48)
µmax→0
µmax
≤ constant
(276)
where the constant on the right-hand side should be indepen-
dent of i. This can be proved as below:
cI−Γi
e
We,0
(cid:1)
(cid:1)
We,0
(cid:13)(cid:13)
(cid:13)(cid:13)
lim
µmax
µmax→0
= lim
(cid:13)(cid:13)µmaxhc(µmax)·1T (γcI−Γe)−1(cid:0)γi
(cid:13)(cid:13)hc(µmax)·1T (γcI−Γe)
−1(cid:0)γi
cI−Γi
µmax→0
(cid:0)(cid:13)(cid:13)γi
cI(cid:13)(cid:13) +(cid:13)(cid:13)Γi
−1(cid:107)
µmax→0hc(µmax) · (cid:107)1(cid:107) · (cid:107)(γcI−Γe)
≤ lim
·
(cid:16)
µmax→0hc(µmax) · N · (cid:107)(γcI−Γe)
·
· (cid:107)We,0(cid:107)
µmax→0hc(µmax)·N·(cid:107)(γcI−Γe)
(cid:13)(cid:13)(cid:1)
(cid:0)ρ(Γe) + (cid:1)i(cid:17)
· (cid:107)We,0(cid:107)
≤ lim
1 + Ce ·
−1(cid:107)
(a)
(b)
e
e
(c)
−1(cid:107)·(1+Ce)·(cid:107)We,0(cid:107)
−1(cid:107) · [1 + Ce] · (cid:107)We,0(cid:107)
≤ lim
= hc(0) · N · (cid:107)(I−Γe)
= constant
L < 1 for
where step (a) uses γc = 1 − µmaxλL + 1
sufficiently small step-sizes, and uses the property that for
any small > 0 there exists a constant C such that (cid:107)X i(cid:107) ≤
C · [ρ(X) + ]i for all i ≥ 0 [49, p.38], step (b) uses the
fact that ρ(Γe) = λ2(A) < 1 so that ρ(Γe) + < 1 for
maxλ2
(277)
2 µ2
(cid:27)
1
λL
+
bv,c
λL
(274)
−11
· (I − Γe)
2 µ2
maxλ2
L → 1 when µmax → 0.
small (e.g., = (1 − ρ(Γe))/2), and step (c) uses γc =
1 − µmaxλL + 1
It remains to prove that condition (129) guarantees the
stability of the matrix Γ, i.e., ρ(Γ) < 1. First, we intro-
duce the diagonal matrices D,0 (cid:44) diag{,··· , N−1} and
D = diag{1, D,0}, where is chosen to be
1
4
(1 − λ2(A))2 ≤
It holds that ρ(Γ) = ρ(D−1
ΓD) since similarity transforma-
tions do not alter eigenvalues. By the definition of Γ in (116),
we have
−1
ΓD = D
−1
Γ0D + µ2
(cid:44) 1
4
11T D
(278)
(279)
−1
D
maxψ0 · D
We now recall that the spectral radius of a matrix is upper
bounded by any of its matrix norms. Thus, taking the 1−norm
(the maximum absolute column sum of the matrix) of both
sides of the above expression and using the triangle inequality
and the fact that 0 < ≤ 1/4, we get
ρ(Γ) = ρ(D
(cid:16)
≤ (cid:107)D
≤ (cid:107)D
= (cid:107)D
1 − −N
= (cid:107)D
1 − −1
(−N − 1)
= (cid:107)D
1 −
4 (−N − 1)
1 − 1
−N
−1
ΓD)
−1
Γ0D(cid:107)1 + (cid:107)µ2
−1
Γ0D(cid:107)1 + µ2
−1
Γ0D(cid:107)1 + µ2
−1
Γ0D(cid:107)1 + µ2
−1
Γ0D(cid:107)1 + µ2
−1
Γ0D(cid:107)1 + µ2
maxψ0 ·
−1
1
µ2
Γ0D(cid:107)1 +
3
maxψ0 · D
maxψ0 · (cid:107)D
maxψ0 ·
maxψ0 ·
maxψ0 ·
11T D(cid:107)1
11T D(cid:107)1
−1 + ··· +
−(N−1)(cid:17)
≤ (cid:107)D
≤ (cid:107)D
−1
−1
maxψ0
(280)
1 +
4
1
Moreover, we can use (117) to write:
−1
Γ0D =
D
γc µmaxhc(µmax)1T D,0
0
−1
,0 ΓeD,0
D
(281)
1−λ2(A)
λ2(A)
where (recall the expression for Γe from (171) where we
replace d2 by λ2(A)):
(282)
µmaxhc(µmax)1T D,0 = µmaxhc(µmax)(cid:2) ··· N−1(cid:3) (283)
−1
,0 ΓeD,0 =
λ2(A)
1−λ2(A)
...
...
...
D
2
2
Therefore, the 1-norm of D−1
Γ0D can be evaluated as
(cid:35)
(cid:34)
(cid:110)
−1
,0 Γ0D,0(cid:107)1 = max
(cid:107)D
γc, λ2(A) + µmaxhc(µmax),
+ µmaxhc(µmax)2,··· ,
+ µmaxhc(µmax)N−1(cid:111)
1 + λ2(A)
1 + λ2(A)
2
2
(284)
28
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
1 + λ2(A)
2
≤
= 1
+
1−λ2(A)
4
+
1−λ2(A)
4
(294)
Substituting (121) and (278) into (292), we find that the latter
conditions are satisfied for
(cid:0)
λL
(cid:107) ¯P [ATU2](cid:107)2∞ + 1
2
(cid:1)
0 < µmax <
0 < µmax <
(cid:107)p(cid:107)2
1λ2
U
(cid:115)
3(1 − λ2(A))2N +1
22N +2ψ0
(295)
Combining (287), (289) and (295) , we arrive at condition
(129).
APPENDIX K
PROOF OF LEMMA 9
Applying the matrix inversion lemma [52] to (116), we get
(I − Γ)
−1 = (I − Γ0 − µ2
−1 +
= (I−Γ0)
−1
maxψ0 · 11T )
maxψ0 · (I−Γ0)−111T (I−Γ0)−1
µ2
maxψ0 · 1T (I − Γ0)−11
1−µ2
(296)
so that
1T (I − Γ)
−1 =
−1
1
(cid:34)
1 − µ2
maxψ0 · 1T (I − Γ0)−11 · 1T (I − Γ0)
(297)
By (117), the matrix Γ0 is a 2×2 block upper triangular matrix
(cid:35)
whose inverse is given by
1T (I − Γe)−1)
(1 − γc)−1
(298)
Substituting (166) into the above expression leads to (265).
Furthermore, from (265), we have
(I − Γe)−1
(I − Γ0)
−1 =
µmaxhc(µmax)
1−γc
0
Since 0 < ≤ 1/4, we have > 2 > ··· > N−1 > 0.
Therefore,
−1
,0 Γ0D,0(cid:107)1 = max
(cid:107)D
γc, λ2(A) + µmaxhc(µ),
+ µmaxhc(µmax)2(cid:111)
1 + λ2(A)
(285)
(cid:110)
2
(cid:110)
Substituting the above expression for (cid:107)D
(280) leads to
−1
,0 Γ0D,0(cid:107)1 into
ρ(Γ) ≤ max
1+λ2(A)
2
+µmaxhc(µmax)2(cid:111)
γc, λ2(A) + µmaxhc(µmax),
µ2
+
−N
maxψ0
1
3
1
3
(286)
We recall from (167) that γc > 0. To ensure ρ(Γ) < 1,
it suffices to require that µmax is such that the following
conditions are satisfied:
−N < 1
(287)
µ2
γc +
maxψ0
1
3
−N < 1
µ2
maxψ0
(288)
λ2(A) + µmaxhc(µmax) +
−N < 1 (289)
1 + λ2(A)
We now solve these three inequalities to get a condition on
µmax. Substituting the expression for γc from (166) into (287),
we get
+ µmaxhc(µmax)2 +
maxψ0
µ2
1
3
2
(cid:16) 1
1 − µmaxλL + µ2
max
−N +
ψ0
1
2(cid:107)p(cid:107)2
1λ2
U
3
the solution of which is given by
0 < µmax <
λL
3 ψ0−N + 1
1
2(cid:107)p(cid:107)2
1λ2
U
(cid:17)
< 1
(290)
(291)
For (288) -- (289), if we substitute the expression for hc(µmax)
from (121) into (288) -- (289), we get a third-order inequality
in µmax, which is difficult to solve in closed-form. However,
inequalities (288) -- (289) can be guaranteed by the following
conditions:
µmaxhc(µmax) <
(1−λ2(A))2
4
,
maxψ0−N
µ2
3
This is because we would then have:
1
3
λ2(A) + µmaxhc(µmax) +
−N
maxψ0
µ2
(1−λ2(A))2
1−λ2(A)
+
4
4
< 1
< λ2(A) +
≤ λ2(A) +
1+λ2(A)
=
2
<
1−λ2(A)
4
(292)
1T (I − Γ0)
−11 =
max
µ−1
(cid:18)
λL− µmax
2 (cid:107)p(cid:107)2
+
λL− µ
1+
hc(µmax)
2(cid:107)p(cid:107)2
1λ2
U
1λ2
U
(cid:19)
1T (I−Γe)
−11
(299)
1−λ2(A)
+
4
1−λ2(A)
4
Substituting (299) into (297), we obtain (264).
APPENDIX L
PROOF OF THEOREM 5
(293)
Taking the squared Euclidean norm of both sides of (74)
and applying the expectation operator, we obtain
Likewise, by the fact that 0 < ≤ 1/4 < 1,
µ2
+ µmaxhc(µmax)2 +
1 + λ2(A)
−N
maxψ0
1
3
2
1 + λ2(A)
1 + λ2(A)
2
<
<
2
+ µmaxhc(µmax) +
+
(1−λ2(A))2
4
+
µ2
1
maxψ0
3
1−λ2(A)
4
−N
E(cid:107) wk,i(cid:107)2 = (cid:107) wc,i(cid:107)2 + E(cid:107) wc,i + (uL,k ⊗ IM )we,i(cid:107)2
− 2 wT
c,i [E wc,i + (uL,k ⊗ IM )we,i]
(300)
which means that, for all i ≥ 0,
(cid:12)(cid:12)E(cid:107) wk,i(cid:107)2 − (cid:107) wc,i(cid:107)2(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)E(cid:107) wc,i + (uL,k ⊗ IM )we,i(cid:107)2
=
CHEN AND SAYED: ON THE LEARNING BEHAVIOR OF ADAPTIVE NETWORKS -- PART I: TRANSIENT ANALYSIS
29
(cid:3)(cid:12)(cid:12)(cid:12)
(b)
(c)
(a)
c,i
(cid:2)E wc,i + (uL,k ⊗ IM )Ewe,i
− 2 wT
(cid:2)
(cid:3)
≤ E(cid:107) wc,i + (uL,k ⊗ IM )we,i(cid:107)2
(cid:107)E wc,i(cid:107) + (cid:107)uL,k ⊗ IM(cid:107) · (cid:107)Ewe,i(cid:107)
+ 2(cid:107) wc,i(cid:107) ·
(cid:3)
(cid:2)E(cid:107) wc,i(cid:107) + (cid:107)uL,k ⊗ IM(cid:107) · E(cid:107)we,i(cid:107)
≤ 2E(cid:107) wc,i(cid:107)2 + 2(cid:107)uL,k ⊗ IM(cid:107)2 · E(cid:107)we,i(cid:107)2
+ 2(cid:107) wc,i(cid:107) ·
(cid:2)(cid:113)E(cid:107) wc,i(cid:107)2 + (cid:107)uL,k ⊗ IM(cid:107) ·
(cid:113)E(cid:107)we,i(cid:107)2(cid:3)
≤ 2E(cid:107) wc,i(cid:107)2 + 2(cid:107)uL,k ⊗ IM(cid:107)2 · E(cid:107)we,i(cid:107)2
+ 2(cid:107) wc,i(cid:107) ·
(cid:113)
(cid:104)(cid:113)EP [ wc,i] + (cid:107)uL,k ⊗ IM(cid:107) ·
= 2EP [ wc,i] + 2(cid:107)uL,k ⊗ IM(cid:107)2 · 1T EP [we,i]
e,∞(cid:1)
(cid:0)1T Γi
+ 2(cid:107) wc,i(cid:107) ·
(cid:104)
eWe,0 + 1T W ub
≤ O(µmax) + 2(cid:107)uL,k ⊗ IM(cid:107)2 ·
(cid:113)
(cid:105)
+ 2γi
max)(cid:1)
(cid:19)(cid:21)
≤ O(µmax) + 2(cid:107)uL,k ⊗ IM(cid:107)2 ·
+ 2γi
(cid:0)1T Γi
eWe,0 + 1T W ub(cid:48)
e,∞
eWe,0 + O(µ2
+ (cid:107)uL,k ⊗ IM(cid:107) ·
1T EP [we,i]
c(cid:107) wc,0(cid:107) ·
c(cid:107) wc,0(cid:107) ·
1T Γi
max)
max)
(cid:20)
O(µ
O(µ
(cid:18)(cid:113)
1
2
1
2
(cid:48)
(f )
(d)
(e)
(cid:112)
(cid:105)
1T Γi
eWe,0 +
O(µ2
max)
+ (cid:107)uL,k ⊗ IM(cid:107) ·
= 2(cid:107)uL,k ⊗ IM(cid:107)2 · 1T Γi
eWe,0
(cid:113)
where step (a) used Cauchy-Schwartz inequality, step (b) used
(cid:107)x+y(cid:107)2 ≤ 2(cid:107)x(cid:107)2 +2(cid:107)y(cid:107)2, step (c) applied Jensen's inequality
to the concave function √·, step (d) used property (157), step
(e) substituted the non-asymptotic bounds (124) and (125) and
the fact that EP [ wc,i] ≤ O(µmax) for all i ≥ 0 from (128),
step (f) used (127) and the fact that √x + y ≤ √x + √y
for x, y ≥ 0, and step (g) used γc < 1 for sufficiently small
step-sizes (guaranteed by (112)).
REFERENCES
[1] J. Chen and A. H. Sayed,
"On the limiting behavior of distributed
optimization strategies," in Proc. Allerton Conf., Monticello, IL, Oct.
2012, pp. 1535 -- 1542.
[2] S. Barbarossa and G. Scutari, "Bio-inspired sensor network design,"
IEEE Signal Process. Mag., vol. 24, no. 3, pp. 26 -- 35, May 2007.
[3] 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.
[4] A. Nedic and D. P. Bertsekas, "Incremental subgradient methods for
nondifferentiable optimization," SIAM J. Optim., vol. 12, no. 1, pp.
109 -- 138, 2001.
[5] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, "Distributed asyn-
chronous deterministic and stochastic gradient optimization algorithms,"
IEEE Trans. Autom. Control, vol. 31, no. 9, pp. 803 -- 812, 1986.
[6] 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.
[7] S. Kar, J. M. F. Moura, and K. Ramanan, "Distributed parameter esti-
mation in sensor networks: Nonlinear observation models and imperfect
communication," IEEE Trans. Inf. Theory, vol. 58, no. 6, pp. 3575 -- 3605,
Jun. 2012.
[8] S. Kar, J. M. F. Moura, and H. V. Poor, "Distributed linear parameter
estimation: Asymptotically efficient adaptive strategies," SIAM Journal
on Control and Optimization, vol. 51, no. 3, pp. 2200 -- 2229, 2013.
[9] 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, Nov. 2010.
[10] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-
agent optimization," IEEE Trans. Autom. Control, vol. 54, no. 1, pp.
48 -- 61, 2009.
[11] A. Nedic and A. Ozdaglar, "Cooperative distributed multi-agent opti-
mization," Convex Optimization in Signal Processing and Communica-
tions, Y. Eldar and D. Palomar (Eds.), Cambridge University Press, pp.
340 -- 386, 2010.
[12] C. Eksin and A. Ribeiro,
"Distributed network optimization with
heuristic rational agents," IEEE Trans. Signal Proc., vol. 60, no. 10,
pp. 5396 -- 5411, Oct. 2012.
[13] C. Eksin, P. Molavi, A. Ribeiro, and A. Jadbabaie, "Learning in network
games with incomplete information: Asymptotic analysis and tractable
implementation of rational behavior," IEEE Signal Process. Mag., vol.
30, no. 3, pp. 30 -- 42, May 2013.
[14] 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.
[15] D. H. Dini and D. P. Mandic,
"Cooperative adaptive estimation of
in Proc. Asilomar Conf.
distributed noncircular complex signals,"
Signals, Syst. and Comput., Pacific Grove, CA, Nov. 2012, pp. 1518 --
1522.
[16] 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.
[17] 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.
[18] 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] 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.
[20] J. Chen and A. H. Sayed, "Distributed Pareto optimization via diffusion
adaptation," IEEE J. Sel. Topics Signal Process., vol. 7, no. 2, pp. 205 --
220, Apr. 2013.
[21] A. H. Sayed, "Diffusion adaptation over networks," in Academic Press
Library in Signal Processing, vol. 3, R. Chellapa and S. Theodoridis,
editors, pp. 323 -- 454, Elsevier, 2014.
[22] A. H. Sayed, "Adaptive networks," Proc. IEEE, vol. 102, no. 4, pp.
460 -- 497, Apr. 2014.
[23] A. H. Sayed, "Adaptation, learning, and optimization over networks,"
Foundations and Trends in Machine Learning, vol. 7, issue 4 -- 5, NOW
Publishers, Boston-Delft, Jul. 2014., pp. 311 -- 801.
[24] J. Chen, Z. J. Towfic, and A. H. Sayed,
"Dictionary learning over
distributed models," IEEE Trans. Signal Process., vol. 63, no. 4, pp.
1001 -- 1016, Feb. 2015.
[25] S. V. Macua, J. Chen, S. Zazo, and A. H. Sayed, "Distributed policy
IEEE Trans. Autom.
evaluation under multiple behavior strategies,"
Control, vol. 60, no. 5, May 2015.
[26] 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.
[27] O. N. Gharehshiran, V. Krishnamurthy, and G. Yin, "Distributed energy-
aware diffusion least mean squares: Game-theoretic learning," IEEE
Journal Sel. Topics Signal Process., vol. 7, no. 5, pp. 821 -- 836, Jun.
2013.
[28] 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.
(cid:104)
+ 2γi
1T Γi
1
2
c(cid:107) wc,0(cid:107) ·
c · (cid:107) wc,0(cid:107) · (cid:107)uL,k ⊗ IM(cid:107) ·
eWe,0
+ 2γi
max) + (cid:107)uL,k ⊗ IM(cid:107) · O(µmax)
O(µ
+ O(µmax) + O(µ2
max)
≤ 2(cid:107)uL,k ⊗ IM(cid:107)2·1T Γi
+2(cid:107) wc,0(cid:107)·(cid:107)uL,k ⊗ IM(cid:107)·
+γi
max)+O(µmax)
eWe,0
eWe,0
(cid:113)
1T Γi
1
2
(g)
c·O(µ
(cid:105)
(301)
30
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. XX, NO. XX, MONTH 2015
wireless sensor networks," IEEE Trans. Signal Process., vol. 61, no. 11,
pp. 2802 -- 2814, Jun. 2013.
[54] B. Widrow, J. M. McCool, M. G. Larimore, and C. R. Johnson Jr, "Sta-
tionary and nonstationary learning characterisitcs of the LMS adaptive
filter," Proc. IEEE, vol. 64, no. 8, pp. 1151 -- 1162, Aug. 1976.
[55] S. Jones, R. Cavin III, and W. Reed, "Analysis of error-gradient adaptive
linear estimators for a class of stationary dependent processes," IEEE
Trans. Inf. Theory, vol. 28, no. 2, pp. 318 -- 329, Mar. 1982.
[56] W. A. Gardner, "Learning characteristics of stochastic-gradient-descent
algorithms: A general study, analysis, and critique," Signal Process.,
vol. 6, no. 2, pp. 113 -- 133, Apr. 1984.
[57] A. Feuer and E. Weinstein,
"Convergence analysis of LMS filters
with uncorrelated gaussian data," IEEE Trans. Acoust., Speech, Signal
Process., vol. 33, no. 1, pp. 222 -- 230, Feb. 1985.
[58] A. J. Laub, Matrix Analysis for Scientists and Engineers, SIAM, PA,
[59] E. Kreyszig, Introductory Functional Analysis with Applications, Wiley,
2005.
NY, 1989.
[60] A. H. Sayed, Adaptive Filters, Wiley, NJ, 2008.
[29] 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.
[30] S. Lee and A. Nedic,
"Distributed random projection algorithm for
convex optimization," IEEE Journal Sel. Topics Signal Process., vol. 7,
no. 2, pp. 221 -- 229, Apr. 2013.
[31] K. I. Tsianos, S. Lawlor, and M. G. Rabbat,
"Consensus-based
distributed optimization: Practical issues and applications in large-scale
machine learning," in Proc. Annual Allerton Conference on Commun.
Control and Comput., Monticello, IL, Oct. 2012, pp. 1543 -- 1550.
[32] D.P. Palomar and M. Chiang, "A tutorial on decomposition methods for
network utility maximization," IEEE J. Sel. Areas Commun., vol. 24,
no. 8, pp. 1439 -- 1451, Aug. 2006.
[33] V. Saligrama, M. Alanyali, and O. Savas,
"Distributed detection in
sensor networks with packet losses and finite capacity links," IEEE
Trans. Signal Proc., vol. 54, no. 11, pp. 4118 -- 4132, Oct. 2006.
[34] J. B. Predd, S. R. Kulkarni, and H. V. Poor, "A collaborative training
algorithm for distributed learning," IEEE Trans. Inf. Theory, vol. 55,
no. 4, pp. 1856 -- 1871, Apr. 2009.
[35] 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.
[36] W. Ren and R. W. Beard, "Consensus seeking in multiagent systems
under dynamically changing interaction topologies," IEEE Trans. Autom.
Control, vol. 50, no. 5, pp. 655 -- 661, May 2005.
[37] 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.
[38] 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.
[39] R. Olfati-Saber, J.A. Fax, and R.M. Murray, "Consensus and cooperation
in networked multi-agent systems," Proc. IEEE, vol. 95, no. 1, pp. 215 --
233, Jan. 2007.
[40] 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.
[41] F. S. Cattivelli and A. H. Sayed, "Self-organization in bird flight forma-
tions using diffusion adaptation," in Proc. 3rd International Workshop on
Computational Advances in Multi-Sensor Adaptive Processing, Aruba,
Dutch Antilles, Dec. 2009, pp. 49 -- 52.
[42] S-Y. Tu and A. H. Sayed,
"Mobile adaptive networks with self-
organization abilities," in Proc. 7th International Symposium on Wireless
Communication Systems, York, United Kingdom, Sep. 2010, pp. 379 --
383.
[43] Y.-W. Hong and A. Scaglione, "A scalable synchronization protocol for
large scale sensor networks and its applications," IEEE J. Sel. Areas
Comm., vol. 23, no. 5, pp. 1085 -- 1099, May 2005.
[44] J. Chen and A. H. Sayed, "On the learning behavior of adaptive networks
-- Part II: Performance analysis," to appear in IEEE Trans. Inf. Theory,
2015 [also available as arXiv:1312.7580, Dec. 2013].
[45] P. Bianchi, G. Fort, and W. Hachem, "Performance of a distributed
stochastic approximation algorithm," IEEE Trans. Inf. Theory, vol. 59,
no. 11, pp. 7405 -- 7418, Nov. 2013.
[46] B. Johansson, T. Keviczky, M. Johansson, and K.H. Johansson, "Subgra-
dient methods and consensus algorithms for solving convex optimization
problems," in Proc. IEEE Conf. Decision and Control (CDC), Cancun,
Mexico, Dec. 2008, IEEE, pp. 4185 -- 4190.
[47] P. Braca, S. Marano, and V. Matta, "Running consensus in wireless
sensor networks," in Proc. 11th IEEE Int. Conf. on Information Fusion,
Cologne, Germany, June 2008, pp. 1 -- 6.
[48] S. S. Stankovic, M. S. Stankovic, and D. M. Stipanovic, "Decentralized
parameter estimation by consensus based stochastic approximation,"
IEEE Trans. Autom. Control, vol. 56, no. 3, pp. 531 -- 543, Mar. 2011.
[49] B. Polyak, Introduction to Optimization, Optimization Software, NY,
1987.
[50] A. Tahbaz-Salehi and A. Jadbabaie,
"A necessary and sufficient
condition for consensus over random networks," IEEE Trans. Autom.
Control, vol. 53, no. 3, pp. 791 -- 795, Apr. 2008.
[51] D. P. Bertsekas and J. N. Tsitsiklis, "Gradient convergence in gradient
methods with errors," SIAM J. Optim., vol. 10, no. 3, pp. 627 -- 642,
2000.
[52] R. A. Horn and C. R. Johnson, Matrix Analysis, Cambridge University
Press, 1990.
[53] F. Iutzeler, C. Philippe, and W. Hachem, "Analysis of sum-weight-like
algorithms for analysis of sum-weight-like algorithms for averaging in
|
1707.04402 | 2 | 1707 | 2018-02-27T09:36:29 | Lenient Multi-Agent Deep Reinforcement Learning | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Much of the success of single agent deep reinforcement learning (DRL) in recent years can be attributed to the use of experience replay memories (ERM), which allow Deep Q-Networks (DQNs) to be trained efficiently through sampling stored state transitions. However, care is required when using ERMs for multi-agent deep reinforcement learning (MA-DRL), as stored transitions can become outdated because agents update their policies in parallel [11]. In this work we apply leniency [23] to MA-DRL. Lenient agents map state-action pairs to decaying temperature values that control the amount of leniency applied towards negative policy updates that are sampled from the ERM. This introduces optimism in the value-function update, and has been shown to facilitate cooperation in tabular fully-cooperative multi-agent reinforcement learning problems. We evaluate our Lenient-DQN (LDQN) empirically against the related Hysteretic-DQN (HDQN) algorithm [22] as well as a modified version we call scheduled-HDQN, that uses average reward learning near terminal states. Evaluations take place in extended variations of the Coordinated Multi-Agent Object Transportation Problem (CMOTP) [8] which include fully-cooperative sub-tasks and stochastic rewards. We find that LDQN agents are more likely to converge to the optimal policy in a stochastic reward CMOTP compared to standard and scheduled-HDQN agents. | cs.MA | cs |
Lenient Multi-Agent Deep Reinforcement Learning
Gregory Palmer
University of Liverpool
United Kingdom
[email protected]
Daan Bloembergen
Centrum Wiskunde & Informatica
The Netherlands
[email protected]
Karl Tuyls
DeepMind and University of Liverpool
United Kingdom
[email protected]
Rahul Savani
University of Liverpool
United Kingdom
[email protected]
ABSTRACT
Much of the success of single agent deep reinforcement learning
(DRL) in recent years can be attributed to the use of experience
replay memories (ERM), which allow Deep Q-Networks (DQNs)
to be trained efficiently through sampling stored state transitions.
However, care is required when using ERMs for multi-agent deep
reinforcement learning (MA-DRL), as stored transitions can become
outdated because agents update their policies in parallel [11]. In
this work we apply leniency [23] to MA-DRL. Lenient agents map
state-action pairs to decaying temperature values that control the
amount of leniency applied towards negative policy updates that
are sampled from the ERM. This introduces optimism in the value-
function update, and has been shown to facilitate cooperation in
tabular fully-cooperative multi-agent reinforcement learning prob-
lems. We evaluate our Lenient-DQN (LDQN) empirically against the
related Hysteretic-DQN (HDQN) algorithm [22] as well as a mod-
ified version we call scheduled-HDQN, that uses average reward
learning near terminal states. Evaluations take place in extended
variations of the Coordinated Multi-Agent Object Transportation
Problem (CMOTP) [8] which include fully-cooperative sub-tasks
and stochastic rewards. We find that LDQN agents are more likely
to converge to the optimal policy in a stochastic reward CMOTP
compared to standard and scheduled-HDQN agents.
KEYWORDS
Distributed problem solving; Multiagent learning; Deep learning
1 INTRODUCTION
The field of deep reinforcement learning has seen a great number
of successes in recent years. Deep reinforcement learning agents
have been shown to master numerous complex problem domains,
ranging from computer games [17, 21, 26, 33] to robotics tasks [10,
12]. Much of this success can be attributed to using convolutional
neural network (ConvNet) architectures as function approximators,
allowing reinforcement learning agents to be applied to domains
with large or continuous state and action spaces. ConvNets are often
trained to approximate policy and value functions through sampling
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.
https://doi.org/doi
past state transitions stored by the agent inside an experience replay
memory (ERM).
Recently the sub-field of multi-agent deep reinforcement learn-
ing (MA-DRL) has received an increased amount of attention. Multi-
agent reinforcement learning is known for being challenging even
in environments with only two implicit learning agents, lacking
the convergence guarantees present in most single-agent learning
algorithms [5, 20]. One of the key challenges faced within multi-
agent reinforcement learning is the moving target problem: Given
an environment with multiple agents whose rewards depend on
each others' actions, the difficulty of finding optimal policies for
each agent is increased due to the policies of the agents being non
stationary [7, 14, 31]. The use of an ERM amplifies this problem,
as a large proportion of the state transitions stored can become
deprecated [22].
Due to the moving target problem reinforcement learning algo-
rithms that converge in a single agent setting often fail in fully-
cooperative multi-agent systems with independent learning agents
that require implicit coordination strategies. Two well researched
approaches used to help parallel reinforcement learning agents
overcome the moving target problem in these domains include hys-
teretic Q-learning [19] and leniency [23]. Recently Omidshafiei et
al. [22] successfully applied concepts from hysteretic Q-learning to
MA-DRL (HDQN). However, we find that standard HDQNs struggle
in fully cooperative domains that yield stochastic rewards. In the
past lenient learners have been shown to outperform hysteretic
agents for fully cooperative stochastic games within a tabular set-
ting [35]. This raises the question whether leniency can be applied
to domains with a high-dimensional state space.
In this work we show how lenient learning can be extended to
MA-DRL. Lenient learners store temperature values that are associ-
ated with state-action pairs. Each time a state-action pair is visited
the respective temperature value is decayed, thereby decreasing the
amount of leniency that the agent applies when performing a policy
update for the state-action pair. The stored temperatures enable
the agents to gradually transition from optimists to average reward
learners for frequently encountered state-action pairs, allowing the
agents to outperform optimistic and maximum based learners in
environments with misleading stochastic rewards [35]. We extend
this idea to MA-DRL by storing leniency values in the ERM, and
demonstrate empirically that lenient MA-DRL agents that learn im-
plicit coordination strategies in parallel are able to converge on the
optimal joint policy in difficult coordination tasks with stochastic
AAMAS'18, July 2018, Stockholm, Sweden
Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani
rewards. We also demonstrate that the performance of Hysteretic
Q-Networks (HDQNs) within stochastic reward environments can
be improved with a scheduled approach.
Our main contributions can be summarized as follows.
1) We introduce the Lenient Deep Q-Network (LDQN) algorithm
which includes two extensions to leniency: a retroactive temper-
ature decay schedule (TDS) that prevents premature temperature
cooling, and a T(s)-Greedy exploration strategy, where the proba-
bility of the optimal action being selected is based on the average
temperature of the current state. When combined, TDS and T(s)-
Greedy exploration encourage exploration until average rewards
have been established for later transitions.
2) We show the benefits of using TDS over average temperature
folding (ATF) [35].
3) We provide an extensive analysis of the leniency-related hyper-
parameters of the LDQN.
4) We propose a scheduled-HDQN that applies less optimism to-
wards state transitions near terminal states compared to earlier
transitions within the episode.
5) We introduce two extensions to the Cooperative Multi-agent
Object Transportation Problem (CMOTP) [8], including narrow
passages that test the agents' ability to master fully-cooperative
sub-tasks, stochastic rewards and noisy observations.
6) We empirically evaluate our proposed LDQN and SHDQN against
standard HDQNs using the extended versions of the CMOTP. We
find that while HDQNs perform well in deterministic CMOTPs, they
are significantly outperformed by SHDQNs in domains that yield a
stochastic reward. Meanwhile LDQNs comprehensively outperform
both approaches within the stochastic reward CMOTP.
The paper proceeds as follows: first we discuss related work
and the motivation for introducing leniency to MA-DRL. We then
provide the reader with the necessary background regarding how
leniency is used within a tabular setting, briefly introduce hysteretic
Q-learning and discuss approaches for clustering states based on
raw pixel values. We subsequently introduce our contributions,
including the Lenient-DQN architecture, T (s)-Greedy exploration,
Temperature Decay Schedules, Scheduled-HDQN and extensions to
the CMOTP, before moving on to discuss the results of empirically
evaluating the LDQN. Finally we summarize our findings, and
discuss future directions for our research.
2 RELATED WORK
A number of methods have been proposed to help deep reinforce-
ment learning agents converge towards an optimal joint policy in
cooperative multi-agent tasks. Gupta et al. [13] evaluated policy
gradient, temporal difference error, and actor critic methods on
cooperative control tasks that included discrete and continuous
state and action spaces, using a decentralized parameter sharing
approach with centralized learning. In contrast our current work
focuses on decentralized-concurrent learning. A recent successful
approach has been to decompose a team value function into agent-
wise value functions through the use of a value decomposition
network architecture [27]. Others have attempted to help concur-
rent learners converge through identifying and deleting obsolete
state transitions stored in the replay memory. For instance, Foerster
et al. [11] used importance sampling as a means to identify outdated
transitions while maintaining an action observation history of the
other agents. Our current work does not require the agents to main-
tain an action observation history. Instead we focus on optimistic
agents within environments that require implicit coordination. This
decentralized approach to multi-agent systems offers a number of
advantages including speed, scalability and robustness [20]. The
motivation for using implicit coordination is that communication
can be expensive in practical applications, and requires efficient
protocols [2, 20, 29].
Hysteretic Q-learning is a form of optimistic learning with a
strong empirical track record in fully-observable multi-agent rein-
forcement learning [3, 20, 37]. Originally introduced to prevent the
overestimation of Q-Values in stochastic games, hysteretic learners
use two learning rates: a learning rate α for updates that increase
the value estimate (Q-value) for a state-action pair and a smaller
learning rate β for updates that decrease the Q-value [19]. However,
while experiments have shown that hysteretic learners perform well
in deterministic environments, they tend to perform sub-optimally
in games with stochastic rewards. Hysteretic learners' struggles
in these domains have been attributed to learning rate β's inter-
dependencies with the other agents' exploration strategies [20].
Lenient learners present an alternative to the hysteretic approach,
and have empirically been shown to converge towards superior
policies in stochastic games with a small state space [35]. Similar to
the hysteretic approach, lenient agents initially adopt an optimistic
disposition, before gradually transforming into average reward
learners [35]. Lenient methods have received criticism in the past
for the time they require to converge [35], the difficulty involved
in selecting the correct hyperparameters, the additional overhead
required for storing the temperature values, and the fact that they
were originally only proposed for matrix games [20]. However,
given their success in tabular settings we here investigate whether
leniency can be applied successfully to MA-DRL.
3 BACKGROUND
Q-Learning. The algorithms implemented for this study are
based upon Q-learning, a form of temporal difference reinforcement
learning that is well suited for solving sequential decision making
problems that yield stochastic and delayed rewards [4, 34]. The
algorithm learns Q-values for state-action pairs which are estimates
of the discounted sum of future rewards (the return) that can be
obtained at time t through selecting action at in a state st , providing
the optimal policy is selected in each state that follows.
Since most interesting sequential decision problems have a large
state-action space, Q-values are often approximated using function
approximators such as tile coding [4] or neural networks [33]. The
parameters θ of the function approximator can be learned from
experience gathered by the agent while exploring their environ-
ment, choosing an action at in state st according to a policy π, and
updating the Q-function by bootstrapping the immediate reward
rt +1 received in state st +1 plus the expected future reward from the
next state (as given by the Q-function):
t − Q(st , at ; θt)(cid:1)∇θt Q(st , at ; θt).
θt +1 = θt + α(cid:0)Y
Q
(1)
Q
Here, Y
is the bootstrap target which sums the immediate reward
t
rt +1 and the current estimate of the return obtainable from the
Lenient Multi-Agent Deep Reinforcement Learning
AAMAS'18, July 2018, Stockholm, Sweden
next state st +1 assuming optimal behaviour (hence the max oper-
ator) and discounted by γ ∈ (0, 1], given in Eq. (2). The Q-value
Q(st , at ; θt) moves towards this target by following the gradient
∇θt Q(st , at ; θt); α ∈ (0, 1] is a scalar used to control the learning
rate.
(2)
Q(st +1, a; θt).
t ≡ rt +1 + γ max
a∈A
Q
Y
Deep Q-Networks (DQN). In deep reinforcement learning [21]
a multi-layer neural network is used as a function approxima-
tor, mapping a set of n-dimensional state variables to a set of m-
: Rn → Rm, where m represents the
dimensional Q-values f
number of actions available to the agent. The network parame-
ters θ can be trained using stochastic gradient descent, randomly
sampling past transitions experienced by the agent that are stored
within an experience replay memory (ERM) [18, 21]. Transitions are
tuples (st , at , st +1, rt +1) consisting of the original state st , the ac-
tion at , the resulting state st +1 and the immediate reward rt +1. The
network is trained to minimize the time dependent loss function
Li(θi),
Li(θi) = Es,a∼p(·)
(3)
where p(s, a) represents a probability distribution of the transitions
stored within the ERM, and Yt is the target:
(cid:104)(Yt − Q(s, a; θt))2(cid:105)
,
Yt ≡ rt +1 + γQ(st +1, argmax
a∈A
Q(st +1, a; θt); θ
′
t).
(4)
Equation (4) is a form of double Q-learning [32] in which the
target action is selected using weights θ, while the target value
is computed using weights θ′ from a target network. The target
network is a more stable version of the current network, with
the weights being copied from current to target network after
every n transitions [33]. Double-DQNs have been shown to reduce
overoptimistic value estimates [33]. This notion is interesting for
our current work, since both leniency and hysteretic Q-learning
attempt to induce sufficient optimism in the early learning phases
to allow the learning agents to converge towards an optimal joint
policy.
Hysteretic Q-Learning. Hysteretic Q-learning [19] is an algo-
rithm designed for decentralised learning in deterministic multi-
agent environments, and which has recently been applied to MA-
DRL as well [22]. Two learning rates are used, α and β, with β < α.
The smaller learning rate β is used whenever an update would
reduce a Q-value. This results in an optimistic update function
which puts more weight on positive experiences, which is shown
to be beneficial in cooperative multi-agent settings. Given a spec-
trum with traditional Q-learning at one end and maximum-based
learning, where negative experiences are completely ignored, at
the other, then hysteretic Q-learning lies somewhere in between
depending on the value chosen for β.
Leniency. Lenient learning was originally introduced by Potter
and De Jong [25] to help cooperative co-evolutionary algorithms
converge towards an optimal policy, and has later been applied
to multi-agent learning as well [24]. It was designed to prevent
relative overgeneralization [36], which occurs when agents gravitate
towards a robust but sub-optimal joint policy due to noise induced
by the mutual influence of each agent's exploration strategy on
others' learning updates.
Leniency has been shown to increase the likelihood of conver-
gence towards the globally optimal solution in stateless coordina-
tion games for reinforcement learning agents [6, 23, 24]. Lenient
learners do so by effectively forgiving (ignoring) sub-optimal ac-
tions by teammates that lead to low rewards during the initial
exploration phase [23, 24]. While initially adopting an optimistic
disposition, the amount of leniency displayed is typically decayed
each time a state-action pair is visited. As a result the agents be-
come less lenient over time for frequently visited state-action pairs
while remaining optimistic within unexplored areas. This transition
to average reward learners helps lenient agents avoid sub-optimal
joint policies in environments that yield stochastic rewards [35].
During training the frequency with which lenient reinforcement
learning agents perform updates that result in lowering the Q-
value of a state action pair (s, a) is determined by leniency and
temperature functions, l(st , at) and Tt(st , at) respectively [35]. The
relation of the temperature function is one to one, with each state-
action pair being assigned a temperature value that is initially set
to a defined maximum temperature value, before being decayed
each time the pair is visited. The leniency function
−K∗Tt(st ,at)
l(st , at) = 1 − e
(5)
uses a constant K as a leniency moderation factor to determine how
the temperature value affects the drop-off in lenience. Following
the update, Tt(st , at) is decayed using a discount factor β ∈ [0, 1]
such that Tt +1(st , at) = β Tt(st , at).
Given a TD-Error δ, where δ = Yt − Q(st , at ; θt), leniency is
applied to a Q-value update as follows:
Q(st , at) =
Q(st , at) + αδ
Q(st , at)
(6)
The random variable x ∼ U(0, 1) is used to ensure that an update
on a negative δ is executed with a probability 1 − l(st , at).
if δ > 0 or x > l(st , at).
if δ ≤ 0 and x ≤ l(st , at).
(cid:40)
Temperature-based exploration. The temperature values main-
tained by lenient learners can also be used to influence the action
selection policy. Recently Wei and Luke [35] introduced Lenient
Multiagent Reinforcement Learning 2 (LMRL2), where the average
temperature of the agent's current state is used with the Boltz-
mann action selection strategy to determine the weight of each
action. As a result agents are more likely to choose a greedy action
within frequently visited states while remaining exploratory for
less-frequented areas of the environment. However, Wei and Luke
[35] note that the choice of temperature moderation factor for the
Boltzmann selection method is a non-trivial task, as Boltzmann
selection is known to struggle to distinguish between Q-Values that
are close together [15, 35].
Average Temperature Folding (ATF). If the agents find them-
selves in the same initial state at the beginning of each episode, then
after repeated interactions the temperature values for state-action
pairs close to the initial state can decay rapidly as they are visited
more frequently. However, it is crucial for the success of the lenient
learners that the temperatures for these state-action pairs remains
sufficiently high for the rewards to propagate back from later stages,
and to prevent the agents from converging upon a sub-optimal
policy [35]. One solution to this problem is to fold the average
temperature for the n actions available to the agent in st +1 into the
AAMAS'18, July 2018, Stockholm, Sweden
which this average temperature T t(st +1) = 1/nn
temperature that is being decayed for (st , at) [35]. The extent to
i =1 Tt(st +1, ai) is
folded in is determined by a constant υ as follows:
Tt +1(st , at) = β
Tt(st , at)
(1 − υ)Tt(st , at) + υT t(st +1)
if st +1 is terminal.
otherwise.
(7)
(cid:40)
Clustering states using autoencoders. In environments with
a high dimensional or continuous state space, a tabular approach
for mapping each possible state-action pair to a temperature as dis-
cussed above is no longer feasible. Binning can be used to discretize
low dimensional continuous state-spaces, however further consid-
erations are required regarding mapping semantically similar states
to a decaying temperature value when dealing with high dimen-
sional domains, such as image observations. Recently, researchers
studying the application of count based exploration to Deep RL
have developed interesting solutions to this problem. For example,
Tang et al. [30] used autoencoders to automatically cluster states
in a meaningful way in challenging benchmark domains including
Montezuma's Revenge.
The autoencoder, consisting of convolutional, dense, and trans-
posed convolutional layers, can be trained using the states stored in
the agent's replay memory [30]. It then serves as a pre-processing
function д : S → RD, with a dense layer consisting of D neurons
with a saturating activation function (e.g. a Sigmoid function) at
the centre. SimHash [9], a locality-sensitive hashing (LSH) function,
can be applied to the rounded output of the dense layer to gener-
ate a hash-key ϕ for a state s. This hash-key is computed using a
constant k × D matrix A with i.i.d. entries drawn from a standard
Gaussian distribution N(0, 1) as
ϕ(s) = sдn(cid:0)A д(s)(cid:1) ∈ {−1, 1}k .
(8)
where д(s) is the autoencoder pre-processing function, and k con-
trols the granularity such that higher values yield a more fine-
grained clustering [30].
4 ALGORITHMIC CONTRIBUTIONS
In the following we describe our main algorithmic contributions.
First we detail our newly proposed Lenient Deep Q-Network, and
thereafter we discuss our extension to Hysteretic DQN, which we
call Scheduled HDQN.
4.1 Lenient Deep Q-Network (LDQN)
Approach. Combining leniency with DQNs requires careful
considerations regarding the use of the temperature values, in par-
ticular when to compute the amount of leniency that should be
applied to a state transition that is sampled from the replay memory.
In our initial trials we used leniency as a mechanism to determine
which transitions should be allowed to enter the ERM. However,
this approach led to poor results, presumably due to the agents de-
veloping a bias during the initial random exploration phase where
transitions were stored indiscriminately. To prevent this bias we use
an alternative approach where we compute and store the amount of
leniency at time t within the ERM tuple: (st−1, at−1, rt , st , l(st , at)t).
The amount of leniency that is stored is determined by the current
temperature value T associated with the hash-key ϕ(s) for state s
Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani
and the selected action a, similar to Eq. (5):
−k×T(ϕ(s),a)
l(s, t) = 1 − e
.
(9)
We use a dictionary to map each (ϕ(s), a) pair encountered to a
temperature value, where the hash-keys are computed using Tang
et al.'s [30] approach described in Section 3. If a temperature value
does not yet exist for (ϕ(s), a) within the dictionary then an entry is
created, setting the temperature value equal to MaxT emperature.
Otherwise the current temperature value is used and subsequently
decayed, to ensure the agent will be less lenient when encountering
a semantically similar state in the future. As in standard DQN the
aim is to minimize the loss function of Eq. (3), with the modification
that for each sample j chosen from the replay memory for which
the leniency conditions of Eq. (6) are not met, are ignored.
Retroactive Temperature Decay Schedule (TDS). During ini-
tial trials we found that temperatures decay rapidly for state-action
pairs belonging to challenging sub-tasks in the environment, even
when using ATF (Section 3). In order to prevent this premature
cooling of temperatures we developed an alternative approach us-
ing a pre-computed temperature decay schedule β0, . . . , βn with a
step limit n. The values for β are computed using an exponent ρ
which is decayed using a decay rate d:
βn = eρ×d t
(10)
for each t, 0 ≤ t < n.
Upon reaching a terminal state the temperature decay schedule
is applied as outlined in Algorithm 1. The aim is to ensure that
temperature values of state-action pairs encountered during the
early phase of an episode are decayed at a slower rate than those
close to the terminal state transition (line 4). We find that main-
taining a slow-decaying maximum temperature ν (lines 5-7) that is
decayed using a decay rate µ helps stabilize the learning process
when ϵ-Greedy exploration is used. Without the decaying maxi-
mum temperature the disparity between the low temperatures in
well explored areas and the high temperatures in relatively unex-
plored areas has a destabilizing effect during the later stages of the
learning process. Furthermore, for agents also using the tempera-
ture values to guide their exploration strategy (see below), ν can
help ensure that the agents transition from exploring to exploiting
within reasonable time. The decaying maximum temperature ν is
used whenever T(ϕ(st−1), at−1) > νt , or when agents fail at their
task in environments where a clear distinction can be made be-
tween success and failure. Therefore TDS is best suited for domains
that yield sparse large rewards.
Applying the TDS after the agents fail at a task could result in the
repeated decay of temperature values for state-action pairs leading
up to a sub-task. For instance, the sub-task of transporting a heavy
item of goods through a doorway may only require a couple of
steps for trained agents who have learned to coordinate. However,
untrained agents may require thousands of steps to complete the
task. If a time-limit is imposed for the agents to deliver the goods,
and the episode ends prematurely while an attempt is made to solve
the sub-task, then the application of the TDS will result in the rapid
decay of the temperature values associated with the frequently en-
countered state-action pairs. We resolve this problem by setting the
temperature values Tt(ϕ(si), ai) > ν to ν at the end of incomplete
Lenient Multi-Agent Deep Reinforcement Learning
AAMAS'18, July 2018, Stockholm, Sweden
Algorithm 1 Application of temperature decay schedule (TDS)
if βnTt(ϕ(si), ai) < νt then
1: Upon reaching a terminal state do
2: n ← 0, steps ← steps taken during the episode
3: for i = steps to 0 do
4:
5:
6:
7:
8:
9:
10: end for
11: ν ← µ ν
Tt +1(ϕ(si), ai) ← βnTt(ϕ(si), ai)
Tt +1(ϕ(si), ai) ← νt
else
end if
n ← n + 1
Figure 1: Lenient-DQN Architecture. We build on the stan-
dard Double-DQN architecture [33] by adding a lenient loss
function (top right, see Section 4.1). Leniency values are
stored in the replay memory along with the state transitions;
we cluster semantically similar states using an autoencoder
and SimHash (bottom left), and apply our retroactive tem-
perature decay schedule (TDS, Algorithm 1). Actions are se-
lected using the T(st)-Greedy exploration method.
runs instead of repeatedly decaying them, thereby ensuring that
the agents maintain a lenient disposition towards one another.
T(st )-Greedy Exploration. During initial trials we encoun-
tered the same problems discussed by Wei and Luke [35] regarding
the selection of the temperature moderation factor for the Boltz-
mann action selection strategy. This led to the development of a
more intuitive T(st)-Greedy exploration method where the aver-
age temperature value T(st) ∈ (0, 1] for a state st replaces the ϵ
in the ϵ-Greedy action selection method. An exponent ξ is used
to control the pace at which the agents transition from explorers
to exploiters. The agent therefore selects action a = arдmaxaQ(a)
with a probability 1 − T(st)ξ and a random action with probability
T(st)ξ . We outline our complete LDQN architecture in Figure 1.
4.2 Scheduled-HDQN (SHDQ)
Hysteretic Q-learners are known to converge towards sub-optimal
joint policies in environments that yield stochastic rewards [35].
However, drawing parallels to lenient learning, where it is desirable
to decay state-action pairs encountered at the beginning of an
episode at a slower rate compared to those close to a terminal state,
we consider that the same principle can be applied to Hysteretic
Q-learning. Subsequently we implemented Scheduled-HDQN with
a pre-computed learning rate schedule β0, . . . , βn where βn is set
to a value approaching α, and for each βt , 0 ≤ t < n, we have βt =
dn−t βn using a decay coefficient d ∈ (0, 1]. The state transitions
encountered throughout each episode are initially stored within a
queue data-structure. Upon reaching a terminal state the n state-
transitions are transferred to the ERM as (st , st +1, rt +1, at , βt ) for
t ∈ {0, . . . , n}. Our hypothesis is that storing β values that approach
α for state-transitions leading to the terminal state will help agents
converge towards optimal joint policies in environments that yield
sparse stochastic rewards.
5 EMPIRICAL EVALUATION
CMOTP Extensions. We subjected our agents to a range of Co-
ordinated Multi-Agent Object Transportation Problems (CMOTPs)
inspired by the scenario discussed in Buşoniu et al. [8], in which
two agents are tasked with delivering one item of goods to a drop-
zone within a grid-world. The agents must first exit a room one
by one before locating and picking up the goods by standing in
the grid cells on the left and right hand side. The task is fully co-
operative, meaning the goods can only be transported upon both
agents grasping the item and choosing to move in the same direc-
tion. Both agents receive a positive reward after placing the goods
inside the drop-zone. The actions available to each agent are to
either stay in place or move left, right, up or down. We subjected
our agents to three variations of the CMOTP, depicted in Figure
2, where each A represents one of the agents, G the goods, and
D-ZONE / DZ mark the drop-zone(s). The layout in sub-figure 2a
is a larger version of the original layout [8], while the layout in
sub-figure 2b introduces narrow-passages between the goods and
the drop-zone, testing whether the agents can learn to coordinate
in order to overcome challenging areas within the environment.
The layout in sub-figure 2c tests the agents' response to stochastic
rewards. Drop-zone 1 (DZ1) yields a reward of 0.8, whereas drop-
zone 2 (DZ2) returns a reward of 1 on 60% of occasions and only 0.4
on the other 40%. DZ1 therefore returns a higher reward on average,
0.8 compared to the 0.76 returned by DZ2. A slippery surface can
be added to introduce stochastic state transitions to the CMOTP,
a common practice within grid-world domains where the agents
move in an unintended direction with a predefined probability at
each time-step.
Setup. We conduct evaluations using a Double-DQN architec-
ture [33] as basis for the algorithms. The Q-network consists of 2
convolutional layers with 32 and 64 kernels respectively, a fully
connected layer with 1024 neurons and an output neuron for each
action. The agents are fed a 16 × 16 tensor representing a gray-
scale version of the grid-world as input. We use the following pixel
values to represent the entities in our grid-world: Aдent1 = 250,
Aдent2 = 200, Goods = 150 and Obstacles = 50. Adam [16] is used
AAMAS'18, July 2018, Stockholm, Sweden
Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani
(a) Original
(b) Narrow-Passage
(c) Stochastic
Figure 2: CMOTP Layouts
to optimize the networks. Our initial experiments are conducted
within a noise free environment, enabling us to speed up the testing
of our LDQN algorithm without having to use an autoencoder for
hashing; instead we apply python's xxhash. We subsequently test
the LDQN with the autoencoder for hashing in a noisy version
of the stochastic reward CMOTP. The autoencoder consists of 2
convolutional Layers with 32 and 64 kernels respectively, 3 fully
connected layers with 1024, 512, and 1024 neurons followed by 2
transposed convolutional layers. For our Scheduled-HDQN agents
we pre-compute β0 to n by setting βn = 0.9 and applying a decay
coefficient of d = 0.99 at each step t = 1 to n, i.e. βn−t = 0.99t βn,
with βn−t being bounded below at 0.4. We summarize the remaining
hyper-parameters in Table 1. In Section 7 we include an extensive
analysis of tuning the leniency related hyper-parameters. We note
at this point that each algorithm used the same learning rate α
specified in Table 1.
Component
Hyper-parameter
Steps between target network synchronization
Setting
0.0001
0.95
5000
250'000
1.0
0.999
0.05
1.0
2.0
-0.01
0.95
1.0
0.999
64
512
DQN-Optimization
ϵ-Greedy Exploration
Leniency
Learning rate α
Discount rate γ
ERM Size
Initial ϵ value
ϵ Decay factor
Minimum ϵ Value
MaxTemperature
TDS Exponent ρ
Temperature Modification Coefficient K
TDS Exponent Decay Rate d
Initial Max Temperature Value ν
Max Temperature Decay Coefficient µ
Autoencoder
HashKey Dimensions k
Number of sigmoidal units in the dense layer D
Table 1: Hyper-parameters
identical actions per state transition. As a result thousands of state
transitions are often required to deliver the goods and receive a re-
ward while the agents explore the environment, preventing the use
of a small replay memory where outdated transitions would be over-
written within reasonable time. As a result standard Double-DQN
architectures struggled to master the CMOTP, failing to coordinate
on a significant number of runs even when confronted with the
relatively simple original CMOTP.
We conducted 30 training runs of 5000 episodes per run for each
LDQN and HDQN configuration. Lenient and hysteretic agents
with β < 0.8 fared significantly better than the standard Double-
DQN, converging towards joint policies that were only a few steps
shy of the optimal 33 steps required to solve the task. Lenient
agents implemented with both ATF and TDS delivered a compara-
ble performance to the hysteretic agents with regards to the average
steps per episode and the coordinated steps percentage measured
over the final 100 steps of each episode (Table 2, left). However,
both LDQN-ATF and LDQN-TDS averaged a statistically significant
higher number of steps per training run compared to hysteretic
agents with β < 0.7. For the hysteretic agents we observe a sta-
tistically significant increase in the average steps per run as the
values for β increase, while the average steps and coordinated steps
percentage over the final 100 episodes remain comparable.
Narrow Passage CMOTP. Lenient agents implemented with
ATF struggle significantly within the narrow-passage CMOTP, as
evident from the results listed in Table 2 (right). We find that the av-
erage temperature values cool off rapidly over the first 100 episodes
within the Pickup and Middle compartments, as illustrated in Fig-
ure 3. Meanwhile agents using TDS manage to maintain sufficient
leniency over the first 1000 episodes to allow rewards to propagate
backwards from the terminal state. We conducted ATF experiments
with a range of values for the fold-in constant υ (0.2, 0.4 and 0.8),
but always witnessed the same outcome. Slowing down the temper-
ature decay would help agents using ATF remain lenient for longer,
with the side-effects of an overoptimistic disposition in stochastic
environments, and an increase in the number of steps required for
convergence if the temperatures are tied to the action selection
policy. Using TDS meanwhile allows agents to maintain sufficient
leniency around difficult sub-tasks within the environment while
being able to decay later transitions at a faster rate. As a result
agents using TDS can learn the average rewards for state transi-
tions close to the terminal state while remaining optimistic for
updates to earlier transitions.
6 DETERMINISTIC CMOTP RESULTS
Original CMOTP. The CMOTP represents a challenging fully
cooperative task for parallel learners. Past research has shown that
deep reinforcement learning agents can converge towards coop-
erative policies in domains where the agents receive feedback for
their individual actions, such as when learning to play pong with
the goal of keeping the ball in play for as long as possible [28].
However, in the CMOTP feedback is only received upon delivering
the goods after a long series of coordinated actions. No immediate
feedback is available upon miscoordination. When using uniform
action selection the agents only have a 20% chance of choosing
Figure 3: Average temperature per compartment
Lenient Multi-Agent Deep Reinforcement Learning
AAMAS'18, July 2018, Stockholm, Sweden
Hyst. β = 0.5
Hyst. β = 0.6
SPE
CSP
SPR
36.4
92%
1'085'982
36.1
92%
1'148'652
36.8
92%
1'408'690
Original CMOTP Results
Hyst. β = 0.7
Hyst. β = 0.8
LDQN ATF
LDQN TDS
Hyst. β = 0.5
528.9
91%
3'495'657
36.9
92%
36.8
92%
1'409'720
1'364'029
45.25
92%
1'594'968
Narrow-Passage CMOTP Results
LDQN ATF
Hyst. β = 0.6
704.9
89%
4'736'936
376.2
90%
3'950'670
LDQN TDS
45.7
92%
2'104'637
Table 2: Deterministic CMTOP Results, including average steps per episode (SPE) over the final 100 episodes, coordinated
steps percentages (CSP) over the final 100 episodes, and the average steps per training run (SPR).
The success of HDQN agents within the narrow-passage CMOTP
depends on the value chosen for β. Agents with β > 0.5 struggle to
coordinate, as we observed over a large range of β values, exem-
plars of which are given in Table 2. The only agents that converge
upon a near optimal joint-policy are those using LDQN-TDS and
HDQN (β = 0.5). We performed a Kolmogorov-Smirnov test with
a null hypothesis that there is no significant difference between
the performance metrics for agents using LDQN-TDS and HDQN
(β = 0.5). We fail to reject the null hypothesis for average steps
per episode and percentage of coordinated steps for the final 100
episodes. However, HDQN (β = 0.5) averaged significantly less
steps per run while maintaining less overhead, replicating previous
observations regarding the strengths of hysteretic agents within
deterministic environments.
7 STOCHASTIC CMOTP RESULTS
In the stochastic setting we are interested in the percentage of runs
for each algorithm that converge upon the optimal joint policy,
which is for the agents to deliver the goods to dropzone 1, yielding
a reward of 0.8, as opposed to dropzone 2 which only returns an
average reward of 0.76 (see Section 5). We conducted 40 runs of
5000 episodes for each algorithm.
As discussed in Section 6, HDQN agents using β > 0.7 frequently
fail to coordinate in the deterministic CMOTP. Therefore, setting
β = 0.7 is the most likely candidate to succeed at solving the sto-
chastic reward CMOTP for standard HDQN architectures. However,
agents using HDQN (β = 0.7) only converged towards the optimal
policy on 42.5% of runs. The scheduled-HDQN performed signifi-
cantly better achieving a 77.5% optimal policy rate. Furthermore
the SHDQN performs well when an additional funnel-like narrow-
passage is inserted close to the dropzones, with 93% success rate.
The drop in performance upon removing the funnel suggests that
the agents are led astray by the optimism applied to earlier tran-
sitions within each episode, presumably around the pickup area
where a crucial decision is made regarding the direction in which
the goods should be transported.
LDQN using ϵ−Greedy exploration performed similar to SHDQN,
converging towards the optimal joint policy on 75% of runs. Mean-
while LDQNs using T(st)-Greedy exploration achieved the highest
percentages of optimal joint-policies, with agents converging on
100% of runs for the following configuration: K = 3.0, d = 0.9,
ξ = 0.25 and µ = 0.9995, which will be discussed in more detail
below. However the percentage of successful runs is related to the
choice of hyperparameters. We therefore include an analysis of
three critical hyperparameters:
• The temperature Modification Coefficient K, that determines
the speed at which agents transition from optimist to average
reward learner (sub-figure 4a). Values: 1, 2 and 3
• The TDS decay-rate d which controls the rate at which tem-
peratures are decayed n-steps prior to the terminal state
(sub-figure 4b). Values: 0.9, 0.95 and 0.99
• T(st)-Greedy exploration exponent ξ, controlling the agent's
transition from explorer to exploiter, with lower values for
ξ encouraging exploration. Values: 0.25, 0.5 and 1.0
(a) Leniency Schedules
(b) TDS
Figure 4: TMC and TDS schedules used during analysis.
We conducted 40 simulation runs for each combination of the
three variables. To determine how well agents using LDQN can cope
with stochastic state transitions we added a slippery surface where
each action results in a random transition with 10% probability
1. The highest performing agents used a steep temperature decay
schedule that maintains high temperatures for early transitions
(d = 0.9 or d = 0.95) with temperature modification coefficients
that slow down the transition from optimist to average reward
learner (K = 2 or k = 3), and exploration exponents that delay
the transition from explorer to exploiter (ξ = 0.25 or ξ = 0.5).
This is illustrated in the heat-maps in Figure 5. When using a TDS
with a more gradual incline (d = 0.99) the temperature values from
earlier state transitions decay at a similar rate to those near terminal
states. In this setting choosing larger values for K increases the
likelihood of the agents converging upon a sub-optimal policy prior
to having established the average rewards available in later states,
as evident from the results plotted in sub-figure 5c. Even when
setting the exploration exponent ξ to 0.25 the agents prematurely
transition to exploiter while holding an overoptimistic disposition
towards follow-on states. Interestingly when K < 3 agents often
converge towards the optimal joint-policy despite setting d = 0.99.
However, the highest percentages of optimal runs (97.5%) were
achieved through combining a steep TDS (d = 0.9 or d = 0.95) with
the slow transition to average reward learner (k = 3) and exploiter
(ξ = 0.25). Meanwhile the lowest percentages for all TDSs resulted
from insufficient leniency (K = 1) and exploration (ξ = 1.0).
Using one of the best-performing configuration (K = 3.0, d = 0.9
and ξ = 0.25) we conducted further trials analyzing the agents'
1Comparable results were obtained during preliminary trials without a slippery surface.
AAMAS'18, July 2018, Stockholm, Sweden
Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani
(a) d = 0.9
(b) d = 0.95
(c) d = 0.99
Figure 5: Analysis of the LDQN hyperparameters. The
heat-maps show the percentage of runs that converged to
the optimal joint-policy (darker is better).
sensitivity to the maximum temperature decay coefficient µ. We
conducted an additional set of 40 runs where µ was increased from
0.999 to 0.9995. Combining T(S)-Greedy with the slow decaying
µ = 0.9995 results in the agents spending more time exploring the
environment at the cost of requiring longer to converge, resulting
in an additional 1'674'106 steps on average per run. However, the
agents delivered the best performance, converging towards the
optimal policy on 100% runs conducted.
Continuous State Space Analysis. Finally we show that se-
mantically similar state-action pairs can be mapped to temperature
values using SimHash in conjunction with an autoencoder. We con-
ducted experiments in a noisy version of the stochastic CMTOP,
where at each time step every pixel value is multiplied by a unique
coefficient drawn from a Gaussian distribution X ∼ N(1.0, 0.01).
A non-sparse tensor is used to represent the environment, with
background cells set to 1.0 prior to noise being applied.
Agents using LDQNs with xxhash converged towards the sub-
optimal joint policy after the addition of noise as illustrated in
Figure 6, with the temperature values decaying uniformly in tune
with ν. LDQN-TDS agents using an autoencoder meanwhile con-
verged towards the optimal policy on 97.5% of runs. It is worth
pointing out that the autoencoder introduces a new set of hyper-
parameters that require consideration, including the size D of the
dense layer at the centre of the autoencoder and the dimensions
K of the hash-key, raising questions regarding the influence of the
granularity on the convergence. We leave this for future work.
Figure 6: Noisy Stochastic CMOTP Average Reward
8 DISCUSSION & CONCLUSION
Our work demonstrates that leniency can help MA-DRL agents
solve a challenging fully cooperative CMOTP using high-dimensional
and noisy images as observations. Having successfully merged
leniency with a Double-DQN architecture raises the question re-
garding how well our LDQN will work with other state of the art
components. We have recently conducted preliminary stochastic
reward CMOTP trials with agents using LDQN with a Prioritized
Experience Replay Memory [1]. Interestingly the agents consistently
converged towards the sub-optimal joint policy. We plan to investi-
gate this further in future work. In addition our research raises the
question how well our extensions would perform in environments
where agents receive stochastic rewards throughout the episode. To
answer this question we plan to test our LDQN within a hunter prey
scenario where each episode runs for a fixed number of time-steps,
with the prey being re-inserted at a random position each time it
is caught [20]. Furthermore we plan to investigate how our LDQN
responds to environments with more than two agent by conducting
CMOTP and hunter-prey scenarios with four agents.
To summarize our contributions:
1) In this work we have shown that leniency can be applied to
MA-DRL, enabling agents to converge upon optimal joint policies
within fully-cooperative environments that require implicit coordi-
nation strategies and yield stochastic rewards.
2) We find that LDQNs significantly outperform standard and
scheduled-HDQNs within environments that yield stochastic re-
wards, replicating findings from tabular settings [35].
3) We introduced two extensions to leniency, including a retroac-
tive temperature decay schedule that prevents the premature decay
of temperatures for state-action pairs and a T(st)-Greedy explo-
ration strategy that encourages agents to remain exploratory in
states with a high average temperature value. The extensions can
in theory also be used by lenient agents within non-deep settings.
4) Our LDQN hyperparameter analysis revealed that the highest
performing agents within stochastic reward domains use a steep
temperature decay schedule that maintains high temperatures for
early transitions combined with a temperature modification coef-
ficient that slows down the transition from optimist to average
reward learner, and an exploration exponent that delays the transi-
tion from explorer to exploiter.
5) We demonstrate that the CMOTP [8] can be used as a benchmark-
ing environment for MA-DRL, requiring reinforcement learning
agents to learn fully-cooperative joint-policies from processing
high dimensional and noisy image observations.
6) Finally, we introduce two extensions to the CMOTP. First we
include narrow passages, allowing us to test lenient agents' abil-
ity to prevent the premature decay of temperature values. Our
second extension introduces two dropzones that yield stochastic
rewards, testing the agents' ability to converge towards an optimal
joint-policy while receiving misleading rewards.
9 ACKNOWLEDGMENTS
We thank the HAL Allergy Group for partially funding the PhD of
Gregory Palmer and gratefully acknowledge the support of NVIDIA
Corporation with the donation of the Titan X Pascal GPU that
enabled this research.
Lenient Multi-Agent Deep Reinforcement Learning
AAMAS'18, July 2018, Stockholm, Sweden
[27] Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vini-
cius Zambaldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl
Tuyls, and Thore Graepel. 2017. Value-Decomposition Networks For Cooperative
Multi-Agent Learning. arXiv preprint arXiv:1706.05296 (2017).
[28] Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Kor-
jus, Juhan Aru, Jaan Aru, and Raul Vicente. 2017. Multiagent cooperation and
competition with deep reinforcement learning. PLoS One 12, 4 (2017), e0172395.
[29] Ming Tan. 1993. Multi-agent reinforcement learning: Independent vs. cooperative
agents. In Proceedings of the tenth international conference on machine learning.
330–337.
[30] Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, Xi Chen, Yan Duan,
John Schulman, Filip De Turck, and Pieter Abbeel. 2016. # Exploration: A Study
of Count-Based Exploration for Deep Reinforcement Learning. arXiv preprint
arXiv:1611.04717 (2016).
[31] Karl Tuyls and Gerhard Weiss. 2012. Multiagent Learning: Basics, Challenges,
and Prospects. AI Magazine 33, 3 (2012), 41–52.
[32] Hado Van Hasselt. 2010. Double Q-learning. In Advances in Neural Information
Processing Systems. 2613–2621.
[33] Hado Van Hasselt, Arthur Guez, and David Silver. 2016. Deep Reinforcement
Learning with Double Q-Learning. AAAI (2016), 2094–2100.
[34] Christopher JCH Watkins and Peter Dayan. 1992. Q-learning. Machine learning
8, 3-4 (1992), 279–292.
[35] Ermo Wei and Sean Luke. 2016. Lenient Learning in Independent-Learner Sto-
chastic Cooperative Games. Journal of Machine Learning Research 17, 84 (2016),
1–42. http://jmlr.org/papers/v17/15-417.html
[36] R Paul Wiegand. 2003. An analysis of cooperative coevolutionary algorithms. Ph.D.
Dissertation. George Mason University Virginia.
[37] Yinliang Xu, Wei Zhang, Wenxin Liu, and Frank Ferrese. 2012. Multiagent-based
reinforcement learning for optimal reactive power dispatch. IEEE Transactions
on Systems, Man, and Cybernetics, Part C (Applications and Reviews) 42, 6 (2012),
1742–1751.
MIT press.
REFERENCES
[1] Ioannis Antonoglou, John Quan Tom Schaul, and David Silver. 2015. Prioritized
Experience Replay. arXiv preprint arXiv:1511.05952 (2015).
[2] Tucker Balch and Ronald C Arkin. 1994. Communication in reactive multiagent
robotic systems. Autonomous robots 1, 1 (1994), 27–52.
[3] Nikos Barbalios and Panagiotis Tzionas. 2014. A robust approach for multi-agent
natural resource allocation based on stochastic optimization algorithms. Applied
Soft Computing 18 (2014), 12–24.
[4] Andrew Barto and Richard Sutton. 1998. Reinforcement learning: An introduction.
[5] Daan Bloembergen, Daniel Hennes, Michael Kaisers, and Karl Tuyls. 2015. Evo-
lutionary dynamics of multi-agent learning: A survey. Journal of Artificial Intelli-
gence Research 53 (2015), 659–697.
[6] Daan Bloembergen, Michael Kaisers, and Karl Tuyls. 2011. Empirical and the-
oretical support for lenient learning. In The 10th International Conference on
Autonomous Agents and Multiagent Systems-Volume 3. International Foundation
for Autonomous Agents and Multiagent Systems, 1105–1106.
[7] Lucian Busoniu, Robert Babuska, and Bart De Schutter. 2008. A comprehensive
survey of multiagent reinforcement learning. IEEE Transactions on Systems, Man,
And Cybernetics-Part C: Applications and Reviews, 38 (2), 2008 (2008).
[8] Lucian Buşoniu, Robert Babuška, and Bart De Schutter. 2010. Multi-agent re-
inforcement learning: An overview. In Innovations in multi-agent systems and
applications-1. Springer, 183–221.
[9] Moses S Charikar. 2002. Similarity estimation techniques from rounding algo-
rithms. In Proceedings of the thiry-fourth annual ACM symposium on Theory of
computing. ACM, 380–388.
[10] Tim de Bruin, Jens Kober, Karl Tuyls, and Robert Babuška. 2015. The importance
of experience replay database composition in deep reinforcement learning. In
Deep Reinforcement Learning Workshop, NIPS.
[11] Jakob Foerster, Nantas Nardelli, Gregory Farquhar, Philip Torr, Pushmeet Kohli,
Shimon Whiteson, et al. 2017. Stabilising Experience Replay for Deep Multi-Agent
Reinforcement Learning. arXiv preprint arXiv:1702.08887 (2017).
[12] Shixiang Gu, Ethan Holly, Timothy Lillicrap, and Sergey Levine. 2016. Deep
Reinforcement Learning for Robotic Manipulation with Asynchronous Off-Policy
Updates. arXiv preprint arXiv:1610.00633 (2016).
[13] Jayesh K Gupta, Maxim Egorov, and Mykel Kochenderfer. 2017. Cooperative
Multi-Agent Control Using Deep Reinforcement Learning. In Proceedings of the
Adaptive and Learning Agents workshop (at AAMAS 2017).
[14] Pablo Hernandez-Leal, Michael Kaisers, Tim Baarslag, and Enrique Munoz de
Cote. 2017. A Survey of Learning in Multiagent Environments: Dealing with
Non-Stationarity. arXiv preprint arXiv:1707.09183 (2017).
[15] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. 1996. Rein-
forcement learning: A survey. Journal of artificial intelligence research 4 (1996),
237–285.
[16] Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimiza-
tion. In Proceedings of the 3rd International Conference on Learning Representations
(ICLR).
[17] Guillaume Lample and Devendra Singh Chaplot. 2017. Playing FPS Games with
Deep Reinforcement Learning. AAAI (2017), 2140–2146.
[18] Long-H Lin. 1992. Self-improving reactive agents based on reinforcement learn-
ing, planning and teaching. Machine learning 8, 3/4 (1992), 69–97.
[19] Laëtitia Matignon, Guillaume J Laurent, and Nadine Le Fort-Piat. 2007. Hysteretic
q-learning: an algorithm for decentralized reinforcement learning in cooperative
multi-agent teams. In Intelligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ
International Conference on. IEEE, 64–69.
[20] Laetitia Matignon, Guillaume J Laurent, and Nadine Le Fort-Piat. 2012. Indepen-
dent reinforcement learners in cooperative Markov games: a survey regarding
coordination problems. The Knowledge Engineering Review 27, 1 (2012), 1–31.
[21] 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), 529–533.
[22] Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan P How, and
John Vian. 2017. Deep decentralized multi-task multi-agent reinforcement learn-
ing under partial observability. In International Conference on Machine Learning.
2681–2690.
[23] Liviu Panait, Keith Sullivan, and Sean Luke. 2006. Lenient learners in cooperative
multiagent systems. In Proceedings of the fifth international joint conference on
Autonomous agents and multiagent systems. ACM, 801–803.
[24] Liviu Panait, Karl Tuyls, and Sean Luke. 2008. Theoretical advantages of le-
nient learners: An evolutionary game theoretic perspective. Journal of Machine
Learning Research 9, Mar (2008), 423–457.
[25] Mitchell A Potter and Kenneth A De Jong. 1994. A cooperative coevolutionary
approach to function optimization. In International Conference on Parallel Problem
Solving from Nature. Springer, 249–257.
[26] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. 2015. Prioritized
experience replay. arXiv preprint arXiv:1511.05952 (2015).
|
1005.4877 | 5 | 1005 | 2015-02-05T09:40:35 | Set-Monotonicity Implies Kelly-Strategyproofness | [
"cs.MA"
] | This paper studies the strategic manipulation of set-valued social choice functions according to Kelly's preference extension, which prescribes that one set of alternatives is preferred to another if and only if all elements of the former are preferred to all elements of the latter. It is shown that set-monotonicity---a new variant of Maskin-monotonicity---implies Kelly-strategyproofness in comprehensive subdomains of the linear domain. Interestingly, there are a handful of appealing Condorcet extensions---such as the top cycle, the minimal covering set, and the bipartisan set---that satisfy set-monotonicity even in the unrestricted linear domain, thereby answering questions raised independently by Barber\`a (1977) and Kelly (1977). | cs.MA | cs |
Set-Monotonicity Implies
Kelly-Strategyproofness
Felix Brandt
Technische Universitat Munchen
Munich, Germany
[email protected]
This paper studies the strategic manipulation of set-valued social choice
functions according to Kelly's preference extension, which prescribes that
one set of alternatives is preferred to another if and only if all elements
of the former are preferred to all elements of the latter.
It is shown that
set-monotonicity -- a new variant of Maskin-monotonicity -- implies Kelly-
strategyproofness in comprehensive subdomains of the linear domain.
In-
terestingly, there are a handful of appealing Condorcet extensions -- such as
the top cycle, the minimal covering set, and the bipartisan set -- that satisfy
set-monotonicity even in the unrestricted linear domain, thereby answering
questions raised independently by Barber`a (1977a) and Kelly (1977).
Keywords: Social Choice Theory, Strategyproofness, Kelly's Preference Extension
JEL Classifications Codes: D71, C70
1. Introduction
One of the central results in microeconomic theory states that every non-trivial so-
cial choice function (SCF) -- a function mapping individual preferences to a collective
choice -- is susceptible to strategic manipulation (Gibbard, 1973; Satterthwaite, 1975).
However, the classic theorem by Gibbard and Satterthwaite only applies to resolute, i.e.,
single-valued, SCFs. The notion of a resolute SCF is rather restricted and artificial.1
For example, consider a situation with two agents and two alternatives such that each
1For instance, Gardenfors (1976) claims that "[resoluteness] is a rather restrictive and unnatural as-
sumption." In a similar vein, Kelly (1977) writes that "the Gibbard-Satterthwaite theorem [. . . ] uses
an assumption of singlevaluedness which is unreasonable" and Taylor (2005) that "If there is a weak-
ness to the Gibbard-Satterthwaite theorem, it is the assumption that winners are unique." This senti-
ment is echoed by various other authors (see, e.g., Barber`a, 1977b; Feldman, 1979b; Bandyopadhyay,
1983a,b; Duggan and Schwartz, 2000; Nehring, 2000; Ching and Zhou, 2002).
1
agent prefers a different alternative. The problem is not that a resolute SCF has to pick
a single alternative (which is a well-motivated practical requirement), but that it has to
pick a single alternative based on the individual preferences alone (see also, Kelly, 1977).
As a consequence, resoluteness is at variance with elementary notions of fairness such as
neutrality and anonymity.
In order to remedy this shortcoming, Gibbard (1977) went on to characterize the class
of strategyproof social decision schemes (SDSs), i.e., aggregation functions that yield
probability distributions over the set of alternatives rather than single alternatives (see
also Gibbard, 1978; Barber`a, 1979). This class consists of rather degenerate SDSs and
Gibbard's characterization is therefore often interpreted as another impossibility result.
However, Gibbard's theorem rests on unusually strong assumptions with respect to the
agents' preferences. In contrast to the traditional setup in social choice theory, which
typically only involves ordinal preferences, his result relies on the expected utility axioms
of von Neumann and Morgenstern (1947), and hence on the existence of linear utility
functions, in order to compare lotteries over alternatives.
The gap between Gibbard and Satterthwaite's theorem for resolute SCFs and
Gibbard's theorem for SDSs has been filled by a number of
impossibility re-
sults with varying underlying notions of how to compare sets of alternatives with
each other (e.g., Gardenfors, 1976; Barber`a, 1977a,b; Kelly, 1977; Feldman, 1979a;
MacIntyre and Pattanaik, 1981; Bandyopadhyay, 1982, 1983b; Duggan and Schwartz,
2000; Barber`a et al., 2001; Ching and Zhou, 2002; Sato, 2008; Umezawa, 2009), many
of which are surveyed by Taylor (2005) and Barber`a (2010). In this paper, we will be
concerned with the one of the weakest (and therefore least controversial) preference ex-
tensions from alternatives to sets due to Kelly (1977). According to this definition, a
set of alternatives is weakly preferred to another set of alternatives if all elements of
the former are weakly preferred to all elements of the latter. A nice aspect of Kelly's
preference extension is that its underlying behavioral assumptions are quite minimalistic
(which strengthens impossibility results), yet reasonable enough to motivate meaning-
ful positive results. Kelly's extension models that the agents are complete unaware of
the tie-breaking mechanism that is used to eventually pick a single alternative. The
question pursued in this paper is whether this uncertainty can be exploited to achieve
strategyproofness.
Barber`a (1977a) and Kelly (1977) have shown independently that all non-trivial SCFs
that are rationalizable via a quasi-transitive relation are manipulable according to Kelly's
extension.2 However, it is it well-known that (quasi-transitive) rationalizability by itself
is unduly restrictive (see, e.g., Mas-Colell and Sonnenschein, 1972). As a consequence,
Kelly (1977) concludes his paper by contemplating that "one plausible interpretation
of such a theorem is that, rather than demonstrating the impossibility of reasonable
strategy-proof social choice functions, it is part of a critique of the regularity [ratio-
nalizability] conditions" and Barber`a (1977a) states that "whether a nonrationalizable
collective choice rule exists which is not manipulable and always leads to nonempty
choices for nonempty finite issues is an open question." Also referring to nonrationaliz-
2Barber`a (1977a) actually uses an extension that is even weaker than that of Kelly (1977).
2
able choice functions, Kelly (1977) writes: "it is an open question how far nondictatorship
can be strengthened in this sort of direction and still avoid impossibility results." The
condition of rationalizability has been significantly weakened in subsequent impossibil-
ity results (MacIntyre and Pattanaik, 1981; Bandyopadhyay, 1982, 1983b). At the same
time, it has been noted that more positive results can be obtained for antisymmetric
(i.e., linear) individual preferences. In particular, it was shown that the omninomina-
tion rule (Gardenfors, 1976), the Pareto rule (Feldman, 1979a,b), the Condorcet rule
(Gardenfors, 1976; Nehring, 2000), and the top cycle (MacIntyre and Pattanaik, 1981;
Bandyopadhyay, 1983a; Sanver and Zwicker, 2012) are strategyproof when preference
are linear (see Remark 1 for more details about these SCFs). However, all these rules
are very indecisive and the latter two may even return Pareto-dominated alternatives.
In this paper, we propose a new variant of Maskin-monotonicity for set-valued SCFs
called set-monotonicity and show that all set-monotonic SCFs are strategyproof in suf-
ficiently rich subdomains of the linear domain. This covers all of the positive results
mentioned above and proves that some -- much more discriminating -- SCFs are strate-
gyproof. Set-monotonicity requires the invariance of choice sets under the weakening
of unchosen alternatives and is satisfied by the omninomination rule, the Pareto rule,
the Condorcet rule, and a handful of appealing Condorcet extensions such as top cycle,
the minimal covering set, and the bipartisan set. Since set-monotonicity coincides with
Maskin-monotonicity in the context of resolute SCFs, this characterization can be seen
as a set-valued generalization of the Muller and Satterthwaite (1977) theorem.
Nehring (2000) has proved a similar extension of the Muller-Sattherthwaite theorem
by showing that Maskin-monotonicity implies strategyproofness of set-valued SCFs in
a sense marginally weaker than that of Kelly (see also Remark 6).3 However, while
Maskin-monotonicity is prohibitive in the general domain, set-monotonicity is not. The
conditions themselves are independent, but we show that Maskin-montonicity implies
set-monotonicity when assuming independence of unchosen alternatives.
We conclude the paper with a number of remarks concerning group-strategyproofness,
stronger preference extensions, weak preferences, weaker domain conditions, and strate-
gic abstention.
2. Preliminaries
Let N = {1, . . . , n} be a finite set of agents, A a finite and nonempty set of alternatives,
and L the set of all linear (i.e., complete, transitive, and antisymmetric) preference
relations over A. For Ri ∈ L , x Ri y denotes that agent i values alternative x at least
as much as alternative y. We write Pi for the strict part of Ri, i.e., x Pi y if x Ri y but
not y Ri x. The (strict) lower contour set of alternative x with respect to Pi is denoted
by L(x, Ri) = {y ∈ A : x Pi y}. For convenience, we will represent preference relations
3According to Nehring's definition, a manipulator is only better off if he strictly prefers all alternatives
in the new choice set to all alternatives in the original choice set. For linear preferences, the two
definitions only differ in whether there can be a single alternative at the intersection of both choice
sets or not.
3
as comma-separated lists. For example, a Pi b Pi c will be written as Ri : a, b, c. Two
distinct alternatives x and y are adjacent in Ri if there is no z with x Pi z Pi y.
i ⊆ R′′
A (Cartesian) domain of preference profiles D is defined as D = Qi∈N
Di ⊆ L N .
The maximal domain L N will be referred to as the general domain. We say that R′′
lies in the comprehensive closure of R and R′ if for all i ∈ N , Ri ∩ R′
i . A
domain D is comprehensive if for all R, R′ ∈ D and R′′ ∈ L N such that R′′ lies in the
comprehensive closure of R and R′, R′′ ∈ D (Nehring, 2000). For example, the domain
of all linear extensions of a fixed partial order is comprehensive. For a given preference
profile R ∈ D, R−i = (R1, . . . , Ri−1, Ri+1, . . . , Rn) denotes the vector of all preference
relations except that of agent i. An alternative x ∈ A is called a Condorcet winner if
{i ∈ N : x Pi y} > n/2 for all y ∈ A \ {x}.
Our central object of study are social choice functions. A social choice function (SCF)
is a function f that maps a preference profile R ∈ D to a nonempty subset of alternatives
f (R). f is resolute if f (R) = 1 for all R ∈ D. A Condorcet extension is an SCF that
uniquely selects a Condorcet winner whenever one exists.
2.1. Monotonicity
We will consider three variants of monotonicity: a weak standard notion and two
strengthenings, one of which was proposed by Maskin (1999) and one of which is new to
this paper. For a given preference profile R, an agent i, and two adjacent alternatives
x, y such that y Pi x, Ri:(x,y) denotes the preference profile in which agent i swapped
alternatives x and y and that is otherwise identical to R.
Definition 1. Let R, R′ ∈ D, i ∈ N , and x, y ∈ A such that R′ = Ri:(x,y). Then, SCF f
satisfies monotonicity, Maskin-monotonicity, or set-monotonicity, if
x ∈ f (R) implies x ∈ f (R′),
(monotonicity)
z ∈ f (R) and y ∈ A \ {z} implies z ∈ f (R′), or
Z = f (R) and y ∈ A \ Z implies Z = f (R′), respectively.
(Maskin-monotonicity)
(set-monotonicity)
The intuitive meaning of these definitions is as follows. An SCF satisfies monotonicity
if a chosen alternative remains in the choice set when it is strengthened with respect to
another alternative; it satisfies Maskin-monotonicity if a chosen alternative remains in
the choice set when weakening another alternative; and it satisfies set-monotonicity if
the choice set is invariant under the weakening of unchosen alternatives.4 5
Clearly, Maskin-monotonicity implies monotonicity and, as will be shown in Propo-
sition 2, set-monotonicity also implies monotonicity. Despite the similar appearance,
set-monotonicity is logically independent of Maskin-monotonicity. Set-monotonicity has
4Sanver and Zwicker (2012) study monotonicity properties for set-valued SCFs in general. None of the
properties they consider is equivalent to set-monotonicity.
5Note that set-monotonicity is in conflict with decisiveness. For instance, non-trivial set-monotonic
SCFs cannot satisfy the (rather strong) positive responsiveness condition introduced by Barber`a
(1977b).
4
a stronger antecedent and a stronger consequence. For example, consider a comprehen-
sive single-agent domain consisting of the relations R1 and R′
1 given by
R1 : a, b, c,
R′
b, a, c.
1 :
and
If we define SCF f by letting f ((R1)) = {c} and f ((R′
monotonicity, but violates set-monotonicity.
letting f ((R1)) = {a, b, c} and f ((R′
Maskin-monotonicity.
1)) = {b, c}, f satisfies Maskin-
If, on the other hand, we define f by
1)) = {b, a}, f satisfies set-monotonicity, but violates
Set-monotonicity coincides with Maskin-monotonicity in the context of resolute SCFs.
Under the condition of independence of unchosen alternatives which is satisfied by var-
ious set-valued SCFs, set-monotonicity is weaker than Maskin-monotonicity. Moreover,
set-monotonicity implies independence of unchosen alternatives. Independence of uncho-
sen alternatives was introduced by Laslier (1997) in the context of tournament solutions
(as "independence of the losers") and requires that the choice set is invariant under
modifications of the preference profile with respect to unchosen alternatives.
Definition 2. An SCF f satisfies independence of unchosen alternatives (IUA) if for
all R, R′ ∈ D such that Ri{x,y} = R′
i{x,y} for all x ∈ f (R), y ∈ A, and i ∈ N ,
f (R) = f (R′).
Proposition 1. Maskin-monotonicity and IUA imply set-monotonicity.
Proof. Let f be an SCF, R ∈ D, i ∈ N , x ∈ A, y ∈ A \ f (R), and R′ = Ri:(x,y).
Maskin-monotonicity implies that f (R) ⊆ f (R′). Now, assume for contradiction that
there is some x′ ∈ f (R′) \ f (R). Since x′ 6∈ f (R), it follows from Maskin-monotonicity
that there is some y′ ∈ A that is strengthened with respect to x′ when moving from R′
to R. Hence, R′ = Ri:(x′,y′), x′ = x, and y′ = y. Since x = x′ 6∈ f (R) and y 6∈ f (R) by
assumption, IUA implies that f (R) = f (R′), a contradiction.
Proposition 2. Set-monotonicity implies monotonicity and IUA.
Proof. We first show that set-monotonicity implies monotonicity. Let f be a set-
monotonic SCF, R ∈ D, i ∈ N , x ∈ f (R), y ∈ A, and R′ ∈ Ri:(x,y). Clearly, in case
y 6∈ f (R), set-monotonicity implies that f (R′) = f (R) and thus x ∈ f (R′). If, on the
other hand, y ∈ f (R), assume for contradiction that x 6∈ f (R′). When moving from R′
to R, y is strengthened with respect to outside alternative x, and set-monotonicity again
implies that f (R) = f (R′), a contradiction. The fact that set-monotonicity implies IUA
is straightforward from the definitions.
2.2. Strategyproofness
An SCF is manipulable if an agent can misrepresent his preferences in order to obtain a
more preferred outcome. Whether one choice set is preferred to another depends on how
the preferences over individual alternatives are to be extended to sets of alternatives.
5
In the absence of information about the tie-breaking mechanism that eventually picks a
single alternative from any choice set, preferences over choice sets are obtained by the
conservative extension bRi (Kelly, 1977), where for any pair of nonempty sets X, Y ⊆ A
and preference relation Ri,
X bRi Y if and only if x Ri y for all x ∈ X and y ∈ Y .
Clearly, in all but the simplest cases, bRi is incomplete, i.e., many pairs of choice sets are
incomparable. The strict part of bRi is denoted by bPi, i.e., X bPi Y if and only if X bRi Y
and x Pi y for at least one pair of x ∈ X and y ∈ Y . For linear preferences, X bPi Y if
and only if x Pi y for all x ∈ X and y ∈ Y with x 6= y. Hence, X ∩ Y ≤ 1.
Definition 3. An SCF is Kelly-strategyproof if there exist no R, R′ ∈ D and i ∈ N
with R−i = R′
−i such that f (R′) bPi f (R).
Kelly-strategyproofness is a very weak notion of strategyproofness. Nevertheless, most
well-known SCFs such as plurality, Borda's rule, Copeland's rule, Slater's rule, or plu-
rality with runoff fail to be Kelly-strategyproof in the general domain (see, e.g., Taylor,
2005, Theorem 2.2.2).
Muller and Satterthwaite (1977) have shown that, in the general domain, a resolute
SCF is strategyproof if and only if it satisfies Maskin-monotonicity. Unfortunately, as
famously shown by Gibbard (1973) and Satterthwaite (1975), only dictatorial or im-
posing resolute SCFs satisfy Maskin-monotonicity in the general domain. However,
Maskin-monotonicity still implies strategyproofness of resolute SCFs in many restricted
domains of interest (see, e.g., Klaus and Bochet, 2013), including the class of compre-
hensive domains considered in this paper (Nehring, 2000).
3. The Result
Theorem 1. Every set-monotonic SCF on a comprehensive domain is Kelly-
strategyproof.
Proof. Let f be a set-monotonic SCF and D a comprehensive domain. We first show
that set-monotonicity is equivalent to a version of set-monotonicity that is not restricted
to pairwise swaps of adjacent alternatives. Rather, we require that choice sets may only
change if an alternative is removed from a lower contour set of a chosen alternative. For
R, R′ ∈ D and i ∈ N with R−i = R′
−i, we say that R′ is an f -improvement over R if
for all x ∈ f (R), L(x, Pi) ⊆ L(x, P ′
i ). We claim that if R′ is an f -improvement over R,
then f (R′) = f (R). The statement can be shown by induction on d(R, R′) = Ri \ R′
i.
The induction basis is trivially satisfied because R = R′ if d(R, R′) = 0. Assume the
statement is true for all R and R′ with d(R, R′) < k and consider R and R′ such
that d(R, R′) = k. Since R 6= R′, there have to be two alternatives x, y ∈ A such
that x Pi y and y P ′
i, we may furthermore
assume that x and y are adjacent in Ri. (However, x and y need not be adjacent in
R′
i.) x 6∈ f (R) because otherwise R′ is not an f -improvement over R. Set-monotonicity
i x. Due to the transitivity of Ri and R′
6
implies that f (Ri:(y,x)) = f (R) and comprehensiveness that Ri:(y,x) ∈ D.
It follows
from d(Ri:(y,x), R′) < k and the induction hypothesis that f (Ri:(y,x)) = f (R′). Hence,
f (R) = f (R′),
Now, for the proof of the statement of the theorem, assume for contradiction that f is
not Kelly-strategyproof. Then, there have to be R, R′ ∈ D, and i ∈ N with R−i = R′
such that f (R′) bPi f (R). The latter obviously entails that f (R) 6= f (R′). The proof
idea is to find some R∗ in the comprehensive closure of R and R′ (and hence in D) such
that R∗ is an f -improvement over both R and R′, which implies that f (R) = f (R′), a
contradiction.
−i
Let m ∈ f (R′) such that L(m, Pi) ∩ f (R′) = ∅. In other words, m is the alternative in
f (R′) that is ranked lowest in Ri. Note that, if f (R) ∩ f (R′) 6= ∅, then f (R) ∩ f (R′) =
{m}. Next, we partition A into the strict lower contour set and the upper contour set
of m with respect to Ri, i.e., L = L(m, Pi) and U = A \ L and define a new preference
profile by letting
R∗
i = RiL ∪ R′
iU ∪ {(x, y) : x ∈ U, y ∈ L},
is ranked as in R′
i and the lower part as in Ri, and R∗ =
i.e., the upper part of R∗
i
(R−i, R∗
i ) = (R′
−i, R∗
i ).
x, y ∈ A with x R∗
R∗
R∗ lies in the comprehensive closure of R and R′ (and hence in D) because for all
i y. For x, y ∈ U , this follows from
i y, we have x Ri y or x R′
iU ; for x, y ∈ L, from R∗
i U = R′
R∗ is an f -improvement over R because f (R) ⊆ L∪{m} and L(x, P ∗
x ∈ L and L(m, P ∗
R∗ is an f -improvement over R′ because f (R′) ⊆ U and L(x, P ∗
x ∈ U . Hence, set-monotonicity implies that f (R∗) = f (R′) 6= f (R), a contradiction.
i ) = L(x, Pi) for all
i ) ⊇ L(m, Pi) = L. Set-monotonicity then implies that f (R∗) = f (R).
i ) ∪ L for all
i L = RiL; and for x ∈ U, y ∈ L from (x, y) ∈ Ri.
i ) = L(x, P ′
We conclude the paper with nine remarks.
Remark 1 (Set-monotonic SCFs). There are a number of rather attractive SCFs
that satisfy set-monotonicity in the general domain L N . In particular, every monotonic
SCF that satisfies the strong superset property (SSP) (i.e., choice sets are invariant
under the removal of unchosen alternatives) also satisfies set-monotonicity.6 Prominent
Condorcet extensions that satisfy both SSP and monotonicity include the Condorcet
rule (which selects a Condorcet winner whenever one exists and returns all alternatives
otherwise), the top cycle (also known as weak closure maximality, GETCHA, or the Smith
set) (Good, 1971; Smith, 1973; Bordes, 1976; Sen, 1977; Schwartz, 1986), the minimal
covering set (Dutta, 1990), the bipartisan set (Laffond et al., 1993) and variations of
6The strong superset property goes back to early work by Chernoff (1954) (where it was called
postulate 5∗) and is also known as bα (Brandt and Harrenstein, 2011), the attention filter axiom
(Masatlioglu et al., 2012), and outcast (Aizerman and Aleskerov, 1995). The term strong superset
property was first used by Bordes (1979). We refer to Monjardet (2008) for a more thorough discus-
sion of the origins of this condition.
7
these (see Laslier, 1997; Dutta and Laslier, 1999; Laslier, 2000; Brandt, 2011).78 Two
other SCFs that satisfy set-monotonicity are the Pareto rule and the omninomination
rule (which returns all alternatives that are top-ranked by at least one agent).
Remark 2 (Coarsenings of Kelly-strategyproof SCFs). For two SCFs f, f ′, we
say that f is a coarsening of f ′ if f ′(R) ⊆ f (R) for all R ∈ D. Kelly's preference
extension has the useful property that X bPi Y implies X ′ bPi Y ′ for all non-singleton
subsets X ′ ⊆ X and Y ′ ⊆ Y . Hence, every coarsening f of a Kelly-strategyproof SCF
f ′ is Kelly-strategyproof if f (R) = f ′(R) whenever f ′(R) = 1. As a consequence, the
(McKelvey) uncovered set (see Duggan, 2013), a coarsening of the minimal covering set
that returns singletons if and only if there is a Condorcet winner, is Kelly-strategyproof,
even though it violates set-monotonicity.9 In light of these comments, it seems interesting
to try to identify inclusion-minimal Kelly-strategyproof SCFs.
Remark 3 (Group-strategyproofness). The proof of Theorem 1 can be straightfor-
wardly extended to show that no group of agents can misstate their preferences in order
to obtain a more preferred outcome.
Remark 4 (Fishburn-strategyproofness).
It has been shown in other work that
Theorem 1 does not carry over to slightly more complete set extensions due to
Fishburn and Gardenfors (Brandt and Brill, 2011).
In fact, Pareto-optimality and
Fishburn-strategyproofness are already incompatible within the class of majoritarian
SCFs (Brandt and Geist, 2014).
Remark 5 (Necessary conditions).
It seems like there are no natural necessary
conditions for Kelly-strategyproofness as long as preferences are linear. For example,
any SCF f such that f (R) > (A/2) + 1 for all R ∈ D satisfies Kelly-strategyproofness
simply because no pair of resulting choice sets is comparable. This observation allows
one to easily construct Kelly-strategyproof SCFs that violate set-monotonicity, Maskin-
monotonicity, or any other reasonable form of monotonicity.
Remark 6 (Weak preferences). When individual preference relations do not have
to be antisymmetric, a number of results mentioned in the introduction rule out the
possibility of reasonable Kelly-strategyproof SCFs. A new result of this kind, which
strengthens some existing theorems, is given as Theorem 2 in the Appendix.
7Remarkably, the robustness of the minimal covering set and the bipartisan set with respect to strategic
manipulation also extends to agenda manipulation. The strong superset property precisely states that
an SCF is resistant to adding and deleting losing alternatives (see also the discussion by Bordes, 1983).
Moreover, both SCFs are composition-consistent, i.e., they are strongly resistant to the introduction
of clones (Laffond et al., 1996). Scoring rules like plurality and Borda's rule are prone to both
types of agenda manipulation (Laslier, 1996; Brandt and Harrenstein, 2011) as well as to strategic
manipulation.
8Another prominent Condorcet extension -- the tournament equilibrium set (Schwartz, 1990) -- was con-
jectured to satisfy SSP and monotonicity for almost 20 years. This conjecture was recently disproved
by Brandt et al. (2013). In fact, it can be shown that the tournament equilibrium set as well as the
related minimal extending set (Brandt, 2011) can be Kelly-manipulated.
9For generalized strategyproofness as defined by Nehring (2000) (see Remark 6), the second condition
is not required and every coarsening of a strategyproof SCF is strategyproof.
8
The proof of Theorem 1 can be adapted for weak preferences to show that set-
monotonicity implies Nehring's generalized strategyproofness, a weakening of Kelly-
strategyproofness (Nehring, 2000).10 Generalized strategyproofness is defined by letting
agent i prefer set X to set Y if and only if x Pi y for all x ∈ X and y ∈ Y .
Remark 7 (Matching markets). Theorem 1 can be straightforwardly extended to
domains in which the indifference relation is fixed for each agent. This is, for example,
the case in matching markets where each agent is typically assumed to be indifferent
between all matchings in which his assignment is identical.
Remark 8 (Connected domains). Theorem 1 does not hold for a weakening of
comprehensiveness, which Nehring (2000) refers to as connectedness. A domain D is
connected if for all i ∈ N , Ri, R′
i ∈ L \
{Ri, R′
i} with R ∩ R′ ⊆ R′′. The
following single-agent SCF f , defined on a connected -- but not comprehensive -- domain
of size three, has been adapted from Nehring (2000) and satisfies set-monotonicity while
it violates Kelly-strategyproofness:
i ∈ Di, the following holds: if there is some R′′
i} with R ∩ R′ ⊆ R′′, then there is some R′′
i ∈ Di \ {Ri, R′
R1 : a, b, c, d,
R′
1 : a, b, d, c,
R′′
b, a, d, c,
1 :
f ((R1)) = {c},
f ((R′
f ((R′′
1)) = {a, b, c, d},
1)) = {b}.
Remark 9 (Strategic abstention). Fishburn and Brams (1983) introduced a partic-
ularly natural variant of strategic manipulation where agents obtain a more preferred
outcome by abstaining the election. Moulin (1988) has shown that every resolute Con-
dorcet extension can be manipulated in this way and thus suffers from the so-called
no-show paradox. However, Moulin's proof strongly relies on resoluteness. If preferences
over sets are given by Kelly's extension, set-valued Condorcet extension that satisfy
Kelly-strategyproofness also cannot be manipulated by abstention under very mild con-
ditions. This is, for instance, the case for all SCFs whose outcome only depends on
pairwise majority margins (which covers all of the Condorcet extensions mentioned in
Remark 1) because any manipulation by abstention can be turned into a manipulation
by strategic misrepresentation.11
Acknowledgments
I am grateful to Florian Brandl, Markus Brill, and Paul Harrenstein for helpful dis-
cussions and comments. This material is based on work supported by the Deutsche
Forschungsgemeinschaft under grants BR 2312/3-3, BR 2312/7-1, and BR 2312/7-2.
Early results of this paper were presented at the 22nd International Joint Conference
on Artificial Intelligence (Barcelona, July 2011). A previous version of this paper, titled
"Group-Strategyproof Irresolute Social Choice Functions," circulated since 2010.
10For generalized strategyproofness, it would also suffice to require a weakening of set-monotonicity in
which the choice set can only get smaller when unchosen alternatives are weakened.
11P´erez (2001) and Jimeno et al. (2009) have proved versions of Moulin's theorem (Moulin, 1988) for
set-valued Condorcet extensions by using stronger assumptions on preference over sets.
9
References
M. Aizerman and F. Aleskerov. Theory of Choice, volume 38 of Studies in Mathematical
and Managerial Economics. North-Holland, 1995.
T. Bandyopadhyay. Threats, counter-threats and strategic manipulation for non-binary
group decision rules. Mathematical Social Sciences, 2(2):145 -- 155, 1982.
T. Bandyopadhyay. Multi-valued decision rules and coalitional non-manipulability. Eco-
nomics Letters, 13(1):37 -- 44, 1983a.
T. Bandyopadhyay. Manipulation of non-imposed, non-oligarchic, non-binary group
decision rules. Economics Letters, 11(1 -- 2):69 -- 73, 1983b.
S. Barber`a. Manipulation of social decision functions. Journal of Economic Theory, 15
(2):266 -- 278, 1977a.
S. Barber`a. The manipulation of social choice mechanisms that do not leave "too much"
to chance. Econometrica, 45(7):1573 -- 1588, 1977b.
S. Barber`a. A note on group strategy-proof decision schemes. Econometrica, 47(3):
637 -- 640, 1979.
S. Barber`a. Strategy-proof social choice. In K. J. Arrow, A. K. Sen, and K. Suzumura,
editors, Handbook of Social Choice and Welfare, volume 2, chapter 25, pages 731 -- 832.
Elsevier, 2010.
S. Barber`a, B. Dutta, and A. Sen. Strategy-proof social choice correspondences. Journal
of Economic Theory, 101(2):374 -- 394, 2001.
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.
North-Holland, 1979.
G. Bordes. On the possibility of reasonable consistent majoritarian choice: Some positive
results. Journal of Economic Theory, 31:122 -- 132, 1983.
F. Brandt. Minimal stable sets in tournaments. Journal of Economic Theory, 146(4):
1481 -- 1499, 2011.
F. Brandt and M. Brill. Necessary and sufficient conditions for the strategyproofness of
irresolute social choice functions. In Proceedings of the 13th Conference on Theoretical
Aspects of Rationality and Knowledge (TARK), pages 136 -- 142. ACM Press, 2011.
F. Brandt and C. Geist. Finding strategyproof social choice functions via SAT solving.
In Proceedings of the 13th International Conference on Autonomous Agents and Multi-
Agent Systems (AAMAS), pages 1193 -- 1200. IFAAMAS, 2014.
10
F. Brandt and P. Harrenstein. Set-rationalizable choice and self-stability. Journal of
Economic Theory, 146(4):1721 -- 1731, 2011.
F. Brandt, M. Chudnovsky, I. Kim, G. Liu, S. Norin, A. Scott, P. Seymour, and
S. Thomass´e. A counterexample to a conjecture of Schwartz. Social Choice and
Welfare, 40:739 -- 743, 2013.
H. Chernoff. Rational selection of decision functions. Econometrica, 22:422 -- 443, 1954.
S. Ching and L. Zhou. Multi-valued strategy-proof social choice rules. Social Choice and
Welfare, 19:569 -- 580, 2002.
J. Duggan. Uncovered sets. Social Choice and Welfare, 41:489 -- 535, 2013.
J. Duggan and T. Schwartz. Strategic manipulability without resoluteness or shared
beliefs: Gibbard-Satterthwaite generalized. Social Choice and Welfare, 17(1):85 -- 93,
2000.
B. Dutta. On the tournament equilibrium set. Social Choice and Welfare, 7(4):381 -- 383,
1990.
B. Dutta and J.-F. Laslier. Comparison functions and choice correspondences. Social
Choice and Welfare, 16(4):513 -- 532, 1999.
A. Feldman. Manipulation and the Pareto rule. Journal of Economic Theory, 21:473 --
482, 1979a.
A. Feldman. Nonmanipulable multi-valued social choice decision functions. Public
Choice, 34:177 -- 188, 1979b.
P. C. Fishburn and S. J. Brams. Paradoxes of preferential voting. Mathematics Magazine,
56(4):207 -- 214, 1983.
P. Gardenfors. Manipulation of social choice functions. Journal of Economic Theory, 13
(2):217 -- 228, 1976.
A. Gibbard. Manipulation of voting schemes. Econometrica, 41:587 -- 602, 1973.
A. Gibbard. Manipulation of schemes that mix voting with chance. Econometrica, 45
(3):665 -- 681, 1977.
A. Gibbard. Straightforwardness of game forms with lotteries as outcomes. Economet-
rica, 46(3):595 -- 614, 1978.
I. J. Good. A note on Condorcet sets. Public Choice, 10:97 -- 101, 1971.
J. L. Jimeno, J. P´erez, and E. Garc´ıa. An extension of the Moulin No Show Paradox
for voting correspondences. Social Choice and Welfare, 33(3):343 -- 459, 2009.
11
J. S. Kelly. Strategy-proofness and social choice functions without single-valuedness.
Econometrica, 45(2):439 -- 446, 1977.
B. Klaus and O. Bochet. The relation between monotonicity and strategyproofness.
Social Choice and Welfare, 40(1):41 -- 63, 2013.
G. Laffond, J.-F. Laslier, and M. Le Breton. The bipartisan set of a tournament game.
Games and Economic Behavior, 5:182 -- 201, 1993.
G. Laffond, J. Lain´e, and J.-F. Laslier. Composition-consistent tournament solutions
and social choice functions. Social Choice and Welfare, 13:75 -- 93, 1996.
J.-F. Laslier. Rank-based choice correspondences. Economics Letters, 52:279 -- 286, 1996.
J.-F. Laslier. Tournament Solutions and Majority Voting. Springer-Verlag, 1997.
J.-F. Laslier. Aggregation of preferences with a variable set of alternatives. Social Choice
and Welfare, 17:269 -- 282, 2000.
I. MacIntyre and P. K. Pattanaik. Strategic voting under minimally binary group decision
functions. Journal of Economic Theory, 25(3):338 -- 352, 1981.
A. Mas-Colell and H. Sonnenschein. General possibility theorems for group decisions.
Review of Economic Studies, 39(2):185 -- 192, 1972.
Y. Masatlioglu, D. Nakajima, and E. Y. Ozbay. Revealed attention. American Economic
Review, 102(5):2183 -- 2205, 2012.
E. Maskin. Nash equilibrium and welfare optimality. Review of Economic Studies, 66
(26):23 -- 38, 1999.
B. Monjardet. Statement of precedence and a comment on IIA terminology. Games and
Economic Behavior, 62:736 -- 738, 2008.
H. Moulin. Condorcet's principle implies the no show paradox. Journal of Economic
Theory, 45:53 -- 64, 1988.
E. Muller and M. A. Satterthwaite. The equivalence of strong positive association and
strategy-proofness. Journal of Economic Theory, 14(2):412 -- 418, 1977.
K. Nehring. Monotonicity implies generalized strategy-proofness for correspondences.
Social Choice and Welfare, 17(2):367 -- 375, 2000.
J. P´erez. The strong no show paradoxes are a common flaw in Condorcet voting corre-
spondences. Social Choice and Welfare, 18(3):601 -- 616, 2001.
M. R. Sanver and W. S. Zwicker. Monotonicity properties and their adaption to irresolute
social choice rules. Social Choice and Welfare, 39(2 -- 3):371 -- 398, 2012.
12
S. Sato. On strategy-proof social choice correspondences. Social Choice and Welfare,
31:331 -- 343, 2008.
M. A. Satterthwaite. Strategy-proofness and Arrow's conditions: Existence and corre-
spondence theorems for voting procedures and social welfare functions. Journal of
Economic Theory, 10:187 -- 217, 1975.
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(1):19 -- 29, 1990.
A. K. Sen. Social choice theory: A re-examination. Econometrica, 45(1):53 -- 89, 1977.
J. H. Smith. Aggregation of preferences with variable electorate. Econometrica, 41(6):
1027 -- 1041, 1973.
A. D. Taylor. Social Choice and the Mathematics of Manipulation. Cambridge University
Press, 2005.
M. Umezawa. Coalitionally strategy-proof social choice correspondences and the Pareto
rule. Social Choice and Welfare, 33:151 -- 158, 2009.
J. von Neumann and O. Morgenstern. Theory of Games and Economic Behavior. Prince-
ton University Press, 2nd edition, 1947.
A. Appendix
For the domain of transitive and complete (but not necessarily antisymmetric) prefer-
ence relations RN , we show that all Condorcet extensions are Kelly-manipulable. This
strengthens Theorem 3 by Gardenfors (1976) and Theorem 8.1.2 by Taylor (2005), who
showed the same statement for a weaker notion of manipulability and a weaker notion of
Condorcet winners, respectively. When assuming that pairwise choices are made accord-
ing to majority rule, this also strengthens Theorems 1 and 2 by MacIntyre and Pattanaik
(1981). However, our construction requires that the number of agents is linear in the
number of alternatives.
Theorem 2. No Condorcet extension is Kelly-strategyproof in domain RN when there
are more than two alternatives.
Proof. Let A = {a1, . . . , am} with m ≥ 3 and consider the following preference profile R
with 3m agents.
In the representation below, sets denote indifference classes of the
agents.
13
R1, R2 : {a2, . . . , am}, a1
R3, R4 : {a1, a3, . . . , am}, a2
R2m−1, R2m : {a1, . . . , am−1}, am
R2m+1 : {a3, . . . , am}, a1, a2
...
...
...
...
R3m−1 : {a1, . . . , am−2}, am−1, am
R3m : {a2, . . . , am−1}, am, a1
For every alternative ai, there are two agents who prefer every alternative to ai and are
otherwise indifferent. Moreover, for every alternative ai there is one agent who prefers
every alternative except ai+1 to ai, ranks ai+1 below ai, and is otherwise indifferent.
Since f (R) yields a nonempty choice set, there has to be some ai ∈ f (R). Due to
the symmetry of the preference profile, we may assume without loss of generality that
a2 ∈ f (R). Now, let
R′
3, R′
4 : a1, {a3, . . . , am}, a2
4). That is, R′ is identical to R, except
and define R′ = (R−3, R′
that agent 3 lifted a1 on top and R′′ is identical to R′, except that agent 4 lifted a1 on
top. Observe that f (R′′) = {a1} because a1 is the Condorcet winner in R′′.
3) and R′′ = (R′
−4, R′
In case that a2 6∈ f (R′), agent 3 can manipulate as follows. Suppose R is the true
preference profile. Then, the least favorable alternative of agent 3 is chosen (possibly
among other alternatives). He can misstate his preferences as in R′ such that a2 is not
chosen. Since he is indifferent between all other alternatives, f (R′) bP3 f (R).
If a2 ∈ f (R′), agent 4 can manipulate similarly. Suppose R′ is the true preference
profile. Again, the least favorable alternative of agent 4 is chosen. By misstating his
preferences as in R′′, he can assure that one of his preferred alternatives, namely a1,
is selected exclusively because it is the Condorcet winner in R′′. Hence, f (R′′) bP ′
4
f (R′).
14
|
1208.6421 | 1 | 1208 | 2012-08-31T08:38:33 | A Novel Service Oriented Model for Query Identification and Solution Development using Semantic Web and Multi Agent System | [
"cs.MA",
"cs.SE"
] | In this paper, we propose to develop service model architecture by merging multi-agentsystems and semantic web technology. The proposed architecture works in two stages namely, Query Identification and Solution Development. A person referred to as customer will submit the problem details or requirements which will be referred to as a query. Anyone who can provide a service will need to register with the registrar module of the architecture. Services can be anything ranging from expert consultancy in the field of agriculture to academic research, from selling products to manufacturing goods, from medical help to legal issues or even providing logistics. Query submitted by customer is first parsed and then iteratively understood with the help of domain experts and the customer to get a precise set of properties. Query thus identified will be solved again with the help of intelligent agent systems which will search the semantic web for all those who can find or provide a solution. A workable solution workflow is created and then depending on the requirements, using the techniques of negotiation or auctioning, solution is implemented to complete the service for customer. This part is termed as solution development. In this service oriented architecture, we first try to analyze the complex set of user requirements then try to provide best possible solution in an optimized way by combining better information searches through semantic web and better workflow provisioning using multi agent systems. | cs.MA | cs | International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
A Novel Service Oriented Model for Query
Identification and Solution Development using
Semantic Web and Multi Agent System
Muneendra Ojha
Indian Institute of Information Technology - Allahabad
Deoghat Jhalwa
Allahabad 211012
ABSTRACT
In this paper, we propose to develop service model architecture by
merging multi-agentsystems and semantic web technology. The
proposed architecture works in two stages namely, Query
Identification and Solution Development. A person referred to as
customer will submit the problem details or requirements which
will be referred to as a query. Anyone who can provide a service
will need to register with the registrar module of the architecture.
Services can be anything ranging from expert consultancy in the
field of agriculture to academic research, from selling products to
manufacturing goods, from medical help to legal issues or even
providing logistics. Query submitted by customer is first parsed
and then iteratively understood with the help of domain experts
and the customer to get a precise set of properties. Query thus
identified will be solved again with the help of intelligent agent
systems which will search the semantic web for all those who can
find or provide a solution. A workable solution workflow is
created and then depending on the requirements, using the
techniques of negotiation or auctioning, solution is implemented
to complete the service for customer. This part is termed as
solution development. In this service oriented architecture, we
first try to analyze the complex set of user requirements then try to
provide best possible solution in an optimized way by combining
better information searches through semantic web and better
workflow provisioning using multi agent systems.
Keywords
Semantic Web, Multi-Agent System, Service Oriented
Architecture
1. INTRODUCTION
Long gone are the days of dumb terminals running set algorithms
and producing results which otherwise if done manually would
have taken much more time. Computer scientists are aiming for
intelligent programs. The ultimate target of artificial intelligenceis
to build programs that are capable of performing like or better
than humans. Since this cannot be achieved in a day, researchers
target towards finding solutions which give better result than the
already existing ones. Acting like humans is an extremely
complex task so AI focuses more on solving smaller problems
which when clubbed together creates the platform for solving or
implementing a real life scenario. Agent system is one such
example where small programs called agents collectively try to
achieve the best possible solution of a complex problem. Ideology
of agent based system revolves around creating complex systems
as a composition of simpler, heterogeneous, distributed
components wherein intelligence comes with ability to observe
and learn from environment. A vast domain of problems is being
addressed using multi agent research and technology [1], [2], [3],
[4], [5].
Although agent based computing has been around for some time
but it is only in the last decade that it has found interest of people
beyond research community. Typical software architectures
containdynamically interacting modules, each with their own
thread and engaging in complex coordinationprotocols [6]. Agent
based computing presents a novel software engineering paradigm
emerging from the amalgamation of object oriented distributed
computing and artificial intelligence [7]. In this paper, we
advocate an agent-oriented paradigm, conceptualizing the analysis
and design of an agent-based system for providing web oriented
services. The remainder of this paper is structured as follows. In
the second section, we provide motivation for this architecture. In
the third section we provide elaborate view of the intelligent
agents, how they perceive, interact and perform. We also
introduce the idea of manual agents and how do autonomous
agents interact with them to optimize the perception. In the fourth
section, we provide the agent oriented model for our proposed
architecture and describe the agent oriented analysis and design.
We also compare it against traditional object oriented design
methodologies. Finally, in the last section, we summarize our
results by providing an insight into the future challenges.
2. MOTIVATION
Most of the real world applications are extremely complex than
they were few years hence. They contain components which are
dynamically interacting with environment through their own
process threads. Most software engineering paradigms are unable
to provide structure that could handle such complexities. It is hard
to identify a general framework that allows the management of
service in a standard platform independent way. So we propose a
middleware architecture that functions as an intermediary between
consumer and service provider. Although the full interoperability
between various middleware platforms is still not achieved, this
architecture serves as a valid starting point to bring distributed
services together. In order to support services within this service
oriented architecture, a lightweight framework is needed that
realizes
interoperability
between
different middleware
technologies. Both the Model Driven Architecture (MDA) [8] and
12
Service Oriented Architecture (SOA) [9] approaches serve to
address the same issue [10]. Let us take an example where a
patient needs to see a doctor. Patient is not sure of which
specialist she should visit or where she could find a good doctor.
Patient uploads the ailment related information over the portal and
requests for a good doctor nearby. System parses the information
and gathers knowledge out of it. Based on this knowledge, it
contacts suitable agents called domain experts, whether
autonomous or manual and tries to clearly identify the disease.
Once it is identified, it again tries to find out the solution, which
in this case would be to find the suitable doctor in a location
assumed to be in the patient’s vicinity. Then it would contact the
doctor, brief him with patient’s problem, take an appointment and
arrange for a ride if needed by the patient at the set date.
Meanwhile,the patient would be able to keep track of the whole
process. Patient can also alter the input details as and when
required so as to keep system updated about her condition. Idea of
the proposed framework is to bring together the vast amount of
information or knowledge spread over Internet in a meaningful
way so as to serve the specific needs of a consumer. So in our
application, SOA better serves the need as it starts from an
approach to distributed computing treating software resource as
service available over the Internet.
Another example can be of a business person willing to setup a
vacuum cleaner factory in some country where it is not a
household necessity. Aim is to build a system intelligent as well
as powerful enough to understand the requirements of setting up a
production unit, technical, financial, social implications, legalities
involved, infrastructure details and any other modalities necessary
for the queried task. No one can argue against the fact that every
related knowledge is available over the Internet and that too free
of cost. Question is “how does an agent understand the
information and derive the relevant knowledge?” Here lies the
actual potential of merging agent oriented approach and Web
Service Architecture (WSA) [11][12] in the context of Semantic
Web. The Semantic Web provides a common framework that
allows data to be shared and reused across application, enterprise,
and community boundaries [13]. It enables knowledge sharing
among heterogeneous, distributed and dynamic sources. Agents
are particularly suitable for performing these kinds of tasks [14].
In the following section, we describe our service architecture and
how semantic web enhances its performance. Then , we impress
upon the reasons that make multi-agent systems a strikingly
appropriate and pertinent solution for the design and development
of a service oriented system over the web. In this context , we
focus upon the role of Agent Oriented Software Engineering
(AOSE) in the Web Service Architecture.
2.1 Semantic Web and Web Service
Architecture
Semantic Web was conceptualized by the inventor of World Wide
Web with a vision to exploit the full potential of data available
over Internet. As per Lee:
“If HTML and the Web made all the online documents look like
one huge book, RDF, schema, and inference languages will make
all the data in the world look like one huge database”[16]
Semantic web describes the relation between things on website as
well as their properties. Instead of visualizing Internet as a web of
pages, it builds a web of data. It starts with defining Uniform
Resource Identifier (URI) which is a string of characters used to
identify name of a resource on Internet. Resource can be anything
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
ranging from books, cars, people, web pages, digital images, or
conceptual things like colors, subject or a metadata. URIs act as
anchor for merging data. In 1997, Metadata vocabularies were
defined and used to make statements. A Resource Description
Framework (RDF) was prepared. Resource can be anything over
Internet identifiable with a URI, Description is the statements
about the resource and Framework was a common model for
statements using variety of vocabularies. An RDF triple (s, p, o) is
defined as “Subject”, “Predicate” and “Object” respectively.
Conceptually, “p” connects “s” and “o”. RDF is a general model
for such triples with machine readable formats like RDF/XML.
RDF is useful but is insufficient in solving all possible
requirements. To counter the limiting possibilities, a new concept
of ontology was introduced. Ontologies define the concepts and
relationships
for describing and
representing an area of
knowledge. A new language Web Ontology Language abbreviated
as “OWL” was formed. OWL is an extra layer of explanation
relying on RDF schema. There are a number of application and
web services using RDF schema and OWL to enhance searches as
well as cater the need of users. A few examples can be listed as
Sun’s White Paper and System Handbook collections, Nokia’s
S60 support portal, Harper’s Online Magazine, Oracle’s virtual
pressroom, Opera’s community site, Dow Jones’ Synaptica [17].
3. THE PROPOSED SYSTEM
Objective is to act as a mediator between human agents who need
some service and human or autonomous agents which provide
such services. Before we start with the formal description of the
model, let us first define few terms that we will use to provide the
comprehensive picture.
Agent: An entity in the system whether human or
autonomous, capable of performing some functionality
will be termed as agent. It shows a fundamental unit of
behavior based on its perception in the environment. It
is intelligent and has a certain level of autonomy to take
decisions on its own. It is social so it can interact with
other agents of the system.
Activities: The functionalities, these agents can perform
will be called as activities. Every agent has some
capability based on which it will perform its intended
functionality. Based on the capability of an agent,
activities can be categorized complex or atomic
Complex activity: A complex activity is one which
cannot be performed by a single agent
the
in
environment. Every complex activity must be dissolved
in smaller activities.
Atomic activity: If we can find even one agent in the
system which is capable enough to carry out complete
activity alone, then that task or activity is considered
atomic. So it does not depend upon the physical view of
a task but ability of an agent to define atomicity of the
task.
Environment: Environment in our system is visualized
as set of tasks assigned to agents. For example agents
involved in the process of request parsing, creating
initial semantics, searching and iteratively interacting
with domain experts belong to one environment. Then
those belonging to negotiation and service provisioning
belong to other environment.
13
Service Provider: A service provider can be any agent
or entity delivering some service to the consumer.
Services may be free or may be charged decided by the
service provider itself.
Output
Problem
Metadata,Seman
ticsDescription
like a central authority
is
Registrar: Registrar
responsible
It has static
for managing agents.
knowledge of the agents, their capabilities, domains,
location and any other information that is relevant to
the services agents provide. It gives a partially
centralized view to the architecture controlling limit of
distributed functionality.
Domain Expert: Domain experts are the members of a
vertical application [18] group defined in semantic web
registered with our system with the registrar module.
They also function as agents in the system. If a group or
organization is registered then a representative agent
interacts on their behalf with the system.
Consumer: In this architecture we define a consumer to
be a human who needs some kind of help. Consumer
submits task request in form of a query. It is assumed
that a consumer needs some kind of service from the
system. Depending on the service, consumer may be
charged by the service provider. Consumer has the
liberty to deny any service before the service has
already started.
3.1 Cognition and rectification of user request
Customer requests do not generally come in the easiest of
machine readable formats. Customer request are handled by the
communicator agent, which passes it onto the query parser
module. Based on the preexisting knowledge base, context
reasoning engine, it tries to parse the query for useful information
and forms the primitive level semantic of it. An XML file
depicting the same is created and sent back to user in a human
readable form through communicator agent as depicted in Figure
1.
Information Servers
Exchange Servers,
Yellow Pages,
Discussion Groups,
Blogs
Online
Databases
Semantic Web and
Web Services
RIF, OWL, RDF
Communicator
Agent
Query Parser
Knowledge Base
Context Reasoning
Engine
Search Engine
Query Moderator
Consumer
Figure 1. Accepting consumer request and parsing iteratively
for identification of the service requested.
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
Sometimes usersthemselvesare not aware of what is needed in a
particular problem. User uploads the problem description and
hopes to get a solution for it. In such cases, agents cannot rely
upon the explanation provided by consumer. What if information
is only half truth? System builds an initial query semantic report
and sends it to the experts groups of the domain identified in
initial report. The initial or primitive understanding is now sent to
the domain experts registered in the system through the Registrar
module. The domain expert or vertical groups’ experts (Figure 2)
are manual agents who have better perceptive capability than an
autonomous programming agent. Domain experts provide their
views on the understanding of problem. They can require more
data or different variety of data for better understanding.
Sometime they can feel that the problem description does not
exactly match their expertise domain so they can refer different
agents or domain in the problem description. Collectively this
works towards obtaining a better cognition of the problem or
Query/ Service
Request
Query Parser
Semantic Web
Initial Query
Semantic
Compiler
Agent
Search Agent
Services Registrar
Vertical
Applications
Domain Experts
Figure 2.Entities involved in the identification of a consumer
query
request submitted by the customer.
If details provided are not sufficient, consumer can be asked to
provide more detail or data. Thus original query is rectified in way
both consumer and agent systems obtain a better comprehension.
This process is an iterative one which repeats itself as long as
either both consumer and domain experts are able to uniquely
identify the query or both system and experts decide that there
cannot be a solution for the query. In either case, system notifies
consumer about the outcome. If it is uniquely identified, system
proceeds to the solution development phase, otherwise stops the
search.
When agent system develops a mutually agreed upon description
of consumer request, it produces a number of descriptor
documents. These
include Problem Description Document,
Vocabulary or Ontology of the request along with any document
which is relevant to identify the task uniquely and precisely.
These documents help in building inference and developing
solution framework.
3.2 Developing solution workflow
A workflow is defined as a pattern of activities enabled by a
systematic organization of resources, defined roles and mass,
energy and information flows, into a work process that can be
documented andlearned. Workflows are designed to achieve
processing intents of some sort, such as physical transformation,
14
processing. Solution
information
or
provision,
service
development phase target creating a workflow which when
executed provides desired services or solution to the consumer.
The workflow is created in such a manner that autonomous agents
can follow it and carry out activities.
First step in problem solving is to find the set of experts who have
this kind of knowledge. Searching through the web becomes more
precise when we have sufficient information and we can target the
search space as problem domain. This is again done through
Registrar. Note that a separate set of experts are required here as it
is not necessary that the expert involved in detecting and
explaining the problem can also be expert in actually solving it.
For example a pathologist has expertise in interpreting and
diagnosing a tumor but lacks surgical experience required to cure
it. Nevertheless we do not alienate the idea of an intersection set
having agents involved both in identification as well as solution
phases. Once experts are found, they are provided complete
problem description and their opinion is sought in creating the
solution workflow. So system undergoes the similar routine of
searching suitable experts as was done in identification phase,
provides them the complete problem detail and iteratively works
with consumer to fine tune the solution workflow. Customer is
always kept in the loop. Automatically creating an optimized
workflow for any problem is still a daunting task. So first a
tentative workflow is created and presented to experts. Experts
supply their opinion and workflow is bettered.
3.3 Providing services
A lot of research has been done in multi agent systems to make
this part completely free from manual intervention. This section
epitomizes the service oriented architecture in its true form. All
the participants in this module are autonomous agents, i.e. self
interested entities that seek to maximize their private utility [19].
We establish two types of agents here, a consumer (an agent
representative of the human consumer) and a provider (again an
agent representative of an individual or an organization providing
service).
In the Agent Oriented Analysis (AOA), tasks, requirements,
services, providers, capabilities and constraints are analyzed from
an agent’s perspective. Service provisioning consists of following
main steps:
1. Workflow description is studied and types of services
required are identified. For each service type a list of
tasks is created. There may be more than one task
possible in any given service type.
2. A list of all possible provider agents is created and
mapped to the service type. Each task is considered
“Atomic” initially i.e., we assume one agent will be able
to finish it to the end. Either agent completes its
assigned task or is unable to complete it; there is no
concept of partially completed task.
If the system could not find an agent having capability
to complete a listed task then such a task is considered a
“Complex” task and needs to be subdivided in smaller
subtasks. This division continues till every task becomes
atomic.
4. For each task negotiate with provider agents depending
upon the criteria of consumer requirement constraints,
and service provider capabilities. Negotiation can be
viewed as an interaction mechanism that aims to resolve
a conflict of interest between various parties through set
3.
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
rules and protocols [20]. Agents’ strategies are very
important over here as agents can cooperate or
sometime compete to achieve the desired provisioning
of services.
a. Cooperative interactions – agents interact and try
to maximize the sum of all their utilities, the social
welfare
b. Competitive interaction – agents interact to satisfy
their own personal preferences. Such agents are
generally called as selfish or self-interested agents
as they try to maximize their own utility function
in order to gain greater rewards or optimize the
result.
Figure 3 depicts clearly the relation between consumer,
Consumer
hasa
Workflow
has a
Task list
Interdependencies
and Constraints
T1
T2
T3
T4
T5
Provider
provides
Service Types
have
Services
Require
remuneration
Service
Registrar
Figure 3.Relation between workflow of consumer,
listed tasks, constraints& interdependencies, services
required for these tasks and their service provider
task, service types, services and provider.
5.
If time is relatively a lesser constraint, auction and
bidding serves the purpose better as we can draw real
life experience from the examples of tender invited by
an organization.
3.4 Design of the service model
As defined earlier in Section 3, any entity performing a unit of
work is treated as an agent in our model. Every agent has a certain
set of defined capabilities. Agents act perceiving the change in its
environment. When it has certain task to complete, it’s the agent’s
responsibility not only to complete the task but also do it in the
best possible of ways.The actions, an agent undertakes to fulfill its
responsibilities is captured by the idea of a utility function. Utility
functions are useful in describing the tradeoff every agent must
make. For example, agent receives payment for searching and
delivering information and incurs cost of the electricity used as
well as the benefits it missed had it been searching and delivering
information to some other consumer. If we translate all these
payments andcosts into utility function then we can easily study
the tradeoffs among them [21]. Agents interact by sending and
15
receiving messages. When an agent initiates interaction by
sending a message, it becomes requester agent and the agent it
sends message to become a provider agent.
Messages are sent to communicate between agents delivering web
services. When an agent wants to avail some service, it sends
request message to the agent representing this particular type of
service with
input requirement semantics. Provider agent
acknowledges the request only if it is interested in fulfilling it and
explains the terms and conditions of providing services. For
example high priority programs which execute immediately cost
higher than those put into batch processing mode. Thus when
requester agent and provider agent become known to each other,
they understand the input semantics and Web Service Description
(WSD) formally, service is provided. Structure of the message is
also defined in WSD. A message can be a simple HTTP GET
request or it can be a SOAP XML. In our model, every service
provider is registered to the Registrar. Initial communication of
knowing the service provider and requester occur through the
Registrar, but once their representative agents acknowledge each
other, they can negotiate freely without intervention of Registrar.
In our system messages are sent through using Simple Object
Access Protocol (SOAP) exchanging XML documents over the
Internet, using Hypertext TransferProtocol (HTTP) [22].
Web Services registration and discovery is done using Universal
Description, Discovery and Integration (UDDI) directory service.
Businesses use Web Services Description Language (WSDL) to
describe relevant information to potential consumers and advertise
services. In our model any activity that results in some fruitful
physical outcome is termed as service. Starting from the very first
interaction of system with consumer through communicator agent
to parsing request, understanding semantics of the request,
consulting domain experts,
identification of
the problem,
searching and consulting solution experts, creating solution
workflow, finding service providers, negotiating, contracting and
getting services completed, all include service definition. Every
expert involved, whether a human agent giving feedback or
suggestion or an autonomous agent searching the semantic web
for some information provides service. Some services are free for
example accessing freely available information on semantic web,
XML
Serialization
RDF
Schema
Web Ontology
Language (OWL)
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
finding suitable service providers through WSDL description
provided in the directory service, contacting providers through
SOAP messages and negotiating a binding contract etc. Only
those services will be charged which has some copyright to it
and/or provider requires some remuneration. When we define a
service in our system, it should be clear that all services are
provided through the Registrar module, which acts as first level of
central authority over collaborative autonomous agents acting
towards achieving solution.
4. ANALYSIS OF THE MODEL
We started building the system with the most essential of
requirements expected from a multi agent system delivering
service. Initial requirements can be classified as functional
requirements and quality requirements.
4.1 Functional requirements
Autonomy:Agents are autonomous as long as it a utility
function to maximize. There are dominant agents whose
task is just to monitor the progress of task. In case an
agent fails to produce output, dominant agents reassign
the task to some other agent. Though dominant agents
do not interfere in the functionality of an individual
agent they do put some kind of organization. Agents are
free to enter the environment and leave it as required.
Social interaction:As per the inherent definition an
agent here is able to communicate with other agents
through message passing. These messages are XML
files describing requirements, requesting services or
reporting status. Though we have put some restriction in
interaction here that agents of one environment would
not be able to interact with different environment
agents.
Collaboration: Agents must collaborate to sort out an
assigned
their utility
to maximize
in order
task
functions. Collaboration can take cooperative form if i t
maximizes the generic output else it can also take a
competitive form when being selfish results in an
Semantic Web
UDDI
Directory Services
Service
Advertisements
MAS
Input Semantic,
Query Descriptor
MAS
Solution
Workflow
MAS
Services
Registrar
Consumer
Services
Registrar
Services
Registrar
Human Agents
Human Agents
Output of services
either in form of
documentation or
physical change in the
environment
b
i
n
d
i
n
g
a
g
r
e
e
m
e
n
t
N
e
g
o
t
i
a
t
i
o
n
s
a
n
d
s
e
r
v
i
c
e
p
r
o
v
i
d
e
r
o
r
o
r
g
a
n
i
z
a
t
i
o
n
s
A
g
e
n
t
s
r
e
p
r
e
s
e
n
t
i
n
g
i
n
d
i
v
i
d
u
a
l
Domain Experts
Solution Provider
Figure 4. Comprehensive view of the service depicting life stages of consumer request including query cognition resulting in
Input semantics, then reconstruction as a solution workflow, and finally resolution in the f orm of services provided
16
optimized result, e.g. in the case of auction, bidding
agents compete with each other resulting in a better deal
for the consumer.
Distributed functionality: Agents act in a distributed
manner when
tasks are
temporally
independent.
Services like communication, data transfer, searching,
negotiating, management etc.
take place without
intervention of any centralized processing unit.
Reusability:We have used several off-the shelf
components in the overall development recognizing
reusability of agent programs.
4.2 Quality requirements
Ease of use:User interface should be easily understood
and easy to use. User can upload query description in
the form of a text file. Status updates are provided back
to user and provisions are made to constantly involve
user in the loop.
Modifiability: Agents can be modified without the need
to change whole architecture. As long as interaction
mechanism is coherent, internal functionality is a black
box for the environment.
Reliability: As long as there is no network failure,
agents are reliable. Agents are able to communicate as
soon as network re-establishes.
Performance: Although requirement criterion specifies
the system to accomplish tasks in a reasonable amount
of time with optimized results but there are so many
smaller modules with open
research problem
interoperating that no claim can be made at this stage.
Security: Security is managed through the Registrar
module of system as only registered entities can use the
system.
Figure 4, depicts the overall architecture with stages of the life
cycle of a consumer request starting from when request is made to
the point when services in terms of satisfactory solution are
provided back to user. Until a binding contract is made, consumer
is free to abandon request anytime. Provisions must be made to
ensure through the stages that the consumer is still interested in
find a solution for her/her query. All the activities should be
abandoned as soon as customer backs off.
AND
FUTURE
5. CONCLUSION
CHALLANGES
In this paper we describe the architecture combining efficiencies
of Multi Agent System with Semantic Web to provide better
query understanding, search, and resolution
in a modular
approach. We cannot claim to have an optimized solution as there
are several open research problems in the system. We provide
software engineering platform for efficient merger of coherent
technological domains. Multi agent systems research is a very hot
field in terms of providing different kind of services in extremely
coordinated fashion as shown in disaster management domain by
the recently concluded Aladdin project [23]. Problems such as
automatic creation of efficient solution workflow, negotiating and
legally binding service providers, remuneration management,
workflow mitigation,
security, data
integration, building
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
commitment, trust, and belief model are some of the areas where
rapid research is going on. We have set a target in our architecture
focusing smaller modules and trying to refine the functionalities
as well as standardize the agent communication on semantic web.
6. REFERENCES
[1] Biswas, P. K. 2007. Towards an agent-oriented approach to
conceptualization. Applied Soft Computing J. Retrieved from
http://dx.doi.org/10.1016/j.asoc.2006.11.009
[2] Brazier, F. M., Keplicz, B. D., Jennings N. R., &Treur, J.
(1995, June 12-14). Formal specification of multi-agent
systems: A real-world case. Proceedings of
the 1st
International Conference on Multiagent Systems (ICMAS
’95), San Francisco.
[3] Brazier, F. M., Dunin-Keplicz, B. M., Jennings, N. R.,
&Treur, J. (1997). DESIRE: Modeling multi-agent systems
in a compositional formal framework. International Journal
of Cooperative Information Systems, 6(1), 67-94.
[4] Jennings, N. R., Faratin,P.,Lomuscio,A. R., Parsons, S.
Sierra, C. and Wooldridge, M. 2001. Automated negotiation:
prospects, methods and challenges. International Journal
ofGroup Decision and Negotiation, 10(2):199-215.
[5] Lind, J. (2000). General concepts of agents and multi-agent
systems. Retrieved from www.agentlab.de
[6] Ciancarini, P., Tolksdorf, R.and Vitali, F. 1999. The World
Wide Web as a place for agents. In M. Wooldridge and M.
Veloso, Ed. Artificial Intelligence Today, Recent Trends and
Developments, LNAI 1600, Springer, Berlin, 175-194.
[7] Odell, J., & Burkhart, R. (1998). Beyond objects: Unleashing
the power of adaptive agents. Tutorial presented at OOPSLA,
Vancouver, B.C.
[8] F. Leymann and D. Roller. Production Workflow: Concepts
and Techniques. Prentice-Hall PTR, Englewood Cliffs, NJ,
1998.
[9] J. Bloomberg. Principles of SOA. Application Development
Trend Magazine, 10(3):22-26, March 2003.
[10] Bocchi,L.,Ciancarini, P., Moretti, R. and Presutti, V. 2008.
On the impact of AOSE in service oriented computing. In
Emerging Methods, Technologies, and Process management
in Software Engineering, A. D. Lucia, F. Ferrucci, G. Tortora
and M. TucciEd. Wiley Interscience, New Jersey, 71-83
[11] D. Booth, H. Haas, F. McCabe, E. Newcomer, M. Champion,
C. Ferris D. Orchard. Web Services Architecture. Technical
Report, World Wide Web Consortium (W3C) 2004. URI –
http://www.w3.org/TR/ws-arch/
[12] Jennings, N. R. (2001). An agent-based approach for
complex software systems. Communications of the ACM,
35-41.
[13] Semantic Web Initiative HomePage (http://www.w3.org/
2001/sw/)
[14] P. Ciancarini and M. Wooldridge, editors. Agent-Oriented
Software Engineering, number 1957 in LNAI, 2001.
17
[15] Tim Berners-Lee, 1999, Weaving the Web: The Original
Design and Ultimate Destiny of the World Wide Web ,
Harper Paperbacks; 1 edition (November 7, 2000)
[16] Herman, I. 2008, An Introduction to Semantic Web,
(Tutorial), 17th International World Wide Web Conference
Beijing, China, 2008
[17] Workflow
Wikipedia.
definition,
http://en.wikipedia.org/wiki/Workflow
URI:
[18] Vertical
URI=
Applications.
http://www.w3.org/standards/semanticweb/applications
[19] Jennings, N. R. 2000. On Agent-Based Software
Engineering. Artificial Intelligence,117(2):277–296.
International Journal of Computer Applications (0975 – 8887)
Volume 35– No.9, December 2011
[20] Mas-Colell, A., Whinston, M. D.and Green, J. R. 1995.
Microeconomic Theory. Oxford University Press, New York.
[21] Vidal, J. M., 2010, Fundamentals ofMultiagent Systems,
URI:
http://multiagent.com/2010/02/multiagent-systems-
textbook.html
[22] Mitra, N. 2003. SOAP Version 1.2 Part 0: Primer. Technical
report, W3C. URI: http://www.w3.org/TR/soap12-part0/
[23] Jennings, N. R., 2010. ALADDIN End of Year Report (Final
Project Report). Southampton, UK: University of
Southampton. URI:
http://www.aladdinproject.org/wp-
content/uploads/2011/02/finalreport.pdf
18
|
1701.01289 | 1 | 1701 | 2017-01-05T12:02:34 | Applying DCOP to User Association Problem in Heterogeneous Networks with Markov Chain Based Algorithm | [
"cs.MA"
] | Multi-agent systems (MAS) is able to characterize the behavior of individual agent and the interaction between agents. Thus, it motivates us to leverage the distributed constraint optimization problem (DCOP), a framework of modeling MAS, to solve the user association problem in heterogeneous networks (HetNets). Two issues we have to consider when we take DCOP into the application of HetNet including: (i) How to set up an effective model by DCOP taking account of the negtive impact of the increment of users on the modeling process (ii) Which kind of algorithms is more suitable to balance the time consumption and the quality of soltuion. Aiming to overcome these issues, we firstly come up with an ECAV-$\eta$ (Each Connection As Variable) model in which a parameter $\eta$ with an adequate assignment ($\eta=3$ in this paper) is able to control the scale of the model. After that, a Markov chain (MC) based algorithm is proposed on the basis of log-sum-exp function. Experimental results show that the solution obtained by DCOP framework is better than the one obtained by the Max-SINR algorithm. Comparing with the Lagrange dual decomposition based method (LDD), the solution performance has been improved since there is no need to transform original problem into a satisfied one. In addition, it is also apparent that the DCOP based method has better robustness than LDD when the number of users increases but the available resource at base stations are limited. | cs.MA | cs |
Applying DCOP to User Association Problem in
Heterogeneous Networks with Markov Chain Based
Algorithm
Peibo Duan∗, Guoqiang Mao∗†‡§, Changsheng Zhang¶ and Bin Zhang¶
∗School of Computing and Communications
University of Technology, Sydney, Australia
†Data61 Australia
§School of Electronic Information Communications, Huazhong Unriversity of Science Technology, Wuhan, China
§Beijing Unriversity of Posts and Telecommunications, Beijing, China
¶School of Computer Application
Northeastern University
Bin Zhang is the correspondence author
Abstract-Multi-agent systems (MAS) is able to characterize
the behavior of individual agent and the interaction between
agents. Thus, it motivates us to leverage the distributed constraint
optimization problem (DCOP), a framework of modeling MAS,
to solve the user association problem in heterogeneous networks
(HetNets). Two issues we have to consider when we take DCOP
into the application of HetNet including: (i) How to set up an
effective model by DCOP taking account of the negtive impact of
the increment of users on the modeling process (ii) Which kind of
algorithms is more suitable to balance the time consumption and
the quality of soltuion. Aiming to overcome these issues, we firstly
come up with an ECAV-η (Each Connection As Variable) model
in which a parameter η with an adequate assignment (η = 3 in
this paper) is able to control the scale of the model. After that, a
Markov chain (MC) based algorithm is proposed on the basis of
log-sum-exp function. Experimental results show that the solution
obtained by DCOP framework is better than the one obtained
by the Max-SINR algorithm. Comparing with the Lagrange dual
decomposition based method (LDD), the solution performance
has been improved since there is no need to transform original
problem into a satisfied one. In addition, it is also apparent that
the DCOP based method has better robustness than LDD when
the number of users increases but the available resource at base
stations are limited.
I. INTRODUCTION
of agent and the cooperation between agents in Multi-agent
System (MAS), a framework, named distributed constraint
optimization problem (DCOP) in terms of constraints that
are known and enforced by distinct agents comes into being
with it. In last decade, the research effort of DCOP has been
dedicated on the following three directions: 1) the development
of DCOP algorithms which are able to better balance the
computational complexity and the accuracy of solution, such
as large neighborhood search method [1]; Markov Chain
Monte Carlo sampling method [2] [3] and distributed junction
tree based method [4] 2) the extension of classical DCOP
model in order to make it more flexible and effective for
practical application, such as expected regret DCOP model [5],
multi-variable agent decomposition model [6] and dynamic
DCOP model [7] 3) the application of DCOP in modeling
environmental systems, such as sensor networks [8], [9], [10],
[11], disaster evacuation [12], traffic control [13], [14] and
resource allocation [15], [16], [3]. In this paper, we take
more attention to the application of DCOP. More precisely,
we leverage DCOP to solve user association problem in the
downlink of multi-tier heterogeneous networks with the aim
to assign mobile users to different base stations in different
tiers while satisfying the QoS constraint on the rate required
by each user.
is generally regarded as a resource allocation problem [17],
[18], [19], [20] in which the resource is defined by the resource
blocks (RBs). In this case, the more RBs allocated to a user, the
larger rate achieved by the user. The methods to solve the user
association problem are divided into centralized controlled and
distributed controlled. With regard to the centralized way, a
central entity is set up to collect information, and then used to
decide which particular BS is to serve which user according
to the collected information. A classical representation of
centralized method is Max-SINR [21]. Distributed controlled
methods attract considerable attention in last decade since they
do not require a central entity and allow BSs and users to make
autonomous user association decisions by themselves through
the interaction between BSs and users. Among all available
methods, the methods based on Lagrange dual decomposation
(LDD) [20] and game theory [22] have better performance.
Hamidreza and Vijay [20] put forward a unified distributed
algorithm for cell association followed by RBs distribution in
a k-tier heterogeneous network. With aid of LDD algorithm,
the users and BSs make their respective decisions based on
local information and a global QoS, expressed in terms of
minimum achievable long-term rate, is achieved. However, the
constraint relaxation and the backtrack in almost each iteration
are needed to avoid overload at the BSs. In addition, as we will
show later, the number of out-of-service users will increase
since a user always selects a best-rate thereby taking up a
large number of RBs and leaving less for others. Nguyen and
Bao [22] proposed a game theory based method in which the
users are modeled as players who participate in the game
of acquiring resources. The best solution is the one which
can satisfy Nash equilibrium (NE). Loosely speaking, such
solution is only a local optima. In addition, it is difficult to
guarantee the quality of the solution.
there is no research of modeling user association problem as
MAS. However, some similar works have been done focusing
on solving the resource management problem in the field
of wireless networks or cognitive radio network [23], [24],
[25] by DCOP framework. These methods can not be directly
applied to user association problem mainly due to the scale of
the models for these practical applications is relatively small.
For instance, Monteiro [24] formalized the channel allocation
in a wireless network as a DCOP with no more than 10 agents
considered in the simulation parts. However, the amount of
users and resource included in a HetNet is always hundreds
and thousands. In this case, a good modeling process along
with a suitable DCOP algorithm is necessary. According to
the in-depth analysis above, it motivates us to explore a good
way to solve user association problem by DCOP. The main
contributions of this paper are as follows:
• An ECAV (Each Connection As Variable) model is pro-
posed for modeling user association problem using DCOP
framework. In addition, we introduce a parameter η with
which we can control the scale (the number of variables
and constriants) of the ECAV model.
• A DCOP algorithm based on Markov chain (MC) is
proposed which is able to balance the time consumption
and the quality of the solution.
• The experiments are conducted which show that
the
results obtained by the proposed algorithm have supe-
rior accuracy compared with the Max-SINR algorithm.
Moreover, it has better robustness than the LDD based
algorithm when the number of users increases but the
available resource at base stations are limited.
The rest of this paper is organized as follows. In Section
II, the definition of DCOP and the system model of user
association problem along with its mixed integer programming
formulation are briefly introduced. In Section III, we illustrate
the ECAV-η model. After that, a MC based algorithm is
designed in section IV. We explore the performance of the
DCOP framework by comparing with the Max-SINR and
LDD methods in Section V. Finally, Section VI draws the
conclusion.
II. PRELIMINARY
This section expounds the DCOP framework and system
model of user association problem along with its mixed integer
programming formulation.
A. DCOP
The definitions of DCOP have a little difference in differ-
ent literatures [26], [4], [5] 1. In this paper, we formalize
the DCOP as a four tuples model < A,V,D,C > where
A = {a1, a2, ..., aA} consists of a set of agents, V =
{v1, v2, ..., vn} is the set of variables in which each variable
vi ∈ V only belongs to an agent a ∈ A. Each variable has
a finite and discrete domain where each value represents a
possible state of the variable. All the domains of different
variables consist of a domain set D = {d1, d2, ..., dn}, where
di is the domain of vi. A constraint c ∈ C = {c1, c2, ..., cC}
is defined as a mapping from the assignments of m variables
to a positive real value:
R(c) : di1 × di2 × ··· dim → R+
(1)
The purpose of a DCOP is to find a set of assignments of
all the variables, denoted as X ∗, which maximize the utility,
namely the sum of all constraint rewards:
(cid:88)
C
argmax
X ∗
R(c)
(2)
B. System Model of User Association Problem
Consider a k-tier HetNet where all the BSs in the same tier
have the same configurations. For example, a two-tier network
including a macro BS (B1) and a femto BS, (B2), is shown in
Fig.1. The set of all BSs is denoted as B = {B1,B2, ...,BNB}
where NB is the total number of BSs. All the BSs in the
kth tier transmit with the same power Pk. The total number
of users is denoted by NU and the set of all users is U =
{U1,U2, ...,UNU}.
With OFDMA technology in LTE-Advanced networks, the
resource, time-frequency, is divided into blocks where each
block is defined as a resource block (RB) including a certain
time duration and certain bandwidth [17]. In this paper, the
resource configured at each BS is in the format of RB so
that its available RBs are decided by the bandwidth and the
scheduling interval duration allocated to that BS. We assume
the BSs in the HetNet share the total bandwidth such that both
intra- and inter-tier interference exist when the BSs allocate
RBs to the users instantaneously.
Assuming the channel state information is available at the
BSs, the SINR experienced by user Uj, served by Bi in the
kth tier is given by
(cid:80)Bl∈B/{Bi} Pkgij + BN0
Pkgij
SINRij =
(3)
In (3), gij is the channel power gain between Uj and Bi,
B/{Bi} represents all
the BSs in B except Bi, B is the
bandwidth and N0 is noise power spectral density. The channel
power gain includes the effect of both path loss and fading.
Path loss is assumed to be static and its effect is captured in
1[26] formalized the DCOP as a three tuples model. [4] adopted a four-
tuples model while [5] used a five tuples model
(a) A simple instance of HetNet
(b) ECAV model
Fig. 1. A simple instance of HetNet with ECAV modeling
commonly used formulation as follows
(cid:88)
(cid:88)
maximize F =
the average value of the channel power gain, while the fading
is assumed to follow the exponential distribution.
From the above, the efficiency of user Uj powered by BS
Bi, denoted as eij, is calculated as
eij = log2(1 + SINRij)
(4)
Given the bandwidth B, time duration T and the scheduling
interval Γ configured at each RB, we attain the unit rate at Uj
upon one RB as follows
uij =
BT eij
Γ
(5)
On the basis of formula (5), the rate received at Uj with
nij RBs provided by Bi in the kth tier is
rij = nijuij
(6)
Associated with each user is a quality-of-service (QoS)
constraint. This is expressed as the minimum total rate the
user should receive. Denoting the rate requiremnt of the jth
user by γj, the minimum number of RBs required to satisfy
γj is calculated by:
min = (cid:100) γj
nij
uij
(cid:101)
(7)
in which (cid:100)·(cid:101) is a ceiling function.
C. Mixed Integer Programming Formulation
The formulations of user association problem by mixed
linear programming are similar in a series of papers (see
the survey literature [19]). in this paper, we present a more
s.t.(cid:88)
(cid:88)
(cid:88)
j∈U
i∈B
xijrij
j∈U
i∈B
xijrij (cid:62) γ,∀Uj ∈ U
xijnij (cid:54) Ni,∀Bi ∈ B
xij (cid:54) 1,∀Uj ∈ U
(8a)
(8b)
(8c)
(8d)
i∈B
nij ∈ {0, 1, ..., Ni},∀Bi ∈ B,∀Uj ∈ U
xij ∈ {0, 1},∀Bi ∈ B,∀Uj ∈ U
(8e)
(8f)
The first constraint ensures the rate QoS requirement from
users. Constraint (8c) indicates that the amount of RBs con-
sumed at the same BS is no more than the total RBs Ni
configurated at the BS. Constraint (8d) guarantees one user
associated with a unique BS. Constraint (8e) guarantees the
number of RBs a BS allocates to a user falls within the range
from zero and Ni. The last constraint (8f) guarantees the
connection between a user and a BS has two states denoted
by a binary variable. The objective function (8a) refers to the
sum of rate rather than a function acted on the rate such as
U (xijrij) (e.g. U (x) = log(1 + x)) in some references. Gen-
erally, two phases are needed to gain the solution including:
1) transforming original problem into a satisfied one through
relaxing Constraint (8e) by nij ∈ {0, nij
min}; 2) the left RBs
in each BS will be allocated to users in order to maximize the
objective function.
III. FORMULATION WITH DCOP
In this section, we expound and illustrate the ECAV model
along with its modified version ECAV-η.
A. ECAV Model
Before giving the formulation based on DCOP, we firstly
introduce the definition of candidate BS:
Definition 1. we declare Bi, i ∈ B is a candidate BS of Uj, j ∈
U if the rate at Uj is above the threshold γ with nij
min RBs
provided by Bi. Simultaneously, nij
min should be less than the
total number of RBs (Ni) configurated at Bi.
i = {0, nij
i , denoted by Dj
After confirming the set of candidate BSs of Uj, j ∈ U,
denoted by, CBj, Uj sends messages to its candidate BSs so
that each Bi, i ∈ NB gets knowledge of its possible connected
users. We define each possible connection between Uj and its
candidate BS Bi as a variable, denoted by V j
i . In this case,
all the variables are divided into NB groups according to the
potential connection between users and different BSs. The do-
min, ..., Ni},
main of each variable V j
i = 0 if no RB is allocated to Uj, otherwise,
where V j
V j
(cid:62) nij
min. We define each group as an agent. Thus, an
n-ary constraint exists among n variables (intra-constraint) to
guarantee that there is no overload at Bi. Note that a user may
have more than one candidate BS, there are constraints (inter-
constraints) connecting the variables affiliated to different
agents on account of the assumption that a unique connection
exists between a user and a BS. Generally speaking,
the
utility (objective) function in the DCOP model is the sum
of constraint rewards which reflects the degree of constraint
violations. We define the reward R(c) of inter- and intra-
constraints in the ECAV model as follows. For ∀c ∈ Cinter
i
i
,V j2
∃V j1
Otherwise
(cid:40) −∞,
−∞,
(cid:88)Vji ∈ ψ(c)rij,
R(c) =
0,
For ∀c ∈ Cintra
R(c) =
i ∈ ψ(c), V al(V j1/j2
i
) > 0(9a)
(9b)
(cid:80)
i ∈ψc
V j
otherwise
V al(V j
(10a)
i ) > Ni
(10b)
i ) represents the assignment of V j
In constraint (9a), ψ(c) is the subset of variables connected
by constraint c. V al(V j
i . A
reward (we use −∞ in this paper) is assigned to the constraints
if there at least two variables are non-zero at the same time
(unique connection between a user and a BS). Otherwise, the
reward is equal to zero. In constraint (10a), the reward is −∞
once there is a overload at the BS. Otherwise, the reward is
the sum of the rates achieved at users.
It
is easy to find that a variable in the ECAV model
with non-zero assignment covers constraint (8b) and (8e) in
the mixed integer programming formulation. Moreover, intra
and inter-constraints respectively cover constraint (8c) and
constraint(8d). Therefore, The global optimal solution X ∗
obtained from the ECAV model is consistent with the one
obtained from the mixed integer programming formulation,
denoted as X 2.
To better understand the modeling process, we recall the
instance in Fig.1 where the candidate BSs of U1 and U2 are
2We say X ∗ is consistent with X when the total rate calculated by objective
function (2) and (8a) is equal. This is because there may be no more than
one optimal solution.
i
1 ,V 2
2 ,V 2
2 and V 4
1 and V 3
the same, denoted as {B1,B2}, while the candidate BSs of U3
and U4 are respectively {B1} and {B2}. We assume the total
RBs configurated at B1 and B2 is 8 and 10. For simplicity, we
assume the rate of each user served by one RB provided by B1
is 0.8 bit/s. And 1 bit/s of each user is served by B2. Then, the
ECAV model is shown in Fig.1(b). There are two agents named
A1 and A2. The variables in A1 are V 1
1 where V j
refers to a connection between user Uj and Bi. Similarly, the
variables in A2 are V 1
2 . Assuming the threshold rate
0.8(cid:101) = 4 RBs needed
is 3 bit/s, we can calculate that at least (cid:100) 3
for the users served by B1, thus the domain of each variable
in A1 is {0, 4, ..., 8}. Also, the domain of each variable in
A2 is {0, 3, ..., 10}. The black lines in each agent are two
intra}. The
3-nry intra-constraints, thus Cintra = {C1
red lines connecting two agents are two intra-constraints, thus
Cinter = {C1
inter to illustrate
how the reward of constraint works in different conditions.
intra, the reward is −∞ when all the variables
Considering C1
associated with C1
intra have the same assignment 4. Thus the
total number of RBs consumed by three users is 12 which is
more than 8 RBs configurated at B1. Otherwise, the reward
is 0.8 × 4 × 3 = 9.6 (bit/s) calculated according to (6).
Considering C1
inter, the reward is −∞ when the assignment
of V 1
1 is 4 because it means
U2 will connect with more than one BSs (B1 and B2), which
violates the assumption of unique connection. Otherwise, the
reward is 0 (9b)). If there is no constraint violated, the final
utility calculated by the objective function is the total rate in
the whole HetNet (constraint (10b)).
2 is 3 and the assignment of V 1
inter}. We use C1
intra and C1
intra,C2
inter,C2
B. ECAV-η Model
The scale of an ECAV model, referring to the number of
agents and constraints, is related to the number of users, BSs
and the candidate BSs hold at each user. However, some
candidate BSs of the user can be ignored because these
BSs are able to satisfy the requirement of the user but with
massive RBs consumed. It can be illustrated by the number
of RBs a BS allocate to a user is inversely proportional to the
geographical distance between them. In this way, we introduce
a parameter η with which we limit the number of candidate
BSs of each user is no more than η. The following algorithms
present the selection of top η candidate BSs (denoted by CB)
and the modeling process of ECAV-η.
Algorithm 1 is the pseudo code for determining CB. It
is executed by each user distributely. More precisely, a user
estimates its total candidate BSs CB by the procedure from
line 5 to 9. Based on 4 to 7, the candidate BSs of a user
is ordered according to the unit number of RBs consumed at
such user served by different BSs (from line 22 to 28). The
time consumption of Algorithm 1 mainly consists of two parts.
One is the dermination of CB with time complexity O(NB).
The other is the ordering operation with time complexity
O(NB2). As a result, the total time expended of Algorithm 1
is O(NB + NB2). With CB, we present the pseudo code in
relation to the building of ECAV-η model.
As for Algorithm 2, it firstly sets up the agents distributely
Algorithm 1 CBj of user Uj, j ∈ U based on η
Input: The information of HetNet (B, U, γ, η)
Output: The set of candidate BS CB based on η
Initialize: CBj ← φ, CBj ← φ
procedure GETALLCANDIDATEBS
for i ∈ B, j ∈ U do
if rij ≥ γ then
(cid:83){Bi}
CBj
end if
end for
end procedure
procedure GETPARTIALCANDIDATEBS
BubbleSort (CBj)
if CBj > η then
j ← CBn
j
for n from 1 to η do
CBn
end for
CBj ← CBj
else
end if
end procedure
procedure BUBBLESORT(CBj)
for m from 1 to CBj do
for n from CBj to m + 1 do
j ) > SIN R(CBn−1
j and CBn−1
if SIN R(CBn
exchane CBn
j
j
) then
end if
end for
end for
end procedure
(cid:46) sorting by SINR
(cid:46) get η candidate BSs
Algorithm 2 ECAV-η
Initialize:
{A,V,D,R} ← φ
R ← φ
procedure SETECAV
for i ∈ B do
end for
for ∀j ∈ U, m ∈ CBj do
A(cid:83){Ai}
V(cid:83){V j
D(cid:83){Dj
C(cid:83){Cj
R(cid:83){R(Cj
C(cid:83){Ci
R(cid:83){R(Ci
i }
i ← {0, nij
Dj
i}
inter}
inter)}
intra}
intra)}
min}
end for
for i ∈ A do
end for
end procedure
(cid:46) elements in DCOP model
(cid:46) utility upon each constraint
(cid:46) based on (9b), (9a)
(cid:46) based on (10a), (10b)
(line 6). It
takes O(1). After that, each user determines
variables, domains as well as inter-constraints from line 8 to
14. This is also carried out in parallel with O(1). Finally, the
intra-constraints are constructed by each agent with O(1) (line
16 to 17). The total time complexity is O(3).
IV. MARKOV CHAIN BASED ALGORITHM
DCOP, to some degree, is a combinatorial optimization
problem in which the variables select a set of values to
maximize the objective function without or with the minimum
constraint violation. We use S to denote the set of all possible
combination of assignments of variables. Also, we call each
element s ∈ S as a candidate solution. Considering an ECAV-
η model in which the four tuples are as follows:
i a connection between Uj and Bi}
iDj
min}}
We are able to rewrite the model in the following way:
• A = {A1,A2, ...,ANB}
• V = {V j
i = {0, N ij
• D = {Dj
• C = Cinter ∪ Cintra
(cid:88)
(cid:88)
max
s∈S
s.t.(cid:64)V j1
i∈B,j∈U
,V j2
i
V j
i ∈ψ(c)
V al(V j
i )
i ∈ ψ(c), V al(V j1/j2
rij > N ij
i
min
) > 0
(11a)
(11b)
(11c)
(cid:88)
(cid:88)
After that, a convex log-sum-exp approximation of (11a) can
be made by:
max
s∈S
i∈B,j∈U
V al(V j
i ) ≈ 1
β
log(
exp(β
V al(V j
i ))
i∈B,j∈U
(12)
where β is a positive constant. We then estimate the gap be-
tween log-sum-exp approximation and (11a) by the following
proposition in [27]:
Proposition 2. Given a positive constant β and n nonnegative
values y1, y2, ..., yn, we have
max
i=1,2,...,n
yi ≤ 1
β
log(
exp(βyi))
≤ max
i=1,2,...,n
yi +
1
β
logn
(13)
In addition, the objective function (11a) has the same optimal
value with the following transformation:
(cid:88)
(cid:88)
(cid:88)
s∈S
s∈S
max
ps(cid:62)0
s.t.
ps
i∈B,j∈U
ps = 1
V als(V j
i )
(14)
i∈B,j∈U V als(V j
the reward with a
i )
simplicity, we use gβ =
solution s. For
is
in which (cid:80)
candidate
(cid:88)
s∈S
n(cid:88)
i=1
i )). Hence, on the ba-
1
sis of formulations (12) and (13), the estimation of (11a) can
be employed by evaluating gβ in the following way:
β log((cid:80)
s∈S exp(β(cid:80)
i∈B,j∈U V al(V j
(cid:88)
(cid:88)
(cid:88)
V als(V j
i∈B,j∈U
s∈S
i ) − 1
β
max
ps(cid:62)0
ps
pslogps
(15)
(cid:88)
s∈S
Assuming s∗ and λ∗ are the primal and dual optimal points
with zero duality gap. By solving the Karush-Kuhn-Tucker
(KKT) conditions [27], we can obtain the following equations:
V als(V j
i ) − 1
β
logps∗ − 1
β
+ λ = 0,∀s ∈ S
s.t.
ps = 1
s∈S
(cid:88)
(cid:88)
i∈B,j∈U
ps∗ = 1
s∈S
λ ≥ 0
(16a)
(16b)
(16c)
(17)
Then we can get the solution of ps∗ as follows:
i∈B,j∈U V als(V j
i ))
i∈B,j∈U V als(V j
i ))
exp(β(cid:80)
(cid:80)
s∈S exp(β(cid:80)
ps∗ =
(cid:88)
On the basis of above transformation,
the objective is
to construct a MC with the state space being S and the
stationary distribution being the optimal solution ps∗ when
MC converges. In this way, the assignments of variables will
be time-shared according to ps∗ and the system will stay
in a better or best solution with most of the time. Another
important thing is to design the nonnegative transition rate
qs,s(cid:48) between two states s and s(cid:48). According to [28], a series
of methods are provided which not only guarantee the resulting
MC is irreducible, but also satisfy the balance equation:
psqs,s(cid:48) = ps(cid:48)qs(cid:48),s. In this paper, we use the following method:
qs,s(cid:48) = α[exp(β
i∈B,j∈U
V als(V j
i ))]−1
(18)
The advantage of (18) is that the transition rate is indepen-
dent of the performance of s(cid:48). A distributed algorithm, named
Wait-and-Hp 3 in [28], is used to get the solution after we
transform DCOP into a MC. However, as the existence of
inter- and intra- constriants in, a checking through the way of
message passing is made in order to avoid constraint violation.
V. EXPERIMENTAL EVALUATION
A. Experimental Setting
In this section, we test the performance of the MC based
algorithm with different assginments of η in the ECAV model.
A simulated environment including a three-tiers HetNet cre-
ated within a 1000m × 1000m square is considered. In the
system, there is one macro BS, 5 pico BSs and 10 femto
3to save space, we advise readers to get more details from literature [28]
BSs with their transmission powers respectively 46, 35, and
20 dBm. The macro BS is fixed at the center of the square,
and the other BSs are randomly distributed. The path loss
between the macro (pico) BSs and the users is defined as
L(d) = 34 + 40log10(d), while the pass loss between femto
BSs and users is L(d) = 37 + 30log10(d). The parameter d
represents the Euclidean distance between the BSs and the
users in meters. The noise power refers to the thermal noise
at room temperature with a bandwidth of 180kHz and equals
to -111.45 dBm. One second scheduling interval is considered.
Without special illustration, 200 RBs are configured at macro
BS, as well as 100 and 50 RBs are configured at each pico
and femto BS. In addition, all the results are the mean of 10
instances.
B. Experimental Results
We firstly discuss the impact of different assignments of η
on the performance of ECAV model from the point of view
of the runtime and the quality of solution. More precisely,
we generate different number of users ranging from 20 to
100 with the step interval of 10. The time consumed by the
MC based algorithm is displayed in Fig.2. It is clear to see
that more time is needed when the number of users increases.
Also, the growth of runtime is depended on the value of η.
Specially, there is an explosive growth of runtime when we set
η from four to five. As previously stated, this is caused by more
candidate BSs considered by each user. However, the quality
of the solutions with different values of η is not obviously
improved according the results in Table I. For instance, the
average rate achieved at each user is only improved no more
than 0.1 bit/s when the number of users are 100 with the values
of η are 3 and 5. It is difficult to make a theoretical analysis
of the realationship between η and the quality of the solution.
We leave this research in future works.
From above analysis, we set η = 3 in the following
experiments in order to balance the runtime and performance
of the solution. In addition, we test the performance of the MC
based algorithm comparing with its counterparts Max-SINR
and LDD based algorithms.
THE AVERAGE RATE (BIT/S) ACHIEVED AT EACH USER
TABLE I
Users
50
80
100
η = 1
12.47
8.12
6.11
η = 2
12.82
8.37
6.37
η = 3
13.10
8.55
6.73
η = 4
13.22
8.68
6.81
η = 5
13.59
8.77
6.83
In Fig.3, we check the connection state between 200 users
and BSs in different tiers. A phenomenon we can observe
from the figure is that there are more or less some users out
of service even we use different kinds of algorithms. It is not
only caused by the limited resource configured at each BSs,
but also related to the positions of such kinds of users. They
are located at the edge of the square and hardly served by any
BS in the system. Further, more users are served by macro
BS in Max-SINR algorithm because a larger SINR always
eixsts between the users and macro BS. As a result, the total
non-served users in Mmax-SINR algorithms are more than the
other two if there is no scheme for allocating the left resource.
On the other hand, the number of non-served users in MC are
less than LDD when η = 3 since the user Uj will select a
BS Bi with the maximal QIij in each iteration of the LDD
algorithm. In other words, the users prefer to connect with
a BS which can offer better QoS even when more resources
are consumed. Therefore, some BSs have to spend more RBs
which leads to the resource at these BSs being more easily
used up.
Fig. 2. The runtime of ECAV-η (η = 1 − 5) with different number of users
in the HetNet
Fig. 5. The CDFs of the rate acheived at users
Fig. 3. The connection between users and BSs according to the allocation
scheme obtained through different algorithms
Fig. 6.
station
The total rate against the number of RBs held at the macro base
In Fig.4, we produce a statistic of the number of non-served
users when we change the total number of users configured
in the HetNet. The average number of non-served users for
each algorithm along with the standard deviation is presented
in the figure. Compared with Fig.3, a more clear results imply
that more than 60 (at worst, around 70) non-served users in
the Max-SINR algorithm. The LDD based algorithm comes
the second with approximate 20 users. The best resutls are
obtained by the MC algorithm with no more than 20 users
even the total users in the HetNet is 240.
Fig. 4. Non-served users in the HetNet with a different number of users
In Fig.5, we compare the cumulative distribution function
(CDF) of the rate. The rate of the users seldomly drops below
the threshold (3 bit/s) when we use the distributed algorithms
(LDD and MC based algorithms), while Max-SINR algorithm
is unable to satisfy the rate QoS constraints. Moreover, the
rate CDFs of the MC based algorithm never lie above the
corresponding CDFs obtained by implementing the Max-SINR
algorithm (the gap is between 6%− 20%). Likewise, At worst
5% gap eixts between the MC based algorithm and LDD when
we set η = 3.
At last, another intesest observation is made by configurat-
ing different number of RBs at macro BS (Fig.6). When we
change the number of RBs from 150 to 250 at macro BS, it is
clear to see that the total rate obtained by LDD is not sensitive
to the variation of the resource hold by macro BS. This result
is also related to the solving process in which two phases are
needed when employing a LDD based algorithm. As we have
discuss in the Introduction section, a solution which can satisfy
the basic QoS requirement will be accepted by the LDD based
algorithm. It finally affects the allocation of left resource at
marco BS. As a result, the algorithm easily falls into the local
optima. This problem, to some degree, can be overcome by the
ECAV model since there is only one phase in the model. With
the ECAV model, a constraint satisfied problem is transformed
into a constraint optimizaiton problem. And the advantage
of DCOP is successfully applied into solving user assocation
problem.
VI. CONCLUSION
An important breakthrough in this paper is that we take
the DCOP into the application of HetNet. More preisely, we
propose an ECAV model along with a parameter η to reduce
the number of nodes and constraints in the model. In addition,
a markov basesd algorithm is applied to balance the quality of
solution and the time consumed. From experimental results,
we can draw a conclusion that the quality of the solution
obtained by the ECAV-3 model solved with the MC based
algorithm is better than the centralized algorithm, Max-SINR
and distributed one LDD, especially when the number of users
increases but they are limited to the available RBs. In future
work, we will extend our research to the following two aspects:
like K-opt [29] and ADOPT [30]
for DCOP,
there are already a theoretial analysis on the
completeness of solution. However, it is still a chanllenge
job in most research of DCOP algorithm, like the MC based
algorithm proposed in this paper. Thus, we will explore the
quality of the solution assoicated with different values of η.
In practice, the BSs in small cells (like pico/femto BSs) have
properties of plug-and-play. They are generally deployed in a
home or small business where the environment is dynamic. In
this way, we should design a DCOP model which is fit for the
variations in the environment such as the mobility of users
and different states (active or sleep) of BSs. To this end, a
stochastic DCOP model can be considered like the one in .
In some algorithms,
REFERENCES
[1] F. Fioretto, F. Campeotto, A. Dovier, E. Pontelli, and W. Yeoh, Large
neighborhood search with quality guarantees for distributed constraint
optimization problems," in Proceedings of the 2015 International Con-
ference on Autonomous Agents and Multiagent Systems.
International
Foundation for Autonomous Agents and Multiagent Systems, 2015, pp.
1835–1836.
[2] F. Fioretto, W. Yeoh, and E. Pontelli, A dynamic programming-based
mcmc framework for solving dcops with gpus," in International Confer-
ence on Principles and Practice of Constraint Programming. Springer,
2016, pp. 813–831.
[3] X. Ge, S. Tu, T. Han, Q. Li, and G. Mao, Energy efficiency of small
cell backhaul networks based on gauss–markov mobile models," IET
Networks, vol. 4, no. 2, pp. 158–167, 2015.
[4] Y. Kim and V. R. Lesser, Djao: A communication-constrained dcop
algorithm that combines features of adopt and action-gdl." in AAAI,
2014, pp. 2680–2687.
[5] T. Le, F. Fioretto, W. Yeoh, T. C. Son, and E. Pontelli, Er-dcops: A
framework for distributed constraint optimization with uncertainty in
constraint utilities," in Proceedings of the 2016 International Conference
on Autonomous Agents & Multiagent Systems.
International Foundation
for Autonomous Agents and Multiagent Systems, 2016, pp. 606–614.
[6] F. Fioretto, W. Yeoh, and E. Pontelli, Multi-variable agent decomposition
for dcops," in Thirtieth AAAI Conference on Artificial Intelligence, 2016.
[7] W. Yeoh, P. Varakantham, X. Sun, and S. Koenig, Incremental dcop
search algorithms for solving dynamic dcops," in The 10th International
Conference on Autonomous Agents and Multiagent Systems-Volume
3.
International Foundation for Autonomous Agents and Multiagent
Systems, 2011, pp. 1069–1070.
[8] V. Lesser, C. L. Ortiz Jr, and M. Tambe, Distributed sensor networks:
A multiagent perspective. Springer Science & Business Media, 2012,
vol. 9.
[9] G. Mao and B. D. Anderson, Graph theoretic models and tools for the
analysis of dynamic wireless multihop networks," in 2009 IEEE Wireless
Communications and Networking Conference.
IEEE, 2009, pp. 1–6.
[10] G. Mao, B. D. Anderson, and B. Fidan, Online calibration of path loss
IEEE,
exponent in wireless sensor networks," in IEEE Globecom 2006.
2006, pp. 1–6.
[11] A. A. Kannan, B. Fidan, and G. Mao, Robust distributed sensor network
localization based on analysis of flip ambiguities," in IEEE GLOBECOM
2008-2008 IEEE Global Telecommunications Conference.
IEEE, 2008,
pp. 1–6.
[12] K. Kinoshita, K. Iizuka, and Y. Iizuka, Effective disaster evacuation by
solving the distributed constraint optimization problem," in Advanced
Applied Informatics (IIAIAAI), 2013 IIAI International Conference on.
IEEE, 2013, pp. 399–400.
[13] T. Brys, T. T. Pham, and M. E. Taylor, Distributed learning and multi-
objectivity in traffic light control," Connection Science, vol. 26, no. 1,
pp. 65–83, 2014.
[14] R. Mao and G. Mao, Road traffic density estimation in vehicular
networks," in 2013 IEEE Wireless Communications and Networking
Conference (WCNC).
IEEE, 2013, pp. 4653–4658.
[15] F. Amigoni, A. Castelletti, and M. Giuliani, Modeling the management
of water resources systems using multi-objective dcops," in Proceedings
of the 2015 International Conference on Autonomous Agents and Mul-
tiagent Systems.
International Foundation for Autonomous Agents and
Multiagent Systems, 2015, pp. 821–829.
[16] P. Rust, G. Picard, and F. Ramparany, Using message-passing dcop
algorithms to solve energy-efficient smart environment configuration
problems," in International Joint Conference on Artificial Intelligence,
2016.
[17] N. Guan, Y. Zhou, L. Tian, G. Sun, and J. Shi, QoS guaranteed resource
block allocation algorithm for lte systems," in Proc. Int. Conf. Wireless
and Mob. Comp., Netw. and Commun. (WiMob), Shanghai, China, Oct.
2011, pp. 307–312.
[18] T. K. Vu, Resource allocation in heterogeneous networks," Ph.D. dis-
sertation, University of Ulsan, 2014.
[19] D. Liu, L. Wang, Y. Chen, M. Elkashlan, K.-K. Wong, R. Schober, and
L. Hanzo, User association in 5g networks: A survey and an outlook,"
IEEE Communications Surveys & Tutorials, vol. 18, no. 2, pp. 1018–
1044, 2016.
[20] H. Boostanimehr and V. K. Bhargava, Unified and distributed QoS-
driven cell association algorithms in heterogeneous networks," IEEE
Wireless Commun., vol. 14, no. 3, pp. 1650–1662, 2015.
[21] Q. Ye, B. Rong, Y. Chen, C. Caramanis, and J. G. Andrews, Towards an
optimal user association in heterogeneous cellular networks," in Global
Communications Conference (GLOBECOM), 2012 IEEE.
IEEE, 2012,
pp. 4143–4147.
[22] V. N. Ha and L. B. Le, Distributed base station association and power
control for heterogeneous cellular networks," IEEE Transactions on
Vehicular Technology, vol. 63, no. 1, pp. 282–296, 2014.
[23] T. L. Monteiro, M. E. Pellenz, M. C. Penna, F. Enembreck, R. D.
Souza, and G. Pujolle, Channel allocation algorithms for wlans using
distributed optimization," AEU-International Journal of Electronics and
Communications, vol. 66, no. 6, pp. 480–490, 2012.
[24] T. L. Monteiro, G. Pujolle, M. E. Pellenz, M. C. Penna, and R. D.
Souza, A multi-agent approach to optimal channel assignment in wlans,"
in 2012 IEEE Wireless Communications and Networking Conference
(WCNC).
IEEE, 2012, pp. 2637–2642.
[25] J. Xie, I. Howitt, and A. Raja, Cognitive radio resource management
using multi-agent systems," in IEEE CCNC, 2007.
[28] M. Chen, S. C. Liew, Z. Shao, and C. Kai, Markov approximation for
combinatorial network optimization," IEEE Transactions on Information
Theory, vol. 59, no. 10, pp. 6301–6327, 2013.
[29] H. Katagishi and J. P. Pearce, Kopt: Distributed dcop algorithm for
arbitrary k-optima with monotonically increasing utility," in Ninth DCR
Workshop, 2007.
[30] P. J. Modi, W.-M. Shen, M. Tambe, and M. Yokoo, Adopt: Asynchronous
distributed constraint optimization with quality guarantees," Artificial
Intelligence, vol. 161, no. 1, pp. 149–180, 2005.
[26] M. Vinyals, J. A. Rodriguez-Aguilar, and J. Cerquides, Constructing a
unifying theory of dynamic programming dcop algorithms via the gen-
eralized distributive law," Autonomous Agents and Multi-Agent Systems,
vol. 22, no. 3, pp. 439–464, 2011.
[27] S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge
university press, 2004.
|
1702.07984 | 3 | 1702 | 2018-10-28T03:19:07 | Iterative Local Voting for Collective Decision-making in Continuous Spaces | [
"cs.MA",
"cs.CY",
"cs.GT"
] | Many societal decision problems lie in high-dimensional continuous spaces not amenable to the voting techniques common for their discrete or single-dimensional counterparts. These problems are typically discretized before running an election or decided upon through negotiation by representatives. We propose a algorithm called {\sc Iterative Local Voting} for collective decision-making in this setting. In this algorithm, voters are sequentially sampled and asked to modify a candidate solution within some local neighborhood of its current value, as defined by a ball in some chosen norm, with the size of the ball shrinking at a specified rate.
We first prove the convergence of this algorithm under appropriate choices of neighborhoods to Pareto optimal solutions with desirable fairness properties in certain natural settings: when the voters' utilities can be expressed in terms of some form of distance from their ideal solution, and when these utilities are additively decomposable across dimensions. In many of these cases, we obtain convergence to the societal welfare maximizing solution.
We then describe an experiment in which we test our algorithm for the decision of the U.S. Federal Budget on Mechanical Turk with over 2,000 workers, employing neighborhoods defined by $\mathcal{L}^1, \mathcal{L}^2$ and $\mathcal{L}^\infty$ balls. We make several observations that inform future implementations of such a procedure. | cs.MA | cs |
Iterative Local Voting for Collective Decision-making in
Continuous Spaces
Nikhil Garg
Stanford University
[email protected]
Vijay Kamble
University of Illinois at Chicago
[email protected]
Ashish Goel
Stanford University
[email protected]
David Marn
University of California, Berkeley
[email protected]
Kamesh Munagala
Duke University
[email protected]
October 30, 2018
Abstract
Many societal decision problems lie in high-dimensional continuous spaces not amenable
to the voting techniques common for their discrete or single-dimensional counterparts. These
problems are typically discretized before running an election or decided upon through negotiation
by representatives. We propose a algorithm called Iterative Local Voting for collective
decision-making in this setting. In this algorithm, voters are sequentially sampled and asked to
modify a candidate solution within some local neighborhood of its current value, as defined by
a ball in some chosen norm, with the size of the ball shrinking at a specified rate.
We first prove the convergence of this algorithm under appropriate choices of neighborhoods
to Pareto optimal solutions with desirable fairness properties in certain natural settings: when
the voters' utilities can be expressed in terms of some form of distance from their ideal solution,
and when these utilities are additively decomposable across dimensions. In many of these cases,
we obtain convergence to the societal welfare maximizing solution.
We then describe an experiment in which we test our algorithm for the decision of the U.S.
Federal Budget on Mechanical Turk with over 2,000 workers, employing neighborhoods defined
by L1,L2 and L∞ balls. We make several observations that inform future implementations of
such a procedure.
1
Introduction
Methods and experiments to increase large-scale, direct citizen participation in policy-making
have recently become commonplace as an attempt to revitalize democracy. Computational and
crowdsourcing techniques involving human-algorithm interaction have been a key driver of this
trend [5, 21, 22, 28]. Some of the most important collective decisions, whether in government or in
business, lie in high-dimensional, continuous spaces -- e.g. budgeting, taxation brackets and rates,
collectively bargained wages and benefits, urban planning etc. Direct voting methods originally de-
signed for categorical decisions are typically infeasible for collective decision-making in such spaces.
Supported by NSF grant nos. CCF-1408784, CCF-1637397, CCF-1637418, and IIS-1447554, ONR grant no.
N00014-15-1-2786, ARO grant no. W911NF-14-1-0526, and the NSF Graduate Research Fellowship under grant no.
DGE-114747. This work benefited from many helpful discussions with Oliver Hinder.
1
Although there has been some theoretical progress on designing mechanisms for continuous decision-
making [8, 24, 27], in practice these problems are usually resolved using traditional approaches --
they are either discretized before running an election, or are decided upon through negotiation by
committee, such as in a standard representative democracy [5, 14, 15, 30, 32].
One of the main reasons for the current gap between theory and practice in this domain is the
challenge of designing practically implementable mechanisms. We desire procedures that are simple
enough to explain and use in practice, and that result in justifiable solutions while being robust
to the inevitable deviations from ideal models of user behavior and preferences. To address this
challenge, a social planner must first make practically reasonable assumptions on the nature and
complexity of feedback that can be elicited from people and then design simple algorithms that
operate effectively under these conditions. Further, while robustness to real-world model deviations
may be difficult to prove in theory, it can be checked in practice through experiments.
We first tackle the question of what type of feedback voters can give. In general, for the types
of problems we wish to solve, a voter cannot fully articulate her utility function. Even if voters
in a voting booth had the patience to state their exact utility for a reasonably large number of
points (e.g. how much they liked each candidate solution on a scale from one to five), there is no
reason to believe that they could do so in any consistent manner. On the other hand, we posit that
it is relatively easy for people to choose their favorite amongst a reasonably small set of options,
or articulate how they would like to locally modify a candidate solution to better match their
preferences. Such an assumption is common and is a central motivation in social choice, especially
implicit utilitarian voting [26].
In this paper, we study and experimentally test a type of algorithm for large-scale preference
aggregation that effectively leverages the possibility of asking voters such easy questions. In this
algorithm that we call Iterative Local Voting (ILV), voters are sequentially sampled and are
asked to modify a candidate solution to their favorite point within some local neighborhood, until
a stable solution is obtained (if at all). With a continuum of voters, no one votes more than once.
The algorithm designer has flexibility in deciding how these local neighborhoods are defined -- in
this paper we focus on neighborhoods that are balls in the Lq norm, and in particular on the cases
and ∞ neighborhoods correspond to bounds on the sum of absolute values of the changes, the sum
of the square of the changes, and the maximum change, respectively.)
where q = 1, 2 or ∞. (For M < ∞ dimensional vectors, the Lq norm (cid:107)x(cid:107)q (cid:44) q(cid:112)(cid:80)m xmq. q = 1, 2
More formally, consider a M-dimensional societal decision problem in X ⊂ RM and a population
of voters V, where each voter v ∈ V has bounded utility fv(x) ∈ R,∀ x ∈ X . Then we consider
the class of algorithms described in Algorithm 1. We study the algorithm class under two plausible
models of how voters respond to query (1), which asks for the voter's favorite point in a local region.
• Model A: One possibility is that voters exactly perform the maximization asked of them, re-
sponding with their favorite point in the given Lq norm constraint set. In other words, they return
a point arg maxx∈{s:(cid:107)s−xt−1(cid:107)q≤rt}fvt(x). Note that by definition of this movement, the algorithm
is myopically incentive compatible: if a voter is the last voter and no projections are used, then
truthfully performing this movement is the dominant strategy. In general, the mechanism is not
globally incentive compatible, nor incentive compatible with projections onto the feasible region.
Simple examples of manipulations in both instances exist.
• Model B: On the other hand, voters may not actually search within the constraint set to find
their favorite point inside of it. Rather, a voter v may have an idea about how to best improve the
current point and then move in that direction to the boundary of the given constraint set. This
model leads to a voter moving the current solution in the direction of the gradient of her utility
2
Algorithm 1: Iterative Local Voting (ILV)
Inputs:
termination time T , norm q for local neighborhood.
Output: Solution x.
Initial solution x0 ∈ X , tolerance > 0, an integer N, initial radius r0 > 0,
• For t ≥ 1, sample a voter vt ∈ V at random from the population; set rt = r0/t and elicit
(cid:48)
t = arg maxx∈{s:(cid:107)s−xt−1(cid:107)q≤rt}fvt(x),
x
(1)
t]X , where [·]X is a projection onto space X ; i.e. ask the voter to
and then compute xt = [x(cid:48)
move to her favorite point within the constraints, and find the projection of the reported
point onto X .
• Stop when either t = T , in which case return xT , or when maxl, m∈{t−N,...,t} xl − xm ≤ , in
which case return x = xt.
gt(cid:107)gt(cid:107)q
, for some gt ∈ ∂fvt(xt−1). Note that ∂f (x) denotes the
function, returning a point xt−1 + rt
set of subgradients of a function f at x, i.e. g ∈ ∂f (x) if ∀y, f (y) − f (x) ≥ gT (y − x).
ILV is directly inspired by the stochastic approximation approach to solve optimization problems
[29], especially stochastic gradient descent (SGD) and the stochastic subgradient method (SSGM).
The idea is that if (a) voter preferences are drawn from some probability distribution and (b) the
response of a voter to the query (1) moves the solution approximately in the direction of her utility
gradient, then this procedure almost implements stochastic gradient descent for minimizing negative
expected utility.
The caveat is that although the procedure can potentially obtain the direction of the gradient
of the voter utilities, it cannot in general obtain any information about its magnitude since the
movement norm is chosen by the procedure itself. However, we show that for certain plausible
utility and voter response models, the algorithm does indeed converge to a unique point with
desirable properties, including cases in which it converges to the societal optimum.
Note that with such feedback and without any additional assumptions on voter preferences (e.g.
that voter utilities are normalized to the same scale), no algorithm has any hope of finding a desirable
solution that depends on the cardinal values of voters' utilities, e.g., the social welfare maximizing
solution (the solution that maximizes the sum of agent utilities). This is because an algorithm that
uses only ordinal information about voter preferences is insensitive to any scaling or even monotonic
transformations of those preferences.
1.1 Contributions
This work is a step in extending the vast literature in social choice to continuous spaces, taking into
account the feedback that voters can actually give. Our main theoretical contributions are as follows:
• Convergence for Lp normed utilities: We show that if the agents cost functions can be
expressed as the Lp distance from their ideal solution, and if agents correctly respond to query (1),
then an interesting duality emerges: for p = 1, 2 or ∞, using Lq neighborhoods, where q = ∞, 2
and 1 respectively, results in the algorithm converging to the unique social welfare optimizing
solution. Whether such a result holds for general (p, q), where q is the dual norm to p (i.e.
1/p + 1/q = 1), is an open question. However, we show that such a general result holds if, in
3
response to query (1), the voter instead moves the current solution in the direction of the gradient
of her utility function to the neighborhood boundary.
• Convergence for other utilities: Next, we show convergence to a unique solution in two cases:
(a) when the voter cost can be expressed as a weighted sum of L2 distances over sub-spaces of the
solution space, under L2 neighborhoods -- in which case the solution is also Pareto efficient, and
(b) when the voter utility can be additively decomposed across dimensions, under L∞ neighbor-
hoods -- in which case the algorithm converges to the median of the ideal solutions of the voters
on each dimension.
We then build a platform and run the first large-scale experiment in voting in multi-dimensional
continuous spaces, in a budget allocation setting. We test three variants of ILV: with L1, L2 and
L∞ neighborhoods. Our main findings are as follows:
• We observe that the algorithm with L∞ neighborhoods is the only alternative that satisfies the
first-order concern for real-world deployability: consistent convergence to a unique stable solution.
Both L1 and L2 neighborhoods result in convergence to multiple solutions.
• The consistent convergence under L∞ neighborhoods in experiments strongly suggests the de-
composability of voter utilities for the budgeting problem. Motivated by this observation, we
propose a general class of decomposable utility functions to model user behavior for the budget
allocation setting.
• We make several qualitative observations about user behavior and preferences. For instance, vot-
ers have large indifference regions in their utilities, with potentially larger regions in dimensions
about which they care about less. Further, we show that asking voters for their ideal budget
allocations and how much they care about a given item is fraught with UI biases and should be
carefully designed.
We remark that an additional attractive feature of such a constrained local update algorithm
in a large population setting is that strategic behavior from the voters is less of a concern: even
if a single voter is strategic, her effect on the outcome is negligible. Further, it may be difficult
for a voter, or even a coalition of voters, to strategically vote; one must reason over the possible
future trajectories of the algorithm over the randomness of future voters. One coalition strategy for
L2 and L∞ neighborhoods, voters trade votes on different dimensions with one another; we leave
robustness to such strategies to future work.
The structure of the paper is as follows. After discussing related work in Section 2, we present
convergence results for our algorithm under different settings in Section 3. In Section 4, we introduce
the budget allocation problem and describe our experimental platform. In Section 5, we analyze
the experiment results, and then we conclude the paper in Section 6. The proofs of our results are
in the appendix.
2 Related Work
Our work relates to various strands of literature. We note that a conference version of this work
appeared previously [13]. Furthermore, the term "iterative voting" is also used in other works to
denote unrelated methods [1, 23].
4
Stochastic Gradient Descent As discussed in the introduction, we draw motivation from the
stochastic subgradient method (SSGM), and our main proof technique is mapping our algorithm
to SSGM. Beginning with the original stochastic approximation algorithm by Robbins and Monro
[29], a rich literature surrounds SSGM, for instance see [4, 19, 25, 31].
Iterative local voting A version of our algorithm, with L2 norm neighborhoods, has been pro-
posed independently several times [3, 9, 17] and is referred to as Normalized Gradient Ascent (NGA).
Instead of directly asking voters to perform query (1), the movement ∇fv(xt−1)
would be estimated
(cid:107)∇fv(xt−1)(cid:107)2
through population surveys to try to compute the fixed point where Ev(cid:104) ∇fv(x)
we work with distributions of voters and for strictly concave utility functions, the movement for
each voter is well-defined for all but a measure 0 set. Then, given a bounded density function of
voters, the expectation is well-defined).
(cid:107)∇fv(x)(cid:107)2(cid:105) = 0. (Note that
This fixed point has been called Directional Equilibrium (DE) in the recent literature [9]. The
movement is equivalent to the movement in this work in the case voters respond according to
Model B and with L2 neighborhoods, and we show in Section 3.3 that, in such cases, the algorithm
converges to a Directional Equilibrium when it converges. We further conjecture that even under
voter Model A, if Algorithm 2 converges, the fixed point is a Directional Equilibrium.
Several properties of the fixed point have been studied, starting from [17] to more recently,
it exists under light assumptions, is Pareto efficient, and has important connections
[9] and [3]:
to the Majority Core literature in economics. Showing that an iterative algorithm akin to ours
converges to such a point has been challenging; indeed, except for special cases such as quadratic
utilities fv(x) = −(x − xv)T Ω(x − xv), with society-wide Ω that encodes the relative importance
and relationships between issues [3], convergence is an open question.
Our algorithm differs from NGA in a few crucial directions, even in the case that the movement
is equivalent: by relating our algorithm to SGD, we are able to characterize the step-size behavior
necessary for convergence and show convergence even when each step is made by a single voter,
rather than after an estimate of the societal normalized gradient. One can also characterize the
convergence rate of the algorithm [25]. Furthermore, the literature has referred to the L2 norm
(or "quadratic budget") constraint as "central to their strategic properties" [2]. In this work, this
limitation is relaxed -- the same strategic property, myopic incentive compatibility, holds for the
other norm constraints for their respective cases.
Finally, because we are primarily interested in designing implementable voting mechanisms, we
focus on somewhat different concerns than the directional equilibria literature. However, we believe
that the ideas in this work, especially the connections to the optimization literature, may prove
useful to work on NGA. To the best of our knowledge, no work studies such an algorithm with other
neighborhoods and under ordinal feedback, or implements such an algorithm.
Optimization without gradients Because we are concerned with optimization without access to
voters' utility functions or its gradients, this work seems to be in the same vein as recent literature
on convex optimization without gradients -- such as with comparisons or with pairs of function
evaluations [10 -- 12, 18]. However, in the social choice or human optimization setting, we cannot
estimate each voter's utility functions or gradients exactly rather than up to a scaling term, and yet
we would like to find some point with good societal properties. This limitation prevents the use of
strategies from such works.
5
Jamieson et al. [18], for example, present an optimal coordinate-descent based algorithm to find
the optimum of a function for the case in which noisy comparisons are available on that function;
in our setting, such an algorithm could be used to find the optimal value for each voter, but not
the societal optimum because each voter can independently scale her utility function. Duchi et al.
[10] present a distributed optimization algorithm where each node (voter) has access to its own
subgradients and a few of its neighbors, but in our case each voter can arbitrarily scale her utility
function and thus her subgradients. Similar problems emerge in applying results from the work
of Duchi et al. [11].
In our work, such scaling does not affect the point to which the algorithm
converges.
Participatory Budgeting The experimental setting for this work, and a driving motivation, is
Participatory Budgeting, in which voters are asked to help create a government budget. Partici-
patory budgeting has been among the most successful programs of Crowdsourced Democracy, with
deployments throughout the world allocating hundreds of millions of dollars annually, and studies
have shown its civic engagement benefits [5, 14, 15, 21, 22, 30, 32].
In a typical election, community members propose projects, which are then refined and voted on
by either their representatives or the entire community, through some discrete aggregation scheme.
In no such real-world election, to our knowledge, can the amount of money to allocate to a project
be determined in a continuous space within the voting process, except through negotiation by
representatives.
Goel et al. [15] propose a "Knapsack Voting" mechanism in which each voter is asked to create
a valid budget under the budget constraint; the votes are then aggregated using K-approval ag-
gregation on each dollar in the budget, allowing for fully continuous allocation in the space. This
mechanism is strategy-proof under some voter utility models. In comparison, our mechanism works
in more general spaces and is potentially easier for voters to do.
Implicit Utilitarian Voting With a finite number of candidates, the problem of optimizing some
societal utility function (based on the cardinality of voter utilities) given only ordinal feedback is
well-studied, with the same motivation as in this work: ordinal feedback such as rankings and subset
selections are relatively easy for voters to provide. The focus in such work, referred to as implicit
utilitarian voting, is to minimize the distortion of the output selected by a given voting rule, over
all possible utility functions consistent with the votes, i.e. minimize the worst case error achieved
by the algorithm due to an under-determination of utility functions when only using the provided
inputs [6, 7, 16, 26]. In this work, we show convergence of our algorithm under certain implicit
utility function forms. However, we do not characterize the maximum distortion of the resulting
fixed point (or even the convergence to any fixed point) under any utility functions consistent with
the given feedback, leaving such analysis for future work.
3 Convergence Analysis
In this section, we discuss the convergence properties of ILV under various utility and behavior
models. For the rest of the technical analysis, we make the following assumptions on our model.
C1 The solution space X ⊆ RM is non-empty, bounded, closed, and convex.
C2 Each voter v has a unique ideal solution xv ∈ X .
6
Spatial, (p, q) = (2, 2), (1,∞), or (∞, 1)
Spatial, (p, q) s.t. 1/p + 1/q = 1
Weighted Euclidean
Decomposable
Model A
Model B
Social Opt. (Thm 1)
?
Social Opt. (Thm 2)
Social Opt. (Thm 1)
Medians (Thm 2)
Table 1: Summary of convergence results
C3 The ideal point xv of each voter is drawn independently from a probability distribution with
a bounded and measurable density function hX .
Under this model, for a solution x ∈ X , the societal utility is given by Ev[fv(x)]. and the social
optimal (SO) solution is any x∗ ∈ arg maxx∈X Ev[fv(x)].
"Convergence" of ILV refers to the convergence of the sequence of random variables {xt}t≥1 to
some x ∈ X with probability 1, assuming that the algorithm is allowed to run indefinitely (this
notion of convergence also implies the termination of the algorithm with probability 1).
In the following subsections, we present several classes of utility functions for which the algorithm
converges, summarized in Table 1. We further formalize the relationship to directional equilibria in
Section 3.3.
3.1 Spatial Utilities
Here we consider spatial utility functions, where the utilities of each voters can be expressed in the
form of some kind of spatial distance from their ideal solutions. First, we consider the following
kind of utilities.
Definition 1. Lp normed utilities. The voter utility function is Lp normed if fv(x) = −(cid:107)x −
xv(cid:107)p,∀x ∈ X .
Under such utilities, for p = 1, 2 and ∞, restricting voters to a ball in the dual norm leads to
convergence to the societal optimum.
Theorem 1. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are Lp normed,
and voters respond to query (1) according to either Model A or Model B. Then, ILV with Lq
neighborhoods converges to the societal optimal point w.p. 1 when (p, q) = (2, 2), (1,∞), or (∞, 1).
The proof is contained in the appendix. A sketch of the proof is as follows. For the given
pairs (p, q), we show that, except in certain 'bad' regions, the update rule xt+1 = arg minx[(cid:107)x −
xvt(cid:107)p : (cid:107)x − xt(cid:107)q ≤ rt] is equivalent to the stochastic subgradient method (SSGM) update rule
xt+1 = xt − rtgt, for some gt ∈ ∂Ev[(cid:107)x − xvt(cid:107)p], and that the probability of being in a 'bad' region
decreases fast enough as a function of rt. We then leverage a standard SSGM convergence result
to finish the proof. One natural question is whether the result extends to general dual norms p, q,
where 1/p + 1/q = 1. Unfortunately, the update rule is not equivalent to SSGM in general, and we
leave the convergence to the societal optimum for general (p, q) as an open question.
Further, note that even if each voter could scale their utility function arbitrarily, the algorithm
would converge to the same point.
However, the general result does hold for general dual norms (p, q) if one assumes the alternative
behavior model.
7
Theorem 2. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are Lp normed,
and voters respond to query (1) according to Model B. Then, ILV with Lq neighborhoods converges
to the societal optimal point w.p. 1 for any p > 0 and q > 0 such that 1/p + 1/q = 1.
The proof is contained in the appendix. It uses the following property of Lp normed utilities: the
Lq norm of the gradient of these utilities at any point other than the ideal point is constant. This
fact, along with the voter behavior model, allows the algorithm to implicitly capture the magnitude
of the gradient of the utilities, and thus a direct mapping to SSGM is obtained. Note that the above
result holds even if we assume that a voter moves to her ideal point xv in case it falls within the
neighborhood (since, as explained earlier, the probability of sampling such a voter decreases fast
enough).
Next, we introduce another general class of utility functions, which we call Weighted Euclidean
utilities, for which one can obtain convergence to a unique solution.
Definition 2. Weighted Euclidean utilities. Let the solution space X be decomposable into K
k=1 dim(xk) = M). Suppose
different sub-spaces, so that x = (x1, . . . , xK) for each x ∈ X (where(cid:80)K
that the utility function of the voter v is
fv(x) = − K(cid:88)k=1
wk
v(cid:107)wv(cid:107)2
(cid:107)xk − xk
v(cid:107)2.
where wv is a voter-specific weight vector, then the function is a Weighted Euclidean utility function.
We further assume that wv ∈ W ⊂ RK
+ and xv are independently drawn for each voter v from a
joint probability distribution with a bounded and measurable density function, with W nonempty,
bounded, closed, and convex.
This utility function can be interpreted as follows: the decision-making problem is decomposable
v for each sub-problem
into K sub-problems, and each voter v has an ideal point xk
k, so that the voter's disutility for a solution is the weighted sum of the Euclidean distances to the
ideal points in each sub-problems. Such utility functions may emerge in facility location problems,
for example, where voters have preferences on the locations of multiple facilities on a map. This
utility form is also the one most closely related to the existing literature on Directional Equilibria
and Quadratic Voting, in which preferences are linear. To recover the weighted linear preferences
case, set K = M, with each sub-space of dimension 1. In this case, the following holds:
v and a weight wk
Proposition 1. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are
Weighted Euclidean, and voters correctly respond to query (1) according to either Model A or
Model B. Then, ILV with L2 neighborhoods converges with probability 1 to the societal optimal
point.
The intuition for the result is as follows: as long as the neighborhood does not contain the ideal
point of the sampled voter, the correct response to query (1) under weighted Euclidean preferences
is to move the solution in the direction of the ideal point to the neighborhood boundary, which, as it
turns out, is the same as the direction of the gradient. Thus with radius rt, the effective movement is
∇fv(xt)
. With (normalized) weighted Euclidean utilities, (cid:107)∇fv(xt)(cid:107)2 = 1 everywhere. As before,
∇fv(xt)2
even if the utilities were not normalized (i.e. not divided by (cid:107)w(cid:107)2), the algorithm would converge
to the same point, as if utility functions were normalized.
8
3.2 Decomposable Utilities
Next consider the general class of decomposable utilities, motivated by the fact that the algorithm
with L∞ neighborhoods is of special interest since they are easy for humans to understand: one can
change each dimension up to a certain amount, independent of the others.
Definition 3. Decomposable utilities. A voter utility function is decomposable if there exists
concave functions f m
v
v (xm).
for m ∈ {1 . . . M} such that fv(x) =(cid:80)M
m=1 f m
If the utility functions for the voters are decomposable, then we can show that our algorithm un-
der L∞ neighborhoods converges to the vector of medians of voters' ideal points on each dimension.
Suppose that hmX is the marginal density function of the random variable xm
v , and let ¯xm be the set
of medians of xm
v . (By set of medians, we mean the set of points such that, on each dimension, the
mass of voters with ideal points above and below.)
Proposition 2. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are de-
composable, and voters respond to query (1) according to either Model A or Model B. Then, ILV
with L∞ neighborhoods converges with probability 1 to a point in the set of medians ¯x.
Although simply eliciting each agent's optimal solution and computing the vector of median
allocations on each dimension is a viable approach in the case of decomposable utilities, deciding
an optimal allocation across multiple dimensions is a more challenging cognitive task than deciding
whether one wants to increase or decrease each dimension relative to the current solution (see
Section 5.2.3 for experimental evidence). In fact, in this case, the algorithm can be run separately
for each dimension, so that each voter expresses her preferences on only one dimension, drastically
reducing the cognitive burden of decision-making on the voter, especially in high dimensional settings
like budgeting.
3.3 Equivalence to Directional Equilibrium
As discussed in Section 2, our algorithm, with L2-norm neighborhoods, is related to an algorithm,
NGA, to find what are called Directional Equilibria in literature. Prior work mostly focuses on the
properties of the fixed point, with discussion of the proposed algorithm limited to simulations. We
show that with the radius decreasing as O( 1
t ), the algorithm indeed finds directional equilibria in
the following sense:
if under a few conditions a trajectory of the algorithm converges to a point,
then that point is a directional equilibrium.
Theorem 3. Suppose that C1, C2, and C3 are satisfied, and let G(x) (cid:44) Ev(cid:104) ∇fv(x)
(cid:107)∇fv(x)(cid:107)2(cid:105). Suppose,
G(x) is uniformly continuous, L2 movement norm constraints are used, and voters move according
to Model B. If a trajectory {x}∞
t=1 of the algorithm converges to x∗, i.e. xt → x∗, then x∗ is a
directional equilibrium, i.e. G(x∗) = 0.
The proof is in the appendix. It relies heavily on the continuity assumption: if a point x is not
a directional equilibrium, then the algorithm with step sizes O( 1
t ) will with probability 1 leave any
small region surrounding x: the net drift of the voter movements is away from the region. We note
that the necessary assumptions hold for all utility functions for which convergence holds, using the
L2 norm algorithm (e.g. weighted Euclidean utilities). It is further possible to characterize other
utility functions for which the equivalence holds: with appropriate conditions on the distribution of
voters and how f differs among voters, the conditions on G can be met.
9
We further conjecture that even under voter Model A, if Algorithm 2 converges, the fixed point
is a Directional Equilibrium. Note that as rt → 0, fv(y) can be linearly approximated by the first
term of the Taylor series expansion around x, for y ∈ {s : s − x2 ≤ rt}. Then, to maximize fv(y)
in the region, if the region does not contain xv voter v chooses y∗ s.t. y∗ − x ≈ rt
, i.e.,
the voter moves the solution approximately in the direction of her gradient to the neighborhood
boundary.
∇f (x)
∇f (x)2
A single step of our algorithm with L2 neighborhoods is similar to Quadratic Voting [20, 34] for
the same reason. Independently of our work, Benjamin et al. [2] formalize the relationship between
the Normalized Gradient Ascent mechanism and Quadratic Voting.
4 Experiments with Budgets
We built a voting platform and ran a large scale experiment, along with several extensive pilots,
on Amazon Mechanical Turk (https://www.mturk.com). Over 4,000 workers participating in total
counting pilots and the final experiment, with over 2,000 workers participating in the final experi-
ment. The design challenges we faced and voter feedback we received provide important lessons for
deploying such systems in a real-world setting.
First we present a theoretical model for our setting. We consider a budget allocation problem
on M items, where the items may include both expenditures and incomes. One possibility is to
define X as the space of feasible allocations, such as those below a spending limit, and to run the
algorithm as defined, with projections. However, in such cases, it may be difficult to theorize about
how voters behave; e.g.
if voters knew their answers would be projected onto a budget balanced
set, they may respond differently.
Rather, we consider an unconstrained budget allocation problem, one in which a voter's utility
includes a term for the budget deficit. Let E ⊆ {1 . . . M}, I = {1 . . . M}\E be the expenditure and
income items, respectively. Then the general budget utility function is fv(x) = gv(x)− d((cid:80)e∈E xe −
(cid:80)i∈I xi), where d is an increasing function on the deficit.
For example, suppose a voter's disutility was proportional to the square of the budget deficit (she
especially dislikes large budget deficits); then, this term adds complex dependencies between the
budget items. In general, nothing is known about convergence of Algorithm 1 with such utilities, as
the deficit term may add complex dependencies between the dimensions. However, if the voter utility
functions are decomposable across the dimensions and L∞ neighborhoods used, then the results of
Section 3.2 can be applied. We propose the following class of decomposable utility functions for the
budgeting problem, achieved by assuming that the cost for the deficit is linear, and call the class
"decomposable with a linear cost for deficit," or DLCD.
Definition 4. Let fv(x) be DLCD if
fv(x) =
M(cid:88)m=1
f m
v (xm) − wv(cid:32)(cid:88)e∈E
xi(cid:33) ,
xe −(cid:88)i∈I
where f m
v
is a concave function for each m and wv ∈ R+.
In the experiments discussed below in the budget setting, ILV consistently and robustly converges
with L∞ norm neighborhoods. Further, it approximately converges to the medians of the optimal
solutions (which are elicited independently), as theorized in Section 3.2. Such a convergence pattern
suggests the validity of the DLCD model, though we do not formally analyze this claim.
10
4.1 Experimental Setup
We asked voters to vote on the U.S. Federal Budget across several of its major categories: National
Defense; Healthcare; Transportation, Science, & Education; and Individual Income Tax (Note that
the US Federal Government cannot just decide to set tax receipts to some value. We asked workers
to assume tax rates would be increased or decreased at proportional rates in hopes of affecting
receipts.)
This setting was deemed the most likely to be meaningful to the largest cross-section of workers
and to yield a diversity of opinion, and we consider budgets a prime application area in general. The
specific categories were chosen because they make up a substantial portion of the budget and are
among the most-discussed items in American politics. We make no normative claims about running
a vote in this setting in reality, and Participatory Budgeting has historically been more successful
at a local level.
One major concern was that with no way to validate that a worker actually performed the task
(since no or little movement is a valid response if the solution presented to the worker was near
her ideal budget), we may not receive high-quality responses. This issue is especially important
in our setting because a worker's actions influence the initial solution future workers see. We thus
restricted the experiment to workers with a high approval rate and who have completed over 500
tasks on Mechanical Turk (MTurk). Further, we offered a bonus to workers for justifying their
movements well, and more than 80% of workers qualified, suggesting that we also received high-
quality movements. The experiment was restricted to Americans to best ensure familiarity with the
setting. Turkprime (https://www.turkprime.com) was used to manage postings and payment.
4.2 Experimental Parameters
Our large scale experiment included 2,000 workers and ran over a week in real-time. Participants
of any of the pilots were excluded. We tested the L1,L2, and L∞ mechanisms, along with a "full
elicitation" mechanism in which workers reported their ideal values for each item, and a "weight" in
[0, 10] indicating how much they cared about the actual spending in that item being close to their
stated value.
To test repeatability of convergence, each of the constrained mechanisms had three copies, given
to three separate groups of people. Each group consisted of two sets with different starting points,
with each worker being asked to vote in each set in her assigned group. Each worker only participates
as part of one group, and cannot vote multiple times.
We used a total of three different sets of starting points across the three groups, such that each
group shared one set of starting points with each of the other two groups. This setup allowed testing
for repeatability across different starting points and observing each worker's behavior at two points.
Workers in one group in each constrained mechanism type were also asked to do the full elicitation
after submitting their movements for the constrained mechanism, and such workers were paid extra.
These copies, along with the full elicitation, resulted in 10 different mechanism instances to which
workers could be allocated, each completed by about 200 workers.
To update the current point, we waited for 10 submissions and then updated the point to their
average. This averaging explains the step-like structure in the convergence plots in the next section.
The radius was decreased approximately every 60 submissions, rt (cid:117) r0
(cid:100)t/60(cid:101). The averaging and slow
radius decay rate were implemented in response to observing in the pilots that the initial few voters
with a large radius had a disproportionately high impact, as there were not enough subsequent voters
11
to recover from large initial movements away from an eventual fixed point (though in theory this
would not be a problem given enough voters). We note that the convergence results for stochastic
subgradient methods trivially extend to cover these modifications: the average movement over a
batch of submissions starting at the same point is still in expectation a subgradient, and the stepped
radius decrease still meets the conditions for valid step-sizes.
4.3 User Experience
Figure 1: UI Screenshot for 1 set of the L2 Mechanism
As workers arrived, they were randomly assigned to a mechanism instance. They had a roughly
equal probability of being assigned to each instance, with slight deviations in case an instance
was "busy" (another user was currently doing the potential 10th submission before an update of
the instance's current point) and to keep the number of workers in each instance balanced. Upon
starting, workers were shown mechanism instructions. We showed the instructions on a separate
page so as to be able to separately measure the time it takes to read & understand a given mechanism,
and the time it takes to do it, but we repeated the instructions on the actual mechanism page as
well for reference.
On the mechanism page, workers were shown the current allocation for each of the two sets in
their group. They could then move, through sliders, to their favorite allocation under the movement
constraint. We explained the movement constraints in text and also automatically calculated for
them the number of "credits" their current movements were using, and how many they had left.
Next to each budget item, we displayed the percentage difference of the current value from the
2016 baseline federal budget, providing important context to workers (The 2016 budget estimate
was obtained from http://federal-budget.insidegov.com/l/119/2016-Estimate and http://
atlas.newamerica.org/education-federal-budget). We also provided short descriptions of what
goes into each budget item as scroll-over text. The resulting budget deficit and its percent change
were displayed above the sliders, assuming other budget items are held constant.
For the full elicitation mechanism, workers were asked to move the sliders to their favorite points
with no constraints (the sliders went from $0 to twice the 2016 value in that category), and then
were asked for their "weights" on each budget item, including the deficit. Figure 1 shows part of
12
the interface for the L2 mechanism, not including instructions, with similar interfaces for the other
constrained mechanisms. The full elicitation mechanism additionally included sliders for items'
weights. On the final page, workers were asked for feedback on the experiment.
A full walk-through of the experiment with screenshots and link to an online demo is available
in the Appendix. We plan on posting the data, including feedback. In general, workers seemed
to like the experiment, though some complained about the constraints, and others were generally
confused. Some expressed excitement about being asked their views in an innovative manner and
suggested that everyone could benefit from participating as, at the least, a thought exercise. The
feedback and explanations provided by workers were much longer than we anticipated, and they
convince us of the procedure's civic engagement benefits.
5 Results and Analysis
We now discuss the results of our experiments.
5.1 Convergence
(a) L1
Figure 2: Solution over time for each mechanism type
One basic test of a voting mechanism is whether it produces a consistent and unique solution,
given a voting population and their behaviors. If an election process can produce multiple, distinct
solutions purely by chance, opponents can assail any particular solution as a fluke and call for a
re-vote. The question of whether the mechanisms consistently converge to the same point thus must
13
2003004005006007008007008009001000110012001300100200300400500600700120013001400150016001700180005010015020020040060080010001200DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration$(Billions)L1Group1,Set0L1Group1,Set1L1Group2,Set0L1Group2,Set1L1Group3,Set0L1Group3,Set1IdealPts.median (b) L2
Figure 2: (Continued) Solution over time for each mechanism type
(c) L∞
14
2003004005006007008007008009001000110012001300100200300400500600700120013001400150016001700180005010015020020040060080010001200DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration$(Billions)L2Group1,Set0L2Group1,Set1L2Group2,Set0L2Group2,Set1L2Group3,Set0L2Group3,Set1IdealPts.median2003004005006007008007008009001000110012001300100200300400500600700120013001400150016001700180005010015020020040060080010001200DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration$(Billions)L∞Group1,Set0L∞Group1,Set1L∞Group2,Set0L∞Group2,Set1L∞Group3,Set0L∞Group3,Set1IdealPts.median (a) L1
Figure 3: Net normalized movement in window of N = 30
be answered before analyzing properties of the equilibrium point itself. In this section, we show that
the L2 and L1 algorithms do not appear to converge to a unique point, while the L∞ mechanism
converges to a unique point across several initial points and with distinct worker populations.
The solutions after each voter for each set of starting points, across the 3 separate groups of
people for each constrained mechanism are shown in Figure 2. Each plot shows all the trajectories
with the given mechanism type, along with the median of the ideal points elicited from the separate
voters who only performed the full elicitation mechanism. Observe that the three mechanisms have
remarkably different convergence patterns. In the L1 mechanism, not even the sets done by the
same group of voters (in the same order) converged in all cases. In some cases, they converged for
some budget items but then diverged again. In the L2 mechanism, sets done by the same voters
starting from separate starting points appear to converge, but the three groups of voters seem to
have settled at two separate equilibria in each dimension. Under the L∞ neighborhood, on the other
hand, all six trajectories, performed by three groups of people, converged to the same allocation
very quickly and remained together throughout the course of the experiment. Furthermore, the
final points, in all dimensions except Healthcare, correspond almost exactly to the median of values
elicited from the separate set of voters who did only the full elicitation mechanism. For Healthcare,
the discrepancy could result from biases in full elicitation (see Section 5.2.4), though we make no
definitive claims. These patterns shed initial insight on how the use of L2 constraints may differ
from theory in prior literature and offer justification for the use of DLCD utility models and the
L∞ constrained mechanism.
One natural question is whether these mechanisms really have converged, or whether if we let
the experiment continue, the results would change. This question is especially salient for the L2
15
0.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00501001502000.00.20.40.60.81.0DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1min(t,N)Pts=max(0,t−N)xs−xs−1rt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)L1Group1,Set0L1Group1,Set1L1Group2,Set0L1Group2,Set1L1Group3,Set0L1Group3,Set1 (b) L2
Figure 3: (Continued) Net normalized movement in window of N = 30.
(c) L∞
16
0.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00501001502000.00.20.40.60.81.0DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1min(t,N)Pts=max(0,t−N)xs−xs−1rt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)L2Group1,Set0L2Group1,Set1L2Group2,Set0L2Group2,Set1L2Group3,Set0L2Group3,Set10.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00.00.20.40.60.81.00501001502000.00.51.01.52.0DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseIteration(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1min(t,N)Pts=max(0,t−N)xs−xs−1rt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)L∞Group1,Set0L∞Group1,Set1L∞Group2,Set0L∞Group2,Set1L∞Group3,Set0L∞Group3,Set1 trajectories, where trajectories within a group of people converged to the same point, but trajectories
between groups did not. Such a pattern could suggest that our results are a consequence of the
radius decreasing too quickly over time, or that the groups had, by chance, different distributions
of voters which would have been corrected with more voters. However, we argue that such does not
seem to be the case, and that the mechanism truly found different equilibria.
We can test whether the final points for each trajectory are stable by checking the net movement
in a window, normalized by each voter's radius, i.e. 1
, for some N. If voters in a
window are canceling each other's movements, then this value would go to 0, and the algorithm would
be stable even if the radius does not decrease. The notion is thus robust to apparent convergence
just due to decreasing radii. The net movement normalized in a sliding window of 30 voters, for
each dimension and mechanism, is shown in Figure 3. It seems to almost die down for almost all
mechanisms and budget items, except for a few cases which do not change the result. We conclude
it likely that the mechanisms have settled into equilibria which are unlikely to change given more
voters.
N(cid:80)t
s=t−N
xs−xs−1
rs
5.2 Understanding Voter Behavior
A mechanism's practical impact depends on more than whether it consistently converges, however.
We now turn our attention to understanding how voters behave under each mechanism and whether
we can learn anything about their utility functions from that behavior. We find that voters under-
stood the mechanisms but that their behaviors suggest large indifference regions, and that the full
elicitation scheme is susceptible to biases that can skew the results.
5.2.1 Voter understanding of mechanisms
One important question is whether, given very little instruction on how to behave, voters understand
the mechanisms and act approximately optimally under their (unknown to us) utility function. This
section shows that the voters behavior follows what one would expect in one important respect: how
much of one's movement budget the voter used on each dimension, given the constraint type.
Regardless of the exact form of the utility function, one would expect that, in the L1 constrained
mechanism, a voter would use most of her movement credits in the dimension about which she cares
most. In fact, in either the Weighted Euclidean preferences case (and with 'sub-space' being a single
dimension) or with a small radius with L1 constraints, a voter would move only on one dimension.
With L2 constraints, one would expect a voter to apportion her movement more equally because she
pays an increasing marginal cost to move more in one dimension (people were explicitly informed
of this consequence in the instructions). Under the Weighted Euclidean preferences model with L2
constraints, a voter would move in each dimension proportional to her weight in that dimension.
Finally, with L∞ constraints, a voter would move, in all dimensions in which she is not indifferent, to
her favorite point in the neighborhood for that dimension (most likely an endpoint), independently
of other dimensions. One would thus expect a more equal distribution of movements.
Figure 4 shows the average movement (as a fraction of the voter's total movement) by each voter
for the dimension she moved most, second, third, and fourth, respectively, for each constrained
mechanism. We reserve discussion of the full elicitation weights for Section 5.2.4. The movement
patterns indicate that voters understood the constraints and moved accordingly -- with more equal
movements across dimensions in L2 than in L1, and more equal movements still in L∞. We dig
deeper into user utility functions next, but can conclude that, regardless of their exact utility
17
functions, voters responded to the constraint sets appropriately.
Figure 4: Average movement in dimension over total movement for each voter, with dimensions
sorted
5.2.2 Large indifference regions
Although it is difficult to extract a voter's full utility function from their movements, the separability
of dimensions (except through the deficit term) under the L∞ constraint allows us to test whether
voters behave according to some given utility model in that dimension, without worrying about the
dependency on other dimensions.
Figure 5 shows, for the L∞ mechanism, a histogram of the movement on a dimension as a frac-
tion of the radius (we find no difference between dimensions here). Note that a large percentage of
voters moved very little on a dimension, even in cases where their ideal point in that dimension was
far away (defined as being unreachable under the current radius). This result cannot be explained
away by workers clicking through without performing the task: almost all workers moved at least
one dimension, and, given that a worker moved in a given dimension, it would not explain smaller
movements being more common than larger movements. That this pattern occurs in the L∞ mech-
anism is key -- if a voter feels any marginal disutility in a dimension, she can move the allocation
without paying a cost of more limited movement in other dimensions. We conclude that, though
voters may share a single ideal point for a dimension when asked for it, they are in fact relatively
indifferent over a potentially large region -- and their actions reflect so.
We further analyze this claim in Appendix Section B, looking at the same distribution of move-
ment but focusing on workers who provided a text explanation longer than (and shorter than,
separately) the median explanation of 197 characters. (We assume that the voters who invested
time in providing a more thorough explanation than the average worker also invested time in moving
the sliders to a satisfactory point, though this assumption cannot be validated.) Though there are
some differences (those who provide longer explanations also tend to use more of their movement),
the general pattern remains the same; only about 40% of workers who provided a long explanation
and were far away from their ideal point on a dimension used the full movement budget. This
pattern suggests that voters are relatively indifferent over large regions.
18
FirstSecondThirdFourth0.00.10.20.30.40.50.60.7FullElicitationWeightsL1L2L∞RankingofdimensionsbymovementforeachvoterMovementasfractionoftotalmovement Furthermore, this lack of movement is correlated with a voter's weights when she was also asked
to do the full elicitation mechanism. Conditioned on being far from her ideal point, when a voter
ranked an item as one of her top two important items (not counting the deficit term), she moved
an average of 74% of her allowed movement in that dimension; when she ranked an item as one of
least two important items, she moved an average of 61%, and the difference is significant through
a two sample t-test with p = .013. We find no significant difference in movement within the top
two ranked items or within the bottom two ranked items. This connection suggests that one can
potentially determine which dimensions a voter cares about by observing these indifference regions
and movements, even in the L∞ constrained case. One caveat is that the differences in effects
are not large, and so at the individual level inference of how much an individual cares about one
dimension over another may be noisy. On the aggregate, however, such determination may prove
useful.
Furthermore, we note that while such indifference regions conflict with the utility models un-
der which the L2 constraint mechanism converges in theory, it fits within the DLCD framework
introduced in Section 4.
Figure 5: Fraction of possible movement in each dimension in L∞, conditioned on distance to ideal
pt. The 'All' condition contains data from all three L∞ instances, whereas the others only from the
instance that also did full elicitation.
5.2.3 Mechanism time
In this section, we note one potential problem with schemes that explicitly elicit voter's optimal
solutions -- for instance, to find the component-wise median -- as compared to the constrained
elicitation used in ILV: it seems to be cognitively difficult for voters. In Figure 6, the median time
per page, aggregated across each mechanism type, is shown. The "Mechanism" time includes a
single user completing both sets in each of the constrained mechanism types, but not does include
the time to also do the extra full elicitation task in cases where a voter was asked to do both
a constrained mechanism and the full elicitation. The full elicitation bars include only voters
who did only the full elicitation mechanism, and so the bars are completely independent. On
average, it took longer to do the full elicitation mechanism than it took to do two sets of any
19
0.00.20.40.60.81.0Movementasfractionofpossiblemovement0.000.050.100.150.200.250.300.350.40HistogrampdfAllNearIdealPtFarfromIdealPt Figure 6: Median time per page
of the constrained mechanisms, suggesting some level of cognitive difficulty in articulating one's
ideal points and weights on each dimension -- even though understanding what the instructions are
asking was simple, as demonstrated by the shorter instruction reading time for the full elicitation
mechanism. The L∞ mechanism took the least time to both understand and do, while the L2
mechanism took the longest to do, among the constrained mechanisms.
This result is intuitive: it is easier to move each budget item independently when the maximum
movement is bounded than it is to move the items when the sum or the sum of the changes squared
is bounded (even when these values are calculated for the voter). In practice, with potentially tens
of items on which constituents are voting, these relative time differences would grow even larger,
potentially rendering full elicitation or L2 constraints unpalatable to voters.
One potential caveat to this finding is that the Full Elicitation mechanism potentially provides
more information than do the other mechanisms. From a polling perspective, it is true that more
information is provided from full elicitation -- one can see the distribution of votes, the disagreement,
and correlation across issues, among other things. However, from a voting perspective, in which the
aggregation (winner) is the only thing reported, it is not clear that this extra information is useful.
Further, much of this information that full elicitation provides can reasonably be extracted from
movements of voters, especially the movements of those who are given a starting point close to the
eventual equilibrium.
5.2.4 UI biases
We now turn our attention to the question of how workers behaved under the full elicitation mech-
anism and highlight some potential problems that may affect results in real deployments. Fig-
ures 7 and 8 show the histogram of values and weights, respectively, elicited from all workers who
did the full elicitation mechanism. Note that in the histogram of values, in every dimension, the
largest peak is at the slider's default value (at the 2016 estimated budget), and the histograms seem
to undergo a phase shift at that peak, suggesting that voters are strongly anchored at the slider's
starting value. This anchoring could systematically bias the medians of the elicited values.
20
WelcomePageInstructionsMechanismFeedback050100150200250FullElicitationL2L1L∞PageTime(Seconds) A similar effect occurs in eliciting voter weights on each dimension. Observe that in Figure 4 the
full elicitation weights appear far more balanced than the weights implied by any of the mechanisms
(for the full elicitation mechanism, the plot shows the average weight over the sum of the weights
for each voter). From the histogram of full elicitation weights, however, we see that this result is
a consequence of voters rarely moving a dimension's weight down from the default of 5, but rather
moving others up.
One potential cause of this behavior is that voters might think that putting high weights on each
dimension would mean their opinions would count more, whereas in any aggregation one would either
ignore the weights (calculate the unweighted median) or normalize the weights before aggregating.
In future work, one potential fix could be to add a "normalize" button for the weights, which would
re-normalize the weights, or to automatically normalize the weights as voters move the sliders.
These patterns demonstrate the difficulty in eliciting utilities from voters directly; even asking
voters how much they care about a particular budget item is extremely susceptible to the user
interface design. Though such anchoring to the slider default undoubtedly also occurs in the L∞
constrained mechanism, it would only slow the rate of convergence, assuming the anchoring affects
different voters similarly. These biases can potentially be overcome by changing the UI design, such
as by providing no default value through sliders. Such design choices must be carefully thought
through before deploying real systems, as they can have serious consequences.
Figure 7: Histogram of values from all full elicitation data. The red vertical lines indicate each
slider's default value (at the 2016 estimated budget).
6 Conclusion
We evaluate a natural class of iterative algorithms for collective decision-making in continuous spaces
that makes practically reasonable assumptions on the nature of human feedback. We first introduce
several cases in which the algorithm converges to the societal optimum point, and others in which
the algorithm converges to other interesting solutions. We then experimentally test such algorithms
in the first work to deploy such a scheme. Our findings are significant: even with theoretical backing,
21
020406080100120140020406080100120140160050100150200250050100150200−50005001000150020002500020406080100120140160DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseFullElicitationValue(Billionsof$)NumberofVoters Figure 8: Histogram of weights from all full elicitation data. The red vertical lines indicate the
sliders' default value of 5.
two variants fail the basic test of being able to give a consistent decision across multiple trials with
the same set of voters. On the other hand, a variant that uses L∞ neighborhoods consistently leads
to convergence to the same solution, which has attractive properties under a likely model for voter
preferences suggested by this convergence. We also make certain observations about other properties
of user preferences -- most saliently, that they have large indifferences on dimensions about which
they care less.
In general, this work takes a significant step within the broad research agenda of understanding
the fundamental limitations on the quality of societal outcomes posed by the constraints of human
feedback, and in designing innovative mechanisms that leverage this feedback optimally to obtain
the best achievable outcomes.
A Mechanical Turk Experiment Additional Information
In this section, we provide additional information regarding our Amazon Mechanical Turk experi-
ment, including a walk-through of the user experience. Furthermore, we have a live demo that can
be accessed at: http://gargnikhil.com/projectdetails/IterativeLocalVoting/. This demo
will remain online for the foreseeable future.
Figures 9 through 13 show screenshots of the experiment. We now walkthrough the experiment:
introduction and the consent agreement.
• Figure 9 -- Welcome page. Arriving from Amazon Mechanical Turk, the workers read an
• Figure 10 -- Instructions (shown are L2 instructions). The workers read the instructions, which
• Figures 11, 13 -- Mechanism page for L2 and Full Elicitation, respectively. For the former,
workers are asked to move to their favorite point within a constraint set, for 2 different budget
points. The "Current Credit Allocation" encodes the constraint set -- as workers move the
are also provided on the mechanism page. There is a 5 minute limit for this page.
22
0204060801001200204060801001200204060801001201400204060801001201401600246810020406080100120140160DeficitIncomeTaxTransportation,Science,&EducationHealthcareDefenseFullElicitationWeightNumberofVoters Figure 9: Page 1 -- Welcome Page for all mechanisms
Figure 10: L2 Page 2 -- Instructions
budget bars, it shows how much of their movement budget they have spent, and on which
items. The other constrained movement mechanisms are similar. For the Full Elicitation
mechanism, voters are simply asked to indicate their favorite budget point and weights. The
instructions are repeated on the mechanism page as well at the top. There is a 10 minute
limit for this page.
• Second mechanism, 30% of workers. Some workers were asked to do both one of the L1,L2, or
L∞, and the Full Elicitation mechanism. For these workers, the Full Elicitation mechanism
shows up after the constrained mechanism.
• Figure 12 -- Feedback page. Finally, workers are asked to provide feedback, after which they
are shown a code and return to the Mechanical Turk website.
B Indifference Regions Additional Information
We now present some additional data for the claim in Section 5.2.2, that voters have large indifference
regions on the space. In particular, Figures 14 and 15 reproduce Figure 5 but with workers who
provided explanations longer (and shorter) than the median response, respectively. This split can
(roughly) correspond to workers who may have answered more or less sincerely to the budgeting
question. We find that the response distribution, as measured by the fraction of possible movement
one used when far away from one's ideal point on a given dimension, are similar.
23
Figure 11: L2 Page 3 -- Mechanism
Figure 12: Page 4 -- Feedback for all mechanisms
24
Figure 13: Full Elicitation Page 3 -- Mechanism
Figure 14: Fraction of possible movement in each dimension in L∞, conditioned on distance to ideal
pt. The 'All' condition contains data from all three L∞ instances, whereas the others only from
the instance that also did full elicitation. This plot only includes those people who provided an
explanation as long or longer than the median explanation provided (197 characters).
25
0.00.20.40.60.81.0Movementasfractionofpossiblemovement0.00.10.20.30.40.5HistogrampdfAllNearIdealPtFarfromIdealPt Figure 15: Fraction of possible movement in each dimension in L∞, conditioned on distance to ideal
pt. The 'All' condition contains data from all three L∞ instances, whereas the others only from
the instance that also did full elicitation. This plot only includes those people who provided an
explanation shorter than the median explanation provided (197 characters).
26
0.00.20.40.60.81.0Movementasfractionofpossiblemovement0.00.10.20.30.40.5HistogrampdfAllNearIdealPtFarfromIdealPt C Proofs
In this appendix, we include proofs for all the theorems in the paper.
C.1 Known SSGM Results
Theorem 4. [25, 33] Let θ ∈ Θ be a random vector with distribution P. Let ¯f (x) = E[f (x, θ)] =
well-defined and finite valued. Suppose that f (·, θ), θ ∈ Θ is convex and ¯f (·) is continuous and finite
valued in a neighborhood of point x. For each θ, choose any g(x, θ) ∈ ∂f (x, θ). Then, there exists
¯g(x) ∈ ∂ ¯f (x) s.t. ¯g(x) = Eθ[g(x, θ)].
(cid:82)Θ f (x, θ)dP (θ), for x ∈ X , a non-empty bounded closed convex set, and assume the expectation is
This theorem says that the expected value of the sub-gradient of the utility at any point x
across voters is a subgradient of the societal utility at x, irrespective of how the voters choose the
subgradient when there are multiple subgradients, i.e., when the utility function is not differentiable.
This key result allows us to use the subgradient of utility function of a sampled voter as an unbaised
estimate of the societal subgradient.
Now, consider a convex function f on a non-empty bounded closed convex set X ⊂ RM, and
use [·]X to designate the projection operator. Starting with some x0 ∈ X , consider the SSGM
update rule xt = [xt−1 − rt(¯gt + zt + bt)]X , where zt is a zero-mean random variable and bt is a
constant, and ¯gt ∈ ∂f (xt). Let Et[·] be the conditional expectation given Ft, the σ-field generated
by x0, x1, . . . , xt. Then we have the following convergence result.
Theorem 5. [19] Consider the above update rule. If
f (·) has a unique minimizer x
t < ∞
r2
∗ ∈ X
rt > 0,(cid:88)t
rt = ∞,(cid:88)t
∃C1 ∈ R < ∞ s.t. (cid:107)∂f (x)(cid:107)2 ≤ C1,∀ x ∈ X
∃C2 ∈ R < ∞ s.t. Et[(cid:107)zt(cid:107)2] ≤ C2,∀ t
∃C3 ∈ R < ∞ s.t. (cid:107)bt(cid:107)2 ≤ C3,∀ t
rt(cid:107)bt(cid:107) < ∞ w.p. 1
(cid:88)t
Then xt → x∗ w.p. 1 as t → ∞.
Note: Jiang and Walrand [19] prove the result for gradients, though the same proof follows for
subgradients. Only the inequality [x∗ − xt]T gt ≤ f (x∗) − f (xt) for gradient gt at iteration t is used,
which holds for subgradients. Boyd and Mutapcic [4] provide a general discussion of subgradient
methods, along with similar results. Shor [31], in Theorem 46, provide a convergence proof for the
stochastic subgradient method without projections and the extra noise terms.
C.2 Mapping ILV to SSGM
As described in Section 3, suppose that hX is the induced probability distribution on the ideal
values of the voters. In the following discussion, we will refer to voters and their ideal solutions
interchangeably.
27
Next, we restate ILV without the stopping condition so that it looks like the stochastic subgra-
dient method. Consider Algorithm 2.
Algorithm 2: ILV
Start at some x0 ∈ X . For t ≥ 1,
• Sample voter vt ∈ V from hV.
• Compute xt = [xt−1 − rtgvt(xt)]X , where rt = r0
t and rtgvt(xt) is movement given by vt.
We want to minimize the societal cost, ¯f (x) = E[fv(x)]. From Theorem 4, it immediately follows
that if each voter v articulates a subgradient of her utility function for all x, i.e. gv(x) ∈ ∂fv(x),
then from Theorem 5, we can conclude that the algorithm converges. However, users may not be
able to articulate such a subgradient. Instead, when the voters respond correctly to query (1) (i.e.
move to their favorite point in the given Lq neighborhood), we have
gvt(xt) =
xt − arg minx[fvt(x) : (cid:107)x − xt(cid:107)q ≤ rt]
rt
.
(2)
Furthermore, for all the proofs, we assume the following.
A. The solution space X ⊂ RM is non-empty, bounded, closed, and convex.
B. Each voter v has a unique ideal solution xv ∈ X .
C. The ideal point xv of each voter is drawn independently from a probability distribution with
a bounded and measurable density function hX on M dimensions: there exists C s.t. ∀ x we
have hX (x) ≤ C. This assumption allows us to bound the probability of errors that occur in
small regions of the space.
C.3 Proof of Theorem 1
Let the disutility, or cost to voter v ∈ V be fv(x) = (cid:107)x − xv(cid:107)p for all x ∈ X . We use the following
technical lemma:
Lemma 1. For q ∈ {1, 2,∞}, there exists K2 ∈ R+ s.t. (cid:107)gv(x) − gt(cid:107)2 ≤ K2, ∀ gt ∈ ∂fv(x) for any
v and x.
The lemma bounds the error in the movement direction from the gradient direction, by noting that
both the movement direction and the gradient direction have bounded norms.
We also need the following lemma, which is proved separately for each case in the following sections.
Lemma 2. Suppose that fv(x) (cid:44) (cid:107)xv − x(cid:107)p and define the function
At (cid:44) I{gvt(xt) /∈ ∂fvt(xt)},
where gvt(xt) is as defined in (2). Then there exists C ∈ R s.t. ∀ n, P(At = 1Ft) ≤ Crt, when
(p = 2, q = 2), (p = 1, q = ∞), or (p = ∞, q = 1).
The lemma can be interpreted as follows: At indicates a 'bad' event, when a voter may not be
providing a true subgradient of her utility function. However, the probability of the event occurring
vanishes with rt, which, as we will see below, is the right rate for the algorithm to converge.
28
Theorem 1. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are Lp normed,
and voters respond to query (1) according to either Model A or Model B. Then, ILV with Lq
neighborhoods converges to the societal optimal point w.p. 1 when (p, q) = (2, 2), (1,∞), or (∞, 1).
Proof. We will show that Algorithm 2 meets the conditions in Theorem 5. Let bt (cid:44) Et[gvt(xt)] − ¯gt
and zt (cid:44) gvt(xt)−Et[gt], for some ¯gt ∈ ∂ ¯f (xt). Then, gvt(xt) can be written as gvt(xt) = ¯gt + zt + bt.
We show that bt, zt meet the conditions in the theorem, and so the algorithm converges.
Let At be the indicator function described in Lemma 2. Then, for some ¯gt ∈ ∂ ¯f (xt),
bt = Et[gvt(xt)] − ¯gt
= Et[gvt(xt)] − Et[gt]
= P(At = 1Ft)(Et[gvt(xt)At = 1] − Et[gtAt = 1])
+ P(At = 0Ft)(Et[gvt(xt)At = 0] − Et[gtAt = 0])
= P(At = 1Ft)(Et[gvt(xt)At = 1] − Et[gtAt = 1])
≤ Crt(Et[gvt(xt)At = 1] − Et[gtAt = 1]).
Theorem 4, i.i.d sampling of v
Lemma 2
Combining with Lemma 1, and the fact that rt = r0/t, we have
(cid:88) rt(cid:107)bt(cid:107) ≤ ∞ and there exists C1 ∈ R < ∞ s.t. (cid:107)bt(cid:107)2 ≤ C1,∀ t.
Finally, note that (cid:107)zt(cid:107) (cid:44) (cid:107)gvt(xt) − Et[gvt(xt)](cid:107) is bounded for each t because the (cid:107)gvt(xt)(cid:107) is
bounded as defined. Thus, all the conditions in Theorem 5 are met for both bt and zt, and the
algorithm converges.
C.4 Proof of Theorem 2
Instead of moving to their favorite point on the ball, voters now instead move in the direction of
the gradient of their utility function to the boundary of the given neighborhood. In this case, we
have:
gvt(xt) =
gvt
(cid:107)gvt(cid:107)q
; for gvt ∈ ∂fvt(xt).
(3)
The key to the proof is the following observation, that the q norm of the gradient of the p norm,
except at the ideal points on each dimension, is constant. This observation is formalized in the
following lemma:
Lemma 3. ∀ (p, q) s.t. p > 0, q > 0, and 1/p + 1/q = 1, (cid:107)∇(cid:107)x − xv(cid:107)p(cid:107)q = 1,∀ x s.t. xm (cid:54)= xm
any m.
Theorem 2. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are Lp normed,
and voters respond to query (1) according to Model B. Then, ILV with Lq neighborhoods converges
to the societal optimal point w.p. 1 for any p > 0 and q > 0 such that 1/p + 1/q = 1.
v for
Proof. Since the probability of picking a voter v such that xm
v for some dimension m is 0, we
have gvt(xt) = gvt for gvt = ∇fvt(xt). Thus we obtain the gradient exactly, and hence Theorem 5
applies with bt = 0 for all t.
t = xm
29
C.5 Proof of Propositions
We now turn our attention to the case of Weighted Euclidean utilities and show that Algorithm 2
converges to the societal optimum. The analogue to Lemma 2 for this case is (proved in the following
subsection):
k=1
wk
v(cid:107)wv(cid:107)2
(cid:107)xk − xk
At (cid:44) I{gvt(xt) /∈ ∂fvt(xt)},
v(cid:107)2, and define the function
Lemma 4. Suppose that fv(x) (cid:44)(cid:80)K
where gvt(xt) is as defined in (2) for q = 2. Then there exists C ∈ R s.t. ∀ n, P(At = 1Ft) ≤ Crt.
Proposition 1. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are
Weighted Euclidean, and voters correctly respond to query (1) according to either Model A or
Model B. Then, ILV with L2 neighborhoods converges with probability 1 to the societal optimal
point.
Proof. The proof is then similar to that of Theorem 1, and the algorithm converges to x∗ =
(cid:80)K
arg min E(cid:104)
k=1 wk
v(cid:107)xk−xk
(cid:107)wv(cid:107)2
v(cid:107)2
(cid:105).
Now, we sketch the proof for fully decomposable utility functions and L∞ neighborhoods.
Proposition 2. Suppose that conditions C1, C2, and C3 are satisfied, the voter utilities are de-
composable, and voters respond to query (1) according to either Model A or Model B. Then, ILV
with L∞ neighborhoods converges with probability 1 to a point in the set of medians ¯x.
t−1 + rt ≤ xm
t−1 < xm
v . On the other hand if xm
t−1 by
Proof. Consider each dimension separately. If xm
v , then the sampled voter decreases
rt as long as xm
v . Thus except for when a voter's ideal solution is too close to
t−1 by rt as long as xm
xm
the current point, the algorithm can be seen as performing SSGM on each dimension separately as
if the utility function was L1 (the absolute value) on each dimension. Thus a proof akin to that of
Theorem 1 with p = 1, q = ∞ holds.
v , then the sampled voter increases xm
t−1 > xm
t−1 − rt ≥ xm
C.6 Proof of Theorem 3
We now show that the algorithm finds directional equilibria in the following sense:
if under a
few conditions a trajectory of the algorithm converges to a point, then that point is a directional
equilibrium.
Theorem 3. Suppose that C1, C2, and C3 are satisfied, and let G(x) (cid:44) Ev(cid:104) ∇fv(x)
(cid:107)∇fv(x)(cid:107)2(cid:105). Suppose,
G(x) is uniformly continuous, L2 movement norm constraints are used, and voters move according
to Model B. If a trajectory {x}∞
t=1 of the algorithm converges to x∗, i.e. xt → x∗, then x∗ is a
directional equilibrium, i.e. G(x∗) = 0.
Proof. Suppose x∗ is not a directional equilibrium, i.e. ∃ > 0 s.t. (cid:107)G(x∗)(cid:107)2 = . Consider a
δ-ball around x∗, Bδ (cid:44) {x : (cid:107)x∗ − x(cid:107)2 < δ}, with δ, 2 > 0 chosen such that ∃m ∈ {1 . . . M} s.t.
∀x ∈ Bδ, sign(Gm(x)) = sign(Gm(x∗)) and Gm(x) > 2, i.e. the gradient in the mth dimension
does not change sign and has magnitude bounded below. Such a δ, 2 exists by the continuity as-
sumption (if x∗ is not a directional equilibrium, at least 1 dimension of G(x∗) is non-zero and thus
30
one can construct a ball around x∗ such that G(x), x ∈ Bδ in that dimension satisfies the conditions).
Now, one can show that the probability of leaving neighborhoods around x∗ goes to 1: ∀t >
0, 0 < δ2 < δ, w.p. 1 ∃τ ≥ t s.t. (cid:107)xτ − x∗(cid:107)2 > δ2.
Suppose xt ∈ Bδ2 (otherwise τ = t satisfies), rk = 1
k .
∆xk
∆xk (cid:44) −rk
∇f vk (xk)
(cid:107)∇f vk (xk)(cid:107)2
∗(cid:107)2
defn of (cid:107) · (cid:107)2
∆xk(cid:107)2 − δ2
∆xk,m
(cid:107)xτ − x
∗(cid:107)2 = (cid:107)xt − x
∗
+
∆xk(cid:107)2
τ(cid:88)k=t
∆xk(cid:107)2 − (cid:107)xt − x
(cid:107) τ(cid:88)k=t
xτ = xt +
τ(cid:88)k=t
≥ (cid:107) τ(cid:88)k=t
≥ (cid:107) τ(cid:88)k=t
∆xk(cid:107)2 ≥ τ(cid:88)k=t
= Ev(cid:34) τ(cid:88)k=t
≥ Ev(cid:34) τ(cid:88)k=t
P r(cid:32) τ(cid:88)k=t
By Hoeffding's inequality,
∆xk,m(cid:35) +
τ(cid:88)k=t
∆xk,m(cid:35) − τ(cid:88)k=t
∆xk,m − Ev(cid:34) τ(cid:88)k=t
∆xk,m(cid:35)
∆xk,m(cid:35)
∆xk,m − Ev(cid:34) τ(cid:88)k=t
∆xk,m − Ev(cid:34) τ(cid:88)k=t
∆xk,m(cid:35) ≥ 3(cid:33) ≤ exp(cid:34)− 2(τ − t)22
k (cid:35)
k=t
3
1
2(cid:80)τ
→ 0 as τ → ∞
Furthermore, by the continuity assumption,
Ev(cid:34) τ(cid:88)k=t
∆xk,m(cid:35) (cid:44) τ(cid:88)k=t
rkGm(xk)
→ ∞ as τ → ∞ while xk ∈ Bδ2
Thus, P r((cid:107)xτ − x∗(cid:107)2 > δ2) → 1 as τ → ∞. Thus, if an infinite trajectory converges to x∗, then
w.p. 1, then x∗ is a directional equilibrium.
C.7 Proofs of Lemmas
Lemma 1 For q ∈ {1, 2,∞}, ∃K2 ∈ R+ < ∞ s.t. (cid:107)gvt − gt(cid:107)2 ≤ K2, ∀ gt ∈ ∂fvt(xt), vt, xt.
31
Proof.
(cid:107)gvt(xt) − gt(cid:107)2 ≤ (cid:107)gvt(xt)(cid:107)2 + (cid:107)gt(cid:107)2
(cid:107)xt − arg minx[(cid:107)x − xvt(cid:107)p : (cid:107)x − xt(cid:107)q ≤ rt](cid:107)2
+ (cid:107)gt(cid:107)2
=
≤ K1 + (cid:107)gt(cid:107)2
≤ K2
rt
for some K1, K2 ∈ R+. The second inequality follows from the fact that for finite M-dimensional
vector spaces, (cid:107)y(cid:107)2 ≤ (cid:107)y(cid:107)1 and (cid:107)y(cid:107)2 ≤ √
M(cid:107)y(cid:107)∞. The third follows from the norm of the subgra-
dients of the p norm being bounded.
Lemma 2, case (p = 2, q = 2).
Proof. Remember that At (cid:44) I{gvt(xt) /∈ ∂fvt(xt)}. Let Bt = I{(cid:107)xvt − xt(cid:107)2 ≤ rt}. We show that
A) Bt = 0 =⇒ At = 0, and B) ∃C ∈ R s.t. P(Bt = 1Ft) ≤ Crt. Then, ∃C ∈ R s.t.
P(At = 1Ft) ≤ Crt.
Part A, Bt = 0 =⇒ gvt(xt) = gt, for some gt ∈ ∂fvt(xt):
First, note that
∂fvt(x) = ∂(cid:107)x − xvt(cid:107)2
=(cid:40){ x−xvt
x (cid:54)= xvt
}
(cid:107)xvt−x(cid:107)2
{g : (cid:107)g(cid:107)2 ≤ 1} x = xvt
=(cid:90)x∈{x:(cid:107)x−xt(cid:107)2≤rt}
=(cid:90)x∈{x:(cid:107)x−xt(cid:107)2≤rt}
≤ Cr2
≤ Crt
t
hXFt(x)dx
hX (x)dx
v drawn independent of history
bounded hX
rt ≤ 1 eventually
32
xvt − xt
(cid:107)xvt − xt(cid:107)2
Definition
If (cid:107)xvt − xt(cid:107)2 > rt, then
Then,
gvt(xt) =
=
[(cid:107)x − xvt(cid:107)2 : (cid:107)x − xt(cid:107)2 ≤ rt] = xt + rt
arg min
x
rt
xvt−xt
xt − arg minx[(cid:107)x − xvt(cid:107)2 : (cid:107)x − xt(cid:107)2 ≤ rt]
xt −(cid:16)xt + rt
rt
xt − xvt
(cid:107)xvt − xt(cid:107)2
(cid:107)xvt−xt(cid:107)2(cid:17)
=
∈ ∂fvt(xt)
Part B, ∃C ∈ R s.t. P(Bt = 1Ft) ≤ Crt:
P(Bt = 1Ft) = P((cid:107)xvt − xt(cid:107)2 ≤ rtFt)
for some C ∈ R < ∞. Note that C depends on the volume of a sphere in M dimensions.
Lemma 2, case (p = 1, q = ∞).
vt − x1
t ),
t )(cid:3)T
vt − xM
t ≤ rt}. We show the the same two parts as in the above proof.
Proof. Let hvt(xt) (cid:44)(cid:2)sign(x1
Let Bt = I{∃m,xm
vt − xm
Part A, Bt = 0 =⇒ gvt(xt) = gt, for some gt ∈ ∂fvt(xt):
First, note that the subgradients are
vt − xm
t ),
sign(xM
sign(xm
. . . ,
. . . ,
∂fvt(x) = ∂(cid:107)x − xvt(cid:107)1
= {g : (cid:107)g(cid:107)∞ ≤ 1, gT (x − xvt) = (cid:107)x − xvt(cid:107)1}
If ∀ m,xm
vt − xm
t > rt, then
[(cid:107)x − xvt(cid:107)1 : (cid:107)x − xt(cid:107)∞ ≤ rt] = xt + rthvt(xt)
arg min
x
Then,
gvt(xt) =
xt − arg minx[(cid:107)x − xvt(cid:107)1 : (cid:107)x − xt(cid:107)∞ ≤ rt]
xt − (xt + rthvt(xt))
rt
Definition
Part B, ∃C ∈ R s.t. P(Bt = 1Ft) ≤ Crt:
P(Bt = 1Ft) = P(∃m : xm
t ≤ rtFt)
rt
=
= −hvt(xt)
∈ ∂fvt(xt)
vt − xm
=(cid:90)x∈{x:∃m,xm−xm
=(cid:90)x∈{x:∃m,xm−xm
≤ Crt
t ≤rt}
t ≤rt}
hXFt(x)dx
hX (x)dx
v drawn independent of history
bounded hX , fixed M, bounded X
for some C ∈ R < ∞. In the last line, C (cid:117) 2M (diameter(X )), based on the volume of the slices
around the ideal points on each dimension.
Lemma 2, case (p = 1, q = ∞).
Proof. Let ¯mt ∈ arg maxm xm
. . . , 0,
vt ), 0, . . . , 0, 0(cid:3)T ,
t + rt}. We show the the same two parts as in the
Let hvt(xt) (cid:44)(cid:2)0, 0,
Let Bt (cid:44) I{∃m (cid:54)= ¯mt : x ¯mt
vt − xm
t ,
sign(x ¯mt
< xm
vt − x ¯mt
t − x ¯mt
vt − xm
t
33
above proofs.
Part A, Bt = 0 =⇒ gvt(xt) = gt, for some gt ∈ ∂fvt(xt):
First, note that when Bt = 0, the set of subgradients is
∂fvt(x) = ∂(cid:107)x − xvt(cid:107)∞
= {hvt(xt)}
Also when Bt = 0,
Then,
[(cid:107)x − xvt(cid:107)∞ : (cid:107)x − xt(cid:107)1 ≤ rt] = xt − rthvt(xt)
arg min
x
gvt(xt) =
=
xt − arg minx[(cid:107)x − xvt(cid:107)1 : (cid:107)x − xt(cid:107)∞ ≤ rt]
xt − (xt − rthvt(xt))
rt
rt
Definition
= hvt(xt)
∈ ∂fvt(xt)
Part B, ∃C ∈ R s.t. P(Bt = 1Ft) ≤ Crt:
P(Bt = 1Ft) = P(I{∃m (cid:54)= ¯mt : x ¯mt
vt − x ¯mt
t
< xm
vt − xm
t + rt}Ft)
hXFt(x)dx
=(cid:90)x∈{x:∃m(cid:54)= ¯mt s.t. x ¯mt
=(cid:90)x∈{x:∃m(cid:54)= ¯mt s.t. x ¯mt
≤ Crt
vt −x ¯mt
t
vt −x ¯mt
t
<xm
vt
−xm
t +rt}
<xm
vt
−xm
t +rt}
hX (x)dx
v drawn independent of history
bounded hX , fixed M, bounded X
for some C ∈ R < ∞. Note that C (cid:117) 2M 2(diameter(X )), based on the volume of the slices around
each dimension.
Lemma 3 ∀ (p, q) s.t. p > 0, q > 0, 1/p + 1/q = 1, (cid:107)∇(cid:107)x − xv(cid:107)p(cid:107)q = 1,∀ x s.t. ∀ m, xm (cid:54)= xm
v .
Proof. If xm (cid:54)= xm
v ,∀ m :
∇m(cid:107)x − xv(cid:107)p = ∇m(cid:32)(cid:88)m
v p(cid:33)1/p
xm − xm
∇mxm − xm
v p
v p)1−1/p
1
p
xm − xm
((cid:80)m xm − xm
v p−1 (∇mxm − xm
v )
(cid:107)x − xv(cid:107)p−1
p
=
=
34
Then
(cid:107)∇(cid:107)x − xv(cid:107)p(cid:107)q = (cid:107)xm − xm
v p−1 (∇mxm − xm
v )
(cid:107)x − xv(cid:107)p−1
p
(cid:107)q
1
1
1
p (cid:32)(cid:88)m (cid:12)(cid:12)xm − xm
p (cid:32)(cid:88)m
xm − xm
(cid:107)x − xv(cid:107)p/q
p
=
=
=
(cid:107)x − xv(cid:107)p−1
(cid:107)x − xv(cid:107)p−1
(cid:107)x − xv(cid:107)p−1
p
v )(cid:12)(cid:12)q(cid:33)1/q
v p−1 (∇mxm − xm
v (p−1)q(cid:33)1/q
(p − 1)q = p
= 1
k=1
wk
v(cid:107)wv(cid:107)2
(cid:107)xk − xk
At (cid:44) I{gvt(xt) /∈ ∂fvt(xt)},
v(cid:107)2, and define the function
Lemma 4 Suppose that fv(x) (cid:44)(cid:80)K
where gvt(xt) is as defined in (2) for q = 2. Then there exists C ∈ R s.t. ∀ n, P(At = 1Ft) ≤ Crt.
t (cid:107)2 ≤ rt}. We show the same two parts for Bt as for the proofs
Proof. Let Bt = I{∃k s.t. (cid:107)xk
for Lemma 2.
Part A, Bt = 0 =⇒ gvt(xt) = gt, for some gt ∈ ∂fvt(xt):
First, note that, when Bt = 0,
vt − xk
K(cid:88)k=1
(cid:107)xk − xk
v(cid:107)2
wk
v(cid:107)wv(cid:107)2
xm − xm
vt − xkm
(cid:107)xkm
vt
t (cid:107)2
∂mfvt(xt) = ∂m
=
wkm
(cid:107)wv(cid:107)2
Also if Bt = 0, then
where km is the subspace that contains the mth dimension
arg min
x (cid:34) K(cid:88)k=1
Then,
wk
(cid:107)wv(cid:107)2
(cid:107)xk − xk
v(cid:107)2 : (cid:107)x − xt(cid:107)2 ≤ rt(cid:35) = xt + rt(cid:34). . . ,
wkm
(cid:107)wv(cid:107)2
vt − xm
xm
vt − xkm
(cid:107)xkm
t (cid:107)2
, . . .(cid:35)
xt − arg minx[(cid:107)x − xvt(cid:107)2 : (cid:107)x − xt(cid:107)2 ≤ rt]
gvt(xt) =
Definition
∈ ∂fvt(xt)
Part B, ∃C ∈ R s.t. P(Bt = 1Ft) ≤ Crt:
rt
35
P(Bt = 1Ft) = P((cid:107)xvt − xt(cid:107)2 ≤ rtFt)
=(cid:90)x∈{x:∃k s.t. (cid:107)xk
=(cid:90)x∈{x:∃k s.t. (cid:107)xk
vt
≤ Cr2
≤ Crt
t
hXFt(x)dx
hX (x)dx
v drawn independent of history
−xk
t (cid:107)2≤rt}
−xk
t (cid:107)2≤rt}
bounded hX
rt ≤ 1 eventually
vt
for some C ∈ R < ∞. Note that C depends on K and M.
References
[1] Stéphane Airiau and Ulle Endriss. Iterated Majority Voting. In Proceedings of the 1st Interna-
tional Conference on Algorithmic Decision Theory, ADT '09, pages 38 -- 49, Berlin, Heidelberg,
2009. Springer-Verlag.
ISBN 978-3-642-04427-4. doi: 10.1007/978-3-642-04428-1_4. URL
http://dx.doi.org/10.1007/978-3-642-04428-1_4.
[2] Daniel Benjamin, Ori Heffetz, Miles Kimball, and Derek Lougee. The relationship between the
normalized gradient addition mechanism and quadratic voting. Public Choice, 172(1):233 -- 263,
July 2017.
ISSN 1573-7101. doi: 10.1007/s11127-017-0414-3. URL https://doi.org/10.
1007/s11127-017-0414-3.
[3] Daniel J. Benjamin, Ori Heffetz, Miles S. Kimball, and Nichole Szembrot. Aggregating Local
Preferences to Guide Marginal Policy Adjustments. The American economic review, 103(3):
605 -- 610, May 2013. ISSN 0002-8282. doi: 10.1257/aer.103.3.605. URL https://www.ncbi.
nlm.nih.gov/pmc/articles/PMC3760035/.
[4] Stephen Boyd and Almir Mutapcic. Subgradient methods. Lecture notes of EE364b, Stanford
University, Winter Quarter, 2007, 2006.
[5] Yves Cabannes. Participatory budgeting: a significant contribution to participatory democracy.
ISSN 0956-2478, 1746-0301. doi:
Environment and Urbanization, 16(1):27 -- 46, April 2004.
10.1177/095624780401600104. URL http://eau.sagepub.com/content/16/1/27.
[6] Ioannis Caragiannis and Ariel D. Procaccia. Voting almost maximizes social welfare despite
limited communication. Artificial Intelligence, 175(9):1655 -- 1671, June 2011. ISSN 0004-3702.
doi: 10.1016/j.artint.2011.03.005. URL http://www.sciencedirect.com/science/article/
pii/S0004370211000506.
[7] 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. URL
https://www.jair.org/media/5282/live-5282-9726-jair.pdf.
[8] Yukun Cheng and Sanming Zhou. A Survey on Approximation Mechanism Design With-
In David Gao, Ning Ruan, and Wenxun Xing, editors, Ad-
out Money for Facility Games.
vances in Global Optimization, number 95 in Springer Proceedings in Mathematics & Statistics,
36
pages 117 -- 128. Springer International Publishing, 2015. ISBN 978-3-319-08376-6 978-3-319-
08377-3. URL http://link.springer.com/chapter/10.1007/978-3-319-08377-3_13. DOI:
10.1007/978-3-319-08377-3_13.
[9] Hun Chung and John Duggan. Directional equilibria. Journal of Theoretical Politics, 30(3):
272 -- 305, July 2018. ISSN 0951-6298. doi: 10.1177/0951629818775515. URL https://doi.
org/10.1177/0951629818775515.
[10] J. C. Duchi, A. Agarwal, and M. J. Wainwright. Dual Averaging for Distributed Optimization:
Convergence Analysis and Network Scaling. IEEE Transactions on Automatic Control, 57(3):
592 -- 606, March 2012. ISSN 0018-9286. doi: 10.1109/TAC.2011.2161027.
[11] J. C. Duchi, M. I. Jordan, M. J. Wainwright, and A. Wibisono. Optimal Rates for Zero-
Order Convex Optimization: The Power of Two Function Evaluations. IEEE Transactions on
Information Theory, 61(5):2788 -- 2806, May 2015.
ISSN 0018-9448. doi: 10.1109/TIT.2015.
2409256.
[12] Abraham D. Flaxman, Adam Tauman Kalai, and H. Brendan McMahan. Online Convex
Optimization in the Bandit Setting: Gradient Descent Without a Gradient. In Proceedings of
the Sixteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA '05, pages 385 --
394, Philadelphia, PA, USA, 2005. Society for Industrial and Applied Mathematics.
ISBN
978-0-89871-585-9. URL http://dl.acm.org/citation.cfm?id=1070432.1070486.
[13] Nikhil Garg, Vijay Kamble, Ashish Goel, David Marn, and Kamesh Munagala. Collaborative
Optimization for Collective Decision-making in Continuous Spaces. In Proceedings of the 26th
International Conference on World Wide Web, WWW '17, pages 617 -- 626, Republic and Canton
of Geneva, Switzerland, 2017. International World Wide Web Conferences Steering Committee.
ISBN 978-1-4503-4913-0. doi: 10.1145/3038912.3052690. URL https://doi.org/10.1145/
3038912.3052690.
[14] Hollie Russon Gilman. Transformative Deliberations: Participatory Budgeting in the United
States. Journal of Public Deliberation, 8(2), 2012. URL http://search.proquest.com/
openview/4c3b7fd3ae888f483cacd031a749131e/1?pq-origsite=gscholar.
[15] Ashish Goel, Anilesh K. Krishnaswamy, Sukolsak Sakshuwong, and Tanja Aitamurto. Knapsack
Voting: Voting mechanisms for Participatory Budgeting. 2016. URL http://web.stanford.
edu/~anilesh/publications/knapsack_voting_full.pdf.
[16] Ashish Goel, Anilesh K. Krishnaswamy, and Kamesh Munagala. Metric Distortion of So-
cial Choice Rules: Lower Bounds and Fairness Properties.
In Proceedings of the 2017
ACM Conference on Economics and Computation, EC '17, pages 287 -- 304, New York, NY,
USA, 2017. ACM.
ISBN 978-1-4503-4527-9. doi: 10.1145/3033274.3085138. URL http:
//doi.acm.org/10.1145/3033274.3085138.
[17] Aanund Hylland and Richard Zeckhauser. A mechanism for selecting public goods when pref-
erences must be elicited. 1980.
[18] Kevin G.
Jamieson, Robert Nowak,
and Ben Recht.
derivative-free
Systems,
of
ing
4509-query-complexity-of-derivative-free-optimization.
In Advances
2012.
optimization.
2672 -- 2680,
pages
in Neural
Query
complexity
Information Process-
http://papers.nips.cc/paper/
URL
37
[19] Libin Jiang and Jean Walrand. Scheduling and Congestion Control for Wireless and Pro-
Synthesis Lectures on Communication Networks, 3(1):1 -- 156, January
ISSN 1935-4185, 1935-4193. doi: 10.2200/S00270ED1V01Y201008CNT006. URL
cessing Networks.
2010.
http://www.morganclaypool.com/doi/abs/10.2200/S00270ED1V01Y201008CNT006.
[20] Steven P. Lalley and E. Glen Weyl. Quadratic Voting. SSRN Scholarly Paper ID 2003531,
Social Science Research Network, Rochester, NY, December 2015. URL http://papers.ssrn.
com/abstract=2003531.
[21] David Timothy Lee, Ashish Goel, Tanja Aitamurto, and Helene Landemore. Crowdsourcing
for Participatory Democracies: Efficient Elicitation of Social Choice Functions.
In Second
AAAI Conference on Human Computation and Crowdsourcing, September 2014. URL https:
//www.aaai.org/ocs/index.php/HCOMP/HCOMP14/paper/view/8952.
[22] Patrice McDermott. Building open government. Government Information Quarterly, 27(4):
401 -- 413, October 2010. ISSN 0740-624X. doi: 10.1016/j.giq.2010.07.002. URL http://www.
sciencedirect.com/science/article/pii/S0740624X10000663.
[23] Reshef Meir, Maria Polukarov, Jeffrey S. Rosenschein, and Nicholas R. Jennings. Convergence
to Equilibria in Plurality Voting. In AAAI, volume 10, pages 823 -- 828, 2010.
[24] H. Moulin. On Strategy-Proofness and Single Peakedness. Public Choice, 35(4):437 -- 455, 1980.
ISSN 0048-5829. URL http://www.jstor.org/stable/30023824.
[25] Arkadi Nemirovski, Anatoli Juditsky, Guanghui Lan, and Alexander Shapiro. Robust stochastic
approximation approach to stochastic programming. SIAM Journal on optimization, 19(4):
1574 -- 1609, 2009.
[26] Ariel D. Procaccia and Jeffrey S. Rosenschein. The distortion of cardinal preferences in voting.
In International Workshop on Cooperative Information Agents, pages 317 -- 331. Springer, 2006.
URL http://link.springer.com/chapter/10.1007/11839354_23.
[27] Ariel D. Procaccia and Moshe Tennenholtz. Approximate Mechanism Design Without Money.
In Proceedings of the 10th ACM Conference on Electronic Commerce, EC '09, pages 177 -- 186,
New York, NY, USA, 2009. ACM. ISBN 978-1-60558-458-4. doi: 10.1145/1566374.1566401.
URL http://doi.acm.org/10.1145/1566374.1566401.
[28] David Quarfoot, Douglas von Kohorn, Kevin Slavin, Rory Sutherland, David Goldstein, and
Ellen Konar. Quadratic voting in the wild: real people, real votes. Public Choice, 172(1):
283 -- 303, July 2017. ISSN 1573-7101. doi: 10.1007/s11127-017-0416-1. URL https://doi.
org/10.1007/s11127-017-0416-1.
[29] Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of
ISSN 0003-4851, 2168-8990. doi:
Mathematical Statistics, 22(3):400 -- 407, September 1951.
10.1214/aoms/1177729586. URL http://projecteuclid.org/euclid.aoms/1177729586.
[30] Anwar Shah, editor. Participatory budgeting. Public sector governance and accountability se-
ries. World Bank, Washington, D.C, 2007. ISBN 978-0-8213-6923-4 978-0-8213-6924-1. OCLC:
ocm71947871.
38
[31] Naum Z. Shor. Nondifferentiable Optimization and Polynomial Problems, volume 24 of Noncon-
vex Optimization and Its Applications. Springer US, Boston, MA, 1998. ISBN 978-1-4419-4792-
5 978-1-4757-6015-6. doi: 10.1007/978-1-4757-6015-6. URL http://link.springer.com/10.
1007/978-1-4757-6015-6.
[32] Yves Sintomer, Carsten Herzberg, and Anja Rocke. From Porto Alegre to Europe: potentials
and limitations of participatory budgeting. International Journal of Urban and Regional Re-
search, 32(1):164 -- 178, 2008. URL http://www.cpa.zju.edu.cn/participatory_budgeting_
conference/english_articles/paper2.pdf.
[33] Volker Strassen. The Existence of Probability Measures with Given Marginals. The Annals of
Mathematical Statistics, 36(2):423 -- 439, April 1965. ISSN 0003-4851, 2168-8990. doi: 10.1214/
aoms/1177700153. URL https://projecteuclid.org/euclid.aoms/1177700153.
[34] T. Nicolaus Tideman and Florenz Plassmann. Efficient Collective Decision-Making, Marginal
Cost Pricing, And Quadratic Voting. SSRN Scholarly Paper ID 2836610, Social Science Re-
search Network, Rochester, NY, August 2016. URL https://papers.ssrn.com/abstract=
2836610.
39
|
1910.10748 | 1 | 1910 | 2019-10-23T18:08:39 | Dynamic multi-agent assignment via discrete optimal transport | [
"cs.MA",
"eess.SY",
"eess.SY",
"math.OC",
"stat.CO"
] | We propose an optimal solution to a deterministic dynamic assignment problem by leveraging connections to the theory of discrete optimal transport to convert the combinatorial assignment problem into a tractable linear program. We seek to allow a multi-vehicle swarm to accomplish a dynamically changing task, for example tracking a multi-target swarm. Our approach simultaneously determines the optimal assignment and the control of the individual agents. As a result, the assignment policy accounts for the dynamics and capabilities of a heterogeneous set of agents and targets. In contrast to a majority of existing assignment schemes, this approach improves upon distance-based metrics for assignments by considering cost metrics that account for the underlying dynamics manifold. We provide a theoretical justification for the reformulation of this problem, and show that the minimizer of the dynamic assignment problem is equivalent to the minimizer of the associated Monge problem arising in optimal transport. We prove that by accounting for dynamics, we only require computing an assignment once over the operating lifetime --- significantly decreasing computational expense. Furthermore, we show that the cost benefits achieved by our approach increase as the swarm size increases, achieving almost 50\% cost reduction compared with distance-based metrics. We demonstrate our approach through simulation on several linear and linearized problems. | cs.MA | cs | Dynamic multi-agent assignment via discrete optimal transport
Koray Gabriel Kachar* and Alex Gorodetsky**
9
1
0
2
t
c
O
3
2
]
A
M
.
s
c
[
1
v
8
4
7
0
1
.
0
1
9
1
:
v
i
X
r
a
Abstract -- We propose an optimal solution to a deterministic
dynamic assignment problem by leveraging connections to the
theory of discrete optimal transport to convert the combinato-
rial assignment problem into a tractable linear program. We
seek to allow a multi-vehicle swarm to accomplish a dynamically
changing task, for example tracking a multi-target swarm. Our
approach simultaneously determines the optimal assignment
and the control of the individual agents. As a result, the
assignment policy accounts for the dynamics and capabilities of
a heterogeneous set of agents and targets. In contrast to a ma-
jority of existing assignment schemes, this approach improves
upon distance-based metrics for assignments by considering
cost metrics that account for the underlying dynamics manifold.
We provide a theoretical justification for the reformulation of
this problem, and show that the minimizer of the dynamic
assignment problem is equivalent to the minimizer of the asso-
ciated Monge problem arising in optimal transport. We prove
that by accounting for dynamics, we only require computing
an assignment once over the operating lifetime -- significantly
decreasing computational expense. Furthermore, we show that
the cost benefits achieved by our approach increase as the
swarm size increases, achieving almost 50% cost reduction
compared with distance-based metrics. We demonstrate our
approach through simulation on several linear and linearized
problems.
I. INTRODUCTION
Our aim is to enable efficient centralized decision making
amongst swarms of agents that are tasked to intercept or track
a swarm of target vehicles. Specifically, we seek an optimal
centralized assignment policy that is capability-aware -- it
can leverage known dynamics of the agents and targets to
make optimal assignments that respect the capabilities of the
agents and targets. We approach this problem by posing an
objective function that accounts for both the high level cost of
all assignments and the low-level costs of the optimal control
policies used by each agent measures. We add differential
constraints arising from vehicle dynamics to complete the
optimization formulation. This approach stands in contrast to
the majority of techniques that use distance-based (or bottle-
neck assignment [7]) cost functions [26], [15], [30].
The approach we take in this work is based on the realiza-
tion of the close relationship between the given problem and
the theory of optimal couplings, or optimal transport [34],
[35]. In the context of probability theory, to which it is often
applied, optimal transport studies the problem of determining
joint distributions between sets of random variables whose
marginal distributions are constrained. In other words it tries
*Koray G. Kachar
is with the Department of Aerospace En-
gineering, University of Michigan, Ann Arbor, MI 48109, USA
[email protected]
**Alex Gorodetsky is an Assitant Professor, Aerospace Engineering, Uni-
versity of Michigan, Ann Arbor, MI 48109, USA [email protected]
to find a coupling that maps a reference measure to a target
measure. Optimal transport has been applied a wide variety
of other areas as well; for instance it has been used to great
effect in the areas of machine learning [9], [11], [19], image
manipulation [16], and Bayesian inference [14].
A. Innovation and Contributions
The fundamental insight we use to relate OT to the present
context is that the set agents may be viewed as a discrete
measure that we seek to map to the discrete measure denoted
by the set of targets. In this way, we consider discrete
optimal transport (DOT). Our context is also different from
the standard DOT problem in that the target measure is
changing and that
the transport of the reference to the
target must respect the differential constraints given by the
dynamics. Our innovation is that we can address these issues
by introducing a new metric that respects the dynamics, as
explored by Ghoussoub et. al. [21], rather than the traditional
unweighted Euclidean metric that underpins the Wasserstein
or "Earth Movers" distance. Our proposed metric uses the
optimal control cost of a single-agent vs. single-target system
as the cost of the proposed assignment. For instance if the
agents will perform LQR reference tracking to intercept
their targets then the LQR cost is used as the transportation
cost. Alternatively, if the agents will solve a pursuit evasion
game, then the transportation cost will be obtained from the
solution to the differential game. In this way, the assignment
becomes aware of the capabilities of the system, including
the differential constraints and the decision making approach
of individual agents.
Our problem is specified by two inputs
1) The dynamics of the agents and their targets
2) A mechanism to evaluate a feedback policy and its cost
for any single agent
Using these two specifications we form a cost function that
is the sum of all individual agent cost functions, and seek
an assignment that minimizes this total cost. Critically, we
see that the cost used for each agent is that of the feedback
policy -- not the distance. Typically, such feedback policies
are obtained to either optimally regulate or operate the
underlying agent. Thus, the cost incurred by an agent that is
following its feedback policy is a more appropriate measure
of optimality than one based on the distance an agent must
travel.
Our approach provides a solution for this problem, and
consists of the following contributions
1) A new capability-aware assignment approach that si-
multaneously optimizes the assignment and the under-
lying feedback controls of the agents
2) A reformulation of the vehicle-target dynamic as-
signment problem as a linear program by leveraging
concepts from discrete optimal transport
The above two contributions are supported via both theo-
retical and simulation results. In particular, we prove that
our cost function can be reformulated into the Monge
problem from optimal transport. This problem can then be
solved via a linear programming approach. The capability-
aware assignment problem is demonstrated to have lower
final cost as compared to a distance-based assignment that
neglects the feedback control policy. We empirically show
that the optimality gap between our approach and distance-
based metrics grows with the number of agents. Finally, we
prove that after formulating the assignment problem in the
DOT framework, it needs only be solved once rather than
repeatedly over the life of the system. As a result, we see
significant computational benefits compared to repeatedly re-
calculating a distance-based assignment.
B. Related work
Assignment and resource allocation problems present
themselves across many disciplines. In the area of multirobot
task assignment, self-organizing map neural-networks were
designed to learn how to assign teams of robots to target
locations under a dual decision-making and path-planning
framework [39]. However, the algorithm proposed in that
work is largely heuristic and does not consider the underlying
capabilities of the assigned robots. Other papers have consid-
ered more general kinematic movements of the formations
in general, rather than individual agent capabilities, and
were able to provide suboptimality guarantees for the overall
assignment [24]. Another approach can be found on [27] that
proposes an approach that is very similar to ours in that it
solves a related linear programming problem. However, that
approach did not consider the effect of general dynamics of
the system or a changing set of targets.
A similar assignment problem arises in from vehicle-
passenger transit scheduling that have become extremely im-
portant in ride sharing applications [10], [13]. Alonso-Mora
et.al [2] investigated dynamic trip-vehicle assignment that is
both scalable and optimal using a greedy guess that refines
itself over time. In general, these problems lack consideration
of underlying dynamics of the resource being assigned or that
task being assigned to. Assignment problems also arise in
wide areas of econometrics dealing with matching population
models to maximize total utility surplus, contract theory, or
risk management [20], [22], [25]. In general, these problems
also do not consider the underlying dynamic nature within
the assignment problem.
One closely related application area that, at times, also
considers the dynamics in completing an assignment is the
so-called weapon-target-assignment (WTA)
[23], [8]. The
WTA problem itself comes in two-forms: the static WTA
and the dynamic WTA. In the static WTA problem, all of
the assignments are made at once with no feedback possible,
whereas the dynamic WTA allows for feedback between
decision making phases [37]. Our approach is related to this
problem as it is a certain mixture of these two; first, it con-
siders explicit dynamic capabilities of the agents and targets
during the assignment problem; and second it potentially
allows for reassignment of the agents during operations. Our
setup can also be viewed as a limiting case of the traditional
WTA in that we assume that once the weapon intercepts
the target it successfully destroys it with 100% probability.
This contrasts to the traditional WTA setting where a weapon
might only have a certain probability of destroying its target.
The traditional WTA assignment problem (with proba-
bilistic survival of targets after interception) has typically
been formulated as a nonlinear integer programming problem
for which exact methods have not yet been found. As
a result, a large number of heuristic and/or approximate
approaches have been developed. For instance approaches
based on approximate dynamic programming [12], fuzzy
reasoning [33], various search heuristics [37], genetic and/or
global optimization approaches [28], [29], network-based
heuristics [1], amongst others have all been studied. In
comparison to these previous works on WTA we provide
several contributions. Our proposed (as far as we are aware
previously unrecognized) link to optimal transport theory can
yields additional theoretical and computational guarantees.
Finally, we review some connections between our pro-
posed approach and existing solutions in robotics and con-
trol. Fredrick et. al. [18] investigated multi-robot path plan-
ning and shape formation underpinned by optimal transport
to prove that the desired formations can be obtained while
maintaining collision-free motions with assurance of conver-
gence to global minima. Similarly, Bandyopadhyay et. al.
[5], [4], [6] describe an approach where swarm members are
placed into bins which have constraints that must satisfied
in order to permit the transition of the agents to neighboring
bins. These motion constraints are representative of the
dynamics or physical limitations present in the system. In
terms of the approach described in this paper, the optimal
transport cost metric is thus a modified L2 distance between
the centroids of the bins subject
to a motion constraint
matrix; if motion is possible, the cost is the L2 distance,
otherwise the cost is the maximum value.
Here we consider a specific setting, one with deterministic
and known dynamics, for which we can prove optimality.
While we do not consider limitations on communication
between agents, this problem has also been consider in the
decentralized decision making context where each vehicle
is making its own decisions. In this case all of the agents
must come to a consensus through different communica-
tion/negotiation strategies, see e.g. [38] for a greedy assign-
ment strategy and [3] for an example of a game-theoretical
formulation.
II. PROBLEM DEFINITION
In this section we define the dynamic agent-target assign-
ment problem. We begin by describing a dynamical system
that describes the evolution of active states, targets, and
destinations. We then provide an optimization problem that
we seek to solve.
A. Dynamical System
represented by a tuple
We limit our presentation to the case of linear, control-
affine systems for clarity of exposition. Our approach and
theory is also valid for nonlinear systems given an ability to
compute policy costs associated with nonlinear controllers.
Let n, m ∈ Z+ denote a positive number of autonomous
agents (resources) and targets. If we consider agent i and
target j, then their states at time t are denoted by xi(t) ∈
X ⊂ Rd, yj(t) ∈ Y ⊂ Rdy, respectively. Agent i takes
actions ui ∈ Rdu .
In our problem, the number of agents and targets can
only decrease with time. We leave consideration of newly
appearing targets to future work. Each object that has not
been removed is termed active so that at t = 0 we have n
active agents and m active targets.
An agent/target pair can become inactive when the agent
successfully intercepts or completes its resource allocation.
Let πx and πy define functions that extract the positions of
the agents and targets from their states. Successful resource
allocation of an agent is defined when (cid:107)πx(xi)−πy(yj)(cid:107) ≤ .
In other words, when the position of agent i is within an
−ball of target j, then both become inactive.
The activity of the agents and targets at each time is
represented by the active sets A(t) and T (t). For instance,
if all agents are active then A(t) = [1, . . . , n], whereas if,
for instance agent i has successfully reached target j then i
is removed from A and j is removed from T . This process
defines an evolution of the active sets. At a given time t,
the active agents and targets evolve according to a stochastic
differential equation
xi(t) = Aixi(t) + Biui(t)
yj(t) = Aiyj(t)
for i ∈ A(t)
for j ∈ T (t).
where Ai ∈ Rd×d and Bi ∈ Rd×du correspond to the
drift of linear dynamics of the agent; and Aj ∈ Rdy×dyu
corresponds to the closed loop linear dynamics of the targets.
Note that here we have assumed linear dynamics; however
this assumption is entirely unnecessary for our theory in
Section V. It is however, more computationally tractable
because it leads to a solution of sets of linear optimal control
problems, for instance LQR or LQI. A more significant
assumption that these dynamics imply is that there is no
interaction between agents, i.e., we do not consider collisions
or other interference effects. We leave this matter for future
work, but note that in the simulation examples in Section VI
we noticed that collisions between agents did not generally
occur.
A(t),T (t), (xi(t))i∈A(t) , (yj(t))j∈T (t)
Finally, the entire state of the system be defined by the
. Let
tuple S(t) =
O denote a set of states S(t) for which there is at least one
active target or destination. Define the exit time τ as the first
time that the state S(t) exists from O.
B. Policies, cost functions, and optimization
(cid:16)
(cid:17)
We seek a feedback policy µ that maps S(t) to a set of
controls ui(t) for all active agents i ∈ A(t). The policy is
µ(S(t)) = (σt, (µi : i ∈ A(t)))
(1)
where σt : A(t) → T (t) is an index function that assigns
active agents to active targets and µi is a feedback control
policy for the individual agents. The goal then, is to deter-
mine an optimal feedback policy of this form.
An optimal feedback policy is one that minimizes
µ∗ = arg min
µ
0
g(t, S(t), µ(S(t)))
,
(2)
where the stage cost mapping the state of the system to the
reals is g and the time τ is the first exit time. The optimal
value function will then be denoted by
(cid:21)
J(S(0)) =
g(t, S(t), µ∗(S(t))dt
(3)
(cid:20)(cid:90) τ
(cid:90) τ
0
The stage cost intends to guide each agent i to its assigned
target j and is therefore represented by the sum
g(t, S(t), µ(S(t))) =
gi(xi, ui, yσt(i)),
(4)
n(cid:88)
i=1,i∈A(t)
where the cost gi assigned to agent i is a function of the
corresponding agent state, the agent control, and the target
to which the agent is assigned. For instance, this cost could
be a quadratic corresponding to an infinite horizon tracking
problem [36]
(cid:0)πx(xi) − πy(yσt(i))(cid:1)T
(cid:0)πxss(xi) − πyss (yσt(i))(cid:1)T
Q(cid:0)πx(xi) − πy(yσt(i))(cid:1) + uT
Q(cid:0)πxss (xi) − πyss (yσt(i))(cid:1)
gi(xi, ui, yσt(i)) =
i Rui
−
+ uT
ssi
Russi
(5)
where Q and R penalize the distance between the agent-
target system and the control and πxss and πyss are the
steady-state values for the agent and assigned-to target,
respectively. These transient and steady-state terms represent
the dual goals of this particular optimal controller, which are
to drive the error of the agent-target to the optimal steady-
state and then keep the agent-target system at this optimal
state.
III. DISCRETE OPTIMAL TRANSPORT
In this section we provide background to discrete optimal
transport, and indicate how it relates to our dynamic assign-
ment problem. We follow the description given by [31].
Let Σn denote a probability simplex so that any a ∈ Σn
belongs to the simplex
(cid:41)
n(cid:88)
i=1
ai = 1
(6)
(cid:40)
Σn ≡
A discrete measure µX is defined by the fixed locations
x1, . . . , xn ∈ X and weights a ∈ Σn, and denoted by
µX =
aiδxi.
(7)
a ∈ Rn
+ :
n(cid:88)
i=1
µX = (cid:80)n
i=1 aiδxi and µY = (cid:80)m
The transport map T between two discrete measures,
j=1 bjδyj , is a surjective
function T : {x1, . . . , xn} → {y1, . . . , ym} that satisfies
(cid:88)
bj =
i,T (xi)=yj
ai ∀j = 1, . . . , m.
(8)
Compactly the above is written as T(cid:93)µx = µy, which implies
that measure µy is the push-forward measure of µx under
the map T .
A. Monge problem
We seek to find optimal assignments for the agent-target
system, and this implies that we seek a map T that minimizes
some the transportation cost. Let the transportation cost be
defined pointwise as
c(x, y) = c(x, T (x)),
∀(x, y) ∈ X × Y.
(9)
The Monge problem then seeks a map T that minimizes
(cid:35)
c(xi, T (xi)) : T(cid:93)µX = µY
.
(10)
To parameterize T, we can define an index function σ :
(cid:74)n(cid:75) →(cid:74)m(cid:75) so that T (xi) = yσ(i), just as in Equation (1).
The problem with optimizing Equation (10) is that it is
not-convex. In general, convexity can be achieved by relaxing
the deterministic nature of the map to allow portions of xi
to be directed towards yj. The resulting stochastic map is
defined by a coupling matrix (transition probability matrix, or
stochastic matrix) P ∈ Rn×m
with pij indicating the portion
of xi assigned to yj. Define a set of allowable coupling
(cid:8)P ∈ Rn×m
matrices
: P 1m = a and P T 1n = b(cid:9) ,
(cid:88)
where 1k denotes a vector of size k filled with ones. The
Monge-Kantorovich optimization formulation then becomes
U(a, b) ≡
+
+
c(xi, yj)pij,
(11)
Lc(a, b) = min
P∈U (a,b)
i,j
and it can be solved with linear programming. Under the con-
ditions given next, the solution to this optimization problem
is equal to the solution of the Monge problem.
B. Matching problem
The matching problem is a particular realization of OT that
has the property that the minimizer of (11) is equal to that
of (10). The formal statement of this equivalence is given
below.
Proposition 1 (Kantorovich for matching (Prop 2.1, [31])):
If n = m and a = b = 1n/n then there exists an optimal
solution for minimizing Equation (11) Pσ∗, which is a
permutation matrix associated to an optimal permutation
σ∗
∈ Perm(n) for Equation (10).
In this setting, we seek a a one-to-one mapping T . The
constraint set U(a, b) becomes the set of doubly stochastic
matrices with entries, and the coupling matrix has elements
(cid:34) n(cid:88)
i=1
min
T
(cid:26) 1/n
0
pij =
if j = σ(i),
otherwise.
(12)
In the context of our assignment problem, this case occurs
when there are an equal amount of agents and targets. A
discrete optimal transport formulation can also be applied
for a relaxation of the Kantorovich problem so that several
agents can be assigned to the same target. This problem can
also guarantee binary coupling matrices (essential for our
application). For further details, we refer to [16].
C. Metrics
The choice of cost c(xi, yj) is problem dependent; how-
ever, the most commonly used cost for optimal transport be-
tween distributions is the Euclidean distance. Parameterized,
by α ≥ 1, it is given by
c(xi, yj) = (cid:107)xi − yj(cid:107)α,
(13)
where (cid:107)·(cid:107) is the Euclidean norm. Using this metric for points,
Lc(a, b) can be viewed as a metric for measures. In this
case, Lc is called the Lα-Wasserstein distance [16]. In the
statistical community, this metric is also called the Earth
movers distance (EMD).
This metric implies that the cost of moving a resource xi
to yj is dominated by the distance between them in Euclidean
space. The total cost of the assignment then becomes a
sum of the distances. In our application to assignment in
dynamical systems, the Euclidean metric may not be most
appropriate because it does not account for the underlying
dynamics of the system. One of our insights is that using
a metric determined by the underlying dynamics of the
problems leads to more optimal assignments.
IV. ASSIGNMENT IN DYNAMIC SYSTEMS WITH DOT
In this section we describe how DOT can be applied to
minimize Equation (2). As previously stated, our goals are to
determine the assignment policy σ that is "capability-aware."
In other words, the assignment policy must account for the
dynamics of the system -- the capabilities of the agents and
the targets.
A direct application of the EMD metric within DOT would
potentially require constant reassignment at each timestep
because the metric makes no accountability of the future
system state. In other words, it would be greedy and simply
assign each agent to minimize the total distance between
agent/target pairs.
In the next two subsections we first describe an algorithm
that leverages the knowledge of the interception strategies
of each agent to make an assignment, and then we provide
and discuss pseudocode to illustrate the flexibility of our
approach .
A. Algorithm
The metric we propose for the transportation cost of
assigning xi
is that corresponding to the optimal
actions of a one-agent-to-one-target optimization problem.
For instance, let us assume that agent i is paired to target
to yj
j = σ(i), then in the 1v1 scenario for policy µi, we have a
total incurred cost of
J µi
iσ(i)(xi(0),yσ(i)(0)) =
(cid:90) τi
gi(t, xi, µi(t, xi(t), yσ(i)(t)), yσ(i))dt
0
The optimal policy is obtained by minimizing this value
function.
Let Jiσ(i) = minµ J µi
iσ(i) correspond to the value function
under the optimal policy. Our proposed transportation cost
is this optimal value function
cdyn(xi(0), yσ(i)(0)) = Jiσ(i)(xi(0), yσ(i)(0)).
(15)
For example, for linear dynamics with quadratic cost the
transportation becomes
cdyn(xi(0), yσ(i)(0)) =
f (xi(0), yσ(i)(0))T Piσ(i)f (xi(0), yσ(i)(0))+
2piσ(i)f (xi(0), yσ(i)(0))−
f (xi(0), yσ(i))T
2piσ(i)f (xi(0), yσ(i))ss
ssPiσ(i)f (xi(0), yσ(i))ss−
(16)
the distance directly by extracting the positions (pos, but an
entire state can also be used), and discards gpol -- the cost
of the actual policy. As a result, this assignment needs to
be continuous recomputed. In Section V we prove that our
approach only requires an assignment to be generated once.
(14)
Algorithm 1 Simulation engine with dynamics-based assign-
ment and LQT tracking
Inputs n agents A, states X = (xi)i∈A, dynamics (fi)i∈A;
Outputs Completion of simulation, all targets are tracked
n targets Y = {y1, . . . , yn}, dynamics (fTi)n
i=1 = assign(X, Y, (fi)n
i−1, LQT, distdyn)
i=1, (fTi )n
σ, (µi)n
i=1
while t ≤ tf do
for (i, xi) ∈ A do
j = σ(i)
yj(t) ← Get target state from environment
ui(t) = µi(t, xi, yj)
x∗
i (t + dt) = ODEsolve(t, t + dt, fi, xi, ui)
A = remove-tracked(A, (xi)n
i=1, (yi)n
j=1)
end for
end while
where f is a function that combines the agent and target
state into suitable form. For instance f (x, y) = x − y can
be used for a reference tracking problem [36]; Piσ(i) is the
solution of the continuous algebraic Riccati equation for the
LQR-based tracker; piσ(i) is the feed-forward control of the
state being tracked; and f (xi(0), yσ(i))ss is a function that
provides the steady-state value for the quadratic agent-target
state.
B. Pseudocode
In this section we provide and describe the pseudocode
for the proposed algorithm. A sample implementation that
makes specific choices about the dynamics and policies is
shown in Algorithm 1. This algorithm takes as inputs all
of the agent states, target states, and dynamics. In Line 1,
the assignment and individual agent policies are obtained by
querying Algorithm 2.
Algorithm 2 performs the optimal transport allocation. Its
inputs are all of the states and their dynamics, an algorithm
for computing the policies for each agent when it is assigned
to some state, and a cost metric. Algorithm 1 makes two
specific choices for these components. First, it uses the linear
quadratic tracker (LQT) developed in [36] that uses linear
dynamics. However, if the dynamics are nonlinear, any other
computable policy can be used. The specific cost metric
is distdyn, which is the dynamics-based distance given
by Algorithm 3. This algorithm, uses the cost of the LQT
policy (16) as the transportation cost. Algorithm 2 has two
steps. First it calls the discrete optimal transport routine with
a pre-specified distance metric to obtain an assignment σ. It
then iterates through all agents and obtains the individual
policy for each agent that follows the assignment.
The high-level Algorithms 1 and 5 demonstrate the differ-
ences between our approach and the standard approach that
uses the distance metric (Algorithm 4). Algorithm 4 evaluates
Algorithm 2 Assignment policy: assign
Inputs Set of agents X = (xi)n
i=1; Set of targets Y =
i=1; agent dynamics fi; target dynamics fTi; control
(yi)n
policy generator pol; 1 vs. 1 function cost calculator
dist
Outputs An assignment policy σ; agent policies (µi)n
1: σ = DOT(X, Y, dist)
2: for i = 1 . . . n do
3:
4: end for
5: return σ, (µi)n
gi, µi = dist(xi, yσi, fi, fTσi
)
i=1
i=1
Algorithm 3 Dynamics-based cost: distdyn
Inputs Agent state x; target state y; agent dynamics f; target
dynamics fT ; control policy generator pol
Outputs The cost g and the optimal policy µ for assigning
agent x to target y
1: g, µ = pol(x, y, f, fT )
2: return g, µ
V. ANALYSIS
In this section we analyze the proposed algorithms. Our
aim is to show that the optimization problem (2) can be
reformulated into the Monge-Kantorovich optimal transport
problem. We follow a two step procedure. First we show
that the optimal assignment policy of Equation (1) does not
change with time, and next we show that the problem is
identical to the Monge problem.
Proposition 2 (Constant assignment policy): The optimal
assignment policy σt for minimizing (2) is fixed over time.
Algorithm 4 Distance-based cost: distemd
Inputs Agent state x; target state y; agent dynamics f; target
dynamics fT ; control policy generator pol
Outputs The cost g and the optimal policy µ for assigning
agent x to target y
1: gpol, µ = pol(x, y, f, fT )
2: g = (cid:107)pos(x) − pos(y)(cid:107)
3: return g, µ
Algorithm 5 Simulation engine with distance-based assign-
ment and LQT tracking
Inputs n agents A, states X = (xi)i∈A, dynamics (fi)i∈A;
Outputs Completion of simulation, all targets are tracked
while t ≤ tf do
X, Y ← extract-active-agents(A)
σ, (µi)n
n targets Y = {y1, . . . , yn}, dynamics (fTi)n
i=1 = assign(X, Y, (fi)n
i−1, LQT, distemd)
i=1, (fTi )n
i=1
for (i, xi) ∈ A do
j = σ(i)
yj(t) ← Get target state from environment
ui(t) = µi(t, xi, yj)
x∗
i (t + dt) = ODEsolve(t, t + dt, fi, xi, ui)
i=1, (yi)n
A = remove-tracked(A, (xi)n
j=1)
end for
end while
In other words, if σtc(i) = j for a fixed i and j at time tc,
then for any t > tc we also have σtc(i) = j.
Proof: We consider the case of two agents first,
and then extend our approach to multiple agents through
induction. We start with the case of two agents x1 and x2 and
two targets y1 and y2. We will compare two policies: a time
varying policy that includes at least one switch µ(2)(t) =
(σt, µ1, µ2), and a second policy µ(2) = (σ, µ1, µ2) that does
not switch and where µ1 and µ2 minimize Equation (14).
We first consider the case where µ(2)(t) both contains
switches and the final assignment is equivalent to the initial
assignment. In this case let t1 denote the time the first switch
occurs and t2 denote the time of the final switch back to the
original assignment. Without loss of generality, assume that
the initial assignment is given by σ0(i) = i. Let A(t) denote
the active agents under the policy and τ and τi denote the
exit times. The cost associated with µ(2) is
Jµ(2)(S(0)) =
gi(xi, ui, yσt(i))dt
(cid:90) τ
n(cid:88)
0
i=1,i∈ A(t)
Letting δi∈A(t) denote an indicator function, we can rewrite
the total cost as a sum over each agent
Jµ(2)(S(0)) =
δi∈A(t)gi(xi, µi(xi), yσt(i))dt
(17)
(cid:90) τ
2(cid:88)
i=1
0
t2, and after t2. Denoting τi denote the exit time of agent i
we have
Jµ(2)(S(0)) =
gi(xi, µi(xi), yσ(i))dt+
2(cid:88)
i=1
(cid:104)(cid:90) t1
(cid:90) t2
0
gi(xi, µi(xi), yσt(i))dt+
gi(xi, µi(xi), yσ(i))dt
(cid:105)
t1
(cid:90) τ1
t2
(cid:80)n
(cid:82) τ1
Finally, suppose that µ(2) = µ(2)(0) = (σ0, µ1, µ2) is a
policy that maintained the original assignment. In this case,
because µ1 and µ2 are optimized for the original assignment,
they clearly result in policies that have lower costs than µ1
and µ2 incurred during the time period from t1 and t2. In
other words, since each agent ends up targeting the same
target that it initially targeted, it is at least more effective
to directly follow the policy to the target
than to have
intermediate deviations to the other target, i.e.,
i=1
Jµ(2)(S(0)) ≥
0 gi(xi, µi(xi), yσ(i))dt = Jµ(2)(S(0)).
An identical argument follows for the case where the final
policy is different than the initial policy. In that case, we
would set µ(2) = µ(2)(τ ).
The case for more than two agents and targets follows
by noticing that any system of n agents can be analyzed
by considering a system of two modified agents. The first
modified agent is the augmentation of the first n − 1, and
the last one is the nth agent. Then the scenario is identical
to the 2v2 assignment and the same argument follows.
Now that we have shown that the optimal assignment is
time independent, we can show that the minimizer of our
our stated optimization problem (2) is the same as that of
the Monge problem (10).
Theorem 1 (Optimization problem equivalence): The op-
timal solution of the assignment problem given by the
optimization problem (2) is equivalent to that obtained by
minimizing the Monge problem (10) when the 1v1 cost
function (15) is used as the transportation cost.
Proof: We use the fact that the optimal policy maintains
a fixed index assignment vector for all time, i.e., σt = σ. Let
S(0) denote the initial state of the system, then the cost for
any initial state can be represented as
gi(xi, ui, yσ(i))dt
δi∈A(t)gi(xi, ui, yσ(i))dt
i=1,i∈A(t)
n(cid:88)
(cid:90) τ
(cid:90) τi
0
0
(cid:90) τ
n(cid:88)
n(cid:88)
n(cid:88)
i=1
i=1
i=1
J(S(0)) =
=
=
≥
gi(xi, ui, yσ(i))dt
0
cdyn(xi(0), yσ(i)(0))
Now we can break up the integral into three section corre-
sponding to the cost before the switch at t1, between t1 and
where the first equality came from Equations (3) and (4) the
second equality follows the same argument as Equation (17);
the third equality follows from the definition of τi; and
the final inequality follows from the definition of cdyn in
Equation (15). Because of the definitions of 1v1 exit times τi,
we implicitly the cost function to only those policies where
the agents reach their targets, i.e., T(cid:93)µA = µT where µA is
the initial distribution of the agents and µT is the distribution
of the targets at interception. Strict equality is obtained when
the policies µi correspond to the optimal policies 1v1 policies
that minimize (14) so that ui(t) are generated by µi. Thus,
we have proved the stated result.
VI. SIMULATION RESULTS
We now numerically demonstrate the effectiveness of our
approach through several simulated examples. In each exam-
ple, we have used the Python Optimal Transport library [17]
to solve the underlying DOT problem. In each case, the
dynamics are integrated via the RK45 integration scheme.
A. Double integrators in three dimensions
In this section we demonstrate that using the dynamics-
based cost function over the standard distance-based Wasser-
stein metric yields significant savings that increases with size
of the system. For the various examples we will consider
agent/target systems of sizes 5 vs. 5, 10 vs. 10, 20 vs. 20
and 100 vs. 100.
This set of examples uses a simple system of double
integrators in three dimensions, where the velocity term is
directly forced. The evolution of the state of each agent
xi = (xi1, xi2, xi3, xi4, xi5, xi6) is given by
for j = 1, 2, 3 and
xij = xi(j+3)
xij = ui(j−3)
for j = 4, 5, 6, where each agent has three control inputs
(one for each dimension). The target dynamics are identical
to the agent dynamics.
Each agent uses an infinite horizon linear-quadratic track-
ing policy of [36] where the stage cost of an assignment is
given by
g(xi, ui, yσ(i) = eiσ(i)(t)T Qieiσ(i)(t) + ui(t)T Riui(t)−
Riussi )
(18)
Qiessiσ(i) + uT
ssi
(eT
ssiσ(i)
where eiσ(i)(t) and essiσ(i) are defined as the transient error
and steady-state error between the ith agent state and an
assigned-to target state, respectively; ui is the control input
to drive the ith agent to the assigned-to target; and ussi
is the control input for the ith agent to the assigned-to
target operating at steady-state conditions. For the weight
matrices we choose Qi = diag(103, 103, 103, 0, 0, 0), where
the nonzero weights correspond to the errors in positions in
each dimension and the zero weights correspond to the errors
in velocity. The control penalty is chosen to be Ri = I3×3.
The targets use an identical tracking policy; however they
track certain fixed positions in space.
The initial conditions of the system consist of the positions
and velocities of each agent and target, a set of stationary
locations that are tracked by the targets, and a set of
assignments from each target to the the stationary location.
These conditions are randomly generated for the following
results.
The initial conditions of the agents consist of uniformly
distributed position and velocity components on an interval
of −1000 to 1000 and −5000 to 5000, respectively. The
initial conditions of the targets position are equivalent, but
with velocity parameters following a uniform distribution
between −1000 to 1000. The terminal target locations were
randomly selected on a uniform distribution between −1000
and 1000.
Fig. 1. Normalized costs incurred by 100 agents tracking 100 targets.
The cumulative costs over time for the assignment policy that uses EMD is
shown by the solid cyan line. The cost of this policy exceeds the optimal
cost given by the dotted black line. This optimal cost is computed by
summing the optimal costs of the value function for each agent under the
optimal assignment. The cumulative costs of the dynamics-based assignment
approaches this optimal value, as expected.
In Figure 1, we show the cumulative control costs incurred
by a system of 100 agents while they are attempting to
tracking 100 targets. Recall that the EMD-based objective
assigns agents to targets with the aim of minimizing the total
Euclidean distance. This assignment does not account for
the dynamics of the agent and as a result, it performs worse
than the dynamics-based assignment which accounts for the
effort to actually get the agent to its assigned target. Me-
chanically, this performance difference results because agents
are either incorrectly assigned at the beginning or because
agents switch assignments over the course of their operations.
For this simulation, the EMD-based policy checks whether
reassignment is necessary every 0.1 seconds.
Because visualizing the movements of 100 agents and
targets is difficult, we demonstrate prototypical movements
for a 5 vs. 5 system in Figures 2 and its X-Y projection 3
These figures demonstrate both the optimal trajectories of
the agents and targets under the dynamics-based optimal
assignment and the sub-optimal trajectories of the EMD-
based assignment. Agents A0 and A1, for example, take
significantly different paths to different targets. The move-
ments corresponding to the EMD-based policy require more
manuevering.
012345Time(s)0.000.250.500.751.001.25NormalizedCostOptimalcostCum.StageCostEMDCum.StageCostDynCost-to-goDyn The dynamics-based policy leverages the dynamic capa-
bilities of the agents to select the targets that each individual
would optimally be able to track over time. Finally, note that
the individual agent controllers that we use are fundamentally
tracking controllers, thus the agents act to match the velocity
and position of their targets. This is the reason why several
maneuvers show the agent passing and then returning to the
target -- for instance A1 to T3 under the EMD policy.
Fig. 2. Trajectories of agents and targets in a 5 vs 5 system. The trajec-
tories for an assignment policy that accounts for the dynamics (Dyn) are
qualitatively different than the assignment policy that uses the Wasserstein
distance (EMD). For labels of each path see the X-Y projection in Figure 3.
Fig. 3. Projection of the trajectories of Figure 2 onto the X-Y plane.
Since the dynamics-based assignment policy selects an
optimal assignment at initial time, it offers significant control
cost benefits over assignments that continually reassign the
agents based on the EMD.
The benefit of dynamics-based assignment grows with the
size of the system. To demonstrate this fact we use perform
Monte Carlo simulations of one hundred realizations of a 5
vs 5, 10 vs 10, and 20 vs 20 system by sampling over initial
conditions. As the complexity of the engagement increases,
the amount of additional control effort required by the EMD
based assignment grows, shown by Figure 4.
Furthermore, Figure 5 illustrates that as the system size
grows the EMD-based policy performs more switches. This
fact contributes to the observed loss in efficiency of the
EMD-based policy.
Fig. 4.
The EMD-based assignment policy becomes increasingly less
effective as the size of the system grows. Histograms of 100 Monte Carlo
simulations obtained by sampling initial conditions for various system sizes
are shown. As the system size increases, the distribution of the difference
between the control effort of the EMD and Dyn based assignments increases.
Fig. 5.
Monte Carlo simulations for 5v5, 10v10, and 20v20 systems
reveal that the average number of Agent-Target assignment switches for
EMD-based assignments positively correlates with the size of the systems.
B. Linearized Quadcopter
We now compare the algorithms on swarms of linearized
quadcopter dynamics [32] that are slightly modified versions
of double integrators. The dynamics of both the agents and
the targets in this case are given by
x = u
y = v
z = w
ψ = r
θ = q
φ = p
u = −gθ + fwx
m
v = gφ + fwy
m
w = fwx−ft
p = τx+τwx
q = τy+τwy
r = τz+τwz
Ixx
m
Iyy
Izz
f (x, u, d) =
x =(cid:2)x y
where the twelve dimensional state space
z ψ θ φ u v w p q
r(cid:3)T
∈ R12
x−100001000y−100001000z−100001000A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4−100001000x−1000−50005001000yA0T0A1T1A2T2A3T3A4T4C0C1C2C3C4A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4AgentPath(Dyn)TargetPathAgentStartTargetStartTerminalStateAgentPath(EMD)0.00.20.40.60.8ControlExpenditureDifference(EMD-Dyn)/Dyn010203040Frequency5v510v1020v205v510v1020v20Agents02468Average#Assign.Switches consists of the position, attitude, translational velocity, and
rotational velocity components of the vehicle. The parame-
ters of the system are m = 0.1kg, Ixx = 0.00062kg-m2,
Iyy = 0.00113kg-m2, Izz = 0.9(Ixx + Iyy) and g =
9.81m/s2, respectively. Linearization was performed under
small oscillation and small angle approximations. Further-
more, we will assume no wind disturbance forces and
torques, d = [fwx, fwy, fwz, τwx, τwy, τwz] = 0. The control
inputs are four dimensional and consist of the forces and
torques u = [ft, τx, τy, τz] that act on the vertical thrust and
angular motions about the three principal axes.
The initial positions and velocities of the agents are
sampled uniformly between −100 to 100 and −500 to
500, respectively. The initial velocities of the targets were
sampled uniformly between −50 to 50. The attitude and
rotational velocity terms for both agents and targets were
uniformly distributed between −2π and 2π and −25 and
25, respectively. The terminal target locations were randomly
selected from a uniform distribution between −100 and 100.
The control parameters for the agents and targets are updated
to Qi = diag(103, 103, 103, 103, 103, 103, 0, 0, 0, 0, 0, 0) and
Ri = I4x4, respectively.
Similar to the double integrator systems, the dynamics-
based assignment policy is able to optimally assign the more
complex quadcopter agents to complete their tracking task
with minimal cost. Figure 6 illustrates the cumulative cost
expended by the agent swarm and once again showcases the
optimality of the dynamics-based assignment method. Unlike
the EMD policy, the complete dynamic information of the
swarm members are used in the decision-making process as
opposed to only the euclidean distance components. In the
end, the EMD-based assignment policy incurs a cost that is
1.7 times greater than the dynamics-based assignment policy.
Figures 7 and Figures 8 reveal the paths taken by the
agents managed by the EMD and Dyn policies. Agents 1
and 4, in particular, are allowed to take advantage of their
initial dynamic states to cheaply track their targets, instead
of being reassigned (by the EMD-based policy) mid-flight
to closer targets that appear. In this case, the reassignment
causes extreme turning maneuvers that require significant
control expense.
Since the linearized quadcopter operates over a d = 12
statespace, the computational cost for performing assign-
ments are more expensive, and since the EMD-based policy
requires checking and updating assignments every time incre-
ment, it requires significantly greater computational expense.
For this problem, the total cost of all reassignments required
0.6 seconds by the EMD policy, a signification porition of
the total simulation time of five seconds.
Fig. 6.
Normalized costs incurred by 5 linearized quadcopter agents
tracking 5 linearized quadcopter targets. The cumulative costs for the
EMD policy exceeds the optimal cumulative costs of the dynamics-based
policy for a system operating realistic dynamics. The dynamics-based policy
continues to settle at the optimal value.
Fig. 7.
Trajectories for a 5 vs 5 system operating linearized quad-
copter dynamics. The dynamics-based policy accounts for the full dynamic
capability of the agents in its assignment, rather than solely relying on
spatial proximity information. This includes leveraging the rotational and
translational information of the vehicle in the decision process.
Fig. 8. Projection of the trajectories of Figure 7 onto the X-Y plane. Agents
1 and 4 are able to use their initial dynamic conditions to optimally track
their targets instead of performing expensive turning maneuvers.
VII. CONCLUSION
In this paper we have demonstrated how to reformulate
a dynamic multi-vehicle assignment problem into a linear
program by linking this problem with the theory of optimal
transport. This theory allows us to prove optimality and to
increase the system efficiency using our approach. In the end,
we have developed an assignment approach that is capability-
aware. The assignment accounts for the capabilities of all the
agents and targets in the system.
One direction of future research is the incorporation of
constraints amongst the various agents to avoid collisions
or other interactions. An extension of DOT theory in this
direction could greatly increase the tractability of numer-
012345Time(s)0.00.51.01.52.0NormalizedCostOptimalcostCum.StageCostEMDCum.StageCostDynCost-to-goDynx−1000100y−1000100z−1000100A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4−50050100x−100−50050100yA0T0A1T1A2T2A3T3A4T4C0C1C2C3C4A0T0A1T1A2T2A3T3A4T4C0C1C2C3C4AgentPath(Dyn)TargetPathAgentStartTargetStartTerminalStateAgentPath(EMD) ous multi-agent swarm operations, for example large scale
formation flight. Another direction for future research is
the incorporation of stochastic dynamics and partial state
information. For either case, the approach described in this
paper can be used as the basis of a greedy or approximate
dynamic programming approach that is traditionally used
for these problems. Finally, we can incorporate learning
into the program where the agents periodically update their
knowledge about the intent of the targets.
VIII. ACKNOWLEDGMENTS
We would like to thank Tom Bucklaew and Dustin Martin
of Draper Laboratory for their helpful guidance and vision
in support of this project. This research has been supported
by Draper Laboratory, 555 Technology Square, Cambridge,
MA 02139.
REFERENCES
[1] Ravindra K Ahuja, Arvind Kumar, Krishna C Jha, and James B
Orlin. Exact and heuristic algorithms for the weapon-target assignment
problem. Operations research, 55(6):1136 -- 1146, 2007.
[2] Javier Alonso-Mora, Samitha Samaranayake, Alex Wallar, Emilio
Frazzoli, and Daniela Rus. On-demand high-capacity ride-sharing
via dynamic trip-vehicle assignment. Proceedings of the National
Academy of Sciences, 114(3):462 -- 467, 2017.
[3] Gurdal Arslan, Jason R Marden, and Jeff S Shamma. Autonomous
vehicle-target assignment: A game-theoretical formulation. Journal of
Dynamic Systems, Measurement, and Control, 129(5):584 -- 596, 2007.
[4] Saptarshi Bandyopadhyay. Novel probabilistic and distributed algo-
rithms for guidance, control, and nonlinear estimation of large-scale
multi-agent systems. PhD thesis, University of Illinois at Urbana-
Champaign, 2016.
[5] Saptarshi Bandyopadhyay, Soon-Jo Chung, and Fred Y Hadaegh.
Probabilistic swarm guidance using optimal transport. In 2014 IEEE
Conference on Control Applications (CCA), pages 498 -- 505. IEEE,
2014.
[6] Saptarshi Bandyopadhyay, Soon-Jo Chung, and Fred Y Hadaegh.
Probabilistic and distributed control of a large-scale swarm of au-
tonomous agents. IEEE Transactions on Robotics, 33(5):1103 -- 1123,
2017.
[7] Rainer E Burkard. Selected topics on assignment problems. Discrete
Applied Mathematics, 123(1-3):257 -- 302, 2002.
[8] Huaiping Cai, Jingxu Liu, Yingwu Chen, and Hao Wang. Survey of
the research on dynamic weapon-target assignment problem. Journal
of Systems Engineering and Electronics, 17(3):559 -- 565, 2006.
[9] Guillermo Canas and Lorenzo Rosasco. Learning probability measures
In Advances in Neural
with respect to optimal transport metrics.
Information Processing Systems, pages 2492 -- 2500, 2012.
[10] Avishai Avi Ceder. Optimal multi-vehicle type transit timetabling and
vehicle scheduling. Procedia-Social and Behavioral Sciences, 20:19 --
30, 2011.
[11] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal
transport. In Advances in neural information processing systems, pages
2292 -- 2300, 2013.
[12] Michael T Davis, Matthew J Robbins, and Brian J Lunday. Ap-
proximate dynamic programming for missile defense interceptor fire
control. European Journal of Operational Research, 259(3):873 -- 886,
2017.
[13] Gonc¸alo Homem de Almeida Correia and Bart van Arem. Solving the
user optimum privately owned automated vehicles assignment problem
(uo-poavap): A model to explore the impacts of self-driving vehicles
on urban mobility. Transportation Research Part B: Methodological,
87:64 -- 88, 2016.
[14] Tarek A El Moselhy and Youssef M Marzouk. Bayesian inference with
optimal maps. Journal of Computational Physics, 231(23):7815 -- 7850,
2012.
[15] Jan Faigl, Miroslav Kulich, and Libor Preucil. Goal assignment
In 2012 IEEE/RSJ
using distance cost in multi-robot exploration.
International Conference on Intelligent Robots and Systems, pages
3741 -- 3746. IEEE, 2012.
[16] Sira Ferradans, Nicolas Papadakis, Gabriel Peyr´e, and Jean-Franc¸ois
SIAM Journal on
Aujol. Regularized discrete optimal transport.
Imaging Sciences, 7(3):1853 -- 1882, 2014.
[17] R´emi Flamary and Nicolas Courty. Pot: Python optimal transport
library, 2017.
[18] Christina Frederick, Magnus Egerstedt, and Haomin Zhou. Multi-
robot motion planning via optimal transport theory. arXiv preprint
arXiv:1904.02804, 2019.
[19] Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya,
and Tomaso A Poggio. Learning with a wasserstein loss. In Advances
in Neural Information Processing Systems, pages 2053 -- 2061, 2015.
[20] Alfred Galichon. Optimal transport methods in economics. Princeton
University Press, 2018.
[21] Nassif Ghoussoub, Young-Heon Kim, and Aaron Zeff Palmer. Optimal
transport with controlled dynamics and free end times. SIAM Journal
on Control and Optimization, 56(5):3239 -- 3259, 2018.
[22] Bryan S Graham. Econometric methods for the analysis of assignment
problems in the presence of complementarity and social spillovers. In
Handbook of social economics, volume 1, pages 965 -- 1052. Elsevier,
2011.
[23] Patrick A Hosein and Michael Athans. Some analytical results for
the dynamic weapon-target allocation problem. Technical report,
MASSACHUSETTS INST OF TECH CAMBRIDGE LAB FOR IN-
FORMATION AND DECISION SYSTEMS, 1990.
[24] Meng Ji, Shun-ichi Azuma, and Magnus B Egerstedt. Role-assignment
in multi-agent coordination. 2006.
[25] Roy Jonker and Anton Volgenant. A shortest augmenting path algo-
rithm for dense and sparse linear assignment problems. Computing,
38(4):325 -- 340, 1987.
[26] Stephen Kloder and Seth Hutchinson. Path planning for permutation-
IEEE Transactions on Robotics,
invariant multirobot formations.
22(4):650 -- 665, 2006.
[27] Kwan S Kwok, Brian J Driessen, Cynthia A Phillips, and Craig A
Tovey. Analyzing the multiple-target-multiple-agent scenario using
optimal assignment algorithms. Journal of Intelligent and Robotic
Systems, 35(1):111 -- 122, 2002.
[28] Zne-Jung Lee, Chou-Yuan Lee, and Shun-Feng Su. An immunity-
based ant colony optimization algorithm for solving weapon -- target
assignment problem. Applied Soft Computing, 2(1):39 -- 47, 2002.
[29] Zne-Jung Lee, Shun-Feng Su, and Chou-Yuan Lee. Efficiently solving
general weapon-target assignment problem by genetic algorithms
IEEE Transactions on Systems, Man, and
with greedy eugenics.
Cybernetics, Part B (Cybernetics), 33(1):113 -- 121, 2003.
[30] Dimitra Panagou, Matthew Turpin, and Vijay Kumar. Decentralized
goal assignment and trajectory generation in multi-robot networks.
arXiv preprint arXiv:1402.3735, 2014.
[31] Gabriel Peyr´e and Marco Cuturi. Computational optimal transport.
Foundations and Trends R(cid:13) in Machine Learning, 11(5-6):355 -- 607,
2019.
[32] Francesco Sabatino. Quadrotor control: modeling, nonlinearcontrol
design, and simulation, 2015.
[33] Mehmet Alper S¸ahin and Kemal Leblebicioglu. Approximating the
optimal mapping for weapon target assignment by fuzzy reasoning.
Information Sciences, 255:30 -- 44, 2014.
[34] C´edric Villani.
Topics in optimal
American Mathematical Soc., 2003.
transportation. Number 58.
[35] C´edric Villani. Optimal transport: old and new, volume 338. Springer
Science & Business Media, 2008.
[36] Jacques L Willems and Iven MY Mareels. A rigorous solution of the
infinite time interval lq problem with constant state tracking. Systems
& control letters, 52(3-4):289 -- 296, 2004.
[37] Bin Xin, Jie Chen, Juan Zhang, Lihua Dou, and Zhihong Peng.
Efficient decision makings for dynamic weapon-target assignment by
virtual permutation and tabu search heuristics. IEEE Transactions on
Systems, Man, and Cybernetics, Part C (Applications and Reviews),
40(6):649 -- 662, 2010.
[38] Brian Yamauchi. Decentralized coordination for multirobot explo-
ration. Robotics and Autonomous Systems, 29(2-3):111 -- 118, 1999.
[39] Anmin Zhu and Simon X Yang. A neural network approach to
dynamic task assignment of multirobots. IEEE transactions on neural
networks, 17(5):1278 -- 1287, 2006.
|
1106.0223 | 1 | 1106 | 2011-06-01T16:17:03 | Decentralized Markets versus Central Control: A Comparative Study | [
"cs.MA",
"cs.AI"
] | Multi-Agent Systems (MAS) promise to offer solutions to problems where established, older paradigms fall short. In order to validate such claims that are repeatedly made in software agent publications, empirical in-depth studies of advantages and weaknesses of multi-agent solutions versus conventional ones in practical applications are needed. Climate control in large buildings is one application area where multi-agent systems, and market-oriented programming in particular, have been reported to be very successful, although central control solutions are still the standard practice. We have therefore constructed and implemented a variety of market designs for this problem, as well as different standard control engineering solutions. This article gives a detailed analysis and comparison, so as to learn about differences between standard versus agent approaches, and yielding new insights about benefits and limitations of computational markets. An important outcome is that "local information plus market communication produces global control". | cs.MA | cs | Journal of Artificial Intelligence Research 11 (1999) 301-333
Submitted 3/99; published 10/99
Decentralized Markets versus Central Control:
A Comparative Study
Fredrik Ygge
EnerSearch AB and Uppsala University
Chalmers Science Park
S-412 88 Gothenburg, Sweden
www.enersearch.se/ygge
[email protected]
Hans Akkermans
AKMC and Free University Amsterdam
Department of Information Management and Software Engineering
Computer Science Division
De Boelelaan 1081a, NL-1081 HV Amsterdam, The Netherlands
[email protected]
Abstract
Multi-Agent Systems (MAS) promise to offer solutions to problems where established,
older paradigms fall short.
In order to validate such claims that are repeatedly made
in software agent publications, empirical in-depth studies of advantages and weaknesses of
multi-agent solutions versus conventional ones in practical applications are needed. Climate
control in large buildings is one application area where multi-agent systems, and market-
oriented programming in particular, have been reported to be very successful, although
central control solutions are still the standard practice. We have therefore constructed and
implemented a variety of market designs for this problem, as well as different standard
control engineering solutions. This article gives a detailed analysis and comparison, so as
to learn about differences between standard versus agent approaches, and yielding new
insights about benefits and limitations of computational markets. An important outcome
is that “local information plus market communication produces global control”.
1. Introduction
When new paradigms arise on the scientific horizon, they must prove their value in com-
parison and competition with existing, more established ones. The multi-agent systems
(MAS) paradigm is no exception. In a recent book on software agents (Bradshaw, 1997),
Norman observes that perhaps “the most relevant predecessors to today’s intelligent agents
are servomechanisms and other control devices”. And indeed, a number of applications for
which multi-agent systems have recently claimed success, are close to the realm of what is
traditionally called control engineering. One clear example is the climate control of large
buildings with many office rooms. Here, Huberman & Clearwater (1994, 1995) have con-
structed and tested a working MAS solution based on a market approach, that they reported
to outperform existing conventional control.
The key question studied in this article is:
in what respect and to what extent are
multi-agent solutions better than their (conventional) alternatives? We believe that the
c(cid:13)1999 AI Access Foundation and Morgan Kaufmann Publishers. All rights reserved.
Ygge & Akkermans
above-mentioned application provides a nice opportunity to study this question in a de-
tailed empirical way. It is practically very relevant, it lends itself to alternative solutions,
and it is quite prototypical for a wide range of industrial applications in distributed re-
source allocation (including energy management applications (Ygge & Akkermans, 1996;
Akkermans & Ygge, 1997; Ygge & Akkermans, 1998), telecoms applications, the file alloca-
tion problem of Kurose and Simha (1989), and the flow problems investigated by Wellman
(1993)).
This article gives a detailed analysis of a published MAS solution to building climate
control, and it compares multi-agent markets with traditional control approaches. We also
introduce an improved and novel multi-agent solution to this problem based on an equilib-
rium market. From our comparative analysis we are able to draw general conclusions about
the suitability of the various approaches. Briefly, we show how computational markets can
be designed that perform as well as centralized controllers having global information about
the total system. However, a major advantage of the market framework is that it achieves
this in a fully decentralized fashion using only locally available data. We finally outline that
it should be possible to come to a more general theory concerning the connections between
markets and conventional control concepts. Here, we show that for the considered type
of applications the quasi-equation “local data + market communication = global control”
holds.
The structure of the article is as follows. After an introduction (Section 2) to market-
oriented programming and available theory, Section 3 gives the problem definition. Section 4
introduces the application domain: it describes the office environment and gives the phys-
ical model for cooling power and the various temperature and outside weather influences.
We then discuss the results of a standard control engineering solution, based on local and
independent integral controllers regulating the building climate (Section 5). Next, we re-
view the market-based approach as put forward by Huberman & Clearwater (1994, 1995)
(Section 6), and validate their claim that this market approach performs better than con-
ventional independent controllers. We subsequently analyze this market protocol in detail
and show that its success is related to the fact that the agents possess global information
before the auction is started (Section 7). In Section 8 we then develop an improved stan-
dard control engineering scheme that also exploits global data. Such a control scheme turns
out to perform much better than the Huberman-Clearwater market. Finally, we propose a
market design of our own based on general equilibrium theory. It performs as well as the
controller having access to global data, but operates on local data only and thus represents
a really decentralized solution (Section 9). Section 10 puts our results into perspective and
summarizes the general conclusions in comparing different approaches.
2. Market-oriented Programming
The use of market mechanisms for resource allocation by computer systems has a rather
long history in computer science, e.g. (Sutherland, 1968), and more recently markets have
been used in a number of different application areas.
302
Decentralized Market Control
There is rather extensive theory available for the relations between optimization and
markets. For example:
• Jennergren (1973) has shown how a price schedule decomposition algorithm can be
used for solving linear programming problems.
• Kurose & Simha (1989) showed how an equilibrium constitutes an optimal solution
to a file allocation problem. A similar result was shown by Bertsekas (1992) for the
assignment problem.
• Bikhchandani & Mamer (1997) showed that a Pareto efficient outcome1 maximizes
the sum of the utilities in an exchange market with agents holding quasi-linear utility
functions2 This is a significant extension to the above, as it also applies to non-linear
problems.3
• The results of Bikhchandani and Mamer were extended to markets that include pro-
duction and uncertainty by Ygge (1998, Chapter 3). The relations between these
types of markets and traditional optimization problems were also made more explicit
in the latter work.
Markets approaches have been applied to a number of different computerized applica-
tions, and some of them are briefly discussed here. Already in 1968 Sutherland proposed an
auction mechanism for allocation of computational resources on a PDP-1 computer (Suther-
land, 1968). Different amounts of money were assigned to different users in accordance with
the importance of their projects. It was reported that the users then over time learn how
to bid properly for computational resources. This basic approach has been further refined
by, for example, Gagliano et al. (Gagliano et al., 1995).
Kurose & Simha (1989) investigate a file allocation problem. In this work the agents
report their marginal utility (for having a certain amount of storage) and its derivative to
an auctioneer, which reallocates the resource using a resource-oriented Newton-Raphson
algorithm (cf. Ygge & Akkermans, 1998) until an equilibrium has been reached in which
the marginal utilities of all agents are the same. Although the paper is based on a number of
microeconomic abstractions, it does not for example utilize prices and there is no trade-off
between different commodities. So, it is debatable whether this approach can be referred
to as really market-based.4
The assigning problem of allocating n objects to n users has been investigated by Bert-
sekas (1992). Each user has a valuation of each object, and cannot be assigned more than
one object. It is reported that each user can be seen as an economic agent, and it is shown
how an auction (which essentially is an English auction) results in an equilibrium. The
1. An allocation is Pareto efficient or Pareto optimal if there is no alternative allocation that makes any
agent better off without making the outcome worse for some other agent (Varian, 1996, p. 15).
2. An example of a quasi-linear utility function is given in Eq. (19).
3. However, one should remember that this theory does not provide any computational advantages. If it is
hard to find the allocation that maximizes the sum of the utilities, then it is also hard to find a Pareto
optimal one (as they are the same in this case).
4. One should also note that the formulation of the Newton-Raphson scheme of this approach is overly
simplistic, and there are much better standard methods available (Press et al., 1994; Ygge & Akkermans,
1998).
303
Ygge & Akkermans
Auctioneer
Trading Agent
Initiate Auction ❍
❍
✟✙
✏
❍
❍❥
✟
✟
✏
✏
No
Express demand
✏✶
Supply = Demand?
❳❳❳❳❳❳③
Yes
Register and implement
Figure 1: High-level view of an equilibrium market mechanism.
assumed agent behavior is that an agent bids v − w for its most highly valued object (i.e.
the object with the highest difference, v, between the valuation and the price). w is the
difference between the valuation and the price of the second most preferred object. There
is however no motivation given for the assumed agent behavior, and it is not clear why any
agent would use this strategy unless it can be externally imposed. Though there is some
economical interpretation of the prices, the approach is not very market-like due to the
rather unrealistic assumptions of agent behavior.
The equilibria in the applications by Kurose and Simha, and Bertsekas are proven to be
optimal. It is not hard to see that the problem of Kurose and Simha can be reformulated
as a proper market with quasi-linear utility functions, and that the competitive equilib-
rium then is equivalent to the equilibrium described by Kurose and Simha, cf. (Ygge, 1998,
Corollary 3.3.1). Similarly, the price vector obtained by Bertsekas clearly constitutes a
competitive equilibrium. At the same time, it is shown by Ygge (Ygge, 1998, Theorem 3.2),
that all separable optimization problems can be formulated in market terms and the (com-
petitive) general equilibrium (if existing) is identical to the optimal solution to the original
optimization problem. Consequently, this recent theory generalizes the earlier theory by for
example Kurose and Simha, and Bertsekas.
Wellman et al. have contributed significantly in developing market-based approaches to
resource allocation into a programming paradigm, which they have given the name market-
oriented programming, e.g. (Wellman, 1993; Mullen & Wellman, 1995; Wellman, 1995, 1996;
Yamaki, Wellman, & Ishida, 1996; Hu & Wellman, 1996; Cheng & Wellman, 1998; Wellman
& Hu, 1998; Walsh, Wellman, Wurman, & MacKie-Mason, 1998). Particularly, the micro-
economic framework of general equilibrium theory has been successfully used as a resource
allocation mechanism. In such a market which we call an equilibrium market agents send
demand functions telling how much they like to consume or produce at different prices.
The auctioneer then tries to establish an equilibrium price vector such that supply meets
demand for all commodities, cf. Figure 1.
The process of submitting parts of the demand function may be iterated if an equilibrium
price is outside the region captured by the submitted demand functions. One such process
is the basic price tatonnement process, cf. e.g. (Cheng & Wellman, 1998), in which demand
functions for the respective commodities are sent to an auctioneer. Each of those demands is
based on expected prices of the other commodities. That is, if those other prices change, a set
of new demand functions may need to be submitted. Once the auctioneer has established
an equilibrium price, the agents will exchange the resources as stated by their bid and
304
Decentralized Market Control
the equilibrium price. (For example, if an agent states that it wants to buy 1/p units of
a commodity—where p is the price in the commodity money—and the equilibrium price
becomes 1, the agent will buy one unit of resource for one unit of money.)
Equilibrium markets have many attractive theoretical properties. For example, if all
agents act competitively5, the outcome is Pareto efficient. Furthermore, if all utility func-
tions are quasi-linear, the outcome is globally optimal (Ygge, 1998, Theorem 3.2), and
in the presence of uncertainty, the outcome maximizes the expected global utility (Ygge,
1998, Theorem 3.5). We note that equilibrium markets be computationally implemented in
a computationally very efficient manner (Ygge, 1998, Chapter 4).
Wellman et al. have applied equilibrium markets to a number of applications, such
as multi-commodity flow problems, design problems, and bandwidth allocation problems.
They have also introduced a market-based approach to scheduling which has many similar-
ities with the assignment problem of Bertsekas described above, but relies on more realistic
assumptions on agent behavior.
The present authors have introduced a market-oriented approach to power load man-
agement (Ygge & Akkermans, 1996; Ygge, 1998; Ygge et al., 1999).
We note that the aim of market-oriented programming in computer science is funda-
In
mentally different from the aim of economic theory. This is visualized in Figure 2.
market-oriented programming, microeconomic theory is taken as given and serves as the
theory for implementation of computational agents. Whether or not the microeconomic
theory actually reflects human behavior is not the critical issue. The important question
is instead how microeconomic theory can be utilized for the implementation of successful
resource allocation mechanisms in computer systems. For example, even though no (or at
least very few) people believe that humans use explicit utility functions when making their
decisions, such functions do appear very useful to concisely represent human preferences for
use in computational agents.
It is obviously very interesting to investigate the use of computational markets for au-
tomating trades between different self-interested parties, where information is private and
is revealed only if there is an expected gain from doing so. However, the use of markets has
also been proposed for standard resource allocation where the true utility/costs for all nodes
that consume and produce resource is assumed to be available (though possibly uncertain
and/or distributed in the system). The main arguments found in the literature for applying
market to these types of problems are:
• The numerous similarities between economic systems and distributed computer sys-
tems suggest that models and methods previously developed within the field of math-
ematical economics can serve as blueprints for engineering similar mechanisms in
distributed computer systems (Kurose & Simha, 1989).
• Auction algorithms are highly intuitive and easy to understand; they can be explained
in terms of economic competition concepts, which are couched on everyday experi-
ence (Bertsekas, 1992).
5. An agent that acts competitively treats prices as exogenous, that is, the impact on the prices due to its
own behavior is negligible (Varian, 1996, p. 516). This is a very reasonable assumption if the market is
of at least moderate size and/or if there is uncertainty about the behavior of the other agents (Sandholm
& Ygge, 1997).
305
Ygge & Akkermans
Micro-economic theory
Study
Economist
Generate
Study
Computer
scientist
Generate
Real World
Human agents on
human markets
Computational agents on
computational markets
Figure 2: A simplified view on the relation between economics and computer science with respect
to microeconomic theory. Economists study humans and generate theory that is used
for the explanation of human economic behavior. Computer scientists doing market-
oriented programming use this theory as a basis for building working computational market
systems.
• Market approaches enable a natural decomposition, both from a software engineering
perspective as well as from a computational perspective (Schreiber et al., 1999; Ygge,
1998, Chapter 15).
• Market approaches are very flexible in that they allow for ongoing addition and dele-
tion of agents. No global changes are required—merely the demand/supply relation
is altered (Ygge, 1998, Chapter 15).
• Markets are informationally efficient in terms of information dimensionality (Jordan,
1982), and the abstractions used are the most natural ones for the user (Ygge, 1998,
Chapter 15).
• The introduction of trading resources for some sort of money enables evaluation of
local performance and valuation of resources, so that it becomes apparent which re-
sources are the most valuable and which agents are using the most of these (Ygge,
1998, Chapter 15).
The above arguments are mainly conceptual and related to software/system design and
engineering issues, and must prove their value from acceptance by software and system
designers. But there are also different and in some respects stronger claims. For exam-
ple, Huberman and Clearwater state (Huberman & Clearwater, 1995) for the application
of building control that: ”While in principle an omniscient central controller with access
to all the environmental and thermal parameters of a building (i.e. a perfect model) could
306
Decentralized Market Control
optimally control it, in practice such knowledge is seldom available to the system. Instead,
partial information about local changes in the variables (such as instantaneous office occu-
pancy, external temperature, and computer use) is the only reliable source that can be used
for controlling the building.” As an alternative to such an omniscient controller Huberman
and Clearwater propose a market-based multi-agent approach to this problem. This raises
a very interesting question: Are there applications in which the information structure is
such that market-based approaches do better than traditional approaches? In this paper
we will carefully examine this issue for a building climate control problem. We will inves-
tigate what information different alternatives (traditional and new market-based) require,
and what the performance of these different approaches are. We will use exactly the same
problem formulation as Huberman and Clearwater in order to reproduce their results and
evaluate other new approaches with their original formulation. This problem formulation
is given in the next section.
3. Problem Definition
The task is to allocate a resource (cold air) in an office building, given the setpoint temper-
atures of the respective offices. As a measure of the success of the allocation, the standard
deviation of the deviation from the setpoint is used (Huberman & Clearwater, 1995), i.e.
StdDev(Ti − T setp) = vuut
1
N
[(Tio − T setp
o
N
Xo=1
) − (hTii − hT setpi)]2,
(1)
o
where h·i denotes the average value of a variable, Tio is the actual temperature of an office,
and T setp
is the setpoint temperature. The index i denotes a time interval under observation
and the index o denotes the office under observation. This naming convention is used
throughout the article. (In addition the index k will also be used to denote time periods.)
It may be debated whether or not this is the best measure, but we stick to it in this
article in order to evaluate the approach taken by Huberman and Clearwater using their
own measure.
4. The Office Environment
The case study we consider for the comparison of decentralized markets versus central
control solutions is the building climate control of a large office environment.
In this section, we present a mathematical-physical model of the office environment.
We first give a conceptual summary so that it is possible to understand the basic ideas of
the model without studying the equations. The offices are attached to a pipe in which the
resource (cold air) is transported as in Figure 3. The characteristics of this system are similar
to the characteristics of a district heating system, but with offices instead of households.
We assume that there are 100 offices in total, and that they are equally distributed towards
East, South, West, and North.
The thermodynamics of the office environment is actually quite simple. Every office is
seen as a storage place for heat, but heat may dissipate to its environment. In the model,
the thermodynamic behavior of an office is equivalent to a basic electrical RC-circuit. Here,
307
Ygge & Akkermans
Noon sun
Morning sun
Afternoon sun
Air out
Air in
Figure 3: Offices and air transportation.
voltage is analogous to temperature, and electrical current is analogous to heat flow. C and
R then respectively denote heat capacitance and thermal resistance.
A good general reference on thermodynamic models as the one we use here is the book
by Incropera & Witt (1990). The heat equations are continuous in time, but are discretized
according to standard procedures from control engineering, cf. (Ogata, 1990). The ontol-
ogy and reusability aspects involved in thermodynamics model construction are discussed
extensively by Borst, Akkermans, & Top (1997).
4.1 Thermodynamic Properties
The resource treated is cooling power. Each office can make use only of a fraction, η, of the
available resource at that office, P avail
, so that
io
P cons
io ≤ η · P avail
io
,
(2)
io
where P cons
is the consumed power. The available resource at one office is equal to the
available resource at the previous office minus the consumed resource at the previous office.
Throughout this article we assume an η of 0.5.
We treat everything in discrete time. The time interval we use in the calculations is
one minute. For all offices the temperature, Tio, is obtained by integrating a differential
equation in discretized form:
i
ko
)/Co,
(3)
Tio = T0,o +
(P heat
ko − P cons
Xk=1
where P heat
described by
ko
is the heating power and Co is the thermal capacitance. The heating power is
P heat
io = (T virt
io − Tio)/Ro,
(4)
308
Decentralized Market Control
where Ro is the thermal resistance and T virt
more detail below.
io
is a virtual outdoor temperature, described in
From Eqs. (3) and (4) we see that there is a feedback loop between the office temperature
and the heating power. Solving for the temperature we obtain
Tio =
1
1 + 1
RoCo
Ti−1,o +
T virt
io
Ro
io
− P cons
Co
, i > 0.
This is the equation for the dynamics of the system that can be directly computed. At the
right-hand side we have known quantities, where Co and Ro are externally given building
parameters, P cons
is the output of the utilized controller (as will be described in detail later
in the article for various different controllers), and T virt
is obtained from the weather model
below.
io
io
(5)
(6)
(7)
4.2 Weather Model
All external weather influences on the office environment are modeled by a virtual tem-
perature, representing the outdoor temperature, sun radiation, etc. We assume that there
is sunshine every day and that the outdoor temperature, T outd, varies from 22 to 35◦C
according to
T outd
i
= 22 + 13 · e−((i·s−4) mod 24−12)2/20,
where s is the length of each time interval expressed in hours, so here s = 1/60.
The virtual temperature, T virt
io
, is described by
T virt
io = T outd
i
+ T sun
o + T f luct
io
,
where T f luct is a random disturbance, thought to represent small fluctuations caused by for
example the wind. T f luct is Gaussian distributed with zero mean and a standard deviation
equal to unity. T sun is the sun radiation component. For the offices located at the East
side T sun is described by
T sun,East
i
= 8 · e−((i·s+4) mod 24−12)2/5,
and correspondingly for the South and the West offices
and
T sun,South
i
= 15 · e−(i·s mod 24−12)2/5,
T sun,W est
i
= 8 · e−((i·s−4) mod 24−12)2/5.
The various temperatures are plotted in Figure 4.
4.3 Office Temperatures without Control
(8)
(9)
(10)
In Figure 5, the temperature for a South-oriented office is plotted with different thermal
resistances, Ro, and thermal capacitances, Co. For simplicity we assume all Ro to be equal
and all Co to be equal. From this figure we observe two things: first, the higher RoCo
309
Ygge & Akkermans
C
e
r
u
t
a
r
p
e
m
e
T
35
30
25
20
15
10
5
0
C
e
r
u
t
a
r
e
p
m
e
T
15
10
5
0
-5
0 3 6 9
2
1
5
1
8
1
1
2
4
2
Time of day, h
C
e
r
u
t
a
r
e
p
m
e
T
50
40
30
20
10
0
0 3 6 9
2
1
5
1
8
1
1
2
4
2
Time of day, h
0 3 6 9
2
1
5
1
8
1
1
2
4
2
Time of day, h
Figure 4: The plot at the left shows the outdoor temperature, T outd
. The middle plot shows the sun
radiation components, T sun, (with peaks at time 8, 12, and 16 hours for offices located
at the East, South, and West sides, respectively). Finally, the outdoor temperature plus
the sun radiation components are plotted at the right.
i
C
e
r
u
t
a
r
e
p
e
m
e
T
45
40
35
30
25
20
15
R=0 or C=0
R= 10, C= 10,
R=10 and C= 20 or R=20 and
C=10
R=20 and C=20
0 2 4 6 8
0
1
2
1
4
1
6
1
8
1
0
2
2
2
Time of day, h
Figure 5: The indoor temperature for an uncontrolled office is plotted for different values of the
thermal resistance and the heat capacitance. Small values of the thermal resistance and
capacitance give higher peaks, while higher values give smoother curves.
the bigger the lag in temperature changes, and second, the higher RoCo the smaller the
fluctuations in temperature. For the simulation experiments in this article we took Ro = 10
and Co = 10. An Ro or a Co equal to zero implies that Tio = T virt
, as can be seen by
letting Ro or Co approach zero in Eq. (5). Clearly, without control the office temperatures
strongly fluctuate and reach unacceptably high values in all cases.
io
310
Decentralized Market Control
5. Control-A: Conventional Independent Controllers
5.1 Integral Control for the Offices
The application of regulating office temperature has a long tradition in control theory. The
most widely known controllers are different variants of the PID controller. The letters PID
denote that the control signal is proportional (P) to the error (that is, to the difference
between the setpoint and the actual value); proportional to the integral (I) of the error; or
proportional to the derivative (D) of the error. Here, we use a variant of an integrating
controller6 of the form
Fio = Fi−1,o + β(Tio − T setp
o
),
(11)
where F is the output signal from the controller, and β is a so-called gain parameter (that
can be set externally in the design of the controller). For the simulations it is assumed
that Fio is limited to a value between zero and three. This is in order to model that the
valves can be closed but that cooling resources are not delivered from one room to another
(the lower bound), and that there is a maximum to the amount of resources that can be
obtained by opening the valve fully (the upper bound). The control signal Fio is sent to the
actuator and the actual P cons
is obtained from
io
P cons
io = ( Fio,
η · P avail
io
Fio ≤ η · P avail
, Fio > η · P avail
io
io
.
(12)
Plots of the office temperatures with different gains are shown in Figure 6. The gain of
the controller is not critical for this application. Too high a gain will result in the controller
overreacting when the temperature exceeds the setpoint, after which it will be under the
setpoint for quite some time. This leads to a larger error than if smaller adjustments are
made. Also, the amplitude of the control signal then gets unnecessarily high, but the system
does not get dangerously unstable. We note that the maximum deviation here is ±0.06◦C.
Thus, controllers using any of the three gains perform very well. In the further calculations
of this article a gain equal to 10 has been adopted.
5.2 The Implications of Limited Resources
So far, we have assumed that the total amount of available resources is unlimited. Now,
we suppose that there is a maximum value for the cooling power that is inserted into the
system. In such a situation, offices that are situated close to the air input will obtain a
sufficient amount of cool air, but those near the end will suffer if totally uncoordinated
controllers are used. Thus, the smaller the total amount of available resources, the larger
the standard deviation will be. This is visualized in Figure 7. As a reasonable figure we
have chosen an upper limit for the total resource amount of 140.
We conclude, as shown by the example, that independent integrating controllers perform
very well when the amount of cooling resources is unlimited. On the other hand, when there
6. The main reason for choosing an integrating controller is that we want to have a setting that is as close
as possible to the setting described by Huberman and Clearwater (1995). Other controllers for this
application may well be considered, but as long as the performance is as good as shown in Figure 6, this
is not at all crucial for the central argument of this article.
311
Ygge & Akkermans
20,06
20,04
20,02
20
19,98
19,96
19,94
1
Gain = 100
Gain = 10
Gain = 1
11
21
Minutes in an observed interval
C
,
e
r
u
t
a
r
e
p
m
e
T
Figure 6: The indoor temperature for an office, utilizing an integral controller, is plotted for different
controller gains. The setpoint temperature is 20◦C.
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
6
5
4
3
2
1
0
0 1 2 3 4 5 6 7 8 9
0
1
4
1
5
1
6
1
7
1
8
1
9
1
0
2
1
2
2
2
3
2
4
2
1
1
2
1
3
1
Time of day, h
Figure 7: The standard deviation (in ◦C) , as defined in Eq. (1) for each interval, i, is plotted
for different amounts of available cold air resources (130, 140, 150, and 160) with 100
offices. The lower the amount of available resources, the higher the standard deviation.
312
Decentralized Market Control
is a shortage of available resources, the standard deviation increases dramatically, yielding
poor performance.
6. Market-A: The Approach by Huberman and Clearwater
A multi-agent systems solution to the problem of building control has been presented by
Huberman and Clearwater (1994, 1995). The approach taken is to model the resource
allocation problem of the building environment as a computational market where agents
buy and sell cooling power resources. The non-separability in terms of agents is ignored.7
The basic idea is that every office is represented by an agent that is responsible for
making good use of its resources, and to this end trades resources with other agents. The
agents send bids to an auctioneer, which calculates a clearing price, and makes sure that
no agent has to buy for a price higher than its bid nor has to sell for a price lower than its
bid.
In this section we give the market protocol proposed by Huberman and Clearwater, and
we reproduce their results. In order to make the paper self-contained, the interested reader
can find the original equations underlying the Huberman-Clearwater market protocol in
Appendix A.
6.1 Market protocol
A bid is constructed as the following tuple:
bid = [sellio, vio, Bio],
(13)
where sell is a boolean variable indicating whether the current bid is a sell bid (true) or
buy bid (f alse), v is the volume traded for, and B is the price that the agent demands (sell
bid) or is prepared to pay (buy bid).
Each of the variables in Eq. (13) is a function of other variables according to
sellio = sellio(Tio, T setp
vio = vio(Ti, Tsetp, α), and
Bio = Bio(Tio, T setp
o
o
, hTii, hT setp
o
i),
(14)
, hTii, hT setp
o
i, mio),
1
, T setp
, . . . , T setp
where Ti = [Ti1, Ti2, . . . , Ti100], Tsetp = [T setp
100 ], α is a strength parameter of
the auction, and mio is called a money parameter. The trade volume is directly proportional
to α and, hence, doubling α doubles the traded volume in an auction. Thus, α is a free
parameter that can be externally set to tune the trade volumes to proper levels. The
money parameter does not correspond to any real currency; it is merely a control variable,
varying between 100 and 200, and it does not have any direct connection to, for example,
the value of the cooling resource. The money cannot be saved between auction rounds
2
7. A resource allocation problem is separable if the total utility of the system of the system can be expressed
as the sum of the utilities of each node (producer/consumer), and the only constraints on how much
resource that can be assigned to each node is only restricted by that their sum must not exceed the total
available resource (Ibaraki & Katoh, 1988). Hence, the problem treated in this paper is not separable as
there are other constraints on how the resource can be distributed. For example, if we assign sufficiently
much to the penultimate agent, the ultimate will have a significant amount of resource available.
313
Ygge & Akkermans
in the Huberman-Clearwater protocol, another indication that it simply functions as an
adjustment parameter in this protocol.
An auction consists of the following three phases:
1. All
agents
Bio(Tio, T setp
o
, hTii, hT setp
o
send
their
bids,
[sellio(Tio, T setp
o
i, mio)], to the auctioneer.
, hTii, hT setp
o
i,
vio(Ti, Tsetp, α),
2. The auctioneer then computes a market clearing price by searching for the following
minimum:
min
pi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xosell=true,Bio≤pi
vio −
Xosell=f alse,Bio≥pi
.
(15)
vio(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
This yields a price such that supply matches demand as closely as possible.
3. Finally, the resources are allocated as dictated by the above bids and the obtained
market price, according to the rule: all sellers that are offering vio for a price, Bio,
that is lower than or equal to the market price will sell vio; and correspondingly for
buyers.8
As an example of the computation of the equilibrium price, assume that the auctioneer
has received the following bids [true, 2, 4], [true, 1, 3], [true, 2, 2], [f alse, 1, 3], [f alse, 2, 2],
[f alse, 2, 1]. Then the market clearing price is 3, leading to acceptance of bid 2, 3, 4, and 5.
(Bid 1 is too high a sell price and bid 6 is too low a buy price.)
6.2 Simulations
Figure 8 shows two plots of a simulation for the period between 3 p.m. and 7 p.m.9 The
initial temperatures for all offices were set to 20◦C. The upper plot is the standard de-
viation when independent integrating controllers are used, and the lower one shows the
agent-based control scheme as defined above. We found that for the adjustment parameter
α (see Eq. (23)) a value of 64 led to the smallest overall standard deviation. The key ob-
servation from the figure is that the agent approach offers at least one order of magnitude
of improvement.
Compared to independent conventional controllers this is indeed a major advance, and
this is the central claim made by Huberman and Clearwater. The market approach given
above leads to a reduced control error and thus to increased comfort compared to stan-
dard controllers. Our simulations thus reproduce and validate the results of Huberman &
Clearwater (1994, 1995).
8. Since the bids are given using discrete volumes, supply will seldomly match demand exactly at the
clearing price. Normally, there will be a very small excess demand or supply. If there is an excess supply,
all buyers that are willing to pay at most the clearing price will buy, but not all sellers that are willing
to accept at least the clearing price can sell. In this situation, a seller is selected randomly from the
valid candidates does deliver only a fraction of its bid. A corresponding procedure is used when there is
a small excess demand.
9. All solutions discussed in this article have been implemented in and simulated by the authors using
C++ on a PC running Windows95. Furthermore, all simulations have been independently recreated and
verified by Bengt Johannesson from various relevant papers (Ygge & Akkermans, 1997; Huberman &
Clearwater, 1995; Clearwater & Huberman, 1994)), as part of a master’s thesis project, using Python
on a PC running Linux.
314
Decentralized Market Control
4,5
4
3,5
3
2,5
2
1,5
1
0,5
0
5
1
Control-A
Market-A
6
1
7
1
Time of day, h
8
1
9
1
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
Figure 8: Standard deviation for independent controllers (top) and agent-based control (bottom).
However, this cannot be the last word in a successful system study. A first question to
be raised is whether we can find out in more detail what the actual reason is for the success
of the above market-based approach. A second relevant question is whether there are
alternative, market-based and/or central control-based, solutions to the building control
problem that are even better. The answer to both questions is yes, as we set out to
demonstrate now.
7. Market-A′: A Suite of Variations
In this section we present a suite of variations on the scheme presented in Section 6. The
main aim is to understand which of many factors involved are actually responsible for the
good performance of the Huberman-Clearwater market approach.
7.1 Deleting the Money Dependency
A first simplification is to remove the adjustment parameter called money. This is done by
setting mio to a very high value, C, regardless of the resource allocation.10 The simplified
market protocol then is:
1. All
agents
Bio(Tio, T setp
o
send
their
bids,
[sellio(Tio, T setp
o
, hTii, hT setp
o
i,
vio(Ti, Tsetp, α),
, hTii, hT setp
o
i, mio = C)], to the auctioneer.
2. The auctioneer computes a market price from
3. Allocate the resource as dictated by the bids and the market clearing price.
min
pi
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Xosell=true,Bio≤pi
315
vio −
Xosell=f alse,Bio≥pi
.
(16)
vio(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Ygge & Akkermans
0,25
0,2
0,15
0,1
0,05
0
5
1
Market-A
Market-A',
No Money
6
1
7
1
Time of day, h
8
1
9
1
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
Figure 9: Standard deviation for the original Huberman-Clearwater scheme and a scheme without
any money dependencies.
Plots of the standard deviation with the original scheme as well as with a scheme where
all money dependencies have been removed are shown in Figure 9. For the scheme without
money, setting the α parameter to 66 turned out to be optimal. The scheme without
money dependencies performs as well as the original scheme. The reason for this is that
Bio’s dependency on mio is completely negligible.11 Hence, the money parameter does not
play any significant role in the success of the Huberman-Clearwater market scheme.
7.2 Deleting the Temperature Dependency
Another factor of interest is the impact of the temperature on the bid prices. We remove
this dependency by setting the bid price to 10 for all selling agents and to 100 for all buying
agents. That is, we now use the following market scheme:
1. All agents send their bids,
[sellio(Tio, T setp
o
, hTii, hT setp
o
to the auctioneer.
i), vio(Ti, Tsetp, α), Bio =( 10,
100, sell = f alse
sell = true
],
2. Now, all prices larger than 10 and smaller than 100 are equally good candidates. If
there is a mismatch between supply and demand, say, supply exceeds demand, the
agents that will sell are picked randomly among the valid candidates, and the resources
are allocated accordingly.
10. More precisely, this is done by setting U (0, m) in Eq. (28) (Appendix A) to a constant value, 2000. This
corresponds to letting mio approach infinity.
11. In actual fact, U (0, m) in Eq. (28) of Appendix A will vary only between 1999.86 and 2000 for all possible
mio.
316
Decentralized Market Control
0,25
0,2
0,15
0,1
0,05
0
15
Market-A',
No Money
Market A',
No Money,
No Temp
16
17
Time of day, h
18
19
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
Figure 10: Standard deviation for the scheme with the money dependencies removed, compared to
a scheme with both the money and temperature dependencies removed.
In Figure 10 the standard deviation is plotted for the scheme without money, mentioned
above, and for a scheme where the dependency on the temperature has been removed as
well. Here, setting the fit parameter α to 65 turned out to be optimal. We see that also
here the performance is as good as that of the original scheme. Note that the temperature
is still used to determine both vio and the boolean variable sell. But the conclusion is that
the temperature dependency does not affect the quality of the market-based solution to the
building control problem.
7.3 Deleting the Auction
Next, we let the agents assign their bids to themselves without any auction whatsoever. This
means that the sum total of the controller outputs, Fio, might sometimes exceed the total
resource amount and sometimes be below that. The physical model is of course still obeyed,
so that the total resource amount actually used, the total consumed cooling power P cons
,
will never exceed the totally available resource amount, as described by Eq. (2). Hence, the
resource updates are described by:
io
F ′
(17)
io = Fi−1,o( +vio(Ti, Tsetp, α), sell = f alse
Fio =
−vio(Ti, Tsetp, α), sell = true
F ′
io < 0
0,
io, 0 ≤ F ′
F ′
io > 3
3,
io ≤ 3
F ′
The result of this simulation is shown in Figure 11. Here, an α of 17 turned out to be
optimal. The surprising conclusion is that the multi-agent scheme without performing any
auction is roughly ten times better than the auction-based Market-A scheme.12
12. From a practical point of view this is a pathological solution, revealing a loophole in the problem
definition of Clearwater and Huberman, cf. Section 3. The revealed loophole of the problem definition is
317
Ygge & Akkermans
0,25
0,2
0,15
0,1
0,05
0
15
Market-A
Market-A', No
Money, No
Temp, No
Auction
16
17
Time of day, h
18
19
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
Figure 11: Standard deviation with the auction mechanisms completely removed, compared to the
original Huberman-Clearwater Market-A scheme.
7.4 Discussion
At first glance, it might seem counterintuitive that performance actually improves signifi-
cantly when the core mechanisms of a market are removed. First we showed that introducing
the market improves performance considerably compared to conventional independent con-
trol, and then we showed that market mechanisms are superfluous. What, then, is the big
difference between the uncoordinated integrating controller, the Control-A scheme, and
the multi-agent scheme without any auction that we ended up with? The simple answer, in
our view, is the access to global data that each agent has, in terms of the average temper-
ature and the average setpoint temperature (as follows from Eq. (14)), in the Market-A
and Market-A′ schemes. Knowing average quantities means that an agent has some infor-
mation about the other agents c.q. offices. This information is not available in the case of
independent conventional controllers. The exploitation of this non-local information makes
the big difference.
that it does not take absolute temperature into account, but only the differences between temperature
(i.e. minimizing standard deviation while potentially sacrificing average temperature). In the proper
approaches introduced in the remainder of this article we do not allow ourselves to take advantage
of this loophole, but minimize differences while using all available resource. That is, all approaches
that are introduced from now on solve a problem which is strictly harder than the problem defined
in Section3. What happens in the Market-A scheme without an auction can be seen from observing the
definition of tio in Eq. (22). It is based on a quotient between the setpoint temperature and the actual
temperature. Then the request volume, Eq. (24), is proportional to this quotient. This always results in
a significant excess supply. (Hint, think of a small example with two offices, both with setpoint 20, and
with temperatures 19 and 21. Now, 20
21 , and hence the selling volume will
exceed the buying volume.) This implies that at each trade occasion, many offices are prevented from
selling (because of poor definitions of the volumes). The offices that are prevented to sell will deviate
significantly, and hence this results in a high value of the measure. When the auction is deleted, all
agents that are below the average distance to the setpoints are now able to get rid of the resource and
hence the differences in temperature will decrease.
19 is farther away from 1 than 20
318
Decentralized Market Control
Thus, the upshot of our factor analysis of the success of the Huberman-Clearwater
market scheme of Section 6 is that the agents have to share crucial office and control
information first, before the auction takes place. This is why one can remove the auction
without any deterioration of the results: a major function of an auction is to share necessary
information through the bidding procedure, but in the Huberman-Clearwater protocol the
needed information has already been shared. A more appropriate description of this market
scheme would therefore be to say that it consists of not three (as described in Section 6),
but four phases, in which the first phase is that all agents send their current temperatures
Tio as well as their setpoint temperatures T setp
to all other agents.
o
In sum, we have now answered the first question raised at the end of the previous section:
What is the reason for the success of the Huberman-Clearwater market scheme, compared
to conventional independent controllers? The single most important factor is contained in
the first step above: before the market procedure is carried out, the agents communicate key
information about their current situation to all other agents. If that is done neither money
nor auction mechanisms are needed. For conventional independent controllers, with which
Huberman and Clearwater compare their own approach, this globally shared information is
not available, however. This explains why these two solutions perform so very differently.
Now, we turn to the second question raised in the previous section: are there alternative
and perhaps better solutions? The above analysis suggests that we look into two different
directions. If one wants to persuade control engineers of the value of agent-based approaches,
it will be more convincing to compare the Huberman-Clearwater scheme not to strictly local
controllers, but to controllers that have access to the same global information as the agents
have. Secondly, this dependency on global information by the agents is a highly undesirable
feature that clearly runs counter to a decentralized computational market philosophy. So,
one should investigate whether it is possible to design a market protocol that exploits only
strictly local information in the agents’ bid construction. These issues we investigate in
the following two sections, first treating central control and subsequently devising a strictly
local market protocol. It will be seen that both yield significantly improved performance.
8. Control-B: A Standard Controller Using Global Data
Having concluded that the access to global data is crucial for performance, it is of course
of interest to analyze what the performance would be of an integrating controller, like the
one introduced in Eq. (11), but now incorporating global data.
We would like the controller to take into account not only its own deviation from its set-
point, but to consider also the deviations of the other offices from their setpoints. Therefore,
the controller in Eq. (11) is extended to
Fio = Fi−1,o + β[(Tio − T setp
o
) − (hTii − hT setpi)],
(18)
where β is set to 10 as previously. P cons
io
is, as before, obtained from Eq. (12).
The plot from the simulation with this controller, compared to the Market-A sim-
ulation, and to the Market-A′ simulation where the auction was removed, is shown in
Figure 12.
We see that the standard deviation is approximately the same for the Control-B and
the Market-A′ schemes. Thus, the Control-B scheme also performs roughly ten times
319
Ygge & Akkermans
0,25
0,2
0,15
0,1
0,05
0
5
1
Market-A
Market-A',
No Money,
No Temp,
No Auction
Control-B
6
1
7
1
Time of day, h
8
1
9
1
i
n
o
i
t
a
v
e
d
d
r
a
d
n
a
t
S
Figure 12: Standard deviation with an integrating controller that utilizes global data compared to
the Huberman-Clearwater Market-A scheme, and to the Market-A′ scheme with the
auction removed.
better than Market-A. An important difference, though, is that Control-B employs a
well-known integrating controller for which there are well-understood methods and theories
for e.g. stability analysis.13 In contrast, the Market-A scheme is not easily analyzable
from a formal theory perspective, since it does not rely on such well established concepts.
While Control-B is relatively clear, see Eq. (18), it is much harder to conceptually grasp
the principles behind Market-A, as it consists of a large number of complicated and not
easily justified equations (see Eqs. (21)–(33) in Appendix A). The main conclusion from
the present section is that, if we enable conventional controllers to act upon the same global
information as the Huberman-Clearwater agents do, the central control approach performs
best and is easiest to understand.
9. Market-B: A Market with Local Data Only
Thus, the computational market Market-A is outperformed by the global control scheme
Control-B. Therefore, an interesting follow-up question is whether a simple and well
performing computational market approach can be devised that does not depend on hav-
ing available global information, in contrast to both these schemes.
In this section we
show that this is indeed the case, such that a decentralized market performs comparably
to fully centralized conventional control. The approach derived in this section relies on
more general theory for the relations between optimization problems and markets–(Ygge,
1998, Chapter 3) and (Bikhchandani & Mamer, 1997) – applicable to this and many other
applications.
13. This holds under the assumption that the characteristic time scale of the variations of the average
temperature is much larger than that of the fluctuation of the temperature. This is a very reasonable
assumption in the present case.
320
Decentralized Market Control
We use a market-oriented programming approach based on an equilibrium market. The
basic protocol for this is (cf. Figure 1):
1. Each agent submits a net demand function zio(p) to the auctioneer, describing what
change in allocation is desired by agent o at price p.
2. The auctioneer computes an equilibrium, by calculating a price pi such that
Po zio(pi) = 0 (or, strictly speaking, Po zio(pi) ≤ ǫ, where ǫ is a small (positive)
numerical tolerance).
3. Each agent receives its demanded resource zio(pi) as calculated at the obtained market
equilibrium price.
The computation of zio(p) and the process of computing the equilibrium are described
further below. Note that this is a truly distributed approach, as it does not rely on any
global data, or communication of temperatures between the agents before the market com-
munication starts.
9.1 The Relation Between Markets and Conventional Controllers utilizing
Global Data
The performance measure for the system is given by Eq. (1). The best system is therefore
one that minimizes this equation. Hence, the most straightforward move one can think of
to come to a market model, is to take this measure as representing the utility function for
the overall system. So, the utility functions for the individual agents are ideally related to
[(Tio − T setp
) − (hTii − hT setpi)]2. 14 However, this is still a formulation containing global
information.
o
Thus, we want to obtain a purely local reformulation, by getting rid of the terms with
hT i containing the global information. We might replace them, though, by terms relating
to the changes in the local resource. In doing so, we take inspiration from the standard
controller equations Eqs. (11) and (18), indicating that we get good results (for uncon-
strained resources) with the update equation Fio = Fi−1,o + φio, when φio has the form
φio = β(Tio − T setp
). The intended interpretation of φio is to represent the output that the
local controller would have delivered if it acted independently with unconstrained resources.
In the market setting each agent updates its control signal Fio through Fio = Fi−1,o + ∆Fio,
where ∆Fio is determined by the outcome of the market. Since the resource is only re-
o=1 ∆Fio = 0. Accordingly, as a step in the
design of a Market-B scheme, based on local data only, we employ the following definition.
distributed among the agents, we have that PN
o
Definition 9.1 Let the utility function for the individual office agents be defined by
u(∆Fio, m) = −α2
o(∆Fio − φio)2 + m,
(19)
where αo is a strength parameter for each office representing its physical properties such as
Ro and Co. (The proper choice of αo is discussed in Theorem 9.3 later on.)
Furthermore, let all agents be price-takers.
14. Since utilities are expressions of preference orderings, they are invariant under monotonic transforma-
tions.
321
Ygge & Akkermans
Due to the simple quadratic form of the utility function in Eq. (19), the net demand,
zio(p), is easily computed for price-taking agents by analytical means. Note that the money
m here are fundamentally different from the money of Market-A, c.f. Section 6. Here the
money represents a commodity and the agent needs to decide how much money to trade
for the commodity cold air. As described above, the money in Section 6 is just a control
variable without an intuitive interpretation.
Theorem 9.1 With the utility functions and behavior given by Definition 9.1, there exists
a unique equilibrium price piPo zio(pi) = 0, and the allocation obtained by the equilibrium
market is Pareto optimal.
Proof. See (Mas-Colell, Whinston, & Green, 1995; Varian, 1992; Takayama, 1985). ✷
In this setting, there is a wide variety of different algorithms that are guaranteed to
converge to the clearing price piPo zio(pi) = 0, in the second step of our equilibrium
market protocol given above. Examples of such algorithms are price tatonnement, Wal-
ras (Cheng & Wellman, 1998), and various Newton-Raphson type methods. A detailed
discussion of communication and computation efficiency aspects of such algorithms is given
by Ygge (1998, Chapter 4).
Theorem 9.2 Any Pareto optimal outcome in a market in which the agents hold the utility
function of Eq. (19) is equivalent to an integrating controller that exploits global infor-
mation as described by the following resource update equation :
∆Fio = φio −
1
α2
o · h1/α2i
hφii,
(20)
given that no agent is at its boundary. Here, h1/α2i and hφii are the average 1/α2
respectively.
o, and φi0
At a Pareto-optimal allocation where no agent is at its bounds, all
Proof.
∂u(∆Fio)/∂∆Fio are equal.
(Assume ∂u(∆Fip)/∂∆Fip < ∂u(∆Fiq)/∂∆Fiq, then at any
price ∂u(∆Fip)/∂∆Fip < p < ∂u(∆Fiq)/∂∆Fiq, there exists a (sufficiently small) amount
of ∆Fi, say ǫ, such that up(∆Fip − ǫ, mip + pǫ) > up(∆Fip, mip) and uq(∆Fiq + ǫ, miq −
pǫ) > uq(∆Fiq, miq). Correspondingly for ∂u(∆Fip)/∂∆Fip > ∂u(∆Fiq)/∂∆Fiq. Hence,
∂u(∆Fip)/∂∆Fip = ∂u(∆Fiq)/∂∆Fiq).
Thus, it will hold for every office that ∆Fio − φio =
α2
N
α2
o
(∆FiN − φiN ). Summing the
. Adding ∆FiN −φiN to
o=1 ∆Fio =
1
α2
o
o=1
o=1 φio = α2
o=1 ∆Fio−PN −1
equations yieldsPN −1
both sides and dividing both sides by N together with the resource constraint (PN
0) yields
N ·h1/α2i hφii. For reasons of symmetry, this equation holds for all offices. ✷
α2
N (∆FiN −φiN )PN −1
N (∆FiN − φiN )PN
o=1 φio
N
−PN
1
α2
o
o=1
N
which is equivalent to ∆FiN = φiN −
1
= α2
Corollary 9.1 For the special case where all αo are equal, and φio = β(Tio−T setp
becomes exactly the Control-B scheme captured by Eq. (18).
o
), Eq. (20)
322
Decentralized Market Control
Consequently,
it also follows that simulation results for an equilibrium Market-B
scheme, based on the utility function of Eq. (19) with the provisions given in the above corol-
lary, are identical to those of the centralized Control-B scheme, cf. Figure 12. This fully
decentralized market scheme thus performs better than the original Huberman-Clearwater
market protocol.
In sum, we see that (under suitable conditions which are fulfilled here) local data plus
market communication is equivalent to conventional central control utilizing global data.
Even though our proof was based on the assumption that the agents are never at their
boundary values, it will be a close approximation in many practical applications. It should
also be noted that managing the boundaries is not required for a successful implementation.
As seen above, omitting the management of boundaries in the current application leads to
Control-B, which was shown to have a very high performance.
9.2 Finding an Optimal Utility Function
In this section we show how an optimal utility function is constructed in the constrained
case, from an optimal controller for the unconstrained case.15 Earlier we noted that it is
debatable whether or not the measure in Eq. (1) is a good one. One argument against it is
that it allows for pathological solutions, as it does not take the average indoor temperature
into consideration.
If we for example increase the indoor temperature in every office by
10◦C, all of them could be unbearably hot, while the measure might be minimized. A
reasonable thing to do here, to prevent these kind of solutions, is to treat the average
temperature as given (the result of using all available resources). If we do so, the following
theorem shows how an optimal utility function is constructed in the constrained case, from
an optimal controller for the unconstrained case.
Theorem 9.3 If Tio is a linear function of ∆Fio,16 and if φio minimizes Eq. (1) in the
unconstrained case, then in a market where the utility functions are described by Eq. (19),
with αo = ∂Tio
, any associated Pareto optimal allocation minimizes Eq. (1) in the con-
∂∆Fio
strained case, if the resource can be independently allocated among the agents (i.e., when
o=1 ∆Fio = 0), and the
the only constraint on ∆Fio is the upper and lower bound, and PN
Proof. Minimizing Eq. (1) boils down to minimizing f (Ti1, Ti2, . . . , Tin) = PN
average temperature is considered as given.
o
T setp
φio minimizes Eq. (1), it holds that
is considered as given, we have that
for ∆Fio = φio. This gives that Tio(φio) = T setp
o=1[(Tio −
)−(hTii−hT setpi)]2, since this is just a monotonic transformation. Due to the fact that
= 0 for ∆Fio = φio. Since the average temperature
) − (hTii − hT setpi)] · ∂Tio
∂φio
o + (hTii − hT setpi). Thus, f can be
o=1[Tio(∆Fio)−Tio(φio)]2. Since Tio is a linear function of ∆Fio, we have that
Tio(∆Fio) = Tio(φio) + ∂Tio
(∆Fio −
∂∆Fio
φio)2. As the reallocation of m does not effect the total (summed) utility, minimizing
· (∆Fio − φio), and hence f becomes PN
rewritten as PN
∂∆Fio(cid:17)2
o=1(cid:16) ∂Tio
= 2[(Tio(φio) − T setp
o
∂f
∂∆Fio
∂f
∂∆Fio
15. The resource is constrained when the action of any agent is limited by Pio, see Eq. (12) and Figure 7. It
is unconstrained when such a limitation is not present, cf. Figure 6.
16. From the thermal model discussed in Section 4 (especially Eqs. (5) and (12)) we note that this is indeed
the case for a reasonably wide range.
323
Ygge & Akkermans
o=1 uo.
io + uo(∆F a
o=1 uo, with αo = ∂Tio
∂∆Fio
f is equivalent to minimizing −PN
Eq. (19) maximizes PN
maximize PN
. Furthermore, we have that
any Pareto optimal allocation in a market where the utility functions are described by
io that does not
io =
ma
is always a Pareto
improvement. Hence, a non-maximized sum implies a non-Pareto optimal allocation, and
a Pareto optimal allocation implies a maximized sum.) Then (Ygge, 1998, Theorem 3.1,
p. 44) implies that any Pareto-optimal allocation in a market where the utility functions
are described by Eq. (19) with αo = ∂Tio
∂∆Fio
(Suppose that there is an allocation ∆F a
io and letting mb
io does. Then reallocating ∆Fio to ∆F b
io,ma
io)
o=1 uo, but that ∆F b
io, mb
io, ma
io) − uo(∆F b
io) +PN
, minimizes f . ✷
o=1 uo(∆F b
io,mb
o=1 uo(∆F a
io)−PN
N
Thus, we have constructed a fully distributed market design that yields an optimal
outcome. Particularly note that Theorem 9.3 was not based on the assumption that an
integral controller was used, rather it said that if φio (the desired resource by the controller)
is optimal, then the proposed utility function generates a globally optimal outcome.
9.3 Discussion
Previously, we saw that the independent controller Control-B that incorporates global
data, viz. the average temperatures, performs very well. In the present section we positively
answered the question if one can construct a market, Market-B, that is based on local
data only and that performs as well.
For this result we have employed a market approach based on general equilibrium theory.
This is of course not the only available mechanism for resource allocation in multi-agent
systems. It seems interesting to try out other mechanisms, like the contract net protocol
(Davis & Smith, 1983), and see if they perform better. However, Theorem 9.3 tells us
that, if we treat the problem of building control as being separable in terms of agents (as
also done by Huberman and Clearwater), there is no better scheme.17 For example, if
we assigned all the resources to an auctioneer, that on its turn would iteratively assign the
total resource in small portions to bidders bidding with their true marginal utility, we would
end up with something close to the competitive equilibrium, but we cannot do better than
Market-B. Furthermore, this would be a computationally extremely inefficient way to
arrive at equilibrium compared to other available methods (Ygge, 1998, Chapter 4). That
is, we can use different mechanisms for achieving the competitive equilibrium, but we can
never hope to find a mechanism that would do better than the Market-B scheme.
17. We note that, as mentioned earlier, the problem is actually not fully separable in terms of agents and
therefore better solutions may exist if this is taken into account. Another observation is that in this article
we have investigated only the case of using the currently available resource as the interesting commodity,
in accordance with the work of Huberman and Clearwater, and we found an optimal mechanism for that.
We note however that extending the negotiations to future resources as well could potentially increase
performance. But this is a different problem setting with different demands on available local and/or
global information items, such as predictions. We have given a solution to this problem in recent work,
see Ygge et al. (1999).
324
Decentralized Market Control
10. Conclusions
We believe that both the approach and the results, as presented in this article, pose an
interesting challenge to the software agent community. Multi-agent systems offer a new
way of looking at information systems development, with a potentially large future im-
pact. However, new approaches must prove their value in comparison and competition with
existing, more established ones. The agent paradigm is no exception.
We have therefore deliberately played the role of the devil’s advocate in this article. In
our view, a key question not yet satisfactorily answered by the software agent community
is:
in what respect and to what extent are multi-agent solutions better than their more
conventional alternatives? This article has shown that arguing in favor of the multi-agent
systems approach does require careful analysis beyond the disciplinary boundaries of com-
puter science. Empirical and comparative studies in the multi-agent literature of the kind
carried out in the present article are all too rare as yet. But as we have shown in techni-
cal detail, established paradigms such as conventional central control cannot be that easily
dismissed. A similar argument holds, by the way, regarding mathematical optimization
techniques in distributed resource allocation problems, cf. our previous discussions (Ygge
& Akkermans, 1996; Ygge, 1998). Many of them are, in a distributed guise, better than
many newly proposed market protocols.
Abstract considerations alone, concerning the general nature of agenthood, autonomy,
rationality, or cooperation, are not sufficient to prove the value of the agent paradigm. Such
theoretical reflections are worthwhile, but do not diminish the need for thorough analysis of
(agent and market) failures and successes in real-life applications. Therefore, we have taken
a different approach, aimed at obtaining experimental data points on the basis of which
convincing software agent claims can be established.
The data point considered in this article is climate control within a large building con-
sisting of many offices. Given a measure, Eq. (1), a local controller approach, Eq. (11),
current temperatures, and the setpoint temperature, we have investigated the problem on
how to properly distribute the resource among the offices. This is a rather prototypical
application relating to the general problem of optimal resource allocation in a highly dis-
tributed environment. This class of problems has already received much attention in the
multi-agent area. Reportedly, this type of application is very suitable for market-oriented
programming (Huberman & Clearwater, 1995). On the other hand, we have devised some
better conventional control engineering solutions, as well as alternative and better market
designs.
The main conclusions of our investigation are:
• The market approach by Huberman & Clearwater (1994, 1995) indeed outperforms
a standard control solution based on local, independent controllers. So, the market-
based multi-agent approach indeed yields a working solution to this type of problem.
Our analysis has shown that the success of this market approach depends on the
agents communicating their local information to all other agents before the auction
starts.
• However, if conventional central control schemes are allowed to exploit the same global
information, they perform even better.
325
Ygge & Akkermans
• We have proposed an alternative market design based on general equilibrium theory
that uses local data only. It performs better than the Huberman-Clearwater market
and as well as a centralized control scheme having access to global information.
• Our general conclusion can be formulated as a quasi-equation: “local information
+ market communication = global control”. This holds under suitable conditions
(existence of market equilibrium, price-taking agents) and its validity has been specif-
ically shown here for the case of building climate control (compare particularly the
Control-B and the Market-B schemes we developed). However, as we intend to
show in forthcoming work, this is a much more generally valid result.
• The important difference is that in computational markets this global information is
an emergent property rather than a presupposed concept, as it is in central control.
In our analysis we have focused on the market approach. It is tempting to ask whether
things are different when a non-market multi-agent approach is followed, say, using the
contract net (Davis & Smith, 1983). As we argued, the answer in our opinion is a straight-
forward no. The goal in the considered class of problems is to find the optimal distributed
solution. Alternative agent approaches, market as well as non-market ones, only change the
multi-agent dynamics on the way to this goal. This might be done in a better or poorer way,
but it is not possible to change this goal itself. The goal state in any multi-agent approach
is, however formulated, equivalent to market equilibrium, the yardstick for having achieved
it is given by some quantitative performance measure as we discussed, and both are stable
across different agent approaches.
Thus, one of the main conclusions of this article is that one must be very careful when
promoting multi-agent approaches to resource allocation problems for which traditional
approaches are available.
In particular, one should be cautious using arguments related
to distribution of information (as was done by Huberman and Clearwater), or other com-
putational aspects. Still, we argue that there are conceptual advantages of market-based
approaches to this type of problems, as well as other advantages such as evaluation of local
performance (cf. Section 2). These advantages do not show off very clearly in the present
application setting – for example, it was assumed that all offices have the same thermal
characteristics and office agents on the market are not added and deleted continuously on
the fly – but they are clearly visible in other more general settings.
A final note is that in this article we have devised an optimal market design given the
problem formulation rather than devised an optimal approach to the general problem of
building control. This is in line with the objective of this paper to give a comparative study
of different possible approaches. That is, we have focused on a published and well-known
problem formulation in order to focus on the subject of markets for resource allocation
alone. As stated earlier in the article, several aspects can be improved, for example:
• The local controllers. I-controllers are seldom used, rather PI- or PID-controllers are
preferred. Furthermore, for these kind of computerized settings, modern digital con-
trollers, based on e.g. pole-placement methods are preferable (Astrom & Wittenmark,
1990).
326
Decentralized Market Control
• The measure. The average value is at least as important as the standard deviation.
(We have as a consequence excluded pathological solutions from our analysis by taking
the average temperature as given in Theorem 9.3).
• The incentives for office agents to reveal their true preferences in order to avoid spec-
ulation. Reasonably, personnel in the offices should have the opportunity to make
trade-offs between comfort and economical value. (It has been shown that generally
speaking it is very difficult to benefit from speculation if the number of agents on the
market is sufficiently large, and/or if uncertainty about the other agents’ behavior
exists (Sandholm & Ygge, 1997).)
• Taking several time periods into account. As each office serves as a storage for heat,
agents can for example gain by using relatively much resource during certain hours
when the total resource need is small.
That is, a more realistic approach to the building control problem is to use minimized
cost as the measure, give the users real incentives to reveal their energy preferences (by
letting them pay their actual energy costs so that attempted speculation will generally
cost money), and take future time periods into account. Elsewhere we have solved the
problem of dealing with simultaneous optimization over different time periods, by means of
a multi-commodity market (Ygge et al., 1999).
In such a more realistic and large-scale setting, the market-based approach is attractive
compared to its alternatives. The abstractions used (prices and demands) are natural and
easily understood by everyone, and they are uniform over all types of agents (even ones
that do not use the resource for controlling a temperature).
A major qualitative conclusion of this article is that “local data plus market communi-
cation yields global control”. This conclusion is based on the discussed application setting
of building control, as an instance of a distributed resource allocation problem. It suggests
that generally for this type of problem both central control engineering and multi-agent
market solutions can be devised that give comparable optimal control quality. A next step
is to prove this more generally in a rigorous mathematical fashion, delineating the precondi-
tions in more detail. Such a proof can be based upon the (continuous and matrix-algebraic)
dynamic systems theory available from control engineering, and especially upon results con-
cerning what is called optimal control. We believe that this indeed can be done in a formal
way, and as a conjecture we state the following general result. Multi-agent equilibrium mar-
kets yield optimal decentralized solutions to distributed resource allocation problems that
are of the same quality, in terms of a given overall systems performance index, as solutions
given by the optimal central (multi-input, multi-output) systems controller that has access
to all relevant local data. These local data involve the total system state and control vectors
(in building control these are the vectors formed from the difference between the actual and
setpoint temperatures for each office, and the cooling power for each office, respectively).
Preconditions for this to be true are: (i) agents act competitively; (ii) the equilibrium ex-
ists; (iii) the systems performance index can be written as a linear combination of local
contributions (implying diagonality of certain matrices; if not, agents are not independent).
These local contributions are directly related to the agents’ utility functions. The agent
approach will be more readily generalizable to large-scale systems and non-linear control
327
Ygge & Akkermans
solutions (linearity is the main case where control engineering can get analytical mathe-
matical expressions). This is indeed a strong and general statement about the relationship
(and even outcome equivalence) between decentralized markets and central control.
Acknowledgments
The present work was mainly carried out at the University of Karlskrona/Ronneby. We
thank Rune Gustavsson and Hans Ottosson for all their support. A special acknowledgment
goes to Olle Lindeberg whose very detailed comments on the draft papers led to significant
improvements in the simulations and whose ideas also helped us in the design of the market
in Section 9. We benefit significantly from several discussions with Michael Wellman. We
also thank Bengt Johannesson, who, as a part of his master’s thesis, went through all details
of this article and independently recreated and verified all the simulations. We thank Arne
Andersson, Eric Astor, and the SoC team for useful comments and discussions of draft
material. This work was partially sponsored by NUTEK and EnerSearch AB.
An earlier version of this paper was presented at the MAAMAW’97 workshop (Ygge &
Akkermans, 1997).
328
Decentralized Market Control
Appendix A. Original Formulation of the Huberman-Clearwater Market
All formulae in this section were directly taken from the papers by Huberman & Clearwa-
ter (1994, 1995).
Trade volumes First, the decision for an agent to buy or sell is based on
tio =
T setp
o
Tio
·
hTii
hT setpi ( tio > 1, seller
tio < 1,
buyer
Then, the total trade volume, V , is calculated from
Vi =
N
Xo=1
1 − tio ,
where N is the number of offices.
Every agent calculates its request volume, v, according to
vio = α
1 − tio
Vi
.
.
(21)
(22)
(23)
When an agent buys or sells its v the actual movement of a valve, called V AV , is
computed from
∆V AVio = f (f lowio, vio, V AVio),
after which the actual V AV position for each interval is updated according to
V AVi+1,o = V AVio + ∆V AVio.
(24)
(25)
The function f in Eq. (24) stems from the physics of the office control, but it is not
easy to derive in the general case and it is neither explicitly given in the original papers
(Clearwater & Huberman, 1994; Huberman & Clearwater, 1995). A reasonable and simple
choice is to assume a linear relationship (the cited papers suggest that they do so in practice
as well), replacing Eq. (24) and Eq. (25) by
Fi+1,o = Fio ± vio,
(26)
where plus or minus depend on whether it refers to an accepted buy or sell bid. Pcio is
obtained from Eq. (12). We have employed Eq. (26) in our simulations. The nature of this
assumption is not crucial to our central line of reasoning.
Bids The bids are based on a marginal utility 18 of the form described by19
U (tio/T setp
o
, mio) = [U (0, mio)](1−tio/T setp
o
) = [U (0, mio)]
(1− hTii
Tio hT setp i
)
,
(27)
18. This is called utility in the work of Huberman and Clearwater. We prefer to use the term marginal utility
instead, because it is directly related to price. This terminology conforms better to that of microeconomic
theory.
19. This notion of a utility function used by Clearwater and Huberman is very much based on the work of
Steiglitz & Honig (1992)
329
Ygge & Akkermans
with
and
U (0, mio) = u3 − (u3 − u1)e−γmio ,
γ = ln(cid:20) u3 − u1
u3 − u2(cid:21) ,
(28)
(29)
where u1 = 20, u2 = 200, and u3 = 2000, and m is the amount of money that an agent has,
given by
With the above relation between Fio and V AV , we rewrite Eq. (30) as
mio = 100(2 − V AVio).
mio = 100(2 −
F max
o
− Fio
F max
o
).
Observing Eqs. (28) and (29), we note that these equations can be simplified to
(30)
(31)
(32)
U (0, mio) = u3 − (u3 − u1)e−γmio = u3 − (u3 − u1)(eγ)−mio =
u3 − (u3 − u1)(cid:16) u3−u1
u3−u2(cid:17)−mio
.
The bids are calculated from multiplying the marginal utility with the previous price,
price, according to
Bi,o = Uio(tio/T setp
o
, mio) · pricei.
(33)
o
o
, mio) will be well above 20001− 2
, mio) is minimized for minimized Tio and maximized hTii
This is the equation given by Huberman and Clearwater. Straightforward application
of Eq. (33) turns out to produce major problems in simulations, however. Equation (27)
shows that U (tio/T setp
hT setpi . As we
can expect Tio to be well above 10◦C and hTii
hT setpi to be well below 2, and U (0, m) ≈ 2000,
we can be sure that U (tio/T setp
10 ≈ 437. Thus, the bidding
price will never be below 437. Then, Eq. (33) tells us that the market price will be at
least price0 · 437i. This leads to numerical overflow after a few iterations. We note however
that, since all agents multiply their bids with the previous price, this has no effect on the
reallocation itself:
it affects only the price level. Therefore, we omit multiplying by the
previous price in our simulations, so that the agents’ bid prices in Eq. (33) equal their
marginal utilities of Eq. (27). Huberman and Clearwater state that they burn all money
after each auction round in order to avoid inflation and overflow, indicating that they may
adopt a similar procedure. In any case, this procedure leads to exactly the same allocations
but avoids numerical overflow.
330
Decentralized Market Control
References
Akkermans, J. M., & Ygge, F. (1997). Smart software as customer assistant in large-scale
In Proceedings of Distribution Automation/Demand
distributed load management.
Side Management (DA/DSM) ’97. PenWell Conferences and Exhibitions.
Bertsekas, D. (1992). Auction algorithms for network flow problems: A tutorial introduc-
tion. Computational Optimization and Applications, pp. 7–66.
Bikhchandani, S., & Mamer, J. W. (1997). Competitive equilibrium in an exchange economy
with indivisibilities. Journal of Economic Theory, 74, 385–413.
Borst, W. N., Akkermans, J. M., & Top, J. L. (1997). Engineering ontologies. International
Journal of Human-Computer Studies, 46, 365–406. ISSN 1071-5819.
Bradshaw, J. M. (Ed.). (1997). Software Agents. AAAI Press/The MIT Press, Menlo Park,
CA.
Cheng, J., & Wellman, M. P. (1998). The walras algorithm—a convergent distributed
implementation of general equilibrium outcomes. Computational Economics, 12, 1–
24.
Clearwater, S., & Huberman, B. A. (1994). Thermal markets for controlling building envi-
ronments. Energy Engineering, 91 (3), 25–56.
Davis, R., & Smith, R. G. (1983). Negotiation as a metaphor for distributed problem
solving. Artificial Intelligence, 20 (1), 63–109.
Gagliano, R. A., Fraser, M. D., & Schaefer, M. E. (1995). Allocation of compting resources.
Communications of the ACM, 38, 88–103.
Hu, J., & Wellman, M. P. (1996). Self-fulfilling bias in multiagent learning. In Tokoro, M.
(Ed.), Proceedings of the Second International Conference on Multi-Agent Systems,
ICMAS’96, pp. 118–125. AAAI Press, Menlo Park, CA.
Huberman, B. A., & Clearwater, S. (1995). A multi-agent system for controlling building
environments. In Lesser, V. (Ed.), Proceedings of the First International Conference
on Multi-Agent Systems, ICMAS’95, pp. 171–176. AAAI Press / The MIT Press,
Menlo Park, CA.
Ibaraki, T., & Katoh, N. (1988). Resource Allocation Problems—Algorithmic Approaches.
The MIT Press, Cambridge, MA.
Incropera, F. P., & Witt, D. P. D. (1990). Fundamentals of Heat and Mass Transfer. Wiley
and Sons, New York. Third Edition, ISBN 0-471-51729-1.
Jennergren, P. (1973). A price schedules decomposition algorithm for linear programming
problems. Econometrica, 41 (5), 965–980.
Jordan, J. S. (1982). The competitive allocation process is informationally efficient uniquely.
Journal of Economic Theory, 28, 1–18.
331
Ygge & Akkermans
Kurose, J. F., & Simha, R. (1989). A microeconomic approach to optimal resource allocation
in distributed computer systems. IEEE Transactions on Computers, 38 (5), 705–717.
Mas-Colell, A., Whinston, M., & Green, J. R. (1995). Microeconomic Theory. Oxford
University Press.
Mullen, T., & Wellman, M. P. (1995). A simple computational market for network informa-
tion services. In Lesser, V. (Ed.), Proceedings of the First International Conference
on Multi-Agent Systems, ICMAS’95, pp. 283–289 San Francisco, CA.
Ogata, K. (1990). Modern Control Engineering. Prentice-Hall, Englewood Cliffs, NJ. Second
Edition, ISBN 0-13-589128-0.
Press, W., Teukolsky, S., Vetterling, W., & Flannery, B. (1994). Numerical Recipies in C.
Cambridge University Press. Second Edition.
Sandholm, T. W., & Ygge, F. (1997). On the gains and losses of speculation in equilibrium
markets. In Proceeding of the Fifteenth International Joint Conference on Artificial
Intelligence, IJCAI 97, pp. 632–638.
Schreiber, A. T., Akkermans, J. M., & al. (1999). Knowledge Engineering and Management.
The MIT Press, Cambridge, MA. In Press.
Steiglitz, K., & Honig, M. L. (1992). Chaotic behavior in an auction-based microeconomic
model. Unpublished Manuscript.
Sutherland, I. E. (1968). A futures market in computer time. Communications of the ACM,
11 (6), 449–451.
Takayama, A. (1985). Mathematical Economics. Cambridge University Press.
Varian, H. R. (1992). Microeconomic Analysis. New York: W. W. Norton. Third Edition.
Varian, H. R. (1996). Intermediate Microeconomics—A Modern Approach. W. W. Norton
and Company, New York. Fourth Edition.
Walsh, W. E., Wellman, M. P., Wurman, P. R., & MacKie-Mason, J. K. (1998). Some
economics of market-based distributed scheduling. In Proceedings of the Eighteenth
International Conference on Distributed Computing Systems, pp. 612–621.
Wellman, M. P. (1993). A market-oriented programming environment and its application to
distributed multicommodity flow problems. Journal of Artificial Intelligence Research,
pp. 1–23.
Wellman, M. P. (1996). Market-oriented programming: Some early lessons. In Clearwater,
S. (Ed.), Market-Based Control: A Paradigm for Distributed Resource Allocation,
chap. 4. World Scientific.
Wellman, M. P. (1995). A computational market model for distributed configuration design.
AI EDAM, 9, 125–133.
332
Decentralized Market Control
Wellman, M. P., & Hu, J. (1998). Conjectural equilibrium in multiagent learning. Machine
Learning, 33, 179–200.
Yamaki, H., Wellman, M. P., & Ishida, T. (1996). A market-based approach to allocating
QoS for multimedia applications.
In Tokoro, M. (Ed.), Proceedings of the Second
International Conference on Multi-Agent Systems, ICMAS’96, pp. 385–392. AAAI
Press, Menlo Park, CA.
Ygge, F. (1998). Market-Oriented Programming and its Application to Power Load Man-
ISBN
agement. Ph.D. thesis, Department of Computer Science, Lund University.
91-628-3055-4, CODEN LUNFD6/(NFCS-1012)/1-224/(1998).
Ygge, F., & Akkermans, J. M. (1996). Power load management as a computational market.
In Tokoro, M. (Ed.), Proceedings of the Second International Conference on Multi-
Agent Systems, ICMAS’96, pp. 393–400. AAAI Press, Menlo Park, CA.
Ygge, F., & Akkermans, J. M. (1997). Making a case for multi-agent systems. In Boman,
M., & de Velde, W. V. (Eds.), Proceedings of MAAMAW ’97, pp. 156–176. Springer
Verlag, Berlin. ISBN-3-540-63077-5.
Ygge, F., & Akkermans, J. M. (1998). On resource-oriented multi-commodity market com-
putations. In Demazeau, Y. (Ed.), Proceedings of the Third International Conference
on Multi-Agent Systems ICMAS’98, pp. 365–371. IEEE Computer Society.
Ygge, F., Akkermans, J. M., Andersson, A., Krejic, M., & Boertjes, E. (1999). The Home-
Bots system and field tests: A multi-commodity market for predictive load manage-
ment. In Proceedings of the Fourth International Conference and Exhibition on The
Practical Application of Intelligent Agents and Multi-Agents (PAAM99), pp. 363–382.
Astrom, K.-J., & Wittenmark, B. (1990). Computer-Controlled Systems - Theory and
Design. Prentice Hall, Englewood Cliffs, NJ. ISBN 0-13-168600-3.
333
|
1203.2315 | 1 | 1203 | 2012-03-11T06:23:44 | Modeling multistage decision processes with Reflexive Game Theory | [
"cs.MA",
"cs.AI"
] | This paper introduces application of Reflexive Game Theory to the matter of multistage decision making processes. The idea behind is that each decision making session has certain parameters like "when the session is taking place", "who are the group members to make decision", "how group members influence on each other", etc. This study illustrates the consecutive or sequential decision making process, which consist of two stages. During the stage 1 decisions about the parameters of the ultimate decision making are made. Then stage 2 is implementation of Ultimate decision making itself. Since during stage 1 there can be multiple decision sessions. In such a case it takes more than two sessions to make ultimate (final) decision. Therefore the overall process of ultimate decision making becomes multistage decision making process consisting of consecutive decision making sessions. | cs.MA | cs |
Modeling multistage decision processes with
Reflexive Game Theory
Sergey Tarasenko
[email protected]
Abstract. This paper introduces application of Reflexive Game Theory
to the matter of multistage decision making processes. The idea behind is
that each decision making session has certain parameters like "when the
session is taking place", "who are the group members to make decision",
"how group members influence on each other", etc. This study illustrates
the consecutive or sequential decision making process, which consist of
two stages. During the stage 1 decisions about the parameters of the
ultimate decision making are made. Then stage 2 is implementation of
Ultimate decision making itself. Since during stage 1 there can be mul-
tiple decision sessions. In such a case it takes more than two sessions to
make ultimate (final) decision. Therefore the overall process of ultimate
decision making becomes multistage decision making process consisting
of consecutive decision making sessions.
Key words: Reflexive Game Theory (RGT), multi-stage decision mak-
ing
1 Introduction
The Reflexive Game Theory (RGT) [1, 2] allows to predict choices of subjects
in the group. To do so, the information about a group structure and mutual
influences between subjects is needed. Formulation and development of RGT
was possible due to fundamental psychological research in the field of reflexion,
which had been conducted by Vladimir Lefebvre [3].
The group structure means the set of pair-wise relationships between sub-
jects in the group. These relationships can be either of alliance or conflict type.
The mutual influences are formulated in terms of elements of Boolean algebra,
which is built upon the set of universal actions. The elements of Boolean algebra
represent all possible choices. The mutual influences are presented in the form
of Influence matrix.
In general, RGT inference can be presented as a sequence of the following
steps [1, 2]:
1) formalize choices in terms of elements of Boolean algebra of alternatives;
2) presentation of a group in the form of a fully connected relationship graph,
where solid-line and dashed-line ribs (edges) represent alliance and conflict rela-
tionships, respectively;
2
Sergey Tarasenko
3) if relationship graph is decomposable, then it is represented in the form of
polynomial: alliance and conflict are denoted by conjunction (·) and disjunction
(+) operations;
4) diagonal form transformation (build diagonal form on the basis of the
polynomial and fold this diagonal form);
5) deduct the decision equations;
6) input influence values into the decision equations for each subject.
Let us call the process of decision making in a group to be a session. Therefore,
in RGT models a single session.
2 Model of two-stage decision making: formation of points
of view
This study is dedicated to the matter of setting mutual influences in a group
by means of reflexive control [4]. The influences, which subjects make on each
other, could be considered as a result of a decision making session previous to
ultimate decision making (final session). We will call the influences, obtained
as a result of a previous session(s), a set-up influences. The set-up influences
are intermediate result of the overall decision making process. The term set-up
influences is related to the influences, which are used during the final session,
only.
Consequently, the overall decision making process could be segregated into
two stages. Let the result of such discussion (decision making) be a particular
decision regarding the matter under consideration. We assume the actual decision
making regarding the matter (final session - Stage 2) is preceded by preliminary
session (Stage 1). Stage 1 is about a decision making regarding the influences
(points of view), which each subject will support during the final session. We call
such overall decision making process to be a two-stage decision making process.
The general schema of a two-stage decision making is presented in Fig.1.
Fig. 1. The general schema of the two-stage decision making.
To illustrate such model we consider a simple example.
Example 1. Let director of some company has a meeting with his advisors.
The goal of this meeting is to make decision about marketing policy for the next
half a year. The background analysis and predictions of experts suggest three
distinct strategies: aggressive (action α), moderate (action β) and soft (action
γ) strategies. The points of view of director and his advisors are formulated
Modeling Mixed Groups of Humans and Robots
3
in terms of Boolean algebra of alternatives. Term point of view implies that a
subject makes the same influences on the others. Director supports moderate
strategy ({α}), the 1st and the 2nd advisors are supporting aggressive strategy
({β}), and the 3rd advisor defends the idea of soft strategy ({γ}). The matrix
of initial influences is presented in Table 1.
Table 1. Matrix of initial points of view (influences) used in Example 1
a
c
b
d
a a {α} {α} {α}
b {α} b {α} {α}
c {β} {β} c {β}
c {γ} {γ} {γ} d
Let director is in a conflict with all his advisors, but his advisors are in
alliance with each other. Variable c represents the Director, variables a, b and d
correspond to the 1st , the 2nd and the 3rd advisor, respectively.
The relationship graph is presented in Fig.2. Polynomial abd + c corresponds
to this graph.
Fig. 2. Relationship graph for a director-advisors group.
After diagonal form transformation the polynomial does not change:
[a][b][d]
[abd]
+[c]
1 + [c]
[abc + d]
= [abd + c]
= abd + c.
Then we obtain four decision equations and their solutions (decision intervals)
(Table 2).
ence matrix:
Next we calculate the decision intervals by using information from the influ-
subject a: (bd + c) ⊇ a ⊇ c ⇒ ({α}{γ} + {β}) ⊇ a ⊇ {β} ⇒ a = {β};
subject b: (ad + c) ⊇ b ⊇ c ⇒ ({α}{γ} + {β}) ⊇ b ⊇ {β} ⇒ b = {β};
subject c: 1 ⊇ c ⊇ abd⇒ 1 ⊇ c ⊇ {α}{α}{γ} ⇒ 1 ⊇ c ⊇ 0 ⇒ c = c;
subject d: (ab + c) ⊇ d ⊇ c ⇒ ({α}c + {β}) ⊇ b ⊇ {β} ⇒ {α, β} ⊇ d ⊇ {β}.
4
Sergey Tarasenko
Table 2. Decision intervals for Example 1
Subject Decision Equations Decision Intervals
a = (bd + c)a + ca (bd + c) ⊇ a ⊇ c
(ad + c) ⊇ b ⊇ c
b = (ad + c)b + cb
b
a
c
d
c = c + abdc
1 ⊇ c ⊇ abd
d = (ab + c)d + cd (ab + c) ⊇ d ⊇ c
Therefore, after the preliminary sessions, the point of view of the subjects
have changed. Director has obtained a freedom of choice, since he can choose
any alternative: 1 ⊇ c ⊇ 0 ⇒ c = c. At the same time, the 1st and the 2nd
advisors support moderate strategy (a = b = {β}). Finally, the 3rd advisor now
can choose between points of view {α, β} (aggressive of moderate strategy) and
{β} (moderate strategy): {α, β} ⊇ d ⊇ {β}.
Thus, the point of view of the 1st and the 2nd advisors is strictly determined,
while the point of view of 3rd advisor is probabilistic.
Next we calculate choice of each subject during the final session, considering
the influences resulting from the preliminary session. The matrix of set-up influ-
ences is presented in Table 3. The intervals in the matrix imply that a subject
can choose either of alternatives from the given interval as a point of view.
Table 3. The matrix of set-up influences for Example 1.
a
a
{β}
a
b
b
{β}
b
c
{β}
{β}
1 ⊇ c ⊇ 0
1 ⊇ c ⊇ 0
c
c {α, β} ⊇ d ⊇ {β} {α, β} ⊇ d ⊇ {β} {α, β} ⊇ d ⊇ {β}
c
d
{β}
{β}
1 ⊇ c ⊇ 0
d
Subject a:
d = {α, β}: (bd + c) ⊇ a ⊇ c ⇒({β}{α, β} + c) ⊇ a ⊇ c ⇒({β} + c) ⊇ a ⊇ c;
d = {β}: (bd + c) ⊇ a ⊇ c ⇒({β}{β} + c) ⊇ a ⊇ c ⇒({β} + c) ⊇ a ⊇ c;
Subject b:
d = {α, β}: (ad + c) ⊇ b ⊇ c ⇒({β}{α, β} + c) ⊇ b ⊇ c ⇒({β} + c) ⊇ b ⊇ c;
d = {β}: (ad + c) ⊇ b ⊇ c ⇒({β}{β} + c) ⊇ b ⊇ c ⇒({β} + c) ⊇ b ⊇ c;
Subject c:
d = {α, β}: 1 ⊇ c ⊇ abd ⇒1 ⊇ c ⊇ {β}{β}{α, β} ⇒1 ⊇ c ⊇ {β};
d = {β}: 1 ⊇ c ⊇ abd ⇒1 ⊇ c ⊇ {β}{β}{β} ⇒1 ⊇ c ⊇ {β};
Subject d:
Modeling Mixed Groups of Humans and Robots
5
(ab + c) ⊇ d ⊇ c ⇒({β}{β} + c) ⊇ d ⊇ c ⇒({β} + c) ⊇ d ⊇ c.
Now we compare the results of a single session with the ones of the two-stage
decision making.
The single session case has been considered above. If the final decision has to
be made after the single session, then the 3rd advisor would be able to choose
alternative {α, β} and realize action α. This option implies that each advisor is
responsible for a particular part of the entire company and can take management
decisions on his own.
Next we consider the decision made after the two-stage decision making.
In such a case, regardless of influence of the 3rd advisor (subject d), choice of
advisors a and b is defined by the interval {β} + c ⊇ x ⊇ c, where x is either a
or b variable. Thus, if director is inactive (c = 0), subjects a and b can choose
either moderate strategy ({β}) or make no decision (0 = {}). The same is true
for subject d.
If the director makes influence {β}, then all advisors will choose alternative
{β}.
The director himself can choose from the interval 1 ⊇ c ⊇ {β} after the final
session. This means that the director can choose any alternative, containing
action β. Thus, occasionally the director can realize his initial point of view as
moderate strategy.
This example illustrates how using the two-stage decision making it is possi-
ble to make ones opponents choose the ones point of view. Meanwhile a person
interested in such reflexive control can still sustain the initial point of view. The
obtained results are applicable in both cases when 1) only the director makes a
decision; or 2) the decision are made individually by each subject.
3 A Model of a multi-stage decision making: set-up
parameters of the final session
Now we consider the two-stage model in more details. In the considered example,
during the preliminary session only the decision regarding the influences has been
under consideration. In general case, however, before the final session has begun,
there can be made decisions regarding any parameters of the final session. Such
parameters include but are not limit to:
1) group structure (number of subjects and relationships between subjects
in a group);
2) points of view;
3) decision to start a final session (a time when the final session should start),
etc.
We call the decision regarding a single parameter to be consecutive decision,
and decisions regarding distinct parameters to be parallel decisions.
Therefore, during the first stage (before the final session) it is possible to
make multiple decisions regarding various parameters of the final session. This
decisions could be both parallel and consecutive ones. We call such model of
decision making to be a multi-stage process of decision making (Fig.3).
6
Sergey Tarasenko
Fig. 3. Multi-stage decision making model.
4 Modeling multi-stage decision making processes with
RGT
Next we consider realization of multi-stage decision making with RGT.
Example 2: Change a group structure. Considering the subject from Example
1, we analyze the case when director wants to exclude the 3rd advisor from the
group, which will make the final decision.
In such a case, there is a single action 1 to exclude subject d from the
group. Then Boolean algebra of alternatives includes only two elements: 1 and
0. Furthermore, it is enough that director just raise a question to exclude subject
d from a group and make influence 1 on each subject: if = 1, then a = 1, b = 1
and d = 1 (Table 2). Thus the decision to exclude subject d from the group
would be made automatically (Fig.4).
Fig. 4. Exclusion of a subject d from a group.
Example 3: Realization of a multi-stage decision making. Let the first deci-
sion discussed during the first stage is a decision regarding influences (points
of view). The next decision was about exclusion of a subject d from the group.
Thus, during the first step the formation (setting-up) of points of view has been
implemented, then the structure of a group was changed. Therefore the group,
which should make a final decision is described by polynomial ab+c. The decision
equations and their solutions are presented in Table 4.
The overall multi-stage decision making process is presented in Fig.5.
We consider that the point of view cannot change without preliminary session
regarding the parameter. Therefore we assume that the points of view do not
change after the change of group structure.
Modeling Mixed Groups of Humans and Robots
7
Table 4. Decision intervals for Example 3
Subject Decision Equations Decision Intervals
(b + c) ⊇ a ⊇ c
(a + c) ⊇ b ⊇ c
a = (b + c)a + ca
b = (a + c)b + cb
a
b
c
c = c + abc
1 ⊇ c ⊇ ab
Fig. 5. Illustration of multi-stage decision making process. The influences are indicated
by the arrow-ends of the ribs. The actual influence is presented near the arrow-end.
Therefore, during the final session the subjects would make the set-up influ-
ences derived from the preliminary session: subjects a and b will make influences
{β} and subject c will have a choice from the interval 1 ⊇ c ⊇ {β}.
Such process is introduced in Fig. 5. During the 1st stage (first step), the
points of view of subjects have been formed. On the 2nd stage (second step), the
decision to exclude subject d from a group has been made. Finally, during the
3rd stage the final decision regarding the marketing strategy has been made.
5 Discussion and conclusion
This study introduces the two-stage and multi-stage decision making processed.
During the last stage the final decision is made. During the earlier stages the
decisions regarding the parameters of a final session are considered.
This study shows how before the final decision making the intermediate de-
cision regarding parameters of the final session can be made and how to overall
8
Sergey Tarasenko
process of decision making could be represented as a sequence of decision making
sessions.
This approach enables complex decision making, which involves numerous
parameters.
The important feature of the multi-stage decision making is that during the
preliminary sessions subjects can convince other subjects to accept their own
point of view. Therefore other subjects can be convince to make decisions bene-
ficial for a particular one. Such approach also allows to distribute the responsi-
bility between all the members of the group, who make the final decision.
The results presented in this study allow to extend the scope of applications of
RGT to modeling of multi-stage decision making processes. Therefore it becomes
possible to perform scenario analysis of various variants of future trends and
apply reflexive control to the management of projects.
References
1. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Cogito-Centre, Moscow (2009)
(in Russian)
2. Lefebvre, V.A.: Lectures on Reflexive Game Theory. Leaf & Oaks, Los Angeles
(2010)
3. Lefebvre, V.A.: Algebra of Conscience. D. Reidel, Holland (1982)
4. Lefebvre, V.A.: The basic ideas of reflexive game's logic. Problems of systems and
structures research. 73 -- 79 (1965) (in Russian)
|
1712.05987 | 1 | 1712 | 2017-12-16T16:44:26 | Priority Rules on ATN (PRT) Intersections | [
"cs.MA"
] | In Autonomous Transit Networks some basic elements influence the throughput: network structure, maximum velocity, number of vehicles etc. Other parameters like station structure, dynamic routing or vehicle behavior on intersections play minor role. Yet in highly congested nets, when vehicles interfere in the traffic, some subtle decisions may influence overall system ridership. We tested the impact of intersection priority rules on passenger waiting time, which measures the throughput. The dependence occurred its relevance in a crowded network. | cs.MA | cs | Waldemar Grabski, Wiktor B. Daszczuk
I
PRIORITY RULES ON ATN (PRT) INTERSECTIONS
In Autonomous Transit Networks some basic elements influence the throughput: network structure, maximum velocity,
number of vehicles etc. Other parameters like station structure, dynamic routing or vehicle behavior on intersections play
minor role. Yet in highly congested nets, when vehicles interfere in the traffic, some subtle decisions may influence overall
system ridership. We tested the impact of intersection priority rules on passenger waiting time, which measures the
throughput. The dependence occurred its relevance in a crowded network.
INTRODUCTION
Autonomous Transit Network ATN (Personal Rapid Transit
PRT) is a communication means using autonomous guided vehicles
traveling on a track separated from urban traffic (typically elevated)
[1]. It consists of nodes: stations, capacitors and intersections,
interconnected by track segments. We assume two types of
segments: road with lower maximum velocity and highway with
higher maximum velocity. Moreover, highway segments to dot
connect stations nor capacitors, only intersections are allowed. The
intersections are of two types: "fork" and "join" ("x" type intersections
are not allowed).
A research concentrates on effectiveness of Personal Rapid
Transit [2][3][4][5][6]. If a network works in equilibrium conditions [7],
the main method of performance improvement is empty vehicle
management [5][7][8] (as little may be done with occupied ones: the
only option is ride-sharing [9][10][11]). Yet, if the number of vehicles
causes congestion, some subtle decision may improve the traffic
one-way road
station
E
F
L
suburbs
roundabout
G
B
city
A
C
D
K
Fig. 1. „City" benchmark
two-way highway
H
capacitor
I
J
3000m
1
I
conditions. We have
the
throughput of ATN network: dynamic routing and intersection priority
rules.
two concepts of
tested
increasing
The question arises: how to identify congestion conditions? Of
course, inserting many vehicles to the network causes they
interference and the throughput drops down. But for research we
need a single parameter that measures congestion degree, and a
single value of this parameter (a threshold) which identifies the
network is overcrowded. We decided to use average squared delay
for this purpose.
The research on dynamic routing is based on insertion of
"Sunday driver" vehicles, which drive slowly and block other
vehicles. As the only point of decision in ATN vehicle run is "fork"
type intersection, we run the route planning algorithm in front of
every such intersection. The Dijkstra planning algorithm [12] is
equipped with segment cost function dependent on segment length,
segment maximum velocity and segment occupation, everyone with
coefficient parameter, The experiments gave inconclusive results,
as output parameter (average passenger waiting time) is not
monotonic as a function of the segment occupation coefficient.
The second investigation concerns priority rules on "join"
intersections. If two vehicles approach an intersection, and if driving
with their actual velocities would cause their bringing too close to
each other (closer than assumed separation), one of the vehicles
must slow down or even stop. Of course, other vehicles that
possibly follow this slowing down vehicle must slow down as well.
This causes a congestion effect.
to
the both segments). The
We assume the fair rule on intersection of one type segments,
if they are both road segments of both highway segments. The
is achieved using slider principle (priority assigned
fairness
alternately
research concerns
intersection of different type segments: road and highway. We
tested the variants of choosing a vehicle which is not forced to slow
down in described situation: on highway, on road or slider principle.
In the research we use the Feniks simulation tool [13],
elaborated in Eco-Mobility project [14]. In the project, several
simulation-based research programs were carried [7][8][15][16]. The
Feniks simulator is still under development, a next step concerns
dual-mode operation, in which the vehicle may leave PRT network
and continue their trips in ordinary urban traffic.
In Section 1 the benchmark network is presented. Section 2
describes the manner of identification of network saturation point.
Simulation results are covered in Section 3. The research is
concluded in Section 4.
1. THE BENCHMARK
The "City" benchmark [17] is presented in Fig. 1. Double-line
segments are two-way highways (two tracks in opposite directions).
Single-line segments are ordinary roads. Circles with letters are
stations while dashed squares are capacitors. Bold circles are
roundabouts allowing to change direction to any of four outcoming
track segments.
It is a model of a town, with central city and suburbs. A highway
surrounds the city and connects suburbs with each other and with
the city. There are 12 stations and 4 capacitors. Total track length is
about 33km.
The traffic parameters are:
– 5 berths in every station,
– Highway max velocity 15m/s,
– Road max velocity 10m/s,
– Max acceleration and deceleration 2m/s2,
– Boarding and alighting: triangle distribution with parameters
(10, 20, 30) seconds,
– separation 10m,
– Origin-destination matrix a matrix filled with equal values
(every trip target chosen randomly).
2.
IDENTIFICATION OF NETWORK SATURATION POINT
Saturation point is a situation in which the vehicles interfere
„disturbing" each other. The disturbing reveals in enlarging of a trip
time.
In papers concerning ATN simulation, some measures of traffic
quality are used. For example, maximum passenger waiting time or
average waiting time is used [18]. Yet, we prefer a single scalar
which unambiguously measures
the network saturation. We
elaborated a synthetic parameter: Average Squared Delay (ASD [7])
(1)
(square root of sum of squared relative delays of full trips (ft)
divided by number of full trips). The delay ft (in [%]) is counted as
actual trip time divided by nominal trip time (with maximal allowed
velocity). When the vehicles do not interfere, in typical PRT network
ASD is about 15%. A designer may designate a threshold of
average squared delay over which the network is assumed to be
congested. The results of ASD for a number of vehicles between 48
and 320 is shown in Table 1 and illustrated in Fig. 2. In our
example, having the basis of ASD=15%, we may define that the
network is congested if ASD rises 25%. This is reached at the
number of vehicles about 240.
Tab. 1. Identification of network saturation point
Number of vehicles
Average Squared Delay [%]
48
72
96
120
144
192
240
280
320
192
15.29
15.71
16.35
16.87
18.08
20.28
24.77
28.76
35.97
20.28
2
I
Fig. 2. Identification of network saturation point
3. SIMULATION RESULTS
For the „City" model (Fig. 1) 240 vehicles were used in
simulations (saturation point). The three input intensities were used:
320, 480 and 960 groups/h (one trip is organized for every group of
passengers of cardinality 1-4 persons). For every input intensity, all
three priority rules were tested:
– Priority for highway,
– Slider principle (alternating priority for highway and for
road),
– Priority for road.
The results show influence of priority rules on the throughput of
PRT network. The influence rises in the input intensity. It is
reasonable because higher input results in more trips to be
organized and the actual congestion of the network is higher.
Because the differences are not clearly visible for input races 320
and 480 groups/h, they are enlarged in Fig. 4. The plots clearly
show, that taking the "priority for road" improves the throughput by
9-30% compared with "priority for highway". Furthermore, "slider"
improves by 11-14% compared with "priority for highway". This
unintuitive behavior (we expected that the highway should be
favored) may be explained such that vehicle queues waiting to
cross the intersection are easier to create on the roads because
there is a lower maximum speed than on the highway.
4. CONCLUSIONS
In the design of a PRT network, many decisions should be
made: on a location of stations and capacitors, on the segments
category (road/highway), which may differ in curvature and slope,
stations size and number of vehicles. In the paper we showed that
in the network working in congestion conditions, even subtle traffic
rules may tune the throughput of ATN system. We tested priority
rules, but station capacity, dynamic routing (with other algorithm
than ours) and similar mechanisms may also improve the quality.
BIBLIOGRAPHY
1.
2.
J. Irving, Fundamentals of Personal Rapid Transit, Lexington
Books, Lexington, MA, 1978. ISBN: 0-669-02520-8
I. Andréasson, Reallocation of Empty PRT vehicles en route,
in: TRB Annu. Meet. Washingt. DC, 12-16 January 2003,
Transportation Research Board, 2003: pp. 1–13.
3. P. Zheng, D. Jeffery, M. McDonald, Development and
evaluation of traffic management strategies for personal rapid
transit, in: Ind. Simul. Conf. 2009, Loughborough, UK, 1-3 June
2009, 2009: pp. 191–195.
4. M. van der Heijden, M. Ebben, N. Gademann, A. van Harten,
Scheduling vehicles in automated transportation systems, OR
Spectr. 24 (2002) 31–58. doi:10.1007/s291-002-8199-x.
J.D. Lees-Miller, R.E. Wilson, Proactive empty vehicle
redistribution for personal rapid transit and taxis, Transp. Plan.
Technol.
17–30.
doi:10.1080/03081060.2012.635414.
(2012)
5.
35
6. O. Chebbi, J. Chaouachi, Reducing the wasted transportation
capacity of Personal Rapid Transit systems: An integrated
model and multi-objective optimization approach, Transp. Res.
Part E Logist. Transp. Rev. 89
(2016) 236–258.
doi:10.1016/j.tre.2015.08.008.
7. W.B. Daszczuk, J. Mieścicki, W. Grabski, Distributed algorithm
for empty vehicles management in personal rapid transit (PRT)
3
I
network,
Adv.
doi:10.1002/atr.1365.
J.
Transp.
50
(2016)
608–629.
8. W.B. Daszczuk, W. Choromański, J. Mieścicki, W. Grabski,
Empty vehicles management as a method
for reducing
passenger waiting time in Personal Rapid Transit networks,
IET Intell. Transp. Syst. 9 (2015) 231–239. doi:10.1049/iet-
its.2013.0084.
I. Andréasson, Ride-Sharing on PRT, in: 10th Autom. People
Movers, Orlando, Florida, 1-4 May 2005, American Society of
Civil
1–7.
doi:10.1061/40766(174)22.
Engineers,
Reston,
2005:
VA,
pp.
9.
10. I.J. Andreasson, Extending PRT Capabilities, in: 12th Autom.
People Movers, Atlanta, GA, 31 May - 3 June 2009, American
Society of Civil Engineers, Reston, VA, 2009: pp. 343–349.
doi:10.1061/41038(343)29.
11. R. Johnson, Doubling Personal Rapid Transit Capacity with
Ridesharing, Transp. Res. Rec. J. Transp. Res. Board. 1930
(2005) 107–112. doi:10.3141/1930-13.
12. E.W. Dijkstra, A note on two problems in connexion with
269–271.
(1959)
Math.
1
graphs,
doi:10.1007/BF01386390.
Numer.
15. M. Kozłowski, W. Choromański, J. Kowara, Parametric
sensitivity analysis of ATN-PRT vehicle (Automated transit
network – personal rapid transit), J. Vibroengineering. 17
(2015)
1436–1451.
https://www.scopus.com/inward/record.uri?eid=2-s2.0-
84946555953&partnerID=40&md5=1b3eb293ddc7bca0d8e9cc
fba1974cdb.
16. W. Choromański, I. Grabarek, M. Kozłowski, Simulation and
Experimental Study of Selected Parameters of
the
Multifunction Steering Wheel in the View of Users' Abilities and
Accuracy of Vehicle Maneuvers, Procedia Manuf. 3 (2015)
3085–3091. doi:10.1016/j.promfg.2015.07.855.
17. J. Mieścicki, W.B. Daszczuk, Proposed benchmarks for PRT
networks simulation, Arch. Transp. 27–28 (2013) 123–133.
18. J. Lees-Miller, J. Hammersley, R. Wilson, Theoretical
Maximum Capacity as Benchmark
for Empty Vehicle
Redistribution in Personal Rapid Transit, Transp. Res. Rec. J.
Transp. Res. Board. 2146 (2010) 76–83. doi:10.3141/2146-10.
13. W.B. Daszczuk, Discrete Event Simulation of Personal Rapid
Transit (PRT) Systems, in: Autobusy-TEST, 2016(6), 2016: pp.
1302–1310. ArXiv: 1705.05237
14. W.
Choromański,
Eco-Mobility,
http://www.eco-
mobilnosc.pw.edu.pl/?sLang=en.
Zasady pierwszeństwa na skrzyżowaniach sieci ATN(PRT)
W sieciach transportowych ATN (Autonomous Transit
Network) na przepustowość mają wpływ jej główne elementy
jak struktura sieci, maksymalna prędkość, liczba pojazdów
itd. Inne parametry jak struktura przystanków, dynamiczny
Fig. 3. The effect of priority rules on the „join" intersection to
average waiting time, with different input rates
Fig. 4. Plot from Fig. 3, except for the case of the largest input
4
I
wybór drogi czy zachowania pojazdów na skrzyżowaniach
grają niewielką rolę. Jednak przy dużym zagęszczeniu
pojazdy przeszkadzają sobie wzajemnie i pewne subtelne
decyzje mogą wpłynąć na zdolność przewozową sieci.
Zbadaliśmy wpływ zasad pierwszeństwa na skrzyżowaniach
na czas oczekiwania na pojazdy, który
jest miarą
przepustowości. Zależność okazała się istotna w zatłoczonej
sieci.
Authors:
Waldemar Grabski, MSc – Warsaw University of Technology,
Institute of Computer Science, Nowowiejska str. 15/19, 00-665
Warsaw, [email protected]
Wiktor B. Daszczuk, PhD – Warsaw University of Technology,
Institute of Computer Science, Nowowiejska str. 15/19, 00-665
Warsaw, [email protected]
5
|
1510.00109 | 3 | 1510 | 2016-03-25T02:17:22 | Confinement Control of Double Integrators using Partially Periodic Leader Trajectories | [
"cs.MA",
"eess.SY",
"math.OC"
] | We consider a multi-agent confinement control problem in which a single leader has a purely repulsive effect on follower agents with double-integrator dynamics. By decomposing the leader's control inputs into periodic and aperiodic components, we show that the leader can be driven so as to guarantee confinement of the followers about a time-dependent trajectory in the plane. We use tools from averaging theory and an input-to-state stability type argument to derive conditions on the model parameters that guarantee confinement of the followers about the trajectory. For the case of a single follower, we show that if the follower starts at the origin, then the error in trajectory tracking can be made arbitrarily small depending on the frequency of the periodic control components and the rate of change of the trajectory. We validate our approach using simulations and experiments with a small mobile robot. | cs.MA | cs | Confinement Control of Double Integrators using Partially Periodic
Leader Trajectories
Karthik Elamvazhuthi, Sean Wilson, and Spring Berman
6
1
0
2
r
a
M
5
2
]
A
M
.
s
c
[
3
v
9
0
1
0
0
.
0
1
5
1
:
v
i
X
r
a
Abstract- We consider a multi-agent confinement control
problem in which a single leader has a purely repulsive effect
on follower agents with double-integrator dynamics. By decom-
posing the leader's control inputs into periodic and aperiodic
components, we show that the leader can be driven so as to
guarantee confinement of the followers about a time-dependent
trajectory in the plane. We use tools from averaging theory and
an input-to-state stability type argument to derive conditions
on the model parameters that guarantee confinement of the
followers about the trajectory. For the case of a single follower,
we show that if the follower starts at the origin, then the error
in trajectory tracking can be made arbitrarily small depending
on the frequency of the periodic control components and the
rate of change of the trajectory. We validate our approach using
simulations and experiments with a small mobile robot.
I. INTRODUCTION
In this work, we present a leader-follower control strategy
for multi-robot applications such as exploration, environmen-
tal monitoring, disaster response, and targeted drug delivery
at the micro-nanoscale. We consider a scenario in which
a leader agent must confine a group of follower agents to
a certain region around a target trajectory while steering
them along this trajectory. For example, a leader agent with
sophisticated sensing, localization, and planning capabilities
may be required to herd a group of followers, which lack
these capabilities, through an environment with obstacles.
The problem that we consider is closely related to con-
tainment control problems [31] in the multi-agent systems
literature. In these problems,
the objective is to design
interaction rules between leader and follower agents so that
the followers are eventually contained in the convex hull
spanned by the leaders. In contrast to most other work on
such problems,
the system that we consider is naturally
unstable for steady-state control inputs due to the repulsive
nature of the interaction between the leader and the followers.
Hence, a stabilization mechanism is needed to achieve con-
finement of the followers. One way to stabilize the system
is to incorporate feedback on the followers' positions and
velocities into the leader trajectories. Our control approach
does not require this feedback, and instead uses open-loop
oscillatory strategies to (practically) stabilize the follower
positions about near-zero velocity conditions.
The use of oscillatory inputs in control theory has a rich
history. There has especially been extensive work on motion
*This research was supported by NSF Award CMMI-1436960 and by
DARPA Young Faculty Award D14AP00054.
The authors are with the School
for Engineering of Matter,
State University, Tempe, AZ
E-mail: {karthikevaz, sean.t.wilson,
and Energy, Arizona
Transport
85287, USA.
spring.berman}@asu.edu
planning of driftless systems. In this work, sinusoidal inputs
of appropriately chosen frequencies are used to produce exci-
tations along different independent directions (corresponding
to Lie brackets of system vector fields) in the system's state
space [37], [24], [11]. In fact, for a certain class of systems
with a specific geometric structure, solutions of some optimal
control problems are of a periodic nature [6], [34]. Similarly,
oscillatory inputs have also played an important role in
stabilization problems [1], [2], [39], [13], [33]. While most
of these works have considered the control of systems with a
range of rest conditions (equilibrium points), there has also
been some work on practical stabilization of systems for
which no such rest condition exists [23]. The application of
oscillatory inputs can also be found in the physics literature.
Kapitza's pendulum is an example of a system whose unsta-
ble equilibrium is rendered stable when subject to oscillatory
vibrations [20]. The method of using oscillatory potentials
for trapping charged particles is also well-known [14], [26].
Our approach is also motivated by animal foraging and
livestock herding behaviors, which have inspired a number of
algorithms and multi-robot control strategies [17], [28], [3],
[22], [12], [5]. The approaches in [17], [5], and [12] are most
similar to ours in their use of sinusoidal leader trajectories
that
introduce repulsion terms into the agent dynamics.
While [17] uses sinusoidal trajectories to mimic the foraging
strategies of dolphins, the analysis is based on Snell's law
and the agent interactions arise from prey "bouncing" off of
predators by specular reflection. In [5] and [12], the follower
population is modeled as a probability density over space and
time, and the objective is to control the support of this density
function. The followers are diffusing at a finite speed, and
their models are of single-integrator type. These two works
derive conditions for confinement to be possible under these
assumptions. Attraction-type potentials are also considered
in [12], and conditions are also derived for the confinement
of followers in dimensions greater than two. Although these
works do not consider the trajectory tracking problem, [5]
does study the existence of control inputs for point-to-point
steering problems.
Other works that consider similar problems use methods
based on consensus protocols and multiple leaders [18], [8].
Multi-agent controllability problems have also received some
attention in related contexts [9], [29], [38], [27], [7].
II. PROBLEM STATEMENT
We consider a system with a single leader agent, whose
position at time t is given by [xl(t),yl(t)]T , and N follower
agents, whose positions at
i =
time t are [xi(t),yi(t)]T ,
1, ...,N. We denote the x and y velocity components of
follower agent i at time t by [vi
y(t)]T . The dynamics of
the leader and followers are defined by the following system
of equations:
x(t),vi
(1)
xl = ux
yl = uy
xi = vi
x
vi
x =
yi = vi
y
A(xi − xl)
[(xi − xl)2 + (yi − yl)2]α − kvi
x
vi
y =
A(yi − yl)
[(xi − xl)2 + (yi − yl)2]α − kvi
y
where A,k,α ∈ R+ and i = 1, ...,N. This model produces
one-way interaction between the leader and the followers
for any non-zero value of A. The interaction potentials are
of gravitational type but repulsive in nature.
We formulate our confinement control problem using the
following definition, in which (cid:107)x(cid:107) is the Euclidean norm of
x ∈ Rn:
Definition II.1. Let γ : [0,1] → R2 and R > 0. Then the
system (32) is said to be R-confinement controllable about
the trajectory γ if there exist T > 0, ux : [0,T ] → R, and uy :
[0,T ] → R such that the solution of (32) satisfies (cid:107)γ(t/T )−
[xi(t),yi(t)]T(cid:107) < R and (cid:107)γ(t/T )− [xl(t),yl(t)]T(cid:107) ≥ R for all
i ∈ {1, ...,N} and all t ∈ [0,T ].
Our objective is to derive conditions on the parameters A,
k, and α and on the trajectory γ(t) and its time derivative
γ(t) which guarantee that system (32) is R-confinement
controllable about the trajectory γ. The requirement that the
leader must maintain a minimum distance from the trajectory
differentiates our approach from other confinement strategies
that use a single leader, which typically employ attraction-
type interactions between the leader and followers. Moreover,
this distance requirement provides a way to prevent collisions
between the leader and the followers without explicitly
modeling the physical dimensions of the agents.
System (32) can be expressed in the affine control form
x = f(x) + ∑i uigi(x), which enables the application of Lie
algebraic conditions of geometric control to study its con-
trollability properties [35], [36]. However, the system has
no fixed point for any time-independent set of control
inputs {ui}. Therefore, one cannot conclude much more
than accessibility at points of interest. While Lie algebraic
conditions for controllability do exist for the case of non-rest
conditions, they require the assumption of global bounds on
the vector fields [15], which is not a valid assumption for
system (32). Alternatively, certain results on controllability
about trajectories can be applied if the admissibility of the
trajectories can be characterized beforehand [4].
We take a different approach by using averaging theory
to show that
the
target trajectory with a particular choice of control inputs.
Averaging theory is built on the principle that the behavior of
the followers can be stabilized about
a dynamical system with a rapidly oscillating vector field can
be accurately represented by the averaged system behavior
over a time interval that is dependent on a perturbation
parameter ε. For small enough ε, one can relate the solutions
of the averaged system, a set of autonomous differential
equations, to the solutions of the original non-autonomous
differential equations. Moreover, the stability of the solutions
is easier to characterize in the averaged system than in the
original system. We apply averaging analysis to our system
in the next section.
III. AVERAGING ANALYSIS
In the forthcoming analysis, we prove that system (32)
is R-confinement controllable about a trajectory γ(t) under
suitable bounds on γ(t) and γ(t). We also demonstrate that
if a follower starts at the origin, then its distance from the
trajectory can be made arbitrarily small while the leader
maintains a distance R from the trajectory. To simplify the
analysis, we consider the case where R = 1 and N = 1,
i.e., there is a single follower. We show that the analysis
also holds for any follower population N, since there are no
interactions between the followers. The results for general
R > 0 follow trivially from an appropriate scaling of the
parameter A.
Denote the initial position of the leader by [xl(0),yl(0)]T =
[1,0]T , and let ux(t) = −ω sinωt and uy(t) = −ω cosωt for
some ω > 0. With these control inputs, one can explicitly
solve for xl(t) and yl(t). Then system (32) can be expressed
in the following reduced form,
x =
vx =
y =
vy =
A(x− cosωt)
vx
[(x− cosωt)2 + (y− sinωt))2]α − kvx
vy
[(x− cosωt)2 + (y− sinωt))2]α − kvy.
A(y− sinωt)
(2)
We change the time variable to τ = ωt and define ε = 1/ω.
Then the solution of (2) satisfies,
(3)
dx
dτ = εvx
dvx
dτ =
dy
dτ = εvy
dvx
dτ =
εA(x− cosτ)
[(x− cosτ)2 + (y− sinτ)2]α − εkvx
εA(y− sinτ)
[(x− cosτ)2 + (y− sinτ)2]α − εkvy.
Now the system can be expressed in the so-called standard
averaging form,
where X = [x,vx,y,vy]T and
f(τ,X) =
dX
dτ = εf(τ,X),
vx
A(x−cosτ)
[(x−cosτ)2+(y−sinτ)2]α − kvx
[(x−cosτ)2+(y−sinτ)2]α − kvy
vy
A(y−sinτ)
(4)
(5)
.
Then we can consider the averaged system,
¯X(0) = ¯X0,
¯X = εfav( ¯X)
where ¯X = [ ¯x, ¯vx, ¯yy, ¯vy]T and
(cid:82) 2π
(cid:82) 2π
0
0
1
2π
1
2π
fav( ¯X) =
¯vx
A( ¯x−cos(s))
[( ¯x−cos(s))2+( ¯y−sin(s))2]α ds− k ¯vx
[( ¯x−cos(s))2+( ¯y−sin(s))2]α ds− k ¯vy
A( ¯y−sin(s))
¯vy
(6)
(7)
.
If the origin of the averaged system is an exponentially
stable equilibrium point, then there exists a neighborhood
of the origin for which solutions of the original system
remain bounded and close to the origin for infinite time. The
following lemma gives conditions for exponential stability
of the origin in the averaged system. The existence and
uniqueness of solutions to the averaged system follow from
this result for initial conditions sufficiently close to the
origin.
Lemma III.1. The origin is an exponentially stable equilib-
rium point of the averaged system (6) for all k > 0 if A > 0
and α > 1.
Proof. The proof is based on the linearization principle. The
Jacobian map for fav( ¯X) at
¯X = 0 is given by (see the
Appendix for computations):
0
0
0
0
0
0
Dfav(0) =
0
1
−A(α − 1) −k
0
0
1
0 −A(α − 1) −k
The spectrum of Dfav(0) can be computed to be
(k2 − 4A(α − 1))
Then the result follows from [21][Corollary 4.3].
σ (Dfav(0)) = − k
2
± 1
2
1
2 .
.
(8)
(9)
The following proposition demonstrates that this result
can be used to obtain estimates for the solutions of system
(4). The proposition gives bounds on solutions of general
systems in which the vector field can be decomposed into
(1) a highly oscillatory, state-dependent component, and (2)
a state-independent component, which is not necessarily pe-
riodic, that evolves on a slower time scale. By converting our
system into this form through a coordinate transformation,
we can use this result to construct control inputs for the
leader to confine followers about a trajectory. The proof is
similar to the proofs of the averaging theorems in [21], [16].
Since our result does not immediately follow from these
theorems, we adapt the initial steps of their proofs to our
problem.
Proposition III.2. Let a function f(t,x,ε) and its partial
derivatives with respect to (x,ε) up to the first order be
continuous and bounded for (t,x,ε) ∈ [0,∞)×U0× [0,ε0] for
every compact set U0 ⊂ U, where U ⊂ Rn, and some ε0 > 0.
Additionally, let g : [0,∞) → Rn be continuous. Suppose that
f is T -periodic in t for some T > 0. Then x ≡ x(t,ε,g) and
z ≡ z(t) evolve according to the following equations:
x = εf(t,x,ε) + εg(t),
x(0,ε,g) = x0,
(10)
(cid:82) T
z(0) = z0,
z = εfav(z),
(11)
where fav(z) = 1
0 f(τ,z,0)dτ. Suppose that the origin z =
T
0 ∈ U is an exponentially stable equilibrium point of the
averaged system Eq. (11), Ω ⊂ U is a compact subset of its
region of attraction, and x0 ∈ Ω. Then there exist positive
constants k, δ∗, ε∗, β1, β2, and µ such that if (cid:107)g(t)(cid:107) ≤ δ for
all t ≥ 0, then the condition (cid:107)x(0,ε,g)− z(0)(cid:107) ≤ µ implies
that
(cid:107)x(t,ε,g)− z(t)(cid:107) ≤ ke−λεt(cid:107)x0 − z0(cid:107) + β1ε + β2δ
(12)
for all ε ∈ (0,ε∗), δ ∈ (0,δ∗), and t ≥ 0.
Proof. Define the functions
h(t,x) = f(t,x,0)− fav(x)
and
u(t,x) =
h(τ,x)dτ.
(cid:90) t
0
(13)
(14)
Since h(t,x) is T -periodic in t and has zero mean,
the
function u(t,x) is T -periodic in t. Hence, u(t,x) is bounded
for all (t,x)∈ [0,∞)×Ω. Moreover, ∂u/∂t and Dxu are given
by
∂u
∂t = h(t,x),
Dxu =
(cid:90) t
0
Dxh(τ,x)dτ.
Consider the change of variables
x = y + εu(t,y).
(15)
(16)
Differentiating both sides of Eq. (16) with respect to t, we
obtain
x = y + ε
∂u
∂t (t,y) + εDyu(t,y)y.
(17)
∂u
∂t
= εf(t,y + εu,ε) + εg(t)− εf(t,y,0)
Substituting Eq. (10) for x into Eq. (17), the new state
variable y satisfies the equation
(cid:2)I + εDyu(cid:3)y = εf(t,y + εu,ε) + εg(t)− ε
(t,x) ∈ [0,∞)×U0 imply that for small enough ε,(cid:2)I +εDyu(cid:3)
y = εfav(y) + εg(t) + ε(cid:2)εDyf(t,y,0)u + ε
−εDyufav(y)− εDyug(t)(cid:3) + O(ε3)
The assumption of bounds on the partial derivatives of f over
is invertible. Hence, the state variable y satisfies
∂f
∂ε (t,y,0)
+εfav(y).
:= εfav(y) + εg(t) + ε2q(t,y,ε,g),
(18)
where q(t,y,ε,g) is not necessarily periodic. From the
assumptions on f(t,x,ε) and g(t), q(t,y,ε,g) is continuous
and bounded for (t,x,ε) ∈ [0,∞)× Ω× [0,ε0].
Now consider the change of variables s = εt. Then the
solutions of Eq. (18) and Eq. (11) satisfy
dy
ds = fav(y) + g(s) + εq(s/ε,y,ε,g)
and
dz
ds = fav(z),
(19)
(20)
the origin z = 0 is an ex-
respectively. By assumption,
ponentially stable equilibrium point of Eq. (20). Then by
[21][Theorem 9.1], there exist positive constants ε∗, δ∗, µ,
β1, β2, δ , k, and λ such that the conditions (cid:107)g(s)(cid:107) < δ and
(cid:107)y0 − z0(cid:107) ≤ µ imply that,
(cid:107)y(s)− z(s)(cid:107) ≤ ke−λ s(cid:107)y0 − z0(cid:107) + β1ε + β2c
(21)
for all ε ∈ (0, ε∗), δ ∈ (0,δ∗), and s ≥ 0. For small enough
ε, the transformation (16) is a one-to-one mapping from
Ω to its image for all t ≥ 0 and ε ∈ [0,εo] for some εo >
0 [32][Lemma 2.8.3]. Therefore, x − y = O(ε). Hence, it
follows from the triangle inequality and the estimate (21) that
there exist positive constants δ∗, ε∗, µ, and c such that the
conditions (cid:107)g(t)(cid:107) ≤ δ and (cid:107)x(0,ε,g)−z(0)(cid:107) ≤ µ imply that
Eq. (12) is true for all ε ∈ (0,ε∗), δ ∈ (0,δ∗), and t ≥ 0.
Note that this result is not the same as general averaging,
which is a generalization of the above proposition for the
case where g = 0 and f(t,x,ε) is required to be only almost-
periodic in t [21], [32], since we have explicit bounds on the
solution of Eq. (11) that depend on ε as well as δ . One can
impose the additional assumption that f(t,x,ε) has partial
derivatives up to the second order. Under this assumption,
g = 0 implies the existence of exponentially stable periodic
orbits that approach the equilibrium point of the averaged
system as ε tends to 0. See, for example, [16]. In this
case, one can interpret
the result as "uniform" input-to-
state stability of the system about ε-dependent exponentially
stable T -periodic orbits. Our result can also be seen as
a special case of the results in [25], where input-to-state
stability of general time-varying systems via averaging is
analyzed using the notion of strong and weak averages.
In the next two results, we use Proposition III.2 to show
the existence of leader control inputs that can locally stabilize
the followers about general trajectories in the plane that start
at the origin. We consider leader inputs with an oscillatory
component, as in Eq. (3), and an additional component
that depends on the target trajectory to be tracked. Using a
coordinate transformation, we show that this system can be
converted to the form in Eq. (11), and hence we can obtain
bounds on the distances between the follower positions and
the target trajectory.
Remark on notation: We say that a function f : U → R, where
U ⊂ Rn, is in C1(U) if it has continuous derivatives up to
order 1.
Lemma III.3. Suppose that γ1 and γ2 are elements in
C1([0,1]) such that γ1(0) = γ2(0) = 0, and define the tra-
jectory γ(t) = (γ1(t),γ2(t)). Additionally, let γδ
2 be
1 and γδ
elements in C1([0,1/δ ]) such that γδ
2 (t) =
γ2(δt) for each t ∈ [0,1/δ ], where δ > 0. Define γδ (t) =
(γδ
1 (t),γδ
2 (t)). Let X = [x,vx,y,vy]T and define F(t,X,ω,γδ )
as
1 (t) = γ1(δt) and γδ
[(x−cosωt−γδ
F(t,X,ω,γδ ) =
2 (t))2]α − kvx
2 (t))2]α − kvy
where A > 0, k > 0, and α > 1. Then consider the system
vx
A(x−cosωt−γδ
1 (t))2+(y−sinωt−γδ
vy
A(y−sinωt−γδ
2 (t))
1 (t))2+(y−sinωt−γδ
[(x−cosωt−γδ
1 (t))
(22)
X = F(t,X,ω,γδ ),
X(0) = X0.
(23)
There exist positive constants r1 > 0, ω1 > 0, and δ1 > 0,
which depend on A, k, and α, such that ω > ω1, δ > δ1
and (cid:107)X0(cid:107) ≤ r1 implies that the solution X(t) to Eq. (23) is
defined over the interval t ∈ [0,1/δ ] and that
2 (t))(cid:107) < 1
(24)
for all t ∈ [0,1/δ ]. Moreover, if X0 = 0, then for each ν > 0
there exist ω2 > 0 and δ2 > 0 such that for all ω > ω2 and
δ ∈ (0,δ2),
(cid:107)(x(t),y(t)− (γδ
1 (t),γδ
(cid:107)(x(t),y(t)− (γδ
1 (t),γδ
2 (t))(cid:107) < ν
(25)
for all t ∈ [0,1/δ ].
Proof. Consider
X − [γδ
1 (t),0,γδ
problem
the
E =
2 (t),0]T . Then we have the initial value
variables
change
of
E = F(t,E,ω,0) + δg(t),
(26)
where g(t) = [ γ1(δt),0, γ2(δt),0]T . Let ε = 1/ω and consider
the change of variable s = t/ε. Then the solution of Eq. (26)
satisfies
E(0) = X0,
dE
ds = εF(s,E(s),1,0) + εδg(εs).
(27)
The estimates in Eq. (24) and Eq. (25) follow from Propo-
sition III.2. Note that the dependence of g on the εs time
scale does not affect the proof, since g is still bounded from
above as required.
For the case where the follower agent starts at the origin,
this proposition implies that one can choose leader control
inputs that drive the follower to track the trajectory γ
arbitrarily closely, albeit over a longer period of time than
if the required bounds on the tracking error were weaker.
This is a stronger result than is needed to ensure general R-
confinement controllability, for which it is sufficient for the
follower agent to be within a distance R of the trajectory.
Theorem III.4. Suppose that [xl(0),yl(0)] = [1,0]T . Then
there exists r > 0 such that (cid:107)[xi,yi,vi
y](cid:107) < r for all i =
1,2....N implies that system (32) is 1-confinement control-
lable about all trajectories γ for which γ ∈ {(γ1,γ2) : γ1,γ2 ∈
C1([0,1]), γ1(0) = γ2(0) = 0}.
x,vi
Proof. Let ω and δ be positive scalars. Consider the follow-
ing control inputs for the leader:
ux(t) = −ω sinωt + δ γx(δt)
uy(t) = ω cosωt + δ γy(δt).
(28)
Integrating these functions from 0 to t for each t ∈ [0,1/δ ],
we can explicitly solve for the state variables xl(t) and yl(t):
xl(t) = cos(ωt) + γx(δt),
yl(t) = sin(ωt) + γx(δt).
(29)
Define F as in Eq. (22) and substitute Eq. (29) into system
(32). Then the state variables xi,vi
y in the solution of
system (32) satisfy
x,yi,vi
dXi
dt = F(t,Xi,ω,γδ ),
for each i∈{1,2, ...,N}, where Xi = [xi,vi
follows from Lemma III.3.
Xi(0) = Xi
0,
y]T . The result
x,yi,vi
(30)
The assumption that
too
restrictive. If [xl(0),yl(0)]T is any element on the unit circle,
then the leader control inputs are
[xl(0),yl(0)] = [1,0]T is not
ux(t) = −ω sin(ωt + φ ) + δ γx(δt)
uy(t) = ω cos(ωt + φ ) + δ γy(δt)
(31)
for an appropriately chosen φ. These inputs will have the
same effect on the system. Due to the continuous dependence
of the solutions of system (32) on its initial conditions
and parameters, it can be argued further that there exists
a neighborhood of the unit circle from which the leader can
always be driven to the circle, after which it can execute the
control inputs defined in Eq. (31).
Remark III.5. We can also consider a more general setting
in which the follower agents are running a consensus pro-
tocol while being confined by the leader. In this case, the
followers' interactions can be modeled using an undirected
simple graph with fixed topology, G = (V , E ), where V is
the set of follower agents and E identifies the pairs of agents
that interact with each other. The graph is not necessarily
connected. Then the x-coordinates of the follower agents'
positions and velocities evolve according to
xi = vi
x
vi
x =
[(xi − xl)2 + (yi − yl)2]α − kvi
x − v j
− ∑
(vi
j∈N (i)
A(xi − xl)
(xi − x j)− ∑
j∈N (i)
x
x) ,
(32)
with similar dynamics for the y-coordinates. As in Eq. (8),
the Jacobian map for the averaged vector field fav,c of the
composite group of followers can be computed as
(cid:20)B 0
(cid:21)
0 B
,
IN
Dfav,c(0) =
(cid:20)
where
B =
0
−βIN − L −kIN − γL
(cid:21)
,
(33)
(34)
λi± = − k
2
in which IN is the identity matrix of size N × N, L is
the Laplacian of the graph G , and β = A(α − 1), γ, and
k are positive scalars. The eigenvalues of Dfav,c(0) can be
computed according to [30][Chapter 4] as:
γµi ±(cid:112)(γµi − k)2 − 4(β − µi)
(35)
where µi is the ith eigenvalue of −L . Since the spectrum of
Dfav,c(0) lies in the open left half-plane of C whenever β =
A(α −1), k, and γ are positive, this implies that confinement
controllability of the leader-follower system is retained under
inter-follower interactions that can be modeled as a graph
with fixed topology.
+
2
,
IV. SIMULATIONS
We validate our confinement control approach with simu-
lations of two scenarios with different parameter values. In
Case 1, the target trajectory is γ(θ ) = [2cos2πθ 2sin2πθ ]T ,
the final time condition is T = 40π, and the other parameters
are A = 1, k = 0.05, α = 10, and ω = 10. In Case 2, we
set γ(θ ) = [10cos4πθ 2πθ ]T , T = 80π, A = 1, k = 0.03,
α = 5, and ω = 20. In both cases, the leader control inputs
are chosen according to Eq. (28), and there are three follower
agents. The initial positions and velocities of the followers
are drawn uniformly randomly from a ball of radius 0.125
(Case 1) or 0.5 (Case 2) that is centered at the origin.
Fig. 1 shows the agent trajectories during the two sim-
ulated scenarios. In both cases, the followers remain well
within a distance of 1 from the target
trajectory, which
confirms that the system is 1-confinement controllable about
this trajectory. The followers in Case 2 display larger initial
oscillations about the target trajectory than the followers in
Case 1. This is because the followers in Case 2 have a lower
damping coefficient k and initial velocities that were chosen
from a larger neighborhood of the origin, and thus tended to
be higher.
We also note that the higher value of ω in Case 2 implies
that the region of attraction of the original system is closer
to the region of the attraction of the averaged system than in
Case 1. This caused the confinement strategy in Case 2 to
be successful for a larger ball of possible initial conditions
than in Case 1. Indeed, we found that the followers in Case
1 did not remain confined when the initial conditions were
drawn from outside the ball of radius 0.125.
V. ROBOT EXPERIMENTS
A. Experimental Setup
We also tested our confinement control approach with
an experiment in which a small differential-drive robotic
platform Pheeno [40] emulated a follower agent that was
confined by a virtual leader. The experiment was performed
in the ASU Autonomous Collective Systems Laboratory. One
robot was placed in a 1.5 m × 2.1 m arena and marked
with a 2D binary identification tag to allow for real-time
position and orientation tracking. The tags were tracked
using one overhead Microsoft LifeCam Studio Webcam with
a resolution of 1920×1080 pixels at a frame rate of 30 FPS.
(a) Case 1
Fig. 2. The robot being herded by a virtual leader along a target trajectory
γE, shown in green. The blue dot is the virtual leader's position, the pink
dot is the center of the leader's orbit, and the dark red dot is the position
of a simulated follower agent. The light blue line on the robot indicates its
orientation.
(b) Case 2
Fig. 1. Evolution of leader and follower agent positions over time. The
followers start at random positions within a neighborhood of the origin.
The 2D binary tags were identified using standard OpenCV
algorithms on a Windows computer.
To accommodate the nonholonomic constraint of the robot
platform, the robot used a PI controller to keep its head-
ing facing away from the virtual leader at all times. This
heading control produces approximate holonomic motion
in a differential-drive platform. To facilitate this motion,
the leader's oscillation frequency was maintained below
the bandwidth of the robots' motors. A desktop computer
updated and transmitted the virtual leader's position and the
robot's global position to the robot via WiFi at a frequency
of 3 Hz.
We defined a circular target trajectory, shown in Fig. 2,
for the robot to track in the experiment. The trajectory was
given by γE (t) = [Rp cos(ωpt) Rp(sinωpt)]T , where Rp =
67 cm and ωp = 0.02. The robot was controlled using the
parameters A = R2α−1
, RL = 38 cm, k = 1, α = 4, and ω = 2.
L
Fig. 3. Time evolution of the positions of the leader agent (blue), simulated
follower agent (red), and robot (black) for an experiment with the target
trajectory γE (green).
B. Results
We compared the trajectory of the robot during the ex-
periment with the corresponding trajectory of a simulated
follower agent, indicated by a dark red dot in Fig. 2, that
moves under the influence of the same leader agent. Note
that while the robot is nonholonomic, the simulated follower
is holonomic. Fig. 3 plots the trajectories of the virtual
leader, the robot, and the simulated follower agent during
the experiment. The robot was initially placed inside the
region of confinement bounded by the leader's path, and the
simulated follower was initialized at the center of the leader's
orbit. The figure illustrates that the robot was successfully
kept inside the region of confinement for the duration of the
experiment.
050100150−3−2.5−2−1.5−1−0.500.51−2−1.5−1−0.500.511.52 tx yFollower1Follower2Follower3Leader0100200300−25−20−15−10−505−1012345678 tx yFollower1Follower2Follower3Leader50100150200250300−10001002003000200400600800100012001400 X (cm)Y (cm) Time (s)RobotTarget TrajectorySimulated FollowerLeader Fig. 3 shows that the trajectory of the robot deviates farther
from γE than the trajectory of the simulated follower. This
is likely due to the nonlinear effects of friction on the shafts
of the robots' motors. The motors require a minimum input
voltage to drive the robot's motion, which can be controlled
at speeds above 2 cm/s. These speeds are generated by the
repulsive interaction potential in the robot's controller when
the distance between the robot and the leader is below a
threshold value. Hence,
the robot moves away from the
leader only when it is within this distance; otherwise, it
is stationary. In contrast, the simulated follower does not
have such a constraint on its speed and will move when
the leader is farther away, causing its trajectory to adhere
more closely to γE. Thus, the experiment shows that when
our confinement strategy is implemented in practice, the
controller design must account for unmodeled dynamics in
the physical platform by allowing only those velocity control
inputs that exceed an appropriate lower bound.
VI. CONCLUSION
We presented a leader-follower control strategy for confin-
ing a group of agents to a region around a target trajectory.
By constructing oscillatory leader inputs, we used averaging
theory to show that follower agents can be confined to a
neighborhood of the origin. From a control design perspec-
tive, one could investigate the quantitative effects of different
parameters on the stability properties of the system. Well-
developed computational tools can be used to estimate the
size of the region of attraction of the origin [10], [19]. One
could also augment this open-loop strategy with feedback to
develop a more robust stabilization mechanism. Incorporat-
ing multiple leaders and inter-agent collision avoidance into
the control strategy is another direction for future work.
VII. ACKNOWLEDGMENT
K.E. thanks Matthias Kawski for useful suggestions re-
garding the proofs.
APPENDIX
Let U be a subset of Rn. The Jacobian map for a function
f : U → Rm at a state x ∈ U is given by:
∂ f1
(x)
∂ xn
...
∂ fm
(x)
∂ xn
(x)
∂ x1
...
∂ fm
(x)
∂ x1
···
...
···
Df(x) =
∂ f1
.
Let U be a subset of the unit sphere in R4, and define the
function f : U → R4 as:
(cid:82) 2π
(cid:82) 2π
f(x) =
1
2π
0
1
2π
0
x2
A(x1−cos(s))
[(x1−cos(s))2+(x3−sin(s))2]α ds− kx2
[(x1−cos(s))2+(x3−sin(s))2]α ds− kx4
A(x3−sin(s))
x4
,
where A,k,α ∈ R+. Define g1(x,t) = x−cos(t), g2(x,t) = x−
sin(t), and h(x,y,t) = g1(x,t)2 + g2(y,t)2. The function f is
=
continuously differentiable everywhere in U and has a well-
defined Jacobian at the origin. We compute the nontrivial
partial derivatives of the Jacobian map Df(x) using Leibniz's
rule for partial derivatives of integrals:
∂ f2
∂ x1
∂ f2
∂ x3
∂ f4
∂ x1
∂ f4
∂ x3
Evaluating these partial derivatives at the origin, we ob-
h(x1,x3,s)α − 2αg1(x1,s)2h(x1,x3,s)α−1
−2αg1(x1,s)g2(x3,s)h(x1,x3,s)α−1
−2αg1(x1,s)g2(x3,s)h(x1,x3,s)α−1
h(x1,x3,s)α − 2αg2(x3,s)2h(x1,x3,s)α−1
(cid:90) 2π
(cid:90) 2π
(cid:90) 2π
(cid:90) 2π
A
2π
A
2π
A
2π
A
2π
h(x1,x3,s)2α
h(x1,x3,s)2α
h(x1,x3,s)2α
h(x1,x3,s)2α
ds.
ds
ds
ds
=
=
=
0
0
0
0
tain:
∂ f2
∂ x1
∂ f2
∂ x3
∂ f4
∂ x1
∂ f4
∂ x3
(0) =
(0) =
(0) =
(0) =
A
2π
A
2π
A
2π
A
2π
0
(cid:90) 2π
(cid:90) 2π
(cid:90) 2π
(cid:90) 2π
0
0
0
(1− 2α sin2(s))ds = − A(α − 1)
−2α sin(s)cos(s)ds = 0
−2α sin(s)cos(s)ds = 0
(1− 2α cos2(s))ds = − A(α − 1).
REFERENCES
[1] John Baillieul. Stable average motions of mechanical systems subject
to periodic forcing. Dynamics and Control of Mechanical Systems:
The falling cat and related problems, 1:1–23, 1993.
[2] Richard E. Bellman, Joseph Bentsman, and Semyon M. Meerkov.
Vibrational control of nonlinear systems: Vibrational stabilizability.
IEEE Trans. on Automatic Control, 31(8):710–716, 1986.
[3] Brandon Bennett and Matthew Trafankowski. A comparative investi-
gation of herding algorithms. In Proc. of the Symp. on Understanding
and Modelling Collective Phenomena (UMoCoP), pages 33–38, 2012.
[4] Rosa Maria Bianchini and Gianna Stefani. Controllability along a
trajectory: A variational approach. SIAM Journal on Control and
Optimization, 31(4):900–927, 1993.
[5] Alberto Bressan and Dongmei Zhang. Control problems for a class of
set valued evolutions. Set-Valued and Variational Analysis, 20(4):581–
601, 2012.
[6] Roger W. Brockett. Control theory and singular Riemannian geometry.
Springer, 1982.
[7] Roger W. Brockett. On the control of a flock by a leader. Proc. of
the Steklov Institute of Mathematics, 268(1):49–57, 2010.
[8] Yongcan Cao, Wei Ren, and Magnus Egerstedt. Distributed contain-
ment control with multiple stationary or dynamic leaders in fixed and
switching directed networks. Automatica, 48(8):1586–1597, 2012.
[9] Marco Caponigro, Massimo Fornasier, Benedetto Piccoli, and Em-
manuel Trélat. Sparse stabilization and control of alignment models.
Mathematical Models and Methods in Applied Sciences, 25(3):521–
564, 2015.
[10] Graziano Chesi. Domain of attraction: analysis and control via SOS
programming. Springer Science & Business Media, 2011.
[11] Yacine Chitour, Frédéric Jean, and Ruixing Long. A global steering
method for nonholonomic systems. Journal of Differential Equations,
254(4):1903–1956, 2013.
[12] Rinaldo M. Colombo and Nikolay Pogodaev. Confinement strategies
in a model for the interaction between individuals and a continuum.
SIAM Journal on Applied Dynamical Systems, 11(2):741–770, 2012.
[13] Jean-Michel Coron. Global asymptotic stabilization for controllable
systems without drift. Mathematics of Control, Signals and Systems,
5(3):295–312, 1992.
[14] Ido Gilary, Nimrod Moiseyev, Saar Rahav, and Shmuel Fishman.
Trapping of particles by lasers:
the quantum Kapitza pendulum.
Journal of Physics A: Mathematical and General, 36(25):L409, 2003.
[15] J.-M. Godhavn, Andrea Balluchi, Lara S. Crawford, and S. Shankar
Sastry. Steering of a class of nonholonomic systems with drift terms.
Automatica, 35(5):837–847, 1999.
[16] John Guckenheimer and Philip Holmes. Nonlinear oscillations, dy-
namical systems, and bifurcations of vector fields, volume 42. Springer
Science & Business Media, 1983.
[17] Musad A. Haque, Amir R. Rahmani, and Magnus Egerstedt. Bio-
logically inspired confinement of multi-robot systems. International
Journal of Bio-Inspired Computation, 3(4):213–224, 2011.
[18] Meng Ji, Giancarlo Ferrari-Trecate, Magnus Egerstedt, and Annalisa
IEEE Trans. on
Buffa. Containment control in mobile networks.
Automatic Control, 53(8):1972–1975, 2008.
[19] Reza Kamyar and Matthew Peet.
Polynomial optimization with
applications to stability analysis and control-alternatives to sum of
squares. arXiv preprint arXiv:1408.5119, 2014.
[20] P. L. Kapitza. Dynamic stability of a pendulum with an oscillating
point of suspension. Journal of Experimental and Theoretical Physics,
21(5):588–597, 1951.
[21] Hassan Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002.
[22] Jyh-Ming Lien, Burchan Bayazit, Ross T. Sowell, Samuel Rodriguez,
and Nancy M. Amato. Shepherding behaviors. In Proc. IEEE Int'l.
Conf. on Robotics and Automation (ICRA), pages 4159–4164, 2004.
[23] Luc Moreau and Dirk Aeyels. Practical stability and stabilization.
IEEE Trans. on Automatic Control, 45(8):1554–1558, 2000.
[24] Richard M. Murray and S. Shankar Sastry. Nonholonomic motion
planning: Steering using sinusoids. IEEE Trans. on Automatic Control,
38(5):700–716, 1993.
[25] D. Neši´c and Andrew R. Teel.
Input-to-state stability for nonlinear
time-varying systems via averaging. Mathematics of Control, Signals
and Systems, 14(3):257–280, 2001.
[26] Wolfgang Paul. Electromagnetic traps for charged and neutral parti-
cles. Reviews of Modern Physics, 62(3):531, 1990.
[27] Benedetto Piccoli, Francesco Rossi, and Emmanuel Trélat. Control
arXiv preprint
to flocking of the kinetic Cucker-Smale model.
arXiv:1411.4687, 2014.
[28] Alyssa Pierson and Mac Schwager. Bio-inspired non-cooperative
multi-robot herding. In Proc. Int'l. Conf. on Robotics and Automation
(ICRA), 2015.
[29] Amirreza Rahmani, Meng Ji, Mehran Mesbahi, and Magnus Egerstedt.
Controllability of multi-agent systems from a graph-theoretic perspec-
tive. SIAM J. on Control and Optimization, 48(1):162–186, 2009.
[30] Wei Ren and Randal W. Beard. Distributed consensus in multi-vehicle
cooperative control. Springer, 1st edition, 2008.
[31] Wei Ren and Yongcan Cao. Distributed coordination of multi-agent
networks: emergent problems, models, and issues. Springer Science
& Business Media, 2010.
[32] Jan A. Sanders, Ferdinand Verhulst, and James A. Murdock. Averaging
methods in nonlinear dynamical systems. Springer-Verlag New York,
2nd edition, 2007.
[33] Andrey Sarychev. Stability criteria for time-periodic systems via high-
order averaging techniques. In Nonlinear Control in the Year 2000,
volume 2, pages 365–377. Springer, 2001.
[34] S. Shankar Sastry and Richard Montgomery. The structure of optimal
controls for a steering problem. In IFAC Symp. on Nonlinear Control
Systems Design (NOLCOS), pages 385–390, 1992.
[35] Eduardo D. Sontag. Mathematical control theory: deterministic finite
dimensional systems, volume 6. Springer Science & Business Media,
2013.
[36] Héctor J. Sussmann. A general theorem on local controllability. SIAM
Journal on Control and Optimization, 25(1):158–194, 1987.
[37] Héctor J. Sussmann and Wensheng Liu. Limits of highly oscillatory
controls and the approximation of general paths by admissible trajec-
tories. In Proc. IEEE Conf. on Decision and Control (CDC), pages
437–442, 1991.
[38] Herbert G. Tanner. On the controllability of nearest neighbor inter-
connections. In Proc. IEEE Conf. on Decision and Control (CDC),
pages 2467–2472, 2004.
[39] Andrew R. Teel, Richard M. Murray, and Gregory C. Walsh. Non-
holonomic control systems: from steering to stabilization with sinu-
soids. International Journal of Control, 62(4):849–870, 1995.
[40] Sean Wilson, Ruben Gameros, Michael Sheely, Matthew Lin, Kathryn
Dover, Robert Gevorkyan, Matt Haberland, Andrea Bertozzi, and
Spring Berman.
Pheeno, a versatile swarm robotic research and
education platform. Accepted to IEEE Robotics and Automation
Letters (RA-L), 2016.
|
1312.2642 | 1 | 1312 | 2013-12-10T02:20:24 | Cellular Automata based Feedback Mechanism in Strengthening biological Sequence Analysis Approach to Robotic Soccer | [
"cs.MA",
"cs.RO"
] | This paper reports on the application of sequence analysis algorithms for agents in robotic soccer and a suitable representation is proposed to achieve this mapping. The objective of this research is to generate novel better in-game strategies with the aim of faster adaptation to the changing environment. A homogeneous non-communicating multi-agent architecture using the representation is presented. To achieve real-time learning during a game, a bucket brigade algorithm is used to reinforce Cellular Automata Based Classifier. A technique for selecting strategies based on sequence analysis is adopted. | cs.MA | cs | Cellular Automata based Feedback Mechanism in Strengthening
biological Sequence Analysis Approach to Robotic Soccer
P. Kiran Sree 1 , G.V.S. Raju2, .S. Viswandha Raju3 and N.S.S.S.N Usha Devi4
1. Associate Professor, Department of Computer Science, S.R.K Institute of Technology, Enikepadu,
Vijayawada,India, [email protected], Mobile: +919959818274.
2. Prof and Head of the Department, Computer Science, Swarnandhra Engineering college,Narasapur.
3. Professor, Department of CSE, Gokaraju Rangaraju Institute of Engineering & Technology
4. Graduate Student of C.S.E,J.N.T.University.
ABSTRACT
This paper reports on the application of sequence
analysis algorithms for agents in robotic soccer and
a suitable representation is proposed to achieve this
mapping. The objective of this research is to
generate novel better in-game strategies with the
aim of
faster adaptation
to
the changing
environment. A homogeneous non-communicating
multi-agent architecture using the representation is
presented. To achieve real-time learning during a
game, a bucket brigade algorithm is used to
reinforce Cellular Automata Based Classifier. A
technique
for
selecting
strategies based on
sequence analysis is adopted.
Keywords: Multi-agent architecture, bucket brigade
algorithm, reinforce learning, Cellular Automata
Classifer.
1. Introduction
Although each domain presents a variety of
approaches, from a research perspective the ideal
domain embodies as many issues as possible. Robotic
soccer is a particularly good domain for studying
multi-agent systems. Originated by Alan Mackworth
[1], it has been gaining popularity in recent years
with several international competitions taking place
[2]. Even though robotic soccer is a game, most real-
world complexities are maintained. Some of the
distinguishing characteristics of the domain include:
real-time, noisy with hidden state, collaborative and
adversarial goals. It is the ideal test bed for
evaluating different machine learning techniques in a
direct manner, as it provides for multiple levels of
evaluation such as evaluation of low-level behavior
e.g. maneuvering on the field as well as evaluation of
strategic responses to changing scenarios.
Several Multi-agent system scenarios are possible:
• Homogeneous, on-communicating
• Homogeneous, communicating
• Heterogeneous, on-communicating
• Heterogeneous, communicating
Each of which have their own set of challenging
research
issues.
In
homogeneous,
non-
communicating multi-agent systems, all of the agents
have the same internal structure including goals,
domain knowledge, and possible actions. They also
have the same procedure for selecting among their
actions. The only differences among agents are their
sensory inputs and the actual actions they take: they
are
situated
differently
in
the world.
A DNA fragment is usually written as a sequence of
letters A, C, T and G – representing the four
nucleotides Adenine, Cytosine, Thymine and
Guanine. DNA is the master behind all the activities
in the cell and is responsible for the synthesis of
proteins. In this paper, Section-2 gives an overview
of related work. A representation scheme for robotic
soccer is presented in Section-3. An architecture
depicting a homogeneous non-communicating agent
utilizing the above representation is proposed in
Section-4. Section-5 discusses experimental results
and Section-6 deals with future directions and
concludes the paper.
2. Related Work
Balch and Arkin [6] use homogeneous, reactive,
non-communicating agents
to study
formation
maintenance in autonomous robots. The robots' goal
is to move together in a military formation such as a
diamond, column, or wedge. They periodically come
across obstacles which prevent one or more of the
robots from moving in a straight line. After passing
the obstacle, all robots must adjust in order to regain
their formation. The actual robot motion is a simple
ICGST-AIML Journal, Volume 8, Issue I, June 200829
PATTERN
PLAYER
Action
Base
Kick towards goal
Pass to team-mate
Move towards ball
Table 2: Distinct patterns and their occurrence in each player
sequence
the provided input sequence against these patterns for
matches. Table 2 gives the different patterns and
occurrences of players.
vectors.
these
of
sum
weighted
Levy and Rosenschein [7] create agents that each act
in service of its own goals. They use game theoretic
techniques to find equilibrium points and thus to
decide how
to act. These agents are clearly
deliberative, as they search for actions rather than
simply retrieving them. There are also several
existing systems and techniques that mix reactive and
deliberative behaviors. One example is the OASIS
system which reasons about when to be reactive and
when to follow goal-directed plans [8]. Another
example is reactive deliberation [9]. As the name
implies, it mixes reactive and deliberative behavior:
an agent reasons about which reactive behavior to
follow under the constraint that it must choose
actions at a rate of 60 Hz. Reactive deliberations was
developed on the first robotic soccer platform.
3. Sequential representation of robotic soccer
The game of robotic soccer involves a vast amount
of data with respect
to both player and ball
movement. The movement of the ball during the
course of a match is tracked to obtain a game
sequence. Also, each individual's contribution to the
game can be assessed by focusing on player
movements during the game, yielding a player
sequence. This is instrumental in characterizing a
player's contribution to the success or failure of the
game. For example, we consider a game played over
a period of 60 minutes and represent this game as a
sequence of length 60, where each letter represents
the player who is in possession of the ball at that
minute. Game 1: bcgad-bccc-g-aab—bbaggd.
and the Player sequence for player 'a' would be
represented as –
The agent has just four actuators for physically
manipulating the world: turn, dash, kick and catch.
Player a: ACTTGC-AG--TTCCCA-... .
The server only executes one of these commands for
each player at the end of each simulator cycle. If an
where '-' denotes an idle action or that the ball is in
agent sends more than one such command during the
motion. A history of games played can be collected
same
cycle,
only
one
is
executed
non-
and analyzed using the sequence analysis algorithms
deterministically.
runs
simulator
the
Since
employed frequently in bio-informatics.
asynchronously from the agents, there is no way to
keep perfect time with the server's cycle. Therefore
3.1 Algorithm for Finding Tandem Repeats
any given command could be missed by the server. It
The concept of finding tandem repeats has been
is up to the agent to determine whether a given
employed here to trace subsequences or regions that
command has been executed by observing the future
frequently repeat in the game and whether such
state of the world.
repetition is ideal for the game.
4. Proposed Work
It involves two subtasks
We propose soccer-playing agent architecture with a
• Finding all possible unique subsequences of a
sequential
analysis
component, which was
specified length.
strengthened by using CA classifier for providing
• Finding the location and number of tandem
feed back during the match.
We utilize the
repeats.
representation described in the previous sections.
The input to the first subtask is the starting and
ending lengths of the pattern and the result is the
creation of a list of unique patterns that range from
the specified lower to the upper range. This list
becomes the input to the second stage which checks
NO. OF
OCCURENCES
4
3
3
5
5
2
2
3
CACC
ACCC
ACCC
ACCC
CCCC
CCCCB
CCCAGCC
CC-G
player-a
player-b
player-d
player-c
player-a
player-b
player-a
player-d
Table 1: Task Table
Turn towards ball
A
C
G
T
ICGST-AIML Journal, Volume 8, Issue I, June 200830 Figure. 1 : Sequence Driven Classifier System Architecture
Figure 2: The soccer server display. Each player is represented as a two-halved circle. The light side is the side towards which the
player is facing. All players are facing the ball, which is in the middle of the field. The black bars on the left and right sides of the
field are the goals
Shooting Behavior
We make a simple strategy for shooting the ball into
the goal. To shoot the ball to the goal, it is important
that the robot can see both ball and goal. Therefore,
the robot must round the ball until the robot can see
both ball and goal with the camera toward the ball.
Finally, the robot kicks the ball strongly. The
concrete procedure of shooting behavior is follows:
1)Find the ball
2)Approach the ball
While approaching the ball
if the area of the ball > 20 then stop
3)Round the ball
d (cid:197)the direction of the goal
switch(d) right: clockwise round the ball(AGGGT)
with the camera toward the ball(ACCCT)
left: counterclockwise round the ball(AAACT)
with the camera toward the ball(TTTAC)
if the robot can see both ball and goal then stop
4)Turn
towards
robot
the
the body of
ball(ATACT)
5) Use the CA Feed Back
6)Kick the ball strongly(AATAA)
the
ICGST-AIML Journal, Volume 8, Issue I, June 200831 following dependency matrix While moving from one
state to other, the dependency matrix indicates on
which neighboring cells the state should depend. So
cell 254 depends on its state, left neighbor, and right
neighbor figure (3). Now we represented the transition
function in the form of matrix. In the case of
complement FMACA we use another vector for
representation of chromosome.
Table 4. CA Fuzzy Rules
Fig3: Matrix Representation
4.1.4 FMACA Based Tree-Structured Classifier
Cellular Automata choose the rules carefully after
trying many other possibilities, some of which caused
the cells to die too fast and others which caused too
many cells to be born. Life balances these tendencies,
making it hard to tell whether a pattern will die out
completely, form a stable population, or grow
forever. Life is just one example of a cellular
automaton, which is any system in which rules are
applied to cells and their neighbors in a regular grid.
A good way to get started in Life is to try out
different patterns and see what happens. Even
completely random starting patterns rapidly turn into
Life objects recognizable to anyone with a little
experience. In this section, we follow a simple-
looking pattern called the R-pantomime. It starts out
with just five cells, but gets complicated very fast.
We can see many of the early discoveries in Life just
by running this one pattern in the applet Like
decision
tree classifiers, FMACA based
tree
structured classifier uses
the distinct k-means
algorithm recursively partitions the training set to get
nodes (attractors of a FMACA) belonging to a single
class. Each node (attractor basin) of the tree is either
a leaf indicating a class; or a decision (intermediate)
node which specifies a test on a single FMACA,
according to equations 1,2. Suppose, we want to
design a FMACA based pattern classifier to classify a
Table 3: The Soccer server agent's commands
4.1. Cellular Automata (CA) and Fuzzy Cellular
Automata (FCA)
A CA [4], [5], [6], consists of a number of cells
organized in the form of a lattice. It evolves in discrete
space and time. The next state of a cell depends on its
own state and the states of its neighboring cells. In a 3-
neighborhood dependency, the next state qi (t + 1) of a
cell is assumed to be dependent only on itself and on
its two neighbors (left and right), and is denoted as
qi(t + 1) = f (qi−1(t), q i(t), qi+1(t)) (1)
where qi (t) represents the state of the ith cell at tth
instant of time, f is the next state function and referred
to as the rule of the automata. The decimal equivalent
of the next state function, as introduced by Wolfram,
is the rule number of the CA cell. In a 2-state 3-
neighborhood CA, there are total 256 distinct next
state functions.
4.1.1 FCA Fundamentals
FCA [2], [6] is a linear array of cells which evolves in
time. Each cell of the array assumes a state qi, a rational
value in the interval [0, 1] (fuzzy states) and changes its
state according to a local evolution function on its own
state and the states of its two neighbors. The degree to
which a cell is in fuzzy states 1 and 0 can be calculated
with the membership functions. This gives more
accuracy in finding the coding regions. In a FCA, the
conventional Boolean functions are AND , OR, NOT.
4.1.2 Dependency Matrix for FCA
Rule defined in equation 1 should be represented as a
local transition function of FCA cell. That rules are
converted into matrix form for easier representation of
chromosomes [16].
Example 1: A 4-cell null boundary hybrid FCA with
the following rule
< 238, 254, 238, 252 > (that is, < (qi+qi+1),
(qi−1+qi+qi+1), (qi + qi+1), (qi−1 + qi) >) applied
from left to right, may be characterized by the
ICGST-AIML Journal, Volume 8, Issue I, June 200832
training set S = {S1, S2, · , SK} into K classes. First, a
FMACA with k-attractor basins
is Once we
formulated the transition function, we can move form
one state to other. For the example 1 if initial state is
P (0) = (0.80, 0.20, 0.20, 0.00) then the next states
will be
P (1) = (1.00 1.00, 0.20, 0.20),
P (2) = (1.00 1.00, 0.40, 0.40),
P (3) = (1.00 1.00, 0.80, 0.80),
P (4) = (1.00 1.00, 1.00, 1.00).
We just introduced the concept of Cellular Automata
in this section. CA is a linear array of cells which
evolves in time. Each cell of the array assumes a state
qi, a rational value in the interval [0, 1] (fuzzy states)
and changes its state according to a local evolution
function on its own state and the states of its two
neighbors. The degree to which a cell is in fuzzy
states 1 and 0 can be calculated with the membership
functions. This gives more accuracy in finding the
coding regions. In a FCA, the conventional Boolean
functions are AND , OR, NOT. Figure 1,2 shows the
architecture of
sequence-driven
the proposed
classifier system. This system differs from standard
classifier systems [5] in two main ways generated.
The training set S is then distributed into k attractor
basins (nodes). Let, S’ be the set of elements in an
attractor basin. If S’ belongs to only one class, then
label that attractor basin for that class. Otherwise, this
process is repeated recursively for each attractor
basin (node) until all the examples in each attractor
basin belong to one class. Tree construction is
reported in [7]. The above discussions have been
formalized in the following algorithm. We are using
genetic algorithm classify the training set.
Algorithm 1: FMACA Tree Building (using
distinct K means algorithms)
Input : Training set S = {S1, S2, · ·, SK}
Output: FMACA Tree.
Partition(S, K)
Step 1: Generate a FMACA with k number of
attractor basins.
Step 2: Distribute S into k attractor basins (nodes).
Step 3: Evaluate the distribution of examples in each
attractor basin (node).
Step 4: If all the examples (S’) of an attractor basin
(node) belong to only one class, then label the
attractor basin (leaf node) for that class.
Step 5: If examples (S’) of an attractor basin belong
to K’ number of classes, then Partition (S’, K’).
Step 6: Stop.
First, the proposed system adds an event-analysis
section and prepares a record of past game
sequences/player sequences that are analyzed using
sequence analysis. Second, the classifier discovery
section using genetic algorithms targets only actions
while the conditions are generated using information
provided by sequence analysis. The systems also
provides for different types of reward – a large
reward for winning matches and smaller rewards that
can be obtained from succeeding in a single play,
such as passing the ball
5. Experimental Results
The Sequence Analysis algorithms were tested on
player and game sequences of varying lengths and
distinct sub sequences yielded the following results
from which various inferences for improving the next
match were made.
Client 1 sends movement commands after every
perception it receives. This strategy works out fine in
cycle t-1; but in cycle t it misses the opportunity to
act because it receives no perceptions; and in cycle
t+1 it sends two movement commands, only one of
which will be executed. Client 2, on the other hand,
successfully sends one movement command every
cycle. Note that in cycle t it must act with no new
perceptual information, while in cycle t+1, it receives
two perceptions prior to acting and one afterwards.
Ideally, it would act after receiving and taking into
account all three perceptions. However, it does not
know precisely when the simulator cycle will change
internally in the soccer server and it cannot know
ahead of time when it will receive perceptions. Thus,
in exchange for the ability to act every simulator
cycle, it sometimes acts with less than the maximal
amount of information about the world. However, as
each simulator cycle represents only a short amount
of real time (simulator_step msec), the world does
not change very much from cycle to cycle, and the
client can act accurately even if it takes some of its
perceptions into account only before its subsequent
action.
Figure 4: A sample period of the server-client
interface over the course of 3 simulator cycles at
times t-1, t, and t+1. The thick central horizontal line
represents the real time as kept by the server. The top
and bottom horizontal lines represent the time-lines
of two separate clients. The dashed arrows from the
server towards a client represent perceptions for that
client. The solid arrows from a client towards the
server represent movement commands sent by that
client. These arrows end at the point in time at which
the server executes the movement commands, namely
the end of the simulator cycle during which they are
sent. Asynchronous sensing and acting, especially
when the sensing can happen at unpredictable
intervals, is a very challenging paradigm for agents to
handle. Agents must balance the need to act regularly
ICGST-AIML Journal, Volume 8, Issue I, June 200833 and as quickly as possible with the need to gather
information about the environment. Along with
asynchronous sensing and action, the soccer server
captures several other real-world complexities, as will
become evident throughout the remainder of this
section.
5.1 Entropy
Entropy is the measure of randomness of a system.
The maximum entropy (close to 1) of a system
signifies chaotic behavior, whereas low entropy
indicates ordered behavior. In case of complex
system, mean entropy is close to the critical value
0·84 with high variance. To measure the entropy, we
select a moving window of 10 time steps (w=10). The
CA rule vector is operated for 10000 time steps from
a random initial state. The mean entropy and the
standard deviation from
the mean have been
computed. For each CA the procedure is repeated for
15 times with different random initial states. The high
entropy with high standard deviation indicates that
the evolved CA performing pattern recognition task
are at the edge of chaos.
5.2 Mutual Information
Mutual information measures the correlation between
patterns generated at a fixed time interval. If a pattern
P1 is the copy of P2, then mutual information
between P1 and P2 is 1. The mutual information
between two statistically independent patterns is 0.
Both the ordered and chaotic CA rules do not create
highly correlated patterns and in effect generate
pattern set with low mutual information. On the other
hand, the complex CA rules create highly correlated
structures producing maximum mutual information.
To measure the mutual information of the patterns
generated by CA, we follow the method proposed.
These results were translated into strength values and
directly modified within the LCS. In order to justify
that this proposed architecture is indeed faster at
learning targeted behavior at a much faster pace,
conditions were
“Clean Slate”
experimental
maintained. The Classifier rule set
initially
is
completely randomly generated – to simulate a blank
knowledge base. The rule set is then generated prior
to the first iteration and the LCS was run for up to
200,000 iterations. Initialization parameters for the
Genetic Algorithm specify the algorithm to run every
4000 iterations – this interval is necessary so that the
strengths of the rule sets are given time to stabilize
after credit apportionment. The entropy and mutual
information of the CA in successive generations of
GA are reported in Figure 5,6 ,7,8
For four different CA size (n= 10, 15, 20, 30). For
each of the cases, the values of entropy and mutual
information reach their steady state once the AIS
FMACA for a given pattern set gets evolved. For
understanding the motion, the initial population (IP)
is randomly generated. All these figures points to the
fact that as the CA evolve towards the desired goal of
maximum pattern recognizing capability, the entropy
values fluctuate in the intermediate generations, but
saturate to a particular value (close to the critical
value 0·84 [245]) when fit rule
is obtained.
Simultaneously, the values of mutual information
fluctuate at the intermediate points prior to reaching
maximum value that remains stable in subsequent
generations. All these figures indicate that the CA
move from chaotic region to the edge of chaos to
perform complex computation associated with pattern
recognition. Table 3 shows the application of the
algorithm for a single game which when applied to
hundreds of games, useful inferences could be made.
The patterns thus identified through the application of
tandem repeats algorithm served to identify the risk
of threats and goals thereby enabling dynamic
decision making in future games with reference to
past history of games played. The goal sequences
identified were xxCCT, where x represents any of the
4 characters. The following is a table showing the
sequences that are likely to occur in the case of a goal
or threat and their percentage of occurrences. The
goal sequences were identified by application of the
algorithm
repeats. These
tandem
finding
for
sequences were in turn checked using the algorithm
for Trans-membrane region to find the percentage of
occurrences of these patterns in multiple numbers of
games.
Table 3: Percentage of occurrence of Goal and Threat sequences
SEQUEN
CES
95%
75%
50%
< 50%
TCCCT CACCT CxCCT CCAT
Goal
(xxCCT)
Threat CTCCC CCACC CCxCC GCAC
The Learning classifier system was first tested by
first evolving completely random classifier rules and
then placed in the environment to gage the learning
rate. Figure 2 shows the proportion of correct actions
proposed by the learning classifier system versus the
time epochs taken. Subsequent experiments were
conducted by constructing a history of 10,000 games
– and the performance of the classifier system was
observed. Figure 7. shows the performance of LCS,
an accuracy rate of nearly 75% was obtained after
merely 100,000 iterations. This is observed to be far
less when compared
to other agent
learning
mechanisms based on reinforcement learning. Given
a good balance between desired accuracy and training
time, this technique will yield a good “knowledge
base” of conditions and actions upon which the agent
will base all its responses.
ICGST-AIML Journal, Volume 8, Issue I, June 200834 6 .Conclusion
The advantage of adopting sequence analysis for
evaluating strategy is that the analysis is capable of
distinguishing individual strategy of an agent as well
the overall strategic play of the team as such. This
approach could easily be adapted towards developing
squad-based tactics of team behavior evolution, by
evolving specific strategies for specific groups - for
example, creating a group of defenders whose overall
goal is to defend their goal, and a squad of offensive
players whose only aim is to target the opposing
team's goal. Furthermore, use of advanced genetic
operations during the discovery stages could enable
the classifier system to discover better actions within
a smaller time-bound and improve the real-time
response of the system. A representation scheme for
robotic soccer in the form of biological sequence has
been presented. Two new algorithms to find repeating
patterns and
ideal game strategies similar
to
biological problems have been developed and tried on
the sequences generated. A number of sequences has
been generated from various games played and the
algorithms proved effective in analysing various
repeating patterns before a goal and finding faulty
moves. The entire system was improved by using the
feed back generated by CA Classifier.
Figure 5: Entropy & Mutation information for n=10
Figure 6: Entropy & Mutation information for n=15
Figure 7: Entropy & Mutation information for n=20
Figure 8: Entropy & Mutation information for n=30
ICGST-AIML Journal, Volume 8, Issue I, June 200835 Figure 9. Graph (Proportion Correct VS Iterations)
7. References
[1] Michael Sahota, Alan K. Mackworth, Rod A.
Barman, and Stewart J. Kingdon , “Real-time
control of soccer-playing robots using off-board
vision: the dynamite testbed”, IEEE
International Conference on Systems, Man, and
Cybernetics, pages 1995 ,3690-3663.
[2] Hiroaki Kitano, Yasuo Kuniyoshi, Itsuki Noda,
Minoru Asada, Hitoshi Matsubara, and Ei-Ichi
Osawa, “ Robocup: A challenge problem for ai ”,
AI Magazine, Volume 18(Issue 1), 1997,pp. 73-
85.
[3] Needleman, S.B., Wunsch C.D., “A general
method applicable to the search for similarities in
the amino acid sequences of two proteins”,
Journal of Molecular Biology Volume 48, 2000,
pp. 443-453.
[4] Smith T.F, Waterman M.S., “Identification of
common molecular subsequences” Journal of
Molecular Biology, Volume 14, 1998,pp.95-197.
[5] Goldberg, D.E. , “ Genetic algorithms in Search
,Optimization
and Machine Learning
”,
Addison-Wesley, 1989,Reading.
[6] Tucker Balch and Ronald C. Arkin. , “ Motor
schema-based formation control for multi agent
Proceedings of
the First
robot
teams ”,
International Conference
on Multi-Agent
Systems ICMAS-95, 1995,pp. 10-16.
[7] Ran Levy and Jeffrey S. Rosenschein , ”A game
theoretic approach to the pursuit problem ”,
Working Papers of
the 11th International
Workshop on Distributed Artificial Intelligence,
1992,pp.195-21.
[8] Anand S. Rao and Michael P. George , ” BDI
agents: From theory to practice.”, Proceedings of
the First International Conference on Multi-
Agent Systems ICMAS-95, 1995, pp. 31
[9] Michael K. Sahota. , “Reactive deliberation:
An architecture for real-time intelligent control
in dynamic environments”, Proceedings of the
Twelfth National Conference on Artificial
Intelligence, pages 1303-1308,1994, pp. 184,
185, 202.
[10] P.Kiran Sree, I .Ramesh Babu ,”Identification of
Protein Coding Regions in Genomic DNA Using
Unsupervised FMACA Based Pattern Classifier”
in International Journal of Computer Science &
Network Security with
ISSN: 1738-7906
Volume Number: Vol.8, No.1,2008.
[11] P.Kiran Sree, R.Ramachandran, “Identification
of Protein Coding Regions in Genomic DNA
Cellular
Fuzzy
Supervised
Using
of
International
Automata”.in
journal
in Computer Science
Advances
and
Engineering, with
ISSN: 0973-6999, Vol:
1,2008.
[12] Eric E. Snyder ,Gary D. Stormo, “ Identification
of Protein Coding Regions
In Genomic
DNA”.ICCS Transactions 2002.
[13] E E Snyder and G D Stormo,”Identification of
coding regions in genomic DNA sequences: an
application of dynamic programming and neural
networks “ Nucleic Acids Res. 1993 February
11; 21(3): 607–613.
[14] P. Flocchini, F. Geurts, A. Mingarelli, and N.
Santoro (2000),“Convergence and Aperiodicity
in Fuzzy Cellular Automata: Revisiting Rule
90,”Physica D.
[15] P. Maji and P. P. Chaudhuri (2004),“FMACA:
A Fuzzy Cellular Automata Based Pattern
Classifier,” Proceedings of 9th International
Conference on Database Systems , Korea, pp.
494–505, 2004.
[16] C.G. Langton (2000), “Self-reproduction in
cellular automata,” Physica D, vol.10, pp.135–
144.
[17] T. Toffoli(1998), “Reversible computing,” in
Automata, Languages and Programming,
ed. J.W. De Bakker and J. Van Leeuwen,
pp.632–644.
[18] G. Vichniac(1994), “Simulating physics with
cellular automata,” Physica D,vol.10, pp.96–
115.
ICGST-AIML Journal, Volume 8, Issue I, June 200836 [19] S. Chattopadhyay, S. Adhikari, S. Sengupta, and
M. Pal (2000), “Highly regular, modular, and
cascadable design of cellular automata-based
pattern classifier,” IEEE Trans. Very Large
Scale Integr. Syst., vol.8, no.6,
[20] J. Fickett (1982), “Recognition of protein coding
regions in dna sequences,”Nucleic Acids Res.,
vol. 10, pp. 5303–5318.
[21] B. E. Blaisdell(1983), “A prevalent persistent
global non randomness that distinguishes
coding and non-coding eukaryotic nuclear dna
sequence,” J. Molec. Evol., vol. 19, pp. 122–
133.
[22] R. Farber, A. Lapedes, and K. Sirotkin(1992),
“Determination of eukaryotic protein
coding regions using neural networks and
information theory,” J. Mol. Biol., vol. 226, pp.
471–479
[23] E. Uberbacher and R. Mural(1991), “Locating
protein-coding regions in human dna sequences
by a multiple sensor-neural network approach,”
Proc. Natl. Acad. Sci., USA, vol. 88, pp. 11261–
11265.
[24] Aickelin U and Cayzer S (2002): The Danger
Theory and Its Application to AIS, Proceedings
1st International Conference on AIS, pp 141-148,
Canterbury, UK.
[25] D. Dasgupta, 1999. Artificial Immune Systems
and Their Application. Berlin, Germany:
Springer-verlag.
[26] Jerne, N. K. (1974), “Towards a Network Theory
of the Immune System”, Ann. Immunol. (Inst.
Pasteur) 125C, pp. 373-389.
Biography
P.KIRAN SREE received his
B.Tech in Computer Science &
Engineering, from J.N.T.U and
in Computer Science &
M.E
Engineering
from
Anna
University. He is pursuing Ph.D
in Computer Science
from
J.N.T.U, Hyderabad. He has
published many
technical papers; both
in
international and national Journals& Conferences
.His areas of interests include Cellular Automata,
Parallel Algorithms, Artificial
Intelligence,
Compiler Design and Computer Networks. He also
wrote books on Analysis of Algorithms, Theory of
Computation and Artificial Intelligence. He was the
reviewer for many International Journals and IEEE
Society Conferences in Artificial Intelligence and
Networks. He was also member
in many
International Technical Committees. He was the
Associate Editor for Asian Journal of Scientific
Research (ISSN: 1992-1454), Journal of Artificial
Intelligence
(ISSN:
1994-5450),
Information
Technology Journal (ISSN: 1812-5646), Journal of
Software Engineering (ISSN: 1819-4311), and
Research Journal of Information Technology. He
was also invited speaker and organizing chairman
for special sessions
in WSEAS
international
conferences. He is the member of C.S.I, I.E.T.E,
I.S.T.E (India), ICST (Europe) and IAENG (U.S.A).
He is now associated with S.R.K Institute of
Technology, Vijayawada
G.V.S.Raju
is working as a
Professor in the Department of
C.S.E, Swarnandhra Engineering
College, Narasapuram, West
Godavari , Dt. He has published
good number of papers
in
international conferences.
journals
Dr S. Viswanadha Raju is
working as a Professor
in
Department of CSE, Gokaraju
Rangaraju Engineering College,
Hyderabad. He has published
good number of papers both in
international conferences and
N.S.S.S.N Usha Devi was a
graduate student of C.S.E from
J.N.T.U. She has published 5
research papers in international
conferences
in
two
and
international journals
ICGST-AIML Journal, Volume 8, Issue I, June 200837 |
1507.03826 | 1 | 1507 | 2015-07-14T12:41:14 | Modelling Structured Societies: a Multi-relational Approach to Context Permeability | [
"cs.MA",
"cs.SI",
"physics.soc-ph"
] | The structure of social relations is fundamental for the construction of plausible simulation scenarios. It shapes the way actors interact and create their identity within overlapping social contexts. Each actor interacts in multiple contexts within different types of social relations that constitute their social space. In this article, we present an approach to model structured agent societies with multiple coexisting social networks. We study the notion of context permeability, using a game in which agents try to achieve global consensus. We design and analyse two different models of permeability. In the first model, agents interact concurrently in multiple social networks. In the second, we introduce a context switching mechanism which adds a dynamic temporal component to agent interaction in the model. Agents switch between the different networks spending more or less time in each one. We compare these models and analyse the influence of different social networks regarding the speed of convergence to consensus. We conduct a series of experiments that show the impact of different configurations for coexisting social networks. This approach unveils both the limitations of the current modelling approaches and possible research directions for complex social space simulations. | cs.MA | cs |
Modelling Structured Societies: a Multi-relational
Approach to Context Permeability
Davide Nunesa,∗, Luis Antunesa
aGUESS/BioISI -- Instituto de Biosistemas e Ciencias Integrativas, Faculdade de Ciencias,
Universidade de Lisboa, 1749-016 Lisboa, Portugal.
Abstract
The structure of social relations is fundamental for the construction of plausible
simulation scenarios. It shapes the way actors interact and create their identity
within overlapping social contexts. Each actor interacts in multiple contexts
within different types of social relations that constitute their social space. In this
article, we present an approach to model structured agent societies with multiple
coexisting social networks. We study the notion of context permeability, using a
game in which agents try to achieve global consensus. We design and analyse two
different models of permeability. In the first model, agents interact concurrently
in multiple social networks.
In the second, we introduce a context switching
mechanism which adds a dynamic temporal component to agent interaction in
the model. Agents switch between the different networks spending more or
less time in each one. We compare these models and analyse the influence of
different social networks regarding the speed of convergence to consensus. We
conduct a series of experiments that show the impact of different configurations
for coexisting social networks. This approach unveils both the limitations of
the current modelling approaches and possible research directions for complex
social space simulations.
Keywords: Social Simulation and Modelling, Agent Societies, Consensus,
Context, Social Networks
'
∗Corresponding author
Email addresses: [email protected] (Davide Nunes),
[email protected] (Luis Antunes)
Preprint submitted to Elsevier
September 5, 2018
1. Introduction
In most simulations of complex social phenomena, agents are considered to in-
habit a space in which structure is very simple. This space has little resemblance
with the social world it is supposed to depict and for which conclusions are sup-
posed to be extrapolated. This simplicity does not come by chance, rather, it
is necessary and desired by the researchers: the problems to be approached are
themselves so complex that whichever factors of complexity can be reduced (or
at least postponed), the reduction is always welcome. So, it is common practice
that geographical space is reduced to a two-dimensional grid, and all social re-
lations between agents are condensed into one more or less structured abstract
relation. Most social simulation and modelling approaches disregard the fact
that we engage in multiple social relations. Moreover, each kind of social rela-
tion can possess distinctive characteristics that include: rich information such
as degree of connectivity, centrality, trust, interactions frequency, asymmetry,
and so on.
To explore the addition of multiple relations and its consequences for the
dissemination of phenomena in social simulations, we put forward to reduce
the emphasis given on agent individual interactions. We accomplished this by
choosing a simple -- and especially neutral -- game to model those interactions.
Our main concern was that the game itself would not provide biases or trends
in the collective phenomenon being studied, so we chose the consensus forma-
tion game, with a straightforward majority rule to decide the outcome of each
individual move.
While most models are simplified descriptions -- reductions -- of real-world
phenomena, many constitute complex systems themselves and thus, we need
techniques such as simulation to explore their properties. We can make a
distinction between two types of complexity: the complexity of real systems
(ontological complexity) and the complexity of models (descriptive complex-
ity) [1]. The levels of abstraction in a simulation model can then range from
data-driven paradigms to more abstract descriptions that allow us to create
what-if-scenarios. The study of this abstract descriptive complexity in simu-
lation models is as valuable as its data-driven counterpart. We may not be
able to make predictions based on the direct application of such models; but,
their study can inspire the engineering of artificial complex systems and reveal
properties with applications beyond the explanation of observable phenomena.
Studying consensus formation can thus advance our understanding of related
real-world social phenomena.
Examples of thematic real-world phenomena in social simulation models in-
clude for instance the joint assessments of policies or, in the context of economics
and politics, the voting problem. Herbert Simon's investigations on this problem
were one of the first stepping stones to the field of social simulation [2]; which
then inspired models (like the one we present in this article) related to evolution
and dissemination of opinions, which we call opinion dynamics. In agent-based
opinion dynamics, agent interactions are guided by social space abstractions.
In some models, the dimensionality and structure of this space is irrelevant
2
(any agent can interact with any other agent). Other models use an underlying
artefact that structures agent neighbourhoods. Axelrod for instance, represents
agent neighbourhoods as a bi-dimensional grid in its model of dissemination
of culture [3]. In an attempt to mimic real social systems, one can also make
use of complex network models to create the infrastructure that guides agent
interaction (see [4] for an example).
In real-world scenarios, actors engage in a multitude of social relations dif-
ferent in kind and quality. Most simulation models don't explore social space
designs that take into account the differentiation between coexisting social net-
works. Modelling multiple coexisting relations was an idea pioneered by Peter
Albin in [5] but without further development. The process of interacting in these
different complex social dimensions can be seen as the basis for the formation
of our social identity [6, 7].
In this article, we explore the of modelling opinion dynamics with multiple
social networks. We look at how the properties of different network models in-
fluence the convergence to opinion consensus. We present a series of models that
use these networks in different ways: which create distinct emergent dynamics.
We want to study the consequences of using multiple social networks at the
same time while maintaining the interaction model as simple and abstract as
possible (following the methodology in most opinion dynamics literature). We
present two models where agents can: interact at the same time in the multiple
networks (choosing partners from any network); or switch between networks
(choosing only partners from their current network).
1.1. Article Structure
This article is organised as follows. In the next section, we present the work
related to opinion dynamics, social space modelling and complex network mod-
els. In the following section, we describe our game of consensus and introduce
our multiple model variations. These are designed to study the notion of con-
text permeability and consensus formation in multiple social networks. Section
4 describes the experimental setup; the set of tool and methodologies followed
to conduct our investigations. In section 5, we present and discuss our results
and compare the different simulation models. Finally, we summarise what we
learned from our experiments and point out future research directions.
2. Related Work
In this section, we present work related to opinion dynamics, social space mod-
elling, and complex network models.
2.1. Opinion Dynamics and Consensus Formation
Formal opinion dynamics models provide an understanding, if not an analysis, of
opinion formation processes. An early formulation of such models was created
as a way to understand complex phenomena found empirically in groups [8].
The work on consensus building (in the context of decision-making) was first
3
studied by DeGroot [9] and Lehrer [10]. Empirical studies of opinion formation
in large populations have methodological limitations. Computational sociology
arises with a set of tools -- simulation models in particular -- to cope with such
limitations. We use multi-agent simulation (MAS) as a methodological frame-
work to study such social phenomena in a larger scale. Most opinion dynamics
models make use of binary opinion values [11, 12], or continuous values [13, 14].
For a detailed analysis over some opinion dynamics models, refer to [15].
Opinion dynamics models allow us to discover under which circumstances
a population of agents reaches consensus or polarisation. Agent-based models
can have broader application outside social simulation though.
In computa-
tional distributed systems, consensus is a means by which processes agree on
some data value needed during computation. Typically, this agreement is the
result of a negotiation process (often with the aid of a mediator). In human soci-
eties, social conventions emerge to deal with coordination and subsequently with
cooperation problems [16]. These conventions are regularities of behaviour that
can turn normative if they come to be persistent solutions to recurrent prob-
lems. MAS are also capable of producing emergent conventions in coordination
or cooperation problems [17]. The decentralised nature of these computational
models of consensus is highly desirable for dynamic control problems. In these
scenarios, creating conventions before hand (off-line), or developing a central
control mechanism for generating them, can be a difficult and intractable task:
either due to the uncertainty and complexity associated with the environment,
or due to the system scale and heterogeneity (which makes it very difficult to
handcraft each component).
Consensus models are also investigated as means to create conventions in
MAS. Shoham and Tennenholtz create a bridge between economic literature
and machine learning by studying a series of models in which multiple agents
are engaged in learning a particular convention [18]. They call this process co-
learning. The complexity of these systems comes from its concurrent nature:
one agent adapts to the behaviour of another agents it has encountered, these
other agents update their behaviour in a similar fashion, which results in a
highly non-linear system dynamics.
In their work, Shoham and Tennenholtz
define the notion of stochastic social games and compare different rules with
the objective of establishing conventions in a decentralised fashion.
In their
stochastic games, they present one particular rule that we use in our models:
Definition 1. External Majority (EM) update rule: adopt action i if so far it
was observed in other agents more often than other action and remain with your
current action in the case of equality.
This EM rule was shown to coincide with Highest Cumulative Reward (HCR),
which is a simple rule that states that "an agent should adopt an action that
has yielded the highest cumulative reward to date." We use the EM rule in our
simulation models for its simplicity and success in the evolution of conventions.
4
2.2. Social Structure in Simulation Models
The usage of a bi-dimensional grid to represent abstract social spaces in simula-
tion models, is one of the most widely used approaches in the agent-based simu-
lation and modelling literature. This has its origins in the "checkerboard" model
introduced in computational social sciences by Schelling and Sakoda [19, 20].
Another famous example that uses bi-dimensional grids as social structure is
the social simulation model of dissemination of culture from Axelrod [3].
Cellular automata (CA) are an example from the area of artificial life that
also makes use of bi-dimensional structures to model neighbourhood interac-
tions. While these models are idealised frameworks, their exploration can bring
us deeper insights than models with high level of descriptive complexity (which
would render their analysis very difficult). The work of Flache and Hegsel-
mann relaxes the standard CA assumptions about the regular bi-dimensional
grids by considering irregular grid structures (Voronoi diagrams) [21]. They
present results on the robustness of some important general properties in re-
lation to variations in the typical grid structure. There are also contributions
that incorporate both the complexity of network models and the behaviour of
dynamic processes. One example is the work in [22], which explores a graph-
based cellular automaton to study the relationship between spatial forms of
urban systems and the robustness of different process dynamics under spatial
change. Moreover, this shows how we can use real geographical information
to construct graph-based cellular automata (thus making a connection between
purely abstract models and data-driven models).
Finally, another type of social space models are the random graph/network
models. Each complex network -- or class of complex networks -- captures spe-
cific topological properties. These properties are found in real-world network
structures. One of the first complex network models is the random graph from
Erdos and R´enyi [23]. Other examples include the model of preferential attach-
ment [24], and the small-world networks [25]. One common issue of these models
is the fact that once generated, the network structure is static. The models are
not really suitable for the description of highly dynamic groups or communities.
An example of a model that takes such dynamics into account is the model of
team formation in [26]. A more in-depth review of network models (and their
properties) is given in [27].
3. Multi-context Models
In this section, we present our modelling approach to explore the concept of
permeability between contexts. We use multiple social networks to represent
the complex social space in which an agent is inserted. In a simulation model,
this setting can be seen as a n-dimensional scenario where each dimension con-
tains a network that represents a different social relation (see figure 1). The
word "context" is used in many different senses and it is by itself subjected to
many analysis [28]. Here, we use the term social context in a simpler and more
strict way. Agents belong to distinct social contexts which are their neighbour-
hoods in these multiple networks. The social context of an agent is thus a set
5
of neighbours currently available for interaction at a particular network or set
of networks. Context permeability is the ability of a particular norm, strategy,
opinion, or trend, to permeate from one social relation to another. Agents be-
longing to disjoint neighbourhoods in different networks serve as bridges between
subsets of the population that wouldn't otherwise be influenced by a consensus
formed in a distant coalition.
Figure 1: Multiple social network structure that shapes the social space
in our simulation models.
We study the notion of context permeability in different simulation models
in which agents interact using a simple consensus game. The society of agents
has to adopt a binary opinion value according to a majority rule. We use the
speed of consensus as a measure for self-organisation and explore the relationship
between different network topologies and this measure.
We partitioned the design space of our experiments using different simu-
lation models. Each model is designed to analyse different aspects of context
permeability. In a first model [29, 30], we study the notion of permeability by
overlapping social networks.
In a second model [12], we analyse the dynam-
ics introduced by switching between social contexts(neighbourhoods in different
networks). This adds a temporal component that changes the dynamics of the
game: agents can only perform encounters with neighbours that are active in
their current context. (It also adds the chance for some agents to be isolated
during the simulation.)
3.1. Simulation and Consensus Game
Our agent-based models are designed as discrete-event models. On each simula-
tion cycle, every agent executes a simulation step. The agents are selected in a
random uniform fashion for execution. This is common practice in agent-based
simulation and guarantees that there is no bias caused by the order in which
the agents are executed. For each simulation model, we present a description
of the individual agent behaviour. Every model we present can be seen as a
binary opinion dynamics model, or a consensus game. In this game, the soci-
ety of agents tries to reach an arbitrary global consensus about two possible
choices or opinion values. During a simulation run, each agent keeps track of
the number of each opinion value of their interaction partners. In each itera-
tion, each agent selects an available neighbour to interact with and observes its
current opinion value. The agent decides to switch its current opinion choice
6
if the observed opinion value becomes the majority of the two possible choices.
The goal of the game is to reach a global consensus, but the particular choice
that gets collectively selected is irrelevant. What is important is that overall
agreement is achieved. In the consensus game we consider, each agent uses the
previously presented external majority (EM) rule to update its opinion value
(see definition 1 in section 2).
The reason we are not interested in which exact option gets selected is the
same reason why we chose such a content-neutral game for the individual inter-
actions. From the very beginning, our research questions were focused on the
collective properties, the structure of the networks involved, and the dynami-
cal outcomes of the simulations: thus leaving out issues related with individual
motivations and desired results (either for any individual agent or the society).
Granted, most applications of this research will often include such goals, for
inquisition into complex social issues, including possibly policy design, in which
the individual rationality must be considered, and will have a very strong influ-
ence on the collective outcomes.
3.2. Context Permeability
The first model of context permeability [29, 30] is designed to study the setting
where an agent is immersed in a complex social world: where it engages in
a multitude of relations with other agents.
In this model, we consider that
the context permeability is created by agents that can belong simultaneously to
several relations. As an example, two agents can be simultaneously family and
co-workers. While some links might not connect two agents directly, others can
contain neighbourhoods in which they are related (either directly or through a
common neighbour). Model 1 describes the behaviour of each agent for each
simulated step.
7
Model 1 Context Permeability
M {Simulation Model / Environment}
A {Current Agent in Execution}
{Randomly select a network}
r ← randomUniform(0,M.networks.length())
network ← M.networks[r]
{Randomly select a neighbour}
neighbours ← network.neighboursOf(A)
r ← randomUniform(0,neighbours.length())
partner ← neighbours[r]
{Update opinion and memory}
{opinions take the values {0,1} -- so they can be used as an index}
A.memory[partner.opinion]++
if (A.memory[partner.opinion] > A.memory[A.opinion]) then
A.opinion ← partner.opinion
end if
In this model, on each time step, each agent select a random neighbour from
any social network available and updates its opinion based on the opinion of the
selected partner using the external majority rule.
3.3. Context Switching
In the context switching model [12], agents are embedded in multiple relations
represented as static social networks and they switch contexts with a probability
ζCi associated with context Ci. The switching probability is the same for every
agent in a given network. Each agent switches from its current neighbourhood
to another network. The agents are only active in one context at a time, and
can only perform encounters with neighbours in the same context. We can think
of context switching as a temporary deployment in another place, such as what
happens with temporary immigration. The network topology is static; when
an agent switches from one network to another, they become inactive in one
network and active in their destination.
8
Model 2 Context Switching
M {Simulation Model / Environment}
A {Current Agent in Execution}
{Get the current network (context is a network index)}
cNetwork ← M.networks[A.context]
{Get the neighbours for the current network}
neighbours ← cNetwork.neighboursOf(A)
{filter by agents active in the same network}
neighbours ← {n n ∈ neighbours ∧ (n.context = A.context)}
{Randomly select a neighbour}
r ← randomUniform(0,neighbours.length())
partner ← neighbours[r]
{Update opinion and memory}
{opinions take the values {0,1} -- so they can be used as an index}
A.memory[partner.opinion]++
if (A.memory[partner.opinion] > A.memory[A.opinion]) then
A.opinion ← partner.opinion
end if
{Switch to another network}
switchingProb ← M.params.switchingProb(A.context)
r ← randomUniform(0,1)
if (r < switchingProb) then
numNets ← M.networks.length()
nextNetworks ← {i (i ∈ [0,numNets[) ∧ (M.networks[i] != A.context)}
r ← randomUniform(0,nextNetworks.length())
A.context ← nextNetworks[r]
end if
In this model, agents select an (active) partner from their current neighbour-
hood. If a partner is available, the agents update their opinion value based on
EM. At the end of the interaction, the agent switches from the current context
to a different one with a probability associated with its current context.
This model describes an abstract way to represent the time spent on each
network using the switching probability ζCi. Here, the permeability between
contexts is achieved using this temporal component. Context switching intro-
duces a notion that has not been explored in the literature so far: the fact that,
although some social contexts can be relatively stable, our social peers are not
always available for interaction and spend different amounts of time in distinct
social contexts.
9
4. Experimental Setup
In this section, we present all the tools and processes necessary to produce
the current research output. The experiments were developed using the MA-
SON [31] simulation framework written in Java. All the code for the simulation
models can be found here [32]. Moreover, all the results presented in this article
were made reproducible by using the statistical computing language R [33] and
the R package knitr [34]. Knitr is used for generating reports that contain the
code, the results of its execution (plots and tables), and the textual description
for such results. All the R code used to produce the data analysis and the
configuration files used to reproduce the data can be found in [35].
For each configuration in our experiments, we performed 100 independent
simulation runs. The results are analysed in terms of value distribution, average
and standard deviation, or variance over the 100 runs. Like we described in
section 3.1, our models are discrete-event models. In each cycle agents execute
a step in a random order: this is so that this dynamical system -- our simulation
model -- is not influenced by the order in which agents are executed. Simulations
end when total consensus is achieved or 2000 steps have passed.
4.1. Social Network Models
The social networks used in our models are k-regular and scale-free. A k-regular
network is a network where all the nodes have the same number of connections.
These networks are constructed by arranging the nodes in a ring and connecting
each node to their next k neighbours. Each network has 2k edges per vertex.
Scale-free networks are networks in which the degree distribution follows
a power law, at least asymptotically. That is, the fraction P (k) of nodes in
the network having k connections to other nodes goes for large values of k
as: P (k) ∼ k−γ where γ is a parameter whose value is typically in the range
2 < γ < 3.
We use the method proposed in [24] by Barab´asi and Albert to construct the
scale-free network instances using a preferential attachment. This model builds
upon the perception of a common property of many large networks: a scale-
free-power-law distribution of node connectivity. This feature was found to be a
consequence of two generic mechanisms: networks expand continuously by the
addition of new vertices, and new vertices attach preferentially to sites that are
already well connected (more commonly known as "preferential attachment").
In these networks, the probability P (k) of two nodes being connected to each
other decays as a power law, following P (k) ∼ k−γ.
The network instances were generated using the b-have network library [36].
This is a Java library that allows the creation and manipulation of network/graph
data structures. It also includes the more commonly used random network mod-
els from the literature.
4.2. Measuring Network Properties
We measure two structural properties of our networks: the average path length
and the clustering coefficient. The average path length measures the typical sep-
aration between two vertices in the graph. The clustering coefficient measures
10
the average cliquishness of the graph neighbourhood. This measure quantifies
how close the neighbours of a node are to forming a clique. Duncan J. Watts and
Steven Strogatz [25] introduced the measure to characterise a class of complex
networks called small-world networks.
These properties characterise structures that can be found in the real scenar-
ios. The small-world networks generated by the models of Watts and Strogatz
for instance, model real-world social networks with short average path length,
local clustering structures, and triadic closures between the nodes [25]. One
shortcoming of this model, is the fact that these networks do not possess node
hubs found in many real-world social networks. (Nevertheless, these models are
designed to create networks with specific properties, not as general models for
all kinds of real social networks)
Since we construct scenarios where we overlap highly clustered k-regular
networks and scale-free networks with node hubs, we are interested in the kind
of properties that emerge from merging these structures -- and if/how they are
comparable to the existing models -- . The networks were produced by the b-
have network library [36], exported to files, and analysed using the igraph R
package [37].
5. Results and Discussion
In this section, we analyse and discuss the results for different experiments
with both the context permeability and context switching models. We start by
analysing the structural properties that arise from combining different networks.
We then observe the impact of different model configurations in the convergence
to consensus. We correlate the outcomes of our context permeability model sim-
ulations with the structural properties of the multi-network structures. We also
explore how the new dynamics introduced by the context switching mechanism
affects the consensus building process.
5.1. Overlapping Network Properties
In this first analysis, we investigate the properties of different network topologies
used in our models. One of the parameters in some experiments is the number
of networks in which the agents interact. Adding more networks implies the
addition of more connections. We show that what is important is not just the
the number of connections, but the properties of the resulting structure (when
we merge the networks). We analyse the networks as follows.
Each network is generated in such a way that the node indexes are ran-
domised. This means that we can have multiple networks with the same topol-
ogy and each node can have different neighbourhoods in different networks.
Also, neighbourhoods do not necessarily overlap due to the node shuffling.
Adding more networks to the social space is not the same as creating a sin-
gle network with twice the number of connections.
11
Figure 2: Overlapping of two k-regular networks, each with k=10.
Table 1: Properties for overlapping of two k-regular networks, each with
k=10.
Nodes Edges Clustering Coef. Avg. Path Length
10-regular 1
10-regular 2
Combined
20-regular
100
100
100
100
1,000
1,000
1,790
2,000
0.711
0.711
0.512
0.731
2.980
2.980
1.638
1.788
Consider figure 2. We created two random k-regular networks with k=10 and
merged them ignoring edges from common neighbours. (In the simulation mod-
els this edge agglutination is not performed, we use it here for the purposes of
network analysis.) Common edges are highlighted with a different colour.
The networks are drawn using the
Kamada-Kawai Layout [38] which treats
the "geometric" (Euclidean) distance be-
tween two vertices in the drawing as the
"graph theoretic" distance between them
in the corresponding graph. We can
see by the results of the layout in fig-
ure 2, that the distance between nodes
in the network has decreased, this is also
confirmed by the reduced average path
length of the combined networks (see ta-
ble 1). To illustrate our point, we anal-
ysed one k-regular network with k = 20
(see figure 3 and the corresponding entry on table 1). The clustering coefficient
of a 20-regular network is approximately the same as the one of a 10-regular
network (higher than the combination of two 10-regular networks). The average
path length also decreases as there are more connections between previously dis-
tant nodes. As we will show, these properties may vary due to the node indexes
being subjected to random permutations.
Figure 3: Single k-regular
network with k=20.
12
10−Regular 1llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll10−Regular 2llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllCombined 10−RegularsllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllK−Regular with K=20llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll Since we model the agent social space as a multitude of networks, we need
to investigate the properties resulting from the merging of these networks. We
do this empirically by taking the network instances used in the simulations, and
analysing the distribution of clustering coefficient and average path length for
the different configurations.
5.1.1. Properties of Overlapping K-Regular Networks
We will now investigate what kind of properties we can get from merging k-
regular networks. First, we analysed the distribution for the average path length
and the clustering coefficient values over 100 network instances (each with 100
nodes). From a box plot preliminary analysis (see figures A.2 and A.1), we can
see that the average path length does not vary much for the 100 instances, as
such, the average makes a good descriptor for these properties.
Figure 4 shows the average value for the average path length of 100 k-regular
network instances (each with 100 nodes). We can see that the average path
length changes more drastically when we go from 1 to 2 networks. This is
precisely the effect we can see in figure 2. Merging these networks at random
effectively creates multiple shortcuts between points that were not connected in
the original k-regular topology. Beyond this point, adding more networks does
not modify this structural property in a significant way. Note that, since each
network has 100 nodes, with k = 50 the network is fully connected (hence the
average path length being 1).
(a)
(b)
Figure 4: Average value for the average path length of 100 instances
of overlapping k-regular networks (containing 100 nodes each) with
k={1,2,3,4,5} (4a), and k= {10,20,30,40,50} (4b).
Figure 5 shows the results for the average clustering coefficient of 100 instances
of overlapping k-regular networks. Adding multiple networks changes the clus-
tering coefficient of the resulting structure. For k ≤ 20, the clustering coefficient
drops when we add a second network. This happens because when we merge
these networks, their connectivity is not high enough to generate shortcuts with
13
lllll51015202512345Number of NetworksAverage of Average Path LengthKl12345lllll1.01.52.02.53.012345Number of NetworksAverage of Average Path LengthKl1020304050 tightly clustered neighbourhoods. The clustering coefficient is computed by
taking the average local clustering coefficient of each node in the network. An
initial highly clustered k-regular network (with all the nodes having the same
structure) is affected when you modify the node neighbourhoods in such a way
that some are more clustered than others. Networks with k = 1 have a ring-like
structure, so there are no nodes forming a clique1 with their neighbours.
(a)
(b)
Figure 5: Average value for the clustering coefficient of 100 instances
of overlapping k-regular networks (containing 100 nodes each) with
k={1,2,3,4,5} (5a), and k = {10,20,30,40,50} (5b).
5.1.2. Properties of Overlapping Scale-free Networks
We also looked at the properties that result from merging scale-free networks
(see figure A.3). Like the previous results, the properties don't vary much
between the 100 different instances. We can then say that specific configurations
display a specific average path length and clustering coefficient both for k-regular
networks and scale-free networks. We plotted the average of the average path
length (figure 6a) and clustering coefficient (figure A.3b) for each overlapping
network configuration.
1A clique is a group of nodes such that every two nodes are connected by an edge.
14
lllll0.00.20.40.612345Number of NetworksAverage of Clustering CoefficientKl12345lllll0.50.60.70.80.91.012345Number of NetworksAverage of Clustering CoefficientKl1020304050 (a)
(b)
Figure 6: Average value for the average path length (A.3a) and clustering
coefficient (A.3b) for 100 instances of overlapping scale-free networks
(containing 100 nodes each).
The parameter d (figure 6) dictates how many edges are added by preferential
attachment (see section4.1). For d = 1, the resulting network is a forest: a
network composed of disjoint tree graphs. Figure 6a shows us that the average
path length still decreases consistently when we add more networks. One of the
characteristics of scale-free networks is its short average path length, thus, for
d ≥ 2, adding more edges does not make a significant difference. Scale-free
networks also have a reduced clustering coefficient (see figure 6b). When we
add more networks the clustering coefficient increases.
5.1.3. Merging K-Regular with Scale-Free Networks
Finally, we look at what happens when we merge k-regular networks with scale-
free networks. We didn't make an exhaustive analysis for this configuration,
but we show what is the resulting structure of such merging. You can see in
figure 7 that this results in a network with a lower average path length due to
the shortcuts created by the scale-free network. It is also less dense than the
structure with two 10-regular networks (figure 2). We consider only the merging
of two networks.
15
lllll23412345Number of NetworksAverage of Average Path LengthDl12345lllll0.00.10.20.30.412345Number of NetworksAverage of Clustering CoefficientDl12345 Figure 7: Overlapping of a 10-regular with a scale-free network with
D=2.
(a)
(b)
Figure 8: Average average path length for 100 instances of heterogeneous
network configurations: one scale-free with d = {1, 2, 3, 4, 5} and one k-
regular network with k = {1, 2, 3, 4, 5} (8a) and k = {10, 20, 30, 40, 50}
(8b). (See the respective box plots in the appendix: figure A.4).
Figure 8 shows the average path length of 100 independently generated in-
stances of combinations of one k-regular network and one scale-free network
with different values k and d respectively. For configurations where the k-regular
networks have low values of k (figure 8a), adding a scale-free network always de-
creases the average path length. Moreover, the higher the d, the lower the path
length (as more connections are formed between different nodes). With higher
values of k (figure 8b) however, increasing the value of d does not decrease the
average path length. This is not surprising given the level of connectivity of this
structure. The first network is already highly dense and connected, so adding
a few extra connections doesn't matter; what does matter is that we have a
scale-free network creating shortcuts between nodes that were not previously
connected.
16
10−Regular NetworkllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllScale−Free with D = 2llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllCombined Networkslllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll1.82.12.42.73.03.312345DAverage of Average Path LengthKl12345lllll1.21.51.812345DAverage of Average Path LengthKl1020304050 (a)
(b)
Figure 9: Average clustering coefficient for 100 instances of hetero-
geneous network configurations: one scale-free with d = {1, 2, 3, 4, 5}
and one k-regular network with k = {1, 2, 3, 4, 5} (A.3b) and k =
{10, 20, 30, 40, 50} (See the respective box plots in the appending: fig-
ureA.5).
Figure 9 shows the results for the clustering coefficient. These are similar to
the results we get for k-regular networks (see figure 5). The difference is that
we are not adding more networks but rather increasing the parameter d of the
scale-free (number of connections added by preferential attachment for each
note). In this case, the effects are qualitatively similar with the ones of merged
k-regular network with low value of k. In contrast, for high values of k (figure
9b), a higher value of d is not enough to make the network more clustered.
This happens because scale-free networks are more sparse and have a very low
number of triangles between nodes.
5.2. Context Permeability
In this section, we discuss the results for various sets of experiments on the con-
text permeability model [29, 30]. First, we used k-regular and scale-free networks
with our model. Each network layer was configured with the same topology. We
analyse the context permeability model in terms of convergence to consensus:
the ratio of simulations that converged to total consensus over 3000 runs; and
how many agent encounters were needed for this to happen.
5.2.1. Convergence Ratio
To analyse the ratio of convergence to consensus for different network configu-
rations, we correlate this ratio with the properties that the combined network
structure exhibits. As we have seen in section 5.1, the properties don't vary
much for different random instances, so we use the average as the descriptor for
these properties.
The following tables show the results of our simulations. Table 2 shows
the percentage of total consensus achieved with 2000 cycles for 3000 runs. For
17
lllll0.10.20.30.40.512345DAverage of Clustering CoefficientKl12345lllll0.50.60.70.80.91.012345DAverage of Clustering CoefficientKl1020304050 k-regular networks with a small k, consensus is rarely achieved with a single
network. However, as soon as we add more networks, consensus is achieved in
a significantly greater number of occasions. These results are especially inter-
esting for the scale-free networks (table 3): as soon as we add more networks,
for a scale-free with d ≥ 2, we can achieve consensus in a lot of occasions. This
is to show that achieving consensus is not just a matter of connectivity. With
scale-free networks, we can achieve consensus more often with less clustered
networks. This convergence is achieved not because of the clustering coefficient
but rather due to lower average path length. This is difficult to visualise us-
ing only the tables and the previous plots so, to confirm this hypothesis, we
measured the Spearman's correlation 2 between the ratio of convergences to con-
sensus, the average average path length, and the average clustering coefficient
for the corresponding network configurations (see table 4).
Table 2: Ratio of convergence to total consensus in 3000 independent
runs with with a number of concomitant networks (# nets.) equal in
kind: k-regular networks with k = {1, 2, 3, 4, 5, 10, 20, 30, 40, 50}.
# nets.
1
2
3
4
1
2
3
4
5
0.00
0.08
0.66
0.88
0.95
0.00
0.59
0.94
0.99
0.99
0.00
0.81
0.98
1.00
1.00
0.01
0.90
0.99
1.00
1.00
value for k
5
10
0.02
0.93
0.99
1.00
1.00
0.18
0.98
1.00
1.00
1.00
20
30
40
50
0.67
1.00
1.00
1.00
1.00
0.97
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
Table 3: Ratio of convergence to total consensus in 3000 independent
runs with a number of concomitant networks (# nets.) equal in kind:
scale-free networks with d = {1, 2, 3, 4, 5}.
# nets.
1
2
3
4
value for d
1
2
3
4
5
0.00
0.28
0.78
0.93
0.97
0.29
0.93
0.99
1.00
1.00
0.86
0.99
1.00
1.00
1.00
0.96
1.00
1.00
1.00
1.00
5
0.99
1.00
1.00
1.00
1.00
2The Spearman's rank correlation coefficient is a non-parametric measure of statistical
dependence between two variables. It assesses how well the relationship between two variables
can be described using a monotonic function.
18
Table 4: Spearman's correlation between convergence ratio (CR), average aver-
age path length (APL) and clustering coefficient (CC) along with the respective
p-value and confidence interval, for a confidence level of 95%.
network model
X
Y
Correlation CI(95%)
k-regular CR APL
k-regular CR
CC
scale-free CR APL
CC
scale-free CR
-0.726
0.264
-0.912
0.633
[-0.836, -0.561]
[-0.015, 0.505]
[-0.961, -0.808]
[0.318, 0.823]
p-value
2.44 × 10−9
6.40 × 10−2
2.28 × 10−10
6.78 × 10−4
5.2.2. Number of Encounters to Achieve Consensus
We now analyse the convergence in terms of average number of meetings during
a simulation run. Since the maximum number of simulation cycles is 2000, the
maximum number of encounters is 2 000 000 (we have 100 agents and each
one performs one encounter per cycle). We show that for some configurations,
the average of encounters is not a good descriptor since it varies greatly from
run to run. Rather than considering the average, it is better to look at the
distributions of these measures. Nevertheless, we included the data with all the
average number of encounters, as well as the respective standard deviations for
different network configurations. You can find this data in appendix B.
First, we present the results for k-regular networks. To observe the distri-
bution of the number of encounters, consider the box plots for k = {1, 2, 3, 4, 5}
(figure 10a) and k = {10, 20, 30, 40, 50} (figure 10b). Figure 10 shows the results
for configurations with a number of networks ≥ 3 (since less networks lead to
much worse results in terms of convergence speed). These plots reveal two types
of situations: the typical runs, in which the number of encounters doesn't vary
much, and some "outlier runs," which present some interesting dynamics, as we
will see later on.
(a)
(b)
Figure 10: Number of meetings over 100 independent simulation runs
for multiple k-regular networks: k = {1, 2, 3, 4, 5} (10a) and k =
{10, 20, 30, 40, 50} (10b)).
19
llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll050000100000150000200000345Number of NetworksNumber of EncountersK12345lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll0200004000060000345Number of NetworksNumber of EncountersK1020304050 Figure 10 shows us results qualitatively similar to what we previously observed
for convergence ratios (see section 5.2.1). Networks with lesser connectivity and
a greater average path length lead to total convergence less often. The simula-
tion outcomes also vary much more in comparison with the other configurations
(k > 1 and networks ≥ 3).
Notice that for k = 3 and networks = 3 (10b), we still have runs that deviate
from what we consider the typical behaviour. Instead of disregarding this as
an outlier, we explore this behaviour to better understand the dynamics of our
External Majority consensus game.
We analysed the network structure from this run and observed that there
were no distinct differences between the network instances from the "outlier
run" and the ones of the "typical" run. This pointed to the consensus game
itself as a possible cause for the irregular behaviour -- not to the network struc-
ture -- . What happens is that the agents reach a state from which converging
towards consensus is significantly harder. Villatoro [39] found that some com-
plex network models lead consensus games to form metastable sub-conventions.
These are very difficult to break and reaching 100% agreement is not as straight-
forward as assumed by previous researchers. We are not concerned with this
difficulty, instead, we want to explore how to describe what happens in simi-
lar cases. Figure 11 shows the consensus progression for two simulation runs a
typical case and the outlier.
(a) Typical Run
(b) Outlier Run
Figure 11: Number of agents with each opinion during a simulation run:
typical run in 11a and "outlier run" in 11b.
We take the outlier in the configuration with k = 3 and networks = 3, and
look at the memory that agents use in the consensus game. As we described
previously, agents record the number of individuals encountered with each opin-
ion value (two possible values in this case). We look at the average difference
between the number of values observed for all the agents. We also look at the
variance to find how the opinion observations evolve throughout the simulation.
Figure 12 shows the average difference between opinion memory for the first 10
steps (which was when the simulation converged for the normal run).
20
02550751000369Simulation StepNumber of Agentsnum−opinion−0num−opinion−1Normal Run02550751000200400600Simulation StepNumber of Agentsnum−opinion−0num−opinion−1Outlier Run (a) Typical Run
(b) Outlier Run
Figure 12: Average and variance of the difference between opinions
observed for the 100 agents: typical run in figure 12a and first steps of
the"outlier run" in figure 12b.
We can see that both the average difference in the opinion memory and
the variance grow faster in the normal run than in the outlier run. This is
not enough to establish a significant difference between the two. Moreover,
the memory differences are quite similar, but the number of agents for each
opinion was quite even. This happens due to the position of agents in the
network and their initial opinion values. These circumstances led the agents to
a initial stability. Note that the opinion difference continues to rise throughout
the simulation along with the variance. The variance starts to increase more
rapidly after 200 steps. The exponential growth reveals that the convergence
to consensus is not done evenly throughout the network, this is one of the
reasons why consensus takes more time to be achieved. Also, the low variance
in the first 200 steps shows us that the opinion strength was evenly distributed:
we have almost the same number of agents for each opinion but the memory
differences are qualitatively the same for both opinions. As soon as this variance
increases, one of the opinions breaks the stable state and convergence towards
global consensus begins.
21
02460.02.55.07.510.0Simulation StepAverage Memory Difference0240.02.55.07.510.0Simulation StepVariance of Opinion Memory Difference0120.02.55.07.510.0Simulation StepAverage Memory Difference0123450.02.55.07.510.0Simulation StepVariance of Opinion Memory Difference Figure 13: Average and variance of the difference between opinions
observed for the 100 agents throughout the outlier run.
In other words, different neighbourhoods in the network lead to different opinion
observation configurations (possibly even creating self-reinforcing structures).
Thus, the influence of heterogeneous connectivity is reflected on the growth of
opinion difference variance.
(a)
(b)
Figure 14: Number of meetings over 100 independent simulation runs
for multiple scale-free networks: with d={1,2,3,4,5} (14a) with a zoom
on the best configurations in 14b.
Figure 14 shows the average number of encounters during 100 independent sim-
ulation runs for scale-free networks. (We don't include the configuration for one
network and d = 1 because these never converge.) The convergence ratio is very
low for the configurations with one network d = 2 and 2 networks with d = 1;
hence the high number of encounters (see also table 3).
These results are similar to those of k-regular in that adding more networks
speeds up the convergence to consensus. The difference is that more "atypi-
cal" runs occur with multiple scale-free networks. Again, this is the effect of
these types of topologies. As we discussed previously, in these less connected
topologies, agents are more prone to be arranged in such a way that progres-
22
01002003000200400600Simulation StepAverage Memory Difference0500010000150000200400600Simulation StepVariance of Opinion Memory Differencelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll05000010000015000020000012345Number of NetworksNumber of Encountersd12345llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll25005000750010000345Number of NetworksNumber of Encountersd2345 sion towards consensus is more difficult to attain due to bottlenecks. Different
regions in the network might converge towards different opinion values.
5.2.3. Heterogeneous Network Configuration
We also performed experiments with mixed network topologies: we used one
scale-free and one k-regular networks. Table 5 shows the convergence ratio for
the heterogeneous network configuration with the multiple values for d and k
for the scale-free and k-regular respectively.
Table 5: Ratio of convergence to total consensus in 3000 independent
runs with two networks: one k-regular and one scale-free network with
different k and d values respectively.
k
d
1
2
3
4
5
1
2
3
4
5
10
20
30
40
0.18
0.63
0.77
0.73
0.87
0.47
0.72
0.86
0.92
0.85
0.55
0.88
0.87
0.92
0.94
0.67
0.89
0.97
0.97
0.94
0.61
0.90
0.98
0.96
0.98
0.76
0.97
0.99
0.96
0.99
0.77
1.00
0.98
0.99
1.00
0.81
1.00
0.99
1.00
1.00
0.75
0.99
1.00
1.00
1.00
50
0.82
1.00
1.00
1.00
1.00
One of the differences between this configuration and the previous ones can be
observed in the simulations with d = 1 and k = 1. The convergence ratio is
higher than with 2 1− regular networks (see table 2) -- but not better than two
scale-free networks. This happens because with k = 1 the network is basically a
ring and has the maximum possible average path length for a connected graph.
Adding two rings improves the convergence ratio but the underlying structure
is still very susceptible to self-reinforcing structures (a connected sub-graph is
basically a line). Adding a scale-free changes this drastically as we are mixing
tree-like network components with a ring.
23
(a)
(b)
Figure 15: Number of meetings over 100 independent simulation runs
for heterogeneous configuration:
one k-regular network with k =
{1, 2, 3, 4, 5} and one scale-free network with d{1, 2, 3, 4, 5}. Figure 15b
shows a zoom on the best configurations.
Figure 15 shows the results for the number of encounters necessary for the
agents to achieve consensus. We focus on the k-regular networks with higher
values of k. We can see that with these two networks, the configurations with
values of d >= 2 for scale-free networks produce drastically better results in
terms of speed of convergence. The data for the average number of encounters,
as well as the respective standard deviations referent to the results in figure 15,
can be found in appendix B, table B.3.
5.3. Context Switching
In this section, we present the results relative to the context switching model.
The difference between this model and the previous one is that the agents no
longer interact in multiple networks at the same time (being able to select any
neighbour at a given simulation step). In this model, they become active in a
single network at a time. Agents can switch to a different network at the end
of each step. The switching mechanism uses a probability associated with each
network. With this new idea of swapping contexts, we covered some space left
undeveloped in Antunes and colleagues original work [29, 30].
As we discussed in section 3.3 (see model 2), the switching probability dic-
tates the frequency with which an agent switches from the current network
after an encounter has been performed. In an abstract manner, this allows us
to model how much time agents spend on each network. A further development
of this model will be to assign different preferences to different networks for
each agent. This can help us model phenomena such as real-world agents that
dedicate more or less time consuming content from different social networks.
For now, we attribute this probability to the network to restrain the model
complexity.
24
lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll05000010000015000020000012345DNumber of EncountersK1020304050lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll250050007500100002345DNumber of EncountersK1020304050 5.3.1. Exploring the Switching Probability
We first look at the influence of the new parameter (switching probability) in the
convergence to consensus. We span the switching probability between 0 and 1 in
intervals of 0.05. We do this for 2 networks with k-regular and scale-free topolo-
gies. Figure 16 shows the results for two k-regular networks: one configuration
with k = 10 and one configuration with k = 30. We use these two configura-
tions to study the relationship between the number of encounters, the switching
probability parameter, and the connectivity of the networks. Remember that
with k = 30 the average path length is also lower (section 5.1.1).
(a) 10-regular Networks
(b) 30-regular Networks
Figure 16: Contour plot for the average number of meets during a sim-
ulation for 100 independent runs: 2 10-regular networks (16a) and 30-
regular networks (16b). See the perspective plots in appendix C.
For smaller values of k, symmetry in the context switching probability (having
the same probability in both networks) is more important if the agent switches
less from one of the networks. Switching less from one network means spending
more time in that network. This means that switching more from the other
network can be disruptive to a neighbourhood that has already converged to a
sub-convention. This can be observed in configurations with k-regular networks
(figure 16a) but it is especially apparent in scale-free networks (figure 17a).
25
500010000100001000010000100001000020000200002000020000500020000700001200001700002e+05500050005000500050005000100000.000.250.500.751.000.000.250.500.751.00Switching for Network 1 regular network, k=10Switching for Network 2 regular network, k=1050000100000150000200000Average Encounters500050005000500010000100001000010000100001000010000500010000200007000011111150005000500050005000500050000.000.250.500.751.000.000.250.500.751.00Switching for Network 1 regular network, k=30Switching for Network 2 regular network, k=30250005000075000100000Average Encounters (a) Scale-Free d = 1
(b) Scale-Free d = 5
Figure 17: Contour plot for the average number of meets during a simu-
lation for 100 independent runs: 2 scale-free networks with d = 1 (17a)
and d = 5 (17b). See the perspective plots in appendix C.
When we increase the connectivity (and lower the average path length),
both in k-regular and scale-free networks, the switching probability becomes
"irrelevant" -- for any probability value, the speed of convergence will be the
same -- . The exception is obviously when considering the switching probability
with value 0. In this case, agents are isolated (they are initially equally dis-
tributed throughout the networks) and reaching consensus becomes very hard.
(Especially because agents are active only in one network and this can create
disconnected graphs.)
One major difference from the previous model is that the agents no longer
interact with a neighbour from any network at any time. The neighbour selec-
tion has to be made from the network they are currently active in. This causes
the agents to spend different amounts of time in different networks. An agent
can be very influential in one network -- being a very central node in that context
-- and marginally important in another:
in a sense its opinion in one network
can be more important to overall convergence than in other network, due to its
position in the network.
To explore the contribution of connectivity to the new context switching
dynamics, we setup two scale-free networks; one with d = 1 and the other with
d = 5. We then spanned the switching probability in increments of 0.05. The
results can be seen in figure 18.
26
120000170000170000199957.580.000.250.500.751.000.000.250.500.751.00Switching for Network 1 scale−free network, d=1Switching for Network 2 scale−free network, d=190000120000150000180000Average Encounters50005000500050005000500050005000500050005000500050005000500010000100001000010000100001000010000500010000100001000010000100001000020000700001200001700005000187865.59500050005000500050000.000.250.500.751.000.000.250.500.751.00Switching for Network 1 scale−free network, d=5Switching for Network 2 scale−free network, d=550000100000150000Average Encounters Figure 18: Contour plot for the average number of meets during a sim-
ulation for 100 independent runs: 2 scale-free networks, the first with
d = 5 and the second with d = 1. See the perspective plots in appendix
C (figure C.5).
In this case it is more important (to convergence speed) to switch from the
network with lower average path length and a forest-like composition (and a
probability of at least the same value or more than the one of the other network
with d = 5). The same can be observed when we mix k-regular with scale-free
networks (see figure 19).
Switching less from the less connected network is bad in both cases:
it is
possible that spending more time in the network with bigger neighbourhoods
allows for a stabler convergence. Sub-conventions can emerge in scale-free net-
works with d = 1 because usually some network regions are isolated by a single
node -- the root of the sub-tree they belong to.
27
500050001000010000100001000010000100001000010000100005000100001000010000100001000010000100002000020000700005000120000170000199979.665000500050005000500050000.000.250.500.751.000.000.250.500.751.00Switching for Network 1 scale−free network, d=5Switching for Network 2 scale−free network, d=150000100000150000Average Encounters Figure 19: Contour plot for the average number of meets during a sim-
ulation for 100 independent runs: one 10-regular network (k=10) and a
scale-free network with d = 1. See the perspective plots in appendix C
(figure C.6).
5.3.2. Comparison with Context Permeability
Finally, to compare the context switching model with the previous model of con-
text permeability, we froze the switching probabilities with values ζ = 0.25, 0.5, 0.75
and varied the number of networks. (We considered these values because they
conveniently characterise the probability in low, medium and high switching.)
We compare the results with the context permeability in terms of average num-
ber of encounters to achieve consensus. Note that in some cases, due to the
distribution of the number of encounters, the average is not an accurate de-
scriptor for the convergence. Nonetheless, configurations with highly variable
outcomes usually produce an average that is qualitatively distinct from the rest
(see figure 10 in section 5.2.2). We took the average number of encounters dur-
ing the simulations to compare the context switching model with the context
permeability model.
28
20000199985.852000020000700001200001200001200001700001700000.000.250.500.751.000.000.250.500.751.00Switching for Network 1 regular network, k=10Switching for Network 2 scale−free network, d=150000100000150000Average Encounters (a)
(b)
(c)
(d)
Figure 20: Average number of encounters during a simulation for 100
independent runs with k-regular topologies in all the network layers.
We compare the results of context switching with context permeability
with a numberof networks = {2, 3, 4, 5}. The switching probability is
the same in all the networks.
Figure 20 shows that for 2 networks the results in terms of number of en-
counters for the context switching approximate those for context permeability.
When we increase the number of networks (n ≥ 3), the configurations with
higher switching probability lead the results to be closer to what happens in
context permeability. This is no surprise since more switching makes agents
switch more often between networks and consequently allows them too choose
more often from different neighbourhoods. This is almost the same as having
a larger neighbourhood to choose from, which is what happened in the context
permeability model.
The surprise was that for values of k ≥ 20, the results where practically the
same for both models independently of the switching probability -- even when we
increased the number of networks, the number of encounters remained around
2000 on average for all the models. Above a certain level of neighbourhood size
(and overlapping, which causes the average path length to drop), the switching
probability becomes less influential. It is still important that we get compara-
29
llllllllll050000100000150000200000123451020304050KAverage EncountersCSll0.250.50.75permeabilityK−Regular Networks With Number of Networks = 2llllllllll0250005000075000100000123451020304050KAverage EncountersCSll0.250.50.75permeabilityK−Regular Networks With Number of Networks = 3llllllllll010000200003000040000123451020304050KAverage EncountersCSll0.250.50.75permeabilityK−Regular Networks With Number of Networks = 4llllllllll100002000030000123451020304050KAverage EncountersCSll0.250.50.75permeabilityK−Regular Networks With Number of Networks = 5 ble results because we are introducing a temporal component that was never
explored before in these types of opinion dynamics models: the fact that the
agents can become active in different networks at different points in time.
(a)
(b)
(c)
(d)
Figure 21: Average number of encounters during a simulation for 100
independent runs with scale-free topologies in all the network layers.
We compare the results of context switching with context permeability
with a numberof networks = {2, 3, 4, 5}. The switching probability is
the same in all the networks.
For scale-free networks, figure 21 shows us similar results. Switching with a
higher probability displays results similar with those of context permeability. In
this case, adding more networks makes so that lowering switching probabilities
displays slightly worse results. This influence comes from the neighbourhoods
being considerably smaller in scale-free networks -- even with values of d = 5 --
so the impact of low switching was bound to be noticed.
An increasing number of networks seems to always reduce the number of
encounters to achieve consensus in both models. This happens due to what
we called permeability. Having multiple points of dissemination that permeate
between different social networks enhances the overall convergence to consensus.
With permeability there are more previously isolated nodes being reached in
30
lllll0e+005e+041e+0512345DAverage EncountersCSll0.250.50.75permeabilityScale−Free Networks With Number of Networks = 2lllll020000400006000012345DAverage EncountersCSll0.250.50.75permeabilityScale−Free Networks With Number of Networks = 3lllll100002000012345DAverage EncountersCSll0.250.50.75permeabilityScale−Free Networks With Number of Networks = 4lllll5000100001500012345DAverage EncountersCSll0.250.50.75permeabilityScale−Free Networks With Number of Networks = 5 sparsely connected networks such as scale-free networks.
6. Conclusions
In this article, we analysed and discussed our modelling framework for multi-
relational models of social spaces. These results show that not only contexts
play an important role in the dissemination of consensus in artificial agent soci-
eties, but also that simple mechanisms can generate a great deal of complexity
(especially in those scenarios where conventions are being co-learned by mul-
tiple agents). Adding more networks to our opinion dynamics model leads to
outcomes where consensus was achieved not only more often but also faster. In
particular, our results show that achieving convergence (i.e. total consensus in a
network) is not a matter of connectivity. With scale-free networks we managed
to achieve consensus more often and quicker than with less clustered networks.
This conclusion is especially important for application (or models) dealing with
real-world social phenomena, many times represented with such networks.
The second innovation presented in this article was the fact that our model
describes an abstract way to represent the time spent on each network using
the switching probability. This temporal component introduces a new dynamic
to the study of opinion dynamics. Considering that in agent neighbourhoods,
agents might not be available at all times -- and spend different amounts of
time in different neighbourhoods -- , this is an important factor to be included in
simulation models (either representing real-world systems with different levels
of abstraction, or artificial agent societies for particular applications).
Future work includes the analysis of different complex network models.
While k-regular and scale-free networks are two of the most pervasive network
examples in social simulation, there has been a growing number of models with
different properties inspired by real-world phenomena that can be considered.
We will also analyse the networks according to different properties such as be-
tweeness centrality of certain nodes. Moreover, we plan to apply what we learned
with our consensus games to cooperation problems. Constructing agents that
can learn both behaviour and coordination procedures is a difficult task (espe-
cially when working with large distributed artificial societies). Consensus games
such as this one can be used to reduce the search space in problems where agents
need to cooperate or coordinate in a decentralised fashion.
While the convergence and number of encounters seem to be correlated with
the average path length, this might not be the only property that plays a role in
faster convergence to consensus. Correlation is just an informal tool that helps
us assert if some monotonic relationship exists. More research is needed to find
out, for instance, if self-reinforced structures -- like the ones reported by [39] --
exist in our multi-network structures. In particular, we are interested in finding
out if permeability mitigates the effects of such phenomenon.
We have seen that context permeability by context overlapping or switching
can provide a fairly straightforward modelling methodology. This can be used
to compose simulation scenarios for more complex social spaces. What remains
31
to be done is to identify more precise contextual structures in real-world net-
works. While capturing the structure of real social network can be a daunting
task, we do have pervasive records of networking activity between social actors:
online social networks. Analysing these can lead us to insights about complex
structural properties that emerge from user activity, and contexts that can be
identified within those networks. A difficult -- but not impossible -- prospect for
further research would be to track similar contexts between different real online
social networks.
Acknowledgements
Work supported by the Funda¸cao para a Ciencia e a Tecnologia under the
grant SFRH/BD/86034/2012 and by centre grant (to BioISI, Centre Reference:
UID/MULTI/04046/2013), from FCT/MCTES/PIDDAC, Portugal
32
A. Overlapping Network Properties
A.1. Homogeneous Network Topology Configurations
(a)
(b)
Figure A.1: Average path length for 100 instances of overlap-
ping k-regular networks(containing 100 nodes each) with
k =
{1, 2, 3, 4, 5} (A.1a), and k = {10, 20, 30, 40, 50} (A.1b). Since there
is barely any variation in the property values for each configuration, the
colours cannot be seen correctly, they are presented in the same order
as the legend nonetheless.
(a)
(b)
Figure A.2: Clustering coefficient
for 100 instances of overlap-
ping k-regular networks (containing 100 nodes each) with
k =
{1, 2, 3, 4, 5} (A.2a) and k = {10, 20, 30, 40, 50} (A.2b). Note that the
variation is bigger for networks with less connectivity. Also, for k = 50,
the network is fully connected.
33
lllllllllllllllllllllllllll51015202512345Number of NetworksAverage Path LengthK12345lllllllllllllllllllllllllll1.01.52.02.53.012345Number of NetworksAverage Path LengthK1020304050lllllllllllllllllllllllllllllll0.00.20.40.612345Number of NetworksClustering CoefficientK12345lllllllllllllllllllllllllllll0.50.60.70.80.91.012345Number of NetworksClustering CoefficientK1020304050 (a) Average path length
(b) Clustering Coefficient
Figure A.3: Average path length and clustering coefficient for 100 in-
stances of overlapping scale-free networks. D is a parameter used to
construct the network. It dictates how many edges are added by prefer-
ential attachment each time a new node is added to the networks. For
more details refer to [24].
A.2. Heterogeneous Network Topology Configurations
(a)
(b)
Figure A.4: Average path length for 100 instances of two overlapping
networks: 1 scale-free with d = {1, 2, 3, 4, 5} and 1 k-regular network
with k = {1, 2, 3, 4, 5} (A.4a), and k = {10, 20, 30, 40, 50} (A.4b). Since
there is barely any variation in the property values for each configu-
ration, the colours cannot be seen correctly in some cases, they are
presented in the same order as the legend nonetheless.
34
llllllllllllllllllllllllllllllllllllll2345612345Number of NetworksAverage Path LengthD12345llllllllllllllllllllllll0.00.10.20.30.412345Number of NetworksClustering CoefficientD12345lllllllllllllllllllllllllllllllll2.02.53.03.512345DAverage Path LengthK12345lllllllllllllllllllll1.21.51.82.112345DAverage Path LengthK1020304050 (a)
(b)
Figure A.5: Clustering coefficient for 100 instances of two overlapping
networks: 1 scale-free with d = {1, 2, 3, 4, 5} and 1 k-regular network
with k = {1, 2, 3, 4, 5} (A.4a), and k = {10, 20, 30, 40, 50} (A.4b). Since
there is barely any variation in the property values for each configu-
ration, the colours cannot be seen correctly in some cases, they are
presented in the same order as the legend nonetheless.
35
llllllllllllllllllllllllllll0.00.10.20.30.40.512345DClustering CoefficientK12345llllllllll0.60.81.012345DClustering CoefficientK1020304050 B. Context Permeability Analysis
Table B.1: Average number of meetings during a simulation run for 100 agents and a maximum of 2000 simulation
cycles for different configuration of k-regular networks.
Number of Networks
k
1
2
3
4
5
10
20
30
40
50
avg.
200000
200000
198033
198093
196042
178207
75541
11703
1694
1788
1
2
sd
0.0
0.0
19670.0
19070.0
27845.6
62301.7
95871.3
43425.0
954.5
786.9
avg.
191237
99744
68555
38359
21287
8066
2338
1952
1646
1724
sd
35780.6
89096.9
82808.5
69305.4
48141.1
28298.4
2739.9
1087.9
821.9
744.7
avg.
97770
22690
10217
5028
3444
2894
2178
1726
1862
1776
3
sd
86648.0
50813.5
33963.1
10680.0
5144.0
6424.5
1588.3
936.6
872.3
827.7
avg.
44827
7514
6539
3710
2834
1800
1936
1930
1806
1842
4
sd
71714.8
21352.6
22227.3
4618.7
4334.5
1148.2
989.4
1096.3
884.6
896.8
avg.
14210
4602
2828
2678
2278
2180
1802
1956
1686
1896
5
sd
31633.6
6116.7
2283.6
2127.4
1921.5
2078.3
823.6
1273.5
647.1
911.4
3
6
Table B.2: Average number of meetings during a simulation run for 100 agents and a maximum of 2000 simulation
cycles for different configuration of scale-free networks.
1
Number of Networks
2
3
4
d
1
2
3
4
5
avg.
sd
avg.
sd
200000
150632
37824
13751
8802
0
78710
64073
33295
18754
141085
21654
12309
4846
2536
83720
41341
36441
7921
1880
avg.
66020
7555
5299
2852
3204
sd
82601
21407
20035
2766
11008
avg.
17712
5890
2804
2370
2166
sd
42444
10889
3791
1441
1269
avg.
14917
3638
2618
2200
1958
5
sd
38343.6
5127.0
2445.4
1433.4
971.6
3
7
Table B.3: Average number of meetings during a simulation run for 100 agents and a maximum of 2000 simulation
cycles for different configuration of 1 k-regular and 1 scale-free networks.
1
2
avg.
sd
avg.
sd
173312
118591
109155
80035
89179
62416
61371
51559
58625
48702
61519
90107
88722
88719
91823
81036
81409
77468
83049
74718
100175
72054
38270
33339
31044
11017
6680
4980
5483
6568
88406
86077
66714
64075
62485
35071
17900
12740
20169
22611
k
1
2
3
4
5
10
20
30
40
50
d
3
avg.
63191
43308
34899
21289
13908
9095
6944
5095
2726
2566
sd
81071
70318
65533
45725
36791
27266
27826
20028
3735
1716
4
5
sd
84549
58760
55589
39516
39490
39798
19832
1355
1503
1619
avg.
51313
43281
23436
24034
9252
5559
2390
2046
2412
1980
sd
67704
72182
49978
52407
28404
20448
2313
1225
1874
1081
avg.
73715
31756
26412
15575
14112
12182
4721
2268
2346
2362
C. Context Switching Analysis
Figure C.1: Perpective plot for the average number of meets during a
simulation for 100 independent runs: 2 10-regular networks (k = 30).
Figure C.2: Perpective plot for the average number of meets during a
simulation for 100 independent runs: 2 30-regular networks (k = 30).
Figure C.3: Perpective plot for the average number of meets during a
simulation for 100 independent runs: 2 scale-free networks with d = 1.
38
Switching for Network 1 regular network, k=100.00.20.40.60.81.0Switching for Network 2 regular network, k=100.00.20.40.60.81.0 Avg. Encounters50000100000150000200000Switching for Network 1 regular network, k=100.20.40.60.81.0Switching for Network 2 regular network, k=100.20.40.60.81.0 Avg. Encounters20000400006000080000Switching for Network 1 regular network, k=100.40.50.60.70.80.91.0Switching for Network 2 regular network, k=100.40.50.60.70.80.91.0 Avg. Encounters10000200003000040000Switching for Network 1 regular network, k=300.00.20.40.60.81.0Switching for Network 2 regular network, k=300.00.20.40.60.81.0 Avg. Encounters2e+044e+046e+048e+041e+05Switching for Network 1 regular network, k=300.20.40.60.81.0Switching for Network 2 regular network, k=300.20.40.60.81.0 Avg. Encounters2000300040005000Switching for Network 1 regular network, k=300.40.50.60.70.80.91.0Switching for Network 2 regular network, k=300.40.50.60.70.80.91.0 Avg. Encounters2000250030003500Switching for Network 1 scale−free network, d=10.00.20.40.60.81.0Switching for Network 2 scale−free network, d=10.00.20.40.60.81.0 Avg. Encounters100000120000140000160000180000Switching for Network 1 scale−free network, d=10.20.40.60.81.0Switching for Network 2 scale−free network, d=10.20.40.60.81.0 Avg. Encounters100000120000140000Switching for Network 1 scale−free network, d=10.40.50.60.70.80.91.0Switching for Network 2 scale−free network, d=10.40.50.60.70.80.91.0 Avg. Encounters90000100000110000120000130000 R
Figure C.4: Perpective plot for the average number of meets during a
simulation for 100 independent runs: 2 scale-free networks with d = 5.
Figure C.5: Perspective plot for the average number of meets during
a simulation for 100 independent runs: 2 scale-free networks, the first
with d = 5 and the second with d = 1.
Figure C.6: Perspective plot for the average number of meets during a
simulation for 100 independent runs: 1 10-regular network (k = 10) and
a scale-free network with d = 1.
39
Switching for Network 1 scale−free network, d=50.00.20.40.60.81.0Switching for Network 2 scale−free network, d=50.00.20.40.60.81.0 Avg. Encounters50000100000150000Switching for Network 1 scale−free network, d=50.20.40.60.81.0Switching for Network 2 scale−free network, d=50.20.40.60.81.0 Avg. Encounters400060008000Switching for Network 1 scale−free network, d=50.40.50.60.70.80.91.0Switching for Network 2 scale−free network, d=50.40.50.60.70.80.91.0 Avg. Encounters3000400050006000Switching for Network 1 scale−free network, d=50.00.20.40.60.81.0Switching for Network 2 scale−free network, d=10.00.20.40.60.81.0 Avg. Encounters50000100000150000Switching for Network 1 scale−free network, d=50.20.40.60.81.0Switching for Network 2 scale−free network, d=10.20.40.60.81.0 Avg. Encounters50000100000150000Switching for Network 1 scale−free network, d=50.40.50.60.70.80.91.0Switching for Network 2 scale−free network, d=10.40.50.60.70.80.91.0 Avg. Encounters50000100000150000Switching for Network 1 regular network, k=100.00.20.40.60.81.0Switching for Network 2 scale−free network, d=10.00.20.40.60.81.0 Avg. Encounters50000100000150000Switching for Network 1 regular network, k=100.20.40.60.81.0Switching for Network 2 scale−free network, d=10.20.40.60.81.0 Avg. Encounters50000100000150000Switching for Network 1 regular network, k=100.40.50.60.70.80.91.0Switching for Network 2 scale−free network, d=10.40.50.60.70.80.91.0 Avg. Encounters50000100000150000 References
[1] C. Emmeche, Aspects of complexity in life and science, PHILOSOPHICA-
GENT- (1997) 41 -- 68.
[2] H. A. Simon, Bandwagon and underdog effects and the possibility of elec-
tion predictions, The Public Opinion Quarterly 18 (3) (1954) pp. 245 -- 253.
[3] R. Axelrod, The Dissemination of Culture, Journal of Conflict Resolution
41 (2) (1997) 203 -- 226.
[4] G. Weisbuch, Bounded confidence and social networks, The European
Physical Journal B - Condensed Matter and Complex Systems 38 (2004)
339 -- 343.
[5] P. S. Albin, The analysis of complex socioeconomic systems, Lexington,
Mass. : Lexington Books, 1975.
[6] S. Roccas, M. B. Brewer, Social identity complexity, Personality and Social
Psychology Review 6 (2) (2002) 88 -- 106.
[7] N. Ellemers, R. Spears, B. Doosje, Self and social identity*, Annual Review
of Psychology 53 (1) (2002) 161 -- 186.
[8] J. French, A formal theory of social power, Psychological Review 63 (1956)
181 -- 194.
[9] M. H. DeGroot, Reaching a Consensus, Journal of the American Statistical
Association 69 (345) (1974) 118 -- 121.
[10] K. Lehrer, Social consensus and rational agnoiology, Synthese 31 (1975)
141 -- 160.
[11] S. Galam, Rational group decision making: A random field ising model t
= 0, Physica A: Statistical Mechanics and its Applications 238 (1) (1997)
66 -- 80.
[12] L. Antunes, D. Nunes, H. Coelho, J. a. Balsa, P. Urbano, Context Switching
versus Context Permeability in Multiple Social Networks, in: Progress in
Artificial Intelligence, Vol. 5816 of Lecture Notes in Computer Science,
2009, pp. 547 -- 559.
[13] G. Deffuant, D. Neau, F. Amblard, G. Weisbuch, Mixing beliefs among
interacting agents, Advances in Complex Systems 3 (01n04) (2000) 87 -- 98.
[14] G. Deffuant, F. Amblard, G. Weisbuch, T. Faure, How can extremism pre-
vail? A study based on the relative agreement interaction model, Journal
of Artificial Societies and Social Simulation 5 (4) (2002) 1.
[15] R. Hegselmann, U. Krause, Opinion dynamics and bounded confidence:
models, analysis and simulation, Journal of Artificial Societies and Social
Simulation 5 (3) (2002) 2.
40
[16] D. K. Lewis, Convention: a philosophical study, Harvard University Press
Cambridge, 1969.
[17] J. Delgado, Emergence of social conventions in complex networks, Artificial
intelligence 141 (1) (2002) 171 -- 185.
[18] Y. Shoham, M. Tennenholtz, Co-learning and the evolution of social activ-
ity, Tech. rep., DTIC Document (1994).
[19] T. C. Schelling, Dynamic models of segregation†, Journal of mathematical
sociology 1 (2) (1971) 143 -- 186.
[20] J. M. Sakoda, The checkerboard model of social interaction, The Journal
of Mathematical Sociology 1 (1) (1971) 119 -- 132.
[21] A. Flache, R. Hegselmann, Do irregular grids make a difference? relaxing
the spatial regularity assumption in cellular models of social dynamics.,
Journal of Artificial Societies and Social Simulation 4 (4).
[22] D. B. O'Sullivan, Graph-based cellular automaton models of urban spatial
processes, Ph.D. thesis, University of London (2000).
[23] P. Erdos, A. R´enyi, On random graphs, Publ. Math. Debrecen 6 (1959)
290 -- 297.
[24] A. Barab´asi, R. Albert, Emergence of scaling in random networks, science
286 (5439) (1999) 509.
[25] D. J. Watts, S. H. Strogatz, Collective dynamics of 'small-world' networks,
Nature 393 (6684) (1998) 440 -- 442.
[26] M. E. Gaston, M. desJardins, Agent-organized networks for dynamic team
formation, in: Proceedings of the fourth international joint conference on
Autonomous agents and multiagent systems, ACM, 2005, pp. 230 -- 237.
[27] D. Nunes, L. Antunes, Consensus by segregation - the formation of local
consensus within context switching dynamics, Proceedings of the 4th World
Congress on Social Simulation.
[28] P. J. Hayes, Contexts in context, in: Context in knowledge representation
and natural language, AAAI Fall Symposium, 1997.
[29] L. Antunes, J. Balsa, P. Urbano, H. Coelho, The challenge of context per-
meability in social simulation, in: Proceedings of the Fourth European
Social Simulation Association, 2007.
[30] L. Antunes, J. Balsa, P. Urbano, H. Coelho, Exploring context permeability
in multiple social networks, in: K. Takadama, C. Cioffi-Revilla, G. Deffuant
(Eds.), Simulating Interacting Agents and Social Phenomena, Vol. 7 of
Agent-Based Social Systems, Springer Japan, 2010, pp. 77 -- 87.
41
[31] S. Luke, C. Cioffi-Revilla, L. Panait, K. Sullivan, G. Balan, MASON: A
Multiagent Simulation Environment, Simulation 81 (7) (2005) 517 -- 527.
[32] D. Nunes, L. Antunes, Context Permeability Modelsdoi:{10.5281/
zenodo.11067}.
URL http://dx.doi.org/10.5281/zenodo.11067
[33] R. D. C. Team, R: A Language and Environment for Statistical Computing,
R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-
07-0 (2008).
URL http://www.R-project.org
[34] Y. Xie, knitr: A comprehensive tool for reproducible research in R, in:
V. Stodden, F. Leisch, R. D. Peng (Eds.), Implementing Reproducible Com-
putational Research, Chapman and Hall/CRC, 2014.
URL http://www.crcpress.com/product/isbn/9781466561595
[35] D. Nunes, L. Antunes, Context Permeability Analysisdoi:{10.5281/
zenodo.11898}.
URL http://dx.doi.org/10.5281/zenodo.11898
[36] D. Nunes, L. Antunes, B-Have Network Librarydoi:{10.5281/zenodo.
11069}.
URL http://dx.doi.org/10.5281/zenodo.11069
[37] G. Csardi, T. Nepusz, The igraph software package for complex network
research, InterJournal Complex Systems (2006) 1695.
URL http://igraph.org
[38] S. Tomihisa, K. Kawai, An algorithm for drawing general undirected
graphs, Information Processing Letters 31 (1) (1989) 7 -- 15. doi:10.
1016/0020-0190(89)90102-6.
[39] D. Villatoro, J. Sabater-Mir, S. Sen, Robust convention emergence in social
networks through self-reinforcing structures dissolution, ACM Trans. Au-
ton. Adapt. Syst. 8 (1) (2013) 2:1 -- 2:21. doi:10.1145/2451248.2451250.
42
|
1903.12086 | 2 | 1903 | 2019-10-22T18:35:40 | Towards a Theory of Systems Engineering Processes: A Principal-Agent Model of a One-Shot, Shallow Process | [
"cs.MA"
] | Systems engineering processes coordinate the effort of different individuals to generate a product satisfying certain requirements. As the involved engineers are self-interested agents, the goals at different levels of the systems engineering hierarchy may deviate from the system-level goals which may cause budget and schedule overruns. Therefore, there is a need of a systems engineering theory that accounts for the human behavior in systems design. To this end, the objective of this paper is to develop and analyze a principal-agent model of a one-shot (single iteration), shallow (one level of hierarchy) systems engineering process. We assume that the systems engineer maximizes the expected utility of the system, while the subsystem engineers seek to maximize their expected utilities. Furthermore, the systems engineer is unable to monitor the effort of the subsystem engineer and may not have a complete information about their types or the complexity of the design task. However, the systems engineer can incentivize the subsystem engineers by proposing specific contracts. To obtain an optimal incentive, we pose and solve numerically a bi-level optimization problem. Through extensive simulations, we study the optimal incentives arising from different system-level value functions under various combinations of effort costs, problem-solving skills, and task complexities. | cs.MA | cs | SUBMITTED TO IEEE SYSTEMS JOURNAL
1
Towards a Theory of Systems Engineering
Processes: A Principal-Agent Model of a One-Shot,
Shallow Process
†School of Mechanical Engineering, Purdue University, West Lafayette, Indiana 47907-2088
Salar Safarkhani†, Ilias Bilionis†∗, Jitesh H. Panchal†
∗Corresponding author, email: [email protected]
9
1
0
2
t
c
O
2
2
]
A
M
.
s
c
[
2
v
6
8
0
2
1
.
3
0
9
1
:
v
i
X
r
a
Abstract -- Systems engineering processes coordinate the effort
of different individuals to generate a product satisfying certain
requirements. As the involved engineers are self-interested agents,
the goals at different levels of the systems engineering hierarchy
may deviate from the system-level goals which may cause budget
and schedule overruns. Therefore, there is a need of a systems
engineering theory that accounts for the human behavior in sys-
tems design. Undertaking such an ambitious endeavor is clearly
beyond the scope of any single paper due to the inherent difficulty
of rigorously formulating the problem in its full complexity along
with the lack of empirical data. However, as experience in the
physical sciences shows, a lot of knowledge can be generated by
studying simple hypothetical scenarios which nevertheless retain
some aspects of the original problem. To this end, the objective of
this paper is to study the simplest conceivable systems engineering
process, a principal-agent model of a one-shot (single iteration),
shallow (one level of hierarchy) systems engineering process. We
assume that the systems engineer maximizes the expected utility
of the system, while the subsystem engineers seek to maximize
their expected utilities. Furthermore, the systems engineer is un-
able to monitor the effort of the subsystem engineer and may not
have complete information about their types or the complexity of
the design task. However, the systems engineer can incentivize the
subsystem engineers by proposing specific contracts. To obtain
an optimal incentive, we pose and solve numerically a bi-level
optimization problem. Through extensive simulations, we study
the optimal incentives arising from different system-level value
functions under various combinations of effort costs, problem-
solving skills, and task complexities. Our numerical examples
show that, the passed-down requirements to the agents increase
as the task complexity and uncertainty grow and they decrease
with increasing the agents' costs.
Index Terms -- systems engineering theory, systems science,
complex systems, game theory, principal-agent model, mechanism
design, contract theory, expected utility, bi-level programming
problem, optimal incentives.
I. INTRODUCTION
C OST and schedule overruns plague the majority of large
systems engineering projects across multiple industry
sectors including power [1], defense [2], and space [3]. As
design mistakes are more expensive to correct during the
production and operation phases,
the design phase of the
systems engineering process (SEP) has the largest potential
impact on cost and schedule overruns. Collopy et al. [4] argued
that requirements engineering (RE), which is a fundamental
part of the design phase, is a major source of inefficiencies
in systems engineering. In response, they developed value-
driven design (VDD) [5], a systems design approach that starts
with the identification of a system-level value function and
guides the systems engineer (SE) to construct subsystem value
functions that are aligned with the system goals. According to
VDD, the subsystem engineers (sSE) and contractors should
maximize the objective functions passed down by the SE
instead of trying to meet requirements.
RE and VDD make the assumption that the goals of the
human agents involved in the SEP are aligned with the
SE goals. In particular, RE assumes that, agents attempt to
maximize the probability of meeting the requirements, while
VDD assumes that they will maximize the objective functions
supplied by the SE. However, this assumption ignores the
possibility that the design agents, as all humans, may have
personal agendas that not necessarily aligned with the system-
level goals.
Contrary to RE and VDD, it is more plausible that the
design agents seek to maximize their own objectives. Indeed,
there is experimental evidence that the quality of the outcome
of a design task is strongly affected by the reward antici-
pated by the agent [6 -- 8]. In other words, the agent decides
how much effort and resources to devote to a design task
after taking into account the potential reward. In the field,
the reward could be explicitly implemented as an annual
performance-based bonus, or, as it is the case most often, it
could be implicitly encoded in expectations about job security,
promotion, professional reputation, etc. To capture the human
aspect in SEPs, one possible way is one needs to follow a
game-theoretic approach [9], [10]. Most generally, the SEP
should can be modeled as a dynamical hierarchical network
game with incomplete information. Each layer of the hierarchy
represents interactions among the SE and some sSEs, or
the sSEs and other engineers or contractors. With the term
"principal," we refer to any individual delegating a task, while
we reserve the term "agent" for the individual carrying out the
task. Note that an agent may simultaneously be the principal in
a set of interactions down the network. For example, the sSE is
the agent when considering their interaction with the SE (the
principal), but the principal when considering their interaction
with a contractor (the agent). At each time step, the principals
pass down delegated tasks along with incentives, the agents
choose the effort levels that maximize their expected utility,
perform the task, and return the outcome to the principals.
The iterative and hierarchical nature of real SEPs makes
them extremely difficult to model in their full generality. Given
SUBMITTED TO IEEE SYSTEMS JOURNAL
2
that our aim is to develop a theory of SEPs, we start from the
simplest possible version of a SEP which retains, nevertheless,
some of the important elements of the real process. Specifi-
cally, the objective of this paper is to develop and analyze a
principal-agent model of a one-shot, shallow SEP. The SEP
is "one-shot" in the sense that decisions are made in one
iteration and they are final. The term "shallow" refers to a
one-layer-deep SEP hierarchy, i.e., only the SE (principal)
and the sSEs (agents) are involved. The agents maximize their
expected utility given the incentives provided by the principal,
and the principal selects the incentive structure that maximizes
the expected utility of the system. We pose this mechanism
design problem [11] as a bi-level optimization problem and
we solve it numerically.
A key component of our SEP model is the quality function
of an agent. The quality function is a stochastic process
that models the principal's beliefs about the outcome of the
delegated design task given that the agent devotes a certain
amount of effort. The quality function is affected by what the
principal believes about the task complexity and the problem
solving skills of the agent. Following our work [12], we model
the design task as a maximization problem where the agent
seeks the optimal solution. The principal expresses their prior
beliefs about the task complexity by modeling the objective
function as a random draw from a Gaussian process prior with
a suitably selected covariance function.
As we showed in [12], conditioned on knowing the task
complexity and the agent type, the quality function is well
approximated by an increasing, concave function of effort
with additive Gaussian noise. However, we will use a linear
approximation for the quality function.
We study numerically two different scenarios. The first
scenario assumes that the SE knows the agent types and the
task complexity, but they do not observe the agent's effort.
This situation is known in game theory as a moral hazard
problem [13]. The most common way to solve a moral hazard
problem is to use the first order approach (FOA) [14]. In the
FOA, the incentive compatibility constraint of the agent is
replaced by its first order necessary condition. However, the
FOA depends on the convexity of the distribution function
in effort which is not valid in our case. There have been
several attempts to solve the principal-agent model where the
requirements of the FOA may fail, nonetheless they must still
satisfy the monotone likelihood ratio property [15].
In the second scenario, we study the case of moral haz-
ard with simultaneous adverse selection [16], i.e., the SE
observes neither the effort nor the type of agents nor the
task complexity. This is a Bayesian game with incomplete
information.. In this case, the SE experiences additional loss
in their expected utility, because the sSEs' can pretend to
have different types. The revelation principle [17] guarantees
that it suffices to search for the optimal mechanism within
the set of incentive compatible mechanisms, i.e., within the
set of mechanisms in which the sSEs are telling the truth
about their types and technology maturity. In this paper, we
solve the optimization problem in the principal-agent model,
numerically with making no assumptions about the quality
function.
The paper is organized as follows. In section II we will
derive the mathematical model of the SEP and we will study
the type-independent and type-dependent optimal contracts.
We will also introduce the value and utility functions. In sec-
tion III, we perform an exhaustive numerical study and show
the solutions for several case studies. Finally, we conclude in
section IV.
II. MODELING A ONE-SHOT, SHALLOW SYSTEMS
ENGINEERING PROCESS
A. Basic definitions and notation
As mentioned in the introduction, we develop a model of a
one-shot (the game evolves in one iteration and the decisions
are final), shallow (one-layer-deep hierarchy) SEP. The SE has
decomposed the system into N subsystems and assigned a
sSE to each one of them. We use i = 1, . . . , N to label each
subsystem. From now on, we refer to the SE as the principal
and the sSEs as the agents. The principal delegates tasks to
the agents along with incentives. The agents choose how much
effort to devote on their task by maximizing their expected
utility. The principal, anticipates this reaction and selects the
incentives that maximize the system-level expected utility.
Let (Ω,F, P) be a probability space where, Ω is the sample
space, F is a σ-algebra, and P is the probability measure.
With ω ∈ Ω we refer to the random state of nature. We use
upper case letters for random variables (r.v.), bold upper case
letters for their range, and lower case letters for their possible
values. For example, the type of agent i is a r.v. Θi taking
Mi discrete values θi in the set Θi ≡ Θi(Ω) = {1, . . . , Mi}.
Collectively, we denote all types with the N-dimensional tuple
Θ = (Θ1, . . . , ΘN ) and we reserve Θ−i to refer to the (N−1)-
dimensional tuple containing all elements of Θ except Θi. This
notation carries to any N-dimensional tuple. For example, θ
and θ−i are the type values for all agents and all agents except
i, respectively. The range of Θ is Θ = ×N
The principal believes that the agents types vary indepen-
dently, i.e., they assign a probability mass function (p.m.f.) on
Θ that factorizes over types as follows:
i=1Θi.
N(cid:89)
N(cid:89)
P[Θ = θ] =
P[Θi = θi] =
piθi,
(1)
type k, for k in Θi. Of course, we must have(cid:80)Mj
for all θ in Θ, where pik ≥ 0 is the probability that agent i has
k=1 pik = 1,
i=1
i=1
for all i = 1, . . . , N.
Each agent knows their type, but their state of knowledge
about all other agents is the same as the principal's. That is,
if agent i is of type Θi = θi, then their state of knowledge
about everyone else is captured by the p.m.f.:
P[Θ−i = θ−iΘi = θi] =
P[Θ = θ]
P[Θi = θi]
=
pjθj .
(2)
(cid:89)
j(cid:54)=i
Agent i chooses a normalized effort level ei ∈ [0, 1] for
his delegated task. We assume that this normalized effort is
the percentage of an agent's maximum available effort. The
units of the normalized effort depend on the nature of the
agent's subsystem. If the principal and the agent are both
SUBMITTED TO IEEE SYSTEMS JOURNAL
3
part of same organization then the effort can be the time
that the agent dedicates to the delegated task in a particular
period of time, e.g., in a fiscal year. On the other hand, if
the agent is a contractor, then the effort can be the percentage
of the available yearly budget that the contractor spends on
the assigned task. We represent the monetary cost of the i-th
agent's effort with the random process Ci (ei). In economic
terms, Ci (ei) is the opportunity cost, i.e., the payoff of the best
alternative project in agent could devote their effort. In general,
we know that the process Ci(ei) should be an increasing
function of the effort ei. For simplicity, we assume that the
cost of effort of the agents is quadratic,
Ci(ei) := ciΘie2
i ,
(3)
with a type-dependent coefficient cik > 0 for all k in Θi.
The quality function of the i-th agent
is a real valued
random Qi(ei) := Qi(ei) process paremeterized by the effort
ei. The quality function models everybody's beliefs about
the design capabilities of agent i. The interpretation of the
quality function is as follows. If agent i devotes to the task
an effort of level ei, then they produce a random outcome
of quality Qi(ei). In our previous work [12], we created a
stochastic model for the quality function of a designer where
we explicitly captured its dependence on the problem-solving
skills of the designer and on the task complexity. In that
work, we showed that Qi(ei) has increasing and concave
sample paths, that its mean function is increasing concave,
and the standard deviation is decreasing with effort, albeit
mildly,
is independent of the problem-solving skills of
the designer, and it only increases mildly with increasing
task complexity. Examining the spectral decomposition of the
process for various cases, we observed that it can be well-
approximated by:
it
these independence assumptions, the state of knowledge of the
principal is captured by the following probability measure:
(cid:3) =
(cid:20)
N(cid:89)
i=1
(cid:90)
Bi
(cid:21)
i=1Bi
piθi
φ(ξi)dξi
,
(6)
P(cid:2)Θ = θ, Ξ ∈ ×N
for all θ ∈ Θ and all Borel-measurable Bi ⊂ R. Assuming
that all these are common knowledge, the state of knowledge
of agent i after they observe their type θi (but before they
observe Ξi) is
P(cid:2)Θ−i = θ−i, ξ ∈ ×N
= P[Θ−i = θ−iΘi = θi](cid:81)N
P[Θ=θ,ξ∈×N
P[Θi=θi]
=
i=1Bi
i=1Bi]
(cid:3)
(cid:12)(cid:12)Θi = θi
(cid:104)(cid:82)
(cid:105)
(7)
i=1
Bi
φ(ξi)dξi
Finally, we use E[·] to denote the expectation of any quantity
over the state of knowledge of the principal as characterized
by the probability measure of Eq. (6). That is, the expectation
of any function f (Θ, Ξ) of the agent types Θ and the state of
nature Ξ is
E[f (Θ, Ξ)] =
f (θ, ξ)
[piθiφ(ξi)] dξ.
(8)
Similarly, we use the notation Eiθi [·] to denote the conditional
expectation over the state of knowledge of an agent i who
knows that their type is Θi = θi. This is the expectation
E[·Θi = θi] with respect to the probability measure of Eq. (2)
and we have:
(cid:90)
(cid:88)
θ∈Θ
RN
N(cid:89)
i=1
Eik[f (Θ, Ξ)] =
f (θi, θ−i, ξ)
(cid:88)
(cid:90)
θ−i∈Θ−i
RN
(cid:81)N
j=1
(cid:2)pjθj φ(ξj)(cid:3)
piθi
dξ.
(9)
Qi(ei) = q0
iΘi
(ei) + σiΘiΞi,
(4)
where, for k in Θi, q0
ik(ei) is an increasing, concave, type-
dependent mean quality function, σik > 0 is a type-dependent
standard deviation parameter capturing the aleatory uncertainty
of the design process, and Ξi is a standard normal r.v. If we
further assume that the time window for design is relatively
small, then the q0
ik(ei) term can be approximated as a linear
function. Therefore, we will assume that the quality function
is:
Qi(ei) = κiΘiei + σiΘiΞi,
(5)
where, κ is inversely proportional to the complexity of the
problem. For instance, a large κ corresponds to a low-
complexity task while a κ corresponds to a high-complexity
task. The standard deviation parameter σ captures the inherent
uncertainty of the design process and depends on the maturity
of the underlying technology. In summary, an agent's type is
characterized by the triplet cost-complexity-uncertainty.
From the perspective of the principal,
the r.v.'s Ξi are
independent of the agents' types Θi as they represent the
uncertain state of nature. A stronger assumption that we
employ is that the Ξi's are also independent to each other.
This assumption is strong because it essentially means that
the qualities of the various subsystems are decoupled. Under
B. Type-independent optimal contracts
We start by considering the case where the principal offers
a single take-it-or-leave-it contract independent of the agent
type. This is the situation usually encountered in contractual
relationships between the SE and the sSEs within the same
organization. The principal offers the contract and the agent
decides whether or not to accept it. If the agent accepts, then
they select their level of effort by maximizing their expected
utility, they work on their design task, they return the outcome
quality back to the principal, and they receive their reward.
We show a schematic view of this type of contracts in Fig.
1a. A contract is a monetary transfer function ti : R → R
that specifies the agent's compensation ti(qi) contingent on
the quality level qi. Therefore, the payoff of the i-th agent is
the random process:
Πi(ei) = ti (Qi (ei)) − Ci(ei).
(10)
We assume that the agent knows their type, but they choose
the optimal effort level ex-ante, i.e., they choose the effort
level before seeing the state of the nature Ξi. Denoting their
monetary utility function by Ui(πi) = uiΘi(πi), the i-th agent
selects an effort level by solving:
∗
θi
ei
= arg max
ei∈[0,1]
Eiθi [Ui (Πi(ei))] .
(11)
SUBMITTED TO IEEE SYSTEMS JOURNAL
4
Let Q∗
i be the r.v. representing the quality function that the
principal should expect from agent i if they act optimally, i.e.,
Q∗
i = Qi(e∗
iΘi
).
(12)
Then the system level value is a r.v. of the form
V = v(Q∗),
(13)
where v : RN → R is a function of the subsystem outcomes
Q∗. We introduce the form of the value function, v(q), in Sec.
III. Note that, even though in this work the r.v. V is assumed
to be just a function of Q∗, in reality it may also depend
on the random state of nature, e.g., future prices, demand for
the system services. Consideration of the latter is problem-
dependent and beyond the scope of this work.
Given the system value V and taking into account
transfers to the agents, the system-level payoff is the r.v.
Π0 = V − N(cid:88)
ti(Q∗
i ).
the
(14)
i=1
If the monetary utility of the principal is u0(π0), then they
should select the transfer functions t(·) = (t1(·), . . . , tN (·))
by solving:
t∗ (·) = arg max
t(·)
E [u0 (Π0)] .
(15)
However, guarantee that they want to participate in the SEP,
the expected utility of the sSEs must be greater than the
expected utility they would enjoy if they participated in
another project. Therefore, the SE must solve Eq. (15) subject
to the participation constraints:
Eiθi [Ui (Πi)] ≥ ¯uiθi ,
(16)
for all possible values of θi, and all i = 1, . . . , N, where ¯uiθi
is known as the reservation utility of agent i.
C. Type-depdenent optimal contracts
By offering a single transfer function, the principal is unable
to differentiate between the various agent types when adverse
selection is an issue. That is, all agent types, independently of
their cost, complexity, and uncertainty attributes, exactly the
same transfer function. In other words, with a single transfer
function the principal is actually targeting the average agent.
This necessarily leads to inefficiencies stemming from prob-
lems such as paying an agent involved in a low-complexity
task more than a same cost and uncertainty agent involved in
a high-complexity task.
The principal can gain in efficiency by offering different
transfer functions (if any exist) that target specific agent types.
For example, the principal could offer a transfer function that
is suitable for cost-efficient, low-complexity, low-uncertainty
agents, and one for cost-inefficient agents, low-complexity,
low-uncertainty, etc., for any other combination that is sup-
ported by the principal's prior knowledge about the types of
the agent population. To implement this strategy the princi-
pal can employ the following extension to the mechanism
of Sec. II-B. Prior to initiating work, the agents announce
their types to the principal and they receive a contract that
matches the announced type. In Fig. 1b, we show how this
type of contract evolves in time. Let us formulate this idea
mathematically. The i-th agent announces a type θ(cid:48)
i in Θi (not
necessarily the same as their true type θi), and they receive the
(·). The payoff
associated, type-specific, transfer function tiθ(cid:48)
to agent i is now:
i
Πi(ei, θ(cid:48)
i) = tiθ(cid:48)
i
(Qi(ei)) − Ci(ei),
(17)
where all other quantities are like before. Given the announce-
ment of a type θ(cid:48)
i, the rational thing to do for agent i is to
i (θi, θ(cid:48)
select a level of e∗
i) by maximizing their expected utility,
i.e., by solving:
e∗
iθiθ(cid:48)
Eiθi[Ui(Πi(ei, θ(cid:48)
(18)
i))].
= arg max
ei∈[0,1]
i
Of course, the announcement of θ(cid:48)
i is also a matter of choice
and a rational agent should select also by maximizing their
expected utility. The obvious issue here is that agents can lie
about their type. For example, a cost-efficient agent (agent with
low cost of effort) may pretend to be a cost-inefficient agent
(agent with high cost of effort). Fortunately, the revelation
principle [17] comes to the rescue and simplifies the situation.
It guarantees that, among the optimal mechanisms, there is
one that is incentive compatible. Thus it will be sufficient if
the principal constraints their contracts to over truth-telling
mechanisms. Mathematically, to enforce truth-telling, the SE
must satisfy the incentive compatibility constraints:
, θ(cid:48)
, θi))] ≥ Eiθi [Ui(Πi(e∗
Eiθi[Ui(Πi(e∗
(19)
i))],
iΘiθi
iΘiθ(cid:48)
i
for all θi (cid:54)= θ(cid:48)
i in Θi. Eq. (19) expresses mathematically that
"the expected payoff of agent i when they are telling the truth
is always greater than or equal to the expected payoff they
would enjoy if they lied."
Similar to the developments of Sec. II-B, the quality that
the SE expects to receive is:
Q∗
i = Qi(e∗
iΘiΘi
),
(20)
where we use the fact
compatible. The payoff of the SE becomes:
that
the mechanism is incentive-
Π0 = V − N(cid:88)
tiΘi(Q∗
i ).
(21)
i=1
Therefore, to select the optimal transfer functions, the SE must
solve:
E [u0(Π0)] ,
max
t(·,·)
(22)
subject to the incentive compatibility constraints of Eq. (19),
and the participation constraints:
Eiθi[Ui(Πi(e∗
(23)
for all θi ∈ Θi, where we also assume that the incentive
compatibility constrains hold.
iΘiΘi
, Θi))] ≥ ¯uiθi,
SUBMITTED TO IEEE SYSTEMS JOURNAL
5
(a)
(b)
Fig. 1: (a): Timing of the contract for type-independent con-
tracts. (b): Timing of the contract for type-dependent contracts.
D. Parameterization of the transfer functions
Transfer functions must be practically implementable. That
is, they must be easily understood by the agent when expressed
in the form of a contract. To be easily implementable, transfer
functions should be easy to convey in the form of a table.
To achieve this, we restrict our attention to functions that
are made out of constants, step functions, linear functions,
or combinations of these.
Despite the fact that including such functions would likely
enhance the principal's payoff, we exclude transfer functions
that encode penalties for poor agent performance, i.e., transfer
functions that can take negative values. First, contracts with
penalties may not be implementable if the principal and the
agent reside within the same organization. Second, even when
the agent is an external contractor penalties are not commonly
encountered in practice. In particular, if the SE is a sensitive
government office, e.g., the department of defense, national
security may dictate that the contractors should be protected
from bankruptcy. Third, we do not expect our theory to be
empirically valid when penalties are included since, according
to prospect theory [18], humans perceive losses differently.
They are risk-seeking when the reference point starts at a loss
and risk-averse when the reference point starts at a gain.
To overcome these issues we restrict our attention to transfer
functions that include three simple additive terms: a constant
term representing a participation payment, i.e., a payment
received for accepting to be part of the project; a constant
payment that is activated when a requirement is met; and a
linear increasing part activated after meeting the requirement.
The role of the latter two part is to incentivize the agent to
meet and exceed the requirements.
We now describe this parameterization mathematically. The
transfer function associated with type k in Θi of agent i is
parameterized by:
aik,0 + aik,1 H (qi − aik,2)
tik (qi) =
+aik,3 (qi − aik,2) H (qi − aik,2) ,
(24)
where H is the Heaviside function (H(x) = 1 if x ≥ 0
and 0 otherwise), and all the parameters aik,0, . . . , aik,3 are
non-negative. In Eq. (24), aik,0 is the participation reward,
aik,1 is the award for exceeding the passed-down requirement,
aik,2 is the passed-down requirement, and aik,3 the payoff
per unit quality exceeding the passed-down requirement. We
will call these form of transfer functions the "requirement
based plus incentive" (RPI) transfer function. In case the
aik,3 = 0, we call it the "requirement based" (RB) transfer
function. At this point, it is worth mentioning that the passed-
down requirement aik,2 is not necessarily the same as the
true system requirement ri, see our reults in Sec. III. As we
have shown in earlier work [10], the optimal passed-down
requirement differs from the true system requirement. For
example, the SE should ask for higher requirements for the
design task with low-complexity. On the other hand, for the
task with high-complexity, the SE should pass down less than
the actual requirement. For notational convenience, we denote
by aik ∈ R4
+ (R = {x ∈ R : x ≥ 0}) the transfer parameters
pertaining to agent i of type k ∈ Θi, i.e.,
aik = (aik,0, . . . , aik,3) .
(25)
Similarly, with ai ∈ R4Mi
pertaining to agent i for all types, i.e.,
+ we denote the transfer parameters
ai = (ai1, . . . , aiMi) ,
(26)
and with a ∈ R4(cid:80)N
+
tively, i.e.,
i=1 Mi
all the transfer parameters collec-
a = (a1, . . . , aN ) .
(27)
E. Numerical solution of the optimal contract problem
The optimal contract problem is a an intractable bi-level,
non-linear programming problem.In particular, the SE's prob-
lem is for the case of type-dependent contracts is to maximize
the expected system-level utility over the class of imple-
mentable contracts, i.e.,
E [u0(Π0)] ,
max
a
subject to
1) contract implementability constraints:
aik,j ≥ 0,
(28)
(29)
for all i = 1, . . . , N, k = 1, . . . , Mi, j = 0, . . . , 3;
2) individual rationality constraints:
e∗
ikl = arg max
ei∈[0,1]
Eik[Ui(Πi(ei, l))],
(30)
for all i = 1, . . . , N, k = 1, . . . , Mi, l = 1, . . . , Mi;
3) participation constraints:
Eik[Ui(Πi(e∗
ikk, k))] ≥ ¯uik,
for all i = 1, . . . , N, k = 1, . . . , Mi; and
4) incentive compatibility constraints:
Eik[Ui(Πi(e∗
ikk, k))] ≥ Eik[Ui(Πi(e∗
ikl, l))],
for all i = 1, . . . , N and k (cid:54)= l in {1, . . . , Mi}.
(31)
(32)
For the case of type-independent contracts, one adds the
constraint aik = ail for all i = 1, . . . , N and k (cid:54)= l in
{1, . . . , Mi} and the incentive compatibility constraints are
removed.
SUBMITTED TO IEEE SYSTEMS JOURNAL
6
Nc(cid:88)
A common approach to solving bi-level programming prob-
lems is to replace the internal optimization with the cor-
responding Karush-Kuhn-Tucker (KKT) condition. This ap-
proach is used when the internal problem is concave, i.e., when
it has a unique maximum. However, in our case, concavity
is not guaranteed, and we resort to nested optimization. We
implement everything in Python using the Theano [19] sym-
bolic computation package exploit automatic differentiation.
We solve the follower problem using sequential least squares
programming (SLSQP) as implemented in the scipy package.
We use simulated annealing to find the global optimum point
of the leader problem. We first convert the constraint problem
to the unconstrained problem using the penalty method such
that:
f (a) = E [u0(Π0)] +
(33)
where gi (·)'s are the constraints in Eqs. (29-32). Maximizing
the f (a) in Eq. 33, is equivalent to finding the mode of the
distribution:
min (gi (a) , 0) ,
i=1
πγ (a) ∝ exp (γf (a)) ,
(34)
we use Sequential Monte Carlo (SMC) [20] method to sample
from this distribution by increasing γ from 0.001 to 50. To
perform the SMC, we use the "pysmc" package [21]. To ensure
the computational efficiency of our approach, we need to use a
numerical quadrature rule to approximate the expectation over
Ξ. This step is discussed in Appendix A. To guarantee the
reproducibility of our results, we have published our code in
an open source Github repository (https://github.com/ebilionis/
incentives) with an MIT license.
F. Value Function and Risk Behavior
We assume two types of value functions, namely,
the
requirement based (RB) and requirement based plus incentive
(RPI). Mathematically, we define these two value functions as:
N(cid:89)
VRB := v0
{H(Q∗
i − 1)} .
and,
VRPI := v0
N(cid:89)
i=1
i=1
{H(Q∗
i − 1)} [1 + 0.2(Q∗
i − 1)] ,
Fig. 2: The RB value function (black solid line) and RPI value
function (green dashed line).
Fig. 3: The utility functions for risk averse (RA) (black solid
line), risk neutral (RN) (green dashed line).
III. NUMERICAL EXAMPLES
In this section, we start by performing an exhaustive numer-
ical investigation of the effects of task complexity, agent's cost
of effort, uncertainty in the quality of the returned task, and
adverse selection. In Sec. III-A1, we study the "moral hazard
only" scenario with the RB transfer and value functions. In
Sec. III-A2, we study the effect of the RPI transfer and value
functions. We study the "moral hazard with adverse selection"
in Sec. III-A3.
A. Numerical investigation of the proposed model
In these numerical investigations we consider a single risk
neutral principal and a risk averse agent. Each case study
corresponds to a choice of task complexity (κ in Eq. (5)),
cost of effort (c in Eq. (3)), and performance uncertainty (σ in
Eq. (5)). With regards to task complexity, we select κ = 2.5
for an easy task and κ = 1.5 for a hard task. For the cost of
effort parameter, we associate c = 0.1 and c = 0.4 with the
(35)
(36)
(37)
respectively. In Fig. 2, we show these two value functions for
one subsystem.
We consider two different risk behaviors for individuals, risk
averse (RA) and risk neutral (RN). We use the utility function
in Eq. (37), for the risk behavior of the agents and principal,
(cid:40)
u (π (·)) =
a − be−cπ(·), for RA
π (·) , for RN,
where c = 2 for a RA agent. The parameters a and b are:
a = b =
1
1 − e−c .
We show these utility functions for the two different risk
behaviors in Fig. 3.
SUBMITTED TO IEEE SYSTEMS JOURNAL
7
low- and high-cost agents, respectively. Finally, low- and high-
uncertainty tasks are characterized by σ = 0.1 and σ = 0.4,
respectively.
Note that, the parameters κiθi, ciθi, and σiθi have two
indices. The first index i is the agent's (subsystem's) number
and the second index is the type of the agent. We begin
with a series of cases with a single agent with a known
type denoted by 1 (moral-hazard-only case studies). In these
cases, the parameters corresponding to complexity, cost and
uncertainty are denoted by κ11, c11, and σ11, respectively.
We end with a series of cases with a single agent but
with an unknown type that can take two discrete, equally
probable values 1 and 2 (moral-hazard-and-adverse-selection
case studies). Consequently, κ11 denotes the effort coefficient
of a type-1 agent 1, κ12 the same for a type-2 agent, and so
on for all the other parameters.
To avoid numerical difficulties and singularities, we replace
all Heaviside functions with a sigmoids, i.e.,
Hα(x) =
1
1 + e−αx ,
(38)
where the parameter α controls the slope. We choose α = 50
for the transfer functions and α = 100 for the value function.
We consider two types of value functions, RB and RPI value
functions, see Sec. II-F. For the RB value function we use
the transfer function of Eq. (24) constrained so aik, 3 = 0
(RB transfer function). In other words, the agent is paid a
constant amount if they achieve the requirement and there is
no payment per quality exceeding the requirement. For the
case of RPI value function, we remove this constraint.
1) Moral hazard with RB transfer and value functions:
Consider the case of a single risk-averse agent of known type
and a risk-neutral principal with an RB value function. In
Fig. 4a, we show the transfer functions for several agent types
covering all possible combinations of low/high complexity,
low/high cost, and low/high task uncertainty. Fig. 4b depicts
the probability that the principal's expected utility exceeds a
given threshold for all these combinations. We refer to this
curve as the exceedance curve. Finally, in tables I and II, we
report the expected utility of the principal for the low and high
cost agents, respectively. We make the following observations:
1) For the same level of task complexity and uncertainty,
but with increasing cost of effort:
a) the optimal passed-down requirement decreases;
b) the optimal payment for achieving the requirement
increases;
c) the principal's expected utility decreases; and
d) the exceedance curve shifts to the left.
Intuitively, as the agent's cost of effort increases, the
principal must make the contract more attractive to
ensure that the participation constraints are satisfied.
As a consequence, the probability that the principal's
expected utility exceeds a given threshold decreases.
2) For the same level of task uncertainty and cost of effort,
but with increasing complexity:
a) the optimal passed-down requirement decreases;
b) the optimal payment for achieving the requirement
increases;
c) the principal's expected utility decreases; and
d) the exceedance curve shifts to the left.
Thus, we see that an increase in task complexity has the
same a similar effect as an increase in the agent's cost
of effort. As in the previous case, to make sure that the
agent wants to participate, the principal has to make the
contract more attractive as task complexity increases.
3) For the same level of task complexity and cost of effort,
but with increasing uncertainty:
a) the optimal passed-down requirement increases;
b) the optimal payment for achieving the requirement
c) the principals expected utility decreases;
d) the exceedance curve shifts towards the bottom
increases;
right.
This case is the most interesting. Here as the uncertainty
of the task increases, the principal must increase the
passed-down requirement to ensure that they are hedged
against failure. At the same time, however, they must
also increase the payment to ensure that the agent still
has an incentive to participate.
4) For all cases considered, the optimal passed down re-
quirement is greater than the true requirement (which is
set to one). Note, however, this is not universally true.
Our study does not examine all possible combinations of
cost, quality, and utility functions that could have been
considered. Indeed, as we showed in our previous work
[10], there are situations in which a smaller-than-the-true
requirement can be optimal.
TABLE I: The expected utility of the principal for low cost
agent with RB value function.
Low Uncertainty
High Uncertainty
Low Complexity
High Complexity
0.97
0.92
0.93
0.79
TABLE II: The expected utility of the principal for high cost
agent with RB value function.
Low Uncertainty
High Uncertainty
Low Complexity
High Complexity
0.89
0.72
0.77
0.45
2) Moral hazard with RPI transfer and value functions:
This case is identical to Sec. III-A1, albeit we use the RPI
value function, see Sec. II-F, and the RPI transfer function,
see Eq. (24). Fig 5a, depicts the transfer functions for all
combinations of agent types and task complexities. In Fig.
5b, we show the exceedance curve using the RPI value and
transfer functions. Finally, in tables III and IV, we report the
expected utility of the principal using the RPI transfer and
value functions for the low and high cost agents, respectively.
The results are qualitative similar to Sec. III-A1, with the
additional observations:
1) For the same level of task complexity, uncertainty and
agent cost, the optimal reward for achieving the re-
quirement decreases compared to the same cases in
SUBMITTED TO IEEE SYSTEMS JOURNAL
8
(a) Transfer functions using RB value function.
(b) Exceedance curve.
Fig. 4: L and H stand for low and high, respectively, Comp. and Unc. stand for complexity and uncertainty, respectively. The
low and high complexity denote the κ11 = 2.5 and κ11 = 1.5, respectively, low and high cost denote c11 = 0.1 and c11 = 0.4,
respectively, low and high uncertainty denote σ11 = 0.1 and σ11 = 0.4, respectively, RA denotes the risk averse agent. (a):
The RB transfer functions for several different agent types with respect to outcome of the subsystem (Q1) for moral hazard
scenario. (b): The exceedance for the moral hazard scenario using the RB transfer function.
Sec. III-A1. Intuitively, as the principal has the option
to reward the agent based on the quality exceeding the
requirement, they prefer to pay less for fulfilling the
requirement. Instead, the principal incentivizes the agent
to improve the quality beyond the optimal passed-down
requirement.
2) The slope of the transfer function beyond the passed-
down requirement is almost identical to the slope of the
value function.
TABLE III: The expected utility of the principal for low cost
agent with RPI value function.
Low Uncertainty
High Uncertainty
Low Complexity
High Complexity
1.2
1.0
1.2
0.89
TABLE IV: The expected utility of the principal for high cost
agent with RPI value function.
Low Uncertainty
High Uncertainty
Low Complexity
High Complexity
0.95
0.76
0.93
0.56
In table V, we summarize our observations for the results
in Sec.
III-A1 and III-A2. In this table, we show how the
passed-down requirement and payment change when we fix
two parameters of the model (we denote it by "fix" in the
table) and vary the third parameter. We denote increase by ↑
and decrease by ↓.
3) Moral hazard with adverse selection: Consider the case
of a single risk-averse agent of unknown type which takes two
possible values, and a risk-neutral principal with a RB value
function. We consider two possibilities for the unknown type:
TABLE V: Summary of the observations.
complexity
↑
fix
fix
agent cost
uncertainty
fix
↑
fix
fix
fix
↑
requirement
↓
↓
↑
payment
↑
↑
↑
1) Unknown cost of effort. Here, we set κ11 = κ12 = 1.5
(p(κ11 = 1.5) = 1), σ11 = σ12 = 0.1 (p(σ11 = 0.1)),
and p(c11 = 0.1) = 0.5 and p(c12 = 0.4) = 0.5
2) Unknown task complexity. For the unknown quality we
assume that p(κ11 = 2.5) = 0.5 and p(κ12 = 1.5) =
0.5, σ11 = σ12 = 0.4 (p(σ11 = 0.4) = 1), and c11 =
c12 = 0.4 (p(c11 = 0.4) = 1).
In this scenario, we maximize the expected utility of the
principal subject to constraints in Eqs. (29-32). The incentive
compatibility constraint, Eq. (32), guarantees that the agent
will choose the contract that is suitable for their true type. In
other words, as there are two agent types' possibilities, the
principal must offer two contracts, see Fig. 1b. These two
contracts must be designed in a way that there is no benefit
for the agent to deviate from their true type, i.e., the contracts
enforce the agent to be truth telling.
Solving the constraint optimization problem yields:
a11 = a12 = (0, 0.29, 1.06),
i.e., the two contracts collapse into one. Note that the resulting
contract is the same as the pure moral hazard case, Sec. III-A1,
for an agent with type κ11 = 1.5, σ11 = 0.1, and c11 =
0.4. In other words, the principal must behave as if there was
only a high-cost agent. That is, there are no contacts that can
differentiate between a low- and a high-cost agent in this case.
A similar outcome occurs for unknown task complexity.
The solution of the constraint optimization problem for this
SUBMITTED TO IEEE SYSTEMS JOURNAL
9
(a) Transfer functions using RPI value function.
(b) Exceedance curve.
Fig. 5: L and H stand for low and high, respectively, Comp. and Unc. stand for complexity and uncertainty, respectively. The
low and high complexity denote the κ11 = 2.5 and κ11 = 1.5, respectively, low and high cost denote c11 = 0.1 and c11 = 0.4,
respectively, low and high uncertainty denote σ11 = 0.1 and σ11 = 0.4, respectively, RA denotes the risk averse agent. (a):
The RPI transfer functions for several different agent types with respect to outcome of the subsystem (Q1) for moral hazard
scenario. (b): The exceedance curve for the moral hazard scenario using the RPI transfer function.
scenario is:
a11 = a12 = (0, 0.08, 1.11),
which is the same as the optimum contract that is offered for
the pure moral hazard case, Sec. III-A1, for an agent with
type κ11 = 1.5, σ11 = 0.4, and c11 = 0.4. Therefore, in this
case the principal must behave as if there the task is of high
complexity.
Note that in both cases above, the collapse of the two
contracts to one contract is not a generalizable property of our
model. In particular, it may not happen if more flexible trans-
fer functions are allowed, e.g., ones that allow performance
penalties.
In Fig. 6, we show the transfer functions for the adverse
selection scenarios with unknown cost and unknown quality.
In tables VI and VII, we show the expected utility of two
types of agents and the principal using the optimum contract
for unknown cost and unknown quality, respectively. To sum
up:
1) The unknown cost:
a) the optimum transfer function for this problem is
as same as that the principal would have offered for
a single-type high-cost agent with c11 = c12 = 0.4
(moral hazard scenario with no adverse selection);
b) the expected utility of the low cost agent (efficient
agent) is greater than that of the high cost agent.
In this case,
the low-cost agent benefits because of
information asymmetry. In other words, the principal
must pay an information rent to the low-cost agent to
reveal their type.
2) The unknown task complexity:
a) the optimum contract in this case is the contract
that the principal would have offered for the single-
type high-complexity task with κ11 = κ12 = 1.5;
b) the expected utility of an agent dealing with a low-
complexity task is greater than that of an agent
dealing with a high-complexity task.
Again, due to the information asymmetry,
the agent
benefits if the task complexity is low. The principal must
pay an information rent to reveal the task complexity.
Fig. 6: The transfer function for the adverse selection scenarios
with unknown cost (solid line) and unknown quality (dashed
line), the agent is risk averse. For unknown cost: κ11 = κ12 =
1.5 with probability 1, σ11 = σ12 = 0.1 with probability 1, and
c11 = 0.1 with probability 0.5 and c12 = 0.4 with probability
0.5. For unknown quality: κ11 = 2.5 with probability 0.5
and κ12 = 1.5 with probability 0.5, σ11 = σ12 = 0.4 with
probability 1, and c11 = c12 = 0.4 with probability 1.
SUBMITTED TO IEEE SYSTEMS JOURNAL
10
TABLE VI: The expected utility of the agent with unknown
cost for two different contracts.
Low Cost Agent (Type 1)
High Cost Agent (Type 2)
E[u1(·)]
0.39
0
E[u0(·)]
0.72
0.72
TABLE VII: The expected utility of the agent with unknown
quality for two different contracts.
Low Complexity (Type 1)
High Complexity (Type 2)
E[u1(·)]
0.52
0
E[u0(·)]
0.45
0.45
B. Satellite Design
In this section we apply our method on a simplified satellite
design. Typically a satellite consists of seven different subsys-
tems [22], namely, electrical power subsystem, propulsion, at-
titude determination and control, on-board processing, teleme-
try, tracking and command, structures and thermal subsystems.
We focus our attention on the propulsion subsystem (N = 1).
To simplify the analysis, we assume that the design of these
subsystems is assigned to a sSE in a one-shot fashion. Note
that, the actual systems engineering process of the satellite
design is an iterative process and the information and results
are exchanged back and forth in each iteration. Our model
is a crude approximation of reality. The goal of the SE is to
optimally incentivize the sSE to produce subsystem designs
that meet the mission's requirements. Furthermore, we assume
that the propulsion subsystem is decoupled from the other
subsystems, i.e., there is no interactions between them, and
that the SE knows the types of each sSE and therefore, there
is no information asymmetry.
To extract the parameters of the model, i.e., a11, σ11, c11,
we will use available historical data. To this end,
let I1
be the cumulative, sector-wide investment on the propulsion
subsystem and G1 be the delivered specific impulse of solid
propellants (Isp). The specific impulse is defined as the ratio
of thrust to weight flow rate of the propellant and is a measure
of energy content of the propellants [22].
Historical data, say D1 = {(I1,i, G1,i)}S
i=1, of these quanti-
ties are readily available for many technologies. Of course,
cumulative investment and best performance increase with
time, i.e., I1,i ≤ I1,i+1 and G1,i ≤ G1,i+1. We model the
relationship between G1 and I1 as:
G1 = G1,S + A1(I1 − I1,S) + Σ1Ξ1,
(39)
where G1,S and I1,S are the current states of these variables,
Ξ1 ∼ N (0, 1), and A1 and Σ1 are parameters to be estimated
from the all available data, D1. We use a maximum likelihood
estimator for A1 and Σ1. This is equivalent to a least squares
estimate for Ai:
A1 = arg min
A1
[G1,S + A1(I1,i − I1,S) − G1,i]2 ,
(40)
(cid:105)2
S(cid:88)
(cid:104)
S(cid:88)
i=1
i=1
and to setting Σ1 equal to the mean residual square error:
Σ1 =
1
S
G1,S + A1(I1,i − I1,S) − G1,i
.
(41)
Now, let Gr
1 be the required quality for the propulsion
subsystem in physical units. The scaled quality of a subsystem
Q1, can be defined as:
Q1 =
G1 − G1,S
1 − G1,S
Gr
,
(42)
with this definition, we get Q1 = 0 for the state-of-the-art, and
Q1 = 1 for the requirement. Substituting Eq. (39) in Eq. (42)
and using the maximum likelihood estimates for A1 and Σ1,
we obtain:
Q1 =
A1
1 − G1,S
Gr
(I1 − I1,S) +
Σ1
1 − G1,S
Gr
Ξ1.
(43)
From this equation, we can identify the uncertainty σ11 in the
quality function as:
σ11 =
Σ1
1 − G1,S
Gr
.
(44)
Finally, we need to define effort. Let T1 represents the time
for which the propulsion engineer is to be hired. The cost
of the agent per unit time is C1. T1 is just the duration of
the systems engineering process we consider. The value C1T1
can be read from the balance sheets of publicly traded firms
related to the technology. We can associate the effort variable
e1 with the additional investment required to buy the time of
one engineer:
e1 =
I1 − I1,S
C1T1
,
(45)
that is, e1 = 1 corresponds to the effort of one engineer for
time T1. Let us assume there are Z engineers work on the
subsystem. Comparing this equation, Eq. (43), and Eq. (5),
we get that the κ11 coefficient is given by:
κ11 =
ZC1T1 A1
1 − G1,S
Gr
.
(46)
To complete the picture, we need to talk about the value V0
(in USD) of the system if the requirements are met. We can
use this value to normalize all dollar quantities. That is, we
set:
v0 = 1,
(47)
and for the cost per square effort of the agent we set:
c11 =
ZC1T1
V0
.
(48)
Finally, we use some real data to fix some of the parameters.
Trends in delivered Isp (G1 (sec.)) and investments by NASA
(I1 (millions USD)) in chemical propulsion technology with
time are obtained from [23] and [24], respectively. The state-
of-the-art solid propellant technology corresponds to a G1,S
value of 252 sec. and I1,S value of 149.1 million USD. The
maximum likelihood fit of the parameters results in a regres-
sion coefficient of A1 = 0.0133 sec. per million USD, and
standard deviation Σ1 = 0.12 sec. The corresponding data and
the maximum likelihood fit are illustrated in Fig. 7. The value
of C1 is the median salary (per time) of a propulsion engineer
which is approximately 120,000 USD / year, according to
the data obtained from [25]. For simplicity, also assume that
T1 = 1 year. Moreover, we assume that there are 200 engineers
SUBMITTED TO IEEE SYSTEMS JOURNAL
11
work on the subsystem, Z = 200. We will examine two case
studies which is summarized in table VIII.
Fig. 7: Satellite case study (propulsion subsystem): Histor-
ical data (1979 -- 1988) of specific impulse of solid mono-
propellants vs cumulative investment per firm. The solid line
and the shaded area correspond to the maximum likelihood
fit of a linear regression model and the corresponding 95%
prediction intervals, respectively.
TABLE VIII: The model parameters for two case studies.
Gr
1
252.2 s
252.25 s
V0
50,000,000 USD
60,000,000 USD
κ11
1.6
1.28
σ11
0.6
0.48
c11
0.5
0.4
Using RB value function, we depict the contracts for these
two scenarios in Fig. 8.
Fig. 8: The contracts for two case studies in satellite design.
IV. CONCLUSIONS
We developed a game-theoretic model for a one-shot shal-
low SEP. We posed and solved the problem of identifying the
contract (transfer function) that maximizes the principal's ex-
pected utility. Our results show that, the optimum passed-down
requirement is different from the real system requirement.
For the same level of task complexity and uncertainty, as the
agent cost of effort increases, the passed-down requirement de-
creases and the award to achieving the requirement increases.
In this way, the principal makes the contract more attractive to
the high-cost agent and ensures that the participation constraint
is satisfied. Similarly, for the same level of task uncertainty
and cost of effort, increasing task complexity results in lower
passed-down requirement and larger award for achieving the
requirement. For the same level of task complexity and cost
of effort, as the uncertainty increases both the passed-down
requirement and the award for achieving the requirement
increase. This is because the principal wants to make sure that
the system requirements are achieved. Moreover, by increasing
the task complexity, the task uncertainty, or the cost of effort,
the principal earns less and the exceedance curve is shifted to
the left. Using the RPI contracts, the principal pays smaller
amount for achieving the requirement but, instead, they pay
for per quality exceeding the requirement.
For the adverse selection scenario with RB value function,
we observe that when the principal is maximally uncertain
about the cost of the agent, the optimum contracts are equiv-
alent to the contract designed for the high cost agent in the
single-type case with no adverse selection. The low-cost agent
earns more expected utility than the high-cost agent. This is
the information rent that the principal must pay to reveal the
agents' types. Similarly, if the principal is maximally uncertain
about the task complexity, the two optimum contracts for the
unknown quality are equivalent to the contract that is offered to
the high-complexity task where there is no adverse selection.
Note that, the equivalence of the contracts in adverse selection
scenario with the contract that is offered in absence of adverse
selection is not universal. If the class of possible contracts is
enlarged, e.g., to allow penalties, there may be a set of two
contracts that differentiate types.
There are still many remaining questions in modeling SEPs
using a game-theoretic approach. First, there is a need to study
the hierarchical nature of SEPs with potentially coupled sub-
systems. Second, true SEPs are dynamic in nature with many
iterations corresponding to exchange of information between
the various agents. These are the topics of ongoing research
towards a theoretical foundation of systems engineering design
that accounts for human behavior.
APPENDIX A
NUMERICAL ESTIMATION OF THE REQUIRED
EXPECTATIONS
For the numerical implementation of the suggested model,
we need to be able to carry out expectations of the form
of Eq. 9 a.k.a. Eq. 8 and Eq. 7. Since, we have at most
two possible types in our case studies, the summation over
the possible types is trivial. Focusing on expectations over
Ξ, we evaluate them using a sparse grid quadrature rule
[26]. In particular, any expectation of the form E[g(Ξ)] is
approximated by:
w(s)g
,
(49)
E[g(Ξ)] ≈ Ns(cid:88)
s=1
(cid:16)
ξ(s)(cid:17)
SUBMITTED TO IEEE SYSTEMS JOURNAL
12
[13] J. A. Mirrlees, "The theory of moral hazard and
unobservable behaviour: Part i," The Review of Economic
Studies, vol. 66, no. 1, pp. 3 -- 21, 1999.
[Online].
Available: http://www.jstor.org/stable/2566946
[14] W. P. Rogerson, "The first-order approach to principal-
agent problems," Econometrica, vol. 53, no. 6, pp.
1357 -- 1367, 1985. [Online]. Available: http://www.jstor.
org/stable/1913212
[15] R. Ke, C. T. Ryan et al., "A general solution method for
moral hazard problems."
[16] R. Myerson, Game Theory: Analysis of Conflict.
Harvard University Press, 1991. [Online]. Available:
https://books.google.com/books?id=1w5PAAAAMAAJ
[17] R. B. Myerson, "Optimal auction design," Mathematics
of operations research, vol. 6, no. 1, pp. 58 -- 73, 1981.
[18] D. Kahneman and A. Tversky, "Prospect
theory.. an
analy-sis of decision under risk," 2008.
[19] Theano Development Team, "Theano: A Python frame-
work for fast computation of mathematical expressions,"
arXiv e-prints, vol. abs/1605.02688, May 2016. [Online].
Available: http://arxiv.org/abs/1605.02688
[20] A. Doucet, S. Godsill, and C. Andrieu, "On sequential
monte carlo sampling methods for bayesian filtering,"
Statistics and computing, vol. 10, no. 3, pp. 197 -- 208,
2000.
[21] Ilias bilionis, "Sequential Monte Carlo working on
top of pymc." [Online]. Available: https://github.com/
PredictiveScienceLab/pysmc
[22] J. R. Wertz, D. F. Everett, and J. J. Puschell, Space
mission engineering: the new SMAD.
Hawthorne,
CA: Microcosm Press : Sold and distributed world-
wide by Microcosm Astronautics Books, 2011, oCLC:
747731146.
[23] "Solid."
[Online]. Available: http://www.astronautix.
com/s/solid.html
[24] "NASA Historical Data Books." [Online]. Available:
https://history.nasa.gov/SP-4012/vol6/cover6.html
[25] "Propulsion Engineer Salaries." [Online]. Available:
https://www.paysa.com/salaries/propulsion-engineer--t
[26] T. Gerstner and M. Griebel, "Numerical
integration
using sparse grids," Numerical Algorithms, vol. 18,
no. 3, p. 209, Jan 1998. [Online]. Available: https:
//doi.org/10.1023/A:1019129717644
where w(s) and ξ(s) are the Ns = 127 quadrature points of
the level 6 sparse grid quadrature constructed by the Gauss-
Hermite 1D quadrature rule.
ACKNOWLEDGMENT
This material is based upon work supported by the National
Science Foundation under Grant No. 1728165.
REFERENCES
[1] G. Locatelli, "Why are megaprojects, including nuclear
power plants, delivered overbudget and late? reasons and
remedies," arXiv preprint arXiv:1802.07312, 2018.
[2] U. S. G. A. Office, "Navy shipbuilding past performance
provides valuable lessons for future investments," 2018.
[3] N. GAO, "Assessments of major projects," 2018.
[4] I. Maddox, P. Collopy, and P. A. Farrington, "Value-based
assessment of dod acquisitions programs," Procedia
Computer Science, vol. 16, pp. 1161 -- 1169, 2013, 2013
Conference on Systems Engineering Research. [Online].
Available: http://www.sciencedirect.com/science/article/
pii/S1877050913001233
[5] P. D. Collopy and P. M. Hollingsworth, "Value-Driven
Design," Journal of Aircraft, vol. 48, no. 3, pp.
749 -- 759, 2011. [Online]. Available: https://doi.org/10.
2514/1.C000311
[6] A. Bandura, "Social cognitive theory: An agentic per-
spective," Annual review of psychology, vol. 52, no. 1,
pp. 1 -- 26, 2001.
[7] M. Shergadwala, I. Bilionis, K. N. Kannan, and J. H. Pan-
chal, "Quantifying the impact of domain knowledge and
problem framing on sequential decisions in engineering
design," Journal of Mechanical Design, vol. 140, no. 10,
p. 101402, 2018.
[8] A. M. Chaudhari, Z. Sha, and J. H. Panchal, "Analyzing
participant behaviors in design crowdsourcing contests
using causal inference on field data," Journal of Mechan-
ical Design, vol. 140, no. 9, p. 091401, 2018.
[9] S. D. Vermillion and R. J. Malak, "Using a principal-
agent model to investigate delegation in systems engi-
neering," in ASME 2015 International Design Engineer-
ing Technical Conferences and Computers and Infor-
mation in Engineering Conference. American Society
of Mechanical Engineers, 2015, pp. V01BT02A046 --
V01BT02A046.
[10] S. Safarkhani, V. R. Kattakuri, I. Bilionis, and J. Panchal,
"A principal-agent model of systems engineering pro-
cesses with application to satellite design," arXiv preprint
arXiv:1903.06979, 2019.
[11] J. Watson, "Contract, mechanism design, and technologi-
cal detail," Econometrica, vol. 75, no. 1, pp. 55 -- 81, 2007.
[12] S. Safarkhani, I. Bilionis, and J. Panchal, "Understanding
the effect of task complexity and problem-solving skills
on the design performance of agents in systems engineer-
ing," the ASME 2018 International Design Engineering
Technical Conferences and Computers and Information
in Engineering Conference, Aug 2018.
|
cs/0001008 | 3 | 0001 | 2003-06-20T14:20:48 | Predicting the expected behavior of agents that learn about agents: the CLRI framework | [
"cs.MA",
"cs.LG"
] | We describe a framework and equations used to model and predict the behavior of multi-agent systems (MASs) with learning agents. A difference equation is used for calculating the progression of an agent's error in its decision function, thereby telling us how the agent is expected to fare in the MAS. The equation relies on parameters which capture the agent's learning abilities, such as its change rate, learning rate and retention rate, as well as relevant aspects of the MAS such as the impact that agents have on each other. We validate the framework with experimental results using reinforcement learning agents in a market system, as well as with other experimental results gathered from the AI literature. Finally, we use PAC-theory to show how to calculate bounds on the values of the learning parameters. | cs.MA | cs |
Autonomous Agents and Multiagent Systems, January, 2003.
Predicting the Expected Behavior of Agents that
Learn About Agents: The CLRI Framework
Jos´e M. Vidal and Edmund H. Durfee
Swearingen Engineering Center, University of South
Carolina, Columbia, SC, 29208
Advanced Technology Laboratory, University of
Michigan, Ann Arbor, MI, 48102
October 29, 2018
Abstract
We describe a framework and equations used to model and predict the
behavior of multi-agent systems (MASs) with learning agents. A difference
equation is used for calculating the progression of an agent's error in its
decision function, thereby telling us how the agent is expected to fare in
the MAS. The equation relies on parameters which capture the agent's
learning abilities, such as its change rate, learning rate and retention rate,
as well as relevant aspects of the MAS such as the impact that agents
have on each other. We validate the framework with experimental results
using reinforcement learning agents in a market system, as well as with
other experimental results gathered from the AI literature. Finally, we
use PAC-theory to show how to calculate bounds on the values of the
learning parameters.
Multi-Agent Systems, Machine Learning, Complex Systems.
1 Introduction
With the steady increase in the number of multi-agent systems (MASs) with
learning agents [9, 5, 10, 31] the analysis of these systems is becoming increas-
ingly important. Some of the research in this area consists of experiments where
a number of learning agents are placed in a MAS, then different learning or sys-
tem parameters are varied and the results are gathered and analyzed in an
effort to determine how changes in the individual agent behaviors will affect the
system behavior. We have learned about the dynamics of market-based MASs
using this approach [34]. However, in this article we will take a step beyond
these observation-based experimental results and describe a framework that can
be used to model and predict the behavior of MASs with learning agents. We
give a difference equation that can be used to calculate the progression of an
c(cid:13) Kluwer Academic Publishers 2002.
PSfrag replacements
w:
δ
δ
δ
δ2
δ3
a:
δ1
Figure 1: The agents in a MAS.
agent's error in its decision function. The equation relies on the values of pa-
rameters which capture the agents' learning abilities and the relevant aspects
of the MAS. We validate the framework by comparing its predictions with our
own experimental results and with experimental results gathered from the AI
literature. Finally, we show how to use probably approximately correct (PAC)
theory to get bounds on the values of some of the parameters.
The types of MAS we study are exemplified by the abstract representation
shown in Figure 1. We assume that the agents observe the physical state of
the world (denoted by w in the figure) and take some action (a) based on their
observation of the world state. An agent's mapping from states to actions is
denoted by the decision function (δ) inside the agent. Notice that the "physical"
state of the world includes everything that is directly observable by the agent
using its sensors. It could include facts such as a robot's position, or the set
outstanding bids in an auction, depending on the domain. The agent does not
know the decision functions of the other agents. After taking action an agent
can change its decision function as prescribed by whatever machine-learning
algorithm the agent is using.
We have a situation where agents are changing their decision function based
on the effectiveness of their actions. However, the effectiveness of their actions
depends on the other agents' decision functions. This scenario leads to the
immediate problem: if all the agents are changing their decisions functions then
it is not clear what will happen to the system as a whole. Will the system
settle to some sort of equilibrium where all agents stop changing their decision
functions? How long will it take to converge? How do the agents' learning
abilities influence the system's behavior and possible convergence? These are
some of the questions we address in this article.
Section 2 presents our framework for describing an agent's learning abilities
and the error in its behavior. Section 3 presents an equation that can be used
to predict an agent's expected error, as a function of time, when the agent is
in a MAS composed of other learning agents. This equation is simplified in
Section 4 by making some assumptions about the type of MAS being modeled.
Section 5 then defines the last few parameters used in our framework -- volatility
and impact. Section 6 gives an illustrative example of the use of the framework.
The predictions made by our framework are verified by our own experiments, as
2
shown in Section 7, and with the experiments of others, as shown in Section 8.
The use of PAC theory for determining bounds on the learning parameters is
detailed in Section 9. Finally Section 10 describes some of the related work and
Section 12 summarizes our claims.
2 A Framework for Modeling MASs
In order to analyze the behaviors of agents in MASs composed of learning
agents, we must first construct a formal framework for describing these agents.
The framework must state any assumptions and simplifications it makes about
the world, the agents, and the agents' behaviors. It must also be mathemati-
cally precise, so as to allow us to make quantitative predictions about expected
behaviors. Finally, the simplifications brought about because of the need for
mathematical precision should not be so constraining that they prevent the ap-
plicability of the framework to a wide variety of learning algorithms and different
types of MASs. We now describe our framework and explain the types of MASs
and learning behaviors that it can capture.
2.1 The World and its Agents
A MAS consists of a finite number of agents, actions, and world states. We
let N denote the finite set of agents in the system. W denotes the finite set of
world states. Each agent is assumed to have a set of perceptors (e.g., a camera,
microphone, bid queue) with which it can perceive the world. An agent uses is
sensors to "look" at the world and determine which world state w it is in; the
set of all these states is W . Ai, where Ai ≥ 2, denotes the finite set of actions
agent i ∈ N can take.
We assume discrete time, indexed in the various functions by the superscript
t, where t is an integer greater than or equal to 0. The assumption of discrete
time is made, for practical reasons, by a number of learning algorithms.
It
means that, while the world might be continuous, the agents perceive and learn
in separate discrete time steps.
We also assume that there is only one world state w at each time, which all
the agents can perceive in its completeness. That is, we asume the enviroment
is accessible (as defined in [26, p46]). This assumption holds for market systems
in which all the actions of all the agents are perceived by all the agents, and
for software agent domains in which all the agents have access to the same
information. However, it might not hold for robotic domains where one agent's
view of the world might be obscured by some physical obstacle. Even in such
domains, it is possible that there is a strong correlation between the states
perceived by each agent. These correlations could be used to create equivalency
classes over the agents' perceived states, and these classes could then be used
as the states in W .
Finally, we assume the environment is determistic [26, p46]. That is, the
agents' combined actions will always have the expected effect. Of course, agent
3
New world wt ∈ D
/Perceive world wt
iSSSSSSSSSSSSSSSSSS
t←t+1
Learn
/Take action δi(wt)
Receive payoff
or feedback.
Figure 2: Action/Learn loop for an agent.
i might not know what action agent j will take so i might not know the eventual
effect of its own individual action.
2.2 A Description of Agent Behavior
In the types of MASs we are modeling, every agent i perceives the state of the
world w and takes an action ai, at each time step. We assume that every agent's
behavior, at each moment in time, can be described with a simple state-to-
action mapping. That is, an agent's choice of action is solely determined by its
current state-to-action mapping and the current world w.
Formally, we say that agent i's behavior is represented by a decision func-
tion (also known as a "policy" in control theory and a "strategy" in game
theory), given by δt
i : W → Ai. This function maps each state w ∈ W to the
action ai ∈ Ai that agent i will take in that state, at time t. This function can
effectively describe any agent that deterministically chooses its action based on
the state of the world. Notice that the decision function is indexed with the
time t. This allows us to represent agents that change their behavior.
The action agent i should take in each state w is given by the target func-
tion ∆t
i : W → Ai, which also maps each state w ∈ W to an action ai ∈ Ai.
The agent does not have direct access to its target function. The target function
is used to determine how well an agent is doing. That is, it represents the "per-
fect" behavior for a given agent. An agent's learning task is to get its decision
function to match its target function as much as possible.
i, one would need to know δt
Since the choice of action for agent i often depends on the actions of other
agents, the target function for i needs to take these actions into account. That
is, in order to generate ∆t
j(w) for all j ∈ N−i and
w ∈ W . These δt
j(w) functions tell us the actions that all the other agents will
take in every state w. For example, in order for one to determine what an agent
should bid in every world w of an auction-based market system, one will need
to know what the other agents will bid in every world w. One can use these
actions, along with the state w, in order to identify the best action for i to take.
i . These changes
in an agent's decision function reflect its learned knowledge. The agents in
the MASs we consider are engaged in the discrete action/learn loop shown in
Figure 2. The loop works as follows: At time t the agents perceive a world
wt ∈ W which is drawn from a fixed distribution D(w). They then each take
the action dictated by their δt
i functions; all of these actions are assumed to
be taken effectively in parallel. Lastly, they each receive a payoff which their
i (w) can change over time, so that δt+1
An agent's δt
6= δt
i
4
/
/
i
o
o
i
j of all other agents j ∈ N−i.
respective learning algorithms use to change the δt
i so as to, hopefully, better
i. By time t + 1, the agents have new δt+1
match ∆t
functions and are ready to
perceive the world again and repeat the loop. Notice that, at time t, an agent's
∆t
i is derived by taking into account the δt
We assume that D(w) is a fixed probability distribution from which we take
the worlds seen at each time. This assumption is not unreasonably limiting.
For example, in an economic domain where the new state is the new good being
offered, or in an episodic domain where the agents repeatedly engage in different
games (e.g. a Prisoner's Dilemma competition) there is no correlation between
successive world states or between these states and the agents' previous actions.
However, in a robotic domain one could argue that the new state of the world
will depend on the current state of the world; after all, the agents probably
move very little each time step.
Our measure of the correctness of an agent's behavior is given by our error
measure. We define the error of agent i's decision function δt
i (w) as
e(δt
i ) = Xw∈W
= Prw∈D[δt
D(w)Pr[δt
i (w) 6= ∆t
i(w)]
i (w) 6= ∆t
i(w)].
(1)
e(δt
i ) gives us the probability that agent i will take an incorrect action; it
is in keeping with the error definition used in computational learning theory
[17]. We use it to gauge how well agent i is performing. An error of 0 means
that the agent is taking all the actions dictated by its target function. An
error of 1 means that the agent never takes an action as dictated by its target
function. Each action the agent takes is either correct or incorrect, that is, it
either matches the target function or it does not. We do not model degrees of
incorrectness. However, since the error is defined as the average over all possible
world states, an agent that takes the correct action in most world states will
have a small error. Extending the theory to handle degrees of incorrectness is
one of the subjects of your continuing work, see Section 11. All the notation
from this section is summarized in Figure 3.
2.3 The Moving Target Function Problem
The learning problem the agent faces is to change its δt
i (w) so that it matches
∆t
i(w). If we imagine the space of all possible decision functions, then agent i's
δt
i and ∆t
i will be two points in this space, as shown in Figure 4. The agent's
learning problem can then be re-stated as the problem of moving its decision
function as close as possible to its target function, where the distance between
the two functions is given by the error e(δt
i ). This is the traditional machine
learning problem.
However, once agents start to change their decision functions (i.e., change
their behaviors) the problem of learning becomes more complicated because
these changes might cause changes in the other agents' target functions. We
end up with a moving target function, as seen in Figure 5. In these systems, it
5
N the set of all agents, where i ∈ N is one particular agent.
W the set of possible states of the world, where w ∈ W is one particular
state.
Ai the set of all actions that agent i can take.
δt
i : W → Ai the decision function for agent i at time t.
It tells which
action agent i will take in each world.
∆t
i : W → Ai the target function for agent i at time t. It tells us what
It takes into account the actions that
action agent i should take.
other agents will take.
e(δt
i ) = Pr[δt
i (w) 6= ∆t
i(w) w ∈ D] the error of agent i at time t. It is the
probability that i will take an incorrect action, given that the worlds
w are taken from the fixed probability distribution D.
Figure 3: Summary of notation used for describing a MAS and the agents in it.
8q
qqqqq
Learn
qqqq
qqq
δt
i
δt+1
i
e(δt+1
i
)
e(δt
i )
∆i
Figure 4: The traditional learning problem.
6
)
i
)
i
)
i
)
i
)
i
)
i
)
i
)
i
)
i
)
i
)
i
)
i
8
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
is not clear if the error will ever reach 0 or, more generally, what the expected
error will be as time goes to infinity. Determining what will happen to an agent's
error in such a system is what we call the moving target function problem,
which we address in this article. However, we will first need to define some
parameters that describe the capabilities of an agent's learning algorithm.
8q
qq
qq
Learn
qqq
qq
qq
q
δt
i
δt+1
i
e(δt
i )
e(δt+1
i
)
∆t
i Move
/ ∆t+1
i
Figure 5: The learning problem in learning MASs.
2.4 A Model of Learning Algorithms
i into δt+1
An agent's learning algorithm is responsible for changing δt
so that it
is a better match of ∆t
i. Different machine learning algorithms will achieve this
match with different degrees of success. We have found a set of parameters that
can be used to model the effects of a wide range of learning algorithms. The
parameter are: Change rate, Learning rate, Retention rate, and Impact; and
they will be explained in this section, except for Impact which will be introduced
in Section 5. These parameters, along with the equations we provide, form the
CLRI framework (the letters correspond to the first letter of the parameters'
names).
i
After agent i takes an action and receives some payoff, it activates its learning
algorithm, as we showed in Figure 2. The learning algorithm is responsible for
using this payoff in order to change δt
i as much
as possible. We can expect that for some w it was true that δt
i(w),
while for some other w this was not the case. That is, some of the w → ai
mappings given by δt
In general, a learning
algorithm might affect both the correct and incorrect mappings. We will treat
these two cases separately.
i (w) might have been incorrect.
i match ∆t
, making δt+1
i into δt+1
i (w) = ∆t
i
We start by considering the incorrect mappings and define the change rate
of the agent as the probability that the agent will change at least one of its
incorrect mappings. Formally, we define the change rate ci for agent i as
∀w Pr[δt+1
i
(w) 6= δt
i (w) δt
i (w) 6= ∆t
i(w)] = ci.
(2)
The change rate tells us the likelihood of the agent changing an incorrect
mapping into something else. This "something else" might be the correct action,
but it could also be another incorrect action. The probability that the agent
changes an incorrect mapping to the correct action is called the learning rate
of the agent. It is defined as li where
∀w Pr[δt+1
i
(w) = ∆t
i(w) δt
i (w) 6= ∆t
i(w)] = li.
(3)
7
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
+
k
8
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
o
/
There are two constraints which must always be satisfied by these two rates.
Since changing to the correct mapping implies that a change was made, the
value of li must be less than or equal to ci, that is, li ≤ ci must always be true.
Also, if Ai = 2 then ci = li since there are only two actions available, so the
one that is not wrong must be right.
The complementary value for the learning rate is 1 − li and refers to the
probability that an incorrect mapping does not get changed to a correct one. An
example learning rate of li = .5 means that if agent i initially has all mappings
wrong it will make half of them match the original target function after the first
iteration.
We now consider the agent's correct mappings and define the retention
rate as the probability that a correct mapping will stay correct in the next
iteration. The retention rate is given by ri where
∀w Pr[δt+1
i
(w) = ∆t
i(w) δt
i (w) = ∆t
i(w)]. = ri.
(4)
We propose that the behavior of a wide variety of learning algorithms can be
captured (or at least approximated) using appropriate values for ci, li, and
ri. Notice, however, that these three rates claim that the w → a mappings
that change are independent of the w that was just seen. We can justify this
independence by noting that most learning algorithms usually perform some
form of generalization. That is, after observing one world state w and the
payoff associated with it, a typical learning algorithm is able to generalize what
it learned to some other world states. This generalization is reflected in the
fact that the change, learning, and retention rates apply to all w's. However, a
more precise model would capture the fact that, in some learning algorithms,
the mapping for the world state that was just seen is more likely to change than
the mapping for any other world state.
The rates are not time dependent because we assume that agents use one
learning algorithm during their lifetimes. The rates capture the capabilities of
this learning algorithm and, therefore, do not need to vary over time.
Finally, we define volatility to mean the probability that the target function
will change from time t to time t + 1. Formally, volatility is given by vi where
∀w Pr[∆t+1
i
(w) 6= ∆t
i(w)] = vi
(5)
In Section 5, we will show how to calculate vi in terms of the error of the other
agents. We will then see that volatility is not a constant but, instead, varies
with time.
3 Calculating the Agent's Error
We now wish to write a difference equation that will let us calculate the agent's
expected error, as defined in Eq. (1), at time t+1 given the error at time t and the
other parameters we have introduced. We can do this by observing that there
are two conditions that determine the new error: whether ∆t+1
i(w) or
(w) = ∆t
i
8
not, and whether δt
and b ≡ δt
i (w) = ∆t
where: a ∧ b, a ∧ ¬b, ¬a ∧ b, and ¬a ∧ ¬b. Formally, this implies that
i (w) = ∆t
i(w),
i(w), we can then say that we need to consider the four cases
i(w) or not. If we define a ≡ ∆t+1
(w) = ∆t
i
Pr[δt+1
i
(w) 6= ∆t+1
i
(w)] =
i
Pr[δt+1
Pr[δt+1
i
(w) 6= ∆t+1
(w) 6= ∆t+1
i
i
(w) ∧ a ∧ b] + Pr[δt+1
(w) ∧ ¬a ∧ b] + Pr[δt+1
i
(w) 6= ∆t+1
i
(w) ∧ a ∧ ¬b]+
(6)
(w) 6= ∆t+1
i
(w) ∧ ¬a ∧ ¬b],
i
since the four cases are exclusive of each other. Applying the chain rule of
probability, we can rewrite each of the four terms in order to get
Pr[δt+1
i
(w) 6= ∆t+1
i
(w)] = Pr[a ∧ b] · Pr[δt+1
(w) 6= ∆t+1
(w) a ∧ b]+
i
Pr[a ∧ ¬b] · Pr[δt+1
Pr[¬a ∧ b] · Pr[δt+1
Pr[¬a ∧ ¬b] · Pr[δt+1
i
i
i
i
(w) 6= ∆t+1
(w) 6= ∆t+1
i
i
(w) a ∧ ¬b]+
(w) ¬a ∧ b]+
(7)
(w) 6= ∆t+1
i
(w) ¬a ∧ ¬b].
We can now find values for these conditional probabilities. We start with the
first term where, after replacing the values of a and b, we find that
Pr[δt+1
i
(w) 6= ∆t+1
i
(w) ∆t+1
i
(w) = ∆t
i(w) ∧ δt
i (w) = ∆t
i(w)] = 1 − ri.
(8)
Since the target function does not change from time t to t + 1 and the agent was
correct at time t, the agent will also be correct at time t + 1; unless it changes
its correct w → a mapping. The agent changes this mapping with probability
1 − ri.
The value for the second conditional probability is
Pr[δt+1
i
(w) 6= ∆t+1
i
(w) ∆t+1
i
(w) = ∆t
i(w) ∧ δt
i (w) 6= ∆t
i(w)] = 1 − li.
(9)
In this case the target function still stays the same but the agent was incorrect.
If the agent was incorrect then it will change its decision function to match
the target function with probability li. Therefore, the probability that it will
be incorrect next time is the probability that it does not make this change, or
1 − li.
The third probability has a value of
Pr[δt+1
(w) 6= ∆t+1
(w) ∆t+1
i
= (ri + (1 − ri) · B)
i
i
(w) 6= ∆t
i(w) ∧ δt
i (w) = ∆t
i(w)]
(10)
In this case the agent was correct and the target function changes. This means
that if the agent retains the same mapping, which it does with probability ri,
then the agent will definitely be incorrect at time t + 1. If it does not retain the
same mapping, which happens with probability 1 − ri, then it will be incorrect
with probability B, where
B = Pr[δt+1
(w) 6= ∆t+1
(w) 6= ∆t
i
∧ δt+1
i
i
i(w)].
(w)δt
i (w) = ∆t
i(w) ∧ ∆t+1
i
(w) 6= ∆t
i(w)
(11)
9
Finally, the fourth conditional probability has a value of
Pr[δt+1
(w) 6= ∆t+1
(w) ∆t+1
i
= (1 − ci)D + li + (ci − li)F,
i
i
(w) 6= ∆t
i(w) ∧ δt
i (w) 6= ∆t
i(w)]
where
D = Pr[δt
F = Pr[δt+1
∧ δt+1
i
i
i (w) 6= ∆t+1
i
(w)δt
i (w) 6= ∆t
i(w) ∧ ∆t+1
i
(w) 6= ∆t
i(w)]
(w)δt
(w) 6= ∆t+1
i
i(w) ∧ δt+1
(w) 6= ∆t
i
(w) 6= δt
i (w)].
i (w) 6= ∆t
i(w) ∧ ∆t+1
i
(w) 6= ∆t
i(w)
(12)
(13)
(14)
This is the case where the target function changes and the agent was wrong.
We have to consider three possibilities. The first possibility is for the agent
not to change its decision function, which happens with probability 1 − ci. The
probability that the agent will be incorrect in this case is given by D. The
second possibility, when the agent changes its mapping to the correct function,
has a probability of li and ensures that the agent will be incorrect the next time.
The third possibility happens, with probability ci − li when the agent changes
its mapping to an incorrect value. In this case, the probability that it will be
wrong next time is given by F .
We can substitute Eqs. (8), (9), (10), and (12) into Eq. (7), substitute the
values of a and b, and expand Pr[a ∧ b] into Pr[a b] · Pr[b], in order to get
E[e(δt+1
i
)] = E[ Xw∈W
D(w)Pr[δt+1
i
(w) 6= ∆t+1
i
i
Pr[∆t+1
+ Pr[∆t+1
+ Pr[∆t+1
(w) = ∆t
(w) = ∆t
(w) 6= ∆t
i(w)δt
i(w)δt
i(w)δt
i (w) = ∆t
i (w) 6= ∆t
i (w) = ∆t
i(w)] · Pr[δt
i(w)] · Pr[δt
i(w)]
i
i
· Pr[δt
i (w) = ∆t
i(w)] · (ri + (1 − ri) · B)
D(w)(
(w)]] = Xw∈W
i (w) = ∆t
i (w) 6= ∆t
i(w)] · (1 − ri)
i(w)] · (1 − li)
(15)
+ Pr[∆t+1
i
(w) 6= ∆t
i(w)δt
i (w) 6= ∆t
i(w)] · Pr[δt
i (w) 6= ∆t
i(w)]
· (1 − ci)D + li + (ci − li)F.
Equation (15) will model any MAS whose agent learning can be described with
the parameters presented Section 2.4 and whose action/learn loop is the same
as we have described. We can use Eq. (15) to calculate the successive expected
errors for agent i, given values for all the parameters and probabilities. In the
next section we show how this is done in a simple example game.
3.1 The Matching game
In this matching game we have two agents i and j each of whom, in every
world w, wants to play the same action as the other one. Their set of actions is
Ai = Aj , where we assume Ai > 2 (for Ai = 2 the equation is simpler). After
every time step, the agents both learn and change their decision functions in
10
accordance to their learning rates, retention rates, and change rates. Since the
agents are trying to match each other, in this game it is always true that ∆t
i(w) =
δt
j(w) and ∆t
i (w). Given all this information, we can find values for
some of the probabilities in Eq. (15) (including values for Equations (11) (13)
(14)) and rewrite (see Appendix A for derivation) it as:
j (w) = δt
i
E[e(δt+1
)] = Xw∈W
+ (1 − cj) · Pr[δt
i (w) 6= ∆t
i(w)] · (1 − li)
D(w){rj · Pr[δt
i (w) = ∆t
i(w)] · (1 − ri)
+ (1 − rj ) · Pr[δt
i (w) = ∆t
i(w)] ·(cid:18)ri + (1 − ri) ·(cid:18) Ai − 2
Ai − 1(cid:19)(cid:19)
(16)
+ cj · Pr[δt
i (w) 6= ∆t
i(w)] ·(cid:18)1 − lj +
cilj(Ai − 1) + li(1 − lj) − ci
Ai − 2
(cid:19)}
We can better understand this equation by plugging in some values and sim-
plifying. For example, lets assume that ri = rj = 1 and li = lj = 1, which
implies that ci = cj = 1. This is the case where the two agents always change
all their incorrect mappings so as to match their respective target functions at
i (w1) = x and δt
time t. That is, if we had δt
j(w1) = y, then at time t + 1 we will
have δt+1
(w1) = y and δt+1
(w1) = x. This means that agent i changes all its
incorrect mappings to match j, while j changes to match i, so all the mappings
stay wrong after all (i.e., i ends up doing what j did before, while j does what
i did before). The error, therefore, stays the same. We can see this by plugging
the values into Eq. (16). The first three terms will become 0 and the fourth
term will simplify to the definition of error, as given by Eq. (1). Since the fourth
term is the only one that is non-zero, we end up with E[e(δt+1
)] = e(δt
j
i
i ).
i
)] = cie(δt
We can also let ci and li (keeping cj = lj = 1) be arbitrary numbers, which
gives us E[e(δt+1
i ). This tells us that the error will drop faster for
a smaller change rate ci. The reason is that i's learning (remember li ≤ ci) in
this game is counter-productive because it is always made invalid by j's learning
rate of 1. That is, since j is changing all its mappings to match i's actions, i's
best strategy is to keep its actions the same (i.e., ci = 0).
i
4 Further Simplification
We can further simplify Eq. (15) if we are willing to make two assumptions.
The first assumption is that the new actions chosen when either δt
i (w) changes
(and does not match the target), or when ∆t
i(w) changes, are both taken from
flat probability distributions over Ai. By making this assumption we can find
values for B, D, and F , namely:
B = D =
Ai − 2
Ai − 1
F =
Ai − 3
Ai − 2
(17)
The second assumption we make is that the probability of ∆t
for a particular w, is independent of the probability that δt
i(w) changing,
i (w) was correct.
11
i (w) changing were correlated since, if δt
In Section 3.1 we saw that in the matching game the probabilities of ∆t
and δt
also wrong, which meant j would probably change δt
∆t
i(w)
j(w) was
j(w), which would change
i (w) was wrong then δt
i(w).
However, the matching game is a degenerate example in exhibiting such tight
coupling between the agents' target functions. In general, we can expect that
there will be a number of MASs where the probability that any two agents i and
j are correct is uncorrelated (or loosely correlated). For example, in a market
system all sellers try to bid what the buyer wants, so the fact that one seller
bids the correct amount says nothing about another seller's bid. Their bids are
all uncorrelated. In fact, the Distributed Artificial Intelligence literature is full
of systems that try to make the agents' decisions as loosely-coupled as possible
[19, 21].
This second assumption we are trying to make can be formally represented
by having Eq. (18) be true for all pairs of agents i and j in the system.
Pr[δt
i (w) = ∆t
= Pr[δt
i(w) ∧ δt
i (w) = ∆t
j(w) = ∆t
i(w)] · Pr[δt
j(w)]
j(w) = ∆t
j(w)]
(18)
Once we make these two assumptions we can rewrite Eq. (15) as:
E[e(δt+1
i
)] = Xw∈W
D(w)(Pr[∆t+1
i
(w) = ∆t
i(w)] · (Pr[δt
i (w) = ∆t
i(w)] · (1 − ri)
+ Pr[δt
i (w) 6= ∆t
i(w)] · (1 − li))
+ Pr[∆t+1
i
(w) 6= ∆t
i(w)] · (Pr[δt
i (w) = ∆t
i(w)] ·(cid:18)ri + (1 − ri) ·(cid:18) Ai − 2
Ai − 1(cid:19)(cid:19)
(cid:19)))
+ Pr[δt
i (w) 6= ∆t
i(w)] ·(cid:18) Ai − 2 − ci + 2li
Ai − 1
Some of the probabilities in this equation are just the definition of vi, and others
simplify to the agent's error. This means that we can simplify Eq. (19) to:
(19)
E[e(δt+1
i
)] = 1 − ri + vi(cid:18) Airi − 1
Ai − 1 (cid:19)
i )(cid:18)ri − li + vi(cid:18) Ai(li − ri) + li − ci
+ e(δt
Ai − 1
(cid:19)(cid:19) (20)
Eq. (20) is a difference equation that can be used to determine the expected
error of the agent at any time by simply using E[e(δt+1
i ) for the
next iteration. While it might look complicated, it is just the function for a line
y = mx + b where x = e(δt
). Using this observation, and the
fact that e(δt+1
) will always be between 0 and 1, we can determine that the
final convergence point for the error is the point where Eq. (20) intersects the
line y = x. The only exception is if the slope equals −1, in which case we will
see the error oscillating between two points.
i ) and y = e(δt+1
)] as the e(δt
i
i
i
12
)
1
+
t
δ
(
e
i
PSfrag replacements
learning
volatility
t+1
i
e(δ
)
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
e(δt
i )
0.8
1
Figure 6: Error progression for agent i, assuming a fixed volatility vi = .2,
ci = 1, li = .3, ri = 1, Ai = 20. We show the error function (e(δt+1
)), as well
as its two components: learning and volatility. The line y = x allows us trace
the agent's error as it starts at .95 and converges to .44.
i
By looking at Eq. (20) we can also determine that there are two "forces"
acting on the agent's error: volatility and the agent's learning abilities. The
volatility tends to increase the agent's error past its current value while the
learning reduces it. We can better appreciate this effect by separating the vi
terms in Eq. (20) and plotting the vi terms (volatility) and the rest of the terms
(learning) as two separate lines. By definition, these will add up to the line
given by Eq. (20). We have plotted these three lines and traced a sample error
progression in Figure 6. The error starts at .95 and then decreases to eventually
converge to .44. We notice the learning curve always tries to reduce the agent's
error, as confirmed by the fact that its line always falls below y = x. Meanwhile,
the volatility adds an extra error. This extra error is bigger when the agent's
error is small since, any change in the target function is then likely to increase
the agent's error.
5 Volatility and Impact
Equation (20) is useful for determining the agent's error when we know the
volatility of the system. However, it is likely that this value is not available
to us (if we knew it we would already know a lot about the dynamics of the
In this section we determine the value of vi in terms of the other
system).
agents' changes in their decision functions. That is, in terms of Pr[δt+1
j],
for all other agents j.
6= δt
j
In order to do this we first need to define the impact Iji that agent j's
13
changes in its decision function have on i's target function.
∀w∈W Iji = Pr[∆t+1
i
(w) 6= ∆t
i(w) δt+1
j
(w) 6= δt
j(w)]
(21)
We can now start to define volatility by first determining that, for two agents
i and j
∀w∈W vt
i
i = Pr[∆t+1
= Pr[∆t+1
+ Pr[∆t+1
i
i
(w) 6= ∆t
(w) 6= ∆t
(w) 6= ∆t
i(w)]
i(w) δt+1
i(w) δt+1
j
j
(w) 6= δt
(w) = δt
j(w)] · Pr[δt+1
j(w)] · Pr[δt+1
j
(w) 6= δt
(w) = δt
j(w)]
j(w)].
j
(22)
The reader should notice that volatility is no longer constant;
it varies
with time (as recorded by the superscript). The first conditional probability
in Eq. (22) is just Iji. The second one we will set to 0, since we are specifically
interested in MASs where the volatility arises only as a side-effect of the other
agents' learning. That is, we assume that agent i's target function changes only
when j's decision function changes. For cases with more than two agents, we
similarly assume that one agent's target function changes only when some other
agent's decision function changes. That is, we ignore the possibility that outside
influences might change an agent's target function.
We can simplify Eq. (22) and generalize it to N agents, under the assumption
that the other agents' changes in their decision functions will not cancel each
other out, making ∆t
∀w∈W vt
i then becomes
i stay the same as a consequence. vt
i = Pr[∆t+1
= 1 − Yj∈N−i
(w) 6= ∆t
(1 − IjiPr[δt+1
i(w)]
j
i
(w) 6= δt
j(w)]).
(23)
We now need to determine the expected value of Pr[δt+1
(w) 6= δt
j(w)] for
j
any agent. Using i instead of j we have
(w) 6= δt
∀w∈W Pr[δt+1
i (w)]
i (w) 6= ∆t
i (w) = ∆t
where the expected value is:
i
= Pr[δt
+ Pr[δt
i(w)] · Pr[δt+1
i(w)] · Pr[δt+1
i
i
(w) 6= ∆t
(w) 6= ∆t
i(w) δt
i(w) δt
i (w) 6= ∆t
i (w) = ∆t
i(w)]
i(w)],
E[Pr[δt+1
i
(w) 6= δt
i (w)]] = cie(δt
i ) + (1 − ri) · (1 − e(δt
i )).
(24)
(25)
We can then plug Eq. (25) into Eq. (23) in order to get the expected volatility
E[vt
i ] = 1 − Yj∈N−i
1 − Iji(cje(δt
j) + (1 − rj) · (1 − e(δt
j))).
(26)
We can use this expected value of vt
i in Eq. (20) in order to find out how the
other agents' learning will affect agent i. In MASs that have identical learning
agents (i.e., their c, l, r, and I rates are all the same and they start with the
same initial error) we can replace the multiplier in Eq. (26) with an exponent
of N − 1. We use this simplification later in Section 8.2.
14
Final Error for i
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
PSfrag replacements
0
0.2
0.4
Iij
0.6
0.8
1
0
0.4
0.2
Error
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
1
0.8
Iji
0.6
Figure 7: Plot of Final Error for agent i, given li = lj = .2, ri = rj = 1,
ci = cj = 1, Aj = Ai = 20.
6 An Example with Two Agents
In a MAS with just two agents i and j, we can use Eq. (26) to rewrite Eq. (20)
as
E[e(δt+1
i
)] = 1 − ri + Iji(cje(δt
j) + (1 − rj) · (1 − e(δt
Ai − 1 (cid:19)
j)))(cid:18) Airi − 1
+ e(δt
i ){ri − li + Iji(cje(δt
j) + (1 − rj) · (1 − e(δt
j)))
(27)
·(cid:18) Ai(li − ri) + li − ci
Ai − 1
(cid:19)}.
We can now use Eq. (27) to plot values for one particular example. Let us
say that li = lj = .2, ci = cj = 1, ri = rj = 1, Aj = Ai = 20 and we let
the impacts Iij and Iji vary between zero and one. Figure 7 shows the final
error, after convergence, for this situation. It shows an area where the error is
expected to be below .1, corresponding to low values for either Iij , Iji or both.
This area represents MASs that are loosely coupled, i.e., one agent's change
in behavior does not significantly affect the other's target function.
In these
systems we can expect that the error will eventually1 reach a value close to zero.
We see that as the impact increases the final error also increases, with a fairly
abrupt transition between a final error of 0 and bigger final errors. This abrupt
transition is characteristic of these types of systems where there are tendencies
for the system to either converge or diverge, and both of them are self-enforcing
behaviors. Notice also that the graph is not symmetric -- Iij has more weight in
determining i's final error than Iji. This result seems counterintuitive, until we
realize that it is j's error that makes it hard for i to converge to a small error.
1Notice that we are not representing how long it takes for the error to converge. This can
easily be done and is just one more of the parameters our theory allows us to explore.
15
1
0.8
0.6
0.4
0.2
)
tj
δ
(
e
PSfrag replacements
0
0
0.2
0.4
0.6
e(δt
i )
0.8
1
Figure 8: Vector plot for e(δt
j), where Ai = Aj = 20, li = lj = .2,
ri = rj = 1, ci = .5, cj = 1, Iij = .1, Iji = .3. It shows the error progression for
a pair agents i and j. For each pair of errors (e(δt
j)), the arrows indicate
the expected (e(δt+1
i ) and e(δt
i ), e(δt
i
), e(δt+1
)).
j
If Iij is high then, if i has a large error then j's error will increase, which will
make j change its decision function often and make it hard for i to reduce its
error. If Iij is low then, even if Iji is high, j will probably settle down to a low
error and as it does i will also be able to settle down to a low error.
If we were about to design a MAS we would try to build it so that it lies
in the area where the final error is zero. This way we can expect all agents to
eventually have the correct behavior. We note that a substantial percentage of
the research in DAI and MAS deals with taking systems that are not inherently
in this area of near-zero error and designing protocols and rules of encounter so
as to move them into this area, as in [25].
The fact that the final error is 1 for the case with Iij = Iji = 1 can seem
non-intuitive to readers familiar with game theory. In game theory there are
many games, such as the "matching game" from Section 3.1, where two agents
have an impact of 1 on each other. However, it is known [3] that, in these
games, two learning agents will eventually converge to one of the equilibria (if
there are any), making their final error equal to 0. This is certainly true, and
it is exactly what we showed in Section 3.1. The same result is not seen in
Figure 7 because the figure was plotted using our simplified Equation. (20),
which makes the simplifying independence assumption given by Eq. (18). This
assumption cannot be made in games such as the matching game because, in
these games, there is a correlation between the correctness of each of the agents
actions. Specifically, in the matching game it is always true that both agents are
either correct, or incorrect, but it is never true that one of them is correct while
the other one is incorrect, i.e., either they matched, or they did not match.
Another view of the system is given by Figure 8 which shows a vector plot
16
PSfrag replacements
r
o
r
r
e
lj = .005
lj = .04
lj = .055
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
αj = .1
αj = .3
αj = .9
PSfrag replacements
αj = .1
αj = .3
αj = .9
r
o
r
r
e
200
400
600
800
1000
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
lj = .005
lj = .04
lj = .055
200
400
600
800
1000
time
time
(a) Experiment
(b) Theory
Figure 9: Comparison of observed and predicted error.
of the agents' errors. We can see how the bigger errors are quickly reduced but
the pace of learning decreases as the errors get closer to the convergence point.
Notice also that an agent's error need not change in a monotonic fashion. That
is, an agent's error can get bigger for a while before it starts to get smaller.
7 A Simple Application
In order to demonstrate how our theory can be used, we tested it on a simple
market-based MAS. The game consists of three agents, one buyer and two seller
agents i and j. The buyer will always buy at the cheapest price -- but the sellers
do not know this fact. In each time step the sellers post a price and the buyer
decides which of the sellers to buy from, namely, the one with the lowest bid.
The sellers can bid any one of 20 prices in an effort to maximize their profits.
The sellers use a reinforcement learning algorithm with their reinforcements
being the profit the agent achieved in each round, or 0 if it did not sell the good
at the time. In this system we had one good being sold (W = 1).As predicted
by economic theory, the price in this system settles to the sellers' marginal cost,
but it takes time to get there due to the learning inefficiencies.
We experimented with different αj rates2 for the reinforcement learning of
agent j, while keeping αi = .1 fixed, and plotted the running average of the
error of agent i. A comparison is shown in Figure 9. Figure 9(a) gives the
experimental results for three different values of αj. It shows i's average error,
over 100 runs, as a function of time. Since both sellers start with no knowledge,
their initial actions are completely random which makes their error equal to .5.
2
α is the relative weight the algorithm gives to the most recent payoff. α = 1 means that
it will forget all previous experience and use only the latest payoff to determine what action
to take.
17
Then, depending on αj, i's error will either start to go down from there or will
first go up some and then down. Eventually, i's error gets very close to 0, as
the system reaches a market equilibrium.
We can predict this behavior using Eq. (27). Based on the game description,
we set Ai = Aj = 20, since there were 20 possible actions. We let ri =
rj = 1 because reinforcement learning with fixed payoffs enforces the condition
that once an agent is taking the correct action it will never change its decision
function to take a different action. The agent might, however, still take a wrong
action but only when its exploration rate dictates it.
We then let Iij = Iji = .17 based on the rough calculation that each agent
has an equal probability of bidding any one of the 20 prices. If ∆t
i = 20 then
Iji for this situation is the probability that j was also bidding 20 or above,
i.e., 1/20, times the probability that j's new price is lower than 20, i.e. 19/20.
Similarly, if ∆t
i = 19 then Iji is equal to 2/20 times 18/20. The average of all
of these probabilities is .17. A more precise calculation of the impact would
require us to find it via experimentation by actually running the system.
Finally, we chose li = lj = ci = cj = .005 for the first curve (i.e., the one that
compares with αj = .1). We knew that for such a low αj the learning and change
rate should be the same. The actual value was chosen via experimentation. The
resulting curve is shown in Figure 9(b). At this moment, we do not possess a
formal way of deriving learning and change rates from α-rates.
For the second curve (αj = .3) we knew that, since only αj had changed
from the first experiment, we should only change lj and cj.
In fact, these
two values should only be increased. We found their exact values, again by
experimentation, to be lj = .04, cj = .4. For the third curve we found the
values to be lj = .055, cj = .8.
One difference we notice between the experimental and the theoretical results
is that the experimental results show a longer delay before the error starts to
decrease. We attribute this delay to the agent's initially high exploration rate.
That is, the agents initially start by taking all random actions but progressively
reduce this rate of exploration. As the exploration rate decreases the discrepancy
between our theoretical predictions and experimental results is reduced.
In summary, while it is true that we found lj and cj by experimentation,
all the other values were calculated from the description of the problem. Even
the relative values of lj and cj follow the intuitive relation with αj that, as
αj increases so does lj and (even more) cj. Section 9 shows how to calculate
lower bounds on the learning rate. We believe that this experiment provides
solid evidence that our theory can be used to approximately determine the
quantitative behaviors of MASs with learning agents.
18
n
o
i
t
c
a
t
n
o
i
j
l
a
m
i
t
p
o
n
a
i
g
n
s
o
o
h
c
f
o
y
t
i
l
i
b
a
b
o
r
P
PSfrag replacements
time
1 - error
li = .1
1
0.95
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
0
Choosing optimal joint actions
Independent learners
Joint action learners
PSfrag replacements
r
o
r
r
e
-
1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
li = .1
5
10
15
20
25
30
35
40
45
50
5
10
15
20
30
Number of Interactions
25
35
40
45
50
(a) Experiment
time
(b) Theory
Figure 10: Comparing theory (b) with results from [6] (a).
8 Application of our Theory to Experiments in
the Literature
In this section we show how we can apply our theory to experimental results
found in the AI and MAS literature. While we will often not be able to com-
pletely reproduce the authors' results exactly, we believe that being able to
reproduce the flavor and the main quantitative characteristics of experimental
results in the literature shows that our theory can be widely applied and used
by practitioners in this area of research.
8.1 Claus and Boutilier
Claus and Boutilier [6] study the dynamics of a system that contains two re-
inforcement learning agents. Their first experiment puts the two agents in a
matching game exactly like the one we describe in Section 3.1 with Ai =
Aj = 2. Their results show the probability that both agents matched (i.e., 1
- e(δt
i )) as time progressed. Since they were using two reinforcement learning
agents, it was not surprising that the curve they saw, seen in Figure 10(a), was
nearly identical to the curve we saw in our experiments with the two buying
agents (Figure 9(a) with αj = αi = .1, except upside-down).
We can reproduce their curve using our equation for the matching game
Eq. (16). The results can be seen in Figure 10(b). Our theory again fails to
account for the initial exploration rate. We can, however, confirm that by time
15 their Boltzmann temperature (the authors used Boltzmann exploration) had
been reduced from an initial value of 16 to 3.29 and would keep decreasing by
a factor of .9 each time step. This means that by time 15 the agents were,
indeed, starting to do more exploitation (i.e., reduce their error) while doing
19
Success
3800
3600
3400
3200
PSfrag replacements
li
final error
theory
experiment
r
o
r
r
e
l
a
n
fi
PSfrag replacements
1
0.8
0.6
0.4
0.2
theory
experiment
50
100
150
200
Update delay
0
0.001
0.01
li
0.1
(a) Original Experiment
(b) Theory and Experiment
Figure 11: Comparing theory (b) with results from [30] (a).
little exploration.
8.2 Shoham and Tennenholtz
Shohan and Tennenholtz [30] investigate how learning agents might arrive at
social conventions. The authors introduce a simple learning algorithm (strategy-
selection rule) called highest cumulative reward (HCR) which their agents use for
learning these conventions. Shoham and Tennenholtz also provide the results of
a series of experiments using populations of learning agents. We try to reproduce
the results they present in their Section 4.1 where they study the "coordination
game" which is similar to our matching game, but with only two actions.
The experiment in question involves 100 agents, all of them identical and all
of them using HCR. At each time instant the agents take one of two available
actions. The aim is for every pair of chosen agents to take the same action
as each other. Agents are randomly made to form pairs. The agents update
their behavior (i.e., apply HCR) after a given delay. The authors try a series
of delays (from 0 to 200) and show that increasing the update delay decreases
the percentage of trials where, after 1600 iterations, at least 95% of the agents
reached a convention. The authors show surprise at finding this phenomenon.
Their results are reproduced in Figure 11(a) (cf. Figure 1 in their article).
The number of actions for all agents is easily set to Ai = 2, which implies that
we must have li = ci. By examining HCR, it is easy to determine that ri = 1
(i.e if an agent took the right action, it will only get more support for it). At
first intuition, one's impulse is to set Iij = 1 for every pair of agents i and j.
However, since there are 100 of them and only pairs of them interact at every
time instant, the real impact is Iij = 1/99.
We will now convert from their units of measurement into ours.
In Fig-
20
ure 11(a) we can see that their x-axis is called the update delay, which we will
refer to as d. This value is the number of time units that pass before the agent
is allowed to learn. For d = 0 the agent learns after every interaction (i.e., on
every time t), while for d = 200 the agent takes the same action for 200 time in-
stances and only learns after every 200 iterations. This means that we must set
li = 1
p(d+1) where p > 0. The value of p depends on their learning algorithm's
performance, but we know that it must be a small number (< 50) greater than
0. Through some experimentation we settled on p = 6 (other values close to
this one give similar results). Since in their graph they look at 0 ≤ d ≤ 200, we
must then look at li where
6 Finally, we find the value of d in terms
of li to be
1206 ≤ li ≤ 1
1
d =
− 1
(28)
1
pli
The y-axis of Figure 11(a) is the success, i.e., number of trials, out of 4000,
where at least 95% of the agents reached a convention. We will refer to this
value as s. We know that in s/4000 of the trials at least 95% of the agents have
error close to 0 (i.e., reaching a convention means that the agents take the right
action almost all the time), and for the rest of the trials the error was greater.
We can approximately map this to an error by saying that in s/4000 of the trials
the error was 0 (a slight underestimate), while in 1 − s/4000 of the trials the
error was 1 (a slight overestimate). We add these two up (the 0 makes the first
term disappear) and arrive at an equation that maps s to e(δt
i ).
e(δt
4000 (cid:19)
i ) ≈ (cid:18) 4000 − s
(29)
The mapping from d to s is given by their actual data. Their data can be
fit by the following function:
s = 3900 − 4d −
(d − 100)2
100
(30)
Plugging Eq. (28) into Eq. (30), and the result into Eq. (29), we finally arrive
at a function that maps their experimental results into our units:
Final error =
4000 −(cid:16)3900 − 4(1/pli − 1) −(cid:16) (1/pli−1−100)2
100
4000
(cid:17)(cid:17)
(31)
for the range
1
1206 ≤ li ≤ 1
6 .
Now that we have values for ci, li, ri, Iij , Ai, a range for li and an equation
that maps their experimental results into our units, we can plot both functions,
as seen in Figure 11(b). The x-axis was plotted on a log-scale in order to better
show the shape of the experiment curve, otherwise it would appear mostly as a
straight line. For our theory curve we used Equations (20) and (26), and iterated
for 1600 time units, just like in the experiment, and plotted the error at that
point. For the experiment curve we used Eq. (31). We plotted both of these
curves in the specified range for li. The reader will notice that our theory was
21
able to make precise quantitative predictions. The maximum distance from our
theory curve to the experimental curve is .05, which means that our predictions
for the final error were, at worst, within 5% of the experimental values. Also, an
error of about 5% was introduced when mapping from their success percentage
s to our error.
8.3 Others
There are several other examples in the literature where we believe our theory
can be successfully applied.
[16, chapter 3.7] gives results of an experiment
where two agents try to find each other in a 100 by 100 grid. He shows that if
the grid has few obstacles it is faster if both agents move towards each other,
while if there are many obstacles it is faster if one of the agents stays still
while the other one searches for it. We believe that the number of obstacles is
proportional to the change rate that the agents experience and, perhaps, to the
impact that they have on each other. When there are no obstacles the agents
never change their decision functions (because their initial Manhattan heuristics
lead them in the correct path). As the number of obstacles increases, the agents
will start to change their decision functions as they move, which will have an
impact on the other agent's target function. If, however, one of them stays put,
this means that his change rate is 0 so the other agent's target function will
stay still and he will be able to reach his target (i.e., error 0) quicker.
Notice that the problem of a moving target that Ishida studies is different
from the problem of a moving target function which we study. It is, however,
interesting to note their similarities and how our theory can be applied to some
aspects of that domain.
Another possible example is given by [29]. They show two Q-learning agents
trying to cooperate in order to move a block. The authors show how different α
rates (β in their article) affect the quality of the result that the agents converge
to. This quality roughly corresponds to our error, except for the fact that their
measurements implicitly consider some actions to be better than others, while
we consider an action to be either correct or incorrect. This discrepancy would
make it harder to apply our theory to their results but we still believe that a
rough approximation is possible. Our future work includes the extension of the
CLRI framework to handle a more general definition of error -- one that attaches
a utility to each state-action pair, rather than the simple correct/incorrect cat-
egorization we use.
9 Bounding the Learning Rate with Sample Com-
plexity
In the previous examples we have used our knowledge of the learning algorithms
to determine the values of the agent's ci, li, and ri parameters. However, there
might be cases where this is not possible -- the learning algorithm might be too
complicated or unknown. It would be useful, in these cases, to have some other
22
measure of the agent's learning abilities, which could be used to determine some
bounds on the values of these parameters.
One popular measure of the complexity of learning is given by Probably
Approximately Correct (PAC) theory [17], in the form of a measure called the
sample complexity. The sample complexity gives us a loose upper bound on
the number of examples that a consistent learning agent must observe before
arriving at a PAC hypothesis.
There are two important assumptions made by PAC-theory. The first as-
sumption is that the agents are consistent learners3. Using our notation, a
consistent learner is one who, once it has learned a correct w → a mapping does
not forget it. This simply means that the agent must have ri = 1. The second
assumption is that the agent is trying to learn a fixed concept. This assumption
makes ∆t+1
i true for all t.
i = ∆t
The sample complexity m of an agent's learning problem is given by
m ≥
1
ǫ (cid:18)ln
H
γ (cid:19) ,
(32)
where H is the size of the hypothesis space for the agent. In other words, H
is the total number of different δi(w) functions that the agent will consider. For
an agent with no previous knowledge we have H = AiW . However, agents
with previous knowledge might have smaller H, since this knowledge might be
used to eliminate impossible mappings. If a consistent learning agent has seen
m examples then, with probability at least (1 − γ), it has error at most ǫ.
While we cannot map the sample complexity m to a particular learning rate
li, we can use it to put a lower bound on the learning rate for a consistent
learning agent. That is, we can find a lower bound for the learning rate of an
agent who does not forget anything it has seen, and who is trying to learn a
fixed target function. Since the agent does not forget anything it has seen, we
can deduce that its retention rate must be ri = 1. Since the target function
is not changing, we know that Pr[∆t+1
(w) =
∆t
i(w)] = 1. We can plug these values into Eq. (19) and simplify in order to
get:
i(w)] = 0 and Pr[∆t+1
(w) 6= ∆t
i
i
We can solve the difference Eq. (33), for any time n, in order to get:
E[e(δt+1
i
)] = e(δt
i ) · (1 − li).
E[e(δn
i )] = e(δ0
i ) · (1 − li)n.
(33)
(34)
We now remember that after m time steps we expect, with probability (1 − γ),
the error to be less than ǫ. Since Eq. (34) only gives us an expected error,
not a probability distribution over errors, we cannot use it to calculate the
likelihood of the agent having that expected error. That is, we cannot calculate
the "probably" (γ) part of probably approximately correct. We will, therefore,
assume that the γ chosen for m is small enough so that it will be safe to say
that, after m time steps, the error is less than ǫ. In a typical application one
3See [24, p162] for a formal definition of a consistent learner.
23
uses a small γ because it guarantees a high degree of certainty on the upper
bound of the error.
Since we can now safely say that, after m time steps, the error is less than ǫ,
we can then deduce that the li for this agent should be small enough such that,
if n = m, then E[e(δn
i )] ≤ ǫ. This is expressed mathematically as:
e(δ0
i ) · (1 − li)m ≤ ǫ.
We solve this equation for li in order to get:
li ≥ 1 −(cid:18) ǫ
e(δ0
i )(cid:19)1/m
(35)
(36)
.
This equation is not defined for e(δ0
i ) = 0. However, given our assumption of
a fixed target function and ri = 1, we already know, from Eq. (33), that if an
agent starts with an error of 0 it will maintain this error of 0 for any future time
t > 0. Therefore, in this case, the choice of a learning rate has no bearing on
the agent's error, which will always be 0.
Equation (36) gives us a lower bound on the learning rate that a consistent
learner must have, given that it has sample complexity m, and based on an
error ǫ and a sufficiently small γ. A designer of an agent that uses a reason-
able learning algorithm can expect that, if his agent has sample complexity m
(for ǫ error), then his agent will have a learning rate of at least li, as given by
Eq. (36). Furthermore, if a designer is comparing two possible agent designs,
each with a different sample complexity but both with similarly powerful learn-
ing algorithms, he can calculate bounds on the learning rates of both agents
and compare their relative performance.
10 Related Work
The topic of agents learning about agents arises often in the studies of com-
In fact, systems where the agents try to adapt to endogenously
plexity [7].
created dynamics are being widely studied [15, 2].
In these systems, like in
ours, the agents co-create their expectations as they learn and change their be-
haviors. Complexity research uses simulated agents in an effort to understand
the complex behaviors of these systems as observed in the real world.
One example is the work of Arthur et. al. [2], who arrive at the conclusion
that systems of adaptive agents, where the agents are allowed to change the
complexity of their learning algorithms, end up in one of two regimes: a sta-
ble/simple regime where it is trivial to predict an agent's future behavior, and a
complex regime where the agents' behaviors are very complex. It is this second
regime that interests complexity researchers the most. In it, the agents are able
to reach some kind of "equilibrium" point in model building complexity. These
same results are echoed by Darley and Kauffman [8] in a similar experiment. In
this article we have not allowed the agents to dynamically change the complex-
ity of their learning algorithms. Therefore, our dynamics are simpler. Allowing
24
the agents to change their complexity amounts to allowing them to change the
values of their c, l, and r parameters while learning.
However, while complexity research is very important and inspiring, it is only
partially relevant to our work. Our emphasis is on finding ways to predict the
behavior of MASs composed of machine-learning agents. We are only concerned
with the behavior of simpler artificial programmable agents, rather than the
complex behavior of humans or the unpredictable behavior of animals.
The dynamics of MASs have also been studied by Kephart et. al. [18]. In
this work the authors show how simple predictive agents can lead to globally
cyclic or chaotic behaviors. As the authors explain, the chaotic behaviors were a
result of the simple predictive strategies used by the agents. Unlike our agents,
most of their agents are not engaged in learning, instead they use simple fixed
predictive strategies, such as "if the state of the world was x ten time units
before, then it will be x next time so take action a". The authors later show
how learning can be used to eliminate these chaotic global fluctuations.
Matari´c [22] has studied reinforcement learning in multi-robot domains. She
notes, for example, how learning can give rise to social behaviors [23]. The
work shows how robots can be individually programmed to produce certain
group behaviors. It represents a good example of the usefulness and flexibility
of learning agents in multi-agent domains. However, the author does not offer
a mathematical justification for the chosen individual learning algorithms, nor
does she explain why the agents were able to converge to the global behaviors.
Our research hopes to provide the first steps in this direction.
One particularly interesting approach is taken by Carmel and Markovitch
[4]. They work on model-based learning, that is, agents build models of other
agents via observations. They use models based on finite state machines. The
authors show how some of these models can be effectively learned via observation
of the other agent's actions. The authors concentrate on the development of
learning algorithms that would let one agent learn a finite-state machine model
of another agent. They have not considered the case where two or more agents
are simultaneously learning about each other, which we study in this article.
However, their work is more general in the sense that they model agents as
state machines, rather than the state-action pairs we use.
Finally, a lot of experimental work has been done in the area of agents
learning about agents [27, 36]. For example, Sen and Sekaran [28] show how
learning agents in simple MAS converge to system-wide optimal behavior. Their
agents use Q-learning or modified classifier systems in order to learn. The
authors implement these agents and compare the performance of the different
learning algorithms for developing agent coordination. Hu and Wellman [14, 12]
have studied reinforcement learning in market-base MASs, showing how certain
initial learning biases can be self-fulfilling, and how learning can be useful but
is affected by an agent's models of other agents. Claus and Boutilier [6] have
also carried out experimental studies of the behavior of reinforcement learning
agents. We have been able to use the CLRI framework to predict some of their
experimental results [33]. Other researchers [32, 20, 13] have extended the basic
Q-learning [35] algorithm for use with MASs in an effort to either improve or
25
prove convergence to the optimal behavior.
We have also successfully experimented with reinforcement learning simula-
tions [34], but we believe that the formal treatment elucidated in these pages
will shed more light into the real nature of the problem and the relative im-
portance of the various parameters that describe the capabilities of an agent's
learning algorithm.
11 Limitations and Future Work
The CLRI framework places some constraints on the type of systems it can
model, which limits its usability. However, it is important to understand that,
as we remove the limitations from the CLRI framework, the dynamics of the
system become much harder to predict. In the extreme, without any limitations
on the agents' abilities, the system becomes a complex adaptive system, as
studied by Holland [11] and others in the field of complexity. The dynamic
behavior of these systems continues to be studied by complexity researchers
with only modest progress. It is only by placing limitations on the system that
we were able to predict the expected error of agents in the systems modeled by
the CLRI framework.
Our ongoing work involves the relaxation of some of the constraints made by
the CLRI framework so that it may become more easily and widely applicable,
without making the system dynamics impossible to analyze. We are targeting
three specific constraints.
1. The values of ci, li, ri, and Iij cannot, in all situations, be mathematically
determined from the system's description. We have found that bounds for
the ci, li, and ri values can often be determined when using reinforcement
learning or supervised learning. However, the bounds are often very loose.
The values of the Iij parameter depend on the particular system. Some-
times it is trivial to calculate the impact, sometimes it requires extensive
simulation.
2. The CLRI framework assumes that an agent's action is either correct or
incorrect. The framework does not allow degrees of correctness. Specifi-
cally, in many systems the agents can often take several actions, any one
of which is equally good. When modeling these systems, the CLRI frame-
work requires the user to designate one of those actions as the correct one,
thereby ignoring some possibly useful information.
3. The world states are taken from a uniform probability distribution which
does not change over time. The environment is assumed to be episodic.
As such, the framework is limited in the type of domains it can effectively
describe.
We are attacking these challenges with some of the same tools used by re-
searchers in complex adaptive systems, namely, agent-based simulations and
26
co-evolving utility landscapes. We believe we can gain some insight into the dy-
namics of adaptive MASs by constructing and analyzing various types of MASs.
We also believe that the next step for the CLRI framework is the replacement of
the current error definition with a utility function. The agents can then be seen
as searching for the maximum value in the changing utility landscape defined
by their utility function. The degree to which the agents are successful on their
climb to the landscape peaks depends on the abilities of their learning algo-
rithm (change rate, learning rate, and retention rate), and the speed at which
the landscape changes as the other agents change their behavior (impact).
The use of utility landscapes will allow us to consider an agent's utility
for any particular action, rather than simply considering whether an action is
correct or incorrect. The landscapes will also allow us to consider systems where
agents cannot travel between any two world states in one time step. That is, the
agents' moves on the landscape will be constrained in the same manner as their
actions or behaviors are constrained the actual system. Finally, the new theory
will likely need to redefine the CLRI parameters. We hope the new parameters
will be easy to derive directly from the values that govern the machine-learning
algorithms' behavior. These extensions will make the new theory applicable to
a much wider set of domains.
12 Summary
We have presented a framework for studying and predicting the behavior of
MASs composed of learning agents. We believe that this framework captures the
most important parameters that describe an agents' learning and the system's
rules of encounter. Various comparisons between the framework's predictions
and experimental results were given. These comparisons showed that the theo-
retical predictions closely match our experimental results and the experimental
results published by others. Our success in reproducing these results allows us
to confidently state the effectiveness and accuracy of our theory in predicting
the expected error of machine learning agents in MASs.
Since our theory describes an agent's behavior at a high-level (i.e., the agent's
error), it is not capable of making system-specific predictions (e.g., predicting
the particular actions that are favored). These types of system-specific predic-
tions can only be arrived at by the traditional method of implementing popula-
tions of such agents and testing their behaviors. However, we expect that there
will be times when the predictions from our theory will be enough to answer a
designer's questions. A MAS designer that only needs to determine how "good"
the agent's behavior will be could probably use the CLRI framework. A de-
signer that needs to know which particular emergent behaviors will be favored
by his agents will need to implement the agents.
Finally, while we have given some examples as to how learning rates can be
determined for particular machine learning implementations, we do not have
any general method for determining these rates. However, we showed how to
use the sample complexity of a learning problem to determine a lower bound on
27
the learning rate of a consistent learning agent. This bound is useful for quickly
ruling out the possibility of having agents with high expected errors and of
stating that an agent's expected error will be, at most, a certain constant value.
Still, if the agent's learning algorithm is much better than the one assumed by
a consistent learner (e.g., the agent is very good at generalizing from one world
state to many others), then these lower bounds could be significantly inaccurate.
A Derivation for Matching Game
If we can assume that the action chosen when an agent changes δt
result does not match ∆t
distribution, then we can say that:
i (w) and the
i(w) (for some specific w) is taken from a flat probability
B =
Ai − 2
Ai − 1
.
(37)
We will now show how to calculate the fourth term in (16). For the matching
game we find that we can set:
D = 1 − lj
F = lj + (1 − lj)(cid:18) Ai − 3
Ai − 2(cid:19) .
(38)
(39)
Having Ai = 2 implies that ci = li, this means that for this case we have
(1 − ci)D + li + (ci − li)F = li + (1 − ci)(1 − lj).
(40)
For the case where Ai > 2, which is the case we are interested in, we can
plug in the values for D and F and simplify, in order to get the fourth term:
(1 − ci)D + li + (ci − li)F = 1 − lj +
cilj(Ai − 1) + li(1 − lj) − ci
Ai − 2
.
(41)
References
[1] W. Brian Arthur, Steven Durlauf, and David Lane, editors. The Economy
as an Evolving Complex System II. Series in the Sciences of Complexity.
Addison-Wesley, Reading, MA, 1997.
[2] W. Brian Arthur, John H. Holland, Blake LeBaron, Richard Palmer, and
Paul Tayler. Asset pricing under endogenous expectations in an artificial
stock market. In Arthur et al. [1].
[3] Ken Binmore. Modeling rational players, part 2. Economics and Philoso-
phy, 4:9 -- 55, 1988.
[4] David Carmel and Shaul Markovitch. Exploration and adaptation in mul-
tiagent systems: A model-based approach. In Proceedings of the Interna-
tional Joint Conference on AI, 1997.
28
[5] A. Chavez and P. Maes. Kasbah: An agent marketplace for buying and sell-
ing goods. In First International Conference on the Practical Application
of Intelligent Agents and Multi-Agent Technology, pages 75 -- 90, 1996.
[6] Caroline Claus and Craig Boutilier. The dynamics of reinforcement learn-
In Proceedings of Workshop on
ing in cooperative multiagent systems.
Multiagent Learning. AAAI Press, 1997.
[7] George A. Cowan, David Pines, and David Meltzer, editors. Complexity:
Metaphors, Models and Reality. Addison-Wesley, Reading, MA, 1995.
[8] Vince M. Darley and Stuart Kauffman. Natural rationality. In Arthur et al.
[1].
[9] Edmund H. Durfee, Daniel L. Kiskis, and William P. Birmingham. The
agent architecture of the University of Michigan Digital Library. IEE Pro-
ceedings on Software Engineering, 144(1):61 -- 71, 1997.
[10] Oren Etzioni. Moving up the information food chain: Deploying softbots
on the world wide web.
In Proceedings of the Thirteenth National Con-
ference on Artificial Intelligence and the Eighth Innovative Applications of
Artificial Intelligence Conference, pages 1322 -- 1326, Menlo Park, August4 --
8 1996. AAAI Press / MIT Press.
[11] John H. Holland. Hidden Order. Addison-Wesley, 1995.
[12] Junling Hu and Michael P. Wellman. Self-fulfilling bias in multiagent learn-
ing. In Proceedings of the Second International Conference on Multi-Agent
Systems, pages 118 -- 125, 1996.
[13] Junling Hu and Michael P. Wellman. Multiagent reinforcement learning:
Theoretical framework and an algorithm. In Proceedings of the Fifteenth
International Conference on Machine Learning, 1998.
[14] Junling Hu and Michael P. Wellman. Online learning about other agents in
a dynamic multiagent system. In Proceedings of the Second International
Conference on Autonomous Agents, 1998.
[15] Alfred Hubler and David Pines. Complexity: Methaphors, Models and Re-
ality, chapter Prediction and Adaptation in an Evolving Chaotic Environ-
ment, pages 343 -- 379. In Cowan et al. [7], 1994.
[16] Toru Ishida. Real-Time Search for Learnig Autonomous Agents. Kluwer
Academic Publishers, 1997.
[17] Michael J. Kearns and Umesh V. Vazirani. An Introduction to Computa-
tional Learning Theory. The MIT Press, Cambridge, MA, 1994.
[18] Jeffery O. Kephart, Tad Hogg, and Bernardo A. Huberman. Collective
behavior of predictive agents. Physica D, 42(2):48 -- 65, 1990.
29
[19] Victor R. Lesser and Daniel D. Corkill. Functionally accurate, cooperative
distributed systems. IEEE Transactions on Systems, Man, and Cybernet-
ics, 11(1):81 -- 96, January 1981.
[20] Michael L Littman. Markov games as a framework for multi-agent rein-
forcement learning. In Proceedings of the Eleventh International Conference
on Machine Learning, pages 157 -- 163. Morgan Kaufmann, 1994.
[21] Jyi-Shane Liu and Katia P. Sycara. Multiagent coordination in tightly
In Victor Lesser, editor, Proceedings of
coupled real-time environments.
the First International Conference on Multi-Agent Systems. MIT Press,
1995.
[22] Maja J. Matari´c.
Issues and approaches in the design of collective au-
tonomous agents. Robotics and Autonomous Systems, 16(2-4):321 -- 331, Dec
1995.
[23] Maja J. Matari´c. Learning social behaviors. Robotics and Autonomous
Agents, 20:191 -- 204, 1997.
[24] Tom M. Mitchell. Machine Learning. McGraw-Hill, 1997.
[25] Jeffrey S. Rosenschein and Gilad Zlotkin. Rules of Encounter. The MIT
Press, Cambridge, MA, 1994.
[26] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Ap-
proach. Prentice Hall, 1995.
[27] Sandip Sen, editor. Working Notes from the AAAI Symposium on Adap-
tation, Co-evolution and Learning in Multiagent Systems, 1996.
[28] Sandip Sen and Mahendra Sekaran.
Individual learning of coordination
knowledge. Journal of Experimental and Theoretical AI, pages 156 -- 170,
1998.
[29] Sandip Sen, Mahendra Sekaran, and John Hale. Learning to coordinate
without sharing information. In Proceedings of the Twelfth National Con-
ference on Artificial Intelligence, 1994.
[30] Yoav Shoham and Moshe Tennenholtz. On the emergence of social con-
ventions: modeling, analysis, and simulations. Artificial Intelligence,
94(1):139 -- 166, 1997.
[31] Peter Stone and Manuela Velos. Multiagent systems: A survery from a
machine learning perspective. Technical Report CMU-CS-97-193, Carnegie
Mellon University, December 1997.
[32] Peter Stone and Manuela Veloso. Team-partitioned, opaque-transition re-
inforcement learning. In Proceedings of the Third International Conference
on Autonomous Agents. Springer-Verlag, 1999.
30
[33] Jos´e M. Vidal. Computational Agents That Learn About Agents: Algorithms
for Their Design and a Predictive Theory of Their Behavior. PhD thesis,
University of Michigan, 1998.
[34] Jos´e M. Vidal and Edmund H. Durfee. Learning nested models in an in-
formation economy. Journal of Experimental and Theoretical Artificial In-
telligence, 10(3):291 -- 308, 1998.
[35] Christopher J. Watkins and Peter Dayan. Q-learning. Machine Learning,
8:279 -- 292, 1992.
[36] Gerhard Weiss, editor. Distributed Artificial Intelligence meets Machine
Learning. Springer, 1997.
31
|
0903.2282 | 1 | 0903 | 2009-03-12T21:49:36 | Multiagent Learning in Large Anonymous Games | [
"cs.MA",
"cs.GT",
"cs.LG"
] | In large systems, it is important for agents to learn to act effectively, but sophisticated multi-agent learning algorithms generally do not scale. An alternative approach is to find restricted classes of games where simple, efficient algorithms converge. It is shown that stage learning efficiently converges to Nash equilibria in large anonymous games if best-reply dynamics converge. Two features are identified that improve convergence. First, rather than making learning more difficult, more agents are actually beneficial in many settings. Second, providing agents with statistical information about the behavior of others can significantly reduce the number of observations needed. | cs.MA | cs | Multiagent Learning in Large Anonymous Games
Ian A. Kash
Computer Science Dept.
Cornell University
[email protected]
Eric J. Friedman
School of Operations
Research and Information
Engineering
Cornell University
[email protected]
Joseph Y. Halpern
Computer Science Dept.
Cornell University
[email protected]
9
0
0
2
r
a
M
2
1
]
A
M
.
s
c
[
1
v
2
8
2
2
.
3
0
9
0
:
v
i
X
r
a
ABSTRACT
In large systems, it is important for agents to learn to act ef-
fectively, but sophisticated multi-agent learning algorithms
generally do not scale. An alternative approach is to find re-
stricted classes of games where simple, efficient algorithms
converge.
It is shown that stage learning efficiently con-
verges to Nash equilibria in large anonymous games if best-
reply dynamics converge. Two features are identified that
improve convergence. First, rather than making learning
more difficult, more agents are actually beneficial in many
settings. Second, providing agents with statistical informa-
tion about the behavior of others can significantly reduce
the number of observations needed.
Categories and Subject Descriptors
I.2.11 [Artificial Intelligence]: Distributed Artificial Intel-
ligence -- Multiagent systems; J.4 [Social and Behavioral
Sciences]: Economics
General Terms
Algorithms, Economics, Theory
Keywords
Multiagent Learning, Game Theory, Large Games, Anony-
mous Games, Best-Reply Dynamics
1.
INTRODUCTION
Designers of distributed systems are frequently unable to
determine how an agent in the system should behave, be-
cause optimal behavior depends on the user's preferences
and the actions of others. A natural approach is to have
agents use a learning algorithm. Many multiagent learning
algorithms have been proposed including simple strategy up-
date procedures such as fictitious play [10], multiagent ver-
sions of Q-learning [25], and no-regret algorithms [5].
However, as we discuss in Section 2, existing algorithms
are generally unsuitable for large distributed systems. In a
distributed system, each agent has a limited view of the ac-
tions of other agents. Algorithms that require knowing, for
Cite as: Title, Author(s), Proc. of 8th Int. Conf. on Au-
tonomous Agents and Multiagent Systems (AAMAS 2009),
Decker, Sichman, Sierra and Castelfranchi (eds.), May, 10 -- 15, 2009, Bu-
dapest, Hungary, pp. XXX-XXX.
Copyright c(cid:13) 2009, International Foundation for Autonomous Agents and
Multiagent Systems (www.ifaamas.org). All rights reserved.
1
example, the strategy chosen by every agent cannot be im-
plemented. Furthermore, the size of distributed systems re-
quires fast convergence. Users may use the system for short
periods of time and conditions in the system change over
time, so a practical algorithm for a system with thousands
or millions of users needs to have a convergence rate that is
sublinear in the number of agents. Existing algorithms tend
to provide performance guarantees that are polynomial or
even exponential. Finally, the large number of agents in the
system guarantees that there will be noise. Agents will make
mistakes and will behave in unexpectedly. Even if no agent
changes his strategy, there can still be noise in agent payoffs.
For example, a gossip protocol will match different agents
from round to round; congestion in the underlying network
may effect message delays between agents. A learning algo-
rithm needs to be robust to this noise.
While finding an algorithm that satisfies these require-
ments for arbitrary games may be difficult, distributed sys-
tems have characteristics that make the problem easier. First,
they involve a large number of agents. Having more agents
may seem to make learning harder -- after all, there are more
possible interactions. However, it has the advantage that the
outcome of an action typically depends only weakly on what
other agents do. This makes outcomes robust to noise. Hav-
ing a large number of agents also make it less useful for an
agent to try to influence others; it becomes a better policy to
try to learn an optimal response. In contrast, with a small
number of agents, an agent can attempt to guide learning
agents into an outcome that is beneficial for him.
Second, distributed systems are often anonymous [1]; it
does not matter who does something, but rather how many
agents do it. For example, when there is congestion on a link,
the experience of a single agent does not depend on who is
sending the packets, but on how many are being sent.
Finally, and perhaps most importantly, in a distributed
system the system designer controls the game agents are
playing. This gives us a somewhat different perspective than
most work, which takes the game as given. We do not need
to solve the hard problem of finding an efficient algorithm
for all games.
Instead, we can find algorithms that work
efficiently for interesting classes of games, where for us "in-
teresting" means "the type of games a system designer might
wish agents to play." Such games should be "well behaved,"
since it would be strange to design a system where an agent's
decisions can influence other agents in pathological ways.
In Section 3, we show that stage learning [9] is robust,
implementable with minimal information, and converges ef-
ficiently for an interesting class of games. In this algorithm,
agents divide the rounds of the game into a series of stages.
In each stage, the agent uses a fixed strategy except that
he occasionally explores. At the end of a stage, the agent
chooses as his strategy for the next stage whatever strategy
had the highest average reward in the current stage. We
prove that, under appropriate conditions, a large system of
stage learners will follow (approximate) best-reply dynamics
despite errors and exploration.
For games where best-reply dynamics converge, our theo-
rem guarantees that learners will play an approximate Nash
equilibrium. In contrast to previous results where the con-
vergence guarantee scales poorly with the number of agents,
our theorem guarantees convergence in a finite amount of
time with an infinite number of agents. While the assump-
tion that best-reply dynamics converge is a strong one, many
interesting games converge under best-reply dynamics, in-
cluding dominance solvable games and games with mono-
tone best replies. Marden et al. [17] have observed that
convergence of best-reply dynamics is often a property of
games that humans design. Moreover, convergence of best-
reply dynamics is a weaker assumption than a common as-
sumption made in the mechanism design literature, that the
games of interest have dominant strategies (each agent has
a strategy that is optimal no matter what other agents do).
Simulation results, presented in Section 4, show that con-
vergence is fast in practice: a system with thousands of
agents can converge in a few thousand rounds. Further-
more, we identify two factors that determine the rate and
quality of convergence. One is the number of agents: having
more agents makes the noise in the systen more consistent
so agents can learn using fewer observations. The other
is giving agents statistical information about the behavior
of other agents; this can speed convergence by an order of
magnitude. Indeed, even noisy statistical information about
agent behavior, which should be relatively easy to obtain
and disseminate, can significantly improve performance.
2. RELATED WORK
One approach to learning to play games is to generalize
reinforcement learning algorithms such as Q-learning [25].
One nice feature of this approach is that it can handle games
with state, which is important in distributed systems.
In
Q-learning, an agent associates a value with each state-
action pair. When he chooses action a in state s, he up-
dates the value Q(s, a) based on the reward he received
and the best value he can achieve in the resulting state s′
(maxa′ Q(s′, a′)). When generalizing to multiple agents, s
and a become vectors of the state and action of every agent
and the max is replaced by a prediction of the behavior
of other agents. Different algorithms use different predic-
tions; for example, Nash-Q uses a Nash equilibrium calcula-
tion [15]. See [22] for a survey.
Unfortunately, these algorithms converge too slowly for
a large distributed system. The algorithm needs to expe-
rience each possible action profile many times to guarantee
convergence. So, with n agents and k strategies, the naive
convergence time is O(kn). Even with a better represen-
tation for anonymous games, the convergence time is still
O(nk) (typically k ≪ n). There is also a more fundamen-
tal problem with this approach: it assumes information that
an agent is unlikely to have. In order to know which value
to update, the agent must learn the action chosen by ev-
ery other agent. In practice, an agent will learn something
2
about the actions of the agents with whom he directly in-
teracts, but is unlikely to gain much information about the
actions of other agents.
Another approach is no-regret learning, where agents choose
a strategy for each round that guarantees that the regret of
their choices will be low. Hart and Mas-Colell [13] present
such a learning procedure that converges to a correlated equi-
librium [21] given knowledge of what the payoffs of every
action would have been in each round. They also provide
a variant of their algorithm that requires only information
about the agent's actual payoffs [14]. However, to guarantee
convergence to within ǫ of a correlated equilibrium requires
O(kn/ǫ2 log kn), still too slow for large systems. Further-
more, the convergence guarantee is that the distribution of
play converges to equilibrium; the strategies of individual
learners will not converge. Better results can be achieved
in restricted settings. For example, Blum et al. [2] showed
that in routing games a continuum of no-regret learners will
approximate Nash equilibrium in a finite amount of time.
Foster and Young [7] use a stage-learning procedure that
converges to Nash equilibrium for two-player games. Ger-
mano and Lugosi [11] showed that it converges for generic n-
player games (games where best replies are unique). Young [26]
uses a similar algorithm without explicit stages that also
converges for generic n-player games. Rather than selecting
best replies, in these algorithms agents choose new actions
randomly when not in equilibrium. Unfortunately, these
algorithms involve searching the whole strategy space, so
their convergence time is exponential. Another algorithm
that uses stages to provide a stable learning environment is
the ESRL algorithm for coordinated exploration [24].
Marden et al. [18, 19] use an algorithm with experimen-
tation and best replies but without explicit stages that con-
verges for weakly acyclic games, where best-reply dynamics
converge when agents move one at a time, rather than mov-
ing all at once, as we assume here. Convergence is based on
the existence of a sequence of exploration moves that lead
to equilibrium. With n agents who explore with probability
ǫ, this analysis gives a convergence time of O(1/ǫn). Fur-
thermore, the guarantee requires ǫ to be sufficiently small
that agents essentially explore one at a time, so ǫ needs to
be O(1/n).
There is a long history of work examining simple learn-
ing procedures such as fictitious play [10], where each agent
makes a best response assuming that each other player's
strategy is characterized by the empirical frequency of his
observed moves. In contrast to algorithms with convergence
guarantees for general games, these algorithms fail to con-
verge in many games. But for classes of games where they
do converge, they tend to do so rapidly. However, most work
in this area assumes that the actions of agents are observed
by all agents, agents know the payoff matrix, and payoffs are
deterministic. A recent approach in this tradition is based
on the Win or Learn Fast principle, which has limited con-
vergence guarantees but often performs well in practice [4].
There is also a body of empirical work on the convergence
of learning algorithms in multiagent settings. Q-learning has
had empirical success in pricing games [23], n-player coop-
erative games [6], and grid world games [3]. Greenwald at
al. [12] showed that a number of algorithms, including stage
learning, converge in a variety of simple games. Marden et
al. [19] found that their algorithm converged must faster in a
congestion game than the theoretical analysis would suggest.
Our theorem suggests an explanation for these empirical ob-
servations: best-reply dynamics converge in all these games.
While our theorem applies directly only to stage learning, it
provides intuition as to why algorithms that learn "quickly
enough" and change their behavior "slowly enough" rapidly
converge to Nash equilibrium in practice.
u) be Lipschitz continuous.1 For definiteness, we use
the L1 norm as our notion of distance when specifying
continuity (the L1 distance between two vectors is the
sum of the absolute values of the differences in each
component). Note that this formulation assumes all
agents share a common utility function.
3. THEORETICAL RESULTS
3.1 Large Anonymous Games
We are interested in anonymous games with countably
many agents. Assuming that there are countably many
agents simplifies the proofs; it is straightforward to extend
our results to games with a large finite number of agents.
Our model is adapted from that of [1]. Formally, a large
anonymous game is characterized by a tuple Γ = (N, A, P, Pr).
• N is the countably infinite set of agents.
• A is a finite set of actions from which each agent can
choose (for simplicity, we assume that each agent can
choose from the same set of actions).
• ∆(A), the set of probability distributions over A, has
two useful interpretations. The first is as the set of
mixed actions. For a ∈ A we will abuse notation and
denote the mixed action that is a with probability 1 as
a. In each round each agent chooses one of these mixed
actions. The second interpretation of ρ ∈ ∆(A) is as
the fraction of agents choosing each action a ∈ A. This
is important for our notion of anonymity, which says
an agent's utility should depend only on how many
agents choose each action rather than who chooses it.
• G = {g : N → ∆(A)} is the set of (mixed) action
profiles (i.e. which action each agent chooses). Given
the mixed action of every agent, we want to know the
fraction of agents that end up choosing action a. For
g ∈ G, let g(i)(a) denote the probability with which
agent i plays a according to g(i) ∈ ∆(A). We can then
express the fraction of agents in g that choose action
a as limn→∞(1/n) Pn
i=0 g(i)(a), if this limit exists. If
the limit exists for all actions a ∈ A, let ρg ∈ ∆(A)
give the value of the limit for each a. The profiles g
that we use are all determined by a simple random
process. For such profiles g, the strong law of large
numbers (SLLN) guarantees that with probability 1 ρg
is well defined. Thus it will typically be well defined
(using similar limits) for us to talk about the fraction
of agents who do something.
• P ⊂ R is a finite set of payoffs agents can receive.
• Pr : A × ∆(A) → ∆(P ) denotes the distribution over
payoffs that results when the agent performs action
a and other agents follow action profile ρ. We use a
probability distribution over payoffs rather than a pay-
off to model the fact that agent payoffs may change
even if no agent changes his strategy. The expected
utility of an agent who performs mixed action s when
other agents follow action distribution ρ is u(s, ρ) =
Pa∈A Pp∈P ps(a) Pra,ρ(p). Our definition of Pr in
terms of ∆(A) rather than G ensures the the game
is anonymous. We further require that Pr (and thus
An example of a large anonymous game is one where, in
each round, each agent plays a two-player game against an
opponent chosen at random. Then A is the set of actions
of the two-player game and P is the set of payoffs of the
game. Once every agent chooses an action, the distribution
over actions is characterized by some ρ ∈ ∆(A). Let pa,a′
denote the payoff for the agent if he plays a and the other
agent plays a′. Then the utility of mixed action s given
distribution ρ is
u(s, ρ) = X
s(a)ρ(a′)pa,a′.
a,a′∈A2
3.2 Best-Reply Dynamics
Given a game Γ and an action distribution ρ, a natural
goal for an agent is to play the action that maximizes his
expected utility with respect to ρ: argmaxa∈A u(a, ρ). We
call such an action a best reply to ρ. In a practical amount
of time, an agent may have difficulty determining which of
two actions with close expected utilities is better, so we will
allow agents to choose actions that are close to best replies.
If a is a best reply to ρ, then a′ is an η-best reply to ρ if
u(a′, ρ) + η ≥ u(a, ρ). There may be more than one η-best
reply; we denote the set of η-best replies ABRη(ρ).
We do not have a single agent looking for a best reply;
every agent is trying to find a one at the same time.
If
agents start off with some action distribution ρ0, after they
all find a best reply there will be a new action distribution ρ1.
We assume that ρ0(a) = 1/A (agents choose their initial
strategy uniformly at random), but our results apply to any
distribution used to determine the initial strategy. We say
that a sequence (ρ0, ρ1, . . .) is an η-best-reply sequence if the
support of ρi+1 is a subset of ABRη(ρi); that is ρi+1 gives
positive probability only to approximate best replies to ρi.
A η best-reply sequence converges if there exists some t such
that for all t′ > t, ρt′ = ρt. Note that this is a particularly
strong notion of convergence because we require the ρt to
converge in finite time and not merely in the limit. A game
may have infinitely many best-reply sequences, so we say
that approximate best-reply dynamics converge if there exists
some η > 0 such that every η-best-reply sequence converges.
The limit distribution ρt determines a mixed strategy that
is an η-Nash equilibrium.
Our theorem shows that learners can successfully learn in
large anonymous games where approximate best-reply dy-
namics converge. The number of stages needed to converge
is determined by the number of best replies needed before
the sequence converges. It is possibly to design games that
have long best-reply sequences, but it practice most games
have short sequences. One condition that guarantees this is
if ρ0 and all the degenerate action distributions a ∈ A (i.e.,
1Lipschitz continuity imposes the additional constraint that
there is some constant K such that Pr(a, ρ)−Pr(a, ρ′)/ρ−
ρ′1 ≤ K for all ρ and ρ′. Intuitively, this ensures that the
distribution of outcomes doesn't change "too fast." This is a
standard assumption that is easily seen to hold in the games
that have typically been considered in the literature.
3
distributions that assign probability 1 to some a ∈ A) have
unique best replies. In this case, there can be at most A
best replies before equilibrium is reached. Furthermore, in
such games the distinction between η-best replies and best
replies is irrelevant; for sufficiently small η, a η-best reply is
a best reply. It is not hard to show that the property that
degenerate strategies have unique best replies is generic; it
holds for almost every game.
3.3 Stage Learners
An agent who wants to find a best reply may not know the
set of payoffs P , the mapping from actions to distributions
over payoffs Pr, or the action distribution ρ (and, indeed, ρ
may be changing over time), so he will have to use some type
of learning algorithm to learn it. Our approach is to divide
the play of the game into a sequence of stages. In each stage,
the agent almost always plays some fixed action a, but also
explores other actions. At the end of the stage, he chooses a
new a′ for the next stage based on what he has learned. An
important feature of this approach is that agents maintain
their actions for the entire stage, so each stage provides a
stable environment in which agents can learn. To simplify
our results, we specify a way of exploring and learning within
a stage (originally described in [9]), but our results should
generalize to any "reasonable" learning algorithm used to
learn within a stage. (We discuss what is "reasonable" in
Section 5.) In this section, we show that, given a suitable
parameter, at the each stage most agents will have learned
a best reply to the environment of that stage.
Given a game Γ, in each round t agent i needs to se-
lect a mixed action si,t. Our agents use strategies that we
denote aǫ, for a ∈ A, where aǫ(a) = 1 − ǫ and aǫ(a′
6=
a) = ǫ/(A − 1). Thus, with aǫ, an agent almost always
plays a, but with probability ǫ explores other strategies uni-
formly at random. Thus far we have not specified what
information an agent can use to choose si,t. Different games
may provide different information. All that we require is
that an agent know all of his previous actions and his pre-
vious payoffs. More precisely, for all t′ < t, he knows his
action at′ (i) (which is determined by si,t′ ) and his payoffs
pt′ (i) (which is determined by Pr(ai,t′ , ρt′), where ρt′ is the
action distribution for round t′; note that we do not as-
sume that the agent knows ρt′ .) Using this information, we
can express the average value of an action over the pre-
vious τ = ⌈1/ǫ2⌉ rounds (the length of a stage).2 Let
H(a, i, t) = {t − τ ≤ t′ < t at′ (i) = a} be the set of recent
rounds in which a was played by i. Then the average value
is V (a, i, t) = Pt′ ∈H(a,i,t) pt′ (i)/H(a, i, t) if H(a, i, t) > 0
and 0 otherwise. While we need the value of H only at
times that are multiples of τ , for convenience we define it
for arbitrary times t.
We say that an agent is an ǫ-stage learner if he chooses
his actions as follows. If t = 0, st is chosen at random from
{aǫ a ∈ A}. If t is a nonzero multiple of τ , si,t = a(i, t)ǫ
where a(i, t) = argmaxa∈A V (a, i, t). Otherwise, si,t = si,t−1.
Thus, within a stage, his mixed action is fixed and at the end
of a stage he updates it to use the action with the highest
average value during the previous stage.
The evolution of a game played by stage learners is not
deterministic; each agent chooses a random si,0 and the se-
2The use of the exponent 2 is arbitrary. We require only
that the expected number of times a strategy is explored
increases as ǫ decreases.
quence of at(i) and pt(i) he observes is also random. How-
ever, with a countably infinite set of agents, we can use the
SLLN to make statements about the overall behavior of the
game. Let gt(i) = si,t. A run of the game consists of a
sequence of triples (gt, at, pt). The SLLN guarantees that
with probability 1 the fraction of agents who choose a strat-
egy a in at is ρgt (a). Similarly, the fraction of agents who
chose a in at that receive payoff p will be Pr(a, ρgt )(p) with
probability 1.
To make our notion of a stage precise, we refer to the se-
quence of tuples (gnτ , anτ , pnτ ) . . . (g(n+1)τ −1, a(n+1)τ −1, p(n+1)τ −1)
as stage n of the run. During stage n there is a stationary
action distribution that we denote ρgnτ . If si,(n+1)τ = aǫ
and a ∈ ABRη(gnτ ), then we say that agent i has learned
an η-best reply during stage n of the run. As the following
lemma shows, for sufficiently small ǫ, most agents will learn
an η-best reply.
Lemma 3.1. For all
large anonymous games Γ, action
profiles, approximations η > 0, and probabilities of error
e > 0, there exists an ǫ∗ > 0 such that for ǫ < ǫ∗ and all n,
if all agents are ǫ-stage learners, then at least a 1−e fraction
of agents will learn an η-best reply during stage n.
Proof. (Sketch) On average, an agent using strategy aǫ
plays action a (1 − ǫ)τ times during a stage and plays all
other actions ǫτ /(n − 1) times each. For τ large, the realized
number of times played will be close to the expectation value
with high probability. Thus, if ǫτ is sufficiently large, then
the average payoff from each action will be exponentially
close to the true expected value (via a standard Hoeffding
bound on sums of i.i.d. random variables), and thus each the
learner will correctly identify an action with approximately
the highest expected payoff with probability at least 1 − e.
By the SLLN, at least a 1 − e fraction of agents will learn
an η-best reply. A detailed version of this proof in a more
general setting can be found in [9].
3.4 Convergence Theorem
Thus far we have defined large anonymous games where
approximate best-reply dynamics converge. If all agents in
the game are ǫ-stage learners, then the sequence ρ0, ρ1, . . . of
action distributions in a run of the game is not a best-reply
sequence, but it is close. The action used by most agents
most of the time in each ρn is the action used in ρn for some
approximate best reply sequence.
In order to prove this, we need to define "close." Our
definition is based on the error rate e and exploration rate
ǫ that introduces noise into ρn. Intuitively, distribution ρ
is close to ρ if, by changing the strategies of an e fraction
of agents and having all agents explore an ǫ fraction of the
time, we can go from an action profile with corresponding
action distribution ρ to one with corresponding distribution
ρ. Note that this definition will not be symmetric.
In this definition, g identifies what (pure) action each
agent is using that leads to ρ, g ′ allows an e fraction of
agents to use some other action, and g incorporates the fact
that each agent is exploring, so each strategy is an aǫ (the
agent usually plays a but explores with probability ǫ).
Definition 3.2. Action distribution ρ (e, ǫ)-close to ρ if
there exist g, g ′, and g ∈ G such that:
• ρ = ρg and ρ = ρg;
4
• g(i) ∈ A for all i ∈ N;
• ρg − ρg′ 1 ≤ 2e (this allows an e fraction of agents
in g ′ to play a different strategy from g);
• for some ǫ′ ≤ ǫ, if g ′(i) = a then g(i) = aǫ′ .
The use of ǫ′ in the final requirement ensures that if two
distributions are (e, ǫ)-close then they are also (e′, ǫ′)-close
for all e′ ≥ e and ǫ′ ≥ ǫ. As an example of the asymmetry
of this definition, aǫ is (0, ǫ) close to a, but the reverse is
not true. While (e, ǫ)-closeness is a useful distance measure
for our analysis, it is an unnatural notion of distance for
specifying the continuity of u, where we used the L1 norm.
The following simple lemma shows that this distinction is
unimportant; if ρ is sufficiently (e, ǫ)-close to ρ then it is
close according to the L1 measure as well.
Lemma 3.3. If ρ is (e, ǫ)-close to ρ, then ρ−ρ1 ≤ 2(e+
ǫ).
Proof. Since ρ is (e, ǫ)-close to ρ, there exist g, g ′, and
g as in Definition 3.2. Consider the distributions ρg = ρ,
ρg′ , and ρg = ρ. We can view these three distributions as
vectors, and calculate their L1 distances. By Definition 3.2,
ρg′ − ρg1 ≤ 2ǫ because an ǫ fraction
ρg − ρg′ 1 ≤ 2e.
of agents explore. Thus by the triangle inequality, the L1
distance between ρ and ρ is at most 2(e + ǫ).
We have assumed that approximate best reply sequences
of ρn converge, but during a run of the game agents will
actually be learning approximate best replies to ρn. The
following lemma shows that this distinction does not matter
if ρ and ρ are sufficiently close.
Lemma 3.4. For all η there exists a dη such that if ρ
is (e, ǫ)-close to ρ, e > 0, ǫ > 0, and e + ǫ < dη then
ABR(η/2)(ρ) ⊆ ABRη(ρ).
Proof. Let K be the maximum of the Lipschitz con-
stants for all u(a, ·) and dη = η/(8K). Then for all ρ that
are (e, ǫ)-close to ρ and all a, u(a, ρ−u(a, ρ) ≤ ρ−ρ1K ≤
2η/(8K)K = η/4 by Lemma 3.3.
Let a /∈ ABRη(ρ) and a′ ∈ argmaxa′∈ABRη (ρ) u(a′, ρ).
Then u(a, ρ) + η < u(a′, ρ). Combining this with the above
gives u(a, ρ) + η/2 < u(a′, ρ). Thus a /∈ ABRη/2(ρ).
Lemmas 3.1 and 3.4 give requirements on (e, ǫ). In the
statement of the theorem, we call (e, ǫ) η-acceptable if they
satisfy the requirements of both lemmas for η/2 and all η-
best-reply sequences converge in Γ.
Theorem 3.5. Let Γ be a large anonymous game where
approximate best-reply dynamics converge and let (e, ǫ) be
η-acceptable for Γ. If all agents are ǫ-stage learners then,
for all runs, there exists an η-best-reply sequence ρ0, ρ1, . . .
such that in stage n at least a 1 − e fraction will learn a best
reply to ρn with probability 1.
Proof. ρ0 = ρ0, so ρ0 is (e, ǫ)-close to ρ. Assume ρn is
(e, ǫ)-close to ρ. By Lemma 3.1 at least a 1 − e fraction will
learn a η/2-best reply to ρn. By Lemma 3.4, this is a η-best
reply to ρn. Thus ρn+1 will be (e, ǫ)-close to ρn+1.
Theorem 3.5 guarantees that after a finite number of stages,
agents will be close to an approximate Nash equilibrium pro-
file. Specifically, ρn will be (e, ǫ)-close to an η-Nash equi-
librium profile ρn. Note that this means that ρn is actually
an η′-Nash equilibrium for a larger η′ that depends on η,e,ǫ,
and the Lipschitz constant K.
Our three requirements for a practical learning algorithm
were that it require minimal information, converge quickly
in a large system, and be robust to noise. Stage learning re-
quires only that an agent know his own payoffs, so the first
condition is satisfied. Theorem 3.5 shows that it satisfies
the other two requirements. Convergence is guaranteed in
a finite number of stages. While the number of stages de-
pends on the game, in Section 3.2 we argued that in many
cases it will be quite small. Finally, robustness comes from
tolerating an e fraction of errors. While in our proofs we
assumed these errors were due to learning, the analysis is
the same if some of this noise is from other sources such
as churn (agents entering and leaving the system) or agents
making errors. We discuss this issue more in Section 5.
4. SIMULATION RESULTS
Theorem 3.5 guarantees convergence for a sufficiently small
exploration probability ǫ, but decreasing ǫ also increases τ ,
the length of a stage. Increasing the length of a stage means
that agents take longer to reach equilibrium, so for stage
learning to be practical, ǫ needs to be relatively large. To
show that ǫ can be large in practice, we tested populations
of stage learners in a number of games where best reply
dynamics converge and experienced convergence with ǫ be-
tween 0.01 and 0.05. This allows convergence within a few
thousand rounds in many games. While our theorem applies
only to stage learning, the analysis provides intuition as to
why a reasonable algorithm that changes slowly enough that
other learners have a chance to learn best replies should con-
verge as well. To test a very different type of algorithm, we
also implemented the no-regret learning algorithm of Hart
and Mas-Collell [14]. This algorithm also quickly converged
close to Nash equilibrium, although in many games it did
not converge as closely as stage learning.
Our theoretical results make two significant predictions
about factors that influence the rate of convergence. Lemma 3.1
tells us that the length of a stage is determined by the num-
ber of times each strategy needs to be explored to get an
accurate estimate of its value. Thus the amount of infor-
mation provided by each observation has a large effect on
the rate of convergence. For example, in a random match-
ing game, an agents payoff provides information about the
strategy of one other agent. On the other hand, if he receives
his expected payoff for being matched, a single observation
provides information about the entire distribution of strate-
gies. In the latter case the agent can learn with many fewer
observations.
A related prediction is that having more agents will lead to
faster convergence, particularly in games where payoffs are
determined by the average behavior of other agents, because
variance in payoffs due to exploration and mistakes decreases
as the number of agents increases. Our experimental results
illustrate both of these phenomena.
We tested the learning behavior of stage learners and no-
regret learners in a number of games, including prisoner's
dilemma, a climbing game [6], the congestion game described
in [12] with both ACP and serial mechanisms, and two differ-
ent contribution games (called a Diamond-type search model
in [20]). We implemented payoffs both by randomly match-
ing players and by giving each player what his expected pay-
off would have been had he been randomly matched (some
5
m
u
i
r
b
i
l
i
u
q
E
m
o
r
f
e
c
n
a
t
s
D
i
6
5
4
3
2
1
0
0
Stage 2
Stage 10
Stage 100
No−Regret 2
No−Regret 10
No−Regret 100
m
u
i
r
b
i
l
i
u
q
E
m
o
r
f
e
c
n
a
t
s
D
i
Stage 2
Stage 10
Stage 100
No−Regret 2
No−Regret 10
No−Regret 100
6
5
4
3
2
1
0.5
1
1.5
2
Time
2.5
x 104
0
0
1
2
3
Time
4
5
x 104
Figure 1: Convergence with the average
Figure 2: Convergence with random matching
payoffs were adjusted to make the games symmetric). Re-
sults were similar across the different games, so we report
only the results for a contribution game.
In the contribution game, agents choose strategies from 0
to 19, indicating how much effort they contribute to a collec-
tive enterprise. The value to an agent depends on how much
he contributes, as well as how much other agents contribute.
If he contributes x and the contribution of the other agents
is y, then his utility is 2xy − c(x), where c(0) = 0, c(1) = 1,
c(x) = (x − 1)2 for x ∈ 2, . . . , 8 and c(x) = x2 + 2n for x > 8.
We considered two versions of this game. In the first, y is
determined by the average strategy of the other agents. In
the second, y is determined by randomly matching the agent
with another agent.
Our implementation of stage learners is as described in
Section 3.3, with ǫ = 0.05 when y is determined by the av-
erage and ǫ = 0.01 when y is determined by random match-
ing. Rather than taking the length of stage τ as 1/ǫ2, we
set τ = 250 and 2000, respectively; this gives better perfor-
mance. Our implementation of no-regret learners is based
on that of Hart and Mas-Colell [14], with improvements sug-
gested by Greenwald et al. [12].
Figure 1 shows the results for learners in the version of the
game where y is the average strategy of other agents. Each
curve shows the distance from equilibrium as a function of
the number of rounds of a population of agents of a given size
using a given learning algorithm. The results were averaged
over 10 runs. Since the payoffs for nearby strategies are
close, we want our notion of distance to take into account
that agents playing 7 are closer to equilibrium (8) than those
playing zero. Therefore, we consider the expected distance
of ρ from equilibrium: Pa ρ(a)a − 8. To determine ρ, we
counted the number of times each action was over the length
of a stage, so in practice the distance will never be zero
due to mistakes and exploration. For ease of presentation,
the graph shows only populations of size up to 100; similar
results were obtained for populations up to 5000 agents.
For stage learning, increasing the population size has a
dramatic impact. With two agents, mistakes and best replies
to the results of these mistakes cause behavior to be quite
chaotic. With ten agents, agents successfully learn, although
mistakes and suboptimal strategies are quite frequent. With
one hundred agents, all the agents converge quickly to equi-
librium strategies and mistakes are rare; almost all of the
distance from equilibrium is due to exploration.
No-regret learning also converges quickly, but the "qual-
ity" of convergence (how close we get to equilibrium) is not
as high. The major problem is that a significant fraction of
agents play near-optimal actions rather than optimal action.
This may have a number of causes. First, the guarantee is
that the asymptotic value of ρ will be an equilibrium, which
allows the short periods that we consider to be far from equi-
librium. Second, the quality of convergence depends on ǫ,
so tight convergence may require a much lower rate of ex-
ploration and thus a much longer convergence time. Finally,
this algorithm is guaranteed to converge only to a correlated
equilibrium, which may not be a Nash equilibrium.
Figure 2 shows the results when agent payoffs are deter-
mined by randomly matching agents. Even for large num-
bers of stage learners, convergence is not as tight and takes
on the order of ten times longer. This is a result of the infor-
mation available to agents. When payoffs were determined
by the average strategy, a single observation was sufficient
to evaluate a strategy, so we could use very short stages. To
deal with the noise introduced by random matching we need
much longer stages. The number of stages to convergence
is similar. Even with longer stages and a large number of
agents, mistakes are quite common. Nevertheless agents do
successfully learn. The performance of no-regret learners is
less affected because they use payoff information from the
entire run of the game, while stage learners discard payoff
information at the end of each stage.
Convergence in the random-matching game takes approx-
imately 20,000 rounds, which is too slow for many applica-
tions. If a system design requires this type of matching, this
makes learning problematic. However, the results of Fig-
ure 1 suggest that the learning could be done much faster
if the system designer could supply agents with more infor-
mation. This suggests that collecting statistical information
about the behavior of agents may be a critical feature for
ensuring fast convergence. If agents know enough about the
game to determine their expected payoffs from this statisti-
cal information, then they can directly learn, as in Figure 1.
Even with less knowledge about the game, statistical infor-
mation can still speed learning, for example, by helping an
agent determine whether the results of exploring an action
were typical or due to the other agent using a rare action.
6
5. DISCUSSION
While our results show that a natural learning algorithm
can learn efficiently in an interesting class of games, there
are many further issues that merit exploration.
Other Learning Algorithms
Our theorem assumes that agents use a simple rule for lear-
ing within each stage:
they average the value of payoffs
received. However, there are certainly other rules for es-
timating the value of an action; any of these can be used
as long as the rule guarantees that errors can be made ar-
bitrarily rare given sufficient time. It is also not necessary
to restrict agents to stage learning. Stage learning guar-
antees a stationary environment for a period of time, but
such strict behavior may not be needed or practical. Other
approaches, such as exponentially discounting the weight of
observations [12, 19] or Win or Learn Fast [4] allow an algo-
rithm to focus its learning on recent observations and pro-
vide a stable environment in which other agents can learn.
Other Update Rules
In addition to using different algorithms to estimate the val-
ues of actions, a learner could also change the way he uses
those values to update his behavior. For example, rather
than basing his new strategy on only the last stage, he could
base it on the entire history of stages and use a rule in the
spirit of fictitious play. Since there are games where fictitious
play converges but best-reply dynamics do not, this could
extend our results to another interesting class of games, as
long as the errors in each period do not accumulate over
time. Another possibility is to update probabilistically or
use a tolerance to determine whether to update (see e.g. [7,
14]). This could allow convergence in games where best-
reply dynamics oscillate or decrease the fraction of agents
who make mistakes once the system reaches equilibrium.
Model Assumptions
Our model makes several unrealistic assumptions, most no-
tably that there are countably many agents who all share
the same utility function. Essentially the same results holds
with a large, finite number of agents, adding a few more
"error terms".
In particular, since there is always a small
probability that every agent makes a mistake at the same
time, we can prove only that no more than a 1 − e fraction
of the agents make errors in most rounds, and that agents
spending most of their time playing equilibrium strategies.
We have also implicitly assumed that the set of agents is
fixed. We could easily allow for churn: agents entering and
leaving the system. A reasonable policy for newly-arriving
agents is to pick a random aǫ to use in the next stage. If
all agents do this, it follows that convergence is unaffected:
we can treat the new agents as part of the e fraction that
made a mistake in the last stage. Furthermore, this tells us
that newly arriving agents "catch up" very quickly. After a
single stage, new agents are guaranteed to have learned a
best reply with probability at least 1 − e.
Finally, we have assumed that all agents have the same
utility function. Our results can easily be extended to in-
clude a finite number of different types of agents, each with
their own utility function, since the SLLN can be applied to
each type of agent. We believe that our results hold even
if the set of possible types is infinite. This can happen, for
example, if an agent's utility depends on a valuation drawn
from some interval. However, some care is needed to define
best-reply sequences in this case.
State
One common feature of distributed systems not addressed
in this work is state. For example, in a scrip system where
agents pay each other for service using an internal currency
or scrip, whether an agent should seek to provide service
depends on the amount of money he currently has [8].
In principle, we could extend our framework to games
with state:
in each stage each agent chooses a policy to
usually follow and explores other actions with probability ǫ.
Each agent could then use some off-policy algorithm (one
where the agent can learn without controlling the sequence
of observations; see [16] for examples) to learn an optimal
policy to use in the next stage. One major problem with this
approach is that standard algorithms learn too slowly for our
purposes. For example, Q-learning [25] typically needs to
observe each state-action pair hundreds of times in practice.
The low exploration probability means that the expected
SA/ǫ rounds needed to explore each even once for each
pair is large. Efficient learning requires more specialized
algorithms that can make better use of the structure of a
problem, but this also makes providing a general guarantee
of convergence more difficult. Another problem is that, even
if an agent explores each action for each of his possible local
states, the payoff he receives will depend on the states of
the other agents and thus the actions they chose. We need
some property of the game to guarantees this distribution
of states is in some sense "well behaved."
Despite these concerns, preliminary results suggest that
simple learning algorithms work well for games with state.
In experiments on a game using the model of a scrip system
from [8], we found that a stage-learning algorithm that uses
a specialized algorithm for determining the value of actions
in each stage converges to equilibrium quickly despite churn
and agents learning at different rates.
Mixed Equilibria
Another restriction of our results is that our agents only
learn pure strategies. One way to address this is to discretize
the mixed strategy space (see e.g. [7]). If one of the resulting
strategies is sufficiently close to an equilibrium strategy and
best-reply dynamics converge with the discretized strategies,
then we expect agents to converge to a near-equilibrium dis-
tribution of strategies. We have had empirical success using
this approach to learn to play rock-paper-scissors.
Unexpected and Byzantine Behavior
In practice, we expect that not all agents will be trying to
learn optimal behavior in a large system. Some agents may
simply play some particular (possibly mixed) strategy that
they are comfortable with, without trying to learn a better
strategy. Others may be learning but with an unanticipated
utility function. Whatever their reasons, if these sufficiently
few such agents are choosing their strategies i.i.d. from fixed
distribtions (or at least fixed for each stage), then our results
hold without change. This is because we already allow an
e fraction of agents to make arbitrary mistakes, so we can
treat these agents as simply mistaken.
Byzantine agents, who might wish to disrupt learning as
much as possible, do not fit as neatly into our framework;
they need not play the same strategy for an entire stage.
7
However, we expect that since correct agents are random-
izing their decisions, a small number of Byzantine agents
should not be able to cause many agents to make mistakes.
6. CONCLUSION
Learning in distributed systems requires algorithms that
are scalable to thousands of agents and can be implemented
with minimal information about the actions of other agents.
Most general-purpose multiagent learning algorithms fail one
or both of these requirements. We have shown here that
stage learning can be an efficient solution in large anonymous
games where approximate best-reply dynamics lead to ap-
proximate pure strategy Nash equilibria. Many interesting
classes of games have this property, and it is frequently found
in designed games. In contrast to previous work, the time
to convergence guaranteed by the theorem does not increase
with the number of agents.
If system designers can find
an appropriate game satisfying these properties on which to
base their systems, they can be confident that nodes can
efficiently learn appropriate behavior.
Our results also highlight two factors that aid conver-
gence. First, having more learners often improves perfor-
mance. With more learners, the noise introduced into pay-
offs by exploration and mistakes becomes more consistent.
Second, having more information typically improves perfor-
mance. Publicly available statistics about the observed be-
havior of agents can allow an agent to learn effectively while
making fewer local observations.
Acknowledgements
EF, IK, and JH are supported in part by NSF grant ITR-
0325453. JH is also supported in part by NSF grant IIS-
0812045 and by AFOSR grants FA9550-08-1-0438 and FA9550-
05-1-0055. EF is also supported in part by NSF grant CDI-
0835706.
7. REFERENCES
[1] M. Blonski. Equilibrium characterization in large
anonymous games. Technical report, U. Mannheim,
2001.
[2] A. Blum, E. Even-Dar, and K. Ligett. Routing
without regret: on convergence to Nash equilibria of
regret-minimizing algorithms in routing games. In
Proc 25th ACM Symp. on Principles of Distributed
Computing (PODC), pages 45 -- 52, 2006.
[3] M. H. Bowling. Convergence problems of general-sum
multiagent reinforcement learning. In
17th Int. Conf. on Machine Learning (ICML 2000),
pages 89 -- 94, 2000.
[4] M. H. Bowling and M. M. Veloso. Rational and
convergent learning in stochastic games. In 17th
Int. Joint Conference on Artificial Intelligence (IJCAI
2001), pages 1021 -- 1026, 2001.
[5] N. Cesa-Bianchi and G. Lugosi. Prediction, Learning
and Games. Cambridge University Press, 2006.
[6] C. Claus and C. Boutilier. The dynamics of
reinforcement learning in cooperative multiagent
systems. In AAAI-97 Workshop on Multiagent
Learning, pages 746 -- 752, 1998.
[7] D. P. Foster and P. Young. Regret testing: Learning
to play Nash equilibrium without knowing you have
an opponent. Theoretical Economics, 1:341 -- 367, 2006.
8
[8] E. J. Friedman, J. Y. Halpern, and I. A. Kash.
Efficiency and Nash equilibria in a scrip system for
P2P networks. In Proc. Seventh ACM Conference on
Electronic Commerce (EC), pages 140 -- 149, 2006.
[9] E. J. Friedman and S. Shenker. Learning and
implementation on the internet. 1998.
[10] D. Fudenberg and D. Levine. Theory of Learning in
Games. MIT Press, 1998.
[11] F. Germano and G. Lugosi. Global Nash convergence
of Foster and Young's regret testing. Games and
Economic Behavior, 60(1):135 -- 154, July 2007.
[12] A. Greenwald, E. J. Friedman, and S. Shenker.
Learning in networks contexts: Experimental results
from simulations. Games and Economic Behavior,
35(1-2):80 -- 123, 2001.
[13] S. Hart and A. Mas-Colell. A simple adaptive
procedure leading to correlated equilibrium.
Econometrica, 68(5):1127 -- 1150, 2000.
[14] S. Hart and A. Mas-Colell. A reinforecement learning
procedure leading to correlated equilibrium. In
G. Debreu, W. Neuefeind, and W. Trockel, editors,
Economic Essays, pages 181 -- 200. Springer, 2001.
[15] J. Hu and M. P. Wellman. Nash Q-learning for
general-sum stochastic games. Journal of Machine
Learning Research, 4:1039 -- 1069, 2003.
[16] L. P. Kaelbling, M. L. Littman, and A. P. Moore.
Reinforcement learning: A survey. J. Artif. Intell. Res.
(JAIR), 4:237 -- 285, 1996.
[17] J. R. Marden, G. Arslan, and J. S. Shamma.
Connections between cooperative control and
potential games. In Proceedings of the 2007 European
Control Conference (ECC), 2007.
[18] J. R. Marden, G. Arslan, and J. S. Shamma. Regret
based dynamics: convergence in weakly acyclic games.
In 6th Int. Joint Conf. on Autonomous Agents and
Multiagent Systems (AAMAS), pages 42 -- 49, 2007.
[19] J. R. Marden, H. P. Young, G. Arslan, and J. S.
Shamma. Payoff-based dynamics for multi-player
weakly acyclic games. SIAM Journal on Control and
Optimization, 2008. To appear.
[20] P. Milgrom and J. Roberts. Rationalizability, learning,
and equilibrium in games with strategic complement-
arities. Econometrica, 58(6):1255 -- 1277, 1900.
[21] M. Osborne and A. Rubenstein. A Course in Game
Theory. MIT Press, 1994.
[22] Y. Shoham, R. Powers, and T. Grenager. Multi-agent
reinforcement learning: a critical survey. Technical
report, Stanford, 2003.
[23] G. Tesauro and J. O. Kephart. Pricing in agent
economies using multi-agent Q-learning. Autonomous
Agents and Multi-Agent Systems, 5(3):289 -- 304, 2002.
[24] K. Verbeeck, A. Now´e, J. Parent, and K. Tuyls.
Exploring selfish reinforcement learning in repeated
games with stochastic rewards. Journal of
Autonomous Agents and Multi-agent Systems,
14:239 -- 269, 2007.
[25] C. J. Watkins and P. Dayan. Technical note
Q-learning. Machine Learning, 8:279 -- 292, 1992.
[26] H. P. Young. Learning by trial and error. Games and
Economic Behavior, forthcoming.
|
1511.07209 | 2 | 1511 | 2016-07-28T09:11:40 | Multi-Agent Continuous Transportation with Online Balanced Partitioning | [
"cs.MA",
"cs.AI",
"cs.RO"
] | We introduce the concept of continuous transportation task to the context of multi-agent systems. A continuous transportation task is one in which a multi-agent team visits a number of fixed locations, picks up objects, and delivers them to a final destination. The goal is to maximize the rate of transportation while the objects are replenished over time. Examples of problems that need continuous transportation are foraging, area sweeping, and first/last mile problem. Previous approaches typically neglect the interference and are highly dependent on communications among agents. Some also incorporate an additional reconnaissance agent to gather information. In this paper, we present a hybrid of centralized and distributed approaches that minimize the interference and communications in the multi-agent team without the need for a reconnaissance agent. We contribute two partitioning-transportation algorithms inspired by existing algorithms, and contribute one novel online partitioning-transportation algorithm with information gathering in the multi-agent team. Our algorithms have been implemented and tested extensively in the simulation. The results presented in this paper demonstrate the effectiveness of our algorithms that outperform the existing algorithms, even without any communications between the agents and without the presence of a reconnaissance agent. | cs.MA | cs | Multi-Agent Continuous Transportation with Online
Balanced Partitioning
(Extended Abstract)
Chao Wang†, Somchaya Liemhetcharat§, Kian Hsiang Low†
Institute for Infocomm Reserach, Agency for Science, Technology and Research, Singapore§
Department of Computer Science, National University of Singapore, Singapore†
{wangchao, lowkh}@comp.nus.edu.sg†, [email protected]§
6
1
0
2
l
u
J
8
2
]
A
M
.
s
c
[
2
v
9
0
2
7
0
.
1
1
5
1
:
v
i
X
r
a
ABSTRACT
We introduce the concept of continuous transportation task
to the context of multi-agent systems. A continuous trans-
portation task is one in which a multi-agent team visits a
number of fixed locations, picks up objects, and delivers
them to a transportation hub. The goal is to maximize
the rate of transportation while the objects are replenished
over time . In this extended abstract, we present a hybrid of
centralized and distributed approaches that minimize com-
munications in the multi-agent team. We contribute a novel
online partitioning-transportation algorithm with informa-
tion gathering in the multi-agent team.
Keywords
Multi-agent system; Partitioning; Transportation; Coopera-
tion
1.
INTRODUCTION
We are interested in multi-agent coordination and con-
tinuous transportation, where agents visit locations in the
environment to transport objects (passengers or items) and
deliver them to a transportation hub. The objects replen-
ish over time, and we consider Poisson model of object re-
plenishment. Poisson model is suitable for scenarios with
independently-occurring objects, such as passengers appear
at the transit stops. Thus, the continuous transportation
task is general and applicable to many real-life scenarios,
e.g., first/last mile problem, package or mail collection and
delivery.
The most similar work has been done is in multi-robot
foraging and delivery problem [2, 3]. A multi-robot team
forages resources from environment to a home location or
delivers items to locations on request while the resources
replenish over time. The goal of continuous foraging is to
maximize the rate of resource foraging. Thus, continuous
foraging has similarities to continuous transportation, and
thus we compare our algorithms with that of [2, 3].
In addition, continuous area sweeping (e.g., [1]) problem
has similarities to continuous transportation, with the main
difference being that the agents have a carrying capacity and
Appears in: Proceedings of the 15th International Conference
on Autonomous Agents and Multiagent Systems (AAMAS 2016),
J. Thangarajah, K. Tuyls, C. Jonker, S. Marsella (eds.),
May 9 -- 13, 2016, Singapore.
Copyright c(cid:13) 2016, International Foundation for Autonomous Agents and
Multiagent Systems (www.ifaamas.org). All rights reserved.
must periodically return to the transportation hub. Hence,
we compare to the benchmark of [1]. Besides, [1] uses area
partitioning algorithm that relies on communications and
negotiations among robots. Thus, it is prone to message
loss and inconsistency. In contrast, our approach does not
require communications among agents, and the partition-
ing is conducted separately. Hence, we present a hybrid of
centralized and distributed approaches, and the efficacy is
demonstrated in the evaluation section.
2. PROBLEM AND APPROACH
In this section, we formally define the multi-agent continu-
ous transportation problem in detail, and give an overview of
our approach. In a continuous transportation task, a multi-
agent team visits a number of fixed locations and transport
objects to a transportation hub.
2.1 Formal Problem Definition
The multi-agent continuous transportation problem can
be defined as follows:
e.g., the autonomous vehicles.
• A = {a1, ..., ak} is the set of transportation agents,
• si, ci, and yi (≤ ci) denote agent ai's speed, maximum
capacity, and current load, i.e., the number of objects
carried.
• L = {l1, ..., ln} is the set of locations, where l0 is the
transportation hub.
• vj,t denotes the number of objects available at location
lj at time step t, e.g., the number of passengers appear
at the transit stop lj.
j,t denotes ai's estimate of vj,t at time step t.
• v(i)
• oj denotes the observation made at location lj. When a
transportation agent ai arrives at a location lj (j > 0),
min(vj,t, ci − yi) objects at lj are picked up by ai, and
ai makes an observation oj of the number of objects
remaining. When ai arrives at l0, all yi objects carried
by ai are transferred to l0.
• D : L × L → R+ is the distance function of the loca-
tions.
• t(ai, lj, lk) =
(cid:108) D(lj ,lk)
(cid:109)
si
is the time steps taken for an
agent ai to move from location lj to lk.
The goal is to maximize the rate of objects delivered to
the transportation hub l0 within T time steps, i.e., maximize
v0,T
T .
2.2 Our Approach
Our approach for solving the continuous transportation
problem is:
• We assume that vj,t follows a known model -- in this
paper, we assume that vj,t follows the Poisson model,
where the number of objects replenished every time
step follows a Poisson distribution with mean λj. How-
ever, the parameters of the models (i.e., λj) are not
known in advance;
• The estimates v(i)
j,t are updated using the model and
observations from transportation agents ai visiting lo-
cation lj;
• Following [2], the transportation agents do not share
their models v(i)
j,t since it could be expensive subject to
the communication bandwidth and team size. Differ-
ent from [2], in our approach, the sharing of destination
and load among agents are not required.
We contribute an online algorithm that partition the loca-
tions based on their 2D-position and estimated replenish-
ment rate, and controls the transportation agents ai, that
use v(i)
j,t to plan their next destination within the cluster of
locations assigned. The algorithm dynamically repartitions
locations based on information gathered by the multi-agent
team.
3. ALGORITHM AND EVALUATION
Our Online Balanced Partitioning (OBP) algorithm be-
gins with statically partition the locations into clusters based
on 2D-position of locations with k-means algorithm. The
algorithm is inspired by algorithms proposed for continuous
foraging [2]. The main difference is that the agents replan
destinations in the cluster of locations assigned, instead of
all the locations. The replanning of destination within each
cluster is based on Greedy Rate [2], i.e., maximizing the
transportation rate. Further, due to partitioning, communi-
cations of destinations and loads are not required.
By only considering the 2D-position of locations, the work-
load of each agent might not be balanced, i.e., the total re-
plenishment rate of some clusters could be so high that the
agents cannot afford it, while some other agents are idle. On
the other hand, the result of standard k-means algorithm
depends on the choice of initial centroid which is randomly
generated. In this case, we introduce online balanced par-
titioning, i.e., balance estimated total replenishment rate of
each cluster with information gathering. The replenishment
rate is not known in advance. The agents use preset esti-
mated replenishment rate to update its estimated number of
objects. The estimate can only be corrected when they visit
the location and make an observation. Since we focus on
Poisson replenishment model, where number of objects re-
plenished per time step follows a mean value, we believe that
estimated replenishment rate can be corrected with contin-
uous observations, i.e., the total number of objects replen-
ished divided by time steps elapsed.
Our online partitioning algorithm dynamically corrects
the estimated replenishment rate by visiting a location and
making observation. Once the deviation of total replenish-
ment rate of clusters are greater than a preset threshold,
repartitioning will be done immediately.
Figure 1 shows the performance of our Online Balanced
Partitioning (OBP) algorithm when the capacities of the
Figure 1: Comparison of our Online Balanced Partitioning
(OBP) algorithm against the benchmark of Greedy Rate
with Expected Observation (GR+EO) and Continuous Area
Sweeping (CAS) algorithm.
agents are 10 and the number of locations are 20. In order
to demonstrate the effectiveness of OBP, we compared OBP
with GR+EO [2] which requires the presence of a reconnais-
sance agent. The solid red, blue, and gray lines show Online
Balanced Partitioning (OBP), Greedy Rate with Expected
Observation (GR+EO), and Random Transportation (R),
and the shaded areas show the standard deviations of these
algorithms.
As the number of agents increase, OBP outperforms all
other algorithms, even GR+EO (p = 3 × 10−23), primar-
ily because of the dynamic partitioning with information
gathering by the multi-agent team. It clearly illustrates the
efficacy of our OBP algorithms over GR. Without the use
of reconnaissance agent and communications among agents,
our OBP algorithm can still outperform GR+EO algorithm.
Acknowledgments
This work was supported by NUS Computer Center through
the use of its high performance computing facilities.
REFERENCES
[1] M. Ahmadi and P. Stone. A multi-robot system for
continuous area sweeping tasks. In Robotics and
Automation, 2006. ICRA 2006. Proceedings 2006 IEEE
International Conference on, pages 1724 -- 1729. IEEE,
2006.
[2] S. Liemhetcharat, R. Yan, and K. P. Tee. Continuous
foraging and information gathering in a multi-agent
team. In Proceedings of the 2015 International
Conference on Autonomous Agents and Multiagent
Systems, pages 1325 -- 1333. International Foundation for
Autonomous Agents and Multiagent Systems, 2015.
[3] S. Liemhetcharat, R. Yan, K. P. Tee, and M. Lee.
Multi-robot item delivery and foraging: Two sides of a
coin. Robotics, 4(3):365 -- 397, 2015.
|
1009.5346 | 1 | 1009 | 2010-09-27T18:20:56 | A Novel Approach for Cardiac Disease Prediction and Classification Using Intelligent Agents | [
"cs.MA",
"cs.AI"
] | The goal is to develop a novel approach for cardiac disease prediction and diagnosis using intelligent agents. Initially the symptoms are preprocessed using filter and wrapper based agents. The filter removes the missing or irrelevant symptoms. Wrapper is used to extract the data in the data set according to the threshold limits. Dependency of each symptom is identified using dependency checker agent. The classification is based on the prior and posterior probability of the symptoms with the evidence value. Finally the symptoms are classified in to five classes namely absence, starting, mild, moderate and serious. Using the cooperative approach the cardiac problem is solved and verified. | cs.MA | cs | A Novel Approach for Cardiac Disease
Prediction and Classification Using
Intelligent Agents
Murugesan Kuttikrishnan
Department of Computer Science and Engineering
Anna University, Chennai, India
[email protected]
Manjula Dhanabalachandran
Department of Computer Science and Engineering
Anna University, Chennai, India
[email protected]
Abstract— The goal is to develop a novel approach for
cardiac disease prediction and diagnosis using intelligent
agents. Initially the symptoms are preprocessed using filter
and wrapper based agents. The filter removes the missing
or irrelevant symptoms. Wrapper is used to extract the
data in the data set according to the threshold limits.
Dependency of each
symptom
is
identified using
dependency checker agent. The classification is based on
the prior and posterior probability of the symptoms with
the evidence value. Finally the symptoms are classified in
to five classes namely absence, starting, mild, moderate
and serious. Using the cooperative approach the cardiac
problem is solved and verified.
Keywords- Traditional Chinese Medicine (TCM), Naïve
Bayesian Classification (NBC), Bayesian Networks (BN).
I. INTRODUCTION
Intelligent agents are a new paradigm
for
developing software applications. More than this,
agent-based computing has been hailed as „the next
significant breakthrough in software development‟
(Sargent, 1992), and „the new revolution
in
software‟ (Ovum, 1994). Currently, agents are the
focus of intense interest on the part of many sub-
fields of computer science and artificial intelligence.
Agents are being used in an increasingly wide
variety of applications, ranging from comparatively
small systems such as email filters to large, open,
complex, mission critical systems such as air traffic
control. At first sight, it may appear that such
extremely different types of system can have little
in common. And yet this is not the case: in both, the
key abstraction used is that of an agent.
First, 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. Autonomy is a difficult
concept to pin down precisely, but we mean it
simply in the sense • that the system should be able
to act without the direct intervention of humans (or
other agents), and should have control over its own
actions and internal state. It may be helpful to draw
an analogy between the notion • of autonomy with
respect to agents and encapsulation • with respect to
object oriented systems. An object encapsulates
some state, and has some control over this state in
that it can only be accessed or modified via the
methods
that
the object provides. Agents
encapsulate state in just the same way. However, we
also think of agents as encapsulating behavior, in
addition to state. An object does not encapsulate
behavior: it has no control over the execution of
methods – if an object x invokes a method m on an
object y, then y has no control over whether m is
executed or not – it just is. In this sense, object y is
not autonomous, as it has no control over its own
actions. In contrast, we think of an agent as having
exactly this kind of control over what actions it
performs. Because of this distinction, we do not
think of agents as invoking methods (actions) on
agents – rather, we tend to think of them requesting
actions
to be performed. The decision about
whether to act upon the request lies with the
recipient. .
An intelligent agent is a computer system
that is capable of flexible autonomous action in
order to meet its design objectives. By flexible, we
mean that the system must be:
responsive: agents should perceive their
environment (which may be the physical
world, a user, a collection of agents, the
Internet, etc.) and respond in a timely
fashion to changes that occur in it.
proactive: agents should not simply act in
response to their environment, they should
be able
to exhibit opportunistic, goal-
directed behavior and take the initiative
where appropriate, and Applications of
Intelligent Agents
social: agents should be able to interact,
when they deem appropriate, with other
artificial agents and humans in order to
complete their own problem solving and to
help others with their activities.
Hereafter, when we use the term „agent‟, it
should be understood that we are using it as an
abbreviation
for
„intelligent
agent‟. Other
researchers emphasize different aspects of agency
(including, for example, mobility or adaptability).
Naturally, some agents may have additional
characteristics, and for certain types of applications,
some attributes will be more important than others.
However, we believe that it is the presence of all
four attributes in a single software entity that
provides the power of the agent paradigm and
which distinguishes agent systems from related
software paradigms – such as object-oriented
systems, distributed sysems, and expert systems
(see Wooldridge (1997) for a more detailed
discussion). By an agent-based system, we mean
one in which the key abstraction used is that of an
agent. In principle, an agent-based system might be
conceptualized in terms of agents, but implemented
without any software structures corresponding to
agents at all. We can again draw a parallel with
object-oriented software, where
it
is entirely
possible to design a system in terms of objects, but
to implement it without the use of an object-
oriented software environment. But this would at
best be unusual, and at worst, counterproductive. A
similar situation exists with agent technology; we
therefore expect an agent-based system to be both
designed and implemented in terms of agents. A
number of software tools exist that allow a user to
implement software systems as agents, and as
societies of cooperating agents. Note that an agent-
based system may contain any non-zero number of
agents. The multi-agent case – where a system is
designed and implemented as several interacting
agents, is both more general and significantly more
complex than the single-agent case. However, there
are a number of situations where the single-agent
case is appropriate.
II. RELATED WORK
Traditional diagnosis in TCM requires long
experiences and a high level of skill, and is
subjective and deficient in quantitative diagnostic
criteria. This seriously affects the reliability and
repeatability
of
diagnosis
and
limits
the
popularization of TCM. So the focal problem that
needs to be solved urgently is to construct methods
or models to quantify the diagnosis in TCM.[1]
Recently, a
few researchers developed some
methods and systems to modernize TCM. But most
of them are built incorporating totally or partially
rulebased reasoning model, which are lack of the
feasibility of implementing all possible inference by
chaining rules and limits their practical applications
in clinical medicines.An attraction
tool
for
managing various forms of uncertainty is Bayesian
networks (BNs) [2], [3] which is able to represent
knowledge with uncertainty
and
efficiently
performing
reasoning
tasks.Naive Bayesian
classifier (NBC) is a simplified form of BNs that
assumes independence of the observations. Some
research results [4], [5], [6] have demonstrated that
the predictive performance of NBC can be
competitive with more complicated classifiers.In
this study, a novel computerized diagnostic model
based on naive Bayesian classifier (NBC)
is
proposed. Firstly, a Bayesian network structure is
learned from a database of cases [7] to find the
symptom set that are dependent on the disease
directly.Secondly, the symptom set is utilized as
attributes of NBC and the mapping relationships
between the symptom set and the disease are
constructed.To
reduce
the dimensionality and
improve the prediction accuracy of diagnostic
model, symptom selection is requisite.Many feature
selection methods, such as filters [8] and wrappers
[9], have developed. But
the dependency
relationships among symptoms and the mapping
relationships between symptom and syndrome are
not considered
these methods, which are
in
important to diagnosis in TCM.
irrelative
To
lower
the
influences of
symptoms, the mutual information between each
symptom and disease
is computed based on
information entropy theory [10], which is utilized to
assess the significance of symptoms.The paper [11]
presents a multiagent system
for supporting
physicians in performing clinical studies in real
time. The multiagent system is specialized in the
controlling of patients with respect
their
to
appointment behavior. Novel types of agents are
designed to play a special role as representatives for
humans in the environment of clinical studies.
OnkoNet mobile agents have been used successfully
for patient-centric medical problems
solving
[12].,emerged from a project covering all relevant
issues, from empirical process studies in cancer
diagnosis/therapy, down to system implementation
and validation. In the paper [13], a medical
diagnosis multiagent system
that
is organized
according to the principles of swarm intelligence is
proposed. It consists of a large number of agents
that interact with each other by simple indirect
communication.
In the paper [14], a system called Feline
composed of five autonomous agents (expert
systems with some proprieties of the agents)
endowed with medical knowledge is proposed.
These agents cooperate to identify the causes of
anemia at cats. The paper [39], also presents a
development methodology for cooperating expert
systems.In the paper [15], a Web-centric extension
to a previously developed expert system specialized
in
the glaucoma diagnosis
is proposed. The
proposed telehealth solution publishes services of
the developed Glaucoma Expert System on the
World Wide Web.Each agent member of the CMDS
system has problems solving capability and capacity
(the notions are defined in [16, 17]). The capacity of
an agent Agf (Agf U MDUAS)consists in the
amount of problems that can be solved by the agent,
using the existent problem solving resources.The
cooperative medical diagnosis problems solving by
the diagnosis system is partially based on the
blackboard-based problem solving [18, 19]. The
problem solving by the BMDS system is similar
with the situations, when more physicians with
different medical specializations plans a treatment
to cure an illness that is in an advanced stage.
Treatments known to be effective for the curing of
the illness in a less advanced stage cannot be
applied.
the entire discussions one can
From
comprehend and classify the medical agent-based
IDSS research [20] into two categories, namely
Clinical Management and Clinical Research.
Clinical Management envelops all clinical systems
that are designed to help the doctor with diagnosing
and deciding on treatment for medical conditions.
the erstwhile envelopes
Clinical Research on
systems
that are used
to research facts and
connections in attempt to detect new trends and
patterns; it covers systems for both diagnosing
patients and treating them.
III. PREPROCESSING AND CLASSIFICATION
3.1 Filter agent
Feature selection, as a preprocessing step to
reducing
effective
learning,
machine
in
is
dimensionality,
removing
irrelevant data,
in-
creasing learning accuracy, and improving result
comprehensibility. In this work, we introduce a
novel concept, predominant correlation, and
propose a faster method which can identify relevant
features as well as redundancy among relevant
features without pair wise correlation analysis. The
efficiency and effectiveness of our method is
demonstrated through extensive comparisons with
other methods using real-world data of high
dimensionality.
3.2 Wrapper agent
In Wrapper based feature selection, the more
states that are visited during the search phase of the
algorithm the greater the likelihood of finding a
feature subset that has a high internal accuracy
while generalizing poorly. It removes the irrelevant
attributes that are below the threshold value.
3.3 Classifier agent
The classifier agent uses the naïve Bayesian
classification
algorithm. Bayesian
network
algorithm is used to classify the collected attributes
in
to five classes (0-normal,1-starting,2-low,3-
mild,4serious). The mutual information between
each symptom and disease is computed based on
information entropy theory. F and C are symptom
and disease. f, c are events of F and C. I(F,C)=(cid:1)P(f,
c)logP(f,c)/P(f)P(c) Suppose I0 is the prior entropy
of C. I0 =(cid:1)(P(c=1)logp(c=1)+p(c=0)logp(c=0))
The significance of each symptom is calculated by
S(F,C)=I(F,C)/I0. All the symptoms are evaluated
and ranked by significance index S(F,C).
Input Attributes Documentation
1 id: patient identification number
2 ccf: social security number (I replaced this with a
dummy value of 0)
3 age: age in years
4 sex: sex (1 = male; 0 = female)
5 painloc: chest pain location (1 = substernal; 0 =
otherwise)
6 painexer (1 = provoked by exertion; 0 =
otherwise)
7 relrest (1 = relieved after rest; 0 = otherwise)
8 pncaden (sum of 5, 6, and 7)
9 cp: chest pain type
--Value 1: typical angina
--Value 2: atypical angina
--Value 3: non-anginal pain
--Value 4: asymptomatic
10 trestbps: resting blood pressure (in mm Hg on
admission to the hospital)
11 htn
12 chol: serum cholestoral in mg/dl
13 smoke: I believe this is 1 = yes; 0 = no (is or is
not a smoker)
14 cigs (cigarettes per day)
15 years (number of years as a smoker)
16 fbs: (fasting blood sugar > 120 mg/dl) (1 = true;
0 = false)
17 dm (1 = history of diabetes; 0 = no such history)
18 famhist: family history of coronary artery disease
(1 = yes; 0 = no)
19 restecg: resting electrocardiographic results
--Value 0: normal
--Value 1: having ST-T wave abnormality (T
wave inversions and/or ST elevation or depression
of > 0.05 mV)
--Value 2: showing probable or definite left
ventricular hypertrophy by Estes' criteria
20 ekgmo (month of exercise ECG reading)
21 ekgday(day of exercise ECG reading)
22 ekgyr (year of exercise ECG reading)
23 dig (digitalis used furing exercise ECG: 1 = yes;
0 = no)
24 prop (Beta blocker used during exercise ECG: 1
= yes; 0 = no)
25 nitr (nitrates used during exercise ECG: 1 = yes;
0 = no)
26 pro (calcium channel blocker used during
exercise ECG: 1 = yes; 0 = no)
27 diuretic (diuretic used used during exercise ECG:
1 = yes; 0 = no)
28 proto: exercise protocol
1 = Bruce
2 = Kottus
3 = McHenry
4 = fast Balke
5 = Balke
6 = Noughton
7 = bike 150 kpa min/min (Not
sure if "kpa min/min" is what was written!)
8 = bike 125 kpa min/min
9 = bike 100 kpa min/min
10 = bike 75 kpa min/min
11 = bike 50 kpa min/min
12 = arm ergometer
29 thaldur: duration of exercise test in minutes
30 thaltime: time when ST measure depression was
noted
31 met: mets achieved
32 thalach: maximum heart rate achieved
33 thalrest: resting heart rate
34 tpeakbps: peak exercise blood pressure (first of 2
parts)
35 tpeakbpd: peak exercise blood pressure (second
of 2 parts)
36 dummy
37 trestbpd: resting blood pressure
38 exang: exercise induced angina (1 = yes; 0 = no)
39 xhypo: (1 = yes; 0 = no)
40 oldpeak = ST depression induced by exercise
relative to rest
41 slope: the slope of the peak exercise ST segment
--Value 1: upsloping
--Value 2: flat
--Value 3: downsloping
42 rldv5: height at rest
43 rldv5e: height at peak exercise
44 ca: number of major vessels (0-3) colored by
flourosopy
45 restckm: irrelevant
46 exerckm: irrelevant
47 restef: rest raidonuclid (sp?) ejection fraction
48 restwm: rest wall (sp?) motion abnormality
0 = none
1 = mild or moderate
2 = moderate or severe
3 = akinesis or dyskmem (sp?)
49 exeref: exercise radinalid (sp?) ejection fraction
50 exerwm: exercise wall (sp?) motion
51 thal: 3 = normal; 6 = fixed defect; 7 = reversable
defect
52 thalsev: not used
53 thalpul: not used
54 earlobe: not used
55 cmo: month of cardiac cath (sp?) (perhaps "call")
56 cday: day of cardiac cath (sp?)
57 cyr: year of cardiac cath (sp?)
58 num: diagnosis of heart disease (angiographic
disease status)
--Value 0: < 50% diameter narrowing
--Value 1: > 50% diameter narrowing
(in any major vessel: attributes 59 through 68
are vessels)
59 lmt
60 ladprox
61 laddist
62 diag
63 cxmain
64 ramus
65 om1
66 om2
67 rcaprox
68 rcadist
69 lvx1: not used
70 lvx2: not used
71 lvx3: not used
72 lvx4: not used
73 lvf: not used
74 cathef: not used
75 junk: not used
76 name: last name of patient
(I replaced this with the dummy string "name")
EXPERIMENTAL RESULTS
IV.
Table: 1 Data Sets Used
S.No. DATA SET NAME NO. OF INSTANCES
303
CLEVELAND
1
294
HUNGARIAN
2
3
SWITZERLAND
123
4
LONG BEACH
200
Table: 2 Preprocessed Results-Cleveland Data Set
PID FILTER WRAPPER
1
2
3
4
5
6
7
8
9
10
12
10
11
14
11
12
22
11
11
9
13
12
7
10
13
12
15
12
10
10
Table: 3 Preprocessed Results-Hungarian Data Set
PID FILTER WRAPPER
1254
1255
1256
1257
10
9
8
7
12
11
11
10
1258
1259
1260
1261
1262
1263
12
11
13
11
10
13
13
10
12
10
12
13
Table: 4 Preprocessed Results-Swiz Data Set
PID FILTER WRAPPER
6
11
3001
8
10
3002
9
11
3003
3004
8
11
10
9
3005
11
10
3006
11
10
3007
3008
12
11
11
11
3009
3010
11
10
Table: 5 Preprocessed Results-Longbeach Data Set
PID FILTER WRAPPER
1
2
3
4
5
6
7
8
9
10
11
10
11
13
14
12
11
12
6
11
13
8
12
11
6
9
10
10
11
14
Table: 6 Prior Probability of Symptoms
START
ING
0.4
0.4
0.34
0.43
0.2
0.4
0.2
0.3
0.4
0.4
0.3
0.2
0.3
0.333
0.12
0.3
0.3
0.3
O.3
MILD MODE
RATE
0.7
0.85
0.7
0.6
0.4
0.5
0.5
0.6
0.5
0.5
0.466
0.4
0.5
0.5
0.3
0.5
0.488
0.4
0.455
0.6
0.7
0.4
0.5
0.3
0.45
0.4
0.5
0.45
0.4
0.4
0.3
0.4
0.388
0.2
0.4
0.4
0.395
0.35
SYMP
TOM
PAINLOC
PAINEXER
RELREST
PNCADEN
CP
TRESTBPS
HTN
CHOL
SMOKE
CIGS
YEARS
FBS
DM
FAMHIST
RESTECG
EKGMO
EKGDAY
EKGYR
DIG
ABSEN
CE
0.3
0.1
0.2
0.1
0.3
0.3
0.2
0.3
0.23
0.14
0.15
0.16
0.3
0.3
0.1
0.2
0.3
0.2
0.23
SERI
OUS
0.9
0.89
0.8
0.7
0.8
0.76
0.7
0.778
0.8
0.9
0.8
0.7
0.888
0.677
0.55
0.6
0.7
0.9
0.8
TABLE: 7 CLASSIFICATION RESULTS
S.NO
.
DATA SET
ABSE
N
CE
1
2
3
4
CLEVELAND
164
HUNGARIAN
188
SWITZERLAN
D
LONG BEACH
8
51
STAR
T
ING
MIL
D
MOD
E
RATE
SER
I
OUS
55
37
48
56
36
26
32
41
35
28
30
42
13
15
5
10
Figure: 1 Preprocessed Chart for Cleveland Data Set
In the current work, input symptom values are
preprocessed using filter and wrapper approach.
Retained symptom values are used to classify
cardiac patients in to five classes viz. Normal,
Starting, Mild, Moderate and Serious. Classified
values determine severity of the cardiac disease.
Future work represents Heuristic seed selection to
increase classification accuracy over the supervised
naïve Bayesian classification. In this approach five
seeds are selected for five classes. Future work
represents diagnosis of cardiac patients using the
co-operative elaboration algorithm. It uses contract
net protocol that allows autonomy and sharing
among the agents to solve all cardiac problems.
Diagnosis results will be accurate in all conditions
and then intelligent based decision support system is
used to retrieve the patient information from the
database.
VI. ACKNOWLEDGEMENT
Primarily, my gratitude goes to Dr. D.Manjula,
A.P, DCSE, Anna University Chennai, who guided
me to carry out this research with interest and
involvement. With profound reverence and high
regards, I thank Dr. P. Narayanasamy, Professor and
HOD, DISE, Anna University Chennai and
Dr.R.Krishnamoorthy, Professor and Dean, Anna
University Trichy, for the motivation provided by
them during this research work which enabled me to
complete this work successfully.
REFERENCES
Figure: 2 Classification Chart for all Data Sets
V. CONCLUSIONS AND FUTURE WORK
[1] B. Ma, “Expert Systems and Knowledge Bases in
Traditional Chinese Medicine”, Beijing Press, Beijing, 1994
.
[2] J. Pearl, “Probabilistic Reasoning in Intelligent Systems:
Networks of Plausible
Inference” Morgan Kaufmann
Publishers, Inc. San Francisco, California, 1988.
051015202512345678910PIDFILTERWRAPPER [3] J. Cheng,, C. Hatzis,, H. Hayashi, M.A. Krogel, S.
Morishita, D. Page, J. Sese, “KDD Cup 2001 report” SIGKDD
Explor. Newsl, 2002. (2), pp.47-64.
[4] W. Buntine, “Learning classification rules using Bayes”, in
Proc. 6th Int. Workshop Machine Learning, 1989, pp. 94–96.
[5] B. Cestnik, “Estimating probabilities: A crucial task in
machine learning,” in Proc. Euro. Conf. Artif. Intell., 1990, pp.
147–149.
[6] P. Langley, W. Iba, and K. Thompson, “An analysis of
1992”, pp. 223–228.
[7] J. Cheng, R. Greiner, J. Kelly, D. A. Bell, W. Liu,
“Learning Bayesian Networks from Data: An Information-
Theory Based Approach”, Artificial Intelligence, 2002,
Vol.137, pp.43-90.
[8] G. Isabelle, “An Introduction to Variable and Feature
Selection”, Journal of Machine Learning Research, Vol. 3,
pp.1157-1182.
[9] R. Kohavi, G.
for Feature
“Wrappers
John,
Selection”,Artificial Intelligence, Vol. 97, No. 1, 1997,
pp.273-324.
[10] C. E. Shannon, W. Weaver, “The Mathematical Theory of
Communication.Urbana, IL: University of Illinois Press,1949.
[11] H. Myritz, G. Lindemann, G. Zahlmann, Hans -Dieter
Burkhard. “Patient scheduling
in clinical studies with
multiagent techniques”. Proceedings of the 2nd International
Workshop on Multi-Agent Systems for Medicine and
Computational Biology, Hakodate, Japan, pages 87–103, 2006
[12] S. Kirn. Ubiquitous healthcare: “The onkonet mobile
agents architecture”. Proceedings of the 3.rd International
Conference
Netobjectdays.
Objects,
Components,
Architectures,Services,
a
for
Applications
and
NetworkedWorld, Aksit, M.,Mezini, M., Unland, R. (Eds.),
Springer-Verlag, Germany, LNCS, (2591), 2003.
[13] R. Ulieru, M. Unland. “A stigmergic approach to medical
diagnosis”.Proceedings of the 2nd International Workshop on
Multi-Agent Systems
for Medicine and Computational
Biology, Hakodate, Japan, pages 87–103, 2006.
[14] G. Weiss. (Ed.)” Multiagent Systems: A Modern
Approach to Distributed Artificial Intelligence” MIT Press
Cambridge, Massachusetts London, 2000.
[15] M. Wooldridge, G. M. P. O‟Hare, R. Elks. Feline – “a
case study in the design and implementation of a co -operating
expert system”. Proceedings of the International Conference
on Expert Systems and their Applications, France, 1991.
[16] A. Ulieru, M. Grabelkovsky. “Telehealth approach for
glaucoma progression monitoring.”International
Journal:
Information Theories and Applications, (10):326–329, 2005.
[17] B. Iantovics, C. Chira, D. Dumitrescu. “Principles of the
Intelligent Agents”Casa Cartii de Stiinta Press, Cluj-Napoca,
2007.
[18] J. Ferber. “Multi-Agent Systems: An Introduction to
Distributed Artificial Intelligence” Addison Wesley, 1999.
[19] V. Jagannathan, R. Dodhiawala, L. S. Baum. (Eds.)
“Blackboard Architectures and Application”. Academic Press,
San Diego, 1989.
[20] Murugesan.K., Manjula. D., "Applications of Intelligent
Agents in Health Care Systems using Decision Support
Systems," Proceedings of International Conference on D igital
Factory (ICDF 2008), CIT, Coimbatore, India, Aug.1113,
2008: p.173.
AUTHORS PROFILE
K. Murugesan, is pursuing his Ph.D in Computer
Science from Anna University and also working for
Anna University as a Superintendent. His research
interests include Data Mining, Multi agent based
systems etc.,
D.Manjula is working for Anna University as an
Assistant Professor. Her research interests include
Natural Language Processing, Text Mining,
Artificial
Intelligence, Databases and Parallel
Computing.
|
1809.07225 | 2 | 1809 | 2019-06-24T10:34:07 | Deterministic limit of temporal difference reinforcement learning for stochastic games | [
"cs.MA"
] | Reinforcement learning in multiagent systems has been studied in the fields of economic game theory, artificial intelligence and statistical physics by developing an analytical understanding of the learning dynamics (often in relation to the replicator dynamics of evolutionary game theory). However, the majority of these analytical studies focuses on repeated normal form games, which only have a single environmental state. Environmental dynamics, i.e., changes in the state of an environment affecting the agents' payoffs has received less attention, lacking a universal method to obtain deterministic equations from established multistate reinforcement learning algorithms.
In this work we present a novel methodological extension, separating the interaction from the adaptation time scale, to derive the deterministic limit of a general class of reinforcement learning algorithms, called temporal difference learning. This form of learning is equipped to function in more realistic multistate environments by using the estimated value of future environmental states to adapt the agent's behavior. We demonstrate the potential of our method with the three well established learning algorithms Q learning, SARSA learning and Actor-Critic learning. Illustrations of their dynamics on two multiagent, multistate environments reveal a wide range of different dynamical regimes, such as convergence to fixed points, limit cycles, and even deterministic chaos. | cs.MA | cs |
Deterministic limit of temporal difference reinforcement learning for stochastic games
Wolfram Barfuss,1, 2, ∗ Jonathan F. Donges,1, 3 and Jurgen Kurths1, 2, 4
1Potsdam Institute for Climate Impact Research, Germany
2Department of Physics, Humboldt University Berlin, Germany
3Stockholm Resilience Centre, Stockholm University, Sweden
4Saratov State University, Russia
(Dated: June 25, 2019)
Reinforcement learning in multiagent systems has been studied in the fields of economic game
theory, artificial intelligence and statistical physics by developing an analytical understanding of
the learning dynamics (often in relation to the replicator dynamics of evolutionary game theory).
However, the majority of these analytical studies focuses on repeated normal form games, which
only have a single environmental state. Environmental dynamics, i.e., changes in the state of an
environment affecting the agents' payoffs has received less attention, lacking a universal method to
obtain deterministic equations from established multistate reinforcement learning algorithms.
In this work we present a novel methodological extension, separating the interaction from the
adaptation time scale, to derive the deterministic limit of a general class of reinforcement learning
algorithms, called temporal difference learning. This form of learning is equipped to function in
more realistic multistate environments by using the estimated value of future environmental states
to adapt the agent's behavior. We demonstrate the potential of our method with the three well
established learning algorithms Q learning, SARSA learning and Actor-Critic learning.
Illustra-
tions of their dynamics on two multiagent, multistate environments reveal a wide range of different
dynamical regimes, such as convergence to fixed points, limit cycles, and even deterministic chaos.
I.
INTRODUCTION
Individual learning through reinforcements is a central
approach in the fields of artificial intelligence [1 -- 3], neu-
roscience [4, 5], learning in games [6] and behavioral game
theory [7 -- 10], thereby offering a general purpose princi-
ple to either solve complex problems or explain behavior.
Also in the fields of complexity economics [11, 12] and so-
cial science [13], reinforcement learning has been used as
a model for human behavior to study social dilemmas.
However, there is a need for improved understand-
ing and better qualitative insight into the characteris-
tic dynamics that different learning algorithms produce.
Therefore, reinforcement learning has also been studied
from a dynamical systems perspective. In their seminal
work, Borgers and Sarin showed that one of the most ba-
sic reinforcement learning update schemes, Cross learn-
ing [14], converges to the replicator dynamics of evolu-
tionary games theory in the continuous time limit [15].
This has led to at least two, presumably nonoverlapping
research communities, one from statistical physics [16 -- 26]
and one from computer science machine learning [27 -- 35].
Thus, Sato and Crutchfield [18] and Tuyls et al. [27] inde-
pendently deduced identical learning equations in 2003.
The statistical physics articles usually consider the de-
terministic limit of the stochastic learning equations, as-
suming infinitely many interactions between the agents
before an adaptation of behavior occurs. This limit can
either be performed in continuous time with differen-
tial equations [17 -- 19] or discrete time with difference
∗ [email protected]
equations [20 -- 22]. The differences between both vari-
ants can be significant [21, 23]. Deterministic chaos was
found to emerge when learning simple [17] as well as
complicated games [25]. Relaxing the assumption of in-
finitely many interactions between behavior updates re-
vealed that noise can change the attractor of the learning
dynamics significantly, e.g., by noise-induced oscillations
[20, 21].
However, these statistical physics studies so far consid-
ered only repeated normal form games. These are games
where the payoff depends solely on the set of current ac-
tions, typically encoded in the entries of a payoff matrix
(for the typical case of two players). Receiving payoff
and choosing another set of joint actions is performed
repeatedly. This setup lacks the possibility to study
dynamically changing environments and their interplay
with multiple agents. In those systems, rewards depend
not only on the joint action of agents but also on the
states of the environment. Environmental state changes
may occur probabilistically and depend also on joint ac-
tions and the current state. Such a setting is also known
as a Markov game or stochastic game [36, 37]. Thus, a
repeated normal form game is a special case of a stochas-
tic game with only one environmental state. Notably,
Akiyama and Kaneko [38, 39] did emphasize the impor-
tance of a dynamically changing environment; however
did not utilize a reinforcement learning update scheme.
The computer science machine-learning community
dealing with reinforcement learning as a dynamical sys-
tem (see Ref.
[28] for an overview) particularly empha-
sizes the link between evolutionary game theory and mul-
tiagent reinforcement learning as a well grounded theo-
retical framework for the latter [28 -- 31]. This dynamical
systems perspective is proposed as a way to gain quali-
tative insights about the variety of multiagent reinforce-
ment learning algorithms (see Ref. [2] for a review). Con-
sequently, this literature developed a focus on the trans-
lation of established reinforcement learning algorithms to
a dynamical systems description, as well as the develop-
ment of new algorithms based on insights of a dynamical
systems perspective. While there is more work on state-
less games (e.g., Q learning [27] and frequency-adjusted
multiagent Q learning [32]), multiagent learning dynam-
ics for multistate environments have been developed as
well, such as piecewise replicator dynamics [34], state-
coupled replicator dynamics [33] or reverse engineering
state-coupled replicator dynamics [35].
Both communities, statistical physics and machine
learning, share the interest in better qualitative insights
into multiagent learning dynamics. While the statistical
physics community focuses more on dynamical proper-
ties the same set of learning equations can produce, it
leaves a research gap of learning equations capable of
handling multiple environmental states. The machine-
learning community, on the other hand, aims more to-
ward algorithm development, but so far has put their
focus less on a dynamical systems understanding. Taken
together, there is the challenge of developing a dynam-
ical systems theory of multiagent learning dynamics in
varying environmental states.
With this work, we aim to contribute to such a dynam-
ical systems theory of multiagent learning dynamics. We
present a novel methodological extension for obtaining
the deterministic limit of multistate temporal difference
reinforcement learning. In essence, it consists of formu-
lating the temporal difference error for batch learning,
and sending the batch size to infinity. We showcase our
approach with the three prominent learning variants of
Q learning, SARSA learning and Actor-Critic learning.
Illustrations of their learning dynamics reveal multiple
different dynamical regimes, such as fixed points, peri-
odic orbits, and deterministic chaos.
In Sec. II we introduce the necessary background and
notation. Section III presents our method to obtain the
deterministic limit of temporal difference reinforcement
learning and demonstrates it for multistate Q learning,
SARSA learning and Actor-Critic learning. We illustrate
their learning dynamics for two previously utilized two-
agents two-actions two-states environments in Sec. IV.
In Sec. V we conclude with a discussion of our work.
II. PRELIMINARIES
We introduce the components (including notation) of
our multiagent environment systems (see Fig. 1), fol-
lowed by a brief introduction of temporal difference rein-
forcement learning.
X(cid:104)◦(cid:105) :=
:=
sa1 ··· X N
X 1
saN · ◦.
(1)
A. Multiagent Markov environments
2
(cid:80)
1, . . . , Ai
A multiagent Markov environment (also called stochas-
tic game or Markov game) consists of N ∈ N agents. The
environment can exist in Z ∈ N states S = {S1, . . . , SZ}.
In each state each agent has M ∈ N available actions
Ai = {Ai
M}, i = 1, . . . , N to choose from. Hav-
ing an identical number of actions for all states and all
agents is notational convenience, no significant restric-
tion. A joint action of all agents is referred to by a ∈ A =
A1×···×AN , the joint action of all agents but agent i is
denoted by a−i ∈ A−i = A1×···×Ai−1×Ai+1×···×AN .
Environmental dynamics are given by the probabilities
for state changes expressed as a transition tensor T ∈
[0, 1]Z×M×...(N times)···×M×Z. The entry Tsas(cid:48) denotes the
probability P (s(cid:48)s, a) that the environment transitions to
state s(cid:48) given the environment was in state s and the
agents have chosen the joint action a. Hence, for all s, a,
s(cid:48) Tsas(cid:48) = 1 must hold. The assumption that the next
state only depends on the current state and joint action
makes our system Markovian. We here restrict ourselves
to ergodic environments without absorbing states (c.f.
Ref. [35]).
The rewards receivable by the agents are given by the
reward tensor R ∈ RN×Z×M×...(N times)···×M×Z. The en-
try Ri
sas(cid:48) denotes the reward agent i receives when the
environment transits from state s to state s(cid:48) under the
joint action a. Rewards are also called payoffs from a
game theoretic perspective.
Agents draw their actions from their behavior profile
sa = P (a i, s) denotes
X ∈ [0, 1]N×Z×M . The entry X i
the probability that agent i chooses action a in state s.
sa = 1 must hold. We
here focus on the case of independent agents, able to
fully observe the current state of the environment. With
correlated behavior (see e.g., Ref. [2]) and partially ob-
servable environments [40, 41] one could extend the mul-
tiagent environment systems to be even more general.
Note that what we call behavior profile is usually termed
policy from a machine-learning perspective or behavioral
strategy from a game theoretic perspective. We chose
to introduce our own term because policies and strate-
gies suggest a deliberate choice which we do not want to
impose.
Thus, for all i and all s, (cid:80)
a X i
B. Averaging out behavior and environment
We define a notational convention, that allows a sys-
tematic averaging over the current behavior profile X and
the environmental transitions T. It will be used through-
out the paper.
Averaging over the whole behavioral profile yields
(cid:88)
(cid:88)
a
Xsa · ◦
··· (cid:88)
a1∈A1
aN∈AN
3
FIG. 1: Multiagent Markov environment (also known as stochastic or Markov game). N agents choose a joint
action a = (a1, . . . , aN ) from their action sets Ai, based on the current state of the environment s, according to their
sa = P (ai, s). This will change the state of the environment from s to s(cid:48) with probability Tsas(cid:48),
behavior profile X i
and provide each agent with a reward Ri
sas(cid:48).
a1∈A1
a−i
(cid:88)
(cid:88)
(cid:124)
(cid:88)
s(cid:48)
(cid:125)
(cid:88)
s(cid:48)∈S
Here, ◦ serves as a placeholder.
If the quantity to be
inserted for ◦ depends on the summation indices, then
those indices will be summed over as well. If the quantity,
which is averaged out, is used in tensor form, then it is
written in bold. If not, then remaining indices are added
after the right angle bracket.
Averaging over the behavioral profile of the other
agents, keeping the action of agent i, yields
X−i(cid:104)◦(cid:105) :=
:=
sa−i · ◦
X−i
··· (cid:88)
(cid:123)(cid:122)
excl. i
aN∈AN
(cid:123)(cid:122)
sa1 ··· X N
X 1
(cid:124)
excl. i
(cid:125)
saN
· ◦ .
(2)
Last, averaging over the subsequent state s(cid:48) yields
T(cid:104)◦(cid:105) :=
Tsas(cid:48) · ◦ :=
Tsa1...aN s(cid:48) · ◦.
(3)
TX(cid:104)◦(cid:105) and
Of course, these operations may also be combined as
TX−i(cid:104)◦(cid:105) by multiplying both summations.
For example, given a behavior profile X, the resulting
effective Markov Chain transition matrix reads X(cid:104)T(cid:105)ss(cid:48),
which encodes the transition probabilities from state s
to s(cid:48). From X(cid:104)T(cid:105)ss(cid:48) the stationary distribution of envi-
ronmental states σ(X) can be computed. σ(X) is the
eigenvector corresponding to the eigenvalue 1 of X(cid:104)T(cid:105)ss(cid:48).
Its entries encode the ratios of the average durations the
agents find themselves in the respective environmental
states.
The average reward agent i receives from state s under
action a, given all other agents follow the behavior profile
sa. Including agent i's behavior profile
X reads
gives the average reward it receives from state s: TX(cid:104)R(cid:105)i
s.
Hence, TX(cid:104)R(cid:105)i
s =(cid:80)
TX−i(cid:104)R(cid:105)i
TX−i(cid:104)R(cid:105)i
sa holds.
sa ·
a X i
C. Agent's preferences and values
Gi(t) = (1 − γi)(cid:80)∞
Typically, agents are assumed to maximize their expo-
nentially discounted sum of future rewards, called return
k=0(γi)kri(t + k), where γi ∈ [0, 1) is
the discount factor of agent i and ri(t + k) denotes the
reward received by agent i at time step t + k. Expo-
nential discounting is most commonly used for its math-
ematical convenience and because it ensures consistent
preferences over time. Other formulations of a return
use e.g., finite-time horizons, average reward settings, as
well as other ways of discounting, such as hyperbolic dis-
counting. Those other forms require their own form of
reinforcement learning.
Given a behavior profile X, the expected return defines
(cid:10)Gi(t) s(t) = s(cid:11)i
TX
time t.
s (X) :=
the state-value function V i
s
which is
independent of
The operation
TX(cid:104). . . s(t) = s(cid:105) denotes the behavioral and environ-
mental average as defined in Eqs. 1 and 3 given that
in the current time step t the environment is in state s.
Inserting the return yields the Bellman equation [42],
s(t+1)(X) s(t) = s
(1 − γi)ri(t) + γiV i
V i
s (X) =
(cid:69)i
(cid:68)
s.
TX
(4)
This recursive relationship between state values de-
clares that the value of a state s is the discounted value
of the subsequent state s(t + 1) plus (1− γi) times the re-
ward received along the way. Evaluating the behavioral
and environmental average TX(cid:104) (cid:105) and writing in matrix
form we get:
Vi(X) = (1 − γi) · TX(cid:104)R(cid:105)i + γi · X(cid:104)T(cid:105) · Vi(X).
(5)
The reward ri(t) received at time step t is evaluated to
reward TX(cid:104)R(cid:105)i
s for state s, since the behavioral and en-
vironmental average was conditioned on starting in state
s(t) = s. The average subsequent state value V i
s(t+1)(X)
Environmentstate transitionstate setstate observationrewardAgentaction setjoint actionfrom behavior profile from the current state s can be expressed as a matrix
multiplication of the effective Markov transition matrix
and the vector of state values: (cid:80)
Vi(X) = (1 − γi)(cid:0)1Z − γi
using matrix inversion
X(cid:104)T(cid:105)(cid:1)−1
A solution of the state values Vi(X) can be obtained
s(cid:48) X(cid:104)T(cid:105)ss(cid:48) · Vi
s(cid:48)(X).
TX(cid:104)R(cid:105)i .
(6)
The computational complexity of matrix inversion makes
this solution strategy infeasible for large systems. There-
fore many iterative solution methods exist [3].
Equivalently, state-action-value functions Qi
(cid:10)Gi(t) s(t) = s, a(t) = a(cid:11)i
sa are de-
fined as the expected return, given agent i applied action
a in state s and then followed X accordingly: Qi
sa(X) :=
sa. Even though this is the
TX
behavioral average over the whole behavioral profile, the
resulting object carries an action index because the oper-
ation is conditioned on the current action to be a(t) = a.
They can be computed via
sa(X) = (1 − γi) TX−i(cid:104)R(cid:105)i
Qi
sa + γi(cid:88)
s (X) =(cid:80)
a X i
s(cid:48)
X(cid:104)T(cid:105)ss(cid:48) · V i
s(cid:48)(X).
(7)
One can show that V i
saQi
sa(X) holds for
the inverse relation of state-action and state values.
D. Learning through reinforcement
In contrast to the typical game-theoretic assumption of
perfect information, we assume that agents know nothing
about the game in advance. They can only gain infor-
mation about the environment and other agents through
interactions. They do not know the true reward tensor
R or the true transition probabilities Tsas(cid:48). They experi-
ence only reinforcements (i.e., particular rewards Ri
sas(cid:48)),
while observing the current true Markov state of the en-
vironment.
In essence, reinforcement learning consists of iterative
behavior changes toward a behavior profile with maxi-
mum state values. However, due to the agents' limited in-
formation about the environment, they generally cannot
compute a behavior profile's true state- and state-action
values, V i
sa(X), as defined in the previous sec-
tion. Therefore, agents use time-dependent state-value
and state-action-value approximations, V i
sa(t),
during the reinforcement learning process.
s (t) and Qi
s (X) and Qi
1. Temporal difference learning
Basically, state-action-value approximations Qi
sa get
iteratively updated by a temporal difference error
TDi
sa(t):
Qi
sa(t + 1) = Qi
sa(t) + αiTDi
sa(t),
(8)
4
with αi ∈ (0, 1) being the learning rate of agent i. These
state-action propensities Qi
sa can be interpreted as esti-
mates of the state-action values Qi
sa.
The temporal difference error expresses a difference in
the estimation of state-action values. New experience is
used to compute a new estimate of the current state-
action value and corrected by the old estimate. The esti-
mate from the new experience uses exactly the recursive
relation of value functions from the Bellmann equation
(Eq. 4),
TDi
(cid:34)
sa(t) = δss(t)δaa(t)·
(cid:124)
(1 − γi)Ri
s(t)a(t)a−i(t)s(t+1) + γi (cid:103)i
(cid:35)
estimate from new experience
(cid:123)(cid:122)
s(t+1) (t)
(cid:125)
− (cid:103)i
(cid:124) (cid:123)(cid:122) (cid:125)
s(t)(t)
old estimate
.
(9)
Here s and a denote the state-action pair whose tempo-
ral difference error is calculated. With s(t), a(t), etc. we
refer to the state, action, etc. that occurred at time step
t. Thus, the notation Ri
s(t)a(t)a−i(t)s(t+1) refers to the
entry of the reward tensor Ri
saa−is(cid:48) when at time step t
the environmental state was s [s(t) = s], agent i chose
action a [a(t) = a], the other agents chose the joint ac-
tion a−i [a−i(t) = a−i] and the next environmental state
was s(cid:48) [s(t + 1) = s(cid:48)]. The (cid:103)i
s(t+1)(t) indicates the state-
value estimate at time step t of the state visited at the
next time step s(t + 1). (cid:103)i
s(t)(t) denotes the state-value
estimate at time step t of the current state s(t). Differ-
ent choices for these estimations are possible, leading to
different learning variants (see below).
The Kronecker deltas δss(t), δaa(t) indicate that the
temporal difference error for state-action pair (s, a) is
only nonzero when (s, a) was actually visited in time
step t. This denotes and emphasizes that agents can
only learn from experience. In contrast, e.g., experience-
weighted-attraction learning [9] assumes that action
propensities can be updated with hypothetical rewards
an agent would have received if they had played a dif-
ferent action than the current action. These two cases
have been referred to as full vs. partial information [16].
Thus, the Kronecker deltas in Eq. 9 indicate a partial
information update. The agents use only information
experienced through interaction.
The state-action-value approximations Qi
sa are trans-
lated to a behavior profile according to the Gibbs-
Boltzmann distribution [1] (also called softmax)
X i
sa(t) =
(cid:80)
sa(t))
exp(βi Qi
b exp(βi Qi
sb(t))
.
(10)
The behavior profile X becomes a dynamic variable
as well. The parameter βi controls the intensity of
choice or the exploitation level of agent i controlling the
exploration-exploitation trade-off. In analogy to statisti-
cal physics, βi is the inverse temperature. For high βi
agents tend to exploit their learned knowledge about the
environment, leaning toward actions with high estimated
state-action value. For low βi, agents are more likely to
deviate from these high value actions in order to explore
the environment further with the chance of finding ac-
tions, which eventually lead to even higher values. Other
behavior profile translations exist as well (e.g., -greedy
[1]).
2. Three learning variants
The specific choices of the value estimates (cid:103) in the
temporal difference error result in different reinforcement
learning variants.
s(t)(t) = Qi
s(t+1)b(t) and (cid:103)i
a. Q learning. For the Q learning algorithm [1, 3],
(cid:103)i
s(t+1)(t) = maxb Qi
s(t)a(t)(t).
Thus, the Q learning update takes the maximum of the
next state-action-value approximations as an estimate for
the next state value, regardless of the actual next action
the agent plays. This is reasonable because the maxi-
mum is the highest value achievable given the current
knowledge. For the state-value estimate of the current
state, the Q learner takes the current state-action-value
approximation Qi
s(t)a(t)(t). This is reasonable because it
is exactly the quantity that gets updated by Eq. 8.
s(t)(t) = Qi
s(t+1)(t) = V i
s(t+1)a(t+1)(t) and (cid:103)i
b. SARSA learning. For SARSA learning [1, 3],
(cid:103)i
s(t+1)(t) = Qi
s(t)a(t)(t),
where a(t + 1) denotes the action taken by agent i at the
next time step a(t + 1) = Ai(t + 1). Thus, the SARSA
algorithm uses the five ingredients of an update sequence
of State, Action, Reward, next State, next Action to per-
form one update. In practice, the SARSA sequence has
to be shifted one time step backward to know what the
actual "next" action of the agent was.
c. Actor-Critic (AC) learning. For AC learning [1,
3], (cid:103)i
s(t)(t). Com-
pared to Q and SARSA learners, it has an additional data
structure of state-value approximations which get sepa-
s (t)+αi·Di
rately updated according to V i
s (t+1) = V i
sa(t).
The state-action-value approximations Qi
sa serve as the
actor which gets criticized by the state-value approxima-
tions V i
s .
Tab. Ia summarizes the values estimates (cid:103) for these
three learning variants. Q and SARSA learning are
structurally more similar compared to the Actor-Critic
learner, which uses an additional data structure of state-
value approximations V i
s .
s(t+1)(t) and (cid:103)i
s(t)(t) = V i
III. DETERMINISTIC LIMIT
So far we gave a brief introduction to temporal dif-
ference reinforcement learning. A more comprehensive
5
[1].
presentation can be found in Ref.
In this section
we will present a novel extension to the methodology of
interaction-adaptation timescales separation to the gen-
eral class of temporal difference reinforcement learning.
In summary, we (i) give a batch formulation of the tempo-
ral difference error, (ii) separate the timescales of interac-
tion and adaptation by sending the batch size to infinity
and (iii) present a resulting deterministic limit conver-
sion rule for discrete time updates. We showcase our
method in the three learning variants of Q, SARSA and
Actor-Critic learning. For the statistical physics com-
munity, the novelty consists of learning equations, ca-
pable of handling environmental state transitions. For
the machine learning community the novelty lies in the
systematic methodology we use to obtain the determin-
istic learning equations. Note that these deterministic
learning equations will not depend on the state-value or
state-action-value approximations anymore, being iter-
ated maps of the behavior profile alone.
Following e.g., Refs. [18, 19, 22], we first combine Eqs.
8 and 10 and obtain
X i
sa(t + 1) =
(cid:80)
X i
sa(t) exp
αiβiTDi
sa(t)
b X i
sb(t) exp
αiβiTDi
sb(t)
(11)
(cid:17)
(cid:17) .
(cid:16)
(cid:16)
Although it appears that only the product αiβi matters
for a behavior profile update, the temporal difference er-
sa may depend only on the exploitation level βi,
ror TDi
as we will show below.
Next, we formulate the temporal difference error for
batch learning.
A. Batch learning
With batch learning we mean that several time steps
of interaction with the environment and the other agents
take place before an update of the state-action-value ap-
proximations and the behavior profile occurs. It has also
been interpreted as a form of history replay [43] which is
essential to stabilize the learning process when function
approximation (e.g., by deep neural networks) is used
[44]. History (i.e., already experienced state, action, next
state triples) is used again for an update of the state-
action-value approximations.
Imagine that the information from these interactions
are stored inside a batch of size K ∈ N. We introduce
the corresponding temporal difference error of batch size
K:
TDi
sa(t; K) :=
1
·(cid:16)
K(s, a)
(1 − γi)Ri
+ γi (cid:103)i
δss(t+k)δaa(t+k)
s(t+k)a(t+k)a−i(t+k)s(t+k+1)
s(t+k+1) (t) − (cid:103)i
s(t)(t)
(12)
(cid:17)(cid:105)
K−1(cid:88)
(cid:104)
k=0
(cid:103)i
s(t+1)(t)
(cid:103)i
s(t)(t)
(cid:103)i
s(t+1)(t)
(cid:103)i
s(t)(t)
6
Q learning
SARSA learning
maxb Qi
s(t+1)b(t)
Qi
s(t)a(t)(t)
Qi
s(t+1)a(t+1)(t)
Qi
s(t)a(t)(t)
Actor Critic (AC) learning
V i
s(t+1)(t)
V i
s(t)(t)
maxQi
nextV i
nextV i
sa(X)
sa(X)
sa(X)
1
βi log X i
sa(t)
1
βi log X i
sa(t)
/
(a) K = 1
(b) K = ∞
TABLE I: Overview of the three reinforcement learning variants. Shown in the columns are the value estimates for
s(t)(t) for both ends of the batch size spectrum: K = 1 and K = ∞.
the next state (cid:103)i
s(t+1)(t) and the current state (cid:103)i
and transition probabilities according to
K−1(cid:88)
δss(t+k)δaa(t+k) →(cid:88)
(cid:88)
1
K(s, a)
k=0
s(cid:48)
a−i
X−i
sa−iTsaa−is(cid:48).
(15)
For example, the immediate reward Ri
in the temporal difference error becomes
time t gets resealed accordingly, as well.
s(t)a(t)a−i(t)s(t+1)
TX−i(cid:104)R(cid:105)i
sa. The
Taking the limit K → ∞ in this way, we choose to
stay in discrete time, leaving the continuous time limit
following Refs. [18, 19, 25] for future work.
C. Three learning variants
Next, we present the deterministic limit of the tempo-
ral difference error of the three learning variants of Q,
SARSA and Actor-Critic learning. Inserting them into
Eq. 11 yields the complete description of the behavior
profile update in the deterministic limit. Table I presents
an overview of the resulting equations and a comparison
to their batch size K = 1 versions.
1. Q learning
three
The temporal difference error of Q learning con-
(i) Ri
(ii)
s(t)a(t)(t). As already stated,
sa under K → ∞.
s(t+1)b(t), and (iii) Qi
s(t)a(t)a−i(t)s(t+1),
TX−i(cid:104)R(cid:105)i
sa(X), which is defined as
terms:
sists of
maxb Qi
s(t)a(t)a−i(t)s(t+1) →
Ri
s(t+1)b(t) → maxQi
(cid:88)
(cid:88)
maxb Qi
X−i
maxQi
sa(X) :=
Qi
s(cid:48)b(X) (16)
sa−iTsaa−is(cid:48) max
b
s(cid:48)
a−i
using the deterministic limit conversion rule (Eq. 15).
Because of the assumption of infinite interactions, we
can here replace the state-action-value approximations
where K(s, a) = max(1,(cid:80)K−1
k=0 δss(t+k)δaa(t+k)) denotes
the number of times the state-action pair (s, a) was vis-
ited. If the state-action pair (s, a) was never visited, then
K(s, a) = 1. The agents interact K times under the same
behavior profile and use the sample average to summarize
the new experience in order to update the state-action-
value approximations:
Qi
sa(t + K) = Qi
sa(t) + αiTDi
sa(t; K).
(13)
The notation TDi
size 1: TDi
sa(t) = TDi
sa(t; 1).
sa(t) denotes a batch update of batch
B. Separation of timescales
We obtain the deterministic limit of the temporal dif-
ference learning dynamics by sending the batch size to
infinity, K → ∞. Equivalently, this can be regarded
as a separation of timescales. Two processes can be dis-
tinguished during an update of the state-action-value ap-
proximations ∆ Qi
sa(t): adaptation
and interaction,
sa(t+1)− Qi
sa(t) := Qi
∆ Qi
(cid:122)
(cid:16)
sa(t) = αiδss(t)δaa(t)·
(cid:124)
(cid:123)(cid:122)
(1 − γi)Ri
interaction
(cid:125)(cid:124)
adaptation
s(t)a(t)a−i(t)s(t+1) + γi (cid:103)i
s(t+1) (t)
(cid:123)
(cid:17)
− (cid:103)i
s(t) (t)
(cid:125)
(14)
By separating the timescales of both processes, we as-
sume that (infinitely) many interactions happen before
one step of behavior profile adaptation occurs.
Under this assumption and because of the assumed er-
godicity one can replace the sample average, i.e., the sum
over sequences of states and actions with the behavior
profile average, i.e., the sum over state-action behavior
s(cid:48)b as defined
sa(t)+consti
s, where consti
s(t+1)b with the true state-action values Qi
For the third term, we invert Eq. 10, yielding Qi
Qi
by Eq. 7.
sa(t) =
(βi)−1 log X i
s is constant in ac-
tions but may vary for each agent and state. Now, one
can show that the dynamics induced by Eq. 11 are in-
variant against additive transformations in the tempo-
sa(t,∞) + consti
ral difference error TDi
s.
Thus, the third term can be converted according to
s(t)a(t)(t) → (βi)−1 log X i
Qi
All together, the temporal difference error for Q learn-
sa(t,∞) → TDi
sa(t).
ing in the deterministic limit reads
s(t)(t) := (cid:80)
think of a conversion V i
(βi)−1(cid:80)
7
s(t)a(t)(t) →
Qi
sa(t). However, the remaining
term is constant in action, and therefore irrelevant for
the dynamics, as we have argued above. Thus, we can
simply put V i
sa(t) log X i
s(t)(t) → 0.
a X i
sa
a X i
All together, the temporal difference error of the Actor-
Critic learner in the deterministic limit reads
acTDi
sa(t,∞) =(1 − γi) TX−i(cid:104)R(cid:105)i
sa(X)
+ γi nextV i
sa
(21)
qTDi
sa(t,∞) =(1 − γi) TX−i(cid:104)R(cid:105)i
sa
+ γi maxQi
sa(X) − 1
βi log X i
sa(t).
2. SARSA learning
IV. APPLICATION TO EXAMPLE
ENVIRONMENTS
(17)
Two of the three terms of the SARSA temporal differ-
ence error are identical to the one of Q learning, leaving
Qi
s(t+1)a(t+1)(t), which we replace by
nextQi
(cid:88)
(cid:88)
(cid:88)
X−i
sa(X) :=
X i
s(cid:48)bQi
s(cid:48)b(X)
sa−iTsaa−is(cid:48)
s(cid:48)
a−i
b
(18)
using again the deterministic limit conversion rule (Eq.
15) and the state-action value Qi
s(cid:48)b(X) of the behavior
profile X according to Eq. 7.
Thus, the temporal difference error for the SARSA
learning update in the deterministic limit reads
sarsaTDi
sa(t;∞) = (1 − γi) TX−i(cid:104)R(cid:105)i
sa(X) − 1
+ γi nextQi
sa
βi log X i
sa(t).
(19)
3. Actor-Critic (AC) learning
For the temporal difference error for AC learning
s(t+1)(t) and (ii)
s(t+1)(t) →
we have to find replacements for (i) V i
V i
s(t)(t). Applying again Eq.
nextV i
15 yields V i
sa defined as
nextV i
sa :=
X−i
sa−iTsaa−is(cid:48)V i
s(cid:48)(X),
(20)
(cid:88)
(cid:88)
s(cid:48)
a−i
using Eq. 6 for the state value V i
s(cid:48)(X). This is the average
value of the next state given that in the current state the
agent took action a. One can show that nextV i
sa(X) =
nextQi
sa(X) from the SARSA update.
The second remaining term belongs to the slower
adaptation timescale or,
in other words, occurs out-
side the batch. Thus, our deterministic limit con-
version rule (Eq.
15) does not apply. We could
In the following we apply the derived deterministic
learning equations in two different environments. Specif-
ically, we compare the three well established temporal
difference learning variants (Q learning, SARSA learn-
ing and Actor-Critic (AC) learning) in two different two-
agent (N = 2), two-action (M = 2) and two-state
(Z = 2) environments: a two-state Matching Pennies
game and a two-state Prisoner's Dilemma. Since the
main contribution of this paper is the derivation of the
deterministic temporal difference learning equations, we
are not trying to make a case with our example environ-
ments beyond a systematic comparison of our learners.
Therefore, we chose environments that have been used
previously in related literature [33 -- 35, 45]. Note also
that we leave a comparison between the deterministic
limit and the stochastic equations to future work, which
would add a noise term to our equations following the
example of Ref. [20].
To measure the performance of an agent's behavior
profile in a single scalar, we use the dot product between
the stationary state distribution σ(X) of the effective
Markov Chain with the transition matrix X(cid:104)T(cid:105) and the
behavior average reward TX(cid:104)R(cid:105)i. Interestingly, we find
this relation to be identical to the dot product of the
stationary distribution and the state value Vi(X):
σ(X) · TX(cid:104)R(cid:105)i = σ(X) · Vi(X).
(22)
This relation can be shown by using Eq. 6 and the fact
that σ(X) is an eigenvector of X(cid:104)T(cid:105).
In the following examples we will only investigate ho-
mogeneous agents, i.e., agents whose parameters will not
differ from each other. We will therefore drop the agent
indices from αi, βi, and γi. The heterogeneous agent case
is to be explored in future work.
A. Two-state Matching Pennies
The single-state matching pennies game is a paradig-
matic two-agents, two-actions game. Imagine the situa-
tion of soccer penalty kicks. The keeper (agent 1) can
8
learners for both learning rates α. For the low learning
rate α = 0.02, Q and SARSA learners reach a fixed point
of playing both actions with equal probability in both
states, yielding a reward of 0.5. Due to the low α, this
takes approximately 600 time steps. In contrast, the re-
ward trajectory of the AC learner appears to be chaotic.
Figure 5 confirms this observation, which we will discuss
in more detail below.
For the high learning rate α = 0.8, both Q and SARSA
learners enter a periodic limit cycle. Differences in the
trajectories of Q and SARSA learner are clearly visible.
The time average reward of this periodic orbit appears
to be approximately 0.5 for each agent, identical to the
reward of the fixed point at lower α. The AC learner,
however, converges to a fixed point after oscillating near
the edges of the phase space. At this fixed point agent
1 plays action 1 in state 1 with probability 1. Thus, it
has trapped the system into state 2. In state 2, agent 1
plays action 2 and agent 2 plays action 1 with probabil-
ity 1 and, consequently, agent 1 receives a reward of 1,
whereas agent 2 receives 0 reward. One might ask, Why
does agent 2 not decrease their probability for playing
action 1, thereby increasing their own reward? And, in-
deed, the arrows of the temporal difference error suggest
this change of behavior profile. However, agent 2 cannot
follow because their behavior is trapped on the simplex
of nonzero action probabilities X 2
2a. For only M = 2 ac-
tions, X 2
21 = 1 thus can no longer change, regardless of
the temporal difference error.
Increasing the discount factor to γ = 0.9, we observe
the learning rate α to set the timescale of learning (Fig.
4). The intensity of choice remained β = 5.0. A high
learning rate α = 0.8 corresponds to faster learning in
contrast to a low learning rate α = 0.02. Also, the ratio
of learning timescales is comparable to the inverse ratio of
learning rates. For both α, Q and SARSA learners reach
a fixed point, whereas the AC learners seem to move
chaotically (details to be investigated below). Comparing
the trajectories between the learning rates α, we observe
a similar shape for each pair of learners. However, the
similarity of the AC trajectories decreases at larger time
steps.
So far, we varied two parameters: the discount factor
γ ∈ [0, 1) and the learning rate α ∈ (0, 1). Combining
Figs. 3 and 4, we investigated all four combinations of a
low and a high γ with a low and a high α. We can sum-
marize that Q and SARSA learners converge to a fixed
point for all combinations of discount factor γ and learn-
ing rate α, except when γ is low and α simultaneously
high. AC dynamics seem chaotic for all combinations of
α and γ.
To investigate the relationship between the parameters
more thoroughly, Fig. 5 shows bifurcation diagrams with
the bifurcation parameters α and γ. Additionally, it also
gives the largest Lyapunov exponents for each learner and
each parameter combination. A largest Lyapunov expo-
nent greater than zero is a key characteristic of chaotic
motion. We computed the Lyapunov exponent from the
FIG. 2: Two-state Matching Pennies. Rewards are
given in black type in the payoff tables for each state.
State-transition probabilities are indicated by (blue)
arrows.
choose to jump either to the left or right side of the goal,
and the kicker (agent 2) can choose to kick the ball also
either to the left or the right. If both agents choose the
identical side, then the keeper agent wins, otherwise the
kicker wins.
In the two-state version of the game, according to Ref.
[35], the rules are extend as follows: In state 1 the situ-
ation is as described in the single-state version. When-
ever agent 1 (the keeper) decides to jump to the left, the
environment transitions to state 2, in which the agents
switch roles: Agent 1 now plays the kicker and agent 2
the keeper. From here, whenever agent 1 (now the kicker)
decides to kick to the right side the environment transi-
tion again to state 1 and both agents switch their roles
again.
Figure 2 illustrates this two-state Matching Pennies
games. Formally, the payoff matrices are given by
in state 2 for s(cid:48) ∈ {1, 2}. State transitions are governed
by (cid:0) T1112 T1122
T1212 T1222
(cid:1) = ( 1 1
Thus, by construction, the probability of transition-
ing to the other state is independent of agent 2's ac-
tion. Only agent 1 has agency over the state transi-
tions. By playing a uniformly random behavior pro-
file (X 1
21) = (0.5, 0.5, 0.5, 0.5), both agents
would obtain an average reward of 0.5 per time step.
21, X 2
11, X 2
11, X 1
With Fig. 3 we compare the temporal difference er-
ror in the behavior space sections for each environmental
state at a comparable low discount factor γ ∈ (0, 1) of
γ = 0.1, as well as learning trajectories for an exem-
plary initial condition for two learning rates α ∈ (0, 1),
a low one (α = 0.02) and a high one (α = 0.8). Overall,
we observe a variety of qualitatively different dynamical
regimes, such as fixed points, periodic orbits and chaotic
motion.
Specifically, we see that Q learners and SARSA learn-
ers behave qualitatively similarly in contrast to the AC
(cid:16) R1
in state 1 and(cid:16) R1
R1
111s(cid:48) ,R2
121s(cid:48) ,R2
111s(cid:48) R1
121s(cid:48) R1
211s(cid:48) ,R2
221s(cid:48) ,R2
211s(cid:48) R1
221s(cid:48) R1
R1
(cid:17)
(cid:17)
0,1 1,0
112s(cid:48)
122s(cid:48)
112s(cid:48) ,R2
122s(cid:48) ,R2
=(cid:0) 1,0 0,1
(cid:1)
=(cid:0) 0,1 1,0
(cid:1)
(cid:1) = ( 0 0
0 0 ) and (cid:0) T2111 T2121
212s(cid:48) ,R2
222s(cid:48) ,R2
212s(cid:48)
222s(cid:48)
T2211 T2221
1,0 0,1
1 1 ) .
state 1action 1(left)action 2(right)action 1(left)action 2(right)1,00,11,00,1Agent 2: KickerAgent 1: Keeperstate 2action 1(left)action 2(right)action 1(left)action 2(right)0,11,00,11,0Agent 2: KeeperAgent 1: Kicker 9
FIG. 3: Three learners in two-state Matching Pennies environment for low discount factor γ = 0.1;
intensity of choice β = 5.0. At the top, the temporal difference errors for the Q learner (Eq. 17), SARSA learner
(Eq. 19 ) and Actor-Critic (AC) learner (Eq. 21) are shown in two behavior phase-space sections, one for each state.
The arrows indicate the average direction the temporal difference errors drive the learner toward, averaged over all
phase-space points of the other state. Arrow colors (and shadings) additionally encode their lengths. Selected
trajectories are shown in the phase-space sections, as well as by reward trajectories, plotting the average reward
11, X 2
21, X 2
value (Eq. 22) over time steps. Crosses in the phase-space subsections indicate the initial behavior
11, X 1
(X 1
21) = (0.01, 0.99, 0.3, 0.4). Circles signal the arrival at a fixed point, determined by the absolute
difference of behavior profiles between two subsequent time steps being below = 10−6. Trajectories are shown for
two different learning rates α = 0.02 (light red) and α = 0.8 (dark blue). The bold reward trajectory belongs to
agent 1 and the thin one to agent 2. Note that the temporal difference error is independent from the learning rate α.
A variety of qualitatively different dynamical regimes can be observed.
FIG. 4: Two-state Matching Pennies environment for high discount factor γ = 0.9; otherwise identical to Fig.
3.
analytically derived Jacobian matrix, iteratively used in a
QR decomposition according to Ref. [46]. See Appendix
A for details.
The largest Lyapunov exponent for Q and SARSA
learners align almost perfectly with each other, whereas
the largest Lyapunov exponent of the AC learners be-
haves qualitatively different. We first describe the behav-
ior of the Q and SARSA learner: For high learning rates
α and low farsightedness γ, Fig. 5 shows a periodic orbit
with few (four) points in phase space. Largest Lyapunov
exponents are distinctly below 0 at those regimes.
In-
creasing the farsightedness γ both learners enter a regime
10
FIG. 6: Varying intensity of choice β under
constant α · β in two-state Matching Pennies
environment for discount factor γ = 0.9. On the left
trajectories of the three learners (Q: green dashed,
SARSA: blue straight, Actor-Critic(AC): red dotted)
are shown in the two behavior space sections, one for
each state. On the right, the corresponding reward
trajectories are shown. The initial behavior was
11, X 2
(X 1
21) = (0.01, 0.99, 0.3, 0.4). The bold
reward trajectory belongs to agent 1 and the thin one
11, X 1
21, X 2
to agent 2. One observes the deterministic limit of
Actor-Critic learning to be invariant under constant
α · β and SARSA learning to converge to AC learning
under β → ∞.
decrease as distinctly as for high γ.
The behavior of the Actor-Critic dynamics is quali-
tatively different from the one of Q and SARSA. The
placement of the fixed points on the natural numbers
grid suggests that the AC learner get confined on one of
the 16 (M N Z) corners of the behavior phase space. No
regularity to which fixed point the AC learner converges
can be deduced. The largest Lyapunov exponent is al-
ways above zero and experiences an overall decreasing
behavior. Similarly, for a decreasing bifurcation param-
eter α, the largest Lyapunov exponent tends to decrease
as well. Different from the bifurcation diagram along γ,
for low α the system might enter a periodic motion, but
only for some parameters α. No regularity can be de-
termined at which parameters α the AC learners enter a
periodic motion. A more thorough investigation of the
nonlinear dynamics, especially those of the Actor-Critic
learner, seems of great interest but is, however, beyond
the scope of this article and leaves promising paths for
future work.
Concerning the parameter β, the intensity of choice,
one can infer from the update equations (Eq. 11 com-
bined with Eq. 19 and Eq. 21) that the dynamics for
the AC learner are invariant for a constant product αβ.
This is because the temporal difference error of the Actor-
Critic learner in the deterministic limit is independent of
FIG. 5: Varying discount factor γ and learning
rate α in two-state Matching Pennies
environment for intensity of choice β = 5.0 for the Q
learners (green crosses), the SARSA learners (blue dots)
and the Actor-Critic (AC) learners (red triangles). On
the left, the discount factor γ is varied with learning
rate α = 0.8, as indicated by the gray vertical lines on
the right. On the right, the learning rate α is varied
with discount factor γ = 0.1 as indicated by the gray
vertical lines on the left. The three top panels show the
visited behavior points during 1000 iterations after a
transient period of 100 000 time steps from initial
behavior (X 1
21) = (0.01, 0.99, 0.3, 0.4).
11, X 2
11, X 1
21, X 2
Visited points are mapped to the function
21 + 4X 1
8X 2
11 on the vertical axes to give a
fuller image of the visited behavior profiles. The bottom
21 + 2X 2
11 + X 1
panel shows the corresponding largest Lyapunov
exponents for the three learners. Overall, Q and
SARSA learner behave qualitatively more similarly
than the Actor-Critic learner.
21, X 2
11, X 1
11, X 2
of visiting many points in phase space around the stable
fixed point (X 1
21) = (0.5, 0.5, 0.5, 0.5). The
largest Lyapuonv exponents are close to zero. With in-
creasing γ the distance around this fixed-point solution
decreases until the dynamics converge from a farsighted-
ness γ slightly greater than 0.5 onward. From there the
largest Lyapunov exponent decreases again for further
increasing γ. The same observations can be made along
a decreasing bifurcation parameter α, except that at the
end, for low α, the largest Lyapunov exponents do not
111s(cid:48) ,R2
121s(cid:48) ,R2
111s(cid:48) R1
121s(cid:48) R1
112s(cid:48) ,R2
122s(cid:48) ,R2
112s(cid:48)
122s(cid:48)
211s(cid:48) ,R2
221s(cid:48) ,R2
211s(cid:48) R1
221s(cid:48) R1
212s(cid:48) ,R2
222s(cid:48) ,R2
212s(cid:48)
222s(cid:48)
R1
(cid:17)
(cid:17)
(cid:1)
(cid:1)
10,0 2,2
=(cid:0) 3,3 0,10
=(cid:0) 4,4 0,10
(cid:1) = ( 0.1 0.9
10,0 1,1
0.9 0.1 ) .
agents chose opposite actions.
11
Figure 7 illustrates these game dynamics. Formally,
the payoff matrices are given by
(cid:16) R1
in state 1 and(cid:16) R1
R1
in state 2 for s(cid:48) ∈ {1, 2}, respectively. The corresponding
state transition probabilities are given by
(cid:0) T1112 T1122
T1212 T1222
(cid:1) =(cid:0) T2111 T2121
T2211 T2221
To be precise, the rewards in each state do not resem-
ble a classical social dilemma situation. This is because if
both agents would alternately cooperate and defect, both
could receive a larger reward per time step compared to
always cooperating. Hence, this stochastic game, as it
was used in Refs. [33 -- 35], presents more a coordination
than a cooperation challenge to the agents. The mul-
tistate environment can here function as a coordination
device.
A behavior profile in which one agent exploits the other
in one state, while being exploited in the other state,
would result in an average reward per time step of 5 for
each agent, e.g., (X 1
21) = (0, 1, 1, 0).
11, X 2
11, X 1
21, X 2
However, for all three learning types with a mid-ranged
farsightedness (γ = 0.45) and an intensity of choice
β = 5.0, the temporal difference error arrows are point-
ing on average toward the lower left defection-defection
point for each state in behavior phase space (Fig. 8). To
see whether the three learning types may converge to the
described defect-cooperate-cooperate-defect equilibrium,
individual trajectories from two exemplary initial condi-
tions and for two learning rates α are shown, a small one
(α = 0.02) and a high one (α = 0.8).
We observe qualitatively different behavior across all
three learners. The Q learners converge to equilibria
with average rewards distinctly below 5, and the SARSA
learners converge to equilibira with average rewards of
almost 5 for both learning rates α and both exemplary
initial conditions. Both Q and SARSA learners con-
verge to solutions of proper probabilistic behavior, i.e.,
choosing action cooperate and action defect with nonva-
nishing chance. The Actor-Critic learners, on the other
hand, converge to the deterministic defect-cooperate-
cooperate-defect behavior described above for the initial
condition shown with the non-dashed lines in Fig. 8 for
both learning rates α (shown in light red and dark blue).
For the other exemplary initial condition, shown with the
dashed lines, it converges to an all-defection solution in
both states for both α.
Interestingly, for all learners, all combinations of initial
conditions and learning rates converge to a fixed point
solution, except for the Q learners with a comparably
high learning rate α = 0.8, which enter a periodic behav-
ior solution for the initial condition with the nondashed
FIG. 7: Two-state Prisoners Dilemma. Rewards
are given in black type in the payoff tables for each
state. State-transition probabilities are indicated by
(blue) arrows.
β. Further, the dynamics of the SARSA learner will con-
verge to the dynamics of the AC learner under β → ∞.
Figure 6 nicely confirms these two observations. Observ-
ing Table I is another way to see this. Since the value
estimate of the future state is identical for SARSA and
AC learning, letting the value estimate of the current
state vanish by sending β → ∞ makes the SARSA learn-
ers approximate the AC learners.
sa).
11, X 2
11, X 1
21, X 2
As mentioned before, β controls the exploration-
exploitation trade-off.
In the temporal difference er-
rors of the Q and SARSA learner it appears in the
term indicating the value estimate of the current state
−1/βi log(X i
If this term dominates the tempo-
ral difference error (i.e., if β is small), then the learn-
ers tend toward the center of behavior space,
i.e.,
(X 1
21) = (0.5, 0.5, 0.5, 0.5), forgetting what
they have learned about the obtainable reward. This
characteristic happens to be favorable in our two-
state Matching Pennies environment, which is why Q
and SARSA learners perform better in finding the
(X 1
21) = (0.5, 0.5, 0.5, 0.5) solution. On
the other hand, if β is large, then the temporal difference
error is dominated by the current reward and future value
estimate. Not being able to forget, the learners might get
trapped in unfavorable behavior, as we can see observ-
ing the Actor-Critic learners. To calibrate β it is useful
to make oneself clear that it must come in units of [log
behavior] / [reward].
21, X 2
11, X 2
11, X 1
B. Two-state Prisoner's Dilemma
The single state Prisoners Dilemma is another paradig-
matic two-agents, two-actions game. It has been used to
model social dilemmas and study the emergence of coop-
eration. It describes a situation in which two prisoners
are separately interrogated, leaving them with the choice
to either cooperate with each other by not speaking to
the police or defecting by testifying.
The two-state version, which has been used as a test
environment also in Refs. [33 -- 35], extends this situation
somewhat artificially by playing a Prisoner's Dilemma in
each of the two states with a transition probability of
10% from one state to the other if both agents chose the
same action, and a transition probability of 90% if both
state 1action 1(coop.)action 2(defect)action 1(coop.)action 2(defect)3,30,102,210,0Agent 2Agent 1state 2action 1(coop.)action 2(defect)action 1(coop.)action 2(defect)4,40,101,110,0Agent 2Agent 1 12
FIG. 8: Two-state Prisoners Dilemma environment for discount factor γ = 0.45; otherwise identical to Fig. 3.
line. The same phenomenon occurred also in the Match-
ing Pennies environment for low farsightedness γ = 0.1,
however, there for both Q and SARSA learners. It seems
to be caused by the comparably high learning rate. A
high learning rate overshoots the behavior update, re-
sulting in a circling behavior around the fixed point. As
in Fig. 3, the time average reward of the periodic orbit
seems to be comparable to the reward of the correspond-
ing fixed point at lower α. Furthermore, we observe the
same time rescaling effect of the learning rate α in Figure
8 as in Fig. 4.
To visualize the influence of the discount factor γ
on the converged behavior, Fig. 9 shows a bifurca-
tion diagram along the bifurcation parameter γ for two
initial conditions. Pluses in blue result from a uni-
formly random behavior profile of (X 1
21) =
(0.5, 0.5, 0.5, 0.5), whereas the crosses in red initially
started from the behavior profile (X 1
21) =
(0.51, 0.49, 0.49, 0.51).
11, X 2
11, X 2
11, X 1
21, X 2
11, X 1
21, X 2
Across all learners, lower discount factors γ correspond
to all-defect solutions, whereas for higher γ the solu-
tions from the initial condition shown with red crosses
tend toward the cooperate-defect-defect-cooperate solu-
tion. For low γ, the agents are less aware of the pres-
ence of other states and find the all-defect equilibrium
solution of the iterated normal form Prisoner's Dilemma.
The state transition probabilities have less effect on the
learning dynamics. Only above a certain farsightedness
do the agents find the more rewarding cooperate-defect-
defect-cooperate solution.
11 and X 2
The observation from Fig. 8 is confirmed that the prob-
ability to cooperate (i.e., here X 1
21 ) is lowest for
the Q learners, midrange for the SARSA learners and 1
for the Actor-Critic learners. One reason for this observa-
tion can be found in the intensity of choice parameter β.
It balances the reward obtainable in the current behavior
space segment with the forgetting of current knowledge
to be open to new solutions. Such forgetting expresses
itself by temporal difference error components pointing
toward the center of behavior space. Thus, a relatively
small β = 5.0 can explain why solutions at the edge of
the behavior space cannot be reached by Q and SARSA
learners. The AC learner misses this forgetting term in
the deterministic limit and can therefore easily enter be-
havior profiles at the edge of the behavior space.
Q and SARSA learners have a critical discount fac-
tor γ above which the cooperate-defect-defect-cooperate
high reward solution is obtained and below which the
all-defect low reward solution gets selected. However,
for increasing discount factors γ up to 1, Q and SARSA
learners experience a drop in playing the cooperative ac-
tion probability.
The Actor-Critic learners approach the cooperate-
defect-defect-cooperate solution in two steps. For in-
creasing γ, first the probability of agent 2 cooperating
in state 2 (X 2
21) jumps from zero to 1 while agent 1 still
defects in state 1. Only after a slight increase of γ does
agent 1 also cooperate in state 1 (X 1
11).
Interestingly, for the uniformly random initial behav-
ior condition shown with blue pluses, there is no criti-
cal discount factor γ and no learners come close to the
cooperate-defect-defect-cooperate solution. Here, only
for γ close to 1 do all cooperation probabilities X i
s1
gradually increase. Furthermore, exactly at those γ,
where the cooperate-defect-defect-cooperate solution is
obtained from the initial behavior condition shown with
red crosses, the solutions from the uniformly random ini-
tial behavior condition (blue pluses) have a largest Lya-
puonv exponent greater than 0. At other values of γ,
the largest Lyapunov exponents for the two initial con-
ditions overlap. This suggests that the largest Lyapunov
exponents greater than zero may point to the fact that
other, perhaps more rewarding solutions may exist in
phase space. A more thorough investigation regarding
this multistability is an open point for future research.
As we have argued above, the two-state Prisoner's
13
FIG. 9: Varying discount factor γ in two-state Prisoners Dilemma environment for learning rate α = 0.2
and intensity of choice β = 5.0 for the Q learners on the left, the SARSA learners in the middle and the Actor-Critic
learners on the right. The four top panels for each learner show the visited behavior points X 1
11, X 2
11, X 1
21, X 2
21
during 1000 iterations after a transient period of 5000 time steps from initial behavior
11, X 2
(X 1
21, X 2
11, X 2
(X 1
21) = (0.51, 0.49, 0.49, 0.51) in red crosses. The bottom panels show the corresponding largest
Lyapunov exponents for the two initial conditions. Above a critical discount factor γ all learners find the high
rewarding solution from the red crosses initial condition, but do not do so from the blue pluses initial condition.
21) = (0.5, 0.5, 0.5, 0.5) in blue pluses and from initial behavior
11, X 1
21, X 2
11, X 1
also capable of solving a cooperation challenge in a
stochastic game setting, for which we adapt a two-state
Prisoner's Dilemma in analogy to Ref.
[45]. Figure 10
confirms previous findings that cooperation emerges only
in the stochastic game, compared to playing each Pris-
oner's Dilemma repeatedly [45]. Further, cooperation
only emerges for sufficiently large farsightedness γ.
V. DISCUSSION
The main contribution of this paper is the develop-
ment of a technique to obtain the deterministic limit
of temporal difference reinforcement learning. Through
our work we have combined the literature on learning
dynamics from statistical physics with the evolutionary
game theory-inspired learning dynamics literature from
machine learning. For the statistical physics community,
the novelty consists of learning equations, capable of han-
dling environmental state transitions. For the machine
learning community the novelty lies in the systematic
methodology we have used to obtain the deterministic
learning equations.
We have demonstrated our approach with the three
prominent reinforcement learning algorithms from com-
puter science: Q learning, SARSA learning, and Actor-
Critic learning. A comparison of their dynamics in pre-
viously used two-agent, two-actions, two-states environ-
ments has revealed the existence of a variety of qualita-
tively different dynamical regimes, such as convergence
to fixed points, periodic orbits and deterministic chaos.
We have found that Q and SARSA learners tend
to behave qualitatively more similar in comparison to
the Actor-Critic learning dynamics. This characteris-
FIG. 10: Cooperation challenge in a two-state
Prisoner's Dilemma. Top panel shows a two-state
Prisoner's Dilemma game, whose state games
individually favor defection. Bottom panels shows the
level of cooperation SARSA learners with α = 0.016,
β = 250 play after reaching a fixed point from the
center of behavior space (X i
sa = 0.5 for all i, s, a) for
varying discount factors γ. Results for Q and AC
learners are similar. Cooperation levels are shown for
the full stochastic game as well as for each individual
state game played repeatedly. For sufficiently large
farsightedness, cooperation can emerge in the stochastic
game, in contrast to the individual repeated games.
Dilemma as it was used in Refs. [33 -- 35] presents rather a
coordination than a cooperation challenge to the agents.
Figure 10 demonstrates that our learning dynamics are
state 1game 1action 1(coop.)action 2(defect)action 1(coop.)action 2(defect)9,9-1,100,010,-1Agent 2Agent 1state 2game 2action 1(coop.)action 2(defect)action 1(coop.)action 2(defect)1,1-1,20,02,-1Agent 2Agent 1 tic results at least partly from our relatively low inten-
sity of choice parameter β, controlling the exploration-
exploitation trade-off via a forgetting term in the tempo-
ral difference errors. Sending β → ∞, the SARSA learn-
ing dynamics approach the Actor-Critic learning dynam-
ics, as we have shown. Overall the Actor-Critic learn-
ers have a tendency to enter confining behavior profiles,
due to their nonexisting forgetting term. This charac-
teristic leaves them trapped at the edges of the behavior
space. In contrast, Q and SARSA learner do not show
such learning behavior. Interestingly, this characteristic
of the AC learners turns out to be favorable in the two-
state Prisoner's Dilemma environment, where they find
the most rewarding solution in more cases compared to
Q and SARSA but hinders the convergence to the fixed
point solution in the two-state Matching Pennies envi-
ronment. Thus, the most favorable level of forgetting
depends on the environment. In order to tune the respec-
tive parameter β, our consideration that it must come in
the unit of [log behavior] / [reward] may be helpful.
We have demonstrated the effect of the learning rate α
adjusting the speed of learning by controlling the amount
of new information used in a behavior profile update.
Thereby, within limits, α functions as a time rescaling.
However, a comparably large learning rate α might cause
an overshooting phenomenon, hindering the convergence
to a fixed point. Instead, the learners enter a limit cy-
cle around that point. Nevertheless, the average reward
of the limit-cycling behavior was approximately equal to
the one of the fixed point obtained at lower α but took
fewer time steps to reach. Thus, perhaps other dynamical
regimes than fixed points, such as limit cycles or strange
attractors, could be of interest in some applications of
reinforcement learning.
We have also shown the effect of the discount factor
γ adjusting the farsightedness of the agents. At low γ
the state transition probabilities have less effect on the
learning dynamics compared to high discount factors.
To summarize the three parameters α, β, and γ: The
level of exploitation β and the farsightedness γ con-
trol where the learner adapts toward in behavior space,
weighting current reward, expected future reward and
the level of forgetting. The learning rate α controls how
fast the learner adapts along these directions.
We hope that our work might turn out useful for the
application of reinforcement learning in various domains,
14
joint-action) learners (cf. Ref.
with respect to parameter tuning, the design of new al-
gorithms, and the analysis of complex strategic interac-
tions using meta strategies, as Bloembergen et al.
[28]
have pointed out. In this regard, future work could ex-
tend the presented methodology to partial observability
of the Markov states of the environment [40, 41], be-
havior profiles with history, and other-regarding agent
(i.e.
[2] for an overview
of other-regarding agent learning algorithms). Also, the
combination of individual reinforcement learning and so-
cial learning through imitation [47 -- 50] seems promising.
Such endeavors would naturally lead to the exploration of
network effects. It is important to note that only a few
dynamical systems reinforcement learning studies have
begun to incorporate network structures between agents
[22, 23].
Apart from these more technical extensions, we hope
that our learning equations will prove themselves useful
when studying the evolution of cooperation in stochastic
games [45]. With stochastic games one is able to explic-
itly account for a changing environment. Therefore, such
studies are likely to contribute to the advancement of
theoretical research on the sustainability of interlinked
social-ecological systems [51, 52].
Interactions, syner-
gies, and trade-offs between social [13, 53] and ecological
[54] dilemmas can be explored using the framework of
stochastic games. More realistic environments, modeling
e.g., the harvesting of common-pool renewable resources
[55, 56] or the prevention of dangerous climate change
[57, 58], for our learning dynamics are likely to prove
themselves useful. Here, it may be of interest to evaluate
the learning process not only in terms of efficiency but
also how close it came to the optimal behavior. Other
paradigms than value optimization may also be impor-
tant [59], such as sustainability or resilience [60].
Python code to reproduce the figures of this article is available
at https://doi.org/10.5281/zenodo.1495091.
ACKNOWLEDGMENTS
This work was developed in the context of the COPAN project
on Coevolutionary Pathways in the Earth System at the Pots-
dam Institute for Climate Impact Research (PIK). We are grate-
ful for financial support by the Heinrich Boll Foundation, the
Stordalen Foundation (via the Planetary Boundaries Research Net-
work PB.net), the Earth League's EarthDoc program and the Leib-
niz Association (project DOMINOES). We thank Jobst Heitzig for
discussions and comments on the manuscript.
Appendix A: Computation of Lyapunov Exponents
We compute the Lyapunov exponents using an iterative QR decomposition of the Jacobian matrix according to Sandri [46]. In the
following we present the derivation of the Jacobian matrix.
Eq. 11 constitutes a map f , which iteratively updates the behavior profile X ∈ RN×M×Z . Consequently, we can represent its derivative
as a Jacobian tensor f(cid:48)(X) ∈ RN×M×Z×N×M×Z .
Let Ai
sa := X i
sa exp
(cid:16)
αiβiTDi
sa(X)
(cid:17)
be the numerator of Eq. 11, and Bi
s :=(cid:80)
f(cid:48)(X) =
A(cid:48)B − B(cid:48)A
B2
b Ai
sb its denominator, i.e. f =: A/B. Hence,
or, more precisely, in components,
s(X) − dBi
Bi
s
dXj
rb
s(X))2
A and B are known, and if A(cid:48) is known, then B(cid:48) is easily obtained by dBi
dXj
rb
df i
sa(X)
dX j
rb
sa(X)
dXj
rb
(Bi
dAi
=
s(X)
(X)Ai
sa(X)
=(cid:80)
.
dAi
sc(X)
dXj
rb
c
the three learner types Q, SARSA and Actor-Critic learning.
. Therefore we need to compute A(cid:48) for
Let us rewrite Ai
sa for the Q learner according to
Ai
sa := (X i
sa)(1−αi) exp
(cid:16)
αiβi TD
1. Q learning
(cid:17)
i
sa(X)
,
where we removed the estimate of the current value from the temporal difference error, leaving the truncated TD error as
TD
i
sa(X) := (1 − γi)
TX−i(cid:104)R(cid:105)i
sa + γi maxQi
sa(X).
Hence, we can write the derivative of A as
Since(cid:80)
dAi
sa(X)
dX j
rb
sa/dX j
c X i
sc = 1, dX i
rb can be expressed as
(cid:16)
(cid:17)(cid:32)
= exp
αiβi TD
i
sa(X)
(1 − αi)(X i
sa)−αi dX i
sa
dX j
rb
+ αiβi(X i
sa)(1−αi) d TD
i
sa(X)
dX j
rb
(cid:33)
.
The derivative of the truncated temporal difference error reads
= δij δsr(2δab − 1).
dX i
sa
dX j
rb
d TD
i
sa(X)
dX j
rb
= (1 − γi)
d
+ γi d maxQi
dX j
rb
sa(X)
.
sa
TX−i(cid:104)R(cid:105)i
dX j
rb
(cid:88)
dX−i
sa−i
dX j
rb
(cid:88)
s(cid:48)
a−i
Let us write the derivative of the reward as
d
TX−i(cid:104)R(cid:105)i
dX j
rb
using Eq. 2 and Eq. 3, where the derivatives dX−i
dX−i
sa−i
dX j
rb
d maxQi
dX j
rb
(cid:88)
(cid:88)
sa(X)
=
s(cid:48)
a−i
sa−i /dX j
sa
=
For the derivative of the maximum next value we write accordingly
rb need to be executed according to Eq. A6.
Tsaa−is(cid:48) max
Qi
s(cid:48)c(X) +
c
X−i
sa−i Tsaa−is(cid:48)
s(cid:48)c(X)
d maxc Qi
dX j
rb
.
Tsaa−is(cid:48) Ri
saa−is(cid:48)
(cid:88)
(cid:88)
s(cid:48)
a−i
15
(A1)
(A2)
(A3)
(A4)
(A5)
(A6)
(A7)
(A8)
(A9)
(A10)
(A11)
(A12)
Let am := arg maxa Qi
sa(X), then
and
dQi
sa(X)
dX j
rb
= (1 − γi)
d
d maxc Qi
dX j
rb
TX−i(cid:104)R(cid:105)i
dX j
rb
sa
For the derivative of the effective Markov Chain transition tensor we can write
dX j
rb
using Eqs. 2 and 3, where again the derivatives dXsa/dX j
For the derivative of the state value let us rewrite Eq. 6 as V i
d X(cid:104)T(cid:105)ss(cid:48)
sc(X)
= δaam
dQi
sa(X)
dX j
rb
d X(cid:104)T(cid:105)ss(cid:48)
s(cid:48)
dX j
rb
+ γi(cid:88)
(cid:88)
s = (1 − γi)(cid:80)
dXsa
dX j
rb
=
a
Tsaa−is(cid:48) ,
s(cid:48) (X) + X(cid:104)T(cid:105)ss(cid:48)
V i
dV i
s(cid:48) (X)
dX j
rb
.
dV i
s (X)
dX j
rb
= (1 − γi)
rb need to be executed according to Eq. A6.
(cid:88)
s(cid:48)(cid:48)
d(M−1
ss(cid:48)(cid:48) )
dX j
rb
s(cid:48) M−1
s(cid:48)(cid:48) + M−1
ss(cid:48)(cid:48)
TX(cid:104)R(cid:105)i
d TX(cid:104)R(cid:105)i
s(cid:48)(cid:48)
dX j
rb
ss(cid:48) TX(cid:104)R(cid:105)i
s(cid:48) with M := (1Z − γi
X(cid:104)T(cid:105)). Thus,
.
(A13)
To obtain the derivative of the inverse matrix M−1 we use (M−1M )(cid:48) = 0 = (M−1)(cid:48)M + M−1M(cid:48) and therefore (M−1)(cid:48) =
−M−1M(cid:48)M−1. For M(cid:48) we write,
dMss(cid:48)
dX j
rb
= −γi d X(cid:104)T(cid:105)ss(cid:48)
dX j
rb
.
(A14)
We obtain the derivative of the reward according to
d TX(cid:104)R(cid:105)i
s
dX j
rb
=
(cid:88)
(cid:88)
s(cid:48)
a
dXsa
dX j
rb
Tsas(cid:48) Ri
sas(cid:48) ,
16
(A15)
using Eq. 1 and Eq. 3, where the derivatives dX i
sa/dX j
rb need to be executed according to Eq. A6.
Now we can compute the Jacobian matrix for the Q learning dynamics in their deterministic limit.
The computation of the Jacobian matrix for the SARSA learning update in its deterministic limit is similar, except the truncated
2. SARSA learning
temporal difference error reads
instead of Eq. A4. Hence,
TD
i
sa(X) := (1 − γi)
d TD
i
sa(X)
= (1 − γi)
d
and
The derivative of(cid:80)
sa(X)
d nextQi
dX j
rb
s(cid:48)cQi
c X i
=
s(cid:48)
a−i
s(cid:48)c(X) reads
(cid:88)
dX j
rb
dX−i
sa−i
dX j
rb
(cid:88)
d[(cid:80)
TX−i(cid:104)R(cid:105)i
TX−i(cid:104)R(cid:105)i
dX j
rb
sa
sa + γi nextQi
sa(X).
sa(X)
,
+ γi d nextQi
dX j
(cid:88)
rb
(cid:88)
X−i
s(cid:48)
a−i
(cid:88)
c
=
(cid:32)
(cid:88)
c
dX i
s(cid:48)c
dX j
rb
(cid:33)
.
dQi
s(cid:48)c
dX j
rb
Tsaa−is(cid:48)
X i
s(cid:48)cQi
s(cid:48)c(X) +
sa−i Tsaa−is(cid:48)
d[(cid:80)
s(cid:48)c(X)]
c X i
s(cid:48)cQi
dX j
rb
s(cid:48)c(X)]
c X i
s(cid:48)cQi
dX j
rb
Qi
s(cid:48)c(X) + X i
s(cid:48)c
All remaining terms have already been given in the previous section for the Q learner Jacobian matrix.
3. Actor-Critic learning
For the Actor-Critic learning update, Eq. A3 reads
Ai
sa := X i
sa exp
(cid:16)
αiβi TD
(cid:17)
i
sa(X)
,
with the truncated temporal difference error
The derivative of the next value estimate is obtained by
dX−i
sa−i
dX j
rb
d nextV i
dX j
rb
sa(X)
=
s(cid:48)
a−i
i
TD
sa(X) := (1 − γi)
(cid:88)
(cid:88)
TX−i(cid:104)R(cid:105)i
sa(X).
sa + γi nextV i
(cid:88)
(cid:88)
s(cid:48)
a−i
Tsaa−is(cid:48) V i
s(cid:48) (X) +
X−i
sa−i Tsaa−is(cid:48)
dV i
s(cid:48) (X)
dX j
rb
.
(A16)
(A17)
.
(A18)
(A19)
(A20)
(A21)
(A22)
The derivative of the next value V i
learning update.
s(cid:48) is given by Eq. A13. These are all terms necessary to compute the Jacobian matrix for the Actor-Critic
[1] Richard S Sutton and Andrew G Barto, Reinforcement
learning: An introduction (MIT Press, 1998).
[2] L. Busoniu, R. Babuska, and B. De Schutter, "A com-
prehensive survey of multiagent reinforcement learning,"
IEEE Transactions on Systems, Man, and Cybernetics,
Part C (Applications and Reviews) 38, 156 -- 172 (2008).
[3] M Wiering and Merendel van Otterlo, Reinforcement
Learning: State-of-the-Art (Springer Berlin Heidelberg,
2012).
[4] Ashvin Shah, "Psychological and neuroscientific con-
nections with reinforcement learning," in Reinforcement
Learning (Springer Berlin Heidelberg, 2012) pp. 507 -- 537.
[5] Demis Hassabis, Dharshan Kumaran, Christopher Sum-
and Matthew Botvinick, "Neuroscience-
intelligence," Neuron 95, 245 -- 258
merfield,
inspired artificial
(2017).
[6] Drew Fudenberg and David K Levine, The Theory of
Learning in Games (MIT Press, 1998).
[7] Alvin E Roth and Ido Erev, "Learning in extensive-form
games: Experimental data and simple dynamic models in
the intermediate term," Games and Economic Behavior
8, 164 -- 212 (1995).
[8] Ido Erev and Alvin E Roth, "Predicting how people play
games: Reinforcement learning in experimental games
with unique, mixed strategy equilibria," The American
Economic Review 88, 848 -- 881 (1998).
[9] Colin Camerer and Teck Hua Ho, "Experience-weighted
attraction learning in normal form games," Econometrica
67, 827 -- 874 (1999).
[10] Colin F Camerer, Behavioral Game Theory: Exper-
iments in Strategic Interaction (Princeton University
Press, 2003).
[11] W Brian Arthur, "On designing economic agents that be-
have like human agents," Journal of Evolutionary Eco-
nomics 3, 1 -- 22 (1993).
[12] W Brian Arthur, "Complexity and the economy," Science
284, 107 -- 109 (1999).
[13] Michael W Macy and Andreas Flache, "Learning dy-
namics in social dilemmas," Proceedings of the National
Academy of Sciences 99, 7229 -- 7236 (2002).
[14] John G. Cross, "A stochastic learning model of economic
behavior," The Quarterly Journal of Economics 87, 239
(1973).
[15] Tilman Borgers and Rajiv Sarin, "Learning through rein-
forcement and replicator dynamics," Journal of Economic
Theory 77, 1 -- 14 (1997).
[16] Matteo Marsili, Damien Challet, and Riccardo Zecchina,
"Exact solution of a modified el farol's bar problem:
Efficiency and the role of market impact," Physica A:
Statistical Mechanics and its Applications 280, 522 -- 553
(2000).
[17] Yuzuru Sato, Eizo Akiyama,
and J Doyne Farmer,
"Chaos in learning a simple two-person game," Proceed-
ings of the National Academy of Sciences 99, 4748 -- 4751
(2002).
[18] Yuzuru Sato and James P Crutchfield, "Coupled replica-
tor equations for the dynamics of learning in multiagent
systems," Physical Review E 67 (2003), 10.1103/phys-
reve.67.015206.
[19] Yuzuru Sato, Eizo Akiyama, and James P Crutchfield,
"Stability and diversity in collective adaptation," Physica
D: Nonlinear Phenomena 210, 21 -- 57 (2005).
[20] Tobias Galla, "Intrinsic noise in game dynamical learn-
ing," Physical Review Letters 103 (2009), 10.1103/phys-
revlett.103.198702.
[21] Tobias Galla, "Cycles of cooperation and defection in im-
perfect learning," Journal of Statistical Mechanics: The-
ory and Experiment 2011, P08007 (2011).
[22] Alex J. Bladon and Tobias Galla, "Learning dynamics
in public goods games," Physical Review E 84 (2011),
10.1103/physreve.84.041132.
[23] John Realpe-Gomez, Bartosz Szczesny, Luca Dall'Asta,
and Tobias Galla, "Fixation and escape times in stochas-
tic game learning," Journal of Statistical Mechanics:
Theory and Experiment 2012, P10022 (2012).
[24] James BT Sanders, Tobias Galla,
and Jonathan L
Shapiro, "Effects of noise on convergent game-learning
dynamics," Journal of Physics A: Mathematical and The-
oretical 45, 105001 (2012).
[25] Tobias Galla and J. Doyne Farmer, "Complex dynam-
ics in learning complicated games," Proceedings of the
National Academy of Sciences 110, 1232 -- 1236 (2013).
[26] Aleksandra Aloric, Peter Sollich, Peter McBurney, and
Tobias Galla, "Emergence of Cooperative Long-Term
Market Loyalty in Double Auction Markets," PloS ONE
11, e0154606 (2016).
[27] Karl Tuyls, Katja Verbeeck,
and Tom Lenaerts, "A
selection-mutation model for q-learning in multi-agent
17
systems," in Proceedings of the Second International
Joint Conference on Autonomous Agents and Multiagent
Systems, AAMAS 2003 (2003) pp. 693 -- 700.
[28] Daan Bloembergen, Karl Tuyls, Daniel Hennes,
and
Michael Kaisers, "Evolutionary dynamics of multi-agent
learning: A survey," Journal of Artificial Intelligence Re-
search 53, 659 -- 697 (2015).
[29] Karl Tuyls and Ann Now´e, "Evolutionary game theory
and multi-agent reinforcement learning," The Knowledge
Engineering Review 20, 63 -- 90 (2005).
[30] Karl Tuyls, Pieter Jan'T Hoen, and Bram Vanschoen-
winkel, "An evolutionary dynamical analysis of multi-
agent learning in iterated games," Autonomous Agents
and Multi-Agent Systems 12, 115 -- 153 (2006).
[31] Karl Tuyls and Simon Parsons, "What evolutionary game
theory tells us about multiagent learning," Artificial In-
telligence 171, 406 -- 416 (2007).
[32] Michael Kaisers and Karl Tuyls, "Frequency adjusted
multi-agent q-learning," in Proceedings of the 9th Inter-
national Conference on Autonomous Agents and Multia-
gent Systems: Volume 1, AAMAS 2010 (2010) pp. 309 --
315.
[33] Daniel Hennes, Karl Tuyls, and Matthias Rauterberg,
"State-coupled replicator dynamics," in Proceedings of
the 8th International Conference on Autonomous Agents
and Multiagent Systems, AAMAS 2009 (2009) pp. 789 --
796.
[34] Peter Vrancx, Karl Tuyls, and Ronald Westra, "Switch-
ing dynamics of multi-agent learning," in Proceedings of
the 7th International Joint Conference on Autonomous
Agents and Multiagent systems, AAMAS 2008 (2008) pp.
307 -- 313.
[35] Daniel Hennes, Michael Kaisers, and Karl Tuyls, "Resq-
learning in stochastic games," in Proceedings of the Adap-
tive and Learning Agents Workshop, ALA 2010 (2010)
pp. 8 -- 15.
[36] Lloyd S Shapley, "Stochastic games," Proceedings of the
National Academy of Sciences 39, 1095 -- 1100 (1953).
[37] J-F Mertens and Abraham Neyman, "Stochastic games,"
International Journal of Game Theory 10, 53 -- 66 (1981).
[38] Eizo Akiyama and Kunihiko Kaneko, "Dynamical sys-
tems game theory and dynamics of games," Physica D:
Nonlinear Phenomena 147, 221 -- 258 (2000).
[39] Eizo Akiyama and Kunihiko Kaneko, "Dynamical sys-
tems game theory II: A new approach to the problem of
the social dilemma," Physica D: Nonlinear Phenomena
167, 36 -- 71 (2002).
[40] Matthijs T. J. Spaan, "Partially observable markov de-
cision processes," in Reinforcement Learning (Springer
Berlin Heidelberg, 2012) pp. 387 -- 414.
[41] Frans A. Oliehoek, "Decentralized POMDPs," in Adap-
tation, Learning, and Optimization (Springer Berlin Hei-
delberg, 2012) pp. 471 -- 503.
[42] Richard Bellman, "A markovian decision process," Indi-
ana University Mathematics Journal 6, 679 -- 684 (1957).
[43] Sascha Lange, Thomas Gabel, and Martin Riedmiller,
"Batch reinforcement learning," in Reinforcement Learn-
ing (Springer Berlin Heidelberg, 2012) pp. 45 -- 73.
[44] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, An-
drei A. Rusu, Joel Veness, Marc G. Bellemare, Alex
Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg
Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik,
Ioannis Antonoglou, Helen King, Dharshan Kumaran,
Daan Wierstra, Shane Legg,
and Demis Hassabis,
"Human-level control through deep reinforcement learn-
ing," Nature 518, 529 -- 533 (2015).
[45] Christian Hilbe, Step´an Simsa, Krishnendu Chatterjee,
and Martin A. Nowak, "Evolution of cooperation in
stochastic games," Nature 559, 246 -- 249 (2018).
[46] Marco Sandri, "Numerical calculation of lyapunov expo-
nents," The Mathematica Journal 6, 78 -- 84 (1996).
[47] A. Bandura, Social learning Theory (Englewood Cliffs,
NJ: Prentice Hall, 1977).
[48] Marco Smolla, R. Tucker Gilman, Tobias Galla,
and
Susanne Shultz, "Competition for resources can explain
patterns of social and individual learning in nature," Pro-
ceedings of the Royal Society B: Biological Sciences 282,
20151405 (2015).
[49] Wolfram Barfuss, Jonathan F Donges, Marc Wieder-
mann,
and Wolfgang Lucht, "Sustainable use of re-
newable resources in a stylized social -- ecological network
model under heterogeneous resource distribution," Earth
System Dynamics 8, 255 -- 264 (2017, P1).
[50] S. Banisch and E. Olbrich, "Opinion polarization by
learning from social feedback," The Journal of Mathe-
matical Sociology 0, 1 -- 28 (2018).
[51] Simon Levin, "The mathematics of sustainability," No-
tices of the American Mathematical Society 60, 1 (2013).
[52] Jonathan F Donges, Ricarda Winkelmann, Wolfgang
Lucht, Sarah E Cornell, James G Dyke, Johan Rock-
strom, Jobst Heitzig, and Hans Joachim Schellnhuber,
"Closing the loop: Reconnecting human dynamics to
earth system science," The Anthropocene Review 4, 151 --
157 (2017).
[53] R M Dawes, "Social dilemmas," Annual Review of Psy-
18
chology 31, 169 -- 193 (1980).
[54] J. Heitzig, T. Kittel, J. F. Donges,
and N. Molken-
thin, "Topology of sustainable management of dynami-
cal systems with desirable states: from defining planetary
boundaries to safe operating spaces in the earth system,"
Earth System Dynamics 7, 21 -- 50 (2016).
[55] Emilie Lindkvist and Jon Norberg, "Modeling experi-
ential learning: The challenges posed by threshold dy-
namics for sustainable renewable resource management,"
Ecological Economics 104, 107 -- 118 (2014).
[56] Caroline Schill, Therese Lindahl,
and Anne-Sophie
Cr´epin, "Collective action and the risk of ecosystem
regime shifts:
insights from a laboratory experiment,"
Ecology and Society 20 (2015), 10.5751/es-07318-200148.
[57] Manfred Milinski, Ralf D Sommerfeld, Hans-Jurgen
Krambeck, Floyd A Reed, and Jochem Marotzke, "The
collective-risk social dilemma and the prevention of simu-
lated dangerous climate change," Proceedings of the Na-
tional Academy of Sciences 105, 2291 -- 2294 (2008).
[58] S. Barrett and A. Dannenberg, "Climate negotiations un-
der scientific uncertainty," Proceedings of the National
Academy of Sciences 109, 17372 -- 17376 (2012).
[59] Wolfram Barfuss, Jonathan F Donges, Steven J Lade,
and Jurgen Kurths, "When optimization for governing
human-environment tipping elements is neither sustain-
able nor safe," Nature communications 9, 2354 (2018,
P4).
[60] Jonathan F. Donges and Wolfram Barfuss, "From math
to metaphors and back again: Social-ecological resilience
from a multi-agent-environment perspective," GAIA -
Ecological Perspectives for Science and Society 26, 182 --
190 (2017, P2).
|
0811.2551 | 3 | 0811 | 2019-07-09T20:25:22 | Modeling Cultural Dynamics | [
"cs.MA",
"cs.AI",
"q-bio.NC"
] | EVOC (for EVOlution of Culture) is a computer model of culture that enables us to investigate how various factors such as barriers to cultural diffusion, the presence and choice of leaders, or changes in the ratio of innovation to imitation affect the diversity and effectiveness of ideas. It consists of neural network based agents that invent ideas for actions, and imitate neighbors' actions. The model is based on a theory of culture according to which what evolves through culture is not memes or artifacts, but the internal models of the world that give rise to them, and they evolve not through a Darwinian process of competitive exclusion but a Lamarckian process involving exchange of innovation protocols. EVOC shows an increase in mean fitness of actions over time, and an increase and then decrease in the diversity of actions. Diversity of actions is positively correlated with population size and density, and with barriers between populations. Slowly eroding borders increase fitness without sacrificing diversity by fostering specialization followed by sharing of fit actions. Introducing a leader that broadcasts its actions throughout the population increases the fitness of actions but reduces diversity of actions. Increasing the number of leaders reduces this effect. Efforts are underway to simulate the conditions under which an agent immigrating from one culture to another contributes new ideas while still fitting in. | cs.MA | cs | Gabora, L. (2008, November). Modeling cultural dynamics. In A. Davis & J. Ludwig (Co-Chairs), Adaptive agents in a
cultural context: Papers from the AAAI Fall Symposium (pp. 18-25). Symposium conducted at the meeting of Association
for the Advancement of Artificial Intelligence (AAAI), Palo Alto, CA. Retrieved from http://arxiv.org/abs/0811.2551
Modeling Cultural Dynamics
Liane Gabora
University of British Columbia
Okanagan campus, 3333 University Way
Kelowna BC, V1V 1V7, CANADA
[email protected]
Abstract
EVOC (for EVOlution of Culture) is a computer model of
culture that enables us to investigate how various factors such
as barriers to cultural diffusion, the presence and choice of
leaders, or changes in the ratio of innovation to imitation affect
the diversity and effectiveness of ideas. It consists of neural
network based agents that invent ideas for actions, and imitate
neighbors' actions. The model is based on a theory of culture
according to which what evolves through culture is not memes
or artifacts, but the internal models of the world that give rise
to them, and they evolve not through a Darwinian process of
competitive exclusion but a Lamarckian process involving
exchange of innovation protocols. EVOC shows an increase in
mean fitness of actions over time, and an increase and then
decrease in the diversity of actions. Diversity of actions is
positively correlated with population size and density, and with
barriers between populations. Slowly eroding borders increase
fitness without sacrificing diversity by fostering specialization
followed by sharing of fit actions. Introducing a leader that
broadcasts its actions throughout the population increases the
fitness of actions but reduces diversity of actions. Increasing
the number of leaders reduces this effect. Efforts are underway
to simulate the conditions under which an agent immigrating
from one culture to another contributes new ideas while still
'fitting in'.
Introduction
What impact do leaders and role models have on the views
and behaviors in a given culture? Is a dictatorial or a
distributed mode of leadership more effective? What is the
effect of complete or semi-permeable barriers to trade and
immigration, or barriers that erode or strengthen with time?
And what implications do these kinds of cultural patterns
have on how best to negotiate, conduct business, or simply
behave in a foreign land? These questions and others are
addressed using a computer model of culture referred to as
EVOC (for EVOlution of Culture). EVOC consists of
neural network based agents that invent ideas for actions,
and imitate neighbors' actions (Gabora, 2008). EVOC is an
elaboration of Meme and Variations, or MAV (Gabora,
1994, 1995), 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. Although MAV has inspired the incorporation of
cultural phenomena (such as imitation, knowledge-based
operators, and mental simulation) into evolutionary search
algorithms (e.g. Krasnogor & Gustafson, 2004), the goal
behind MAV was not to solve search problems, but to gain
insight into how ideas evolve. It used neural network based
agents that could (1) invent new ideas by modifying
previously learned ones, (2) evaluate ideas, (3) implement
ideas as actions, and (4) imitate ideas implemented by
neighbors. Agents evolved
in a cultural sense, by
generating and sharing ideas for actions, but not in a
biological sense; they neither died nor had offspring. The
approach can thus be contrasted with computer models of
the interaction between biological evolution and individual
learning (Best, 1999, 2006; Higgs, 2000; Hinton &
Nowlan, 1987; Hutchins & Hazelhurst, 1991).
MAV
successfully modeled how
'descent with
modification' can occur in a cultural context, but it had
limitations arising from the outdated methods used to
program it. Moreover, although new ideas in MAV were
generated making use of acquired knowledge and pattern
detection, the name 'Meme and Variations' implied
acceptance of the notion that cultural novelty is generated
randomly, and that culture evolves through a Darwinian
process operating on discrete units of culture, or 'memes'.
Problems with memetics and other Darwinian approaches
to culture have become increasingly apparent (Boone &
Smith, 1998; Fracchia & Lewontin, 1999; Gabora, 2004,
2006, 2008; Jeffreys, 2000). One problem is that natural
selection prohibits the passing on of acquired traits (thus
you don't inherit your mother's tattoo).1 In culture,
however, 'acquired' change -- that is, modification to ideas
between the time they are learned and the time they are
expressed -- is unavoidable. Darwinian approaches must
assume that elements of culture are expressed in the same
form as that in which they are acquired. Natural selection
also assumes that lineages do not intermix. However,
1 That isn't to say that inheritance of acquired traits never
occurs in biological evolution; it does. However to the
extent that this is the case natural selection cannot provide
an accurate model of biological evolution. Because
inheritance of acquired traits is the exception in biology
not the rule, natural selection still provides a roughly
accurate model of biological evolution.
because ideas cohabit a distributed memory with a
multitude of other ideas, they are constantly combining to
give new ideas, and their meanings, associations, and
implications are constantly revised.
It has been proposed what evolves through culture is not
discrete memes or artifacts, but the internal models of the
world that give rise to them (Gabora, 2004), and they
evolve not through a Darwinian process of competitive
exclusion but a Lamarckian process involving exchange of
innovation protocols
(Gabora, 206, 2008). EVOC
incorporates this in part by allowing agents to have
multiple interacting needs, thereby fostering complex
actions that fulfill multiple needs. Elsewhere (Gabora,
2008) the results of experiments using different needs
and/or multiple needs are described.
This paper describes other experiments carried out with
EVOC that were not possible to carry out with MAV.
These experiments investigate how cultural evolution is
affected by leadership, and by affordances of the agents'
world, such as world shape and size, population density,
and barriers that impede information flow, and potentially
erode with time.
Architecture
EVOC consists of an artificial society of agents in a two-
dimensional grid-cell world. It is written in Joone, an
object oriented programming environment, using an open
source neural network library written in Java. This section
describes the key components of the agents and the world
they inhabit.
The Agent
Agents consist of (1) a neural network, which encodes
ideas for actions and detects trends in what constitutes a fit
action, and (2) a body, which implements actions. In MAV
there was only one need -- to attract a mate. Thus actions
were limited to gestures that attract mates. In EVOC agents
can also engage in tool-making actions.
The Neural Network. The core of an agent is a neural
network, as shown in Figure 1. It is composed of six input
nodes that represent concepts of body parts (LEFT ARM,
RIGHT ARM, LEFT LEG, RIGHT LEG, HEAD, and
HIPS), six matching output nodes, and six hidden nodes
that represent more abstract concepts (LEFT, RIGHT,
ARM, LEG, SYMMETRY and MOVEMENT). Input
nodes and output nodes are connected to 'hidden' nodes of
which they are instances (e.g. RIGHT ARM is connected
to RIGHT.) Activation of any input node increases
activation of the MOVEMENT hidden node. Opposite-
direction activation of pairs of limb nodes (e.g. leftward
motion of one arm and rightward motion of the other)
activates the SYMMETRY node.
The neural network learns ideas for actions. An idea is a
pattern of activation across the output nodes consisting of
six elements that instruct the placement of the six body
parts. Training of the neural network is as per (Gabora,
1995). In brief, the neural network starts with small
random weights, and patterns that represent ideas for
actions are presented to the network. Each time a pattern is
presented, the network's actual output is compared to the
desired output. An error term is computed, which is used to
modify the pattern of connectivity in the network such that
its responses become more correct. Since the neural
network is an autoassociator, training continues until the
output is identical to the input. At this point training stops
and the run begins. The value of using a neural network is
simply that trends about what makes for a fit action can be
detected using the symmetry and movement nodes (see
below). The neural network can also be turned off to
compare results to those obtained using instead of a neural
network a simple data structure that cannot detect trends,
and thus invents ideas at random.
they adapt
Knowledge-based Operators. Brains detect regularity and
build schemas with which
the mental
equivalents of mutation and recombination to tailor actions
to 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. Since a new idea for an action is not learned
unless it is fitter than the currently implemented action,
newly learned actions provide valuable information about
what constitutes an effective idea. This information is used
by knowledge-based operators to probabilistically bias
invention such that new ideas are generated strategically as
opposed to randomly. Thus the idea is to translate
knowledge acquired during evaluation of an action into
educated guesses about what makes for a fit action.
Two rules of thumb are used. The first rule is: if
movement is generally beneficial, the probability increases
that new actions involve movement of more body parts.
Each body part starts out at a stationary rest position, and
with an equal probability of changing to movement in one
direction or the other. If the fitter action codes for more
movement, increase the probability of movement of each
body part. Do the opposite if the fitter action codes for less
movement. This rule of thumb is based on the assumption
FROM
NEIGH-
BOR
Head
Left Arm
Right
Arm
Left Leg
Right
Leg
Hips
INPUT
NODES
Symmetry
Move-
ment
Arm
Leg
Right
Left
HIDDEN
NODES
Head
Left Arm
Right
Arm
TO
Left Leg
BODY
Right
Leg
Hips
OUTPUT
NODES
Figure 1. The neural network. See text for details.
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 = movement node activation for current action
am2 = movement node activation for new action
p(im)i = probability of increased movement at body part i
p(dm)i = probability of decreased movement at body part i
realistic mating displays, and exhibits a cultural analog of
epistasis. In biological epistasis, the fitness conferred by
the allele at one gene depends on which allele is present at
another gene. In this cognitive context, epistasis is present
when the fitness contributed by movement of one limb
depends on what other limbs are doing.
The fitness of an action with respect to the second need,
the need to make tools, uses a second fitness function, F2,
and is calculated as in (Gabora, 2008).
IF (am2 > am1)
THEN p(im)i = MAX(1.0, p(im)i + 0.1)
ELSE IF (am2 < am1)
THEN p(im)i = MIN(0.0, p(im)i - 0.1)
p(dm)i = 1 - p(im)i
The second rule of thumb is: if fit actions tend to be
symmetrical (e.g. left arm moves to the right and right arm
moves to the left), the probability increases that new
actions are symmetrical. This generalization is biologically
sensible, since many useful actions (e.g. walking) entail
movement of limbs in opposite directions, while others
(e.g. pushing) entail movement of limbs in the same
direction. This rule is implemented in a manner 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 algorithm.
The Body. If the fitness of an action is evaluated to be
higher than that of any action learned thus far, it is copied
from the output nodes of the neural network that represent
concepts of body parts to a six digit array that contains
representions of actual body parts, referred to as the body.
Since it is useful to know how many agents are doing
essentially the same thing, when node activations are
translated into limb movement they are thresholded such
that there are only three possibilities for each limb:
stationary, left, or right. Six limbs with three possible
positions each gives a total of 729 possible actions. Only
the action that is currently implemented by an agent's body
can be observed and imitated by other agents.
The Fitness Functions
Agents evaluate the effectiveness of their actions according
to how well they satisfy needs using a pre-defined equation
referred to as a fitness function. Agents have two possible
needs. The fitness of an action with respect to the need to
attract mates is referred to as F1, and it is calculated as in
(Gabora, 1995). F1 rewards actions that make use of trends
detected by the symmetry and movement hidden nodes and
used by knowledge-based operators to bias the generation
of new ideas. F1 generates actions that are relatively
Incorporation of Cultural Phenomena
In addition
to knowledge-based operators, discussed
previously, agents incorporate the following phenomena
characteristic of cultural evolution as parameters that can
be turned off or on (in some cases to varying degrees):
•
Imitation. Ideas for how to perform actions spread
when agents copy neighbors' actions. This enables
them to share effective, or 'fit', actions.
Invention. This code enables agents to generate new
actions by modifying their initial action or a previously
invented or imitated action using knowledge-based
operators (discussed previously).
•
simulation.
Before
• Mental
to
implementing an idea as an action, agents can use the
fitness function to assess how fit the action would be if
it were implemented.
committing
The World
MAV allowed only worlds that were toroidal, or 'wrap-
around'. Moreover, the world was always maximally
densely populated, with one agent per cell. In EVOC the
world can be either toroidal or square, and as sparsely or
densely populated as desired, with agents placed in any
configuration. EVOC also allows the creation of complete
or semi-permeable permanent or eroding borders that
decrease the probability of imitation along a frontier.
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 the knowledge-based operators, and (3) implement
a new action. To invent a new idea, the current action is
copied to the input layer of the neural network, and this
previous action is used as a basis from which to generate a
new one. For each node the agent makes a probabilistic
decision as to whether change will take place. If it does,
the direction of change is stochastically biased by the
knowledge-based operators using the activations of the
SYMMETRY and MOVEMENT nodes. Mental simulation
is used to determine whether the new idea has a higher
fitness than the current action. If so, the agent learns and
implements the action specified by the new idea.
To acquire an idea through imitation, an agent randomly
chooses one of its neighbors, and evaluates the fitness of
the action the neighbor is implementing using mental
simulation. If its own action is fitter than that of the
neighbor, it chooses another neighbor, until it has either
observed all of its immediate 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 layer, learned, and implemented.
Fitness of actions starts out low because initially all
agents are immobile. Soon some agent invents an action
that has a higher fitness than doing nothing, and this action
gets imitated, so fitness increases. Fitness 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.
The Graphical User Interface
The graphical user interface (GUI) makes use of the open-
source charting project, JFreeChart, enabling variables to
be user defined at run time, and results to become visible
as the computer program runs. Figure 2 shows the topmost
output panel using the mating fitness function (F1). At the
upper left one specifies the Invention to Imitation Ratio.
This refers to the probability that a given agent, on a given
iteration, invents a new idea for an action, versus the
probability that it imitates a neighbor's action. Below that
is Rate of Conceptual Change, where one specifies the
degree to which a newly invented idea differs from the one
it was based on. Below that is Number of Agents, which
allows the user to specify the size of the artificial society.
Below that is where one specifies Number of Iterations, i.e.
the duration of a run. The agents that make up the artificial
society can be accessed individually by clicking the
appropriate cell in the grid on the upper right. This enables
one to see such details as the action currently implemented
by a particular agent, or the fitness of that action. The
graphs at the bottom plot the mean idea fitness and
diversity of ideas. Tabs shown at the top give access to
other output panels of the GUI.
Summary of Previous Results
EVOC closely replicates
the results of experiments
conducted with MAV (Gabora, 1995). The graph on the
Figure 2. Output panel of GUI using F1. See text for details.
bottom left of Figure 2 shows the increase in fitness of
actions. The graph on the bottom right of Figure 2 shows
the increase and then decrease in the diversity of actions.
Other MAV results that are replicated with EVOC include:
• Fitness increases most quickly with an invention to
imitation ratio of approximately 2:1.
•
• For the agent with the fittest actions, however, the less
it imitates, the better it does.
Increasing the invention-to-imitation ratio increases the
diversity of actions. If increased much beyond 2:1, it
takes more than twice as many iterations for all agents
to settle on optimal actions.
• As in biology, epistatically linked elements take longer
to optimize. (As explained earlier, in the present
context epistasis refers to the situation where the effect
on fitness of what one limb is doing depends on what
another is doing.)
• The program exhibits drift -- the term biologists use to
refer to changes in the relative frequencies of alleles
(forms of a gene) as a statistical byproduct of randomly
sampling from a finite population (Wright, 1969). With
respect to culture, the term pertains not to alleles but to
possible forms of a component of an idea (e.g. if the
idea is to implement the gesture 'wave', one can do this
with one's left hand or one's right).
These results show that concepts from biology are useful in
the analysis of cultural change, but that culture also
exhibits phenomena that have no biological equivalent.
Previous work on EVOC focused on the effects of
changing the need, and integrating multiple needs (Gabora,
2008). It was found that changing the need (modeled as a
change in the fitness function) or giving agents multiple
needs to fulfill, does not change the overall pattern of
results. Mean fitness of actions still increases gradually,
and diversity of actions rises and then falls, exhibiting the
typical inverted U-shaped curve, the magnitude of which is
a function of population size. However, increasing the
number of needs consistently results in a higher diversity
of actions, and tends to decrease mean fitness with respect
to any given need.
Experiments
We now outline the results of current experiments with
EVOC. Unless stated otherwise, the world is toriodal and
consists of 100 cells, with maximum density (one agent per
cell), no broadcasting, no barriers to idea flow, a 1:1
invention to imitation ratio, and a 0.17% probability of
change to any body part during invention (since, with six
body parts, on average each newly invented action differs
from the one it was based on with respect to one body
part).
Complete and Semi-permeable Barriers
Throughout history, the flow of ideas has been impeded
geographical barriers and political/cultural borders. It is
possible to simulate this in EVOC by reducing the
probability of imitation between agents on opposite sides
of a barrier. Barriers were found to increase latency to
converge on fit actions, and to increase diversity, by
effectively dividing the population. Interesting results are
achieved when barriers erode over time such that the
probability of imitation by agents on opposite sides is
initially zero but increases over the duration of a run,
simulating globalization. Figure 3 shows the diversity of
actions implemented after 4 iterations with an eroding
barrier. Eroding barriers foster specialization -- honing in
on unique solutions -- on different sides of the border,
followed by sharing of the best to reach a diverse final set.
Figure 3. Diversity of actions after four iterations with
8x8 grid and eroding barrier between 3rd and 4th
columns. Different actions represented by different
colored cells (which will appear in print as different
shades of grey). Invention to imitation ratio of agents to
right of border twice is that of agents to the left.
Diversity similarly affected by whether the shape of the
world is square -- which simulates a situation where the
flow of ideas is bounded by natural or political boundaries
-- versus toroidal -- which simulates the situation where
such barriers are overcome through globalization. A global
(toroidal) world accelerates fitness and increases diversity
in the short term but decreases it in the long term. This
makes sense; agents at the edges of a square world have
fewer neighbors, and thus more opportunity to retain
deviant actions.
Effect of Population Density
EVOC allows not just the shape of the world to be
changed, but how densely populated it is. Figure 4
illustrates the diversity of actions over a run with different
population densities.
The lower the population density, the more the typical
inverted-U shaped action diversity curve is disrupted; both
the
less
initial peak and subsequent decline are
2/3 of Cells Occupied
4/5 of Cells Occupied
All Cells Occupied
1/3 of Cells
Occupied
Figure 4. Effect of varying population density on
diversity of actions.
pronounced. Further analysis reveals
that decreasing
population density fosters the existence of small isolated
clusters that are unable to learn from one another and share
effective actions, impairing the ability of the society to
converge on only the fittest actions.
Broadcasting
Broadcasting allows the action of a leader, or broadcaster,
to be visible to not just immediate neighbors, but all
agents, thereby simulating the effects of media such as
public performances, television, radio, or internet, on
patterns of cultural change. When broadcasting is turned
on, an agent is no longer limited to its immediate neighbors
as potential role models. Each agent adds the broadcaster
as a possible source of actions it can imitate. A particular
agent can be chosen as the broadcaster before the run, or
the broadcaster can be chosen at random, or the user can
specify that the agent with the fittest action is the
broadcaster. Broadcasting can be intermittent, or continued
throughout the duration of a run. Broadcasting does not
have a significant effect on the fitness of actions, but as
shown in Figure 5, it accelerates convergence on optimal
actions, and consistently reduces diversity.
By varying the number of broadcasters EVOC allows
simulation of the effect on fitness and diversity of ideas of
a dictorial style of leadership (one broadcaster) versus a
distributed style of leadership (multiple leaders). In figure
6 we see how adding the presence of a broadcaster
(comparing column 1 without broadcaster to column 2 with
broadcaster) decreases the diversity of actions. This i seen
clearly looking to the lowest row: 20 iterations. Whereas
without a broadcaster there are eight different actions, and
41% of agents are executing the most popular action, with
a broadcaster there are five different actions, and 84% of
agents are executing the most popular action. However, as
shown in columns 3 to 6, these trends become increasingly
reversed the greater the number of broadcasters. With five
broadcasters, the society converges on nine different
actions, and the percentage of agents executing the most
popular action is down to 31%. These data potentially
speak to the changing effect of media on society. With
only one television or radio station, the effect of media
may have been to make opinions and behaviors more
homogeneous. However with the proliferation of different
radio and television stations, as well as web-based media,
available from not just local sources but around the world,
the effect might well be the reverse: an explosion of
different views and behaviors.
Discussion
the
effects
investigating
This paper has given an overview of factors impacting the
spread of ideas and behaviors that can be investigated with
a computer model of cultural evolution, focusing on new
results
broadcasting
of
(leadership), population density, and
the shape and
penetrability (e.g. presence of boundaries) of the terrain.
Results suggest that properties of the world can have as
great an impact on the evolution of culture as properties of
the agents themselves. The results also show that the
benefits of leadership with respect to enhanced fitness of
ideas may be tempered by decreased diversity of ideas.
This echoes previous simulation findings that leadership
can have adverse effects when agents can communicate
(Gigliotta, Miglino, & Parisi, 2007).
• Building Blocks. Agents will implement actions that
that
function
evaluates
A primary aim of future work will be to examine the
distinctively human phenomenon of cultural open-
endedness. Although presently agents' actions become
more complex and adapted over time, and change is
cumulative in that new actions build on existing ones, once
agents settle on some subset of optimal actions, the
program comes to a standstill. Future versions will use a
fitness
actions differently
depending on the relative strengths of the different needs.
The strength of a need will be a function of both how many
iterations have passed since execution of an action that
satisfied that need, and the degree to which that action
satisfied that need. It is expected that the program will not
come to a standstill because once an agent has filled one
need it will change the kind of action it implements to
satisfy another. Moreover to avoid that agents still zero in
on predictable subsets of actions that fulfill these needs,
future versions of EVOC will incorporate the following:
• Context-sensitive concepts. We plan to move to a more
subsymbolic level, incorporating how constellations of
activated microfeatures are influenced by context
(Aerts & Gabora, 2005a,b; Gabora, Rosch, & Aerts,
2008). This will allow for a richer repertoire of actions.
• Chained Actions. Agents will be allowed to chain
actions into arbitrarily long action sequences.
Figure 6. Diversity of actions over a run with 0, 1, 2, 3, 4, and 5 broadcasters. Different actions are represented by
differently colored cells (which will appear in the printed version as different shades of grey). In all cases there is an
increase followed by a decrease in diversity over time (moving down any column from the first iteration, at the top, to the
20th iteration, at the bottom). However, the decrease is less pronounced the more broadcasters there are. These runs used a
toroidal, maximally dense world, with a 10x10 grid. In this run, broadcasters were chosen at random every iteration, and
when their were multiple broadcasters, agents selected the broadcaster whose action was most similar to their own.
cumulatively modify their world using building blocks
to create structures that satisfy needs, and add to (or
destroy) structures made by others.
With these modifications it is expected that there will no
longer be an a priori limit to the number or complexity of
actions. The role of each of these modifications in bringing
about genuine cultural evolution will be assessed. The
effort will be judged successful if cultural change is not
just cumulative, but cumulative in a way that responds to
needs and situations, and open-ended, such that one
innovation creates niches for the invention of others (as
cars paved the way for seat belts and gas stations).
Further experiments with eroding barriers has potential
implications for the impact of free trade on global diversity
the conditions under which
of ideas, and for investigating the complex relationship
between creativity and culture (Kaufman & Sternberg,
2006). Future efforts will also focus on a more in-depth
analysis of
immigrant
contributes to the fitness and diversity of ideas versus the
conditions under which the immigrant's actions are so
different that they merely stand out and do not contribute
in a productive way. A wider range of needs will be made
available in order to determine the relationship between
degree of similarity between needs of the native and
immigrant populations and latency of the immigrant
population to 'fit in'. Other questions will also be
investigated, such as 'How does the probability of 'fitting
in' change as a function of the number of immigrants?
These issues are timely, and have potential implications for
how individuals should go about negotiating, conducting
business, and simply behaving in a foreign land. Even
when such simulations do not provide specific directives,
they help us to think in more precise terms about the
issues.
Acknowledgments
Thanks to Martin Denton and Jillian Dicker for their work
on EVOC. This project is funded by the Social Sciences
and Humanities Research Council of Canada (SSHRC).
References
Aerts, D. & Gabora, L. 2005a. A state-context-property
model of concepts and their combinations I: The
structure of the sets of contexts and properties.
Kybernetes 34(1&2): 151 -- 175.
Aerts, D. & Gabora, L. 2005b. A state-context-property
model of concepts and their combinations II: A Hilbert
space representation. Kybernetes 34(1&2): 176 -- 205.
Best, M. 1999. How culture can guide evolution: An
inquiry into gene/meme enhancement and opposition.
Adaptive Behavior 7(3): 289 -- 293.
Best, M. 2006. Adaptive value within natural language
discourse. Interaction Studies 7(1), 1-15.
Boone, J. L., Smith, E. A. 1998. Is it evolution yet? A
archaeology. Current
evolutionary
critique of
Anthropology, 39, S141 -- 73.
Fracchia, J. & Lewontin, R. C. 1999. Does culture evolve?
History and Theory 38(4): 52 -- 78.
Gabora, L. 1994. A computer model of the evolution of
culture. In R. Brooks & P. Maes (Eds.) Proceedings of
the 4th International Conference on Artificial Life, July
4-6, Boston MA.
Gabora, L. 1995. Meme and Variations: A computer model
of cultural evolution. In L. Nadel & D. Stein (Eds.)
1993 Lectures in Complex Systems, pp. 471−486.
Reading MA: Addison-Wesley.
Gabora, L. 2004. Ideas are not replicators but minds are.
Biology & Philosophy 19(1): 127−143.
Gabora, L. 2006. The fate of evolutionary archaeology:
Survival or extinction? World Archaeology 38(4):
690−696.
Gabora, L. 2008. The cultural evolution of socially situated
cognition. Cognitive Systems Research 9(1 -- 2): 104 --
113.
Gabora, L. 2008. EVOC: A computer model of cultural
evolution. In V. Sloutsky, B. Love & K. McRae
(Eds.), 30th Annual Meeting of the Cognitive Science
Society. Washington DC, July 23-26, North Salt
Lake, UT: Sheridan Publishing.
Gabora, L., Rosch, E., & Aerts, D. 2008. Toward an
ecological theory of concepts. Ecoogical Psychology
20(1): 84−116.
Gigliotta, O. Miglino, O. & Parisi, D. 2007. Groups of
agents with a leader. Journal of Artificial Societies and
Social
10(4).
http://jasss.soc.surrey.ac.uk/10/4/1.html
Simulation,
Gilbert, D. & Morgner, T. JfreeChart. http://www.jfree.
org/jfreechart/index.html
Higgs, P. G. 2000. The mimetic transition: a simulation
study of the evolution of learning by imitation.
Proceedings of
the Royal Society B: Biological
Sciences 267(1450): 1355 -- 1361.
Hinton, G. E. 1981. Implementing semantic networks in
parallel hardware. In Parallel Models of Associative
Memory. Hillsdale NJ: Erlbaum Press.
Hinton, G. E. & Nowlan, S. J. 1987. How learning can
guide evolution. Complex Systems 1: 495 -- 502.
Holland, J. K. 1975. Adaptation in Natural and Artificial
Systems. Ann Arbor: University of Michigan Press.
Hutchins, E. & Hazelhurst, B. (1991). Learning in the
cultural crocess. In Langton, C., Taylor, J., Farmer, D.,
& Rasmussen, S. (Eds.) Artificial Life II. Redwood
City, CA: Addison-Wesley.
Jeffreys, M. (2000). The meme metaphor. Perspectives in
Biology and Medicine 43(2): 227 -- 242.
Karmiloff-Smith, A. 1992. Beyond modularity: A
developmental perspective on cognitive science.
Cambridge MA: MIT Press.
Kaufman, J. C., & Sternberg, R. J. (Eds). 2006. The
international handbook of creativity. Cambridge:
Cambridge University Press.
Krasnogor, N. & Gustafson, S. 2004. A study on the use of
in memetic algorithms. Natural
"self-generation"
Computing 3(1): 53 -- 76.
Marrone, P. Java Object Oriented Neural Engine (JOONE).
www.jooneworld.com
Wright, S. 1969. Evolution and
the genetics of
Populations. Chicago: University of Chicago Press.
|
1210.5936 | 1 | 1210 | 2012-10-22T15:37:36 | Mod\'elisation multi-niveaux dans AA4MM | [
"cs.MA"
] | In this article, we propose to represent a multi-level phenomenon as a set of interacting models. This perspective makes the levels of representation and their relationships explicit. To deal with coherence, causality and coordination issues between models, we rely on AA4MM, a metamodel dedicated to such a representation. We illustrate our proposal and we show the interest of our approach on a flocking phenomenon. | cs.MA | cs | Modélisation multi-niveaux dans AA4MM
B. Camusa,b
J. Siebertc
C. Bourjota,b
V. Chevriera,b
[email protected] [email protected] [email protected]
[email protected]
a Université de Lorraine, LORIA, UMR 7506 , Vandoeuvre les Nancy, F-54506 France
b Inria, Villers-Lès-Nancy, F-54600 France
c Lab-STICC, Ecole Nationale d'Ingénieurs de Brest, France
Résumé
Dans cet article, nous proposons de représen-
ter un phénomène multi-niveaux sous la forme
de plusieurs modèles en interaction. Cette vi-
sion structure la réflexion en rendant expli-
cites les niveaux de représentation et leurs
relations. Pour répondre aux défis de cohé-
rence, causalité et coordination entre ces mo-
dèles, nous nous appuyons sur le méta-modèle
AA4MM dédié à ce type de représentation. Un
phénomène de flocking permet d'illustrer notre
propos et de monter l'intérêt de cette ap-
proche.
Cette démarche pose des bases conceptuelles
relativement simples pour aborder une ques-
tion encore largement ouverte de la modélisa-
tion des systèmes complexes
Mots-clés : modélisation et simulation multi-
niveaux ; émergence ; méta-modèle
Abstract
In this article, we propose to represent a mul-
ti-level phenomenon as a set of interacting
models. This perspective makes the levels of
representation and their relationships explicit.
To deal with coherence, causality and coordi-
nation issues between models, we rely on
AA4MM, a metamodel dedicated to such a
representation. We illustrate our proposal and
we show the interest of our approach on a
flocking phenomenon.
Keywords: Modeling and simulation multi-
level representation, emergence, metamodel.
1. Introduction
Lorsque l'on aborde la modélisation de phé-
nomènes collectifs, que ce soit en sciences
humaines et sociales ou en biologie, il est sou-
vent nécessaire de représenter le système selon
plusieurs niveaux (Troitzsch, 1996).
Par exemple en biologie (Sloot, 2010) (Uhr-
macher, 2009) (Ballet, 2012), le système à
modéliser fait "naturellement" intervenir plu-
sieurs niveaux ; tels celui des molécules,
membranes, cellules, groupes de cellules, etc. ;
qu'il est nécessaire de représenter et ainsi pou-
voir étudier les liens entre ces niveaux.
Les systèmes multi-agents font partie des ap-
proches intégrant à la fois les aspects indivi-
duels et collectifs d'un système. Ils ont ainsi
permis la modélisation de phénomènes où des
individus en interactions locales font émerger
un comportement collectif et permettent de
répondre à des questions quant aux relations
entre comportements individuels et propriétés
collectives.
De manière plus générale, cette question s'ins-
crit dans l'étude des systèmes complexes. Ain-
si dans (Chavalarias, 2008), les auteurs écri-
vent : "La complexité des systèmes naturels et
sociaux provient de l’existence de plusieurs
niveaux d’organisation correspondant à diffé-
rentes échelles spatio-temporelles. L’un des
principaux défis de la science des systèmes
complexes consiste à développer … des mé-
thodes de modélisation capables de saisir
toutes
les dynamiques d’un système par
l’intégration de ses activités à de nombreux
niveaux, souvent organisés hiérarchiquement."
Cet article tente d'apporter une première ré-
ponse à ce défi.
Après avoir situé la problématique de la modé-
lisation multi-niveaux, nous proposons de re-
présenter un phénomène multi-niveaux sous la
forme de plusieurs modèles (ici on se limitera
aux niveaux micro et macro) en interaction
(partie 2). Cette vision structure la réflexion en
rendant explicites les niveaux de représenta-
tion et leurs relations. Après avoir décrit un
exemple de phénomène multi-niveaux simple,
le flocking (partie 3), nous nous appuyons
(partie 4) sur les concepts du méta-modèle
AA4MM (Agent et Artefact pour la Multi-
Modélisation) dédié à ce type de représenta-
tion pour réaliser une preuve de concepts au-
tour de l'exemple choisi (partie 5).
2. La modélisation multi-niveaux
L’étude des phénomènes collectifs nécessite
donc au moins deux niveaux de représenta-
tions (individuel et collectif). Les SMA offrent
un paradigme intéressant car on a deux ni-
veaux de discours qui sont le niveau micro
(l'agent) et le niveau macro (le système). Ce-
pendant seul le niveau micro y est traditionnel-
lement explicitement représenté, le niveau
macro se contente le plus souvent d'être obser-
vé mais n'est pas défini en tant que tel. Par
conséquent, ce niveau n'existe que par la pré-
sence d'un observateur extérieur au système et
ne prend pas place dans un processus de modé-
lisation. C'est ce que (Quijano, 2010) décrivent
comme une approche "mono-niveau dans la
conception et bi-niveaux dans l’analyse des
comportements produits".
Il existe toutefois un certain nombre de tra-
vaux qui vont au delà de cette approche.
(David, 2009, David 2010, David 2011) pro-
posent une démarche pour réifier les propriétés
émergentes. Une première étape (processus
d'introspection) consiste en la détection des
propriétés émergentes. Cette détection s'appuie
sur des connaissances du phénomène étudié.
Ensuite ces propriétés sont réifiées (si besoin)
sous la forme d'un agent émergent qui sera
doté de propriétés et comportements. Enfin, il
est possible de représenter l'influence de la
propriété émergente sur le système sous la
forme d'un élément d'interposition qui pourra
modifier la perception et/ou l'influence des
agents au niveau micro.
(Quijano, 2010) proposent une première ana-
lyse des manières dont peuvent être couplées
des organisations multi-agents multi-niveaux à
partir de différentes expériences de modélisa-
tion. Ils distinguent trois catégories de sys-
tèmes selon le type de couplage qui est réalisé
entre les modèles associés à ces niveaux. Ils
retiennent qu'il faut pouvoir intégrer des mo-
dèles existants aux formalismes potentielle-
ment hétérogènes, ainsi que de pouvoir détec-
ter, réifier et détruire dynamiquement les
structures émergentes.
Traounez (Traounez, 2005) analyse le système
au niveau micro pour détecter puis réifier des
propriétés émergentes (en l'occurrence des
tourbillons). De plus, il utilise une structure de
donnée récursive pour représenter leur envi-
ronnement à des échelles spatiales différentes.
Cette représentation peut se rapprocher de
celle de (Marilleau, 2008). Néanmoins, ces
représentations correspondent à un change-
ment d'échelle (plus ou moins précise) mais
pas un changement de niveau de description.
Ces travaux considèrent tous que pour étudier
un phénomène complexe à différents niveaux
de description, il est nécessaire de représenter
explicitement chacun des niveaux, de détecter
puis réifier (Bonabeau, 1997) (si besoin) les
propriétés collectives, et enfin, il doit être pos-
sible de contraindre le niveau micro à partir de
considérations du niveau macro.
(David, 2010) (Traounez, 2005) proposent
différents concepts mais
ils sont pensés
comme intégrés dans un tout. Ils n'abordent
les différents couplages
qu'insuffisamment
possibles entre les niveaux (tels qu'évoqués
dans (Quijano, 2010)). Notamment, les ques-
tions de comment représenter un phénomène à
partir de modèles différents, de comment for-
maliser les couplages entre modèles qui cor-
respondent aux relations entre niveaux de des-
criptions sont encore ouvertes.
(Bourgine, 2008) propose un cadre relative-
ment général qui apporte un début de réponse
à ces questions. Cette vision fait intervenir
explicitement les modèles utilisés pour chaque
niveau, ainsi que les relations entre ces ni-
veaux. Le principe général est résumé par le
schéma en figure 1.
L'idée sous-jacente est la suivante : on peut
décrire un phénomène à un niveau macro si
l’on est capable de caractériser ce phénomène
par un ensemble d'informations (noté Y) et que
l'on peut décrire l'évolution de Y temporelle-
ment sans avoir à faire référence au niveau
micro. Cette vision exprime les différents élé-
ments qui interviennent lors de la modélisation
multi-niveaux et la manière dont ils s'articu-
lent :
• X (resp. Y) correspond à l'ensemble
des informations représentant le phénomène au
niveau micro (resp. macro).
f (resp g) est une fonction qui décrit la
•
dynamique du phénomène au niveau micro
(resp. macro).
• e est une fonction d'interprétation
(émergence) permettant le passage du niveau
de description micro à celui macro. Elle cor-
respond aux méta-connaissances de (David,
2010).
i correspond à l'immergence, c'est à
•
dire, à l'influence du niveau macro sur le ni-
veau micro (ce que (David, 2010) appelle
fonction d'interposition). Cette influence modi-
fie le comportement au niveau micro.
Si cette vision permet de situer et mettre en
relation chacun des éléments, elle n'indique
rien quant à leurs expressions (Quels forma-
lismes pour exprimer e, i f ou g? Quelles cor-
respondances entre t et t' ? Etc.).
Figure 1 : Relations (d'après (Bourgine, 2008))
entre niveaux de description et évolution tempo-
relle du phénomène.
Nous proposons de faire un pas vers une clari-
fication de ce principe en envisageant chaque
niveau comme un modèle et les relations entre
eux comme des interactions.
La partie suivante présente un exemple jouet
de modélisation multi-niveaux faisant interve-
nir les différents éléments évoqués par (Bour-
gine, 2008). Nous décrivons ensuite le méta-
modèle AA4MM qui nous permet de décrire
cette représentation comme une société de
modèles interagissant ; puis de l'implanter et la
simuler.
3. Exemple de modélisation multi-
niveaux
Cet exemple volontairement simple servira à
illustrer notre proposition de modélisation
multi-niveaux et à montrer comment s'instan-
cient les différents concepts évoqués.
Les deux niveaux de modélisation du
3.1.
flocking
Nous prenons comme exemple les nuées d'oi-
seaux. Nous considérons deux niveaux de des-
cription : celui des oiseaux (microscopique) et
celui des nuées (macroscopique).
La modélisation au niveau microscopique re-
prend le modèle multi-agent développé dans
(Wilensky, 1998). L'état d'un oiseau est défini
par un identifiant, une position et une orienta-
tion. Son comportement se déduit de ses com-
portements de cohésion, alignement et sépara-
tion vis à vis des autres oiseaux. La figure 2
visualise un état du niveau micro.
Après avoir détecté les nuées, celles-ci seront
réifiées dans le modèle macro par un agent
situé au centre de gravité de la nuée et orienté
selon la direction moyenne des oiseaux qui la
composent, la taille est fonction de la disper-
sion. Ainsi, au fur et à mesure que les nuées
apparaissent ou disparaissent, on met à jour le
modèle au niveau macroscopique.
Figure 2: vue du niveau micro.
Au niveau macroscopique, une nuée est modé-
lisée par un agent caractérisé par une position,
une orientation et une taille. Nous supposons
de manière arbitraire que le comportement
d'une nuée se déduit de ses comportements de
cohésion, alignement et séparation vis à vis
des autres nuées.
Ce modèle macroscopique est une adaptation
du modèle de (Wilensky, 1998) qui prend no-
tamment en compte la taille de l'agent. Par
simplification, il ne gère pas les fusions ni les
séparations des nuées. La figure 3 visualise un
état du niveau macro.
Les influences mutuelles
3.2.
L’influence du niveau micro sur le niveau ma-
cro est fonction de l’interprétation que nous
ferons d'une nuée (Cavagna, 2008) : si plu-
sieurs oiseaux sont suffisamment rapprochés et
possèdent des orientations proches, ils consti-
tueront une nuée évoluant dans un déplace-
ment commun.
Figure 3: vue du niveau macro.
Nous supposons que l’influence du niveau
macro sur le niveau micro est la suivante : tous
les oiseaux appartenant à une nuée se dépla-
cent de la même manière et ont la même orien-
tation. Le modèle macroscopique calcule le
déplacement de chaque nuée (vecteur de tran-
slation v ). Nous utilisons ce vecteur de dépla-
cement global pour déplacer, au niveau micro
chacun des oiseaux appartenant à la nuée. On
remarquera qu'il faut donc étendre le modèle
de (Wilensky, 1998) pour prendre en compte
ce deuxième mode de déplacement au niveau
microscopique.
Enfin, nous supposons également dans un
premier temps que les dynamiques évoluent à
la même vitesse.
La correspondance entre ce multi-modèle et le
schéma inspiré de (Bourgine, 2008) s'effectue
comme suit :
•
•
• X correspond aux caractéristiques des
oiseaux,
• Y correspond aux caractéristiques des
nuées,
f est le modèle de déplacement des oi-
seaux,
• g est le modèle de déplacement des
nuées,
• e est le mécanisme de détection des
nuées,
i est l'influence des nuées sur le com-
portement d’un oiseau.
4. Instanciation avec AA4MM
Nous proposons de représenter ce modèle mul-
ti-niveaux sous la forme de modèles (ici on se
limitera aux niveaux micro et macro) en inte-
raction (émergence et immergence). Pour cela,
nous nous appuyons sur AA4MM, un méta-
modèle dédié à ce type de représentation.
Le méta-modèle AA4MM
4.1.
AA4MM (Siebert, 2011) modélise un phéno-
mène complexe comme un ensemble de mo-
dèles en interaction (ce que nous appellerons
par la suite un multi-modèle). Il s’inscrit dans
le paradigme multi-agent : à tout modèle est
associé un agent et les interactions sont sup-
portées par des artefacts (Ricci, 2007).
On retrouve ce type d'approche également
dans Reiscop (Desmeulles, 2009), Ioda (Kube-
ra, 2011) ou encore Geamas (Marcenac, 1998).
Les dynamiques du phénomène (micro/macro)
sont représentées par des modèles différents.
La dynamique globale du phénomène est si-
mulée grâce à
l'interaction des modèles.
L’originalité vis à vis d'autres approches de
multi-modélisation est d'envisager les interac-
tions de manière indirecte, supportées par un
environnement, et ainsi de modéliser explici-
tement le partage d'informations entre modèles
(couplage structurel).
L'intérêt d'un méta-modèle dans notre cas est
de disposer d'un cadre conceptuel dans lequel
nous pourrons décrire les niveaux de représen-
tation et la manière dont ils interagissent, c'est
à dire modéliser un système sous la forme de
dynamiques à différents niveaux ainsi que de
leurs relations (influences).
D'un point de vue méthodologique, cette vi-
sion oblige à expliciter les choix de modélisa-
tion relatifs à chacun des modèles et à leurs
interactions dès la phase de conception du
multi-modèle et nous contraint sur la manière
de les exprimer. La contrepartie est que
AA4MM possède des spécifications opéra-
tionnelles et des algorithmes prouvés concer-
nant la cohérence temporelle entre modèles qui
permettent d'implanter le multi-modèle et de le
simuler en ne devant coder qu'un nombre res-
treint de fonctions spécifiques à l’application
visée. En effet, les spécifications opération-
nelles permettent de disposer d’un intergiciel.
Cette approche a donné lieu à des preuves de
concepts avec Netlogo (Siebert, 2010) et une
application dans le cadre des réseaux mobiles
ad-hoc (Leclerc, 2010) notamment en réutili-
sant des modèles existants et hétérogènes.
Représentation multi-niveaux avec
4.2.
AA4MM
Nous allons maintenant expliciter la représen-
tation multi-niveaux à l’aide des composants
de AA4MM. Il repose sur trois concepts à par-
tir desquels il est possible de décrire un multi-
modèle :
1. le m-agent contrôle un modèle et prend
en charge les aspects dynamiques des
interactions de ce modèle avec les
autres modèles (figure 4a),
2. chacune de ces interactions (entre m-
agents) est réifiée par un artéfact de
couplage (figure 4b),
3. enfin l'artéfact d'interface réifie les in-
teractions entre un m-agent et son mo-
dèle (figure 4c).
a b c
Figure 4 : symboles des composants du méta-
modèle AA4MM (a) m-agent, (b) artéfact de cou-
plage, (c) artéfact d'interface et son modèle m.
Le comportement d’un m-agent correspond au
cycle :
•
lecture des informations en provenance
des artefacts de couplage,
• mise-à-jour et exécution de son modèle
via l'artéfact d'interface,
• écriture d’informations à destination
des artefacts de couplage.
Un artéfact de couplage propose des opéra-
tions d'écriture (un m-agent écrit les informa-
tions à transmettre), de transformations (chan-
gement d'unités par exemple), et de lecture (un
m-agent lit des informations).
Un artéfact d'interface autorise les opérations
suivantes sur le modèle : initialisation des
données, mise à jour des caractéristiques indi-
viduelles, exécution d'un pas de simulation.
Démarche de modélisation d'un phé-
4.3.
nomène multi-niveaux
La création d'un multi-modèle via AA4MM se
déroule en plusieurs étapes que nous déclinons
ci-après dans le cas de la modélisation multi-
niveaux.
La première étape consiste à définir le graphe
d'interaction entre les modèles, c'est à dire,
quelles informations sont échangées.
Dans notre cas, ce graphe se déduit assez sim-
plement. Chaque niveau de représentation est
associé à un modèle (noté m pour le niveau
micro et M pour le niveau macro). Les interac-
tions sont les relations d'émergence (notée e)
de micro vers macro et d'immergence (notée i)
de macro vers micro (voir figure 5).
Figure 5 : Graphe des relations entre modèles.
La seconde étape consiste à exprimer ce
graphe à l’aide des concepts de AA4MM. Ain-
si à chaque modèle, et donc à chaque niveau,
correspond un m-agent et un artefact
d’interface. Le m-agent Am contrôle (gère) le
modèle du niveau micro et le m-agent AM con-
trôle le modèle du niveau macro. A chaque
relation entre modèles correspond un artefact
de couplage entre leurs m-agents correspon-
dants (cf. figure 6).
La mise en oeuvre de ce diagramme nécessite
des extensions de AA4MM par rapport à ses
spécifications initiales. En détaillant notre
exemple, nous soulignons les manques et indi-
quons1 la manière dont nous y répondons ci-
après.
Concernant les artéfacts d'interface, nous réuti-
liserons ceux existant pour Netlogo dans la
bibliothèque AA4MM.
Dans une démarche bottom-up, initialement le
niveau micro comporte un certain nombre
d’oiseaux créés aléatoirement. Le m-agent Am
fait exécuter un pas de simulation, récupère les
identifiants, positions et orientations des oi-
seaux et les écrit dans l’artefact de couplage e.
L'artefact de couplage e reçoit donc du m-
agent Am la liste des positions et orientation de
chacun des oiseaux et les interprète pour cons-
truire des nuées. Ceci est réalisé par un algo-
rithme d’identification de cluster avec deux
paramètres : un seuil de proximité et un seuil
1 La place disponible dans cet article ne nous
permet pas de fournir l'intégralité des spécifi-
cations. Nous nous contentons d'indiquer les
modifications qualitativement.
d’orientation. Ces informations (liste de nuées
avec centre de gravité, orientation et taille)
sont alors disponibles pour le m-agent AM.
Figure 6 : Diagramme AA4MM de la représenta-
tion multi-niveaux.
Dans sa conception originelle, un artéfact de
couplage conserve le cardinal des informations
qui lui sont fournies lorsqu'il les transmet. La
relation d'émergence oblige à concevoir un
nouveau type d'artéfact (dit d'interprétation)
qui dans le cas de la relation d'émergence ré-
duit les informations (ici en taille). De plus, il
se peut qu'il n'y ait aucune nuée.
Le m-agent AM lit les informations interprétées
par l'artéfact de couplage e ; met à jour les
nuées répertoriées, ajoute les nouvelles et sup-
prime celles disparues dans le simulateur ; fait
exécuter un pas de simulation ; récupère les
nouvelles positions des nuées ; calcule les vec-
teurs de déplacement des nuées puis écrit ces
informations dans l'artéfact de couplage i.
On constate que le m-agent doit pouvoir ajou-
ter ou supprimer des nuées dans le modèle,
qu'il doit également traiter l'absence d'informa-
tions en provenance de l'artéfact d'émergence.
Nous avons, en conséquence, adapté le com-
portement du m-agent et ses spécifications à
ces besoins. L'artefact d’interface est étendu
pour supporter ajout et suppression de nuées
dans le modèle.
L'artéfact de couplage i reçoit une liste de
nuées accompagnées chacune de son vecteur
de déplacement, transforme ces informations
pour fournir une liste d'oiseaux accompagnées
de leur vecteur déplacement et les rend dispo-
nibles au m-agent Am.
Cet artéfact est également un artéfact d'inter-
prétation car il y a augmentation de la quantité
d'informations transmises.
Le comportement du m-agent Am consiste à lire
les données en provenance de l'artéfact de
couplage i, à faire exécuter un pas de simula-
tion en tenant compte des deux modes de dé-
placement, à récupérer les identifiant, position
et orientation de chacun des oiseaux, puis à
écrire ceux-ci dans l'artéfact de couplage e.
Comme il est possible qu'il n'y ait pas de nuée,
ce m-agent doit être étendu pour traiter l'ab-
sence d'information en provenance de l'artéfact
d'immergence.
5. Expérimentations
L'objectif des expérimentations est de montrer
la faisabilité de notre proposition et d'en illus-
trer l'intérêt et les possibilités.
Figure 7 : exécution du multi-modèle.
Nous avons donc implanté l'exemple du flock-
ing. La figure 7 correspond à une copie d'écran
de la simulation multi-niveaux. Il faut noter
qu’ici les modèles s’exécutent dans deux ins-
tances Netlogo (Wilensky, 1999) séparées.
Concernant l'intérêt de l'approche au niveau du
"multi-modélisateur", nous montrons son pou-
voir explicatif (cf 5.1), sa capacité à construire
différentes déclinaisons du multi-modèle par
changement de modèles (cf 5.2), de couplages
(cf 5.3) ou de dynamique temporelle (cf 5.4) et
à comparer ces déclinaisons (cf 5.5).
Les propriétés prouvées de AA4MM (cohé-
rence, causalité et coordination entre modèles
(Siebert, 2011)) facilitent le développement
des différentes déclinaisons.
Niveaux micro-macro explicites
5.1.
Figure 8 : croisement de plusieurs nuées (gauche
niveau micro, droite niveau macro).
L'existence de deux niveaux de description
distincts utilisant chacun un vocabulaire diffé-
rent (Muller, 2004) permet de rendre compte
de phénomènes, comme par exemple, expri-
mer explicitement le croisement entre plu-
sieurs nuées (invisible au niveau micro mais
visible au niveau macro, voir figure 8).
Influence du modèle macro sur le
5.2.
phénomène
Outre le modèle M décrit en section 3, nous
avons utilisé deux autres modèles (dérivés de
M) pour le niveau macro : M1 utilise un fort
facteur de séparation et de faibles facteurs de
cohésion et alignement ; M2 utilise de forts
facteurs de cohésion et alignement et un faible
facteur de séparation.
Phénomène sans immergence
5.3.
Nous souhaitons également étudier le compor-
tement du phénomène sans immergence. L'im-
plantation se déduit à partir de la figure 9 : les
composants inutiles sont supprimés et les m-
agents modifiés de la manière suivante : AM se
contente de lire les informations en prove-
nance de e, Am n'a plus besoin de lire les don-
nées en provenance de i (lequel a disparu).
Cette déclinaison sera notée m en figure 10.
Yt
Yt
e
e
xt
Xt+=
f
Figure 9 : couplage sans immergence, ni compor-
tement macro.
Couplage avec des échelles tempo-
5.4.
relles différentes
Nous reprenons notre exemple tel qu'en sec-
tion 3 mais en supposant maintenant des
échelles de temps différentes : il y a quatre pas
de simulation au niveau micro pour un au ni-
veau macro (noté ici M3).
Nous revisitons l'interprétation que nous fai-
sons de l'immergence : le déplacement macro
est décomposé linéairement en quatre dépla-
cements micro. L'artéfact de couplage i trans-
forme donc les données qu'il reçoit pour cons-
truire quatre listes associant à un identifiant un
quart de vecteur déplacement. Au niveau de la
cohérence temporelle, nous tirons parti de l'al-
gorithme d'exécution prouvé de AA4MM qui
gère les relations temporelles et ne modifions
que i et la correspondance temporelle au ni-
veau de AM3.
Comparaison des déclinaisons du
5.5.
multi-modèle
Dans notre exemple, l'évolution du nombre de
de nuées dépend de plusieurs facteurs : le
comportement individuel micro, et en cas
d'immergence du niveau macro au travers des
paramètres du comportement des nuées.
La figure 10 montre une comparaison de
l'influence de ces différents facteurs. Chaque
point d'une courbe correspond à 100 exécu-
tions.
Dans une démarche de modélisation, ce type
d'étude est important pour comprendre les re-
lations entre niveaux. Notre approche permet
de situer précisément quelle partie du multi-
modèle est impliquée.
Figure 10 : comparaison du nombre de nuées selon
l'influence de différents facteurs. Les notations m,
M, M1, M2 et M3 font référence aux modèles pré-
sentés précédemment.
6. Conclusion
Dans cet article, nous avons apporté une pre-
mière réponse au problème de la représenta-
tion multi-niveaux de phénomènes complexes.
Notre proposition, au niveau conceptuel, s'ins-
pire de (Bourgine, 2008) qui envisage la repré-
sentation multi-niveaux comme un couplage
entre plusieurs systèmes dynamiques, ce que
nous exprimons de façon modulaire sous la
forme de plusieurs modèles multi-agents en
interaction. Pour répondre aux défis de cohé-
rence, causalité et coordination entre ces mo-
dèles, nous nous appuyons sur le méta-modèle
AA4MM dédié à ce type de représentation.
AA4MM nous offre de par son approche mo-
dulaire une flexibilité que nous pouvons ex-
ploiter dans le cadre d’une modélisation multi-
niveaux. Cela permet facilement le change-
ment de modèles, le changement d’échelle
temporelle, le changement de couplage entre
niveaux ; ainsi que la réutilisation de modèles
existants.
La démarche sous-jacente propose de partir du
problème de modélisation multi-niveaux, pour
établir un graphe de
relations
(émer-
gence/immergence) entre les deux niveaux
micro/macro que nous traduisons sous la
forme d’un diagramme AA4MM. Ce dia-
gramme est alors
implanté à partir de
l’intergiciel de AA4MM.
Ce travail pose des bases conceptuelles relati-
vement simples pour aborder une question
encore largement ouverte de la modélisation
des systèmes complexes. Nous envisageons
d'étendre cette approche à des systèmes qui ne
sont plus simplement bi-niveaux (micro-
macro), et de la confronter à d'autres phéno-
mènes dont les relations d'émergence et d'im-
mergence prennent d'autres formes.
Remerciements
Les auteurs remercient l'un des relecteurs pour
ses remarques constructives.
Références
Ballet P. BioDyn, une approche interdisciplinaire
pour la création et le développement de logiciels
intégrés de et de simulation en biologie computation-
nelle. Mémoire d'HDR, Université de Bretagne Occi-
dentale, 2012.
Bonabeau, E., Dessalles, J-L. Detection and emer-
gence. Intellectica, 25 (2), 85-94, 1997.
Bourgine, P. Distributed Problem Solving in natural
and artificial complex systems. Invited talk in Engi-
neering Societies in the Agents World IX, 9th Inter-
national Workshop, ESAW 2008.
Cavagna A., Giardina I., Orlandi A., Parisi G., Pro-
caccini A. The STARFLAG handbook on collective
animal behaviour: Part II, three-dimensional analy-
sis. Animal Behaviour. Volume: 76, Issue: 1, Pages:
237-248, 2008.
Chavalarias D., Bourgine P., Perrier E., Amblard F.,
Arlabosse F., Auger P., Baillon J.-B., Barreteau O.,
Baudot P., Bouchaud E. French Roadmap for com-
plex Systems 2008-2009, French National Network
for Complex Systems (RNSC), Paris Ile-de France
Complex Systems Institute (ISC-PIF) and IXXI,
Entretiens de Cargèse 2008, 2008.
David D., Courdier R. See Emergence as a Me-
taknowledge, a Way to Reify Emergent Phenomena
in Multiagent Simulations?, International Conference
on Agents and Artificial Intelligence (ICAART’09),
pp. 564-569, Porto, Portugal, Jan. 19-21, 2009.
David D. Prospective Territoriale par Simulation
Orientée Agent, Thèse de Doctorat, Université de La
Réunion, 2010.
David D., Payet, R. Courdier. Réification de zones
urbaines émergentes dans un modèle simulant
l’évolution de la population à La Réunion, Journées
Francophones sur les Systèmes Multi-Agents (JFS-
MA’11), pp. 63-72, Valenciennes, France, 17-19
Octobre 2011.
Desmeulles G., Bonneaud S., Redou P., Rodin V.,
Tisseau J. In virtuo experiments based on the multi-
interaction system framework: the reiscop meta-
model. CMES, Computer Modeling in Engineering &
Sciences, 2009.
Gil-Quijano J., Hutzler G., Louail T. Accroche-toi au
niveau, je retire l'échelle, Eléments d'analyse sur la
modélisation multi-niveaux dans
les simulations
multi-agents. Revue d'Intelligence Artificielle, Vol.
24, no 5, pp. 625-648, 2010.
Kubera Y., Mathieu P., Picault S. IODA: an interac-
tion-oriented approach for multi-agent based simula-
tions Autonomous Agents and Multi-Agent Systems,
pp. 1-41, 2011.
Leclerc T., Siebert J., Chevrier V., Ciarletta L., Fes-
tor O. Multi-Modeling and Co-Simulation-based
Mobile Ubiquitous Protocols and Services Develop-
ment and Assessment. In 7th International ICST Con-
ference on Mobile and Ubiquitous Systems (Mobiq-
uitous 2010), 2010.
Muller J.-P. Emergence of Collective Behaviour and
Problem Solving. A. Omicini, P. Petta, and J. Pitt
(Eds.): ESAW 2003, LNAI 3071, pp. 1–21, 2004.
Springer-Verlag, Berlin Heidelberg, 2004.
Marcenac P, Giroux S. GEAMAS: A Generic Archi-
tecture for Agent-Oriented Simulations of Complex
Processes. Applied Intelligence, Volume 8, Number
3, pp. 247-267, 1998.
Marilleau N., Cambier C., Drogoul A., Chotte J.-L.,
Perrier E., Blanchard E.. Environnement multi-
échelle à base de fractales pour la modélisation agent
d'écosystèmes. in Proc. JFSMA'08 René Mandiau et
Pierre Chevailler (eds), pp23-32, Cépaduès, 2008.
Ricci A., Viroli M., Omicini A. Give agents their
artifacts : the a&a approach for engineering working
environments in mas. In AAMAS '07: Proceedings of
the 6th international joint conference on Autonomous
agents and multiagent systems, pp601-603, New
York, USA, 2007.
Siebert J. Approche multi-agent pour la multi-
modélisation et le couplage de simulations. Applica-
tion à l'étude des influences entre le fonctionnement
des réseaux ambiants et le comportement de leurs
utilisateurs. Thèse de l'Université Henri Poincaré -
Nancy 1, September 2011.
Siebert J., Ciarletta L., Chevrier V. Agents and arte-
facts for multiple models co-evolution. Building
complex system simulation as a set of interacting
models. In Proceedings of the International Confer-
ence on Autonomous Agents and Multiagent Sys-
tems, in proc. of AAMAS 2010, May 10, 2010, To-
ronto, Canada, pp509-516.
Sloot P., Hoekstra A. Multi-scale modelling in com-
putational biomedicine Briefings in Bioinformatics,
Volume 11, Issue 1, pp. 142-152, 2010.
Traounez, P. Contribution à la modélisation et à la
prise en compte informatique de niveaux de descrip-
tions multiples. Application aux écosystèmes aqua-
tiques. Thèse de l'Université du Havre. 2005.
Troitzsch, K. Multilevel Simulation. In Troitzsch, K.,
Mueller, U., Gilbert, G., Doran, J., eds., Social Sci-
ence Microsimulation. Springer (1996) 107–120
Uhrmacher A., Degenring D., Zeigler B. Discrete
Event Multi-level Models for Systems Biology. in C.
Priami et al. (Eds.), Transactions on Computational
Systems Biology, LNBI 3380, pp. 66–89, Springer-
Verlag Berlin Heidelberg, 2009.
Wilensky, U. NetLogo Flocking model. http://ccl.
northwestern.edu/netlogo/models/Flocking. Center
for Connected Learning and Computer-Based Model-
ing, Northwestern University, Evanston, IL, 1998.
Wilensky, U. NetLogo. http://ccl.northwestern.edu/
netlogo/. Center for Connected Learning and Com-
puter-Based Modeling, Northwestern University,
Evanston, IL, 1999.
|
0705.1757 | 1 | 0705 | 2007-05-12T10:11:57 | Scalability and Optimisation of a Committee of Agents Using Genetic Algorithm | [
"cs.MA"
] | A population of committees of agents that learn by using neural networks is implemented to simulate the stock market. Each committee of agents, which is regarded as a player in a game, is optimised by continually adapting the architecture of the agents using genetic algorithms. The committees of agents buy and sell stocks by following this procedure: (1) obtain the current price of stocks; (2) predict the future price of stocks; (3) and for a given price trade until all the players are mutually satisfied. The trading of stocks is conducted by following these rules: (1) if a player expects an increase in price then it tries to buy the stock; (2) else if it expects a drop in the price, it sells the stock; (3)and the order in which a player participates in the game is random. The proposed procedure is implemented to simulate trading of three stocks, namely, the Dow Jones, the Nasdaq and the S&P 500. A linear relationship between the number of players and agents versus the computational time to run the complete simulation is observed. It is also found that no player has a monopolistic advantage. | cs.MA | cs | SCALABILITY AND OPTIMISATION OF A COMMITTEE OF AGENTS
USING GENETIC ALGORITHM
T. Marwala*, P. De Wilde*, L. Correia**, P. Mariano**, R. Ribeiro** V. Abramov***, N. Szirbik***, J.
Goossenaerts***
*Electrical and Electronic Engineering Department, Imperial College, Exhibition Road, London SW7
2BT, England
**FCT- Dept. Informatics, Universidade Nova Lisboa, 2825-114 Monte Caparica, Portugal
***Information & Technology Department, Faculty of Technology Management, Eindhoven University
of Technology, The Netherlands
ABSTRACT
A population of committees of agents that learn
by using neural networks is implemented to
simulate the stock market. Each committee of
agents, which is regarded as a player in a game,
is optimised by continually adapting
the
architecture of
the agents using genetic
algorithms. The committees of agents buy and
sell stocks by following this procedure: (1)
obtain the current price of stocks; (2) predict the
future price of stocks; (3) and for a given price
trade until all the players are mutually satisfied.
The trading of stocks is conducted by following
these rules: (1) if a player expects an increase in
price then it tries to buy the stock; (2) else if it
expects a drop in the price, it sells the stock; (3)
and the order in which a player participates in
the game is random. The proposed procedure is
implemented to simulate trading of three stocks,
namely, the Dow Jones, the Nasdaq and the
S&P 500. A linear relationship between the
number of players and agents versus
the
computational
time
to
run
the complete
simulation is observed. It is also found that no
player has a monopolistic advantage.
1.
In this work a population of committees of
agents is used to simulate the stock market.
Each committee of agents is viewed as a player
in a game. These players compete and interact.
The committee of agents [1] are optimised using
Introduction
genetic algorithm [2]. Perrone and Cooper [1]
introduced a committee of networks, which
optimises the decision-making of a population
of networks. They achieved this by assuming
that the trained networks were available and
then assigning to each network a weighting
factor, which indicates the contribution that the
network has on the overall decision of a
population of networks. The limitation of their
proposal is that in a situation where the problem
is changing such as the stock market the method
is not flexible enough to allow the dynamic
evolution of the population of networks. This
paper seeks
to fill
the
limitation on
the
committee procedure by ensuring
the
that
individual networks that make the committee are
allowed to dynamically evolve as the problem
evolves. This is achieved by employing genetic
algorithms [2]. The parameters defining the
architecture of the networks, such as the number
of hidden units, which form a committee, are
defined as design variables and are allowed to
evolve as the problem evolves. The network
attributes that are fit to survive replace those that
are not fit. On implementing genetic algorithm
to select the fit individuals three steps are
followed [2]: (1) crossover of network attributes
within the population; (2) Mutation of each
individual attributes; (3) and reproduction of the
successful attributes. The simple crossover, the
binary mutation and roulette wheel reproduction
techniques are used. Finally the proposed
technique is implemented to simulate the trading
of three stocks. The scalability of the number of
Structure of the simulation technique
agents and players in the simulations with
respect to computational time are investigated.
The evolution of
the complexity of
the
simulation as the players participate in more
trading is also investigated.
2.
The structure that is proposed consists of
committees of agents forming a player in the
stock market. Each agent contributes equally to
the overall decision-making of the player. The
simulation framework consists of a population
of these players that compete for fixed number
of stocks.
The agents learn using neural
networks. The schematic illustration of the
committee of agents is shown in Figure 1. The
structure of each agent evolves using genetic
algorithm such that its contribution to the overall
the
function of a committee adapts
to
evolutionary
time-varying nature of
the
problem. The characteristics of the agents that
evolve are the number of hidden units. The
numbers of hidden units are constrained to fall
within a given space, in this study 1 and 10. The
output activation function is randomly chosen to
be either linear or logistic [3].
Agent 1
Agent 2
Agent 3
. . . . . . .
Agent N
N
O 1
(cid:1)=
N
i
Agent
i
Figure 1: Schematic diagram representing the
committee of agents.
•
•
•
Each committee of agents known as a player has
trade stocks with other players. When prices of
stocks are announced, the players trade by
following these rules:
•
Once a price is announced, the committees
look at the current price and the future
price of stocks. The future price is
determined from the agents that that learn
using neural networks. For a player, the
predicted price is the average of the
prediction of each agent within
that
particular player.
If the predicted price of a stock is lower
than the current price, then the player tries
to sell the stock. If the predicted price for
the stock is higher than the current price,
then the committee tries to buy the stock.
At any given stage the committee is only
prepared to sell the maximum of 40% of
the volume of stocks it has.
The amount of stocks that a committee
buys or sells depends on, amongst other
factors,
the
If
the predicted price.
predicted price of a particular stock is x-%
higher
than
the current price,
the
committee tries to acquire x-% of the
volume available on the market of that
particular stock.
This simulation is started by choosing the
number of players that participate in the trading
of stocks together with the number of agents that
form a player. Then the agents are trained by
randomly assigning the number of hidden units
to fall in the interval [1 10]. The output
activation units are randomly chosen to be linear
or logistic. The agents are trained using the data
from the previous 50 trading days. The trained
agents are grouped into their respective players
and are then used predict the next price given
the current price. The players trade stocks using
the rules that are be described later on in the
paper. The schematic diagram of the simulation
is shown in Figure 2.
Mutation of hidden units of agents that make a
player
For each player agents that make are successful
reproduce at a high probability
Choose number of players and agents
Start
After 50 days have elapsed the performance of
each agent is evaluated and the number of
hidden units are transformed into 8 bits and
within each player exchange bits with other
players, a process called crossover. Thereafter,
the agents mutate at low probability. The
successful agents are duplicated while the less
successful ones are eliminated. Then the
networks are retrained again and the whole
process is repeated. When a price is announced
trading of stocks
the
is conducted until
consensus is reached [4]. At this state, the
Stop
Figure 2: Schematic diagram of the trading simulation.
Train the networks by randomising the network architecture. The number of
hidden units is randomly chosen
Evaluate the performance of networks on unseen data by letting
the players trade stocks
Crossover of hidden units of agents that make a
player
Evaluate the performance of networks
on new data by letting the players trade
for given prices of stocks
Introduction to Genetic Algorithms
overall wealth of the committees does not
increase as a result of trading.
3.
Genetic algorithms were inspired by Darwin’s
theory of natural evolution. In natural evolution,
members of the population compete with each
other to survive and reproduce. Evolutionary
successful individuals reproduce while weaker
members die. As a result the genes that are
successful are likely going to spread within the
population. This natural optimization method is
used in this paper to optimize the decision of a
committee of agents in Figure 1. The basic
genetic algorithm suggested by Holland [4] is
implemented.
The algorithm acts on a
population of binary-string
chromosomes.
These chromosomes are obtained by using the
Gray algorithm [5]. Each of these strings is
discretised representation of a point in the search
space. Here we are searching for the most
optimum combination of architectures that form
a committee and that give the least errors.
Therefore the objective function is the error
given by committee of agents. On generating a
new population three operators are performed:
(1)
crossover;
(2) mutation;
(3)
and
reproduction. Just like in natural evolution, the
probability of mutation occurring is lower than
that of crossover or reproduction. The crossover
operator mixes genetic
information
in
the
population by cutting pairs of chromosomes at
random points
along
their
length
and
exchanging over the cut sections. This operator
has a potential of joining successful operators
together. Simple crossover [4] is implemented
in this paper. The mutation operator picks a
binary digit of the chromosomes at random and
inverts it. This has the potential of introducing
population
to
the
new
information.
Reproduction
takes successful chromosomes
and reproduces them in accordance to their
fitness function. The fit parameters are allowed
to reproduce and the weaker parameters are
eliminated. This is done using the roulette
wheel [2] procedure.
4. Application: Simulating
market
stock
the
The example that is considered in this study is
the trading of three stocks, namely: (1) the Dow
Jones; (2) NASDAQ; (3) and S&P 500. The
time-histories of the stocks are downloaded
from the Internet and used to train agents. For a
given set of price of these stocks the committee
of agents predict the future prices of stocks. The
following procedure is followed:
mnt
.
=
mnt
-
t(mn
in
4.
3.
D
p
prices
change
p
)
1. Using the time histories of the stocks
generate N numbers of players trading M
number of stocks. Each player consists of
k-trained networks that form a committee.
2. Given the current price of pmt of the mth
stock the at time t, the nth player predicts
the price of the mth stock pmn(t+1) at time
t+1.
Evaluate
p(
the
+1
)
p
mnt
For the nth player mth stock, calculate the
factor
decision
D=
df
Qp
is
the
1 where Q
mnt
imi
mnt
„=
mi,
maximum quantity of stocks available for
the mth stock.
5. Randomly choose the order in which the
players are allowed to play.
For
player
mnt <
df
)
df
min(
max(
buy the stock.
7. Repeat 3
8.
Set t = t+1;
9. Go to 4.
It should be noted that on implementing this
procedure the total number of stocks available is
kept constant [6]. The sample results showing
the net worth of players are shown in Figure 5.
This figure indicates that sometimes the players
with successful strategies do not necessarily
dominate indefinitely. This is due to the fact
that strategies that are successful in one time
frame are not necessarily successful at a later
time.
6
m
if
then sell else
6.
)
mnt
x 107
h
t
r
o
w
t
e
N
5
4
3
2
1
0
0
600
Day
200
400
Figure 3: The net worth of each player versus
trading days.
800
1000
1200
Scalability
5.
In this section some aspect about the scalability
of the simulations are studied. It is found that
the method proposed is scalable. However,
generally it was observed that the computational
time increases with the increase in number of
agents and players. A linear relationship exists
between the average computational time taken
to run the complete simulation and the number
of players as well as the number of agents that
form a player. This may be seen in Figures 4
and 5.
900
800
700
600
500
400
300
200
)
s
e
t
u
n
i
m
U
P
C
(
n
o
i
t
a
l
u
m
i
s
e
t
e
l
p
m
o
c
e
h
t
n
u
r
o
t
n
e
k
a
t
e
m
i
t
U
P
C
5
10
15
30
25
20
Number of agents that makes a player: number of players is 20
35
40
Figure 4: CPU time versus the number of
players.
200
180
160
140
120
100
80
60
40
)
s
e
t
u
n
i
m
U
P
C
(
n
o
i
t
a
l
u
m
i
s
e
t
e
l
p
m
o
c
e
h
t
n
u
r
o
t
n
e
k
a
t
e
m
i
t
U
P
C
5
10
Figure 5: CPU time versus the number of agents.
15
30
25
20
Number of a players: the number of agents in a player is 4
35
40
6.
Complexity
In this section we present a study of the
complexity of the populations of agents that
make players of the game. In this paper we
view complexity as the measure of a degree of
variation in a population of agents. Each species
of agents form a dimension in space. Each
dimension has a variation indicating the level of
complexity of a population of that species. In
mathematical for this definition of complexity
may be written as:
N
s
=
i e
(cid:1)
i
=
i
1
Complexity
(1)
Figure 6 shows the average number of hidden
units of an agent that forms a player versus the
number of training conducted. This figure
shows that as the system evolves the number of
hidden units for a given player steadily
decreases and stabilizes around 3.
s
t
i
n
u
n
e
d
d
i
h
f
o
r
e
b
m
u
n
e
g
a
r
e
v
A
10
9
8
7
6
5
4
3
2
1
2
4
6
8
18
16
14
12
10
Number of adaptations − this player has 4 agents
20
22
24
Figure 6: Average number of hidden units of an
agent that forms a player versus the number of
training conducted.
committee of agents is found to be a feasible
approach to modelling a player in the stock
market.
Acknowledgement
This work has been carried in the framework of
research project No. IST-1999-10304 supported
by Commission of the European Communities,
European Union.
Reference:
1. Perrone, M.P. and Cooper, L.N., (1993)
When
networks
disagree:
ensemble
methods for hybrid neural networks Ed.
R.J. Mammone, Artificial Neural Networks
for Speech and Vision, Chapman and Hall,
London, UK, 126-142.
2. Goldberg, D.E., Genetic algorithms in
search, optimization and machine learning,
Addison-Wesley, Reading, MA, 1989.
3. Bishop, C.M., “Neural Networks
for
Pattern Recognition,” Oxford University
Press, 1995.
4. Holland, J., “Adaptation in natural and
artificial systems,” University of Michigan
Press, 1975.
5. Michalewicz, Z., “Genetic algorithms +
data structures = evolution programs,”
Springer-Verlag, 1996.
6. Mas-Colell, A., Whinston, M.D., and
Green, J., Microeconomic theory. Oxford
University Press, 1995.
5
4.8
4.6
4.4
4.2
4
3.8
3.6
t
u
p
t
u
o
r
a
e
n
i
l
/
c
i
t
s
i
g
o
l
h
t
i
w
s
t
n
e
g
a
=
s
e
n
i
l
d
e
t
t
o
d
/
d
i
l
o
s
−
y
t
i
x
e
l
p
m
o
C
3.4
0
5
10
15
Adptations − 10 players 4 agents
20
25
Conclusion
Figure 7: Complexity of species made of linear
and logistic output functions versus the number
of training conducted. Solid: Linear; Dashed:
Logistic
The graph showing the complexity of two
species i.e. logistic and linear output activation
functions versus time of training is shown in
Figure 7. This figure shows that on average the
complexity of the agents with linear output units
decreases as the time elapses. The complexity
of the agents with logistic output units is
approximately constant.
7.
A
simulation of
the
stock market was
successfully conducted. It is found that the
number of players and agents that form a player
that participate in the trading game are directly
proportional to the computational time taken to
run the simulation. It is further observed that no
player has the monopolistic advantage on the
prediction of the stock market. The simulation
also shows that as the time of the trading
elapses, the complexity of the players decreases.
This is due to the fact that as the time of trading
elapses the players become more adapted to the
time-varying nature of the problem, thereby
developing common features. Optimising a
|
0910.1865 | 1 | 0910 | 2009-10-09T21:27:36 | Towards Participatory Design of Multi-agent Approach to Transport Demands | [
"cs.MA"
] | The design of multi-agent based simulations (MABS) is up to now mainly done in laboratories and based on designers' understanding of the activities to be simulated. Domain experts have little chance to directly validate agent behaviors. To fill this gap, we are investigating participatory methods of design, which allow users to participate in the design the pickup and delivery problem (PDP) in the taxi planning problem. In this paper, we present a participatory process for designing new socio-technical architectures to afford the taxi dispatch for this transportation system. The proposed dispatch architecture attempts to increase passenger satisfaction more globally, by concurrently dispatching multiple taxis to the same number of passengers in the same geographical region, and vis-avis human driver and dispatcher satisfaction. | cs.MA | cs | IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
ISSN (Online): 1694-0784
ISSN (Print): 1694-0814
Towards Participatory Design of Multi-agent
Approach to Transport Demands
10
Yee Ming Chen1 , Bo-Yuan Wang
Department of Industrial Engineering and Management
Yuan Ze University
135 Yuan-Tung Rd., Chung-Li, Tao-Yuan.
Taiwan, ROC.
1E-MAIL: [email protected]
Abstract
The design of multi-agent based simulations (MABS) is up to
now mainly done in laboratories and based on designers’
understanding of the activities to be simulated. Domain experts
have little chance to directly validate agent behaviors. To fill this
gap, we are investigating participatory methods of design, which
allow users to participate in the design the pickup and delivery
problem (PDP) in the taxi planning problem. In this paper, we
present a participatory process for designing new socio-technical
architectures to afford the taxi dispatch for this transportation
system. The proposed dispatch architecture attempts to increase
passenger satisfaction more globally, by concurrently dispatching
multiple taxis to the same number of passengers in the same
geographical region, and vis-`a-vis human driver and dispatcher
satisfaction.
Key words: Multi-agent, Pickup and Delivery Problemt,
Simulation.
1. Introduction
Transportation systems are among the most ubiquitous and
complex large-scale systems found in modern society. In
many urban cities, public transportation systems are
broadly accepted as the preferred transportation alternative
for commuting to work, performing errands, or traveling to
social events. Taxis are a convenient means of public
transport in many countries. In providing quality customer
service, fast and efficient fleet dispatching is essential
[1,2]. The major focus of taxi dispatch systems has been
primarily on reaching individual passengers in the shortest
time possible to enhance passenger satisfaction. In the taxi
system, most fares start with a phone call. A customer
calls the taxi central and provides his or her location and
desired destination. The taxi central assigns the delivery
task to a taxi driver and drives his taxi to the customer.
Once the customer gets aboard the taximeter starts running
until the destination is reached. Abstracting from this
normal situation we touch the topic of the Pickup and
Delivery Problem (PDP) [3,4,5]. Various researches have
been done in this field to optimize the performance or
profit of the taxi companies. This means reaching the
passengers via the shortest real-time paths possible.
However,
merely increasing individual passenger
satisfaction, as is the current practice, is a local endeavor,
in that it entails assigning the nearest taxi to a customer
prioritized in a first come first serve queue, without
considering the effects of the assignment on other awaiting
customers in the request queue. To improve taxi fleet
service performance, ideally, we should simultaneously
and optimally assign taxis to service all customer bookings
that are made within the time window. This is a
challenging problem confronting current taxi dispatch
systems.
The classic way to solve PDP’s is to use heuristic
algorithms [6]. This is a centralized approach in which a
algorithms calculates the best routes for all the vehicles. It
results in a fully optimized solution for delivering the
goods in the minimum amount of time and costs with the
resources available. The disadvantage is that you need to
know all pickup and delivery points in advance. To deal
with dynamic environment new approaches were
introduced. One of them is the multi-agent approach. In
multi-agent systems several agents work together to find
the best solution for a problem [7].
In this paper, we focus on the multi-agent based
simulations used for a class of problems that treats the
real-time collaboration of human beings having very
different perceptions and actions. We are implementing an
agent-based simulator for this collaborative work. Each
actor is modeled by an agent which is itself replaceable by
a human actor in the simulations. We model the interaction
between the multi- agent by using AnyLogic [8].
2. Taxi dispatch architecture
In the current centralized taxi dispatch architecture in use
by a taxi operator in most of urban cities, incoming taxi
service requests are queued on a first come first serve basis
at the control center. For each passenger (mobile phone
booking) request, available taxis in the vicinity of the
IJCSI
IJCSI
IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
11
passenger location are considered, and a taxi among them
is dispatched to service it, upon the human driver
accepting the booking job (Figure 1.). An efficient way is
to assign a nearby taxi that can traverse the shortest-time
path to the customer location, computed using real-time
traffic information.
Figure 1. Taxi dispatch architecture
3. Simulation tool for participatory taxi
dispatch system design
With the aim of integrating the experts in the design
process, we have implemented an early version of the
simulator to be designed, based on the multi-agent
approach used for taxi dispatch system design. Towards
this end, we propose using passenger, taxi, and control
center agents capable of collaboration. And, to effectively
utilize such software agents for taxi dispatch, we deploy
them in a multi-agent architecture for PDP tasks, proposed
as depicted in Figure 2.
Figure 2. Proposed multi-agent based simulation for participatory taxi
dispatch
3.1 Agents interface with users
An essential element for this type of “participatory taxi
dispatch” is the existence of user interfaces. Their goal is
not so much to help with the evaluation of the
collaborative procedures but more to contribute so that the
users can improve the agent behavior design [9,10]. They
thus provide a perspective on the way in which the agents
make their decisions. Each interface for a user is attached
to an agent which can play alone the role of this user or
IJCSI
IJCSI
IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
12
can assist him. To achieve this, an attentive study of this
agent capacity is needed for efficient human-agent (user-
assistant) interactions.
A user plays the PDP task firstly to understand the agent
decision-making mechanisms, secondly to validate this
mechanism and finally to improve it. While executing the
simulator, he observes the actions performed by other
agents and his assistant. This human-agent relation leads to
a kind of interactive learning [11,12]. To reach their
learning application, a process consisting of three agents is
envisaged (Figure 3):
Figure 3(a) Definition of a Control Center agent
Figure 3(b) Definition of a passenger agent
Figure 3(c) Definition of a taxi agent
We now summarize our major findings as the followings:
The agent is designed to be able to build a log of
validated/amended actions during interaction sessions
with the user. The formed log must be structured in
such a way the user can use it to correct agent
behaviors.
The action log is re-structured so that any agent can
use it to learn its behaviors.
The multi-agent are given the capacity to learn offline
their behavior from the validated/amended actions
conserved in the log.
The learning techniques are applied in the on-line
mode, i.e. during human-agent interactions.
All participating agents are in one of the designated areas
of operation, and the taxi queues at the control center are
updated accordingly. The communication protocol
supporting the dispatch operations can then be prescribed
as the figure 4.
Figure 4 Participating agents interactions
IJCSI
IJCSI
IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
13
4. Simulation and performance evaluation
To study the performance of the proposed multi-agent
based simulation for participatory taxi dispatch, we
conducted personal computer simulations on AnyLogic
(Figure 5(a).)
(http://kem.iem.yzu.edu.tw/
),
which
simulating taxi operations in a selected urban road network
of reasonable complexity. We focused on operational
efficiency, in terms of passenger waiting time versus
empty cruising time. In our simulations, passenger waiting
time is measured from the moment a customer raises a
request to the moment an assigned taxi arrives to pick up
the passenger; empty taxi cruising time is measured from
the moment it is available to the moment it accepts (or
commits to service) a negotiated assignment.
cruising times)( (Figure 5(b).) gathered from the control
center and taxi agents were saved into a text file for off
line performance analysis.
The taxi company has to employ their taxis to satisfy the
passengers around a city. In this city, taxi company
provides PDP service mode. However, smaller taxi fleet
scale will cause higher passenger idle time. Contrary,
larger taxi fleet scale will cause higher taxi idle time. So
what’s the most suitable taxi fleet scale is a very
important issue for taxi company. On the other hand, the
path planning of taxi fleet also affects the passenger
waiting time and taxi idle time. There are two existing
paths planning: Shortest distance and least time. Therefore,
what’s the most suitable path planning for taxi fleet is also
an important issue.
In order to deal with all of issues mentioned above, we
have to construct the multi-agent based simulation for
participatory taxi dispatch to simulate the operation of the
participatory taxi system in this city. And our performance
measures are passenger average waiting time, participatory
taxi average idle time and participatory taxi average
waiting time in queue. Then we will collect and analyze
the data obtained from simulation. Finally, according to
analysis results, we will make recommendations for the
taxi company. The details of decision analysis for taxi
PDP operation system simulation are described as follows.
Figure 5(a). The taxi dispatch within a road network in the proposed
simulator
Figure 5(b). The simulation of taxi PDP operation system
The simulation data (raw passenger waiting and empty
Figure 6 Taxi fleet scale
From figure 6, we can obviously realize that if the scale of
taxi fleet is less than 7, the passenger average waiting time
is higher than others. Contrary, if the scale of taxi fleet is
more than 7, the taxi average idle time is too high. The
most important thing is what the most feasible scale of
taxi fleet is. We decide the scale of taxi fleet is 11. The
reason is that when taxi fleet scale is larger than or equals
to 11, the passenger average waiting time begins
converging.
Second, after deciding the scale of electric taxi fleet, the
company wants to know which paths planning is better or
IJCSI
IJCSI
IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
14
more suitable for its’ taxi fleet and taxi company
operation. In order to find out which paths planning is
better; we vary the participatory taxi number from 5 to 20
to observe the variances of passenger average waiting time
and taxi average idle time.
scale of taxi fleet increases, the participatory taxi average
idle time also increases. Contrary, if the scale of taxi fleet
decreases, the participatory taxi average idle time will
decrease. So we can make a brief summary: The taxi
company applies least time as path planning is more
suitable for taxi fleet and passengers. Even though the taxi
fleet scale increases causes the participatory taxi average
idle time significantly increases, it also represents that the
small scale of taxi fleet can achieve the lower passenger
average waiting time.
1.2000
1.0000
0.8000
0.6000
0.4000
0.2000
0.0000
s
e
m
i
T
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
T axi fleet scale
Figure 7 Passenger average waiting time
From figure 7, we can clearly know all of the
passenger average waiting time is better under least time
path planning. And there is a trend in figure 7. First, if the
scale of taxi fleet increases, the passenger average waiting
time will decrease. Contrary, if the scale of taxi fleet
decreases, the passenger average waiting time will
increase. Second, we find out that if the scale of taxi fleet
is equal to or large than 11, the passenger average waiting
time begins converging. In brief, the company adds more
participatory taxi in the taxi fleet that only decreases small
passenger average waiting time.
s
e
m
i
T
10.0000
9.0000
8.0000
7.0000
6.0000
5.0000
4.0000
3.0000
2.0000
1.0000
0.0000
Least T ime
Shortest Distance
5
6
7 8
9 10 11 12 13 14 15 16 17 18 19 20
T ax i fleet s cale
Figure 8 Taxi average idle time
In figure 8, we can clearly know most of the
participatory taxi average idle time is worse under least
time path planning. And there is a trend in figure 8. If the
Least T ime
Shortest D istance
5. Conclusions
The new idea of multi-agent based simulation for
automating participatory taxi dispatch is introduced in this
paper. The proposed participatory taxi dispatch
architecture
realizes the idea using collaborative
assignment through a proposed multi-agent architecture.
Using GIS simulations on an urban road network model,
we evaluated the performance of the proposed dispatch
system, and showed that, even on a basic infrastructure
(Fig. 1), the distributed multi-agent system approach is
promising in terms of improvements in operational
efficiency over the existing centralized approach.
Acknowledgements
This research work was sponsored by the National Science
Council, R.O.C., under project number NSC97-2221-
E-155-039.
[3]
References
[1] D. H. Lee, H. Wang, R. L. Cheu, and S. H. Teo, “A taxi
dispatch system based on current demands and real-time
traffic information,” Transportation Research Record, vol.
1882, pp. 193–200, 2004.
[2] Z. Liao, “Real-time taxi dispatching using global
positioning systems,” Communications of the ACM, vol. 46,
no. 5, pp. 81–83, May 2003.
J. Kozlak, J.C. Créput, V. Hilaire and A. Koukam,” Multi-
agent approach to dynamic pick-up and delivery problem
with uncertain knowledge about future transport demands”.
Fundamental Informaticae 71, p. 27-36,2006.
[4] N. Neagu, K. Dorerand M. Calisti,” Solving distributed
delivery problems with agent-based
technologies and
constraint satisfaction techniques”, Whitestein Technologies
AG, Zurich, Switzerland, 2006.
[5] C. Cubillos, B. Crawford, and N. Rodríguez, ,”Distributed
Planning for the On-Line Dial-a-Ride Problem”. Lecture
Notes in Computer Science (including subseries Lecture
Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics) 4742 LNCS: pp 160-169,2007.
J. Renaud, F. F. Boctor, and J. Ouenniche. “A heuristic for
the pickup and delivery traveling salesman problem”,
Comput. Oper. Res.,27(9): pp.905–916, 2000.
[6]
IJCSI
IJCSI
IJCSI International Journal of Computer Science Issues, Vol. 4, No. 1, 2009
15
[7] H. Laıchour, S. Maouche, R. Mandiau,” Traffic control
assistance in connection nodes: multi-agent applications in
urban transport systems”, Comput. Int. Sci. J. 1 ,pp.37–42,
2002.
[8] Xjtechnologies
software Anylogic® v5.2 package.
http://www.xjtek.com/,2005.
[9] H. Ezzedine, A. Trabelsi, C. Kolski,” Modelling of agent
oriented interaction using petri nets, application to HMI
design for transport system”, P. Borne, E. Craye, N.
Proceedings CESA2003 IMACS
Dangourmeau (Eds.),
Multiconference Computational Engineering in Systems
Applications, Lille, France, July 9–11, Ecole Centrale Lille,
Villeneuve d’Ascq, , pp. 1–8, 2003.
[10] H. Ezzedine, A. Trabelsi and C. Kolski,” Modeling of an
interactive system with an agent-based architecture using
Petri nets, application of the method to the supervision of a
Mathematics and Computers in
transport system”.
Simulation, 70:pp 358–376, 2006.
[11] H. Ezzedine, C. Kolski, and A. Pe’ninou, “ Agent-oriented
design of human-computer interface:application to
supervision of an urban transport network” . Engineering
Applications of Artificial Intelligence 18: pp 255–270, 2005.
[12] M. Majid, A. Aickelin, and P. O. Siebers, “ Human
behaviors modeling for discrete event and agent based
simulation : A Case Study”. Annual Operational Research
pp 111-158,2007.
Yee Ming Chen is a professor in the Department of
Industrial Engineering and Management at Yuan
Ze University, where he carries out basic and
applied research in agent-based computing. His
current research interests include soft computing,
supply chain management,
and system
diagnosis/prognosis.
Bo-Yuan Wang is a graduated student in the
Department of Industrial Engineering and
Management at Yuan Ze University,, where he
is studying basic and applied research in
agent-based programming.
His current
research
interests
include web-based
programming and system analysis/synthesis.
IJCSI
IJCSI
|
1307.0284 | 2 | 1307 | 2013-07-03T09:06:09 | The effectiveness of altruistic lobbying: A model study | [
"cs.MA",
"cs.SI",
"eess.SY",
"math.OC",
"physics.soc-ph"
] | Altruistic lobbying is lobbying in the public interest or in the interest of the least protected part of the society. In fact, an altruist has a wide range of strategies, from behaving in the interest of the society as a whole to the support of the most disadvantaged ones. How can we compare the effectiveness of such strategies? Another question is: "Given a strategy, is it possible to estimate the optimal number of participants choosing it?" Finally, do the answers to these questions depend on the level of well-being in the society? Can we say that the poorer the society, the more important is to focus on the support of the poorest? We answer these questions within the framework of the model of social dynamics determined by voting in a stochastic environment. | cs.MA | cs | The effectiveness of altruistic lobbying: A model study
P.Yu. Chebotarev, Z.M. Lezina, A.K. Loginov, Ya.Yu. Tsodikova
Institute of Control Sciences of the Russian Academy of Sciences, Moscow, Russia
1. Introduction
Altruistic lobbying is lobbying in the public interest or in the interest of the least protected part of the
society. This “or” is the point: it is necessary to bear in mind that an altruist has a wide range of strategies, from
behaving in the interest of the society as a whole to the support of the most disadvantaged ones. How can we
compare the effectiveness of such strategies? The second question is: “Given a strategy, is it possible to assess
the optimal number of participants choosing it?” Finally, do the answers to these questions depend on the level
of well-being in the society? For example, can we say that the poorer the society, the more important is to focus
on the support of the poorest? We answer these questions within the framework of the model of social
dynamics determined by voting in a stochastic environment [1–5].
Consider a society consisting of n members (agents); each of them is characterized by a real scalar
interpreted as the utility (or capital) level; a negative value is naturally interpreted as a debt. Let the initial
distribution of capital be given. A proposal of the environment is a vector of increments of individual capitals.
Let these increments be random variables; in the simplest case, we assume that they are independent and
identically distributed with mean μ and standard deviation σ. The environment generates a series of proposals
each of which is put to the vote. Every agent takes part in the voting and has one vote. With some voting
procedure, the profile of votes is converted into a collective decision: the proposal is approved or rejected. The
approved proposals are implemented: the participants get the capital increments specified in the proposal.
Considering a large series of votes, one can explore the dynamics of the vector of agents’ capital: in different
environments, with different social attitudes, and with various voting procedures. An interesting version of the
model is that with ruining: the agents whose capital values become negative are ruined.
2. A study of altruistic lobbying
Consider a society consisting of two categories of agents: selfish and altruistic ones. An egoist votes for
a proposal if and only if this proposal increases his capital.
As noted in the introduction, the altruists have a wide range of strategies. Consider the following type
of strategy. We order all agents by the increase of the current capital value. Since the capital increments are
real-valued random variables, we can assume that all capitals are different. Let an integer be fixed. For
the current proposal, one calculates the total capital increment of poorest agents. An altruist supports the
proposal if and only if this total is positive, no matter what happens to his own capital. If , then the
altruist supports those proposals that enrich the society as a whole. When is smaller, then he supports more
or less numerous “lower” stratum of the society.
Let K be the same initial capital of all the agents. Assume that the distribution of capital increments is
Gaussian: ( ) Consider the case where bankrupt agents drop out. Let m be the number of steps of voting
constituting the “game”. The strategy of altruists is determined by . As a criterion of the effectiveness of this
strategy we consider the relative number of participants that “survive” till the end of the game. Another
important parameter is the number (percentage) of altruists in the society.
We study the following question: “How the effectiveness of altruists’ strategy depends on their
proportion and the threshold at different parameters μ and σ?” In other words, “What is the optimal
proportion of altruists in the society and what is their optimal strategy in a favorable (μ > 0), neutral (μ = 0),
and unfavorable (μ < 0) environment?”
Some simulation results are as follows. Let the society consist of agents; the number of
altruists, m, varies. The experiments are conducted at σ = 12; game lasts 500 steps. On the horizontal axes in
Fig. 1, we have the number of altruists, m, and their strategy parameter ; along vertical axis, the relative
number of agents at the end of the game. In the diagrams, μ and the initial capital K of the agents are also
shown. Some results are as follows.
1. In a neutral environment (μ = 0) and with the support of the poorest ( is small), the number of
altruists should be small, or they waste their influence, resulting in ruining players outside the “support
screen”. At a high , we have saturation with the increase of the number of altruists. With K = 40, the
optimum is reached at = 50 and m > 40 (at higher values of m, saturation holds). In other words,
altruists should support the lower half of the community. With the extension of the support screen, the
effectiveness of the altruistic strategy declines.
Fig. 1.
2. In an unfavorable environment (μ = –0.5) and a high initial capital (K = 130), the picture is very
similar. The only significant difference is that there is no pronounced effect of reducing the
effectiveness of the strategy with the increase of the right border of the support screen. Maximum is
reached at n0 = 30 and m = 10, but it does not change significantly with an increase of these parameters.
3. In an unfavorable environment (μ = –0.5) and a lower initial capital (K = 40), the maximum is reached
at n0 = 80 and m = 60. This is a very interesting conclusion: in a really harsh environment, we have to
support everyone!
4. Finally, in a supportive environment with a high initial capital (μ = 0.5, K = 130), the picture changes
dramatically: we should support a thin layer of the poorest; the others will “resurface themselves”.
The main conclusion is simple: in rich and affluent societies, support of the poorest is effective. The
poorer is the society, the more “wasteful” is the support of the poorest. If you focus on this support, then the
zone of trouble grows. A more detailed analysis of the results reveals some more subtle phenomena as well.
References
1. V.I. Borzenko, Z.M. Lezina, A.K. Loginov, Ya.Yu. Tsodikova, and P.Yu. Chebotarev. Strategies of voting in
stochastic environment: Egoism and collectivism // Automation and Remote Control, 2006, V. 67. No. 2.
P. 311–328.
2. P. Yu. Chebotarev. Analytical expression of the expected values of capital at votin g in the stochastic
environment // Automation and Remote Control, 2006. V. 67. No. 3. P. 480–492.
3. P.Yu. Chebotarev, A.K. Loginov, Ya.Yu. Tsodikova, Z.M. Lezina, V.I. Borzenko. Snowball of cooperation and
snowball communism // Proceedings of the Fourth International Conference on Control Sciences. Moscow ,
2009. P. 687–699 (in Russian).
4. P. Yu. Chebotarev, A. K. Loginov, Ya. Yu. Tsodikova, Z. M. Lezina, V. I. Borzenko. Analysis of collectivism and
egoism phenomena within the context of social welfare // Automation and Remote Control, 2010. V. 71. No. 6.
P. 1196–1207.
5. P.Yu. Chebotarev, A.K. Loginov, Ya.Yu. Tsodikova, Z.M. Lezina, V.I. Borzenko. Voting in a stochastic
environment: The case of two groups // Automation and Remote Control, 2011. V.72. No.7. P. 1537–1547.
|
1509.05112 | 1 | 1509 | 2015-09-17T03:33:26 | Models and Representations for Fractal Social Organizations | [
"cs.MA"
] | Our subject is oriented towards investigation of potential ways of societal organization, that allow for collective intelligent organization and management of resources. The main objective of such organizations is the exploration of the social energy from the existing societies. We conjecture that an organizational model that fulfills the mentioned requirements is the Fractal Social Organization (FSO). Our goal is to prove and verify the effectiveness of this model by performing various simulations using the NetLogo environment, a tool that allows agent-based rapid prototyping. We begin by simulating trivial real life activities that demonstrate the main properties of the core unit of the FSO, namely the SoC. Further, more complex scenarios involving various nested SoCs are simulated. Two main simulation models are presented, allowing us to obtain preliminary results using the FSO concepts as a potential solution. In the first simulation model we demonstrate that by the use of FSO properties the individuals may benefit by receiving more qualitative healthcare services, while in the second simulation model we show how it might be possible to improve the fall detection systems by the use of FSO mechanism. | cs.MA | cs |
UNIVERSITY OF ANTWERP
Faculty of Science
Department of Mathematics and Computer Science
2014-2015
Models and Representations for Fractal Social
Organizations
Arianit Pajaziti
A thesis submitted in partial fulfillment
of the requirements for the degree of
Master of Computer Science:
Computernetworks and Distributed Systems
Promoter:
Vincenzo De Florio
Co-promoter:
Chris Blondia
Contents
List of figures
List of Tables
1 Introduction
1.1 Contributions and goals
. . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Structure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Service oriented Communities
2.1 Mutualistic relationships in SoC . . . . . . . . . . . . . . . . . . . . .
3 Fractal Social Organizations
4 First model
iii
vi
1
4
5
7
9
13
18
4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2
"Walk in park" activity . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.3 Car sharing activity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4
"Go to office" activity . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.5
"Houses on fire" event . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.6 Case study: Health care services . . . . . . . . . . . . . . . . . . . . . 32
4.6.1 Traditional Case
. . . . . . . . . . . . . . . . . . . . . . . . . 32
i
4.6.2 FSO Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.6.3
"Health care" event . . . . . . . . . . . . . . . . . . . . . . . . 33
4.6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
ii
5 Second model
43
5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Simulation Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.3 Simulation Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Conclusions
Future Work
A Exemplary FSO structural organization
60
62
66
List of Figures
2.1
In a) the SOA architecture and b) a simple representation of SoC [9].
2.2 SoC representation [8].
. . . . . . . . . . . . . . . . . . . . . . . . . .
8
9
2.3 Abstract representation of a Mutualistic relationship in SoC [12]. . . . 11
2.4 Merging of two SoC members in a group performing the same activ-
ity [12].
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1 Exemplary FSO structure [8].
. . . . . . . . . . . . . . . . . . . . . . 14
4.1 A screenshot of the simulation area, with the initialized agents.
. . . 19
4.2 A screenshot of the work-space from the NetLogo environment used
for simulations.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.3 A representation of the hierarchy of Service-oriented Communities
(SoC) that forms the FSO. . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Mutualistic relationship in the "Walk in park" activity.
. . . . . . . . 21
4.5 A representation of the flow of "walk in park" activities in the se-
quence diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.6 Message exchange between the agents and SoC coordinator.
. . . . . 23
4.7 Mutualistic Relationship in car sharing activity.
. . . . . . . . . . . . 24
4.8 A representation of the flow of car sharing activities in the sequence
diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
iii
iv
4.9
Individuals performing the "go to office" activity by various means
of transportation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.10 A representation of the flow of "walk in park" activities in the se-
quence diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.11 A simple graph showing the usage of various means of movement of
individuals for performing the "go to office" activity.
. . . . . . . . . 27
4.12 FSO collaboration: Agents from two distinct SoCs handling the
"House burning" event. . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.13 Agents from the Local community SoC handling the "House burning"
event solely.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.14 Results from the execution of "Burning houses" event with the use
of FSO infrastructure.
. . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.15 Results from the execution of "Burning houses" event without using
the FSO infrastructure for collaboration between SoCs.
. . . . . . . . 31
4.16 A Netlogo screenshot of phases that "individual 256" goes through in
case of "health-care" events.
. . . . . . . . . . . . . . . . . . . . . . . 35
4.17 Netlogo screenshot of phases that "individual 256" goes through in
case of "health-care service" events. . . . . . . . . . . . . . . . . . . . 36
4.18 A sequence diagram of the flow of messages in the case where the
main Local Hospital handles the "healthcare" services exception. . . . 37
4.19 A sequence diagram of the flow of messages in the case where "noam-
bulances" exception is handled.
. . . . . . . . . . . . . . . . . . . . . 37
4.20 A sequence diagram of the flow of messages in the case where "nodoc-
tors" exception is handled . . . . . . . . . . . . . . . . . . . . . . . . 37
4.21 Comparison of patients treated/died ratio for each of the three sce-
narios for various threshold values.
. . . . . . . . . . . . . . . . . . . 40
4.22 Average Querying time for each of the three scenarios for various
threshold values.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.1 NetLogo screenshot of the simulation area with initialized agents.
. . 47
v
5.2 A representation of the FSO structure of the simulation model. . . . . 48
5.3 Number of alarm requests handled by the system for S1 and S2 with
a number of IC's ranging from 0 to 40.
. . . . . . . . . . . . . . . . . 55
5.4 FP ratio for S1 and S2 with a number of IC's ranging from 0 to 40.
. 56
5.5 FN ratio for S1 and S2 with a number of IC's ranging from 0 to 40.
. 56
5.6 Average social costs of S1 and S2 with a number of IC's ranging from
0 to 40.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.7 Average waiting times with S1 and S2 and a number of IC's ranging
from 0 to 40.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
List of Tables
5.1 Pseudo-code of the DA agents. . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Pseudo-code of the CA agents. . . . . . . . . . . . . . . . . . . . . . . 50
5.3 Pseudo-code of the IC agents.
. . . . . . . . . . . . . . . . . . . . . . 51
5.4 Pseudo-code of the MA agents.
. . . . . . . . . . . . . . . . . . . . . 52
5.5 Pseudo-code of the PC agents. . . . . . . . . . . . . . . . . . . . . . . 53
5.6 Results of S1 scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.7 Results of S2 scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . 59
vi
Chapter 1
Introduction
Present-day technological development has reached a level where people are sur-
rounded by digital devices and information at every moment. They have become
an inseparable part of our lives. By the use of the inter-connectivity capabilities of
such devices we are exposed to an extensive range of information resources and
services. As a result endless communication and collaboration forms have pio-
neered. The newly established communication channels have become our source
of knowledge, leading to new forms of socialization. This has set the ground for
knowledge-based [1], cyber-physical societies [2] with enormous communication and
organization opportunities. These socio-technical systems characterized by virtually
instantaneous dissemination of knowledge have produced novel forms of collective
intelligence and social interaction. The complexity level of such socio-technical sys-
tems or societies keeps growing in time, and as a consequence, state of the art
organizational and structural approaches are in demand in order to maximize the
output, while deriving efficiency.
The emergence of sought properties such as economic and social welfare for all;
sustainability with respect to natural ecosystems; and especially manageability and
resilience, highly depends on the way social organizations are designed. The fast
growth of the human population is inducing a number of unprecedented critical
situations on the natural ecosystems. In particular, natural resources are becoming
scarcer, and the old ways in the organization of the human action reveal all their
limitation. A more intelligent way of organization and management of resources is
1
2
urgently called for. The ultimate achievement would be a society operating as a
large pool of cooperating, intelligent, mobile resources. This requires learning how
to engineer social-energy-aware solutions.
A typical case in point is given by traditional organizations operating in domains
such as healthcare and crisis management. A common assumption characterizing
those organizations is the adoption of a strict client-server model. This produces at
least the following major consequences:
Fragility. The artificial distinction: an active and a passive side of society severely
affects quality-of-emergence [3, 26]. This classification introduces penalties in
the performance of the society, meaning that only a subset of the social actors
is available to serve the whole set. As well known, the fast growth and the
progressive aging of the human population are introducing shortages in the
subset of "service providers". The problem is no more the ever increasing
social costs. The management of a fragile society that is unable to serve
its citizens is of the most concern. Provided the current trends the state of
unmanageability is down the line.
Absence of a referral service. Although provided with an incomplete view of
the current state of the server and its availability, it is the expected from
the client to identify which server to bind to.
It is the users responsibility
to know, e.g., which emergency service to invoke, which civil organization to
refer to, which hospital to call first, and so on. There are existing Referral
services, however they mainly cover a single domain (i.e., healthcare) and
very specific cases. An example observed in [4] is the seamless transfer of
patient information from a primary to a secondary practitioner. Such services
mostly possess an incomplete view of the available resources, because of their
specialization.
Lack of unitary responses to complex requests. To the best of our knowledge
there is no referral service that is capable of providing with composite response
to complex requests such that the action, knowledge, and assets of multiple
servers are combined and orchestrated in an automatic way. Even electronic
referral systems in use today are mostly limited [5] and only provide prede-
fined services in specific domains. Observed examples of such systems include
3
SHINE [6] and SHINE OS+ [7]. As a consequence, in the face of complex
servicing requests calling for the joint action of multiple servers, the client is
basically left on its own. Societal organizations do not provide unitary re-
sponses nor assist the client in composing and managing them. Reasons for
this may be found in lack of awareness and also improper shift of service
identification responsibilities from the server to the client.
The described vulnerabilities of current organizations urge the need for mutating
our organizational paradigms and assumptions. An organizational approach that
enables the utilization of the full potential of our societies is on demand. Conse-
quently, the artificial distinction between an active and a passive subset should be
removed or at least significantly reduced. Instead of restricting and declining the
participation of roles, the organizations should promote the participation of roles,
thus act as a provider and enabler. Moreover, the organization should stimulate the
cooperation among the role players at all levels, from the citizens to the governing
institutions. The knowledge should flow among the organization's participating en-
tities highlighting the needs and requirements. By exchanging such information, the
organization will have the capability of self-orchestrating responses, which lead to
right and timely decisions.
Our society's challenge is to define one such organizational model, while preserving
the fundamental aspects of the identity of the organization. A further requirement
is learning how to guarantee the resilience of our evolved organizations.
A potential answer for new ways of societal organization was proposed by De Flo-
rio [8]. The proposed organizational model represents a hierarchical distributed
organization of socio-technical systems, called "Fractal Social Organization" (FSO).
The FSO is a nested organizational structure, built of the lower-level organizations,
named Service-oriented Communities (SoC) [9]. SoCs serve as the FSOs building
block.
FSO takes the responsibility as a generalized referral service for the orchestration
of complex social services.
In FSO, the rigid client-server scheme of traditional
organizations is replaced by service orientation, while bottleneck-prone hierarchies
are replaced by communities of peer-level members. Role appointment is not static
and directed by the organization, but rather voluntary and context-driven.
It is
our conjecture that the just stated new design assumptions allow for the creation
1.1 Contributions and goals
4
of smarter societal organizations able to match the complexity of our new complex
world. In the next section we state the main goals and contributions of this thesis.
1.1 Contributions and goals
Our subject is oriented towards investigation of potential ways of societal organiza-
tion, that allow for collective intelligent organization and management of resources.
The main objective of such organizations is the exploration of the social energy from
the existing societies. We conjecture that an organizational model that fulfills the
above mentioned requirements is the Fractal Social Organization (FSO). Our goal is
to prove and verify the effectiveness of this model by performing various simulations
using the NetLogo environment[13], a tool that allows agent-based rapid prototyp-
ing. The development in NetLogo environment is based on the Logo programming
language.
We begin by simulating trivial real life activities that demonstrate the main proper-
ties of the core unit of the FSO, namely the SoC. Further, more complex scenarios
involving various nested SoCs are simulated. Two main simulation models have been
implemented.1 The first one simulates a virtual world, where people perform daily
activities, while the normal flow is perturbed either by some natural event, which
causes damage on their properties or they start having health issues, and as a result
healthcare services are demanded. We compare the efficiency of the FSO approach
in contrast to the Traditional social ways of organization. A detailed description of
this model is provided in Chapter 4.
The second model simulates an Ambient Assisted Living (AAL) organization, where
elder persons residing at their home are in need for healthcare services. By struc-
turing additional agents based on the FSO approach we try to improve the accuracy
of traditional ITC system. A detailed description is provided in Chapter 5.
The objectives of the simulations are:
• Design and implementation of FSO models that mimic social communities.
• Validation of the properties and behaviors of such organizational approaches
in the healthcare domain.
1Link for downloading the simulation files:
https://www.dropbox.com/s/y2vlb1ftmm5a3ns/simulation_files.rar?dl=0
1.2 Structure
5
• Collection of measures and quantitative evidence of the performance of the
FSO organizations based on the simulated models. In particular we are inter-
ested in service turn-around-times and service coverage.
• Assessment of resilience and trustworthiness with respect to timeliness, and
availability. Goal is to verify that the strategies enacted by the FSO do not
impact negatively on the design goals of existing, "traditional" organizations
(e.g., the current hierarchical organization of a hospital).
The work conducted during this thesis led to the writing of two papers in close
collaboration with the Promoter of the thesis. The first one with the title "How
Resilient Are Our Societies? Analyses, Models, and Preliminary Results" has been
accepted in the proceedings of the "Third World Conference on Complex Systems"
(WCCS15)2. The second paper with the title "Tapping Into the Wells of Social
Energy: An Example Based on Falls Identification" is submitted in "The 17th In-
ternational Conference on Information Integration and Web-based Applications &
Services" (iiWAS15)3.
1.2 Structure
The thesis is organized into four main chapters. In Chapter 2 we begin by describing
the main characteristics and properties of Service oriented Communities (SoC). In a
separate section we describe the establishment of Mutualistic Relationships within
the SoC communities. In Chapter 3 we give the main definitions of the Fractal Social
Organizations and describe their structure, based on the SoCs. Further, Chapter 4
introduces the first simulation model. Here we discuss the activities/events and the
implications they have on the individuals agents. The main focus is in the Health
care services case study. The results of the performed experiments are provided in
a separate section within this chapter. Chapter 5 describes the second simulation
model, namely the Ambient Assisted Living organization and the provision of health
care services to elder persons agents. After the description of the case study a
detailed interpretation of the results is performed, where we discuss the extent to
2Web: http://mscomplexsystems.org/wccs15/
3Web: http://www.iiwas.org/conferences/iiwas2015/
1.2 Structure
6
which our goals are met. Finally we give a conclusion of the thesis and the Future
work in separate sections.
Chapter 2
Service oriented Communities
A Service oriented Community (SoC) represents a socio-technical system composed
of peer entities that share services under the supervision of a coordinating agent
[9]. These peer entities can be anything, including human beings, cyber-physical
things etc. The participants of a SoC are named "Members". Members have dis-
tinct capabilities and goals. In [8] the characteristics of the members are named as
"Viewpoints". Within a SoC the members do not have specific roles assigned, in
terms of service providers or receivers. Depending on the context a member may
benefit by receiving a certain service, while another one provides services. In this
sense upon the occurrence of an event, members react accordingly, thus becoming
active. By becoming active they play a role in the community. The service re-
quests, or offered services are published to a service registry. The same stands for
events. The service registry is handled by the coordinating member of the SoC.
Based on the service registry information the coordinator is capable of estimating
the capabilities and availability of community members. Thus for a given particular
situation a member is assigned to play a role that enables an action. By enrolling
a member for a particular action, a certain cost is estimated. Enrollments are done
with a goal that might change from case to case. Sometimes a focus is given in the
speed of realization, while on others safety may be the priority or cost-effectiveness.
The optimization goals may be individual or social oriented. The member of the
community that hosts the service registry and semantic processing can be seen as
a representation of the domain of community. When an event triggers or a situa-
tion happens the coordinator is responsible for recognizing the similarities between
7
8
members and binding them together. An advantage of such organization is that
it brings great communication and linkage between the members, with the aim of
maximizing resource usage.
As observed in [9], the above mentioned organization reminds us of Service Oriented
Architecture (SOA) used in distributed systems. The participating entities or mem-
bers here have the role of a service, which is able to provide or request for an action.
The coordination is done depending on the situation at hand and the roles available.
Figure 2.1 highlights the similarities between SoC and SOA.
Figure 2.1: In a) the SOA architecture and b) a simple representation of SoC [9].
The interaction of different roles within a SoC creates Mutualistic relationships.
This behavior is described in [12] where it was concluded that different actions
performed by an entity have an effect on the domain of other entities being it
positive or negative. A discussion of Mutualistic Relationships in SoC is provided
in the next section. A more detailed representation of the organizational structure
of a SOC is depicted in Figure 2.2.
It is observed that SoC functions as a flat society, or a cloud of social resources
orchestrated by the focal point -- the service coordinator. A drawback of such
organization is that if the number of members that are under the management of
a single coordinator becomes very large, then we would have an overload of the
functions of the coordinator, which results in slow service times. Additionally, the
concentration of orchestration functions on a single node results in a single point
of failure issue. These issues are overcome by the Fractal Social Organizations,
described in Chapter 3. FSO represent a scalable version of the SoC.
2.1 Mutualistic relationships in SoC
9
Figure 2.2: SoC representation [8].
2.1 Mutualistic relationships in SoC
Mutualistic relationships represent the basis of collaboration between species in
nature.
In a mutualistic relationship the participating entities benefit from each
other. An example in nature is the one of bees and flowers. The bees fly from
flower to flower to gather the nectar, necessary for nutrition of their colonies. As
they land in a flower the pollen residing on the flower is attached to their hairy
bodies. When the bee moves to the next flower, some of the pollen attached to their
body is rubbed off their body falling on the other flower, thus pollinating the plant.
The pollination is the process of pollen transfer from flower to flower allowing for
plant reproducing. Like this both entities benefit from the process, the bees gather
the food, while the plants get to be reproduced. This exchange of services between
participating entities represents a typical case of mutualistic relationship, where each
of them benefits from the process. A mutualistic relationship can involve more than
one participating couple, thus creating a chain. De Florio in [12] introduces some
core definitions of a mutualistic relationship model. In what follows we will describe
these main definitions.
2.1 Mutualistic relationships in SoC
10
Action function. Given two domains or systems X and Y, AX represents a set of
actions that can be performed in X, while AY corresponds to actions in Y. The
bijective function:
act : AX → Ay
is a mapping of actions from domain X to Y.
(2.1)
Evaluation function. Given a system X and it's action set AX the following func-
tion:
evalX : AX → IX
(2.2)
maps actions, to a semantic evaluation of significance in domain X. The eval-
uation resides within at least one of the three classes: positive, meaning that
the action was beneficial, neutral - action is evaluated as insignificant, and
negative - action is disadvantageous.
Mutualistic precondition. Given two systems X and Y, corresponding action
sets AX and AY, and IX, IY, the mutualistic precondition between X and Y
fulfills the following conditions:
∃a ∈ AX : evalX(a) ≥ 0 ∧ evalY (act(a)) > 0
∃b ∈ AY : evalY (b) ≥ 0 ∧ evalX(act−1(b)) > 0
(2.3)
(2.4)
The first condition states that an action in X is evaluated as positive and
neutral, and its impact in domain Y is positive, while the second condition
implies the vice-versa, an action in Y evaluated as positive/neutral has a
beneficial impact in domain X.
Mutualistic relationship. Two domains or systems X and Y are in mutualis-
tic relation, when X and Y trigger individual actions manifested as a form
of social behavior that fulfill the mutualistic preconditions. The mutualistic
relation between two domains is written as X R Y.
Extended Mutualistic relationship. Finally, in [12] the notion of Extended mu-
tualistic relationship is defined. Given two systems X and Y by removing two
conditions from formula (2.3) and (2.4) the author describes the systems where
2.1 Mutualistic relationships in SoC
11
the establishment of a mutualistic relationship, has a certain cost for the entity.
The modified preconditions are modified as follows:
∃a ∈ AX : evalY (act(a)) > 0
∃b ∈ AY : evalX(act−1(b)) > 0
(2.5)
(2.6)
Figure 2.3: Abstract representation of a Mutualistic relationship in SoC [12].
Service oriented Communities or SoCs use semantic description of members services
to allow their matching, manifested as a mutualistic or extended mutualistic rela-
tionship. In Figure 2.3 an abstract representation of a mutualistic relationship is
depicted. The figure shows two actors part of two different systems. The activity
A1 triggered by actor1 involves the provision of services that are requested by ac-
tor2, while activity A2 triggered by actor2 provisions services requested by actor1.
If there are multiple members part of a SoC that provide the same services, and
the corresponding activities fulfill each other, the activities can be merged into a
so-called group activity, and the members are bound together. An example from
daily life, would be the one where two individuals are willing to go for a walk with
someone. Both members are providing and requesting the same services, thus they
can be bound by the walking activity. This merging of two members is depicted in
Figure 2.4 as a green dashed box.
2.1 Mutualistic relationships in SoC
12
Figure 2.4: Merging of two SoC members in a group performing the same activ-
ity [12].
The newly formed member of the SoC provides walking services and is open to other
members (member3 or member4 from Figure 2.4) that might request the offered
service. This group of members requests for a place to go for walking, e.g. a park.
A 4th member that offers such service (location) can come in help. The creation
and functioning of members of a SoC as groups sets the basis for the Fractal Social
Organizations. FSO are described in detail in the following chapter.
Chapter 3
Fractal Social Organizations
Fractal Social Organizations (FSO), introduced in [8], represent a class of socio-
technical systems characterized by a distributed, bio-inspired hierarchical architec-
ture. FSO can be described as a greater organizational structure, whose building
blocks are the SoCs. The SoCs of various levels and sizes communicate with each
other, thus enabling the information flow over all FSO levels. This set of rules that
allows for spontaneous emergence of "social overlay networks" (SON) is called a
"canon" [14], [17]. The FSO structure, can be seen as a nested SoC, whose mem-
bers are other SoCs. The SoC components at the same time act as individual and
social entities.
The key characteristic of FSO is the peer-to-peer approach for the collaborating
nodes, meaning that based on the given situation, a node can play the role of a
coordinating agent, or a simple participating member offering certain services. The
communication between the various hierarchical levels is realized by the use of the
so-called "exception" messages, explained more formally in what follows. These
messages are triggered by the coordinators of a lower level, which make decisions
whether an event should involve other entities part of the hierarchy or not. This
approach allows for multiple redundant responses for a given situation. An example
of FSO structural organization is observed in Figure 3.1. The SoC are placed at the
layer members.
FSO allow for aggregation of services offered by humans and technological compo-
13
14
Figure 3.1: Exemplary FSO structure [8].
nents. Given a certain event, the coordinator starts searching for roles by matching
their service by semantic matching for solving the problem. Whenever the solution
cannot be found an exception propagates to the next level of FSO. The procedure is
repeated until a solution is found. The response results in the SON (social-overlay
network) -- ad hoc SoC. We now introduce the main components of an FSO.
Agent. An agent is a member of the system that has an active behavior. It can
be a person, a digital device, a cyber-physical system, or a collection of these
entities. Agents have certain characteristics (for instance, mobility) and can
be equipped with assets (e.g. objects).
Community. FSO may be described as a hierarchy of the Service-oriented Com-
munities (SoC). Communities can be associated either to functional places
(e.g. a flat, a building [10], a hospital) or a region characterized by the occur-
rence of some event, such as an accident or a fire. Seen from a higher level
of the FSO a SoC is also an agent. Agents belong to communities. Based on
the said definitions one might easily derive that a child community can be a
participating member of another larger, parent community. This leads to the
term of nested communities.
Community representative or coordinator. A SoC community is represented
by one of its agents. This agent hosts the service registry of the SoC and is
seen as the representation of the domain of the community.
15
If the given community is of a small size, concretely an agent together with its
assets we call it an "individual SoC" (or iSoC). The community representative
is at the same time a coordinator of the community it represents, and a member
of the parent community.
Role. The Agents or members of an FSO can take a role, meaning that they perform
actions that belong to a certain class of activities (e.g. doctor-, driver-, nurse-,
or patient-specific activities).
Notification. Agents publish notifications, which can be of various natures. The
notifications are received by the community representative, which reacts ac-
cordingly depending on the type of notification. A notification can be a simple
message that informs about the status of an agent, a request for service, iden-
tification of a certain event etc.
Match. For each received notification, the community representative verifies [11]
whether the new notification and those already known "enable" actions.
Action. Actions are performed by agents. An action represents a series of steps
undertaken by the agents that initiate activities. By performing actions the
agents play a certain Role within the community, thus becoming an active
part of it. Every initiation of an action is followed by a notification to the
community representative.
Exception. For a given action or event if there is a shortage of roles capable of han-
dling it, a message is triggered from the community representative towards the
higher levels of the FSO. This message is called an Exception. The exception
message traverses the higher level communities until all the roles are allocated
or until a flooding threshold is overcome and a failure is issued.
Social Overlay Network. Once we have an enabled action, the participating
agents become a new temporary SoC. The lifespan of the new born SoC in lim-
ited to the duration of the activity. If the activity involves agents from distinct
SoCs, the new temporary SoC is coupled of agents from different communi-
ties. The newly formed community brings together nodes from different layers
of the FSO hierarchy. This temporal organization is called a Social Overlay
Network (SON).
16
FSO is based on inter- and intra-community cooperation. This allows for establish-
ment of mutualistic relationships between the members of within a community or
distinct communities. The dynamic infrastructure of FSO allows for cooperation
across the scales of the service, thus substituting the existing rigid hierarchies. The
following points describe a general scenario of event handling by FSO:
• An event takes place. Participating agents of an SoC sense the change.
• The agents become active, thus implicitly play a role. The change of the
status of an agent is followed by a notification published to the community
representative.
• The notification reaches the SoC representative -- for instance, a healthcare
institution.
• The community representative handles the notification and performs local
analysis for managing the response.
• An optimal response is found within the community. All roles are allocated
within the current "circle", thus the response is enacted and the state is ad-
justed. If there is no optimal response found the following point is performed.
• Exception message is propagated requesting for roles capable of accomplishing
the need. The messages traverse through the upper layers of the FSO hierarchy
until the response is finally enabled or a failure is declared.
• During the process execution, new knowledge is accrued both locally and glob-
ally. The response is adjusted according to the alternatives offered at various
levels of FSO.
Cases that require the above described organizational properties are quite
widespread.
In [26] the Katrina hurricane case that took place in New Orleans
is analyzed. The occurrence of such an event led to multiple and diverse conse-
quences. Several "social layers" were disrupted in parallel. This triggered multiple
responses from the different "social layers", starting from: the Private circle (individ-
uals, families, neighbors etc.), Private organizations, Local institutions (City Police,
Fire brigades, Flood rescue etc.), State organizations (Department of emergency
17
management), to end with the National circle. One major problem that followed
the event was the coordination of these social layers. A conclusion has been derived
showing that the lack of coordination between different institutional organizations
of different levels lead to catastrophic consequences to the human society of that
area. Moreover, the classification of the communities in informal respondents (the
private cycle) and institutional respondents resulted in conflicting goals and actions.
It can be easily observed that previously introduced properties of the FSO closely
match the organizational and coordination requirements of the given scenario.
In the upcoming chapters we describe two simulation models that are implemented
based on the given FSO concepts.
Chapter 4
First model
4.1 Overview
In this chapter, we describe one of our simulation models, its implementation and
functionality. For simulations the NetLogo environment [13] is used. The simula-
tion environment is coupled of a simulation area that functions as a "virtual world",
where the initialized agents perform their actions. It also contains a set of controls
that allow for adjustment of input parameters, such as the number of agents, and
various graphs that display the simulation measurements. The agents can move
freely throughout the simulation area in discrete-time steps called "ticks". A vi-
sualization of the simulation area containing the initialized agents can be seen in
Figure 4.1, while the complete work-space is depicted in Figure 4.2.
In our simulation model we define SoCs of various scale and specializations, which
together form a greater organizational structure, namely the FSO. Fig. 4.3 repre-
sents the hierarchical community structure used in the simulation experiments. The
atomic components of the defined organization are the individual SoCs, or iSoCs.
An iSoC represents a "small version" of a SoC and can be represented by a single
individual (agent) and his/her personal belongings, e.g. an individual and his per-
sonal vehicle. In this case the individual is the coordinator or the representative
of its iSoC, while the vehicle is a member of the given iSoC. We define agents of
diverse natures including Individuals, Doctors, Firefighters, and Taxi drivers. A
complete structure of the designed FSO and the participating members can be seen
18
4.1 Overview
19
Figure 4.1: A screenshot of the simulation area, with the initialized agents.
in Appendix A. The number of agents given in the table is not definitive, it changes
depending on the simulation scenario. The emphasis is on the individuals agents
and their actions or activities. By performing an activity individuals implicitly play
a Role within the Community. Whenever an activity is performed by an individual,
the representative of the corresponding Community records the state of the agent.
In cases when there are activities that request for resources that aren't available
within the community, the coordinator triggers "exception" messages to the up-
per level of the organization. This allows the coordinating agent to manage the
underlying resources and establishment of Mutualistic relationships.
As the simulation starts each individual starts performing an activity. The activity
can be a simple process like "going to market" or a more complex one as "going
to office", where the citizen has to be in office, at a certain time. For the latter,
the individual has to perform time, speed and distance calculations to find out the
fastest way of being there.
4.1 Overview
20
Figure 4.2: A screenshot of the work-space from the NetLogo environment used for
simulations.
There are six types of activities defined. On each "tick" of the simulation if there is
an idle individual, an activity is triggered for him. The activity type is chosen ac-
cording to the random1 function of NetLogo. For each activity we chose a triggering
probability of 0.18, except for the "health care" activity/event for which the chosen
probability is 0.09. This because it is considered that the chances for an individual
to get sick or be in need of healthcare services are smaller compared to other daily
activities. The activities have a certain lasting time measured in "ticks". A detailed
description of each of the activities is provided in the following sections. Our focus
is on the "health care" events, while the other simpler activities come in use for
simulating daily life activities, and help describing the main properties of SoC and
FSO. The simplest activities are those of "talk on the phone", and "go to market"
which are performed solely by the individual and do not involve any collaboration
between the agents. In the next section we discuss the "Walk in park" activity.
Figure 4.3: A representation of the hierarchy of Service-oriented Communities (SoC)
that forms the FSO.
1Random numbers: http://ccl.northwestern.edu/netlogo/docs/programming.html#random
4.2 "Walk in park" activity
21
4.2 "Walk in park" activity
In Section 2.1 we gave the main properties of Mutualistic Relationships. Here we
demonstrate the establishment of Mutualistic Relationships in terms of "Walk in
park" activity.
(a) Two agents
(b) Multiple agents
Figure 4.4: Mutualistic relationship in the "Walk in park" activity.
One of the basic activities to be performed by the Individuals is going for a walk in
park. The activity can be performed individually or in group. The former simulates
the situation where a person decides to go walking in the park alone, while the latter
simulates a group of individuals walking together in the park. In order to realize the
group activity the individuals must show interest of chatting with others while walk-
ing. The group activity is defined as a composite activity, containing the chatting
sub-activity. Once an individual chooses to perform such an activity, an informa-
tional message is sent to the corresponding SoC coordinator, containing information
regarding the individuals preferences for the sub-activity and its current state. In
cases when the individual sends a positive message, meaning that is interested to
perform the activity together with someone, the coordinator starts checking for a
matching activity (another individual performing the same activity, and willing to
collaborate). If there is a match, a Mutualistic Relationship between both individ-
uals is established and they proceed performing the activity together. Figure 4.4
depicts an extracted screen-shot from the simulated scenario. In Figure 4.4(a) the
establishment of the Mutualistic Relationship between two individuals (agents) can
4.2 "Walk in park" activity
22
be seen. The individuals performing such activity are colored in pink, same as the
relation which is visualized as a two-way arrow between the two.
As described in Section 2.1 a Mutualistic Relationship can be established between
two ore more agents. For the latter case, more than two individuals should be will-
ing to perform the "Walk in park" activity, while chatting at the same time. An
example of such scenario is depicted in Fig. 4.4b). Whenever one of the agents
finishes the activity he/she proceeds with the other activities, and the links to the
other agents are disconnected, thus ending the relationship.
A detailed representation of the flow of messages from the activity request sent from
individuals towards the coordinator, until the establishment of Mutualistic Relation-
ships is depicted in the sequence diagram in Figure 4.5. The green boxes represent
the process where the establishment of a Mutualistic relationship is performed, while
the vertical colored bold lines represent the duration of the activity.
Figure 4.5: A representation of the flow of "walk in park" activities in the sequence
diagram.
In the next section we describe the establishment of the Mutualistic Relationship in
a scenario where not only Individuals agents are involved.
4.3 Car sharing activity
23
4.3 Car sharing activity
Another activity that can be performed by individuals is traveling to a certain lo-
cation, within the two dimensional virtual world. It is assumed that the location is
too far for the individual to be reached on foot, thus a car is needed for transporta-
tion. A new type of agents introduced in the simulation are the cars. An individual
may or may not posses a car in his iSoC circle. In the former case he can travel to
a certain destination, and also offer sharing the ride. The destination of the indi-
vidual in need for transportation must be in the vicinity of the destination of the
individual that is offering to share the ride in order for the activity to be performed
together. Both individuals benefit from the performed activity, as one party gets
the transportation, while the traveling costs are decreased.
Figure 4.6: Message exchange between the agents and SoC coordinator.
The Individuals that are to perform the car sharing activity or "go to location"
action, as defined in the simulation, inform the corresponding SoC coordinator about
their state. In case the coordinator finds a match, meaning that there is an individual
offering to share a ride, and another one in need for transportation, the activity is
performed together resulting with the establishment of a Mutualistic Relationship.
We model such scenario in our simulation. The message exchange between the
agents extracted from the log of the simulation is depicted in Figure 4.6.
For the individuals that aren't in possession of a car and want to perform such an
activity an invalidation threshold for the activity is defined. Once such threshold
is exceeded the activity is canceled. Another type of activity can be performed
afterwards. Figure 4.7 depicts the three main phases of car sharing activity.
Another representation of the flow of the car sharing activities performed by the in-
4.4 "Go to office" activity
24
(a) The communication between
(b) The individual owning the
the agents is established.
car picks up the other Indi-
vidual.
(c) The car owner drops the
Individual, and continues rid-
ing towards its destination.
Figure 4.7: Mutualistic Relationship in car sharing activity.
dividuals is presented in the Sequence diagram in Figure 4.8. Same as in the previous
cases the green square shows the moment of Mutualistic Relationship establishment
between the two individuals, where an individual provides a certain service, which
turns out to satisfy the request of another individual. The next section introduces
another type of agent, which is specifically related to transportation of individuals.
4.4 "Go to office" activity
A somewhat more complex activity performed by individuals is the "go to office"
activity. There are three possible ways to perform the "go to office" activity: on
4.4 "Go to office" activity
25
Figure 4.8: A representation of the flow of car sharing activities in the sequence
diagram.
foot, by car, or by taxi. Figure 4.9 depicts an exemplary scenario extracted from
the simulation where the individuals are going towards the office in one of the three
above mentioned means of transportation.
Figure 4.9: Individuals performing the "go to office" activity by various means of
transportation.
Immediately after such an activity is triggered for an individual, a calculation of
the distance from its position to the office's location is performed. It is assumed
that the individuals already have knowledge about the location of their office. The
activity has a defined starting time. In order to reach the office at a particular time,
the individual has to choose a way of transportation that allows him to be there
on time. The decision is made after a distance-time calculation is performed by
the individual. In cases when there is enough time available to reach the office the
individual will go on foot. This type of activity can be performed independently,
4.4 "Go to office" activity
26
thus doesn't involve any resource sharing or collaboration. In case the individual is
late to walk, and possesses a car, it can be used to reach the office faster. However,
some individuals lack of personal cars. In this case the exception message is raised
towards the coordinator of the SoC, which tries to locate agents capable of resolving
the given situation within their circle. This could be another individual, willing
to share the ride (car sharing activity) or the newly defined type of agents within
the SoC, the taxi-drivers together with their cars. These agents wander around
the simulation area all the time, aiming to help individuals that are in need of
transportation. The transport request is forwarded by the coordinator towards an
available taxi. The taxi starts moving towards the individual immediately, and
after picking him up they move together towards the office. Again a Mutualistic
relationship is established within the Local residents circle where for a particular
cost the taxi-driver helps the individual with transportation. There might be cases
when no taxis are available immediately, thus the individual will be late for work,
but again less time is lost. Figure 4.10 shows a sequence diagram of the "go to
office" activities performed by the individuals.
Figure 4.10: A representation of the flow of "walk in park" activities in the sequence
diagram.
Running the simulation for 15000 ticks, with 4 offices, 60 Individuals, with only 15 of
them owning cars, and 10 Taxi-drivers/Taxis, the results as depicted in Figure 4.11
are achieved.
It can be observed that the dominating form of transportation used by the indi-
viduals was on foot, followed by the other two forms of transportation, which have
similar allocation. The scenario where the individuals are transported with the help
of taxidrivers shows the social benefits of such organization. The number of times
a civilian is transported by taxi represents the number of established mutualistic
4.5 "Houses on fire" event
27
Figure 4.11: A simple graph showing the usage of various means of movement of
individuals for performing the "go to office" activity.
relationships. In the graph this is depicted by the yellow line. In the next section
we describe a scenario, where the collaboration between two distinct SoCs within
the FSO is involved. As we will see next, it is possible for members of distinct SoCs
to establish Mutualistic Relationships.
4.5 "Houses on fire" event
The "Houses on fire" event mimics the scenario where some catastrophic natural
event takes place that has an impact on the individuals properties, and as a result
the houses catch on fire. In our simulator we define an event generator, that triggers
"house burning" events where a number of houses start burning.
After a few cycles from the start of the simulation, some random houses will start
catching on fire, thus disrupting the daily activities of the individuals. The number
of events to be generated or houses to catch on fire is given initially as an input value
upon initialization of simulation. The process is repeated after a certain amount of
simulation cycles ("ticks") until all the houses are either completely or partially
burned.
Each house has its own "health" level which starts at 100, and as it continues
4.5 "Houses on fire" event
28
burning, the value decreases by a "firelevel" damage value, which initially varies
within 1 to 5 range, and after each cycle it might be increased by 1 unit with
probability 0.5. Once the "health" level of the house reaches 0, it is considered that
the house is completely burned. The "firelevel" value is used to describe the damage
that the fire is causing to the house.
Whenever there is a "house burning" event, if an individual is within the observing
radius2 of 3 of the burning house he will postpone his current activity in order to
provide help. The postponed activity is continued later. In the simulation area an
arrow connecting the individual to the burning house is shown and this operation
will start decreasing the "firelevel" value of the house by 0.5. This mimics the
"help" provided by the individual to extinguish the fire.
In case the individual
realizes that the fire is getting out of control ("health" level of house is below 80),
and the situation cannot be handled by itself solely, an "exception" message is
sent to the Local residents coordinator. Given the current situation and the lack
of resources for helping for fire extinguishing the SoC coordinator forwards the
"exception" towards the Level 2 Emergency response organization. Among others,
this organization contains the Firefighters organization member within its circle.
The complete structure of the FSO was introduced earlier in Figure 4.3.
The Firefigters SoC is composed of Fire-trucks and Firefighters. These members are
also organized in iSoC fashion, where one fire-truck can hold a number of firefighters
ranging from 1 to 4. The number of firefighters in a truck has an impact on the time
that takes to extinguish the fire in a burning house.
The response of the firefighters organization comes immediately, by assigning a fire-
truck and a certain number of firefighters to go and extinguish the fire at the given
house in the "exception" message. This can be observed in the simulation area when
one of the white trucks will turn red and start moving towards the house on fire
(with color red). The firetrucks have greater capabilities for extinguishing fire, thus
they will decrease the "firelevel" value of the burning house by a value of 2 - 5 units
depending on the number of firefighters in the fire truck.
Two simulations are performed with regard to "Houses on fire" event. The number
of initialized agents is as follows: 50 Houses, 50 Individuals, 10 Fire Trucks, and
35 Firefighters. The "House burning" events are generated 10 per cycle, every 100
cycles. The number of activities to be performed by individuals is six. We run
2Netlogo measurement unit used to measure the distance of an agent to other objects.
4.5 "Houses on fire" event
29
(a) A view of the individual and fire-truck helping extin-
guish the fire of the house.
(b) A view of the scene after the intervention of the individ-
ual and the fire-truck. Fire is extinguished, the fire-truck is
freed, while the individual continues his previous activity.
Figure 4.12: FSO collaboration: Agents from two distinct SoCs handling the "House
burning" event.
the simulation for two scenarios: the first one there is no use of the FSO infras-
tructure, thus no "exception" messages are transmitted to firefighters organization
and individuals cope with the triggered events solely, and in the second one the
FSO infrastructure is established and the SoCs will collaborate. An extract of the
simulation area showing the agents handling the "Burning houses" event for each
scenario is shown in Figures 4.12, and 4.13. In the scenario where no FSO infras-
tructure is used, we see a group of three individuals are helping to extinguish the
fire for the given house, thus they succeed. This isn't always the case as the number
of individuals that are in the vicinity of a burning house isn't that high for regular
cases, and most of the times the houses end up completely burned. Such behavior is
4.5 "Houses on fire" event
30
also reflected in the simulation outcomes of the scenario explained in what follows.
(a) A view of a group of individuals helping extinguish the
fire of the house.
(b) A view of the scene after the intervention of the group
of individuals. Fire is extinguished, while the individuals
continue their previous activities.
Figure 4.13: Agents from the Local community SoC handling the "House burning"
event solely.
The metric used to describe the outcome of the simulations is the total number of
completely burned houses. The higher the number, the worse the result.
In the
case where FSO collaboration is enabled the number of completely burned houses is
29, while in the case where there is no FSO infrastructure established this number
reaches 49. The results for each scenario are depicted in the graphs in Figures 4.15,
and 4.14 respectively. In the first scenario we see that the "private social circle"
responds to the event solely, while in the second we scenario we have a joint col-
laboration between the private and the "local institutions circle". By establishing a
proper communication mechanism between the two level SoC communities we ob-
serve that the community benefits with less waste of resources. This reminds of the
4.5 "Houses on fire" event
31
Figure 4.14: Results from the execution of "Burning houses" event with the use of
FSO infrastructure.
Figure 4.15: Results from the execution of "Burning houses" event without using
the FSO infrastructure for collaboration between SoCs.
Katrina hurricane scenario described in chapter 3. It is observed that by a better
organization and collaboration between the "social layers" a better management of
resources is established. In the next section we discuss the Health care services case
study, which represents one of our main focuses of the thesis. We show that by the
full use of FSO capabilities better health care services, in terms of time management
and resource allocation can be offered to individuals.
4.6 Case study: Health care services
32
4.6 Case study: Health care services
A key cause of inefficiency in the interaction between users and social organizations
is often the fact that said interaction is almost completely managed by the indi-
vidual or by individual parts of the overall care system. The partial and outdated
knowledge that the individual has about the available social resources and their state
often translates in time-consuming and sequence querying or polling of services. The
major added value of FSO lies in the fact that it constructs a unitary Whole that
takes responsibility to assemble a coherent response to complex requests for service,
thus avoiding the inefficiencies.
Based on many real-life scenarios including those described in [15],[16] the formula-
tion of the following simulation scenarios was inspired:
Mary has just delivered her baby when she and her husband Joseph are told that
the little one suffers from unexpected complications requiring specific support that
is not available at the private clinic where they are now.
4.6.1 Traditional Case
Following the practices of traditional systems the initial step is to call the emergency
service. A verbal description of the situation and the immediate requirements is
provided. At the emergency service premises the case is tagged as life critical and
immediate attention is taken. The operators of the emergency service might have
knowledge regarding the number of hospitals in the vicinity and their specialties.
However, no further information is known in terms of the hospital's resources and
their availability. In search of this information, calls are made to hospitals, according
to some ordered list. Through the calls, a potential hospital that might be able to
handle the case is identified. An ambulance is assigned for transporting the little
one and her parents to the chosen hospital. Upon reaching the hospital, the doctors
discover that the situation is more complex than the way it had been described
on the phone. Due to the incomplete initial knowledge, it is concluded that the
needed treatment is not available at the current hospital. Calls are made to other
hospitals from the list. Another one is identified as a possible candidate to supply
the necessary treatment, however they lack a specific medical device crucial for
the current case. A third hospital located in a further distance is reached on the
4.6 Case study: Health care services
33
phone. From the conversation it is understood that the hospital is in possession of
the necessary medical device. If a complete knowledge of the resources and their
availability had been available, a mutualistic collaboration between the hospitals
would have fulfilled all the prerequisites for dealing with the case.
4.6.2 FSO Cooperation
A notification of the type "service request" is published from the emergency service
operator, thus reaching the coordinator of the corresponding SoC [8, 9, 18]. Having
partial knowledge about the event the SoC triggers an exception for lack of roles.
The exception reaches the regional SoC which includes the originating SoC and other
hospitals. At this level a more complete view of the available resources is available.
The regional SoC discovers that none of the members can fulfill the request on
their own. However, a combination of a set of roles from distinct SoCs (hospitals) is
observed to provide with a response. The members of the identified SoCs are merged,
thus becoming a SON (a temporary SoC). The new SoC elects a coordinator and
enriches the request with new roles necessary to make it possible for the SON to
deal with the case at hand. An ambulance is dispatched for transporting the little
one and his family, while at the same time another transporting vehicle is assigned
to bring the necessary medial devices to the hospital where the treatment will take
place.
4.6.3 "Health care" event
From all the previously simulated activities the type of events or activities that we
focus are the ones where Individuals are in need of health-care services. In these
cases, we say that the Individual takes the role of the patient.
Throughout the simulation among the standard activities, "health care" events are
triggered by Individuals, reporting that they are in need of health-care services.
The severity of patient's condition in need for treatment is given as a value in the
1 - 10 range. Values 1 - 3 represent a minor illness, and it is considered that the
patient can pay a visit to the local hospital by itself, whereas for values 4 - 10 an
ambulance is required to transport the Individual to one of the hospitals. There are
several hospitals defined, each of them in possession of a certain number of doctors,
4.6 Case study: Health care services
34
ambulances and medical appliances. A doctor can treat all three minor illnesses,
and is an expert for three other diseases. In order to treat the patient the doctor
uses corresponding medical appliances. In some cases the treatment of a particular
condition of patient requires the use of several medical appliances simultaneously.
The request for health-care services issued by a patient must be handled within a
defined threshold, otherwise it is considered that the condition degrades and results
in patient's death. In our "virtual world" this case is mimicked by removing the
Individual from the simulation. Based on the way the health-care request is handled
we simulate three distinct scenarios, namely the Traditional Organization Case, the
FSO case, and Perfect Oracle case.
In the Traditional Organization case, the individual in need for health-care services
picks a random hospital and triggers the request for visit. In cases when the value of
patient's condition is larger than 3 the hospital responds by sending an ambulance
to transport him. Once the patient reaches the hospital a doctor is assigned to treat
the patient, and based on the patient's condition the appropriate medical appliances
needed for treatment are allocated.
It might happen that there are no available
resources within the hospital to treat the patient, e.g. no available doctors that are
experts on treating a particular condition, or there are no medical appliances. In
these cases a "polling" in the other hospitals is performed. The patient's request
is forwarded to the next hospital, without an exact knowledge about the hospital's
resource availability. The procedure is repeated until the patient reaches a hospital
that is able to offer the appropriate treatment, or the request exceeds the predefined
threshold, and the patient dies.
The FSO approach treats the health-care service events in a different manner. First,
the Individual in need of health-care service raises an exception from its iSoC to the
representative of the corresponding SoC (Local residents community). The repre-
sentative or the coordinating agent handles the exception and checks for available
resources able to cope with the given event within its circle. In case of a negative
response the exception is forwarded to the upper level representative, namely the
"Emergency response" agent. Located at a higher hierarchical level the "Emergency
response" has broader knowledge regarding the agents that might be able to deal
with the situation at hand. After locating "Local hospital A" (the main local hospi-
tal) as a potential helping agent, the notification is directed towards it. The "Local
Hospital A" is well informed about its underlying agents, and resources.
In the
4.6 Case study: Health care services
35
(a) Patient waiting for ambu-
(b) Ambulance transporting
lance.
the patient.
(c) The arrival of patient in
hospital.
Figure 4.16: A Netlogo screenshot of phases that "individual 256" goes through in
case of "health-care" events.
positive case it replies to the exception by assigning a Doctor and an Ambulance (if
necessary) to deal with the given condition of the patient. If it happens that the
given Local Hospital is out of available resources capable of handling the given sit-
uation, or in case the condition of the patient is complex and the hospital isn't able
to cope with it, an exception towards "Regional hospitals" level is triggered. The
"Regional hospital" will be set in charge of finding another hospital in possession
of missing resources. With this setup the request will traverse the agents spread
throughout the layers of FSO organization, until the needed resources are found.
This process is manifested by the establishment of a new SON, or a temporary SoC,
coupled of shared resources from distinct SoCs, and it is the key feature behind FSOs
effectiveness. This represents a case where Mutualistic Relationship is established
4.6 Case study: Health care services
36
among members of distinct communities. A graphical representation of the phases
a patient goes through, before the start of its treatment is presented in Fig. 4.16
Fig. 4.17 shows an extracted command log from the simulator representing the
message flow between the agents. Fig. 4.17a) shows a trivial case of SON, where
the received exception, is handled within the local hospital SoC, without a need for
additional resources. We call this case an Infra-Community Cooperation. Fig. 4.17b)
represents a more complex case, where resources need to be shared between two local
hospitals, and this represents the Inter-community cooperation case.
(a) Infra-Community.
(b) Inter-community cooperation.
Figure 4.17: Netlogo screenshot of phases that "individual 256" goes through in case
of "health-care service" events.
In our simulation model we define three types of exceptions for local hospitals,
namely:
• lack of doctors capable of treating a particular type of patients disease
• no ambulances available, to transport patients with severe condition
• medical appliances needed to treat the patient missing.
In this way, by triggering exceptions whenever resources are missing within a partic-
ular hospital an interacting channel between hospitals is established, allowing them
to function as a greater organization, where resources are shared accordingly with
4.6 Case study: Health care services
37
an optimized usage.
The sequence diagrams that depict the flow of messages in cases where the exception
is handled within local hospital, local hospital lacks of ambulances for transporting
the Individual, and local hospital lacks of doctors capable of curing the Individual
are depicted in Figures 4.18,4.19, 4.20 respectively.
Figure 4.18: A sequence diagram of the flow of messages in the case where the main
Local Hospital handles the "healthcare" services exception.
Figure 4.19: A sequence diagram of the flow of messages in the case where "noam-
bulances" exception is handled.
Figure 4.20: A sequence diagram of the flow of messages in the case where "nodoc-
tors" exception is handled
Another scenario considered is the Perfect Oracle. Here the Individuals are assumed
to have perfect knowledge regarding the hospitals and services offered. However
in this setup there is no resource sharing between the hospitals. To simulate the
communication between Individuals and Hospitals the FSO "infrastructure" is used.
4.6 Case study: Health care services
38
A typical scenario would be one where the Individual in need of health-care services
reaches the hospital and once the doctor starts the treatment, it finds out that
there are no medical appliances available needed for treatment. In this case since
the hospitals don't share the resources, the patients request is disregarded and it is
directed to another hospital. The process is repeated until the Individual reaches
the appropriate hospital that offers treatment, otherwise the request is invalidated.
The health care service request is divided in two phases, the Querying time which
represents the time form the moment an Individual is in need of treatment until the
moment when all the needed resources are allocated, and the second phase which is
the Treatment time and it is the time it takes for a patient to get cured from the
moment when the treatment begins. In our experiments we are interested in the
former as it represents the measurement that helps differentiating the efficiency of
the approaches. The aim of simulations is to set a comparison point between the
different approaches and highlight the efficiency of FSO, when dealing with complex
situations involving agents generating requests of various natures. The experimental
results and a comparison between the introduced approaches is discussed in the next
section.
4.6.4 Results
In this section we evaluate the benefits of FSO with respect to Perfect oracle and
Traditional organization. The experiment outcomes are generated from the NETL-
ogo simulations. The simulations are run for all three scenarios for 3000 "ticks", and
with a fixed number of community members. There are 4 Local hospitals defined, as
well as 15 Doctors, 8 Ambulances, and 70 Medical appliances. The Doctors, Ambu-
lances and Medical appliances are distributed throughout the 4 hospitals according
to the NetLogo random function, generated by a deterministic process. This means
that using the same random "seed" the experiments can be reproduced. This ap-
proach was useful in our experiments since we wanted to have the same distribution
of resources for all three models, in order to have an equitable comparison.
To evaluate the model we use two main metrics, namely the Average Querying
Time, and the Number of Patients which could not receive the treatment within the
defined threshold, and thus die. The Average Querying time represents the average
4.6 Case study: Health care services
39
time an Individual in need for "health care" services has to wait until he receives the
necessary treatment. We ran the experiments for each scenario with three different
thresholds, as depicted in Fig. 4.21, 4.22. For each specified threshold the simulation
is ran 5 times, with a distinct number of Individuals, starting with 60 until 140, with
an increasing step of 20. Additionally for the FSO we use the SON metric, which
describes the number of times a solution to the given situation is solved as a result of
inter-community cooperation. This number also represents the usage of communities
social energy. Contrary to FSO, for the Traditional Organization where there is no
cooperation between the Hospitals, we measure the number of failures a patient
faces until the matching hospital for treatment is found.
Figure 4.21 shows a comparison of the number of treated patients, and the number
of patients died for each scenario. For example, in the simulation experiment with
a Threshold of 150, for 140 Individuals the following results are archived:
FSO 189 patients are treated, out of which 56 have died,
Perfect Oracle 177 treated, 71 died,
Traditional organization 175 treated 93 died.
Not only the number of patients treated in FSO is higher, but also the number of
deaths is lower. It can be observed that in all cases FSO performs better, followed
by Perfect Oracle, and Traditional Organization. In the FSO case the number of
patients that fail to receive the treatment is always lower, due to the sharing of re-
sources between the hospitals, while for Perfect Oracle this number is higher because
the patient knows at which hospital he should go, but if the hospital lacks neces-
sary medical appliances it doesn't borrow them from other hospitals, it redirects
the patient instead. In the traditional organization we have no interaction between
hospitals neither. However, in this case the patient chooses the hospitals randomly,
without any knowledge about the available doctors, ambulances or medical appli-
ances. This poor level of information leads to delays in resource discovery, thus the
number of patients that fail to receive treatment is the highest.
The fast service response-time of FSO, also allows for a higher number of patients
to be treated for the same number of simulation cycles. This is observed in the
experiment results too.
4.6 Case study: Health care services
40
(a) Threshold 150.
(b) Threshold 200.
(c) Threshold 250.
Figure 4.21: Comparison of patients treated/died ratio for each of the three scenarios
for various threshold values.
Another observation if we compare the scenarios from the three graphs of Figure
4.21, is that with an increase of the Threshold the number of patients that die
decreases, same as the number of patients treated. This because the request for
"health care" services has more time to receive a response, thus less requests are in-
validated. The number of patients treated also decreases, as the number of generated
cases decreases slightly, because of the longer period of service reception.
Fig. 4.22 depicts the performance of all three models with respect to Average Query-
ing time. If we analyze the querying time results, again we see that FSO outperforms
4.6 Case study: Health care services
41
the other two models. Such outcome comes as a result of exceptions usage, which
allow for fast inter-community communication and resource allocation. It can be
observed that by decreasing the number of Individuals, the FSO improves drasti-
cally, and the number of patients that fail to receive the treatment decreases two to
three times.
(a) Threshold 150.
(b) Threshold 200.
(c) Threshold 250.
Figure 4.22: Average Querying time for each of the three scenarios for various
threshold values.
The SON metric used for FSO ranges from around 100 up to 220 depending on the
number of Individuals participating in the simulation. The higher the number of
Individuals, the higher the SON metric. This due to the increase of number of col-
laborations between communities. On the other hand the number of service failures
4.6 Case study: Health care services
42
in the Traditional Organization ranges from around 50 up to 200. By increasing
the number of Individuals the number of failures increases too. This because the
number of times a patient fails to reach the appropriate hospital increases with a
higher number of Individuals.
Chapter 5
Second model
5.1 Overview
Falls present the major public health problem among elderly persons. According
to [19] falls are the "most significant cause of injury for elderly person" -- more
specifically "the most serious life-threatening events that can occur" in the 65+ age
group [20]. The key factor that determines the severity of a fall is the time until the
treatment is delivered [22],[20].
The number of systems aimed at fall detection has had a dramatic increase over
the recent years. Most of them are based on devices such as accelerometers, gyro-
scopes, or other sensors. A major limitation of such systems lies on the difficulties
of the algorithms and their implementations to provide a reliable assessment for a
fall under real life conditions. Other approaches are oriented towards smart-phones
and their rich set of embedded sensors. [23],[24],[25]. Despite the continuing tech-
nological progress, it is still very difficult to find a monitoring system that is able
to determine whether a person has actually fallen or whether, for example, he or
she has changed his position very quickly. A possible improvement might come by
the use of redundancy. For example two accelerometers of different technology and
design are coupled together [19]. As a result we would have an improved sensitiv-
ity. However, this solution comes with an impact on the social costs spent for the
verification of the false alarms.
There are two possible orientations when one tries to approach this issue. One
43
5.2 Simulation Model
44
way may be to try and improve the current ITC monitoring systems. There have
been numerous research actions, with the intention to improve both sensitivity and
specificity parameters, and yet no standardized solution has been found [22]. An-
other possibility it to search for alternative solutions. Here we discuss one way to
do so. Our proposal is to go beyond the purely technological solution and involve
the humans in the system. Consequently, a socio-technological system is introduced
and analyzed by means of the provided multi-agent simulation model. By the use
of the organizational structure of the FSO we show that the appointment of fall
verification tasks to a group of human agents allows to overcome, to some extent,
the limitations and inefficiencies of purely ITC-based solutions.
In what follows we briefly introduce a number of key measurements that will help
us evaluate the outcomes of our simulation model:
False positive ratio False Positive (FP) ratio or false alarm rate is the probability
of concluding that an event occurred, whereas this is not the case. The alarm
is fired, although the event did not take place.
Specificity Specificity is the probability that the event did not take place and the
system did not fire. Specificity is equal to 1 − FP.
False Negative ratio False Negative (FN) ratio is the probability that the system
indicates that an event has not occurred, when in fact it took place. It can
also be described as a missed alarm.
Sensitivity Sensitivity is the probability that the event took place and the system
did fire. Sensitivity is equal to 1 − FN.
In the next section we will describe the main characteristics of the implemented
simulation model.
5.2 Simulation Model
Again for the setup of the simulation model we make use of the NetLogo environ-
ment. The "virtual world" is now coupled of passive agents named the Elderly
persons. Each of these agents is associated with one or two monitoring devices,
5.2 Simulation Model
45
depending on the simulation scenario. The Elderly agents are positioned in a static
position in the simulation area, as an assumption of their living area (e.g. home).
There is one type of event that occurs, namely the fall of Elderly agents or the "fall
event". The two monitoring devices, (e.g. accelerometer, gyroscope) are in charge
of identifying such events, and raising alarms towards the Community Agents. How-
ever, these devices aren't always accurate identifying the "fall event". The alarms
are collected by the first level Community agent or the so-called "Middleware 1",
which then forwards them to the next level agent, namely the "Middleware 2".
"Middleware 2" is responsible for allocation of available Informal Carers for visit-
ing the Elderly persons whose alarm has been triggered. The Informal Carers are
agents that wander around the simulation area until they receive messages from the
Community agent that asks them to go towards a certain location, where an Elder
person is residing. Once at the Elder agent's site, the Informal carer verifies if the
triggered alarm was true.
At the same time the alarm message is forwarded to the "Middleware 3" community
agent, which represents the Hospital coupled of moving agents or Ambulances, and
Professional carers or Doctors. In cases where the triggered alarm is True, the inter-
vention of Ambulances is required to transport the Elderly person to the hospital,
for receiving the necessary treatment.
In what follows a formal definition of the
above described agents is given.
Elderly Agents. Elderly Agents (EA) are agents representing elderly or impaired
persons residing in their house where their condition is monitored by Device
Agents. In case of true fall detection the EA take the role of patient, thus
professional care is requested.
Professional Carers. Professional Carers (PC) represent agents able to supply
certified healthcare services to Elderly agents. Such agents are institutional,
and provide professional service (e.g the doctors of a hospital).
Informal Carers and Verification. Informal Carers (IC) are mobile agents ca-
pable of providing non-professional services. It is assumed that IC can pay a
visit to an EA and report whether the EA has truly experienced a fall or not.
This is called in what follows verification.
Device Agents. Device Agents (DA) are simple devices (e.g. accelerometers, mo-
5.2 Simulation Model
46
tion sensors, cameras etc.) that are able to provide monitoring services. DA's
are attached to EA agents, thus monitor their condition. A DA triggers an
alarm when it ascertains, with a certain probability, that an EA has fallen.
Another example of DA is a motion sensor stating that an observed EA is not
moving.
DA are characterized by a non-zero probability of false positives and false neg-
atives. Thus for instance in some cases an accelerometer may detect a fall
when this is not true, while in some other cases the accelerometer may not
detect a true case of fall.
Mobility Agents. Mobility Agents (MA) are agents able to provide mobility or
transportation services to other agents. In our case such agents are represented
by ambulances.
Community Agents. Community Agents (CA) are agents managing a "circle" of
other agents. A detailed description of such agents was provided in Chapter
3.
A visualization of the simulation area containing the initialized agents can be seen
in Figure 5.1. The green "person" shape represents the EA agents, while the con-
nected devices represent the DAs. The ICs can be seen close to their cars, used for
transportation. The white building represents the hospital, and connected to it we
have the MAs and PCs.
In terms of the FSO approach, the core unit of the given organizational structure is
the iSoC coupled of an EA, and corresponding DAs. Further all iSoCs form a Service-
oriented Community (SoC) coordinated by the Level 1 CA. The IC agents reside
within the Level 2 CA forming another SoC, similarly to the Level 3 CA coupled of
MA and PC. The agents associated with a CA may send it notifications. Whenever
there is a need for communication between the CAs of distinct levels, the "exception"
messages are triggered. It is assumed that these messages are transmitted reliably
and instantaneously. The overall FSO structure is depicted in Figure 5.2
In order to quantify the results of the various simulation scenarios described in the
upcoming section, we introduce the following metrics:
Social Cost. We define Social Cost (SC) as the number of cycles that a social
5.2 Simulation Model
47
Figure 5.1: NetLogo screenshot of the simulation area with initialized agents.
resource is used to intervene either for a true alarm or for a false positive
(FP).
Cumulative Social Cost. Cumulative Social Cost (CSC) is defined as the overall
number of cycles used by a community to deal with true alarms and FP's
throughout a (simulated) time interval T .
Waiting Time. Waiting Time (WT) is the number of cycles elapsed from the mo-
ment an alarm is triggered, being it true or false.
Cumulative Waiting Time. (CWT) is the overall number of cycles elapsed from
the moment an alarm is triggered, being it true or false. It is the sum of all
the individual WT's occurred throughout a simulation run.
5.3 Simulation Scenarios
48
Figure 5.2: A representation of the FSO structure of the simulation model.
5.3 Simulation Scenarios
Based on the participation of agents on the simulation, we distinguish three main
scenarios:
S1 : Only accelerometers. In this scenario we use only one DA to monitor the state
of EA. In addition, there are no ICs defined meaning that there is no Level 2
CA. This represents the simplest scenario, where once an alarm is triggered,
the message is directly transmitted to the Level 3 CA or the hospital. The CA
is responsible for allocating the MAs to pay a visit to the EA. In case treatment
needed the EA is transported to the hospital, where a PC is assigned to handle
the case at hand. We measure FP rate, FN rate, Sensitivity, Specificity, CSC
and NCSC during a given T . CSC refers to total number of cycles in which
MAs are serving the EAs.
S2 : Accelerometers combined with a second sensor (camera or gyroscope [19]). This
scenario is the same as S1, with a difference in the DA number. Here we use two
DAs, instead of one. This affects the FP, FN rate, Sensitivity and Specificity.
We measure the same as in S1.
S3 : One or two DAs and Informal carers [21, 26]. In these scenarios the complete
FSO structure as described earlier is used. S3 with one DA represents the S1
with added ICs, while S3 with two DAs is the S2 with added ICs. We measure
as in S1 and S2, with various numbers of ICs. Here as described in the results
section we observe a better usage of Social resources.
5.3 Simulation Scenarios
49
We consider the following configuration of agents for the given scenarios:
• 30 EA's residing in their houses. House locations are assigned in cells chosen
pseudo-randomly in the virtual world.
• In scenario S1, 1 DA (an accelerometer) is deployed with each EA. In scenario
S2, an additional DA (e.g., a gyroscope) is added for each EA.
• 1 hospital (Level 3 CA), located in a cell chosen as described above.
• 6 PC's and 5 MA's (ambulances) are located with the hospital.
Scenario S3 extends scenarios S1 and S2 with the addition of IC's. We perform
simulations with an increasing step of 5 IC's from 0 to 40. The IC's are initially
assigned to pseudo-randomly chosen cells and then wander randomly within the
boundaries of the virtual world. The IC's receive requests for verification from
Level 2 CA. Once this happens, they quit pseudo-random wandering and move from
their current position to the verification place.
The major steps in the general scheme of execution for the three scenarios are
represented as pseudo-code in Tables 5.1, 5.2, 5.3, 5.4, and 5.5.
Table 5.1 shows the pseudo-code of the procedure of DA agents.
It represents a
Simple "reflex": if a fall is suspected for the corresponding EA, an alarm is triggered
towards the local CA. The alarm can be raised only when the EA is residing home.
The only moment when EAs might not be there are the cases when they had a
Procedure DA// Code of the DA agents.
// Reaction to false or true alarms.
begin
When ((this.Fired == TRUE) and (EA.NonFallingPeriod == 0) )
this.Send(CA, ALARM, this);
EndWhen
end
Table 5.1: Pseudo-code of the DA agents.
5.3 Simulation Scenarios
50
Procedure L1-CA // Code of the Level 1 CA, reaction to alarm message
begin
When (∃d ∈ DA (this.AlarmReceived(d) == TRUE))
Parallel
// In parallel, raise exceptions to Level2-CA and Level3-CA
Do exceptionEAs[] ← this.Exception(d, ALARM);
this.Send(exceptionEAs, L2-CA);
this.Send(exceptionEAs, L3-CA); EndDo; EndParallel
EndWhen
end
Procedure L2-CA // Code of the Level 2 CA, allocation of IC agents.
begin
When (∃i ∈ IC (i.getAvailable() == TRUE))
// Assign IC to visit EA for which alarm is raised
forEach d in exceptionEAs;
Do this.assign(i, d); i.setAvailable(FALSE); EndDo
EndForEach
EndWhen
end
Procedure L3-CA // Code of the Level 3 CA, allocation of MA and PC agents.
begin
When (∃p ∈ PC (p.getAvailable() == TRUE)
&& ∃m ∈ MA (m.getAvailable() == TRUE) )
// Assign MA to visit EA, and a PC in case intervention needed
forEach d in exceptionEAs;
Do this.assign(p, m);
p.setAvailable(FALSE); m.setAvailable(FALSE);EndDo
EndForEach
EndWhen
end
Table 5.2: Pseudo-code of the CA agents.
5.3 Simulation Scenarios
51
treatment earlier in hospital and are on their way back. Thus, the Non-falling
period is used.
The pseudo-code of the CA agents is presented in Table 5.2. It can be observed
that procedure L1-CA handles the alarms received from DAs. Further it redirects
them as exceptions to Level 2 and Level 3 CA. The L2-CA procedure allocates the
available ICs to perform the verification, while L3-CA allocates the MA and PC
agents based on the received "exception" messages.
The IC's wander aimlessly through the virtual world until a verification request is
received. The request is served, possibly canceling an ongoing emergency call. The
IC procedure can be seen in Table 5.3.
Procedure IC // Code of the IC agents.
// Reaction to alarm received by L2-CA.
begin
When (this.AlarmReceived(L2-CA, d) == TRUE))
Do
// The IC agent goes to the cell the alarm came from
this.GoTo(location(d));
// Once there, it verifies whether
// it was a case of FP (false positive)
If (this.Verification() == FALSE)
// If that is the case, it informs L3-CA to cancel the
// emergency call, and free the PC and MA resources,
// thus reducing the waste in social energy
this.Send(L3-CA, CANCEL_ALARM, d); EndIf
EndDo
// if no alarm is received, the IC agent just wanders pseudo-randomly
Otherwise this.Wander();
EndWhen
end
Table 5.3: Pseudo-code of the IC agents.
5.3 Simulation Scenarios
52
The MA agents wait for emergency calls. When one such call arrives, the MA agents
(ambulances) start moving towards the EA for which the alarm was raised. Reaching
that location is a time- and resource-consuming action. Said time and resources are
wasted in the face of a FP event. In the case of Scenario S3 the loss is reduced if
an IC agent reaches the location earlier than the MA and verifies the occurrence of
an FP. In such a case, the IC cancels the call, thus sparing some social energy. The
pseudo-code of MA agents can be seen in Table 5.4.
Finally the PC procedure is presented in Table 5.5. The PC's check for the arrival
of the MA carrying the EA agent. Once there they start the treatment. When the
treatment process finishes the EA's are freed, and PC's can be assigned to another
case.
Procedure MA// Code of the MA agents.
// Reaction to alarm received by the L3-CA.
begin
When (this.AlarmReceived(L3-CA, d) == TRUE)
Do
this.GoTo(location(d));
// Abort the emergency call if some IC canceled the alarm
While (this.AlarmCanceled(IC, d) == FALSE);
// Once at the location of the alarm,
// verify the situation, if needed bring EA to hospital
If (this.Verification() == FALSE)
this.Send(L3-CA, CANCEL_ALARM, d);
L3-CA.Send(L2-CA, CANCEL_ALARM, d);
Else
this.Bring(EA, location(L3-CA)); EndIf
EndWhile
EndDo
EndWhen
end
Table 5.4: Pseudo-code of the MA agents.
5.3 Simulation Scenarios
53
Each scenario is executed for 10000 simulation cycles (ticks). Every EA experiences
a "true fall" with probability 1
600, while the DA agents trigger false positive alarms
with probability 1
500, and false negatives with probability 100
500.
We now describe the results obtained in the three above-sketched scenarios.
Procedure PC // Code of the PC agents.
// Treatment of EA, once in hospital premises.
begin
When (MA.arrived(d) == TRUE))
Do
// The PC agent starts treating the EA
this.startTreatment(d);
// Once the, EA treatment finishes, it is free to go
When (d.treated() == TRUE))
// a non-falling period c is assigned to EA agent
d.setNonFallingPeriod(c);
EndWhen
EndDo
EndWhen
end
Table 5.5: Pseudo-code of the PC agents.
5.4 Results
5.4 Results
54
We start by running the S1 and S2 scenarios without any IC, then for each scenario
we run a number of experiments with various number of ICs starting from 0 to 40
(S3 scenario). As mentioned previously there are various measurement metrics used
for evaluation. The results of the S1 scenarios are shown in Table 5.6, while for S2
in Table 5.7.
We start by inspecting the FP, FN, Sensitivity, and Specificity parameters. If we
compare scenario S1 to S2 without ICs, we can observe that in S2 the number of
FPs increases for more than 100 FP units. This due to the addition of a second
DA. When there are two types of DA (sensors) per EA defined, the alarm or FP
is triggered as a result of an OR operation of the alarm of each DA. Even, if the
alarm triggered is true in only one of the sensors we will have a FP. Thus, 3 out of 4
possible combinations will be positive, resulting in a larger number of FPs. Further,
by adding the IC agents in the S3 scenarios we see that FP number increases. This
because the IC agents confirm the FP alarms in a faster pace. Thus, for the same
simulation period of 10000 "ticks" more FP requests are handled (see Figure 5.3).
For S2 with ICs the number of FPs fluctuates in the 620 range, while for S1 with
ICs around 320. The FN values in S1 are almost twice of those in S2. This because
in S2 an AND operation is performed between the FN outcome of DA1 and DA2.
As a result the probability of having a FN decreases in S2, and only when we have a
FN from both alarms the FN is reported. Comparing these two scenarios with their
corresponding S3 cases we see that the addition of ICs doesn't have any effect. This
because ICs cannot identify FNs. The graph in Figure 5.4 depicts the FP ratios of
S1, and S2 with various number of ICs. The FN ratios are visualized in Figure 5.5.
A larger number of FPs and a smaller number of FNs, provides with a more Sensitive
system. This because by triggering more alarms, more cases where the alarm is truly
positive are covered. On the other hand there is no difference observed with regard
to the Specificity parameter for each scenario. From the results in Table 5.6 and
Table 5.7 we see that for S1 with or without volunteers the average Sensitivity is
79.55%, while for S2 it increases to an average of 89.92%. The Specificity values
stay the same for all scenarios with an average of around 99%.
We proceed the evaluation with a comparison of the Social costs for S1 and S2,
5.4 Results
55
Figure 5.3: Number of alarm requests handled by the system for S1 and S2 with a
number of IC's ranging from 0 to 40.
while we continue increasing the number of IC's. Comparing S1 without any IC,
with the scenarios where the IC's are added we observe a significant decrease in the
cost of MAs. This because the ICs help verifying the condition of the EA agents,
thus reducing the cost of MAs. MAs have to intervene only when there has been
a true fall of the EA. The same applies for the S2 scenario. Figure 5.6 depicts the
In the case
relation of average MA cost with the IC number for both scenarios.
of S1, the minimum is reached for 20 IC's after which average social cost appears
to stabilize. Another metric that shows the impact of ICs in the MA cost is the
number of alarm verification's. In all scenarios where IC are present the number of
verification's performed by MA is reduced drastically.
An important metric is the average time until a triggered alarm for an EA gets
verified, being it true or false. The average waiting time reflects the reaction time
of the system for a given event. The best results are obtained with a number of 10
ICs. Additional effort beyond this value produces little improvements. With IC =
10 no significant differences are observed between S1 and S2. A graph showing the
average waiting time in report to IC number is given in Figure 5.7.
All in all, we can observe that the involvement of IC agents in the fall detection
5.4 Results
56
Figure 5.4: FP ratio for S1 and S2 with a number of IC's ranging from 0 to 40.
system using the FSO approach, helps reducing the MA social costs, while at the
same time the response time to fall events is fastened.
Figure 5.5: FN ratio for S1 and S2 with a number of IC's ranging from 0 to 40.
5.4 Results
57
Figure 5.6: Average social costs of S1 and S2 with a number of IC's ranging from 0
to 40.
Figure 5.7: Average waiting times with S1 and S2 and a number of IC's ranging
from 0 to 40.
10.47192982
11.24570447
12.09023941
14.09893993
14.62110727
17.43339254
21.44165171
31.91917293
Avg.WT118.6578947
42.28421053
42.30068729
43.35359116
42.99116608
41.43771626
44.51509769
45.2962298
48.73684211
68.25910931
Avg.MAcost
230
0
568
570
5969
5400
24102
99.79
79.86
235
0
582
582
6545
5963
24619
99.78
83.21
231
1
540
543
6565
6023
23541
99.8
81.69
222
0
566
566
7980
7414
24333
99.79
76.71
220
0
576
578
8451
7874
23951
99.79
74.91
227
6
555
563
9815
9252
25062
99.79
83.21
225
8
540
557
11943
11351
25230
99.8
81.65
211
33
472
532
16981
0
25928
99.81
77.09
193
296
0
494
58617
0
33720
99.79
77.68
MAinterventions
MAverifications
ICverifications
Reqs.Handled
CWT
CSC(volunteers)
CSC(ambulances)
Specificity
Sensivity
0.000347549
0.000294217
0.000320124
0.000405769
0.000433369
0.000275393
0.000308676
0.000367909
0.000379999
FNratio
0.5964
0.0058
0.034
166825
230
58
340
0.591
0.0048
0.0344
163097
238
48
344
0.5727
0.0052
0.0311
162385
232
52
311
0.6042
0.0068
0.0342
167515
224
68
342
0.6176
0.0074
0.0357
170681
221
74
357
0.595
0.0046
0.0335
166988
228
46
335
0.5924
0.0051
0.033
165171
227
51
330
0.6015
0.0063
0.032
171175
212
63
320
0.6052
0.0056
0.0299
147313
195
56
299
FPratio
Avg.FN/tick
Avg.FP/tick
TNnumber
TPnumber
FNnumber
FPnumber
S1-40ICs
S1-35ICs
S1-30ICs
S1-25ICs
S1-20ICs
S1-15ICs
S1-10ICs
S1-5ICs
S1
Table5.6:ResultsofS1scenarios
10.51679307
10.62693157
12.11886587
13.61702128
15.2124183
17.18848758
21.67351598
46.97374702
Avg.WT203.5503356
27.21126761
27.9205298
28.38167939
28.92721165
28.38126362
30.28103837
30.98287671
36.09665871
65.13255034
Avg.MAcost
232
0
922
923
9707
8785
25116
99.59
89.51
232
1
904
906
9628
8723
25296
99.59
89.47
239
1
913
917
11113
10199
26026
99.58
89.19
226
3
887
893
12160
11269
25832
99.6
89.51
237
2
913
918
13965
13048
26054
99.58
91.72
235
5
877
886
15229
14344
26829
99.59
89.86
228
37
824
876
18986
18114
27141
99.59
89.41
219
67
745
838
39364
33426
30249
99.58
90.73
161
420
0
596
MAinterventions
MAverifications
ICverifications
Reqs.Handled
121316
CWT
0
38819
99.59
89.89
CSC(volunteers)
CSC(ambulances)
Specificity
Sensivity
0.000185062
0.000185191
0.000193582
0.000170334
0.000145634
0.000186859
0.000184776
0.000164411
0.000192829
FNratio
0.7226
0.003
0.0667
162078
256
30
667
0.7185
0.003
0.0651
161965
255
30
651
0.7208
0.0031
0.0661
160108
256
31
661
0.7323
0.0028
0.0654
164355
239
28
654
0.7222
0.0023
0.0663
157907
255
23
663
0.7099
0.0029
0.0629
155168
257
29
629
0.7203
0.0029
0.0631
156918
245
29
631
0.7195
0.0024
0.0604
145952
235
24
603
0.7013
0.002
0.0418
103699
178
20
418
FPratio
Avg.FN/tick
Avg.FP/tick
TNnumber
TPnumber
FNnumber
FPnumber
S2-40ICs
S2-35ICs
S2-30ICs
S2-25ICs
S2-20ICs
S2-15ICs
S2-10ICs
S2-5ICs
S2
Table5.7:ResultsofS2scenarios
Conclusions
Evolution is a process that has constantly accompanied the humanity. The same
should apply for the traditional organizations. Evolving the traditional organizations
while maintaining the identity of the intended services represents a great challenge
for the mankind. With an ever increasing number of population and with scarcer
natural resources a more intelligent way of management and organization is a must.
The time to explore smarter and more efficient ways of organization is now.
In this thesis we presented two simulation models and preliminary results using the
FSO concepts as a potential solution to the given problems. It was indicated that
FSO provides us with an example of the social-energy-aware solutions we referred to
above. The provided simulation models have shown that FSO's dynamic hierarchical
organization optimally orchestrates all participating entities by the use of exceptions
mechanism, thus overcoming the stiffness of the traditional organizations. By the
use of this interconnecting structure of roles of various natures a significantly im-
proved agility is reached. The inclusion of intelligent mobile agents offers enormous
collaboration opportunities and resource sharing, thus leading to systems able to
tap into the great wells of social energy of our societies.
In the first simulation model we showed that by the use of FSO properties the in-
dividuals may benefit by receiving more qualitative healthcare services. This is due
to the increase of accuracy of service provider identification. At the same time the
service receiving times improved due to the inter-connectivity of organizations, al-
lowing for fast flow of information. By providing a service that involves entities of
various organizations a better usage of resources is archived, thus exploiting more
of the social energy. The diversity of FSO applications is shown by the number of
examples where such approach might have an application. In the second simulation
model we showed that by the usage of FSO mechanism fall detection systems can
be improved. Depending on the specificity and sensitivity of devices as well as on
60
5.4 Results
61
the willingness and availability of human beings, we showed how it is possible to
improve the social costs, make better use of the social resources, and reduce the
average time to respond to identified falls.
In the next section we describe the future work.
Future work
The subjects where FSO approach can get an application are numerous. However,
with a limited amount of time, choices for the subjects to be handled had to be
made. As the work progressed, several interesting ideas about various simulation
models came up. Some of them we did implement, but for the others a substantial
amount of effort and time was required thus they were left out.
An improvement to the current simulation models could be the addition of more
detailed characteristics to the agents, which would help mimicking the real life sce-
narios in a more accurate way. By having more information about the underlying
agents the community representatives could be improved to provide with semantic
analysis when dealing with given situations. Further the emergence of novel and pos-
sibly unexpected service modes, where two or more primary agents (elderly people)
fulfill each others requirements thus creating a self-serve coupling could be explored.
This would allow for analyzing the emergence of social self-service mechanisms and
their social impact.
As a future work the existing models could be complemented with protocols, algo-
rithms and components that allow for a practical implementation of the Service-
oriented Communities and Fractal Social Organizations. The idea is switching from
the current abstract models, to concrete software artifacts. A possible solution could
be the implementation of a middleware based on Web Services to manage intra- and
inter-SoC cooperation.
The topic is very broad and future research can take many directions.
62
Bibliography
[1] D. A. Bray, K. Croxson, W. H. Dutton, and B. Konsynski. Sermo: A community-
based, knowledge ecosystem. In: Oll distributed problem-solving networks con-
ference, 2008 (February).
[2] H. Zhuge. Cyber physical society. Pages 18 of: Semantics knowledge and grid
(skg), 2010.
[3] V. De Florio. Reflections on organization, emergence, and control in sociotech-
nical systems. Lexington Robert, MacDougall [Online]. Available: https://
arxiv.org/abs/1412.6965 (2015)
[4] Y. Kim, A. Chen, E. Keith, H. F. Yee Jr, M. Kushel. Not perfect, but better:
Primary care providers' experiences with electronic referrals in a safety net health
system. Journal of General Internal Medicine 24(5) (2009) 614-619
[5] L. Shaw, D. de Berker. Strengths and weaknesses of electronic referral: com-
parison of data content and clinical value of electronic and paper referrals in
dermatology. British Journal of General Practitioners 57 (2007) 223-224
[6] Anonymous:
Secured health information network and exchange (SHINE)
[Online]. Available:
http://www.gsma.com/connectedliving/wp-content/
uploads/2012/05/Smart-mHealth.pdf (2012)
[7] Anonymous: SHINE OS+ [Online]. Available: http://www.shine.ph (2015)
[8] V. De Florio, M. Bakhouya, A. Coronato, and G. Di Marzo Serugendo, Mod-
els and Concepts for Socio-technical Complex Systems: Towards Fractal Social
Organizations, Systems Research and Behavioral Science, vol. 30, no. 6, 2013.
63
64
[9] V. De Florio, and C. Blondia Service-oriented communities: Visions and con-
tributions towards social organizations. Lecture Notes in Computer Science, vol.
6428. Springer Berlin / Heidelberg. 10.1007/978-3-642-16961-8 51, 2010.
[10] V. De Florio, H. Sun, J. Buys, and C. Blondia: On the impact of fractal or-
ganization on the performance of socio-technical systems. In: Proc. of the 2013
Int.l Workshop on Intelligent Techniques for Ubiquitous Systems (ITUS 2013),
Vietri sul Mare, Italy, IEEE (2013) 1940: International Secretariat, Institute of
Pacific Relations.
[11] H. Sun, V. De Florio, C. Blondia Implementing a role based mutual assistance
community with semantic service description and matching. In: Proc. of the Int.l
Conference on Management of Emergent Digital EcoSystems (MEDES) 2013.
[12] V. De Florio, H. Sun, and M. Bakhouya, Mutualistic relationships in service-
oriented communities and fractal social organizations, CoRR, vol. abs/1408.7016,
2014. [Online]. Available: http://arxiv.org/abs/1408.7016
[13] U. Wilensky, NetLogo [computer software], Available online: http://ccl.
northwestern.edu/netlogo/ 1999.
[14] K. Ryu. Fractal-based reference model for self-reconfigurable manufacturing sys-
tems. Ph.D. thesis, Pohang University of Science and Technology, Korea. August
2003.
[15] Spica, G.: Catania, ospedali pieni: muore neonata. La Repubblica [Online].
Available:
http://palermo.repubblica.it/cronaca/2015/02/12/news/
manca_il_posto_in_rianimazione_negli_ospedali_di_catania_neonata_
muore_in_ambulanza-107141635/?ref=HREA-1 16 February 2015
[16] Ballabio, L.: Giussano, ha un infarto in ospedale:
il pronto soc-
corso
e
chiuso
e
lui muore
cercando
aiuto.
Il Giorno.
[Online].
Available:
http://www.ilgiorno.it/monza/cronaca/2013/11/29/
989314-infarto-morto-prontosoccorso-chiuso.shtml 29 November 2013
[17] H. Warnecke and M. Huser, The fractal company: a revolution in corporate
culture. Springer, 1993.
65
[18] V. De Florio, A. Coronato, M. Bakhouya, and G. Di Marzo Serugendo, Service-
oriented communities: Models and concepts towards fractal social organizations.
In: Proceedings of the 8th International Conference on signal, image technology
and internet based systems (SITIS 2012), IEEE (2012).
[19] Q. T. Huynh, U. D. Nguyen, S. V. Tran, A. Nabili, and B. Q. Tran, Fall
detection system using combination accelerometer and gyroscope in Proc. of the
Second Int.l Conf. on Advances in Electronic Devices and Circuits (EDC 2013),
2013.
[20] N. El-Bendary, Q. Tan, F. C. Pivot, and A. Lam, Fall detection and prevention
for the elderly: a review of trends and challenges, International Journal on Smart
Sensing and Intelligent Systems, vol. 6, no. 3, 2013.
[21] H. Sun, V. De Florio, N. Gui, and C. Blondia, The missing ones: Key in-
gredients towards effective ambient assisted living systems, Journal of Ambient
Intelligence and Smart Environments, vol. 2, no. 2, April 2010.
[22] D. Cusano and T. Doris. Accelerometers, false positives/negatives and fall de-
tection, September 2014.
[23] F. Sposaro , G. Tyson iFall: an Android application for fall monitoring and
response. In Proceedings of the Annual International Conference of the IEEE
Engineering in Medicine and Biology Society. Minneapolis: Institute of Electrical
and Electronics Engineers; 2009:6119-6122.
[24] Dai J, Bai X, Yang Z, Shen Z, Xuan D. Mobile phone-based pervasive fall de-
tection. Pers Ubiquitous Comput 2010, 14:633-643.
[25] Abbate S, Avvenuti M, Bonatesta F, Cola G, Corsini P, Vecchio A. A
smartphone-based fall detection system. Pervasive Mob Comput 2012, 8:883-899.
[26] V. De Florio, H. Sun, and C. Blondia, Community resilience engineering: Re-
flections and preliminary contributions, in Software Engineering for Resilient
Systems, ser. Lecture Notes in Computer Science, I. Majzik and M. Vieira, Eds.
Springer International Publishing, 2014, vol. 8785, pp. 18. [Online]. Available:
http://dx.doi.org/10.1007/978-3-319-12241-01
Appendix A
Exemplary FSO structural organi-
zation
66
FSO-structure.txt
Emergency response members
firefighter's_coordinator 37
people's_coordinator 36
localhospital 8
localhospital 7
People's coordinator members
[taxi 160 taxidriver 154]
[taxi 164 taxidriver 157]
[taxi 165 taxidriver 152]
[taxi 170 taxidriver 151]
[taxi 171 taxidriver 159]
[taxi 172 taxidriver 160]
[taxi 173 taxidriver 158]
[taxi 159 taxidriver 155]
[individual 144]
[individual 139 car 151]
[individual 140]
[individual 142]
[individual 143 car 149]
[individual 138 car 148]
[individual 103 car 199]
[individual 131 car 101]
[individual 161 car 131]
[individual 120 car 166]
[individual 147 car 150]
[individual 146]
[house 116 fire-detector 60]
[house 133 fire-detector 56]
[house 96 fire-detector 78]
[house 104 fire-detector 44]
Firefighters's coordinator members
[ftruck 166 firefighter 172 firefighter 169]
[ftruck 167 firefighter 171]
[ftruck 168 firefighter 170 firefighter 173 ]
localhospital 8 members
medicalappliance 27
medicalappliance 25
medicalappliance 31
medicalappliance 33
medicalappliance 28
medicalappliance 26
67
L2
L1
L1
L1
L1
L1
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L0
L1
L0
L0
L0
L1
L0
L0
L0
L0
L0
L0
doctor 20
doctor 24
transportvan 16
transportvan 13
transportvan 12
transportvan 14
ambulance 10
ambulance 9
ambulance 11
localhospital 7 members
medicalappliance 34
medicalappliance 30
medicalappliance 29
doctor 22
doctor 23
transportvan 17
transportvan 19
68
L0
L0
L0
L0
L0
L0
L0
L0
L0
L1
L0
L0
L0
L0
L0
L0
L0
|
1509.00737 | 1 | 1509 | 2015-09-02T15:16:42 | A Game-theoretic Formulation of the Homogeneous Self-Reconfiguration Problem | [
"cs.MA",
"cs.GT",
"eess.SY"
] | In this paper we formulate the homogeneous two- and three-dimensional self-reconfiguration problem over discrete grids as a constrained potential game. We develop a game-theoretic learning algorithm based on the Metropolis-Hastings algorithm that solves the self-reconfiguration problem in a globally optimal fashion. Both a centralized and a fully distributed algorithm are presented and we show that the only stochastically stable state is the potential function maximizer, i.e. the desired target configuration. These algorithms compute transition probabilities in such a way that even though each agent acts in a self-interested way, the overall collective goal of self-reconfiguration is achieved. Simulation results confirm the feasibility of our approach and show convergence to desired target configurations. | cs.MA | cs | A Game-theoretic Formulation of the Homogeneous
Self-Reconfiguration Problem
Daniel Pickem1, Magnus Egerstedt2, and Jeff S. Shamma3
5
1
0
2
p
e
S
2
]
A
M
.
s
c
[
1
v
7
3
7
0
0
.
9
0
5
1
:
v
i
X
r
a
Abstract -- In this paper we formulate the homogeneous
two- and three-dimensional self-reconfiguration problem over
discrete grids as a constrained potential game. We develop a
game-theoretic learning algorithm based on the Metropolis-
Hastings algorithm that solves the self-reconfiguration problem
in a globally optimal fashion. Both a centralized and a fully
distributed algorithm are presented and we show that the only
stochastically stable state is the potential function maximizer,
i.e. the desired target configuration. These algorithms compute
transition probabilities in such a way that even though each
agent acts in a self-interested way, the overall collective goal of
self-reconfiguration is achieved. Simulation results confirm the
feasibility of our approach and show convergence to desired
target configurations.
I. INTRODUCTION
Self-reconfigurable systems are comprised of individual
agents which are able to connect to and disconnect from one
another to form larger functional structures. These individual
agents or modules can have distinct capabilities, shapes, or
sizes, in which case we call it a heterogeneous system (for
example [8]). Alternatively, modules can be identical and
interchangeable, which describes a homogeneous system (see
[11]). In this paper, we will present algorithms that reconfig-
ure homogeneous systems and treat self-reconfiguration as a
two- and three-dimensional coverage problem.
Self-reconfiguration is furthermore understood to solve the
following problem. Given an initial geometric arrangement
of cubes (called a configuration) CI and a desired target
configuration CT ,
the solution to the self-reconfiguration
problem is a sequence of primitive cube motions that re-
shapes/reconfigures the initial into the target configuration
(see Fig. 1). By configuration we mean a geometric arrange-
ment of a set of agents. The problem setup is then the
following.
• The environment E is a finite two- or three-dimensional
• N agents P = {1, 2, . . . , N} move in discrete steps
• Each agent has a restricted action set Ri which contains
discrete grid, i.e. E ⊆ Z2 or E ⊆ Z3.
through that grid.
only a subset of all its possible actions Ai.
This research was sponsored by AFOSR/MURI Project #FA9550-09-1-
0538 and ONR Project #N00014-09-1-0751.
1D. Pickem is a Ph.D. student in Robotics, Georgia Institute of Technol-
ogy, Atlanta, USA [email protected]
2M. Egerstedt is with the Faculty of Electrical and Computer Engineering,
Georgia Institute of Technology, Atlanta, USA [email protected]
3J. S. Shamma is with the School of Electrical and Computer
Technology, Atlanta, USA
and with King Abdullah University
Saudi Arabia
Engineering, Georgia
[email protected],
of
[email protected].
(KAUST), Thuwal,
and Technology
Institute
Science
of
• An agent's utility or reward Ui(a ∈ A) is inversely
proportional to the distance to the target configuration.
Many approaches to self-reconfiguration have been presented
in the literature, each with certain short-comings. There have
been centralized solutions (both in planning and execution,
e.g. [19]), distributed solutions that required a large amount
of communication (see [8]) or precomputation (see [7], [17]),
approaches that were either focused on locomotion (e.g. [5])
or functional target shape assemblies (e.g. [12]). Distributed
approaches have often relied on precomputation of rulesets
(see [17]), policies (e.g. [7]), or entire sets of paths/folding
schemata of agents (e.g. [6]).
In this paper, we present a fully decentralized approach
to homogeneous self-reconfiguration for which no central
decision maker is required. Currently, each agent needs to
know the location and shape of the target configuration to
compute the utility function of its actions and choose an
action. However, no (in the two-dimensional case) or limited
communication (in the three-dimensional case) is required to
successfully complete a reconfiguration.
The rest of this paper is organized as follows. Section
II discusses relevant related work. Section III presents the
system setup and theoretical formulation of the problem.
In Section IV we discuss the completeness of deterministic
reconfiguration, which is used in Section V to prove the
existence of a unique potential function maximizer. In Sec-
tion VI we decentralize the stochastic algorithm and present
simulation results in Section VII. Section VIII concludes the
paper.
II. RELATED WORK
In this section we want to highlight decentralized ap-
proaches in the literature that bear resemblance to methods
presented in this paper. Especially relevant to the presented
results in this paper are homogeneous self-reconfiguration
approaches such as [5], and [11], which employ cellular
automata and manually designed local rules to model the
system. Similar work in [7] shows an approach for locomo-
tion through self-reconfiguration represented as a Markov
decision process with state-action pairs computed in a dis-
tributed fashion. The presented algorithms are decentralized
but only applicable to locomotion and not the assembly of
arbitrary configurations.
Precomputed rules are also used in [17], in which graph
grammars for self-reconfiguration are automatically gener-
ated. Similarly, in [9] a decentralized approach is presented
based on graph grammatical rules, in which automatically
generated graph grammars are used to assemble arbitrary
Fig. 1: Example of self-reconfiguration sequence from a random 2D configuration (left) to another random 2D configuration
(right). Approximately every 20th time step is shown.
acyclic target configurations. Whereas these approaches are
able to assemble arbitrary target configurations, they rely on
precomputing rulesets for every target configuration.
The algorithms presented in this paper are inspired by
the coverage control literature, specifically game-theoretic
formulations such as [1], [13], and [21]. Both static sensor
coverage as well as dynamic coverage control with mobile
agents are discussed. Note, however, that agents in these
papers are limited to movement
in two dimensions and
operate with a different motion model and most importantly
different constraints.
We address these problems with a decentralized algorithm
that does not rely on precomputed rulesets, can be used for
locomotion as well as the assembly of arbitrary two- and
three-dimensional shapes, can handle changing environment
conditions as well as changing target configurations, and is
scalable to large number of modules due to its decentralized
nature.
III. PROBLEM FORMULATION
In this work, we represent agents as cubic modules that
move through a discrete lattice or environment E = Zd
in discrete steps.1 Without loss of generality these cubes
have unit dimension. Therefore, an agent's current state
or action ai (in a game theoretic sense) is an element
ai ∈ Zd. Note that an agent's action is equivalent to its
position in the lattice. Cubes can be thought of as geometric
embeddings of the agents in our system. A collection of
agents is furthermore called a configuration. Therefore, a
configuration C composed of N agents is a subset of the
representable space ZdN (see [18]). Moreover, we will deal
with homogeneous configurations, in which all agents have
the same properties and are completely interchangeable.
1Since we present
throughout this paper, the dimensionality d will be d ∈ {2, 3}.
two- and three-dimensional self-reconfiguration,
A. Motion Model
In the sliding cube model (see [4], [5], [17], [18], [19]), a
cube is able to perform two primitive motions - a sliding and
a corner motion. In general, a motion specifies a translation
along coordinate axes and is represented by an element m ∈
Zd. A sliding motion is characterized by (cid:107)ms(cid:107)L1 = 1, i.e.
ms,i = 1 for one and only one coordinate i ∈ 1, . . . d, which
translates a cube along one coordinate axis. A corner motion
on the other hand is defined by (cid:107)mc(cid:107)L1 = 2 such that mc,i =
1 for exactly two coordinates i ∈ 1, . . . d, which translates a
cube along two dimensions.
B. Game-theoretic Formulation
In this section we formulate the homogeneous self-
reconfiguration as a potential game (see [16]), which is
a game structure amenable to globally optimal solutions.
Generally, a game is specified by a set of players i ∈
P = {1, 2, . . . , N}, a set of actions Ai for each player,
and a utility function Ui(a) = Ui(ai, a−i) for every player
i. In this notation, a denotes a joint action profile a =
(a1, a2, . . . , aN ) of all N players, and a−i is used to denote
the actions of all players other than agent i.
In a constrained game, the actions of agents are constrained
through their own and other agents' actions. In other words,
given an action set Ai for agent i, only a subset Ri(a)) ⊂ Ai
is available to agent i. A constrained potential game is
furthermore defined as follows.
Definition 1: A constrained exact potential game (see
[21]) is a tuple G = (P,A, {Ui(.)}i∈P ,{Ri(.)}i∈P , Φ(A)),
where
• P = {1, . . . , N} is the set of N players
• A = A1 × ··· × AN is the product set of all agents'
• Ui : A → R are the agents' individual utility functions
• Ri : A → 2Ai is a function that maps a joint action to
action sets Ai
a restricted action set for agent i
Additionally, the agents' utility functions are aligned with a
global objective function or potential Φ : A → R if for all
agents i ∈ P, all actions ai, a(cid:48)
i ∈ Ri(a), and actions of other
agents a−i ∈(cid:81)
j(cid:54)=i Aj the following is true
i, a−i) − Ui(ai, a−i) = Φ(a
i, a−i) − Φ(ai, a−i)
(cid:48)
(cid:48)
Ui(a
The last condition of Def. 1 implies an alignment of
agents' individual incentives and the global goal. Therefore,
under unilateral change (only agent i changes its action from
ai to a(cid:48)
i) the change in utility for agent i is equivalent to the
change in the global potential Φ. This is a highly desirable
property since the maximization of all agents' individual
utilities yields a maximum global potential. We can now
formulate the self-reconfiguration problem in game theoretic
terms and show that it is indeed a constrained potential game.
Definition 2: Game theoretic self-reconfiguration can be
formulated as a constrained potential game, where the indi-
vidual components are defined as follows:
• The set of players P = {1, 2, . . . , N} is the set of all
• The action set of each agent Ai = Zd is a set of discrete
N agents in the configuration.
• The utility function of each agent
lattice positions (or a finite or infinite subset of Zd).
is Ui(a) =
dist(ai,CT )+1. Here, CT is the target configuration and
dist(ai,CT ) = minaj∈CT (cid:107)ai − aj(cid:107).
(cid:88)
• The restricted action sets Ri(a ∈ A) are computed
• The global potential Φ(a ∈ A) =
according to Section III-C.
Ui(a).
1
i∈P
Note that the utility of an agent is independent of all other
agents' actions and depends exclusively on its distance to
the target configuration. An agent's action set, however, is
constrained by its own as well as other agents' actions. The
goal of the game theoretic self-reconfiguration problem is to
maximize the potential function, i.e.
a∈A Φ(a) = max
max
a∈A
Ui(a)
(cid:88)
This can be interpreted as a coverage problem where the
goal of all agents is to cover all positions in the target
configuration. Therefore maximizing the potential is equiv-
alent to maximizing the number of agents that cover target
positions ai ∈ CT . The following propositions shows that
this formulation indeed yields a potential game.
Proposition 1: The self-reconfiguration problem in Def.
2 constitutes a constrained potential game with Φ(a) =
Ui(a) and Ui(a) =
1
dist(ai,CT )+1.
i , a−i), where a0
i∈P
Proof: Let the agents' utility functions Ui(ai, a−i) =
Φ(ai, a−i) − Φ(a0
i denotes the null action
of agent i, which is equivalent to removing agent i from
the environment. Then an agent's utility is its marginal
contribution to coverage (Wonderful Life Utility, see [1]), or
in other words, the grid cells covered exclusively by agent i.
But since each agent covers exactly the grid cell it currently
(cid:88)
i∈P
occupies, the following holds.
Ui(ai, a−i) = Φ(ai, a−i) − Φ(a0
i , a−i)
(cid:88)
j∈P
Uj(a) − (cid:88)
j∈P\{i}
=
Therefore,
Uj(a) = Ui(a)
i, a−i) − Φ(ai, a−i)
i, a−i) − Ui(ai, a−i) = Φ(a
(cid:48)
(cid:48)
Ui(a
As we will see in Section VI, this potential game structure
allows us to derive a decentralized version of the presented
learning algorithm.
C. Action Set Computation
a) 2D reconfiguration:
A core component of constrained potential games is the
computation of restricted action sets. Unlike previous work
(see for example [14] and [21]), agents in our setup are
constrained not just by their own actions, but also those of
others. In this section we present methods for computing
restricted action sets such that agents comply with motion
constraints as well as constraints imposed by other agents.
In the two-dimensional case
agents are restricted to motions on the xy-plane. Unlike
in previous work (see [17] and [18]) where we required a
configuration to remain connected at all times, in this work,
agents are allowed to disconnect from all (or a subset of)
other agents. This approach enables agents to separate from
and merge with other agents at a later time. To formalize
this idea, we first review some graph theoretic concepts.
Definition 3: Let G = (V, E) be the graph composed of
N nodes with V = {v1, v2, . . . , vN}, where node vi repre-
sents agent or location i. Then G is called the connectivity
graph of configuration C if E = V × V with eij ∈ E if
(cid:107)ai − aj(cid:107) = 1.
This definition implies that two nodes vi, vj in the connectiv-
ity graph are adjacent, if agent or location i and j are located
in neighboring grid cells. Note that a connectivity graph can
be computed for any set of grid positions, whether these
positions are occupied by agents or not. We furthermore use
the notions of paths on graphs and graph connectivity in the
usual graph theoretic sense. Note that G is not necessarily
connected as (groups of) agents can split off. Therefore, G
generally consists of connected components Ci such that
G = {C1, C2, . . . , Cm}. Since the edge set E of G is time-
varying, the number m of connected components changes
with time as well. Based on the connectivity graph G and
the current joint action, we now define the function Ri :
A → 2Ai, which maps from the full joint action set to a
restricted action set for agent i and is based on the following
two definitions of sets of primitive actions.
i ∈ Zd \ a−i : (cid:107)ms(cid:107)L1 = 1(cid:9), where
motions is Ms = (cid:8)a(cid:48)
(cid:8)a(cid:48)
i ∈ Zd \ a−i : (cid:107)mc(cid:107)L1 = 2 ∧ mc,j ∈ {0, 1}(cid:9),
Definition 4: The set of all currently possible sliding
Definition 5: The
corner
possible
j ∈ [1, . . . , d] and mc = a(cid:48)
set
motions
i − ai.
currently
=
where
is Mc
ms = a(cid:48)
i − ai.
all
of
i
Note that Ms and Mc in Def. 4 and Def. 5 are equally
applicable to 2D and 3D. These definitions encode the
motion model outlined in Section III-A and allow us to define
the restricted action set in two dimensions as follows.
Definition 6: The two-dimensional restricted action set is
(a) = Ms ∪ Mc.
given by R2D
This definition ensures that agent i can only move to unoccu-
pied neighboring grid positions a(cid:48)
i through sliding or corner
motions (or stay at its current position ai - see Algorithm
1 and Algorithm 2). All other agents replay their current
actions a−i.
b) 3D reconfiguration: Whereas in the 2D case agents
were allowed to move to all unoccupied neighboring grid
cell regardless of connectivity constraints,
in the three-
dimensional case we introduce the requirement of ground-
edness. An agent is immobile, if executing an action would
remove groundedness from any of its neighbors. Grounded-
ness requires a notion of ground plane, which is defined as
follows.
Definition 7 (Ground Plane): The ground plane is the set
SGP = {s ∈ E : sz = 0} where E ⊆ Z3 and the
corresponding connectivity graph is GGP = (VGP , EGP )
with eij ∈ EGP if (cid:107)si − sj(cid:107) = 1.
Note that the ground plane is defined as the xy-plane and
its connectivity graph GGP is, by definition, connected.
Positions s ∈ SGP are not allowed to be occupied by
agents, therefore ai ∈ Ai \ SGP ∀i ∈ P. Using the graph
GGP , we define G(cid:48) = (V (cid:48), E(cid:48)) as V (cid:48) = V ∪ VGP and
E(cid:48) = V (cid:48) × V (cid:48) such that eij ∈ E(cid:48) if for vi, vj ∈ V (cid:48) we
have (cid:107)ai − aj(cid:107)L1
= 1. Note that G(cid:48) represents the current
configuration including the ground plane, and ai represents
an action of an agent or an unoccupied position in the ground
plane.
Definition 8 (Groundedness): An agent i is grounded if
there exists a path on G(cid:48) from vi ∈ V ⊂ V (cid:48) to some vk ∈
VGP ⊂ V (cid:48), where vi represents agent i in the connectivity
graph G (see Def. 3). A configuration C is grounded if every
agent i ∈ P is grounded.
The idea behind groundedness hints at an embedding of
a self-reconfigurable system in the physical world, where
agents cannot choose arbitrary positions in the environment
(e.g. float
in free space). More importantly, we use the
notion of groundednes to prove completeness of determin-
istic reconfiguration in Section IV and irreducibility of the
underlying Markov chain in Section V.
An agent can verify groundedness in a computationally cheap
way through a depth-first search, which is complete and
guaranteed to terminate in time proportional to O(N ) in a
finite space. The notion of groundedness also informs the
restricted action set computation. If all neighbors Ni =
{vj ∈ V : eij ∈ E} (adjacency according to G in Def.
3) of agent i can compute an alternate path to ground (other
than through agent i) then agent i is allowed to move. To
formalize this idea, let G−i = (V−i, E−i) with V−i =
V ∪ VGP \{vi} and E−i = V−i × V−i such that eij ∈ E−i if
for vi, vj ∈ V−i we have (cid:107)ai − aj(cid:107)L1
= 1. G−i is therefore
the connectivity graph of the current configuration including
C
c1
c2
C
c2
c1
c3
SGP
(a) A movement of agent c1 would
remove groundedness of the agent
c2.
SGP
(b) Agenet c1 can move with-
out breaking the groundedness con-
straint for the agents c2 and c3.
Fig. 2: Examples of grounded configurations and feasible
motions of cubes.
the ground plane without agent i. Subsequently, R3D
defined as follows.
i
(a) is
i
i
(a) = {ai}.
Definition 9: The three-dimensional restricted action set
(a) = Ms ∪ Mc if all agents vj ∈ Ni are grounded on
R3D
G−i. Otherwise, R3D
This definition encodes the same criteria as the two-
dimensional action set with the additional constraint of
maintaining groundedness (see Fig. 2). If agent i executing
an action would leave any of its neighbors ungrounded, agent
i is not allowed to move.
IV. DETERMINISTIC COMPLETENESS
Theorem 1 (Completeness of 2D reconfiguration):
In this section we establish completeness of deterministic
reconfiguration in two and three dimensions. We will show
that for any two configurations CI and CT there exists a
deterministically determined sequence of individual agent
actions such that configuration CI will be reconfigured into
CT . These results are required to show irreducibility of the
Markov chain induced by the learning algorithm outlined in
Section V. Irreducibility guarantees the existence of a unique
stationary distribution and furthermore a unique potential
function maximizer. We first show completeness of 2D
reconfiguration.
Any given two-dimensional configuration CI
can be
reconfigured into any other two-dimensional configuration
CT ,
there exists a finite sequence of configurations
{CI = C0,C1, . . . ,CM = CT} such that two consecutive
configurations differ only in one individual agent motion.
Proof: Without loss of generality, assume that CI and
CT do not overlap, i.e. for no ci ∈ CI is it true that also
ci ∈ CT . Additionally, assume that CI and CI are separated
along one dimension k ∈ {ex, ey, ez} (with ex, ey, ez being
the basis vectors of the lattice), i.e. ∀ci,k ∈ CI we have
that ci,k < cj,k, ∀cj ∈ CT . Then at each time step t,
select the agent i whose current position ci ∈ C is closest
to an unoccupied position cj ∈ CT . Plan a deterministic
i.e.
i → c1
i →
path of primitive agent motions pi = {ci = c0
··· → cm
i = cti} using a complete path planner such as A*.
Note that such a path always exists since we don't require
agents to remain connected to any other agents. Therefore,
the path planning problem is reduced to single agent path
planning on a discrete finite grid, which is complete because
A* is complete. This greedy selection process of the agent-
target pairs together with a complete path planning approach
suffices to reconfigure any two-dimensional configuration
into any other two-dimensional configuration (similar to
flood-fill algorithms).
The result in Theorem 1 holds for any configuration, even
configurations that consist of multiple connected compo-
nents. Before we can show a similar result for the 3D case
we need to introduce a graph theoretic result.
Lemma 1: According to Lemma 6 in [19], any finite graph
with at least two vertices contains at least two vertices which
are not articulation points.2
a
finite
of
exists
grounded
i , . . . , am
i , s(cid:107)L1
Proof: Without
3D configuration
loss of generality, assume that
i } such that for some s ∈ SGP (cid:107)am
reconfigured into a 2D configuration C2D
Int,
Theorem 2 (Completeness of 3D to 2D reconfiguration):
CG,3D can
Any finite
be
i.e.
sequence
there
configurations
Int} such that two consecutive
{CG,3D = C0,C1, . . . ,CM = C2D
configurations differ only in one individual agent motion.
the
connectivity graph of CG,3D consists of one connected
component. In any finite grounded 3D configuration, there
always exists an agent i ∈ P with a non-empty restricted
action set Ri(a) > 0.. Agent i is therefore mobile and there
exists a finite path of individual agent motions pi = {ai =
= 1,
a0
i , a1
i.e. the agent's final action is a position on the ground plane
SGP .
Let the subset of agents Pz>1 ⊂ P contain those agents
whose positions are not on the ground plane and Gz>1 =
(Vz>1, Ez>1) the corresponding connectivity graph. Further-
more, let G(cid:48)(cid:48) = (V (cid:48)(cid:48), E(cid:48)(cid:48)) be such that V (cid:48)(cid:48) = Vz>1 ∪{vGP},
where vGP is a single node representing all
the agents
on the ground plane and E(cid:48)(cid:48) such that eij ∈ E(cid:48)(cid:48)
if for
vi, vj ∈ V (cid:48)(cid:48) we have (cid:107)ai − aj(cid:107)L1
= 1. According to Def. 9,
an agent i is mobile if it is not an articulation point in the
connectivity graph G(cid:48)(cid:48) (see Fig. 3). In G(cid:48)(cid:48), vGP may or may
not be a non-articulation points, but according to Lemma
1, in every connected graph there always exist at least two
non-articulation points. Therefore, there exists at least one
agent i that has a non-empty restricted action set. For that
agent, one can compute a deterministic action sequence that
moves agent i to the ground plane. In other words, at each
iteration t we transfer one vertex from Vz>1 to vGP such
z>1 − 1 until Vz>1 = 0 and all agents
that V t
have been moved to vGP . This process terminates in a finite
number of time steps because the initial configuration CG,3D
is finite. The result is a 2D configuration C2D
Int representing
G(cid:48)(cid:48).
z>1 = V t−1
C G
Gz>1
G
G(cid:48)(cid:48)
SGP
Fig. 3: Example of a grounded configuration CG, the ground
plane SGP , associated connectivity graph G. Gz>1 repre-
sents all agents not on the ground plane, while all agents on
the ground plane are represented by a single node in G(cid:48)(cid:48)
I
.
Proof:
Corollary 1: Any finite grounded 3D configuration CG,3D
can be reconfigured into any other finite grounded 3D
configuration CG,3D
T
Since, according to Theorem 2, any finite
grounded 3D configuration CG,3D
can be reduced to an in-
termediate 2D configuration C2D
Int in a finite number of steps,
the reverse is also true - any finite grounded 3D configuration
CG,3D
Int in
T
a finite number of steps. According to Theorem 1, any 2D
configuration C2D
Int can be reconfigured into any other 2D
configuration C2D(cid:48)
Int . Therefore, there exists a deterministic
finite action sequence from CG,3D
can be assembled from some 2D configuration C2D(cid:48)
to CG,3D
.
I
I
T
V. STOCHASTIC RECONFIGURATION
In this and the following section we present a stochastic
reconfiguration algorithm that is fully distributed, does not
require any precomputation of paths or actions, and can
adapt to changing environment conditions. Unlike log-linear
learning ([2]), which cannot handle restricted action sets,
and variants such as binary log-linear learning ([1], [13],
[14]), which can only handle action sets constrained by
an agent's own previous action,
the presented algorithm
guarantees convergence to the potential function maximizer
even if action sets are constrained by all agents' actions.
Our algorithm is based on the Metropolis-Hastings al-
gorithm ([15],[10]), which allows the design of transition
probabilities such that the stationary distribution of the un-
derlying Markov chain is a desired target distribution, which
we choose to be the Gibbs distribution. This choice enables a
distributed implementation of the learning rule in Theorem
3 through the potential game formalism (see Corollary 2).
The Metropolis-Hastings algorithm guarantees two results:
existence and uniqueness of a stationary distribution. We
will use these properties to show that the only stochastically
stable state is x*, the potential function maximizer.
Theorem 3: Given any two states xi and xj representing
global configurations, the transition probabilities
2An articulation point is a vertex in a graph whose removal would
disconnect the graph.
pij =
1
τ (Φ(xj )−Φ(xi))
if e
1
τ (Φ(xj )−Φ(xi)) qji
qij
≤ 1
o.w.
qjie
qij
(cid:40)
e
e
1
τ
1
τ
1
τ
1
τ
Φ(x)
Φ(x)
(cid:80)
(cid:80)
x(cid:48)∈X e
x(cid:48)∈X e
Φ(x(cid:48)) .
guarantee that the unique stationary distribution of the un-
derlying Markov chain is a Gibbs distribution of the form
P r[X = x] =
Proof: Let X be a finite state space containing all
possible states of configurations composed of N agents.3
On that state space, let the desired target distribution be
Φ(x(cid:48)) with Φ defined in
π(x) = P r[X = x] =
Def. 2. By applying the Metropolis-Hastings algorithm, we
can compute transition probabilities P = {pij} such that π
is the stationary distribution of P , i.e. π = πP .
In the Metropolis-Hastings algorithm, a transition proba-
bility is represented as pij = g(xi → xj)α(xi → xj), where
g(xi → xj) is the proposal distribution and α(xi → xj) is
the acceptance distribution. Both are conditional probabilities
of proposing/accepting a state xj given that the current state
is xi.
Let agent k achieve the transition from state xi to xj
through action ak ∈ Rk(a). Then one possible choice for
the proposal distribution is g(xi → xj) = qij = 1Rk , ∀j ∈
{1, . . . ,Rk},
i.e. a random choice among all available
actions of agent k. According to Hastings ([10]), a popular
choice for the acceptance distribution is the Metropolis
choice αij = min
. Note that unlike in the original
formulation ([15]), we don't assume that qij = qji (see Fig.
4 for an illustration of qij and qji). These choices result in
the following transition probabilities:
1, πj qji
πiqij
(cid:110)
(cid:111)
(cid:40)
pij =
πj
πi
qji
qij
if πj qji
πiqij
o.w.
≤ 1
1
τ
1
τ
Φ(xi)
(cid:80)
e
x(cid:48)∈X e
It is easily verified that these pij satisfy the detailed bal-
ance equation πipij = πjpji and thus guarantee the existence
of a stationary distribution (see [15] and [10]). The resulting
pij follow from the definition of πi =
Φ(x(cid:48)) and
similarly πj.
Uniqueness of the stationary distribution follows from the
irreducibility of the Markov chain induced by P = {pij},
which is the case for our choice of proposal and acceptance
distribution because they assign a nonzero probability to
every action in any restricted action set. By Theorem 2 and
Corollary 1 we know that any state xj can be reached from
any other state xi and vice versa. Thus any action path has
nonzero probability and irreducibility follows.
Theorem 3 applies equally for 2D and 3D configuration.
However, for 3D reconfiguration, the proof relies implicitly
on the notion of groundedness to show irreducibility of the
underlying Markov chain (through the computation of R3D
).
The following theorem requires the definition of stochastic
stability.
Definition 10 (Stochastic Stability [20]): A state xi ∈ X
is stochastically stable relative to a Markov process P
if the following holds for the stationary distribution π
lim→0 π
xi > 0.
i
3Such a space is finite if we assume translation and rotation invariance
of the configuration as well as a finite environment.
State xi
State xj
State xj(cid:48)
a1
a2
a1
a2
a1
a3
Rk = 2
Rk(cid:48) = 1
Rk(cid:48)(cid:48) = 3
qij = 1Rk = 1
2
qji = 1Rk(cid:48) = 1
qji(cid:48) = 1Rk(cid:48)(cid:48) = 1
3
Fig. 4: Example of forward and reverse actions with their
associated proposal probabilities qij, qji, and q(cid:48)
ji. Note that
xi, xj, x(cid:48)
j are states of the entire configuration, and agent k
is the currently active agent.
Note that the Markov process is defined through the transi-
tion probabilities in Theorem 3 and the stationary distribution
is a Gibbs distribution. Furthermore is equivalent to the
learning rate τ in the following proof.
Theorem 4: Consider the self-reconfiguration problem in
Def. 2. If all players adhere to the learning rule in Theorem
3 then the unique stochastically stable state x∗ is the state
that maximizes the global potential function.
x(cid:48)∈X e
Proof: Note that
this result holds by definition of
the desired target distribution, which is a Gibbs distribution
centered at the state of maximum global potential and defines
the probability of being in a state x as P r[X = x] =
(cid:80)
Φ(x(cid:48) ) . The learning rate or temperature τ represents
the willingness of an agent to make a suboptimal choice (i.e.
explore the state space). As τ → 0, P r[X = x] → 0 for all
states x (cid:54)= x∗ which are not potential function maximizers
(see [2] and [14]).
Note that the maximum global potential is achieved when
all agents are at a target position ai ∈ CT . Algorithm 1
shows an implementation of Theorem 3. In Algorithm 1,
similarly to the Metropolis-Hastings algorithm in [10], pii =
Φ(x)
1
τ
1
τ
e
1 −(cid:80)
j(cid:54)=i pij.
VI. A DECENTRALIZED ALGORITHM
One shortcoming of Algorithm 1 is its centralized nature
that requires the computation of a global potential function
Φ(xi) and depends on the entire current configuration xi.
A decentralized algorithm is desirable to execute the global
learning rule on a team of agents. The formulation of the
self-reconfiguration problem as a potential game allows us to
rewrite the transition probabilities in a decentralized fashion
as follows.
Corollary 2: The global learning rule of Theorem 3 can
be decentralized such that each agent can execute it with
local information only.
xi as (a(cid:48)
Proof: Note that for agent k, we can express xj and
k, a−k) and (ak, a−k) respectively. According to
input : Target configuration CT
Start clock (see [3])
while True do
if Clock ticks then
(cid:110)
Compute current restricted action set Rk
Select a(cid:48) ∈ Rk with probability q = 1Rk
Compute α = min
if α = 1 then
xt+1 = a(cid:48)
τ (U (a(cid:48))−U (a))(cid:111)
Rk
Rk(cid:48) e
1,
1
a(cid:48) with probability α
a with probability 1 − α
(cid:40)
else
xt+1 =
end
end
input : Current and target configuration C and CT
while True do
(cid:110)
Randomly pick an agent k in state xi
Compute restricted action set Rk
Select ai→j,k ∈ Rk with probability qij = 1Rk
Compute αij = min
if αij = 1 then
xt+1 = xj
else
τ (Φ(xj )−Φ(xi))(cid:111)
1, qji
qij
e
1
(cid:40)
xj with probability αij
xi with probability 1 − αij
xt+1 =
end
end
Algorithm 1: Global game-theoretic learning algorithm.
Note that state xj is the result of agent k applying
action ai→j,k and xi and xj refer to states of the entire
configuration. Also note that qii /∈ Rk but pii (cid:54)= 0.
end
Algorithm 2: Self-reconfiguration using game-theoretic
learning local version that each agent executes. Note that
xt, xt+1 refer to consecutive states of the active agent.
Proposition 1, we can then rewrite Φ(xj)−Φ(xi) as follows.
k) − Uk(ak) = Uk(a
k, a−k) − Uk(ak, a−k)
(cid:48)
(cid:48)
Uk(a
k, a−k) − Φ(ak, a−k)
(cid:48)
= Φ(a
= Φ(xj) − Φ(xi)
(cid:40)
Therefore, we can rewrite the transition probabilities as
follows.
pij =
1
τ (Uk(a(cid:48)
k)−Uk(ak))
if e
1
τ (Uk(a(cid:48)
k)−U (ak)) qji
qij
o.w.
qjie
qij
≤ 1
Since qij, qji, Uk(a(cid:48)
k), as well as Uk(ak) can be computed
with local information, so too can the transition probabilities.
The stationary distribution of the process described by pij is
the same Gibbs distribution as in Theorem 3.
Note that
local can mean multiple hops, because the
computation of restricted action sets requires to maintain
groundedness of all neighboring agents. Verifying ground-
edness requires a DFS search to the ground plane, which in
the worst case can take N − 1 hops.
Algorithm 2 shows a decentralized implementation of
Algorithm 1 and Corollary 2. Note that a transition from
configuration xi to xj is accomplished by agent k executing
action a(cid:48) ∈ Rk starting at its current location a. Therefore,
we can interpret qij as a transition probability for a forward
action and qji as a reverse action (see Fig. 4).
VII. IMPLEMENTATION
The global version of the algorithm was implemented
and evaluated in Matlab. For the simulations shown in Fig.
5, we used a τ = 0.001 that struck a balance between
greedy maximization of agent utilities and exploration of the
state space through suboptimal actions. Agents' actions are
restricted according to the action set computation outlined
in Section III-C and the motion model in Section III-A.
Restricted action sets depend on the agents joint action and
the environment - in our simulations only the ground plane.
Agents' positions are initialized above the ground plane such
that their z-coordinate z ≥ 1. In a straightforward extension
to this algorithm, obstacles can be added to restrict
the
environment even further.
Fig. 5 shows convergence results of Algorithm 1 of
configurations containing 10, 20, and 30 agents. Four types
of reconfigurations have been performed: 2D to 2D, 2D to
3D, 3D to 2D, and 3D to 3D. The vertical lines in Fig.
5 represent
the average time to convergence of all four
types of reconfigurations of a certain size (e.g. the leftmost
line represents average convergence of a configuration of 10
agents). Convergence is achieved, if the configuration reaches
a global potential of Φ = N, i.e. every agent has a utility
of Ui = 1. Note that in the scenarios of Fig. 5, the target
configuration was offset from the initial configuration by a
translation of 10 units along the x-axis. One can observe that
at the beginning of each reconfiguration the global potential
ramps up very fast (within a few hundred time steps), but the
asymptotic convergence to the global optimum can be slow
(see the case 3D to 2D for 30 agents). An example of a 2D
to 2D reconfiguration sequence is shown in Fig. 1.
VIII. CONCLUSION
In this paper, we have applied the potential game for-
malism to the self-reconfiguration problem and developed a
stochastic algorithm that converges to the global potential
function maximizer. Under the assumptions of grounded-
ness, we have introduced a decentralized approach to self-
reconfiguration that requires little communication, does not
rely on a centralized decision maker or precomputation,
and converges to the global optimum. Our simulation re-
sults suggest that this formulation of the self-reconfiguration
problem is indeed a feasible approach and can solve self-
reconfiguration in a decentralized fashion.
Fig. 5: Convergence times for different types of configurations and sizes ranging from 10 to 30 agents.
[13] Yusun Lee Lim. Potential game based cooperative control in dynamic
environments. Master's thesis, Georgia Institute of Technology, 2011.
[14] Jason R. Marden and Jeff S. Shamma. Revisiting log-linear learning:
Asynchrony, completeness and payoff-based implementation. Games
and Economic Behavior, 75(2):788 -- 808, 2012.
[15] Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth,
Augusta H. Teller, and Edward Teller. Equation of state calculations
by fast computing machines. Journal of Chemical Physics, 21:1087 --
1092, 1953.
[16] Dov Monderer and Lloyd S Shapley. Potential games. Games and
Economic Behavior, 14(1):124 -- 143, 1996.
[17] Daniel Pickem and Magnus Egerstedt. Self-reconfiguration using
In Analysis and Design
graph grammars for modular robotics.
of Hybrid Systems (ADHS), 4th IFAC Conference on, Eindhoven,
Netherlands, June 2012.
[18] Daniel Pickem, Magnus Egerstedt, and Jeff S Shamma. Complete
heterogeneous self-reconfiguration: Deadlock avoidance using hole-
free assemblies. In Distributed Estimation and Control in Networked
Systems (NecSys'13), 4th IFAC Workshop on, volume 4, pages 404 --
410, September 2013.
[19] Daniela Rus and Marsette Vona.
Crystalline robots: Self-
reconfiguration with compressible unit modules. Autonomous Robots,
10(1):107 -- 124, Jan. 2001.
[20] Peyton H. Young. The evolution of conventions. Econometrica,
January 1993.
[21] Minghui Zhu and Sonia Mart´ınez. Distributed coverage games for
energy-aware mobile sensor networks. SIAM Journal on Control and
Optimization, 51(1):1 -- 27, 2013.
REFERENCES
[1] Gurdal Arslan, Jason R Marden, and Jeff S Shamma. Autonomous
vehicle-target assignment: A game-theoretical formulation. Journal
of Dynamic Systems, Measurement, and Control, 129(5):584 -- 596,
September 2007.
[2] Lawrence E. Blume. The statistical mechanics of strategic interaction.
Games and economic behavior, 5(3):387 -- 424, 1993.
[3] Stephen P. Boyd, Arpita Ghosh, Balaji Prabhakar, and Devavrat Shah.
IEEE Transactions on Information
Randomized gossip algorithms.
Theory, 52(6):2508 -- 2530, 2006.
[4] Zack Butler, Keith Kotay, Daniela Rus, and Kohji Tomita. Cellular
In
automata for decentralized control of self-reconfigurable robots.
Proc. of the ICRA 2001 workshop on modular robots, pages 21 -- 26,
2001.
[5] Zack Butler, Keith Kotay, Daniela Rus, and Kohji Tomita. Generic
decentralized control for lattice-based self-reconfigurable robots. The
International Journal of Robotics Research, 23(9):919 -- 937, 2004.
[6] Kenneth C. Cheung, Erik D. Demaine, Jonathan Bachrach, and Saul
Griffith. Programmable assembly with universally foldable strings
(moteins). IEEE Transactions on Robotics, 27(4):718 -- 729, 2011.
[7] Robert Fitch and Zack Butler. Million module march: Scalable
The International
locomotion for large self-reconfiguring robots.
Journal of Robotics Research, 27(3-4):331 -- 343, 2008.
[8] Robert Fitch, Zack Butler, and Daniela Rus. Reconfiguration planning
for heterogeneous self-reconfiguring robots. In Intelligent Robots and
Systems, 2003.(IROS 2003). Proceedings. 2003 IEEE/RSJ Interna-
tional Conference on, volume 3, pages 2460 -- 2467, October 2003.
[9] Michael J. Fox. Distributed Learning in Large Populations. PhD
thesis, Georgia Institute of Technology, August 2012.
[10] W Keith Hastings. Monte carlo sampling methods using markov chains
and their applications. Biometrika, 57(1):97 -- 109, 1970.
[11] Keith Kotay and Daniela Rus.
locomotion for a self-
reconfiguring robot. In Robotics and Automation, 2005. ICRA 2005.
Proceedings of the 2005 IEEE International Conference on, pages
2963 -- 2969, 2005.
Efficient
[12] Haruhisa Kurokawa, Kohji Tomita, Akiya Kamimura, Shigeru Kokaji,
Takashi Hasuo, and Satoshi Murata. Distributed self-reconfiguration
of M-TRAN III modular robotic system. The International Journal of
Robotics Research, 27(3-4):373 -- 386, 2008.
010002000300040005000600070008000900010000051015202530Timetoconvergenceusingτ=0.001Time stepsGlobal Potential Size 10: 2D to 2DSize 10: 2D to 3DSize 10: 3D to 2DSize 10: 3D to 3DSize 20: 2D to 2DSize 20: 2D to 3DSize 20: 3D to 2DSize 20: 3D to 3DSize 30: 2D to 2DSize 30: 2D to 3DSize 30: 3D to 2DSize 30: 3D to 3D |
1801.06740 | 1 | 1801 | 2018-01-20T23:31:14 | Knowledge Representation for High-Level Norms and Violation Inference in Logic Programming | [
"cs.MA",
"cs.AI"
] | Most of the knowledge Representation formalisms developed for representing prescriptive norms can be categorized as either suitable for representing either low level or high level norms.We argue that low level norm representations do not advance the cause of autonomy in agents in the sense that it is not the agent itself that determines the normative position it should be at a particular time, on the account of a more general rule. In other words an agent on some external system for a nitty gritty prescriptions of its obligations and prohibitions. On the other hand, high level norms which have an explicit description of a norm's precondition and have some form of implication, do not as they exist in the literature do not support generalized inferences about violation like low level norm representations do. This paper presents a logical formalism for the representation of high level norms in open societies that enable violation inferences that detail the situation in which the norm violation took place and the identity of the norm violation. Norms are formalized as logic programs whose heads specify what an agent is obliged or permitted to do when a situation arises and within what time constraint of the situation.Each norm is also assigned an identity using some reification scheme. The body of each logic program describes the nature of the situation in which the agent is expected to act or desist from acting. This kind of violation is novel in the literature. | cs.MA | cs | Knowledge Representation for High-Level Norms and Violation Inference in Logic
Programming
B.O. AKINKUNMI1 and Moyin F. BABALOLA2
1Dept of Computer Science,
University of Ibadan, Ibadan, Nigeria.
2Dept of Computer Studies,
The Polytechnic, Ibadan, Nigeria.
ABSTRACT
Most of the Knowledge Representation formalisms developed for representing prescriptive
norms in multi-agent normative systems can be categorized as either suitable for representing
high-level or low level norms. We argue that low-level norm representation do not advance
the cause of autonomy in agents in the sense that it is not the agent itself determining the
normative position it should be in at a particular time, on the account of a more general rule.
In other words the agent depends on some external system for a nitty-gritty prescription of its
obligations and prohibitions. On the other hand, high-level norms which have an explicit
description of a norm's precondition and have some form of implication, do not, as they exist
in the literature, support generalized inferences about violation like low level norm
representation formalisms do.
This paper presents a logical formalism for representation of high level norms in open
societies that enable violation inferences that detail the situation in which the norm violation
took place and the identity of the norm violated.
Norms are formalized as logic programs, whose heads each specify what an agent is obliged
or permitted to do when some situation arises and within what time constraint of the situation.
Each norm is also assigned an identity, using a reification scheme. The body of each logic
programs usually describes the nature of the situation in which the agent is expected to act or
desist from acting. This kind of violation inference is novel in the literature.
Keywords: Multi-Agent Systems, Norms, Logic programs, Knowledge Representation
1. INTRODUCTION
Norms are used in MAS to cope with autonomy, different beliefs, interests and desires of the
agents that cohabit in the system. Based on the normative descriptions and the actual (past
and present) actions of the agents, the system should detect the deviating behavior of any
agent in the system.
According to Marin and Sartor(1999), there are four clear properties that can be ascribed to a
norm of conduct. These are: applicability, (pre-)condition, effect and validity. The effect of a
norm is what an agent must do: the obligation to carry out an action or a prohibition from
carrying out an action. The condition is a situation that arises warranting the effect of the
norm. The applicability of a norm is the time that a condition must hold in order for the
effect to hold, while the validity of a norm is determined by its inclusion in the current body
1
of code of conduct at that time.
We classify most of the norms of conduct that have appeared in the literature into two
classes. Low level norms are those that specify a normative position or effect of a norm and
the time when (or by which) an agent is expected to conform to it. Examples of works that
formalize only low level norms include Sadri(2006), Stratulat et al(2001) and King et
al(2017). We call this kind of norm representation as such because they instruct the agent on
what to do and when to do it, rather than have the agent decide on when to act and for what
reason. As it were, the condition for a norm's effect to take place is defined by a clock alarm
signaling the arrival of the specified time. The violation of such a norm takes place when that
time interval expires without the agent effecting the appropriate reaction stipulated by the
norm's effect (i.e. either carry out a specific action or desist from doing so).
On the other hand there are high-level norms that specify an explicit description of both the
effect of the norm and the condition warranting it as well as the relationship between the time
of the effect and the time of the warranting condition. Examples of work from the literature
that formalized only high level norms include the work of Governatori, Rotolo and
Sartor(2005) as well as Artikis and Sergot(2010).
Our argument in this paper is that high-level norms are more appropriate for modeling real
life norms of conduct than are low level norms. This is because low-level formalisms cannot
capture many real life norms that require a detailed description of a warranting situation than
simply the striking of the clock. In addition to this, real life norms of conduct also carry time
constraint between the time the agents should effect his obligation or prohibition and the time
of the warranting condition.
The violation of a high level norm is the failure of an agent to carry out the effect of the norm
within the time constraint of the warranting situation. Furthermore, it is pertinent to
determine the validity of the violated norm(s) as at the time the violation took place, because
an invalid norm cannot be violated. A norm is said to be valid at a point in time if it is part
(member) of the system of norms in consideration at that point in time. The interval, during
which the norm is valid, is called the external time of that norm (Marin and Sartor, 1999).
A particular instance of low level norm representation is the formalism of King et al. (2017)
which treats a normative fluent as a relationship between an event that must hold by a certain
deadline, so that an obligation for an action a to be taken by a deadline d is the fluent obl(a,
d). Such an obligation is "discharged" if such an action takes place before the deadline or
"violated" otherwise.
This paper presents a logic programming based knowledge representation formalism for
norms that captures a norm's condition as a situation, its effect as an the obligation or
prohibition of an action and the temporal constraint between them as a reified entity that
captures qualitative and quantitative constraint. Section 2 discuses related works from the
literature. Section 3 discusses the language of the formalism. Particularly, section 3.2.3
2
discusses the temporal constraint structures used to represent the reified temporal constraints
expected between an agent's action and situation. Section 3.4 discusses how these temporal
constraints can be checked using constraint satisfaction rules expressed as Logic programs.
Section 3.5 shows how to use all these to make inferences about norm violation. Section 3.6
compares our formalism with some existing ones and points out parallel concepts from other
papers.
2. NORM REPRESENTATION IN THE LITERATURE
According to Sergot(1990), there are two kinds of norms in legal reasoning. These are
qualification norms and norms of conduct. Qualification norms define normative or legal
properties and relationships of agents, while norms of conduct define how an ideal agent is
expected to behave under certain circumstances. Consequently, unlike qualification norms,
norms of conduct (or prescriptive norms) take cognizance of the fact that there is a difference
between how an agent is expected to behave and how it might choose to behave in the real
world. The departure between how an agent is expected to behave and how it eventually
behaves is a violation of that expectation or norm.
Much work has been carried out on the logical representation of such norms of conduct.
Notable papers in this regard include the works of Stratulat et al (2001), Marin and
Sator(1999), Sadri et al (2006) and Panagiotidis et al (2009). As Artikis(2003) observed on
the work of Marin and Sator, the paper focused on three major issues namely: norm
applicability, norm effect and norm validity. Norm applicability describes the time interval
during which a norm's (pre-) condition must hold in order for its effect to hold. The effect of
a norm is the action an agent is expected to take or avoid taking when the norm's
precondition takes place. Norm validity at a time describes the recognition of that norm as
part of the body of legislation at a time. Marin and Sartor went on to identify the limitations
of Kowalski and Sergot' s Event Calculus in solving these problems in the process of
formalizing norms and presented an extension of Event Calculus for handling these problems.
In Stratulat et al (2001) the norms are represented by what seem like a reified ternary
predicate. For example an obligation is a triple O(Agent, Action-type, Time-Interval) that
carries the information that an agent is obliged to carry out an action of the type Action-type
during the time interval. For example in Stratulat's language the statement
holds(j, O(ag, a, k)) ………Axiom ST1
states that an obligation for an agent to carry out an action a at the time interval k holds
during time interval interval j. The interval k is the internal time of the norm in the language
of Marin and Sator while the interval j is the external time during which the norm is in effect.
Presumably k should be a subinterval of j. A violation of such a norm is inferred when the
agent is found at a future time to have failed to carry out such an expected action within the
prescribed time interval.
3
Similarly, Sadri et al(2006)'s representation of an obligation is the atom:
obliged(act(Act, Actor, Parameters), T, TC) or
prohibited(act(Act, Actor, Parameters), T, TC)
Thus an assertion that the actor Actor carries out (or is prohibited from carrying out) the
action Act at time T which must satisfy the temporal constraint TC. Sadri et al's temporal
constraints show the time points that bound by the time e g 10 > T > 17.
In addition, for most of the norms represented by Sadri et al(2006), the norm's condition for
an agent to be obliged to act or refrain from acting is set by time itself. Stratulat et al(2001)'s
paper also shared this limitation. For example: you can't park your car here between 10am
and 5 pm. Rather this kind of representation is inadequate for representing a norm like: you
can't park your car here from the moment it starts raining until it's 15 minutes after the rain.
While norms of the earlier type can be regarded as low level norms in the sense that they can
only be applied within a specific time frame, those of the latter kind can be regarded as high
level norms, in the sense that they apply anytime it rains. In that case it's up to the agent to
decide whether the norm is applicable in a particular situation and at what particular time an
agent finds itself in a deontic position.
A particular kind of low level norm is the normative fluent used by King et al(2017). Their
normative fluent is written as ob(a, d), denoting an obligation to carry out an action of type a
before a deadline d. The problem with this kind of deadline norms is that such norms can be
satisfied by actions of similar type undertaken before this norm becomes active. The problem
with this representation is that unless the record of an action that discharged a similar fluent
in the past is removed from the knowledge base as soon as it is discharged, it can
inadvertently discharge future obligations.
This first gap is bridged in this paper by representing the time interval constraints using time
point images (TPI), so that indefinite times and constraints can be represented. The second
gap is also bridged here by using situations to model a norm's condition. A situation can be
described by the propositional fluents (e.g. a raining fluent) that characterize it. Similarly a
situation can be associated with a certain time in which it holds as done by Pinto (1994). This
will make it possible to describe the conditions under which an agent cannot park his/her car
for each of the two norm examples given above.
There are two major examples of norm representation formalisms that are indeed high level
norms in the literature. The first is the temporalized formalization of legal norms by
Governatori, Rotolo and Sartor(2005). In this formalization, a norm's condition has an
implication relationship with the norm's effect (for example, the fact that an action is
obligatory), with each side of the implication associated with relative times. A format for
such rules looks like the following:
4
condition : t effect : t k
This rule reads like: if the condition holds at time t, then the effect (such as an action
becoming obligatory) holds at time t k. The first problem with this kind of formalization is
that it leaves no room for reasoning about violations of the norm (as pointed out by the
authors themselves), particularly because there is no way of identifying the norm being
violated. The best that can be done is to identify the violation of the effect of the norm such
as an obligation to carry out an action at some time relative to the time the condition arose.
The second problem with it is that temporal relations between conditions and effects can be
more complex than can be represented by constraints involving two time points. Firstly when
conditions arise, they usually have a starting point and a termination point. Therefore an
agent may have an obligation to act at some point relative to (i.e. before or after) the start or
finish or the time interval the condition lasted for. For example a robo-sweeper may be
obliged to start sweeping no later than 5 minutes after snowing stops. Another example is
when an agent is obliged to carry out an action 10 minutes before the scheduled end of an
event.
Another example of a formalism we consider appropriate for high level norms in the
literature is one based on the Event Calculus and implemented in PROLOG by Aritkis and
Sergot(2010). An example of a norm that can be represented in Artikis and Sergot's
formalism is: A faculty is empowered to supervise three years after earning a doctorate. It is
rendered thus:
happens(earnedDoctorate(F) = true, T – 1096),
holdsAt(enrolled(X) = true, T).
holdsAt( pow(F, supervise(F, X)) = true, T)
The representation of the robosweeper example (i.e. Robosweeper is obliged to start
sweeping no later than 5 minutes after snowing stops) however, reveals some problems with
EC representations such as this one. It can be rendered thus in Artikis and Sergot's EC
representation thus:
holdsAt(obl(Robosweeper, startweep) = true, T)
holdsAt(stop(snowing), T1),
T1 T,
T1 +5 > T.
The problem with EC representations of norms stems from the fact that actions in EC are
instantaneous. As such, when confronted with actions that have a duration such as sweeping,
one can only capture the commencement and/or termination of such actions when necessary.
A good representation of this norm however, should not only capture the commencement of
the sweeping action, but the action in its entirety. The only way we can capture the sweeping
action in an EC representation is to use the start and stop functions and treat sweep as an
5
event like snowing. However, from an agent's point of view it is critical to distinguish
between an event like snowing and an action like sweeping.
Nonetheless, Both Governatori et al as well as Artikis and Sergot(2010) in their
representations captured the essence of a high-level norm which are a norm's condition and
the essence of a temporal relationship between them. However, having a norm's condition
and effect appear on either sides of an implication as done by both Governatori et al(2005)
and Artikis and Sergot(2010), makes it impossible to give an identity to a particular high
level norm.
In general many real life prescriptive norms are high-level norms. Besides, truly autonomous
agents must be able to work with high level norms. Take for instance a real-life scenario. In
an electronic institution where there is a norm that obliges an agent to supply an order within
forty eight (48) hours after receiving the acknowledgement of payment for the order. The
condition required for the agent to carry out the supply of the order (norm's effect) is
receiving the acknowledgment of payment for the order (norm's condition).
Although this kind of norm can easily be formalized in the logic of Governatori et al, as:
Pay-ack(o) : t Obl(supply(o)) : t + 48
We can only talk about the violation of the specific instances of the effect of this norm which
is an obligation to supply an order at some given time (a low-level norm), and not the
violation of the norm itself. Inferring which high level norm has been violated is critical to
helping a sincere agent improve on its ability to be law-abiding. Communicating such norm
violations also requires that each norm be identified so that the identity of the norm can be
communicated to the erring agent.
Finally, inferring a norm violation requires that one determines whether or not the norm in
question is active or not as only an active norm can be violated. This requires that we be able
to have assertions about a norm's enactment and (if applicable) the repeal dates. This is only
possible if each norm has an identity that can be used to refer to it. Such identities can even
be quantified over in order to form general rules about norms.
Thus the representation of the norm used in this paper carries the following information:
The identity of the agent involved,
A normative fluent such as obl(a) or pro(a) that indicates that an obligation to
carry out an action a or a prohibition from carrying it out. This is the norm's
effect.
A situation whose description will capture the nature of the condition
warranting the effect of the norm,
6
The reified time constraint between the action and situation and
The norm's identity.
A normative token is treated as the relation bringing the other five elements together. This
formalization eliminates the need for a normative implication of the sort championed by
Governatori et al.
3. REPRESENTATION OF HIGH LEVEL NORMS
A norm is basically a rule. Each of these normative rules help in making inferences about
what an agent is expected to do or desist from doing within a certain time frame. The
inferences made from these normative rules are referred to as normative tokens.
A normative token is when an agent finds it is expected to effect a normative fluent, within
some time constraint, and because of, the arising of some named situation in conforming to a
named normative rule. A normative token relation is represented by a predicate NormPos
thus:
NormPos : Agent NormFluent Situation TC Norm-Id Boolean
(where the normative fluent is derived from an application of the function obl to an action
type)
Norm-id is an identifier for the norm rule that produced the actual normative token which is
an obligation. Every obligation that is produced by the norm bears the same norm-id. As such
every rule that helps to infer obligations, prohibitions and permissions has a unique norm-id
that it carries. This kind of rule naming is referred to by the term rule reification which is
similar in spirit to the notion of Davidson's reification (Galton 1991), This is illustrated with
Norm 3.1 and 3.2 as examples of rules generating this kind of normative tokens.
In sections 3.1 – 3.3, the language employed is a many sorted standard first order predicate
logic with equality and with standard semantics for operators or (), and (), negation ()
existential quantifier (), universal quantifier () implication( ) and equivalence( ). This
is needed to express the properties of relations such as "partially describe" or "fully
describe". In section 3.4 and 3.5, we resort to expressing norms with a language that is close
to standard logic programs with equality and typical non-classical implication represented by
if as well as, non-classical negation represented by .
Time is viewed as a linear and discrete model in which we recognize both time intervals and
instants to be known as the sort Interval or Instants. The relations among two time points are
represented by the infix predicates: <, >, , with the signatures:
<, >, , : Instant Instant Boolean
7
The beginning and ending of Intervals represented by the functions begin and end
respectively; of Intervals are Instants as the signature below shows:
begin : Interval Instants
end : Interval Instants
In addition, we will use all of Allen(1984)'s binary interval relations represented by
predicates: Meets, Before, Starts, Ends, During, Overlaps and their inverse relations. In
addition, we will use other binary relations built on these such as Within (interpreted as Starts
or During or Ends) and (improper) Subinterval (interpreted Within or Equals) by
Koomen(1989).
3.1 Normative Fluents and Action Types and Tokens
The basic sorts introduced in this subsection are normative fluents and Action types as well
as Action Tokens. A normative fluent is either the obligation on the part of an agent to carry
out an action of a certain type or the prohibition on the part of an agent from carrying out an
action of a certain type. The permission on the part of an agent to carry out an agent can be
taken as the negation of the prohibition to carry out the same action. Because normative
fluents are reified, the notion of obligation and prohibitions will be treated as the functions
obl and pro respectively. These are both reckoned to have the following signature:
obl, pro: ActionType NormativeFluent
In other words, the obl and pro functions each take an action type as argument and returns a
normative fluent. Apart from these there is a notion of negation for normative fluents which
is the function with the signature:
: NormativeFluent NormativeFluent
The meaning of obl(a), where a is an action type a should be read as "permission not to
carry out an action of type a", while the meaning of pro(a) should be taken to be
"permission to carry out an action of type a". Norms that carry this kind of normative fluents
cannot be violated.
Action types themselves carry a functional structure, so that obl and pro can be regarded as
higher order functions. An action type is usually a function denoted by a verb that describes
the generic type of the action, while its argument will be the object of the action. For example
an action type repair the car12 is the functional token repair(Car12). Thus the obligation to
repair the car12 is the normative fluent: obl(repair(Car12)).
So far we have been talking about an action type. However, there are instances when a
8
norm's condition needs to be described by the occurring of an instance of an event type. In
that case we define the time when an actual event takes place using a function timeA which
given an action returns a time interval, with the signature:
timeA : Action Interval
Finally, assertions that make an actual event belong to a particular event type, are represented
by the predicate Type-of-Action with the signature:
Type-of-Action : Action ActionType Boolean
3.2 Situations, Events, Processes
The basic sorts introduced here are Fluents, Situations, Events Processes and Time Intervals.
In this paper, the conditions of a norm are modeled by situations. Our situation are similar to
that described by Schubert(2000). A situation represents a complete state of the world with a
time frame. Each fluent describes a specific aspect of a situation. In Schubert's FOL**
language there are two kinds of relations that may exist between a fluent f and a situation s.
In one relation a fluent completely describes the situation written as f **s, while in the other
the fluent partially describes the situation f *s. In FOL**, the relation between a situation and
a fluent that holds in it, is in general, an entailment relation i.e. s ╞ f. Unlike FOL** however,
our fluents which we will treat as functional tokens in our language are atomic propositions
that do not require the use of quantifiers eg. Car17 is red, John loves Mary, etc. Speaking of
functions, there is a time function for situations in this formalism which given a situation
returns the time interval over which the situation persists. This function is timeS with the
following signature:
timeS: Situation Interval.
In the reified situation calculus of Pinto(1994), the Holds predicate is used to represent the
relation between a fluent and the situation in which it holds. Holds in Pinto's reified situation
calculus does not in any way, suggest that a fluent completely describes a situation. As such
we will thus use Holds(f, s) to connote the relation f partially describes s.
Schubert(2000) has argued the need for having situations that are completely described by
specific fluents so that accurate causation relations can be expression between situations. The
problem with only allowing situations to be partially described is that often causation
relations are between certain aspects of situations. Other aspects of those situations are
dormant in that relationship. This underscores the need for a completely describes relation
between fluents and situation. Schubert renders this relation in FOL** as **. In order to
respect the tradition from which this relation came from, we will use Holds** as relation
between a fluent and the situation it completely describes so that Holds**(f, s) means f
9
completely describes s. There are two specific relations between Holds and Holds**
described below:
Axiom 3.2.1
The relation "completely describes" is a specialization of "partially describes".
f, s. Holds**(f, s) Holds(f, s)
Axiom 3.2.2
If f partially describes s, there exists another situation of the same time as s, which f
completely describes.
f, s. Holds(f, s) s1. Holds**(f, s1) timeS(s) = timeS(s1)
We will define an implication relationship among fluents as the predicate ImplyF thus:
Axiom 3.2.3
An implication relation exists among between a fluent and another, if the fact that it holds in
any situation implies that the other holds in the same situation.
f, f1. ImplyF(f, f1) s. Holds(f, s) Holds(f1, s)
A simple instance of the implication relationship between fluents will exist between a raining
fluent and a wet fluent. The following axiom further clarifies the difference between the
relations denoted by predicates Holds** and Holds.
Axiom 3.2.4
If f completely describes a situation, then no other fluent other than those implied by f can
hold in that situation.
f, s. Holds**(f, s) (f1. f f1 ( ImplyF(f, f1) Holds(f1, s) ))
Two special instances of fluents are the occurring of an event or the progression of a process.
The fluents are represented by the functional structures: occurring(e) and prog(p)
respectively, where occurring and prog are functions are applied on event token e and
process p respectively giving rise to fluents with signatures:
occurring: Event Fluent
prog : Process Fluent
Each of these can partially or fully describe a situation. Just like situations both events and
processes, have are eventualities with have an inherent time property. These are represented
by time functions timeE and timeP for events and processes respectively, with the following
signatures:
10
timeE : Event Interval
timeP : Process Interval
When an event or a process describes a situation, the time of that situation is the same as that
of the event or process.
Axiom 3.2.5
When the occurring of an event or the progression of a process describes a situation, the time
of the situation is the same as the time of the event or that of the process as the case may be.
(a) Holds(occurring(e), s) timeE(e) = timeS(s)
(b) Holds(prog(p), s) timeP(p) = timeS(s)
It is particularly important to be able to have situations completely described by these kinds
of fluents, because such fluents actually give rise to the presence of other fluents. When the
fact of an event occurring is a fluent that describes a situation, then the fact that certain agents
play a role in that event is also a fluent that partially describes the situation. For the purpose
of describing the relations between an actual event and an event type, as is needed in domain
Axioms 3.2.6 and 3.2.7 below, we resort to using a predicate EventType thus:
e, s. Holds**(occurring(e), s) EventType(e, Class)
!a. Holds(role(a, Teacher), s)
Axiom 3.2.6
If a classroom event occurs within a situation, then there must be a particular agent that
plays the role of a teacher in that situation.
Axiom 3.2.7
If a classroom event occurs within a situation, then there must be at least one agent that plays
the role of a student in that situation.
Finally it is important to distinguish between a process and events has been done in the
literature and summarized in Galton(2016). Our definition of processes is close to that of
Sowa(2000) as cited by Galton. A process can be discrete or continuous. Discrete processes
each involves a finite sequence of either event occurrences or states (described by fluents).
Each step can either be an event (or action), a state of affairs (like in a natural process like a
fermentation or decay) or another process which is a process by itself. A continuous process
can be treated as constituted by a single event. Herein lies the difference between our view of
e, s. Holds**(occurring(e), s) EventType(e, Class)
a. Holds(role(a, Student), s)
11
processes and Sowa's who breaks each continuous process into initiation, continuation and
termination phases.
Definition 3.2.8
A situation s is fully characterized by a process if Either that situation is fully characterized
by either the occurring of an event Or there are two smaller situations whose times cover s
and the first situation is fully characterized either by the occurrence of an event or a fluent
and the second is also a process.
p, s. Holds**(prog(p), s)
(e Holds**(occurring(e), s))
s1.(e Holds**(occurring(e), s1) f Holds**(f, s1))
s2, p1. Holds(prog(p1), s2) Cover(timeS(s1), timeS(s2), timeS(s))
(where j, k, m. Cover(j, k, m) Starts(j, m) Finishes(k, m) Meets(j, k)
see (Koomen 1989) )
Either way, Axiom 3.2.7 above captures the definition of a process.
3.3 Temporal Constraints
The basic sorts introduced here are time point images (TPI) and Time Constraints (TC). A
TPI is a relative point with an anticipated time interval. It can be the beginning (B), End (E)
or any relative time point within the anticipated time interval. Therefore a TPI may be either
of the constants {B, E} or an application of the time displacement function tdisp to either B
or E and an integer. B and E signify the beginning and end of a time integer. The signatures
are given below:
Our approach to composing constraints here is therefore is to let each basic constraint relate
some time point image (i.e. its beginning, end, or some point in between) of the time interval
of the effect of a norm which is the time of the action of the type specified, with some time
point image of the effect of the norm. A basic temporal constraint is composed by the
application one of the functions such as eq (equal), ge (greater than or equal to), le (less than
or equal to), gt greater than) and lt (less than) to an ordered pair of TPI, the first relating to
the action (or norm's effect) while the second is related to the situation (norm's condition)
with the signatures:
tdisp: TPI Integer TPI
eq, ge, gt, le, lt : TPI TPI TC
The functions and, or and not introduced here are used to make temporal constraints with
natural Boolean meanings and the following signatures:
12
not : TC TC
and, or : TC TC TC
For example:
eq(B, E) means the first interval ends exactly when the second interval begins.
eq(tdisp(B, 3), E) means the second interval ends 3 units into the beginning of the first
interval.
Other temporal constraints can be composed from other temporal constraints by the functions
and, or as well as neg. For example, and(eq(B, B), eq(E, E)) describes two intervals that are
equal.
Each of Allen(1984)'s qualitative relations can be represented by our TCS. The following are
the equivalences between Allen's interval relations and our TCS.
Before is equivalent to
Overlaps is equivalent to
Contains is equivalent to
Starts is equivalent to
Finishes is equivalent to
Meets is equivalent to
lt(E, B)
and(lt(B, B), lt(E,E))
and(lt(B, B), gt(E, E))
and(eq(B, B), lt(E, E))
and(gt(B, B), eq(E, E))
eq(E, B)
A major advantage of the TCS representation is that it can represent constraints that combine
both qualitative and quantitative relationships. Examples of such relationships are:
Starts not later than 4 units of time into is equivalent to le(B, tdisp(B, 4))
Ends 10 units of time into
is equivalent to eq(E, tdisp(B, 10)
These are the kinds of constraints that many real-life norms may contain as holding between
norm conditions and their effects.
Finally we must state that in using the context of normative positions, each basic constraint's
argument is pair of the action-type's time interval and the situation's time interval. For
example the following norm, identified by Norm001, means the agent Robosweeper has an
obligation to begin sweeping snow not later than 4 time units after a situation in which it
snowed.
NormPos(Robosweeper, obl(Sweep_Snow), Snowing, and(gt(B, E), le(B, disp(E,4))),
Norm001)
3.4 Normative Rules and Rule Reification
13
Description of the situation that appears in NT
Normative-Token (NT) if
In our formalism, norms or normative rules are represented as Logic programs in which the
heads are normative tokens and the bodies contain the description of norm's conditions which
are represented by situations. i.e.
Each normative token is an assertion that an agent is to conform to a normative fluent, in
some named situation (which is later described in the body of the clause), within some
temporal constraint.
This representation eliminates the need for a normative implication in the sense used by
Governatori et al(2005) and Artikis and Sergot(2010) because unlike these previous
formalisms in which a norm's condition and its effect are antecedent and consequent in an
implication. This is achieved by making the norm's condition, some situation s and the
norm's effect, some normative fluent, terms for the 5-tuple normative token(reification). Thus
when a situation arises whose description matches what appears in the body of the normative
rule, a normative token is inferred which is derived from the head of the normative rule.
This reification of both the condition and effect of norms is the main difference between this
formalism and that of Governatori et al. and Artikis and Sergot. For these both of these
formalisms, it is virtually impossible to reify a norm's conditions without resorting to
situations as we have done.
The last term of a normative token is an identifier for the normative rule that produced it.
This value assigned to this term is the identification for the normative rule from which the
normative token was inferred. Although the identifier only appears in the head of the
normative rule, it is in fact an attribute of the normative rule and it is the identifier for the
rule. Indeed every normative token derived from the same normative rule bears that same
identity. This is a special kind of reification for rules we shall call rule reification.
As we will see from the examples in the next section, it is often necessary to describe other
situations that have some qualitative relations with the situation being described.
3.5 Real life Norms
In this section we present our formalism's representation of some real life norms. It is a rather
trivial exercise to show that the formalism is able to represent all low-level norms used to test
a formalism such as that Stratulat et al (2001). We will return to that question later. At this
point our focus is on representing high level norms.
Norm 3.1
A teacher assigned to teach a class must arrive either on time or not later than 10 minutes
14
into the time of a class.
a, v, s,
NormPos(a, obl(arrive-at(v)), s, and(le(E, tplus(B,10)), ge(E, B)), OB101) if
Ǝ e.
Holds**(occurring(e), s)
EventType(e, Class)
Holds(venue(e, v), s)
Holds(role(a, Teach), s)).
Norm 3.2
Student must register for his/her courses in a semester within one month of the
commencement of the semester.
a, o, s.
NormPos(a, obl(register-for(a,sem))), s, and(and(ge(E, B), le(E,B, 30)), ge(B,B)), OB102) if
Ǝs1, s2.Holds**(studentship(a),s1)
ProcessType(sem, Semester)
Holds**(prog(sem), s)
Within(timeS(s), timeS(s1))
Holds(on-suspension(a, sem), s).
A prohibition is a norm that disallows an agent from carrying out an action of a certain type
within some time interval that has a temporal relation with the warranting situation
(condition) of the prohibition. The structure of a prohibition is the same as that of a normative
position except that the normative fluent is derived from the application of the pro function to
an action type.
The norm 3.3 and norm 3.4 illustrate examples of rules that infer such normative tokens.
Norm 3.3
Student must not be allowed to come in for an examination thirty (30) minutes after the
commencement of the examination.
a, s, e.
NormPos(a, pro(arrive(v)), s, gt(E, tplus(30)}, PR0103), if
Holds**(occuring(e), s)
EventType(e, Examination))
Holds(venue(e,v), s)
Holds(role(a, Candidate), s).
Norm 3.4
It is prohibited for members of university community to release confidential document of the
university to public domain without authorization.
15
a, doc, s,
Prohibition(a, pro(release(doc)), s, or(and(lt(B, E), ge(B, B)), and(le(E, E), gt(E, B))),
PRO014) if
Ǝs1,s2, u.
Holds**(alive(doc), s1)
Holds**(statusdoc(doc, Confidential),s)
Holds**(employ(u, a), s2)
University(u)
Owns(doc, u)
Subinterval(timeS(s), timeS(s1))
Subinterval(timeS(s), timeS(s2))
A permission is a norm that allows an action by an agent as a result of a certain situation
arising. In this formalization, we will treat permissions as the fluent negation of prohibitions
denoted by the function .
Examples of this is given in norm 3.5 and norm 3.6.
Norm 3.5
Members of the university community are permitted to put on their official identity card while
on duty.
a, s.
NormPos(a, pro( put_id-on(a)), s, and(ge(B, B), le(E, E)), PER010) if
Ǝu, s1
Holds**(onDuty(a), s)
Holds**(employ(a, u) s1)
University(u)
Within(timeS(s), timeS(s1)
Norm 3.6
Lecturer is permitted to give reading books on his assigned course to the student at the
beginning of lecture in a semester.
a, b, s.
NormPos(a, pro(give(b)), s, le(E, B), PER012) if
Ǝco.
Holds**(prog(co), s)
ProcessType(co, Course)
Holds(role(a, Teacher), s)
Holds(role(b, ReadingBook), s)
Having represented norms, it is important to discuss the representation of the validity of those
norms. Norm validity is determined by the date of enactment and . Let us take as an example
16
the constitutive norm from the formalization of the British Nationality Act 1981 (Segot et al.,
1986).
The norm any person born in the UK becomes a British citizen is a norm that was only valid
until 1981. A representation of that norm in the proposed language is presented thus:
x, s
NormPos(HMG, obl(grant-citizenship(x)), s, eq(E, B), NBB-1) if
Holds**(born-in(x, UK), s)
However, any agent interpreting that norm on behalf of Her Majesty's Government (HMG)
must be aware that the validity of the norm is from 1950 and 1981 i.e.
Enact(NBB-l, 1950)
Repeal(NBB-1, 1981)
As such it is only if the norm situation s happened within that interval of validity that the
norm NBB-1 is valid. Validity is an important condition for deciding norm violation as we
shall demonstrate.
3.5 Satisfying Temporal Constraints
Our attention now turns to determining when two time intervals satisfy a certain temporal
constraints and for this purpose, we will define constraint satisfaction rules (CSR) that will
determine whether an ordered pair of intervals satisfies a temporal constraint.
Intuitively, it is easy to see how the satisfaction of any of our basic constraints can translate
into logical relations that involve two time points. Similarly the satisfaction of a compound
constraint can be broken down into the satisfaction of constituent constraints.
CSRs for compound temporal constraints are recursive in the sense that they are composed
from other CSRs as illustrated by the following axiom:
Axiom 3.5.1
tc1, tc2, j, k.
Satisfy-Cons( j, k, and(tc1, tc2)) if
Two similar rules exist for the or function. That is embodied in axiom 5.2(a) and (b) below:
Axiom 3.5.2 (a)
j, k, tc1, tc2.
Satisfy-Cons( j, k, tc1)
Satisfy-Cons(j, k, tc2).
Satisfy-Cons( j, k, or(tc1, tc2)) if
Satisfy-Cons(j, k, tc1)
Axiom 3.5.2 (b)
j, k, tc1, tc2.
17
Satisfy-Cons(j, k, or(tc1, tc2)) if
Satisfy-Cons( j, k, tc2)
The other non-recursive rules handle basic temporal constraints. These rules represent a
simple translation of each of the temporal constraint functions into the equivalent relation.
The following axioms 5.3 to 5.6 are all examples of the lt constraint.
Axiom 3.5.3
j, k, t1, t2.
Satisfy-Cons(j, k, lt(tdisp(B, t1), disp(B, t2)) ) if
begin(j) + t1 < begin(k) +t2
Axiom 3.5.4
j, k.
Satisfy-Cons(j, k, lt(B, E)) if
begin( j) < end(k)
Axiom 3.5.5
j, k.
Satisfy-Cons(j, k, lt(E, B)) if
end(j) < begin( k)
Axiom 3.5.6
j, k.
Satisfy-Cons(j, k, lt(E, E)) if
end(j) < end(k)
The next section discusses norm how constraint satisfaction rules is used to infer norm
violations.
3.6 NORM CONFORMANCE AND VIOLATIONS
In order to make inferences about norm violation we need to know the existence of three
fundamental conditions. The first condition will be the existence of a normative token such as
an agent's obligation or prohibition to carry out or desist from carrying out an action. The
second condition will be the failure to conform to the requirement of the normative token.
The third condition is the validity of the normative rule at the time the condition held.
The condition for a norm to be valid with respect to a situation is for the situation to take
place while the norm's enactment remains in force. This is formalized as:
Axiom 3.6.0
norm-id, s.
Validwrt(norm-id, s) if
Enact-at(norm-id, t)
t <= begin(timeS(s))
t1. (t t1 end(timeS(s)) Repeal(norm-id, t1))
18
The conformance of an agent to a norm requiring an agent's obligation requires the existence
of an action of the expected type to be carried by the agent within the required time limit.
This is formalized by the following clause:
Axiom 3.6.1
a, norm-id, s.
Conform(a, norm-id, s) if
act-type, tc, act.
NormPos( a, obl( act-type), s, tc, norm-id)
Type-of-Action(act, act-type)
Actor(a, act)
Validwrt(norm-id, s)
Satisfy-Cons(timeA(act), timeS(s), tc).
The violation of an obligation takes place when in the occurrence of the situation within the
validity period of the rule that generated the normative token, the agent implicated is unable
to carry out the needed action within the required time constraint. That is formalized as the
following clause:
Axiom 3.6.2
a. norm-id, s.
Violate(a, norm-id, s) if
act-type, tc.
NormPos(a, obl(act-type), s, tc, norm-id)
Validwrt(norm-id, s)
act (Type-of-Action (act, a-type)
Actor(a, act)
Satisfy-Cons(timeA(act), timeS(s), tc) )
In the case of a prohibition, a violation takes place when in the case of the occurrence of the
situation, the implicated agent carries out the forbidden action during a time interval that
satisfies the constraint with the time of the situation. Such a violation is formalized thus:
Axiom 3.6.3
p, norm-id, s.
Violate(a, norm-id, s) if
act-type, tc, act.
NormPos( p, pro(act-type), s, tc, norm-id)
Type-of-Action(act, act-type)
Actor(a, act)
Validwrt(norm-id, s)
Satisfy-Cons(timeA(a), timeS(s), tc).
Similarly conformance to a norm that is a prohibition is the absence of the prescribed action
19
within the prescribed time constraint. That is contained in the following axiom:
Axiom 3.6.4
a. norm-id, s.
Conform(a, norm-id, s) if
act-type, tc.
NormPos(a, pro(act-type), s, tc, norm-id)
Validwrt(norm-id, s)
act (Type-of-Action (act, act-type)
Actor(a, act)
Satisfy-Cons(timeA(act), timeS(s), tc) )
act-type, tc. NormPos(a, pro(act-type), s, tc, norm-id)
However, an agent needs not do anything else to conform to a permission to either carry out
or avoid an action. That idea is formalized in Axiom 3.6.5.
Axiom 3.6.5
a, norm-id, s.
Conform(a, norm-id, s) if
Axiom 3.6.6
a, norm-id, s.
Conform(a, norm-id, s) if
This approach contrasts with timed violations that was implemented by Stratulat et al[21]. In
their work, a violation of an obligation is said to have taken place at a time t, if an agent has
an obligation to carry out an action during the time interval (t 1, t2) and as at time t which is
later than t2 the action has not yet been taken. We believe that this approach of theirs is
flawed because a violation is best understood as occurring under some particular
circumstances rather than as occurring at some particular time. For example it is more
meaningful to infer that:
act-type, tc.
NormPos(a, obl(act-type), s, tc, norm-id)
Mr X violated the of promptness obligation with regard to the CSC777 class of 1
April 2016
than to infer that:
Mr X is reckoned to have violated an obligation to report at the Faculty Lecture
theatre between 10 and 10:10 on 1 April 2016.
4. COMPARISON WITH EXISTING FORMALISMS
The norm representation formalism that is closest to that developed in the paper is the one
offered by Panagiotidi, Nieves and Vazquez-Salceda(2009). In their representation, a norm is
a 6-ary relation which consists of:
20
The actor
The (deontic) modality, which may either be an obligation or a permission.
The activating condition for the norm
The maintenance condition for the norm
Deontic Statement and,
The deactivating condition
The activating condition is the equivalent to the norm's precondition. The deontic statement
is a description of the state change that the agent under obligation is expected to effect. The
deactivating condition is a sign that the state change has been effected. A norm becomes
activated when the activating condition becomes true and remains activated until the
deactivating condition becomes true. The deactivating condition is satisfied when the agent
has successfully effected the requirement of the norm. The maintenance condition captures
the deadline for an active norm to become deactivated. Therefore a violation occurs when an
active norm remains active when its maintenance condition has become untrue (i.e. the
deadline has been passed while the deactivating condition is yet unsatisfied by the agent).
However there are significant differences between the approach presented in this paper and
their approach. Firstly, the representation in this paper does not only capture activating
conditions; rather, it captures named instances of their occurrences as situations. So that it is
possible to have named instances of an event or process type as a situation. That is not the
case with the Panagiotidi et al's representation. Again their inference of violation is based on
the observing changes in condition and not necessarily on the occurrence or non-occurence of
actions.
The norm in Panagiotidi et al about the obligation of a repair company rc to repair a car
within 4 days of taking the car to the garage will be rendered thus:
NormPos(rc, obl(repair(car17)), s, and(ge(B, B), le(E, disp(E, 4))), OBRC1) if
Holds(arrive(car17, Garage), s)
We wish to note that our representation allows us to state the car repair that will satisfy the
new obligation cannot precede the arrival of the car at the garage. This way we can avoid
mistaking a previous repair done on car17, with the repair required for a new visit to the
garage. This is an advantage that our formalism has over formalisms such as King et al
(2017) and Panagiotidi et al (2009) that depend on deadlines.
Hashmi, Governatori and Moe(2016) present a taxonomy of obligations that expands on a
simple categorization of obligations due to Governatori (2010). A maintenance obligation is
one in which a norm's effect must continue throughout the time of the obligation (or some
specified situation). The norm represented by example 2 in this paper is one such norm.
It is also the case that 4.1 is an example of Governatori's achievement obligation.
21
An achievement obligation is one in which an actor is expected to carry out the action at least
once within the time limit. An example of this which can be represented in the language of
this paper is:
The Prime Minister must pay the Queen a visit during the parliamentary break.
The representation for this in the representation language presented in this paper would treat
the parliamentary break as an event whose occurring constitutes the warranting situation and
the agent whose visitation action to the Queen must fall within that situation is the Prime
Minister.
Finally in this section, in order to demonstrate that our formalism is versatile enough to
represent low level norms, we present a representation for a norm from Stratulat et al (2001)
that obliges a taxpayer to pay up within the first 31 days of the year, thus:
NormPos(a, obl(pay-tax), s, or(eq(E, B), le(E, disp(B, 31)), OBLTAX1) if
y( begin(timeS(s)) = (January, 1 y)
end(times(s)) = (December, 31, y))
Holds(taxpayer(a), s).
The good side of our representation is that any violation can report on the year for which a
violation took place and the identifier for the particular tax code violated, which in
representation is OBLTAX1.
5. CONCLUSION
This paper has identified two broad categorizations for prescriptive norms in the literature.
High level norm representations capture the norm's condition and effect as well as the
temporal constraints between them. The norm's condition is treated as a situation warranting
the need for an agent to either carry out or refrain from carrying out an action of a certain
type. The effect of the norm is the action that the agent is obliged or permitted to take or
prohibited from taking.
This contrasts with other approaches, which accounts a norm as the obligation or permission
to take an action or the prohibition from taking an action within a particular time interval.
The key disadvantage of low-level norms is that an agent's autonomy is curtailed if s/he
needs an external prescription in order to know what actions s/he is expected or allowed to
take at specific times. On the other hand, an agent with a high level norm representation of
norms can determine from that representation, at what times an actions they have an
obligation to act or prohibited from acting. The difference between the two approaches is
somewhat like that between low-level programming and high-level programming.
22
The major disadvantage of high level norms as they appear in the existing literature is that
generalized violation inference is practically impossible. The formalism for high-level norm
representation presented in this paper fixes that problem by relaxing the strict normative
implication relationship between a norm's condition and its effect as done by Governatori et
al(2005) and Artikis and Sergot(2010). The formalism also makes it possible for norms to be
identified through rule reification.
Another key advantage of low level norms it easily enables an agent to detect when there are
contradictory norms as shown by Artikis(2003). A contradiction occurs when there two
different norms, one of which obliges an agent to carry out an action of a certain type, and the
other prohibits actions of the same type at around the same time. In order to detect
contradictory norms, there is a need for a logical mechanism to convert high level norms into
low level norms on an as needed basis.
REFERENCES
International Joint Conference on Artificial Intelligence, pp 1177-1182, Morgan
6
Allen, J F(1984) A General Model of Action and Time. Artificial Intelligence 23(2) : 123-
154 , July 1984.
Artikis, A (2003) Executable Specification of Open Norm-Governed Societies, PhD thesis,
Dept of Electrical Engineering, Imperial College, University of London.
Artikis A and Sergot M J (2010) Executable Specification of Open Multi-agent Systems
Logic Journal of the IGPL 18(1): 31-65
Galton, A.(1991). Reified Temporal Theories and How to Unreify them. In Proceedings of
the 12th
Kaufmann 1991.
Galton A.(2016) The Ontology of Time and Processes. Presentation at the Third
International School on Applied Ontology, Bozen-Bolzano, Italy, June 27 –July 1, 2016.
Governatori, G. (2010) Law, Logic and Business Processes. In Proceedings of the Third
International Workshop on Requirement Engineering in Law, (RELaw 2010), pp 1 -10
Sydney Australia
Governatori, G., Rotolo A. and Sartor G.(2005) Temporalised Normative Positions in
Defeasible Logic In Sartor, G.(ed.) The Tenth International Conference on AI and Law,
Proceedings of the Conference, pp. 25-34, June 6-11, 2005, Bologna , Italy.
Governatori, G. (2010) Law, Logic and Business Processes. In Proceedings of the Third
International Workshop on Requirement Engineering in Law, (RELaw 2010), pp 1 -10
Sydney Australia.
Hashmi M., Governatori G. and Moe, T. W. (2016), Normative Requirements for Regulatory
23
Compliance: An Abstract Formal Framework. Information Systems Frontiers, 18(3): 429-
455, 2016.
R. Marín and G. Sartor(1999) Time and Norms: A Formalisation In The Event Calculus. In
Proceedings of Conference on Artificial Intelligence and Law (ICAIL), pages 90–100. ACM
Press, 1999.
King T.C., De Vos M., Dignum V., Jonker C.M., Li, T. and Riemsdijk,B.(2017) Automated
Multi-level Governance Compliance Checking. Autonomous Agents and Multi-Agent Systems
31(6): 1283-1343, 2017.
Koomen, J. (1989), Reasoning about Recurrence, PhD thesis, Dept of Computer Science,
University of Rochester, Rochester, New York.
Panagiotidi S, Nieves, J.C. and Vazquez-Salceda, J. (2009). A framework to model norm
dynamics in Answer Set Programming. Proceedings of the Second Multi-Agent Logics,
Languages, and Organisations Federated Workshops, Turin, Italy, September 7-10, 2009.
http://ceur-ws.org/Vol-494/famaspaper8.pdf
Pinto J. A. (2004) Temporal Reasoning in the Situation Calculus, PhD Dissertation,
University of Toronto, Ont., Canada.
Sadri, F, Stathis, K. and F. Toni (2006) Normative KGP Agents Computational &
Mathematical Organization Theory 12(2-3): 101-126
L.K. Schubert(2000). The Situations We Talk About. In Jack Minker (ed) Logic-Based
Artificial Intelligence, Volume 597, The Springer International Series in Engineering and
Computer Science pp 407-439
M. Sergot(1990) The representation of law in computer programs: a survey and comparison.
In T. Bench-Capon,(Ed), Knowledge Based Systems and Legal Applications. Academic Press,
1990.
24
|
1910.13852 | 1 | 1910 | 2019-10-30T13:51:47 | Linear Speedup in Saddle-Point Escape for Decentralized Non-Convex Optimization | [
"cs.MA",
"cs.LG",
"math.OC",
"stat.ML"
] | Under appropriate cooperation protocols and parameter choices, fully decentralized solutions for stochastic optimization have been shown to match the performance of centralized solutions and result in linear speedup (in the number of agents) relative to non-cooperative approaches in the strongly-convex setting. More recently, these results have been extended to the pursuit of first-order stationary points in non-convex environments. In this work, we examine in detail the dependence of second-order convergence guarantees on the spectral properties of the combination policy for non-convex multi agent optimization. We establish linear speedup in saddle-point escape time in the number of agents for symmetric combination policies and study the potential for further improvement by employing asymmetric combination weights. The results imply that a linear speedup can be expected in the pursuit of second-order stationary points, which exclude local maxima as well as strict saddle-points and correspond to local or even global minima in many important learning settings. | cs.MA | cs | LINEAR SPEEDUP IN SADDLE-POINT ESCAPE FOR DECENTRALIZED NON-CONVEX
OPTIMIZATION
Stefan Vlaski and Ali H. Sayed
School of Engineering, ´Ecole Polytechnique F´ed´erale de Lausanne
ABSTRACT
Under appropriate cooperation protocols and parameter choices,
fully decentralized solutions for stochastic optimization have been
shown to match the performance of centralized solutions and re-
sult in linear speedup (in the number of agents) relative to non-
cooperative approaches in the strongly-convex setting. More re-
cently, these results have been extended to the pursuit of first-order
stationary points in non-convex environments.
In this work, we
examine in detail the dependence of second-order convergence
guarantees on the spectral properties of the combination policy for
non-convex multi agent optimization. We establish linear speedup
in saddle-point escape time in the number of agents for symmetric
combination policies and study the potential for further improve-
ment by employing asymmetric combination weights. The results
imply that a linear speedup can be expected in the pursuit of second-
order stationary points, which exclude local maxima as well as
strict saddle-points and correspond to local or even global minima
in many important learning settings.
Index Terms -- Non-convex optimization, saddle-point, second-
order stationarity, minima, decentralized algorithm, centralized al-
gorithm, diffusion strategy.
1. INTRODUCTION AND RELATED WORK
We consider a collection of K agents, where each agent k is
equipped with a local stochastic cost function:
d∇J
which is obtained by employing a weighted combination of instanta-
neous approximations using all K realizations available at time i−1.
This construction requires the evaluation of K (stochastic) gradients
per iteration. If computational constraints limit the number of gradi-
ent evaluations per iteration to one, we can instead randomly sample
an agent location k from the available data and let:
c,1
(wcent
i−1 ) = ∇Qk(wcent
i−1 ; xk,i−1), with prob. pk,
(5)
The evident drawback of such a simplified centralized strategy is that
only one sample is processed and a large number of samples is dis-
carded at every iteration. We can hence expect the construction (4)
to result in better performance relative to the simplified choice (5).
When communication constraints limit the exchange of information
among agents, we can instead appeal to decentralized strategies. For
the purpose of this work, we shall focus on the standard diffusion
strategy, which takes the form:
φk,i = wk,i−1 −µ∇Qk(wk,i−1; xk,i−1)
(6a)
(6b)
where aℓk denote convex combination coefficients satisfying:
aℓk = 1,
aℓk = 0 if ℓ /∈ Nk
(7)
aℓkφℓ,i
wk,i =
NXℓ=1
aℓk ≥ 0, Xℓ∈Nk
Jk(w) , E Qk(w; xk)
(1)
where w ∈ RM denotes a parameter vector and xk denotes the
random data at agent k. We construct the global cost function:
The symbol Nk denotes the set of neighbors of agent k. When the
graph is strongly-connected, it follows from the Perron-Frobenius
theorem that the combination matrix A has a spectral radius of one
and a single eigenvalue at one with corresponding eigenvector [2]:
J(w) ,
pkJk(w)
(2)
Ap = p, 1
Tp = 1,
pk > 0
(8)
KXk=1
9
1
0
2
t
c
O
0
3
]
A
M
.
s
c
[
1
v
2
5
8
3
1
.
0
1
9
1
:
v
i
X
r
a
one, i.e.PK
where the pk ≥ 0 denote convex combination weights that add up to
k=1 pk = 1. When data realizations for xk can be aggre-
gated at a central location, descent along the negative gradient of (2)
can be approximated by means of a centralized stochastic gradient
algorithm of the form [1, 2]:
Comparing the diffusion strategy (6a) -- (6b) to the centralized con-
structions (4) or (5), we observe that the adaptation step (6a) carries
the same complexity per agent as the simplified construction (5).
However, since these computations are performed at K agents in
parallel, and the information is diffused over the network through the
combination step (6b), we expect the diffusion strategy (6a) -- (6b) to
outperform the simplified centralized strategy (5) and more closely
match the full construction (4). In fact, the spectral properties (8) of
the combination weights (7) allow us to establish the following rela-
wcent
i = wi−1 −µd∇J (wcent
i−1 )
(3)
where d∇J(·) denotes a stochastic gradient approximation con-
structed at time i − 1. One possible construction is to let:
tion for the weighted network mean wc,i ,PK
k=1 pk wk,i [2, 3]:
c,K
(wcent
i−1 ) ,
pk∇Qk(wcent
i−1 ; xk,i−1)
(4)
wc,i = wc,i−1 −µ
pk∇Qk(wk,i−1; xk,i−1)
(9)
KXk=1
d∇J
KXk=1
This work was supported in part by NSF grant CCF-1524250.
Emails:{stefan.vlaski, ali.sayed}@epfl.ch.
which almost corresponds to the centralized recursion (3) -- (4) with
the full gradient approximation (4) with the only difference being
that the stochastic gradients are evaluated at the individual iterates
wk,i−1 instead of the weighted network centroid wc,i−1. So long
as the iterates wk,i−1 cluster around the network centroid, and under
appropriate smoothness conditions on the (stochastic) gradients, it is
hence to be expected that the network centroid (9) will match the
performance of the full gradient approximation (4). This intuition
has been studied in great detail and formalized for strongly convex
cost functions, establishing that all iterates wk,i in (6a) -- (6b) will
actually match the centralized full gradient approximation (4) both
in terms of convergence rate [3] and steady-state error [4], which
implies a linear improvement over the simplified construction (5)
in terms of the number of agents [2] when employing a symmetric
combination policy for which pk = 1
K .
More recently, these results have been extended to the pursuit
of first-order stationary points in non-convex environments [5, 6] for
consensus and the exact diffusion algorithm [7]. First-order station-
ary points can include saddle-points and even local maxima and can
generate a bottleneck for many optimization algorithms and problem
formulations [8]. Hence, the purpose of this work is is to establish
that linear speedup can also be expected in the escape from saddle-
points and pursuit of second-order stationary points for non-convex
optimization problems. To this end, we refine and exploit recent re-
sults in [9, 10].
1.1. Related Works
Strategies for decentralized optimization include incremental strate-
gies [11], and decentralized gradient descent (or consensus) [12], as
well as the diffusion algorithm [2, 3, 13]. A second class of strate-
gies is based on primal-dual arguments [7, 14 -- 18]. While most of
these algorithms are applicable to non-convex optimization prob-
lems, most performance guarantees in non-convex environments are
limited to establishing convergence to first-order stationary points,
i.e., points where the gradient is equal to zero [5, 6, 19 -- 21].
Landscape analysis of commonly employed loss surfaces has
uncovered that in many important settings such as tensor decomposi-
tion [22], matrix completion [23], low-rank recovery [24], as well as
certain deep learning architectures [25], all local minima correspond
to global minima and all other first-order stationary points have a
strict-saddle property, which states that the Hessian matrix has at
least one negative eigenvalue. These results have two implications.
First, while first-order stationarity is a useful result in the sense that it
ensures stability of the algorithm, even in non-convex environments,
it is not sufficient to guarantee satisfactory performance, since first-
order stationary points include strict saddle-points, which need not
be globally or even locally optimal. On the other hand, establishing
the escape from strict saddle-points, is sufficient to establish conver-
gence to global optimality in all of these problems.
These observations have sparked a number of works examin-
ing second-order guarantees of local descent algorithms. Strategies
for the escape from saddle-points can generally be divided into one
of two classes. First, since the Hessian at every strict-saddle point,
by definition, contains at least one negative eigenvalue, the descent
direction can be identified by directly employing the Hessian ma-
trix [26] or through an intermediate search for the negative curva-
ture direction [27, 28]. The second class of strategies leverages the
fact that perturbations in the initialization [29] or the update direc-
tion [22, 30 -- 32] cause iterates of first-order algorithms to not get
"stuck" in strict saddle-points, which can be shown to be unstable.
Recently these results have been extended to decentralized optimiza-
tion with deterministic gradients and random initialization [33] as
well as stochastic gradients with diminishing step-size and decay-
ing additive noise [34] as well as constant step-sizes [9, 10]. We
establish in this work, that the saddle-point escape time of the dif-
fusion strategy (6a) -- (6b) decays linearly with the number of agents
in the network when symmetric combination policies are employed
and show how asymmetric combination policies can result in fur-
ther improvement when agents have access to estimates of varying
quality.
2. MODELING CONDITIONS
We shall be employing the following common modeling condi-
tions [2, 22, 31, 34]. See [9, 10] for a discussion.
Assumption 1 (Smoothness). For each k, the gradient ∇Jk(·) is
Lipschitz, namely, for any x, y ∈ RM :
k∇Jk(x) − ∇Jk(y)k ≤ δkx − yk
(10)
Furthermore, Jk(·) is twice-differentiable with Lipschitz Hessian:
k∇2Jk(x) − ∇2Jk(y)k ≤ ρkx − yk
(11)
For each pair of agents k and ℓ, the gradient disagreement is
bounded, namely, for any x ∈ RM :
k∇Jk(x) − ∇Jℓ(x)k ≤ G
(12)
Assumption 2 (Gradient noise process). For each k, the gradient
noise process is defined as
sk,i(wk,i−1) =d∇J k(wk,i−1) − ∇Jk(wk,i−1)
and satisfies
(13)
(14a)
(14b)
E {sk,i(wk,i−1)F i−1} = 0
E(cid:8)k sk,i(wk,i−1)k4F i−1(cid:9) ≤ σ4
k
where we denote by F i the filtration generated by the random pro-
cesses wk,j for all k and j ≤ i and for some non-negative constants
σ4
k. We also assume that the gradient noise processes are pairwise
uncorrelated over the space conditioned on F i−1.
Assumption 3 (Lipschitz covariances). The gradient noise process
has a Lipschitz covariance matrix, i.e.,
Rs,k(wk,i−1) , Ensk,i(wk,i−1)sk,i(wk,i−1)TF i−1o (15)
satisfies
kRs,k(x) − Rs,k(y)k ≤ βRkx − ykγ
(16)
for some βR and 0 < γ ≤ 4.
We shall also make the simplifying assumption.
Assumption 4 (Gradient noise lower bound). The gradient noise
covariance Rs,k(x) at every agent is bounded from below:
Rs,k(x) ≥ σℓ,kI
(17)
This condition can be loosened significantly by requiring a gradient
noise component to be present only in the vicinity of strict saddle-
points and only in the local descent direction, see e.g. [10, 31]. Nev-
ertheless, the simplified condition can always be ensured for exam-
ple by adding a small amount of isotropic noise, similar to [22, 30]
and will be sufficient for the purpose this work.
3. CONVERGENCE ANALYSIS
3.1. Noise Variance Relations
The performance guarantees established in [9, 10] depend on the sta-
tistical properties of the weighted gradient noise term:
si ,
KXk=1
pk sk,i(wk,i−1)
(18)
Under assumptions 1 -- 4, we can refine the bounds from [9]:
Lemma 1 (Variance Bounds). Under assumptions 1 -- 4 we have:
p2
kσ2
k
E(cid:8)ksik2F i−1(cid:9) ≤
KXk=1
KXk=1
k,ℓ! I ≤ E si s
p2
kσ2
T
i ≤ KXk=1
k! I
p2
kσ2
(19)
(20)
and will focus on the dependence on the combination policy further
below. Adjusting the theorems in [9, 10] to account for the variance
bounds (19) -- (20), we obtain:
Theorem 1 (Network disagreement (4th order)). Under assump-
tions 1 2, the network disagreement is bounded after sufficient itera-
tions i ≥ io by:
4
4
(1 − kJ T
ǫ k)4 kV T
Rk
E(cid:13)(cid:13)(cid:13)Wi −(cid:16)1pT ⊗ I(cid:17) Wi(cid:13)(cid:13)(cid:13)
≤ µ4 kVLk4(cid:13)(cid:13)J T
ǫ(cid:13)(cid:13)4
k(cid:17) + o(µ4)
where(cid:13)(cid:13)J T
ǫ(cid:13)(cid:13) = λ2(A) + ǫ ≈ λ2(A) denotes the mixing rate of the
L(cid:9), io = log(cid:0)o(µ4)(cid:1)/log(cid:0)(cid:13)(cid:13)J T
ǫ(cid:13)(cid:13)(cid:1) and o(µ4)
ǫ = col(cid:8)1T, V T
adjacency matrix, A = VǫJ V −1
V −1
denotes a term that is higher in order than µ4.
K 2(cid:16)G4 + max
ǫ with Vǫ = row {p ⊗ I, VR} and
(24)
σ4
k
Proof. The argument is an adjustment of [9, Theorem 1].
Proof. Relations (19) and (20) follow from the pairwise uncorrelat-
edness condition in assumption 2 after cross-multiplying.
From (19) we observe that the average noise term (18) driv-
ing the network centroid experiences a variance reduction. Specif-
ically, in the case when pk = 1/K and σk = σ we would obtain
E(cid:8)ksik2F i−1(cid:9) ≤ σ2/K. This K-fold reduction in gradient noise
variance is at the heart of the improved performance established for
strongly-convex costs [2] and in the pursuit of first-order stationary
points [5]. We shall establish in the sequel that this improvement also
holds in the time required to escape from undesired saddle-points.
3.2. Space Decomposition
This result ensures that the entire network clusters around the
network centroid wc,i after sufficient iterations, allowing us to lever-
age it as a proxy for all agents.
Theorem 2 (Descent relation). Beginning at wc,i−1 in the large
gradient regime G, we can bound:
E {J(wc,i) wc,i−1 ∈ G}
≤ E {J(wc,i−1) wc,i−1 ∈ G} − µ2 c2
π
+
O(µ3)
πG
i−1
(25)
as long as πG
stants are listed in definition 1.
i−1 = Pr {wc,i−1 ∈ G} 6= 0 where the relevant con-
Definition 1 (Sets). The parameter space RM is decomposed into:
Proof. The argument is an adjustment of [9, Theorem 2].
1
c2
c1(cid:18)1 +
G ,(cid:26)w : k∇J(w)k2 ≥ µ
π(cid:19)(cid:27)
H ,nw : w ∈ GC , λmin(cid:0)∇2J(w)(cid:1) ≤ −τo
M ,nw : w ∈ GC , λmin(cid:0)∇2J(w)(cid:1) > −τo
k(cid:17) = O(cid:16)PK
where τ is a small positive parameter, 0 < π < 1 is a pa-
rameter to be chosen, c1 = 1
2 (1 − 2µδ) = O(1) and c2 =
δ
kσ2
k=1 p2
We also define the probabilities πG
i
Pr {wc,i ∈ H} and πM
have πG
k(cid:17). Note that GC = H ∪ M.
,
, Pr {wc,i ∈ M}. Then for all i, we
2(cid:16)PK
, Pr {wc,i ∈ G}, πH
i
(21)
(22)
(23)
k=1 p2
i + πM
i = 1.
i + πH
kσ2
i
Points in the complement of G have small gradient norm and hence
correspond to approximately first-order stationary points. These
points are further classified into strict-saddle points H, where the
Hessian has a significant negative eigenvalue, and second-order sta-
tionary points M. Pursuit of second-order stationary points requires
descent for points in G as well as H.
3.3. Performance Guarantees
Due to space limitations, we forego a detailed discussion on the
derivation of the second-order guarantees of the diffusion algo-
rithm (6a) -- (6b) and refer the reader to [9, 10]. We instead briefly
list the guarantees resulting from the variance bounds (19) -- (20)
Theorem 3 (Descent through strict saddle-points). Suppose
πH
i⋆ 6= 0, i.e., wc,i⋆ is approximately stationary with significant
negative eigenvalue. Then, iterating for is iterations after i⋆ with
log(cid:18)2M
is =
k)
k σ2
k,ℓ(cid:17)
σ2
(PK
(cid:16)PK
k=1 p2
k=1 p2
O(µτ )
k
+ 1(cid:19)
(26)
guarantees
E {J(wc,i⋆+is ) wc,i⋆ ∈ H}
≤ E {J(wc,i⋆ ) wc,i⋆ ∈ H} −
µ
2
M KXk=1
k! +
p2
kσ2
o(µ)
πH
i⋆
(27)
Proof. The argument is an adjustment of [10, Theorem 1].
Theorem 2 ensures descent in one iteration as long as the gra-
dient norm is sufficiently large, while 3 ensures descent even for
first-order stationary points, as long as the Hessian has a negative
eigenvalue in a number of iterations is that can be bounded. This
ensures efficient escape from strict saddle-points. We conclude:
Theorem 4. For sufficiently small step-sizes µ, we have with prob-
ability 1 − π, that wc,io ∈ M, i.e.,
k∇J(wc,io )k2 ≤ O µ KXk=1
k!!
p2
kσ2
(28)
and λmin(cid:0)∇2J(wc,io )(cid:1) ≥ −τ in at most io iterations, where
2 (J(wc,0) − J o)
io ≤
µ2δ(cid:16)PK
k=1 p2
kσ2
is
k(cid:17) π
Proof. The argument is an adjustment of [10, Theorem 2].
4. COMPARATIVE ANALYSIS
4.1. Step-Size Normalization
Note that in Theorem 4, both the limiting accuracy (28) and conver-
gence rate (29) depend on the combination policy and network size
k. To facilitate comparison, we shall normalize
kσ2
the step-size in (6a):
throughPK
k=1 p2
µ′ ,
µ
k=1 p2
kσ2
k
PK
Under this setting, Theorem 4 ensures a point io satisfying
k∇J(wc,io )k2 ≤ O (µ)
io ≤
2 (J(wc,0) − J o)
and λmin(cid:0)∇2J(wc,io )(cid:1) ≥ −τ in at most:
KXk=1
+ 1(cid:19)
log(cid:18)2M
(PK
(cid:16)PK
iterations with
k)
kσ2
k,ℓ(cid:17)
σ2
µ2δπ
is =
k=1 p2
k=1 p2
O(µτ )
k
k! is
p2
kσ2
KXk=1
k!
p2
kσ2
(30)
(31)
(32)
(33)
(29)
the derivation, we assume that the gradient noise is approximately
isotropic, i.e., σ2
k,ℓ so that (33) can be simplified to:
k ≈ σ2
is ≈ O PK
k=1 p2
kσ2
k
µτ
!
(35)
Then, we can formulate the following optimization problem to mini-
mize the escape time is over the space of valid combination policies:
min
A
KXk=1
p2
kσ2
k s.t. aℓk ≥ 0, Xℓ∈Nk
aℓk = 1, aℓk = 0 if ℓ /∈ Nk,
Ap = p, 1
Tp = 1,
pk > 0.
(36)
This precise optimization problem has appeared before in the pursuit
of asymmetric combination policies that minimize the steady-state
error of the diffusion strategy (6a) -- (6b) in strongly-convex environ-
ments [2]. Its solution is available in closed form and can even be
pursued in a decentralized manner, requiring only exchanges among
neighbors [2].
Theorem 6 (Metropolis-Hastings Combination Policy [2]). Un-
der the step-size normalization (30), the asymmetric Metropolis-
Hastings combination policy minimizes the approximate saddle-
point escape time (35). It takes the form:
ℓk =(
ao
ℓ } ,
σ2
k
max{nkσ2
k,nℓσ2
1 −Pm∈Nk\{k} ao
mk,
ℓ ∈ Nk,
ℓ = k.
(37)
where nk = Nk denotes the size of the neighborhood of agent k.
5. SIMULATIONS
We construct a sample landscape to verify the linear speedup in the
size of the network indicated by the analysis in this work. The loss
function is constructed from a single-layer neural network with a
linear hidden layer and a logistic activation function for the output
layer. Penalizing this architecture with the cross-entropy loss gives:
J(w1, W2) = E log(cid:16)1 + e−γwT
1 W2 h(cid:17) +
ρ
2
kw1k2 +
ρ
2
kW2k2
F
(38)
where w1 and W2 denote the weights of the individual layers, h ∈
RM denotes the feature vector, and γ ∈ {±1} is the class vari-
able. It can be verified that this loss has a single strict saddle-point at
w1 = W2 = 0 and global minima in the positive and negative quad-
rant, respectively [10]. We show the evolution of the function value
at the network centroid under the step-size normalization rule (30)
and observe a linear speedup in K, consistent with (34) while not-
ing no significant differences in steady-state performance, which is
consistent with (31).
Fig. 1: Linear speedup in saddle-point escape time.
Note that the normalization of the step-size causes (31) to become in-
dependent of(cid:16)PK
k=1 p2
kσ2
k(cid:17), allowing for the fair evaluation of (32)
and (33) as a function of the number of agents.
4.2. Linear Speedup Using Symmetric Combination Weights
When the combination matrix A is symmetric, i.e., A = AT, it fol-
lows that pk = 1
K [2]. For simplicity, in this section, we shall also
assume a uniform data profile for all agents, i.e., that σk = σ and
σℓ,k = σℓ for all k. We obtain:
Theorem 5 (Linear Speedup for Symmetric Policies). Under the
step-size normalization (30), and for symmetric combination poli-
cies A = AT with the uniform data profile σ2
ℓ,k = σ2
ℓ
for all k, the escape time simplifies to:
k = σ2 and σ2
log(cid:16)2M σ2
σ2
ℓ
O(µτ )
+ 1(cid:17)
σ2
K
is =
= O(cid:18) 1
µτ K(cid:19)
(34)
Proof. The result follows immediately after cancellations.
4.3. Benefit of Employing Asymmetric Combination Weights
In this subsection, we show how employing asymmetric combina-
tion weights can be beneficial in terms of the time required to escape
saddle-points when the data profile across agents is no longer uni-
form. In particular, we will no longer require the upper and lower
bounds σ2
ℓ,k to be common for all agents, and no longer re-
quire the combination policy to be symmetric. Instead, to simplify
k and σ2
6. REFERENCES
[1] B. T. Polyak, Introduction to Optimization, Optimization Soft-
ware, 1997.
[2] A. H. Sayed, "Adaptation, learning, and optimization over net-
works," Foundations and Trends in Machine Learning, vol. 7,
no. 4-5, pp. 311 -- 801, July 2014.
[3] J. Chen and A. H. Sayed, "On the learning behavior of adaptive
networks - Part I: Transient analysis," IEEE Transactions on
Information Theory, vol. 61, no. 6, pp. 3487 -- 3517, June 2015.
[4] J. Chen and A. H. Sayed, "On the learning behavior of adaptive
networks -- Part II: Performance analysis," IEEE Transactions
on Information Theory, vol. 61, no. 6, pp. 3518 -- 3548, June
2015.
[5] X. Lian, C. Zhang, H. Zhang, C.-J. Hsieh, W. Zhang, and
J. Liu, "Can decentralized algorithms outperform centralized
algorithms? A case study for decentralized parallel stochastic
gradient descent," in Advances in Neural Information Process-
ing Systems 30, pp. 5330 -- 5340. 2017.
[6] H. Tang, X. Lian, M. Yan, C. Zhang, and J. Liu, "d2: Decen-
tralized training over decentralized data," in Proceedings of
the 35th International Conference on Machine Learning, 2018,
vol. 80, pp. 4848 -- 4856.
[7] K. Yuan, B. Ying, X. Zhao, and A. H. Sayed, "Exact diffusion
for distributed optimization and learningPart I: Algorithm de-
velopment," IEEE Transactions on Signal Processing, vol. 67,
no. 3, pp. 708 -- 723, Feb 2019.
[8] S. S. Du, C. Jin, J. D. Lee, M. I. Jordan, B. P´oczos, and
A. Singh, "Gradient descent can take exponential time to es-
cape saddle points," in Proceedings of the 31st International
Conference on Neural Information Processing Systems, 2017,
pp. 1067 -- 1077.
[9] S. Vlaski and A. H. Sayed, "Distributed learning in non-convex
environments -- Part I: Agreement at a Linear rate," submitted
for publication, available as arXiv:1907.01848, July 2019.
[10] S. Vlaski and A. H. Sayed, "Distributed learning in non-convex
environments -- Part II: Polynomial escape from saddle-points,"
submitted for publication, available as arXiv:1907.01849, July
2019.
[11] D. P. Bertsekas, "A new class of incremental gradient methods
for least squares problems," SIAM J. Optim., vol. 7, no. 4, pp.
913 -- 926, April 1997.
[12] S. SundharRam, A. Nedic, and V. V. Veeravalli, "Distributed
stochastic subgradient projection algorithms forconvex opti-
mization," Journal of Optimization Theory and Applications,
vol. 147, no. 3, pp. 516 -- 545, Dec 2010.
[13] S. Vlaski, L. Vandenberghe, and A. H. Sayed, "Regularized
diffusion adaptation via conjugate smoothing," available as
arXiv:1909.09417, September 2019.
[14] D. Jakovetic, J. Xavier, and J. M. F. Moura,
"Coopera-
tive convex optimization in networked systems: Augmented
lagrangian algorithms with directed gossip communication,"
IEEE Transactions on Signal Processing, vol. 59, no. 8, pp.
3889 -- 3902, Aug 2011.
[15] J. C. Duchi, A. Agarwal, and M. J. Wainwright, "Dual aver-
aging for distributed optimization: Convergence analysis and
network scaling," IEEE Transactions on Automatic Control,
vol. 57, no. 3, pp. 592 -- 606, March 2012.
[16] M. Jaggi, V. Smith, M. Tak´ac, J. Terhorst, S. Krishnan, T. Hof-
mann, and M. I. Jordan, "Communication-efficient distributed
dual coordinate ascent,"
in Proc. International Conference
on Neural Information Processing Systems, Montreal, Canada,
2014, pp. 3068 -- 3076.
[17] K. I. Tsianos and M. G. Rabbat, "Distributed dual averag-
ing for convex optimization under communication delays," in
Proc. American Control Conference (ACC), Montreal, Canada,
June 2012, pp. 1067 -- 1072.
[18] W. Shi, Q. Ling, G. Wu, and W. Yin,
"EXTRA: An ex-
act first-order algorithm for decentralized consensus optimiza-
tion," SIAM Journal on Optimization, vol. 25, no. 2, pp. 944 --
966, 2015.
[19] P. Di Lorenzo and G. Scutari, "NEXT: in-network nonconvex
optimization," IEEE Transactions on Signal and Information
Processing over Networks, vol. 2, no. 2, pp. 120 -- 136, June
2016.
[20] T. Tatarenko and B. Touri, "Non-convex distributed optimiza-
tion," IEEE Transactions on Automatic Control, vol. 62, no. 8,
pp. 3744 -- 3757, Aug. 2017.
[21] Y. Wang, W. Yin, and J. Zeng, "Global convergence of ADMM
in nonconvex nonsmooth optimization," Journal of Scientific
Computing, vol. 78, no. 1, pp. 29 -- 63, Jan. 2019.
[22] R. Ge, F. Huang, C. Jin, and Y. Yuan, "Escaping from saddle
pointsonline stochastic gradient for tensor decomposition," in
Proc. of Conference on Learning Theory, Paris, France, 2015,
pp. 797 -- 842.
[23] R. Ge, J. D. Lee, and T. Ma, "Matrix completion has no spuri-
ous local minimum," in Advances in Neural Information Pro-
cessing Systems, 2016, pp. 2973 -- 2981.
[24] R. Ge, C. Jin, and Y. Zheng, "No spurious local minima in non-
convex low rank problems: A unified geometric analysis," in
Proceedings of the 34th International Conference on Machine
Learning, 2017, pp. 1233 -- 1242.
[25] K. Kawaguchi, "Deep learning without poor local minima," in
Advances in Neural Information Processing Systems, pp. 586 --
594. 2016.
[26] Y. Nesterov and B.T. Polyak, "Cubic regularization of newton
method and its global performance," Mathematical Program-
ming, vol. 108, no. 1, pp. 177 -- 205, Aug 2006.
[27] C. Fang, C. J. Li, Z. Lin, and T. Zhang,
"SPIDER: Near-
optimal non-convex optimization via stochastic path-integrated
differential estimator," in Proc. of NIPS, pp. 689 -- 699. Mon-
treal, Canada, 2018.
[28] Z. Allen-Zhu and Y. Li, "NEON2: Finding local minima via
first-order oracles," in Proc. of NIPS, pp. 3716 -- 3726. Mon-
treal, Canada, Dec. 2018.
[29] J. D. Lee, M. Simchowitz, M. I. Jordan, and B. Recht, "Gra-
dient descent only converges to minimizers," in 29th Annual
Conference on Learning Theory, New York, 2016, pp. 1246 --
1257.
[30] C. Jin, P. Netrapalli, R. Ge, S. M. Kakade and M. I. Jordan,
"Stochastic gradient descent escapes saddle points efficiently,"
available as arXiv:1902.04811, Feb. 2019.
[31] H. Daneshmand, J. Kohler, A. Lucchi and T. Hofmann,
available as
"Escaping saddles with stochastic gradients,"
arXiv:1803.05999, March 2018.
[32] S. Vlaski and A. H. Sayed,
"Second-order guarantees of
stochastic gradient descent in non-convex optimization," sub-
mitted for publication, available as arXiv:1908.07023, August
2019.
[33] A. Daneshmand, G. Scutari and V. Kungurtsev, "Second-order
guarantees of distributed gradient algorithms," available as
arXiv:1809.08694, Sep. 2018.
[34] B. Swenson, S. Kar, H. V. Poor and J. M. F. Moura, "An-
available as
nealing for distributed global optimization,"
arXiv:1903.07258, March 2019.
|
1802.03858 | 1 | 1802 | 2018-02-12T01:31:18 | Machine Learning-based Variability Handling in IoT Agents | [
"cs.MA",
"cs.LG"
] | Agent-based IoT applications have recently been proposed in several domains, such as health care, smart cities and agriculture. Deploying these applications in specific settings has been very challenging for many reasons including the complex static and dynamic variability of the physical devices such as sensors and actuators, the software application behavior and the environment in which the application is embedded. In this paper, we propose a self-configurable IoT agent approach based on feedback-evaluative machine-learning. The approach involves: i) a variability model of IoT agents; ii) generation of sets of customized agents; iii) feedback evaluative machine learning; iv) modeling and composition of a group of IoT agents; and v) a feature-selection method based on manual and automatic feedback. | cs.MA | cs | Machine Learning-based Variability Handling in IoT Agents
Nathalia Nascimento12, Paulo Alencar2, Carlos Lucena12 and Donald Cowan2
1 Laboratory of Software Engineering.
Pontifical Catholic University of Rio de Janeiro. Rio de Janeiro, Brazil.
2 David R. Cheriton School of Computer Science.
University of Waterloo. Waterloo, ON, Canada.
nnascimento,[email protected], palencar,[email protected]
8
1
0
2
b
e
F
2
1
]
A
M
.
s
c
[
1
v
8
5
8
3
0
.
2
0
8
1
:
v
i
X
r
a
Abstract
Agent-based IoT applications have recently been
proposed in several domains, such as health care,
smart cities and agriculture. Deploying these ap-
plications in specific settings has been very chal-
lenging for many reasons including the complex
static and dynamic variability of the physical de-
vices such as sensors and actuators, the software
application behavior and the environment in which
the application is embedded. In this paper, we pro-
pose a self-configurable IoT agent approach based
on feedback-evaluative machine-learning. The ap-
proach involves:
i) a variability model of IoT
agents; ii) generation of sets of customized agents;
iii) feedback evaluative machine learning; iv) mod-
eling and composition of a group of IoT agents; and
v) a feature-selection method based on manual and
automatic feedback.
1 Introduction
Based on the Google Trends tool [Google, 2018], the Internet
of Things (IoT) [Atzori et al., 2012] is emerging as a topic
that is highly related to robotics and machine learning.
In
fact, the use of learning agents has been proposed as an ap-
propriate approach to modeling IoT applications [do Nasci-
mento and de Lucena, 2017b]. These types of applications
address the problems of distributed control of devices that
must work together to accomplish tasks [Atzori et al., 2012].
This has caused agent-based IoT applications to be consid-
ered for several domains, such as health care, smart cities,
and agriculture. For example, in a smart city, software agents
can autonomously operate traffic lights [do Nascimento and
de Lucena, 2017b; Santos et al., 2017], driverless vehicles
[Herrero-Perez and Martinez-Barbera, 2008] and street lights
[do Nascimento and de Lucena, 2017a].
Agents that can interact with other agents or the envi-
ronment in which the applications are embedded are called
embodied agents [Brooks, 1995; Marocco and Nolfi, 2007;
Nolfi et al., 2016; do Nascimento and de Lucena, 2017a].
The first step in creating an embodied agent is to design its
interaction with an application's sensors and actuators, that
is, the signals that the agent will send and receive [Nolfi et al.,
2016]. As a second step, the software engineer provides this
agent with a behavior specification compatible with its body
and with the task to be accomplished. However, to specify
completely the behaviors of a physical system at design-time
and to identify and foster characteristics that lead to benefi-
cial collective behavior is difficult [Mendonc¸a et al., 2017].
To mitigate these problems, many approaches [Marocco and
Nolfi, 2007; Oliveira and Loula, 2014; Nolfi et al., 2016;
do Nascimento and de Lucena, 2017a] have proposed the use
of evolving neural networks [Nolfi and Parisi, 1996] to enable
an embodied agent to learn to adapt their behavior based on
the dynamics of the environment [Nolfi and Parisi, 1996].
The ability of a software system to be configured for differ-
ent contexts and scenarios is called variability [Galster et al.,
2014]. According to [Galster et al., 2014], achieving variabil-
ity in software systems requires software engineers to adopt
suitable methods and tools for representing, managing and
reasoning about change.
However, the number and complexity of variation points
[Pohl et al., 2005] that must be considered while model-
ing agents for IoT-based systems is quite high [Ayala et al.,
2015]. Thus, "current and traditional agent development pro-
cesses lack the necessary mechanisms to tackle specific man-
agement of components between different applications of the
IoT, bearing in mind the inherent variability of these systems"
[Ayala et al., 2015].
In this paper, we propose a self-configurable IoT agent ap-
proach based on feedback-evaluative machine-learning. The
approach involves:
(i) a variability model for IoT agents;
(ii) generation of sets of customized agents; (iii) feedback-
evaluative machine-learning; (iv) modeling and composition
of a group of IoT agents; and (v) a feature-selection method
based on both manual and automatic feedback.
1.1 Motivation: Variability in IoT Agents
In an Internet of Things application suite, there are sev-
eral options for physical components and software behaviors
for the design of a physical agent [del Campo et al., 2017;
Ayala et al., 2015]. According to existing experiments [Vega
and Fuks, 2016; Soni and Kandasamy, 2017] and our ex-
perience with the IoT domain [do Nascimento et al., 2015;
Briot et al., 2016; Do Nascimento et al., 2016; do Nasci-
mento and de Lucena, 2017b; do Nascimento and de Lucena,
2017a], we introduce possible variants of an IoT embodied
agent in Table 1. For example, the physical devices may vary
Table 1: IoT Agents Variability.
Behavior Variability
Analysis Architecture
(Neural Network Variability)
Behavior/
Constraint
Variability
Number and type of
communication signals
Notification types
(e.g. alerts)
Thresholds to activate
notifications
Number layers
Number neurons
per layer
Activation
Function (e.g.
linear, sigmoid)
Properties (e.g.
WTA, feedback)
Body Variability
Number of sensors
Type of sensors (e.g. temperature,
humidity, motion, lighting, gases)
Calibration of sensors
(e.g. temperature detector
range, range of
presence detection,
reaction time, range
of colors detection)
Energy Consumption
Sensors
Battery life
Communication device
Range of communication
devices (e.g. short range,
long range)
Number and type of motors
Number and type of
actuators (e.g. alarm)
IoT Application Logic -
connection between the
inputs and outputs (e.g. if the
lighting sensor value is zero,
then turn on the light, if the
temperature sensor is below
zero, then turn on the heater)
Architecture (e.g.
full connected, output layer
connected to all of the hidden
units)
in terms of the types of sensors, such as temperature and hu-
midity, and in terms of actuators. Each sensor can also vary
in terms of brands, changing such parameters as energy con-
sumption and battery life. The three main variation points we
have identified as shown in Table 1 illustrate the complexity
of IoT agent-based applications.
Thus, the complexity of the behavior of the agent will vary
based on the physical components that are operated by the
agent. For example, if an agent is able to activate an alarm,
which kinds of alerts can this agent generate? If this agent is
able to communicate, how many words is this agent able to
communicate? If this agent is able to control the temperature
of a room, what are the threshold values set to change the
room's temperature?
In addition, we also need to deal with variants in agent ar-
chitecture that the agent uses to sense the environment and
behave accordingly. For example, this architecture can be
a decision tree, a state machine or a neural network. Many
approaches [Marocco and Nolfi, 2007; Nolfi et al., 2016;
do Nascimento and de Lucena, 2017a] use neuroevolution,
which is "a learning algorithm which uses genetic algorithms
to train neural networks" [Whiteson et al., 2005a]). This type
of network determines the behavior of an agent automatically
based on its physical characteristics and the environment be-
ing monitored. A neural network is a well-known approach to
provide responses dynamically and automatically, and create
a mapping of input-output relations [Haykin, 1994], which
may compactly represent a set of "if..then" conditions [do
Nascimento and de Lucena, 2017a], such as: "if the temper-
ature is below 10◦C, then turn on the heat." However, finding
an appropriate neural network architecture based on the phys-
ical features and constraint behavior that were selected for an
agent, is not easy. To model the neural network, we also need
to account for its architectural variability, such as the acti-
vation function, the number of layers and neurons and prop-
erties such as the use of winner-take-all (WTA) as a neural
selection mechanisms [Fukai and Tanaka, 1997] and the in-
clusion of recurrent connections [Marocco and Nolfi, 2007].
With respect to variabilities, [Marocco and Nolfi, 2007],
performed two experiments with the same embodied agents,
varying only the neural network architectures and neural ac-
tivation functions. In the first experiment, they used a neural
network without internal neurons, while in the second exper-
iment, they used a neural network with internal neurons and
recurrent connections. In addition, they also used different
functions to compute the neurons' outputs. Based only on the
neural network characteristics, they classified the robots from
the first experiment as reactive robots (i.e. "motor actions can
only be determined on the basis of the current sensory state"),
and non-reactive robots (i.e. "motor actions are also influ-
[Marocco
enced by previous sensory and internal states").
and Nolfi, 2007] analyzed whether the type of neural archi-
tecture influenced the performance of a team of robots. They
showed that the differences in performance between reactive
and non-reactive robots vary according to the environmental
conditions and how the robots have been evaluated.
[Oliveira and Loula, 2014] investigated symbol represen-
tations in communication based on the neural architecture
topology that is used to control an embodied agent. They
found that the communication system varies according to how
the hidden layers connect the visual inputs to the auditory in-
puts.
These findings have helped us to conclude that to support
the design of IoT embodied agents, we need to account for
the variability of the physical body, the behavior constraints,
and the architecture that analyses the inputs.
2 Approach
We aim to support the development of IoT embodied agents
by designing a platform to support i) handling variability in
IoT embodied agents, ii) selecting the physical components
that will compose each agent, and iii) finding their appropri-
ate behavior according to their bodies and the scenario where
they will be applied. Figure 1 depicts the high-level model of
our proposed approach to self-configurable agents.
Figure 1: High-level model of the self-configurable agent approach
to generate embodied agents.
Basically, this platform or agent factory contains five mod-
ules: i) a manual control that allows an IoT expert to select
the first set of features manually; ii) a reconfigurable sys-
tem that contains the features that can be used to compose
the set of agents incorporating feature-oriented domain anal-
ysis (FODA) [Pohl et al., 2005] to model the software's vari-
ability; iii) the creation of a set of agents containing the se-
lected features that are also able to use a neural network to
learn about the environment; iv) a module for evaluating feed-
back tasks, by investigating the performance of the group of
agents in the application scenario after the learning execu-
tion (depending on the evaluation result, the control module
can trigger the machine learning algorithm to reconfigure the
set of features); and v) a machine-learning module to select
autonomously a new set of physical, behavior and neural net-
work features.
2.1 Current Implementation
The current implementation of our architecture consists of
two main parts. First (subsection 2.1.1), a human-in-the-loop
selects the set of physical, behavior and neural network fea-
tures for the group of IoT agents. Second (subsection 2.1.2),
based on the features that were selected by the human, a
neuroevolution-based algorithm is used to remove the irrel-
evant physical features and discover the agent's behavior.
During the second step, the neuroevolution-based algo-
rithm considers a specific environment to discover the ap-
propriate behaviors that enable a set of agents to achieve a
collective task on that environment. After finding an appro-
priate behavior (i.e. the weights and topology of the neural
network), the initial phase of the learning process is complete.
However, an unexpected change in environment may force
this process to be re-executed as all variation points can be
affected. If this environmental change makes it necessary to
add a new sensor to the agents' body, the way that the agents
perceive the environment may also be reconfigured, and the
learning process in the second step will also need to be re-
executed. In addition, if the environment changes dynami-
cally, there is a need to identify which variation points will be
affected and how to handle the associated variability.
2.1.1 Changing / Adding Features - Changing the search
space for the neuroevolution-based algorithm
According to the FODA notation, features can be classified
as mandatory, optional and alternative. Alternative features
are not to be used in the same instance, such as the range of
communication devices, the number of words to be communi-
cated or the maximum number of hidden layers. For example,
in the beginning of the experiment, if we select a neural net-
work as the decision architecture, we must choose one of the
features that defines the maximum number of hidden layers
that this neural network can have, such as "two" or "three."
So, if "two" is selected, the search space for learning will be
limited to the use of two hidden layers. If the communication
system of the agents is limited to one word, the learning algo-
rithm will not be able to test other solutions that could involve
the communication of more than one word.
Thus, the current search space to be used by the learning
algorithm has been limited by the set of features that were se-
lected to compose the embodied agents. However, there are
three situations for which this search space may need to be
changed or expanded: i) the learning algorithm does not find
a good solution using this set of features, making it necessary
to select alternative choices for some features (i.e. selecting
a different activation function for the neural network) to re-
configure the set of agents; ii) the user changes some require-
ments of the agent-based system, making it necessary to add
new unpredicted features to the feature model, as described in
[Sharifloo et al., 2016]; and iii) the learning algorithm found
an appropriate solution for the agents in the environment (i.e.
the collection of agents are achieving their tasks in the appli-
cation environment), but the environment changed dynami-
cally, unexpectedly decreasing the performance of the agents.
In this step, there is a need to control the search space that
will be used by the neuroevolution-based algorithm for the
next step (described in subsection 2.1.2). This control con-
sists of selecting the set of features to compose the system.
For instance, a human-in-the-loop has performed this selec-
tion and reconfiguration. But our goal (and we designed our
architecture for this purpose) is to enable an automatic recon-
figuration of the system. In such a case, if the agents face
an unexpected environmental change, a learning algorithm
can be used to select a new set of features to compose the
group of agents and execute the neuroevolution-based algo-
rithm again. In this situation, we proposed the use of a learn-
ing algorithm to reconfigure a neural network (i.e. selecting
another activation function), which can be seen as an auto-
Reconfigurable System for Agent CustomizationGenerates embodied agentsTo live in a scenarioSimulated or real scenarioControllerSensors, actuators and behaviorsMachine LearningReconfigure knowledge(Autonomous Features Selection) EvaluationResultsFeedback LoopAgent ID: xxx1Set with n agentsAgent FactoryManualControlBody featuresBehavior featuresManual Features Selection23514Neural Network featuresSmart Fruit MonitoringInputSensorOutputTemperatureGasMethaneHydrogenHumidityLightingPredictionTime to spoil Neural network#Hidden layers132WinnerTakes All(WTA)AtivationfunctionLinearSigmoidThreshold#neuronsFeedbackDecision In this scenario, we consider a set of street lights distributed
in a neighborhood. These street lights need to learn to save
energy while maintaining the maximum visual comfort in the
illuminated areas. For more details concerning this applica-
tion scenario, see [do Nascimento and de Lucena, 2017a].
3.1 Feature-Oriented Domain Analysis (FODA)
Figure 3 illustrates the use of FODA to express IoT agent
variability in a public lighting application.
matic machine-learning approach (Auto-ML) [Muneesawang
and Guan, 2002].
2.1.2 Using neuroevolution to discard irrelevant features
and discover the agent's current behavior
As described previously, each agent contains a neural net-
work to make decisions. The weights, the topology, the in-
put and output features of this neural network are determined
based on an evolutionary algorithm. This algorithm makes
changes based on the performance evaluation of the agents in
the application.
We implemented this neuroevolution algorithm based
on the Feature Deselective NeuroEvolution of Augmenting
Topologies (FD-NEAT) proposed by [Tan et al., 2009]. But
instead of starting with a minimal architecture with inputs di-
rectly connected to the output layers, without a hidden layer
as proposed by the traditional NEAT and FD-NEAT meth-
ods, we started with a three-layer neural network with all con-
nections. In addition, we decided that a connection removal
means a zeroed weight between two neurons, as illustrated in
2.
Figure 2: Removing input features and other neuronal connections.
If the weight is zero, in our implementation, the neural net-
work's output will not influence the activation function of the
next neuron. In this case, the hidden layer will always exist,
which can make the search more complicated. To mitigate
this complexity, we established positive and negative thresh-
olds for the weight setting in order to stimulate the connection
removals. So, only the connections with higher contributions
will remain during the evolutionary process. For example, in
a weight range of [-2;+2], connections with "0.2" or "-0.1"
weights are examples of connections that will be removed. In
such a case, if all connections between a sensor input and the
hidden layer are removed, this input feature will be discarded.
3 Illustrative Example: Smart Street Lights
To illustrate the variability dimensions of an IoT agent-based
application, we selected and implemented one of the simplest
examples from the IoT domain: a smart street light appli-
cation. Even in a simple experiment of lighting control, we
found many different prototypes in the literature [Carrillo et
al., 2013; De Paz et al., 2016; do Nascimento and de Lu-
cena, 2017a]. For example, [Carrillo et al., 2013] provided
lights with cameras for image processing, while [De Paz et
al., 2016] provided them with ambient light sensors, and [do
Nascimento and de Lucena, 2017a] provided lights with am-
bient light and motion-detection sensors.
Figure 3: Feature model of a smart light agent.
As shown in this figure, even in a simple IoT agent, you
may need to consider many variation points to create an IoT
agent. According to the model, the input, decision and out-
put are mandatory features. But the selection of sensors to
compose the body of the agent is optional. If you decide to
use sensors, you must select at least one of the sensors, such
as the light sensor. In addition, if you select the light sensor
feature, you must select which brand will be used. Depend-
ing on the selected light sensor brand, your agent will be able
to sense very small changes in light or detect a full range of
colors [Intorobotics, 2018].
3.2 Selecting Physical and Neural Network
Features
An IoT expert selected three physical inputs and two phys-
ical outputs to measure and operate each one of the street
lights. The expert also added one behavior output: namely,
the agents could ignore messages received from neighboring
street lights. In addition, the engineer selected a neural net-
work with one hidden layer with five units as the initial net-
work for each agent with the sigmoid function as the activa-
tion function of this neural network.
Figure 4 depicts the three-layer neural network that was
generated based on the selected features. The input layer in-
cludes four units that encode the activation level of the sen-
sors and the previous output value of the listening decision
output. The output layer contains three output units: (i) lis-
teningDecision, that enables the smart lamp to receive signals
from neighboring street lights in the next cycle; (ii) wire-
lessTransmitter, a signal value to be transmitted to neigh-
boring street lights; and (iii) lightDecision, that switches the
light's OFF/DIM/ON functions.
OutputsHiddenNodesInputs0.00.00.00.01.5-1.3Smart LightInputOutputCommunicationReceiverNeural Network#Hidden layers132WinnerTakes All(WTA)AtivationfunctionLinearSigmoidBinary#neuronsFeedbackCommunicationTransmitter tostreet lightsDistance500 m1 KmSensorMotionLightBroken LampDetectorActivate Lamp#levelsof light intensity#signals2323NegativeweightsbiasReLuR1. lightSensor=0 → lamp = 1…unkown rulesComposition rulesCameraMaximum weightDecisionState MachineDecision treeTSL235RColorPALVCNL4000 Figure 4: Neural network resulted from the first feature-selection
interaction.
3.3 Learning about the environment
During the training process, the algorithm evaluates the op-
tions for weights of the network based on energy consump-
tion, the number of people that finished their routes before the
simulation ends, and the total time spent by people moving
during their trip. Therefore, each weight-set trial is evaluated
after the simulation ends based on the following equations:
pP eople =
(completedP eople × 100)
totalP eople
(totalEnergy × 100)
(1)
(2)
(3)
)
pEnergy =
( 11×(timeSimulation×totalSmartLights)
10
(totalT imeT rip × 100)
pT rip =
(( 3×timeSimulation
) × totalP eople)
(2)
f itness = (1.0×pP eople)−(0.6×pT rip)−(0.4×pEnergy)
(4)
in which pP eople is the percentage of people that com-
pleted their routes by the end of the simulation out of the total
number of people participating in the simulation; pEnergy is
the percentage of energy that was consumed by street lights
out of the maximum energy value that could be consumed
during the simulation. We also considered the use of the wire-
less transmitter to calculate energy consumption; pT rip is the
percentage of the total duration time of people's trips out of
the maximum time value that their trip could consume; and
f itness is the fitness of each representation candidate that
encodes the neural network.
Environmental Setting
As illustrated in Figure 5, in this first step, the scenario was
bright during the entire period that the agents were learn-
ing about the environment. After some learning interactions,
the agents developed an appropriate behavior to achieve their
tasks in this version of the application. As a result, the
neuroevolution-based algorithm discarded all the neuronal
connections between the light sensor and the hidden units.
As the environment was always sunny, an obvious behav-
ior for the agent is to turn off the light, whether a person was
present or not. As expected, the agents produced this behav-
ior. We were also expecting that the number of hidden neu-
rons would be considerably decreased because of the simplic-
ity of the task. But only one hidden neuron and one feature
input were removed. This behavior occurred because the IoT
expert selected a sigmoid function as an activation function.
Figure 5: Learning the environment.
As known, the output of the sigmoid function is not zero
when its input is zero (i.e.
if the sigmoid input is zero, the
LED will be turned on). Thus, the learning algorithm found a
configuration to assure that the LED remains turned off.
3.4 Reconfiguring the set of features
The learning algorithm did not find an appropriate
solution
The IoT expert was expecting a fitness performance higher
than 75% and a simpler architecture to use in the real devices.
But after several interactions of the learning algorithm with
the environment, the highest fitness performance achieved by
the learning algorithm was 72%.
As the human-in-the-loop was not satisfied with this result,
he/she reconfigured the first set of features that was used to
compose the IoT agents. For instance, an alternative choice
of the activation function of the neural network was selected:
namely, the binary activation function with threshold. As a
result, the performance result quickly increased by more than
5% and the human-in-the-loop obtained a simpler architec-
ture.
The environment unexpectedly changed
Figure 6: Reconfiguring the set of features.
After a time, a change in the environment occurred. Now,
these agents are operating in an environment in which some-
times the background light can be bright and at other times
dark. As a result, the performance of the set of agents con-
Motion sensorLight sensorWireless receptorWireless transmitterLED (3 levels)Turn on WirelessReceiverPrevious decision (turn onwireless)Input layerHidden layerOutput layerThe performance of each neural network configurationis evaluated40455055606570751471013161922252831343740434649525558616467707376798285PERFORMANCE AVERAGEBEST PERFORMANCEConfigurationassigned toagentsMotion sensorLight sensorWireless receptorWireless transmitterLED (3 levels)Turn on WirelessReceiverPrevious decision (turn onwireless)Input layerHidden layerOutput layer203040506070809016111621263136414651566166717681869196101106111116121126131PERFORMANCE AVERAGEBEST PERFORMANCENeuroevolutionThe neuralNetwork is reconfiguredEnvironmental change decreases agents performanceAgents mustLearn againMotion sensorLight sensorWireless receptorWireless transmitterLED (3 levels)Turn on WirelessReceiverPrevious decision (turn onwireless)Input layerHidden layerOutput layerEnvironmentchanged siderably decreased, as shown in Figure 6. The learning algo-
rithm continued its training from its last state, but the current
analysis architecture was not a viable option for this new sit-
uation.
The human-in-the-loop evaluated this decreased perfor-
mance, and then reconfigured the system. For instance, the
expert could have selected a new sensor, but he/she main-
tained the number of sensor inputs, but selected different
variants for the neural network, such as "two" as the max-
imum number of neurons in the hidden layer and the sig-
moid activation function. Then, the learning algorithm was
re-executed and the agents learned to cope with this environ-
mental change.
4 Related Work
[Whiteson et al., 2005b; Tan et al., 2009; Diuk et al., 2009;
Nguyen et al., 2013; Ure et al., 2014; del Campo et al., 2017]
are some of the examples that apply feature selection to han-
dle variability in learning agent-based systems. For example,
[Diuk et al., 2009] propose an approach that uses reinforce-
ment learning algorithms for structure discovery and feature
selection while actively exploring an unknown environment.
To exemplify the use of the proposed algorithm, the authors
present the problem of a unique robot that has to decide which
of multiple sensory inputs such as camera readings of surface
color and texture, and IR sensor reading are relevant for cap-
turing an environment's transition dynamics.
However, most of these approaches do not address the
problem of environmental- or user-based reconfiguration, in
which a new set of features may be selected based on envi-
ronmental changes, expanding or changing the search space
of the group of learning agents. In addition, as most of these
approaches load all features into the agent, they do not ad-
dress the problem of dealing with mandatory, optional and
alternative features.
[Sharifloo et al., 2016] provides one of the few solutions
that propose an approach for feature selection and feature set
reconfiguration. They presented a theoretical approach that
proposes the use of reinforcement learning for feature selec-
tion, and a reconfiguration guided by system evolution where
the user creates new features to deal with changes in system
requirements. They do not consider the changes that can hap-
pen dynamically in the environment, which can be handled
by an automatic module by testing alternative choices of fea-
tures.
In addition, most of these approaches do not characterize
In fact, [Galster et
variability in their application domain.
al., 2014] observed that most approaches for variability han-
dling are not oriented to specific domains. These approaches
are potentially widely applicable. However, [Galster et al.,
2014] consider that for a variability approach to cover com-
plex domains, it is necessary to create domain-specific solu-
tions. Therefore, [Galster et al., 2014] consider the extension
of variability approaches for specific domains as a promising
direction for future work.
5 Contributions and Ongoing Work
We provided an approach through which a software en-
gineer with expertise in IoT agents co-worked with a
neuroevolutionary-based algorithm that can discard features.
First, the software engineer provided the initial configuration
of the agent-based system, using personal expertise to select
a set of features. Then, a neuroevolutionary-based algorithm
was executed to remove those features that were selected by
the developer, but shown to be irrelevant to the application
during the simulation.
However, after an unexpected environmental change that
was not considered by the software engineer during the ini-
tial design time, the previous solution found by the neuroevo-
lutionary algorithm stopped to work. Then, after evaluating
the environmental changes, the software engineer had three
options: i) to add a new feature to the feature model; ii) to se-
lect alternative choices of some features, including a different
neural network architecture and properties, then starting the
learning process again; and iii) to maintain the set of features
and just reactivate the learning algorithm to continue from its
last state.
In addition, to handling variability in learning for IoT
agents, we identified the main variation points of these kinds
of applications, including the variants that can be involved in
a neural network design. We also provided a feature-oriented
variability model, which is an established software engineer-
ing module.
The proposed approach is an example of a human-in-
the-loop approach in which a machine-learning automated-
procedure assists a software developer in completing his/her
task. Our next step is to enable the use of a learning tech-
nique to reconfigure the set of features based on environmen-
tal changes automatically. As we proposed a hybrid architec-
ture, we can use this learning technique only to reconfigure
the variants related to one of the variation points, such as the
neural network properties. In such an instance, we can have
a human-in-the-loop responsible for handling the body and
behavior variability of the IoT agents.
Acknowledgments
This work has been supported by CAPES scholarship/Pro-
gram 194/Process: 88881.134630/2016-01 and the Labora-
tory of Software Engineering (LES) at PUC-Rio. It has been
developed in cooperation with the University of Waterloo,
Canada. Our thanks to CNPq, CAPES, FAPERJ and PUC-
Rio for their support through scholarships and fellowships.
References
[Atzori et al., 2012] Luigi Atzori, Antonio Iera, Giacomo
Morabito, and Michele Nitti. The social internet of things
(siot)–when social networks meet the internet of things:
Concept, architecture and network characterization. Com-
puter networks, 56(16):3594–3608, 2012.
[Ayala et al., 2015] Inmaculada Ayala, Mercedes Amor,
Lidia Fuentes, and Jos´e M Troya. A software prod-
uct line process to develop agents for the iot. Sensors,
15(7):15640–15660, 2015.
[Briot et al., 2016] Jean-Pierre Briot, Nathalia Moraes
de Nascimento, and Carlos Jos´e Pereira de Lucena. A
multi-agent architecture for quantified fruits: Design and
experience. In 28th International Conference on Software
Engineering & Knowledge Engineering (SEKE'2016),
pages 369–374. SEKE/Knowledge Systems Institute, PA,
USA, 2016.
[Brooks, 1995] Rodney A Brooks. Intelligence without rea-
The artificial life route to artificial intelligence:
son.
Building embodied, situated agents, pages 25–81, 1995.
[Carrillo et al., 2013] C Carrillo, E Diaz-Dorado, J Cidr´as,
A Bouza-Pregal, P Falc´on, A Fern´andez, and A ´Alvarez-
S´anchez. Lighting control system based on digital camera
for energy saving in shop windows. Energy and Buildings,
59:143–151, 2013.
[De Paz et al., 2016] Juan F De Paz, Javier Bajo, Sara
Rodr´ıguez, Gabriel Villarrubia, and Juan M Corchado. In-
telligent system for lighting control in smart cities. Infor-
mation Sciences, 372:241–255, 2016.
[del Campo et al., 2017] In´es del Campo, Victoria Mart´ınez,
Flavia Orosa, Javier Echanobe, Estibalitz Asua, and Koldo
Basterretxea. Piecewise multi-linear fuzzy extreme learn-
ing machine for the implementation of intelligent agents.
In Neural Networks (IJCNN), 2017 International Joint
Conference on, pages 3363–3370. IEEE, 2017.
[Diuk et al., 2009] Carlos Diuk, Lihong Li, and Bethany R
Leffler. The adaptive k-meteorologists problem and its ap-
plication to structure learning and feature selection in rein-
forcement learning. In Proceedings of the 26th Annual In-
ternational Conference on Machine Learning, pages 249–
256. ACM, 2009.
[do Nascimento and de Lucena, 2017a] Nathalia
Moraes
do Nascimento and Carlos Jos´e Pereira de Lucena.
Engineering cooperative smart things based on embodied
In Adaptive Hardware and Systems (AHS),
cognition.
2017 NASA/ESA Conference on, pages 109–116. IEEE,
2017.
[do Nascimento and de Lucena, 2017b] Nathalia Moraes
do Nascimento and Carlos Jos´e Pereira de Lucena. Fiot:
An agent-based framework for self-adaptive and self-
organizing applications based on the internet of things.
Information Sciences, 378:161–176, 2017.
[do Nascimento et al., 2015] Nathalia Moraes do Nasci-
mento, Carlos Jos´e Pereira de Lucena, and Hugo Fuks.
Modeling quantified things using a multi-agent system. In
Web Intelligence and Intelligent Agent Technology (WI-
IAT), 2015 IEEE/WIC/ACM International Conference on,
volume 1, pages 26–32. IEEE, 2015.
[Do Nascimento et al., 2016] Nathalia Moraes Do Nasci-
mento, Marx Leles Viana, and Carlos Jos´e Pereira de Lu-
cena. An iot-based tool for human gas monitoring. 2016.
Shigeru
Tanaka. A simple neural network exhibiting selective
activation of neuronal ensembles: from winner-take-all to
winners-share-all. Neural computation, 9(1):77–97, 1997.
[Fukai and Tanaka, 1997] Tomoki
Fukai
and
[Galster et al., 2014] Matthias Galster, Danny Weyns, Dan
Tofan, Bartosz Michalik, and Paris Avgeriou. Variabil-
ity in software systems-a systematic literature review.
IEEE Transactions on Software Engineering, 40(3):282–
306, 2014.
[Google, 2018] Google.
https://trends.google.com/, January 2018.
Google
trends.
[Haykin, 1994] Simon Haykin. Neural networks: a compre-
hensive foundation. Prentice Hall PTR, 1994.
[Herrero-Perez and Martinez-Barbera, 2008] D
Herrero-
Perez and H Martinez-Barbera. Decentralized coordi-
nation of automated guided vehicles (short paper).
In
AAMAS 2008, 2008.
[Intorobotics, 2018] Intorobotics.
budgeted
com-
mon
sensors.
https://www.intorobotics.com/common-budgeted-arduino-
light-sensors/, January 2018.
Most
light
and most
arduino
[Marocco and Nolfi, 2007] Davide Marocco and Stefano
Nolfi. Emergence of communication in embodied agents
evolved for the ability to solve a collective navigation
problem. Connection Science, 19(1):53–74, 2007.
[Mendonc¸a et al., 2017] M´arcio Mendonc¸a, Ivan R Chrun,
Fl´avio Neves, and Lucia VR Arruda. A cooperative ar-
chitecture for swarm robotic based on dynamic fuzzy cog-
nitive maps. Engineering Applications of Artificial Intelli-
gence, 59:122–132, 2017.
[Muneesawang and Guan, 2002] Paisarn Muneesawang and
Ling Guan. Automatic machine interactions for content-
based image retrieval using a self-organizing tree map
IEEE transactions on neural networks,
architecture.
13(4):821–834, 2002.
[Nguyen et al., 2013] Trung Nguyen, Zhuoru Li, Tomi Si-
lander, and Tze Yun Leong. Online feature selection
for model-based reinforcement learning. In International
Conference on Machine Learning, pages 498–506, 2013.
[Nolfi and Parisi, 1996] Stefano Nolfi and Domenico Parisi.
Learning to adapt to changing environments in evolving
neural networks. Adaptive behavior, 5(1):75–98, 1996.
[Nolfi et al., 2016] Stefano Nolfi, Josh Bongard, Phil Hus-
bands, and Dario Floreano. Evolutionary Robotics, chap-
ter 76, pages 2035–2068. Springer International Publish-
ing, Cham, 2016.
[Oliveira and Loula, 2014] Emerson Oliveira and Angelo
Loula. Symbol interpretation in neural networks: an in-
vestigation on representations in communication. In Pro-
ceedings of the Annual Meeting of the Cognitive Science
Society, volume 36, 2014.
[Pohl et al., 2005] Klaus Pohl, Gunter Bockle, and Frank J
van Der Linden. Software product line engineering: foun-
dations, principles and techniques. Springer Science &
Business Media, 2005.
[Santos et al., 2017] Fernando Santos, Ingrid Nunes, and
Ana LC Bazzan. Model-driven engineering in agent-based
modeling and simulation: a case study in the traffic signal
control domain. In Proceedings of the 16th Conference on
Autonomous Agents and MultiAgent Systems, pages 1725–
1727. International Foundation for Autonomous Agents
and Multiagent Systems, 2017.
[Sharifloo et al., 2016] Amir Molzam Sharifloo, Andreas
Metzger, Cl´ement Quinton, Luciano Baresi, and Klaus
Pohl. Learning and evolution in dynamic software prod-
uct lines. In Software Engineering for Adaptive and Self-
Managing Systems (SEAMS), 2016 IEEE/ACM 11th Inter-
national Symposium on, pages 158–164. IEEE, 2016.
[Soni and Kandasamy, 2017] Gulshan Soni and Selvaradjou
Kandasamy. Smart garbage bin systems–a comprehensive
survey. In International Conference on Intelligent Infor-
mation Technologies, pages 194–206. Springer, 2017.
[Tan et al., 2009] Maxine Tan, Michael Hartley, Michel Bis-
ter, and Rudi Deklerck. Automated feature selection in
neuroevolution. Evolutionary Intelligence, 1(4):271–292,
2009.
[Ure et al., 2014] N Kemal Ure, Girish Chowdhary, Yu Fan
Chen, Jonathan P How, and John Vian. Distributed learn-
ing for planning under uncertainty problems with hetero-
geneous teams. Journal of Intelligent & Robotic Systems,
74(1-2):529, 2014.
[Vega and Fuks, 2016] Katia Vega and Hugo Fuks. Beauty
Technology: Designing Seamless Interfaces for Wearable
Computing. Springer, 2016.
[Whiteson et al., 2005a] Shimon Whiteson, Nate Kohl, Risto
Miikkulainen, and Peter Stone. Evolving soccer keepaway
players through task decomposition. Machine Learning,
59(1-2):5–30, 2005.
[Whiteson et al., 2005b] Shimon Whiteson, Peter Stone,
Kenneth O Stanley, Risto Miikkulainen, and Nate Kohl.
Automatic feature selection in neuroevolution. In Proceed-
ings of the 7th annual conference on Genetic and evolu-
tionary computation, pages 1225–1232. ACM, 2005.
|
1806.09757 | 1 | 1806 | 2018-06-26T01:59:14 | Adaptive guaranteed-performance consensus design for high-order multiagent systems | [
"cs.MA"
] | The current paper addresses the distributed guaranteed-performance consensus design problems for general high-order linear multiagent systems with leaderless and leader-follower structures, respectively. The information about the Laplacian matrix of the interaction topology or its minimum nonzero eigenvalue is usually required in existing works on the guaranteed-performance consensus, which means that their conclusions are not completely distributed. A new translation-adaptive strategy is proposed to realize the completely distributed guaranteed-performance consensus control by using the structure feature of a complete graph in the current paper. For the leaderless case, an adaptive guaranteed-performance consensualization criterion is given in terms of Riccati inequalities and a regulation approach of the consensus control gain is presented by linear matrix inequalities. Extensions to the leader-follower cases are further investigated. Especially, the guaranteed-performance costs for leaderless and leader-follower cases are determined, respectively, which are associated with the intrinsic structure characteristic of the interaction topologies. Finally, two numerical examples are provided to demonstrate theoretical results. | cs.MA | cs |
Adaptive guaranteed-performance consensus design for
high-order multiagent systems✩
Jianxiang Xia, Jie Yanga, Hao Liub,∗, Tang Zhenga
aHigh-Tech Institute of Xi'an, Xi'an, 710025, P.R. China
bSchool of Astronautics, Beihang University, Beijing, 100191, P.R. China
Abstract
The current paper addresses the distributed guaranteed-performance consensus design
problems for general high-order linear multiagent systems with leaderless and leader-
follower structures, respectively. The information about the Laplacian matrix of the inter-
action topology or its minimum nonzero eigenvalue is usually required in existing works
on the guaranteed-performance consensus, which means that their conclusions are not
completely distributed. A new translation-adaptive strategy is proposed to realize the
completely distributed guaranteed-performance consensus control by using the structure
feature of a complete graph in the current paper. For the leaderless case, an adaptive
guaranteed-performance consensualization criterion is given in terms of Riccati inequali-
ties and a regulation approach of the consensus control gain is presented by linear matrix
inequalities. Extensions to the leader-follower cases are further investigated. Especially,
the guaranteed-performance costs for leaderless and leader-follower cases are determined,
respectively, which are associated with the intrinsic structure characteristic of the interac-
tion topologies. Finally, two numerical examples are provided to demonstrate theoretical
results.
Keywords: Multiagent systems, adaptive consensus, guaranteed-performance control,
gain regulation.
1. Introduction
In the last two decades, the distributed cooperative control of multiagent systems has
received great attention by researchers from different fields, such as formation control,
✩This work was supported by the National Natural Science Foundation of China under Grants 61374054,
61503012, 61703411, 61503009, 61333011 and 61421063 and Innovation Foundation of High-Tech Institute
of Xi'an (2015ZZDJJ03), also supported by Innovation Zone Project (17-163-11-ZT-004-017-01).
∗Corresponding author. Tel.: +86 10 62790793.
Email address: [email protected] (Hao Liu)
Preprint submitted to Information Sciences
June 27, 2018
multiple mobile robot systems, distributed computation and distributed sensor networks
[1]-[18], et al. Consensus is a fundamental problem in cooperative control, which re-
quires that all agents achieve an agreement on certain variables of interest by designing
distributed consensus protocols. Generally speaking, according to the structures of inter-
action topologies, consensus can be categorized into the leaderless one and the leader-
follower one. When leader-follower consensus is achieved, the states of all followers track
the state trajectories of the leader. However, for the leaderless case, all agents jointly deter-
mine the agreement state via interacting with each other, which is also called the consensus
function. Some very interesting and significant works were finished in [19]-[33], where
the consensus regulation performance was not considered.
For many practical multiagent systems, it is not only required that they can achieve
consensus, but also satisfy certain consensus performance. Hu et al. [34] showed a typi-
cal example; that is, when multiple mobile autonomous vehicles perform a specific patrol
task, the distance performance is critically important due to limited resource or utility max-
imization. The consensus performance can be modeled as certain cost functions, which can
be usually divided into two types: the individual cost function and the global cost function.
For the individual cost function shown in [35] and [36], each agent has a local objective
function and some global goals can be achieved by optimizing the objective function of
each agent. For the global cost function, some whole performance index is minimized in a
distributed manner. For first-order multiagent systems, Cao and Ren [37] proposed a linear
quadratic global cost function to achieve optimal consensus. For second-order multiagent
systems, a global cost function was constructed by state errors among neighboring agents
in [38]. For high-order multiagent systems, guaranteed-cost consensus criteria were pre-
sented in [39–42], which intrinsically realize suboptimal consensus control. It should be
pointed out that the above consensus results with certain cost functions are not completely
distributed since they are associated with the Laplacian matrix of the interaction topol-
ogy or its nonzero eigenvalues, which are global information of a multiagent system as a
whole.
For high-order multiagent systems without consensus performance constraints, Li et
al. [43] proposed a very interesting dilation-adaptive strategy to eliminate the impacts of
the nonzero eigenvalue of the Laplacian matrix and showed that the interaction strengths
among agents may be monotonically increasing and converge to some finite steady-state
value, but the quantitative value cannot be determined. Since the dilation factor is inversely
proportional to the minimum nonzero eigenvalue, it may be very large when the algebraic
connectivity is small and cannot be precisely determined. However, the precise value of the
dilation factor is required to optimize the global cost function. Hence, the dilation-adaptive
strategy is no longer valid to deal with guaranteed-performance consensus. To the best
of our knowledge, the following three challenging problems on guaranteed-performance
2
consensus are still open: (i) How to realize guaranteed-performance consensus control
in a completely distributed manner; (ii) How to regulate the consensus control gain and
consensus performance between any two agents; (iii) How to determine the guaranteed-
performance cost when the interaction strengths among agents adaptively change and may
be moronically increasing.
In the current paper, we intend to propose a distributed guaranteed-performance con-
sensus scheme without using the Laplacian matrix and its minimum nonzero eigenvalue.
Firstly, the new adaptive guaranteed-performance consensus protocols for leaderless cases
and leader-follower cases are constructed, respectively, where interaction weights among
neighboring agents are adaptively regulated by the state error between each agent and its
neighbors for the leaderless case, but interaction weights between the leader and followers
are time-varying and interaction weights among followers are fixed for the leader-follower
case. Then, for leaderless and leader-follower cases, adaptive guaranteed-performance
consensualization criteria are given, respectively, and the upper bounds of the guaranteed-
performance costs are determined, respectively, which are independent of the adaptively
regulated interaction weights. Finally, by combining with the linear matrix inequality
(LMI) techniques, the regulation approaches of the consensus control gain are respectively
given for leaderless and leader-follower multiagent systems by adjusting the translation
factors.
Compared with closely related literatures on guaranteed-performance consensus for
high-order linear multiagent systems, the current paper has three unique features. Firstly,
the current paper proposes a novel translation-adaptive strategy to realize completely dis-
tributed guaranteed-performance consensus. However, the guaranteed-performance con-
sensus criteria in [39–42] require the global information and the dilation-adaptive strategy
in [43] cannot be used to deal with guaranteed-performance constraints. Secondly, by
the intrinsic structure property of the interaction topology, the regulation approach of the
consensus control gain is proposed and the guaranteed-performance cost is determined in
the current paper, while the consensus control gain cannot be regulated in [39–42], where
the approaches to determine the guaranteed-performance cost are no longer valid when
interaction weights are adaptively regulated. Thirdly, the consensus protocol in the current
paper regulates the consensus performance between any two agents except the leader, but
the consensus protocol in [39–42] can only ensure the consensus performance between
neighboring agents.
The remainder of the current paper is organized as follows. Section 2 presents leaderless
adaptive guaranteed-performance consensualization criteria, determines the guaranteed-
performance cost, and gives a consensus control gain regulation approach. In Section 3,
leader-follower adaptive guaranteed-performance consensus problems are investigated and
the associated consensualization criteria are proposed. Two numerical examples are given
3
to demonstrate theoretical results in Section 4 and some concluding remarks are presented
in Section 5.
Notations: Rd denotes the real column vector space of dimension d and Rd×d stands for
the set of d× d dimensional real matrices. IN represents the identity matrix of dimension N
and 1N is an N-dimensional column vector with all entries equal to 1. 0 stands for the zero
column vector with a compatible dimension. QT denotes the transpose of Q. RT = R > 0
and RT = R < 0 mean that the symmetric matrix R is positive definite and negative
definite, respectively. RT = R ≥ 0 and RT = R ≤ 0 mean that the symmetric matrix R
is positive semidefinite and negative semidefinite, respectively. The notation ⊗ stands for
the Kronecker product. The symmetric elements of a symmetric matrix are denoted by the
symbol *.
2. Leaderless adaptive guaranteed-performance consensualization
This section first presents the problem description for leaderless adaptive guaranteed-
performance consensus design. Then, the associated consensualization criteria are given
in terms of Ricatti inequalities, the guaranteed-performance cost is determined and an
approach to regulate the consensus control gain is shown in terms of LMIs.
2.1. Problem description for leaderless cases
A connected undirected graph G is used to describe the interaction topology of a multi-
agent system, where each agent is denoted by a node, the interaction channel between any
two nodes is represented by an edge and the interaction strength is denoted the edge weight
likwik(t), where lii = 0, lik = 1 if agent k is a neighbor of agent i and lik ≡ 0 otherwise, and
wik(t) > 0 with wik(0) > 0 is a nondecreasing function designed later. The Laplacian matrix
of G is defined as Lw(t) = D(t)−W(t), where W(t) = [likwik(t)]N×N denotes the weight matrix
of G and D(t) = diag{d1(t), d2(t), · · · , dN(t)} with di(t) = PN
k=1,k,i likwik(t) (i = 1, 2, · · · , N)
represents the in-degree matrix of G. Since G is connected, zero is its single eigenvalue
and the other eigenvalues are positive as shown in [44].
The dynamics of each agent is described by the following general high-order linear
model:
xi(t) = Axi(t) + Bui(t) (i ∈ {1, 2, · · · , N}) ,
where A ∈ Rd×d, B ∈ Rd×p, N is a positive integer standing for the number of agents, and
xi(t) and ui(t) are the state and the control input of agent i, respectively.
A leaderless adaptive guaranteed-performance consensus protocol is proposed as fol-
(1)
lows
4
wik(t) (xk(t) − xi(t)),
ui(t) = Ku Pk∈Ni
wik(t) = (xk(t) − xi(t))T Kw (xk(t) − xi(t)) ,
Jr = 1
N
N
(xk(t) − xi(t))T Q (xk(t) − xi(t)) dt,
N
Pk=1R +∞
0
Pi=1
(2)
where Ku ∈ Rp×d and Kw ∈ Rd×d are gain matrices with KT
w = Kw ≥ 0, Ni denotes the
neighbor set of agent i and QT = Q > 0. Here, it is assumed that there exists an upper
bound γik of wik(t) (i, k ∈ {1, 2, · · · , N}), which can be determined by (2) and is related to the
state error between two neighboring agents and the unknown gain matrix Kw. Protocol (2)
has two new features. The first one is that wik(t) with wik(0) > 0 is adaptively regulated and
is nondecreasing since Kw is symmetric and positive semidefinite. If agent k is a neighbor
of agent i, then wik(t) is a practical interaction weight from agent k to agent i. Otherwise,
wik(t) can be regarded as a virtual interaction weight from agent k to agent i. The second
one is that the consensus performance of all agents instead of neighboring agents can be
regulated.
The definition of the leaderless adaptive guaranteed-performance consensualization is
given as follows.
Definition 1. Multiagent system (1) is said to be leaderless adaptively guaranteed-perfor-
mance consensualizable by protocol (2) if there exist Ku and Kw such that limt→+∞(xi(t)−
xk(t)) = 0 (i, k = 1, 2, · · · , N) and Jr ≤ J∗r for any bounded initial states xi(0) (i = 1, 2, · · · , N),
where J∗r is said to be the guaranteed-performance cost.
Remark 1. In protocol (2), Jr(t) denotes the consensus performance regulation term,
which can be realized by choosing a proper matrix Q. The jth diagonal element of Q
stands for the optimization weight of the jth component of the state error between agents
k and i and the other elements of Q represent the coupling relationships among the corre-
sponding components of the state error. Especially, for practical multiagent systems, the
matrix Q is usually chosen as a diagonal matrix. In this case, a bigger optimization weight
can guarantee the smaller squared sum of the associated component of the state error by
the controller design approaches, which are often based on the Riccati inequality. More
detailed explanations and theoretical analysis about the impacts of the parameters to the
control performance for isolated systems and multiagent systems can be found in [37] and
[46], respectively.
In the following, we design gain matrices Ku and Kw such that multiagent system
(1) achieves leaderless adaptive guaranteed-performance consensus, and determine the
guaranteed-performance cost J∗r . Furthermore, an approach to regulate the consensus con-
trol gain is proposed and the consensus motion is determined.
5
2.2. Adaptive guaranteed-performance consensus design for leaderless cases
N(t)iT
1 (t), xT
2 (t), · · · , xT
protocol (2) can be written in a compact form as
, then the dynamics of multiagent system (1) with
where Lw(t) is the Laplacian matrix of the interaction topology. Since the interaction topol-
Let x(t) = hxT
x(t) = (cid:0)IN ⊗ A − Lw(t) ⊗ BKu(cid:1) x(t),
ogy is connected, there exists an orthonormal matrix U = h1N. √N, ¯Ui such that
U T Lw(0)U = " 0
where ∆w(0) = ¯U T Lw(0) ¯U = diag{λ2, λ3, · · · , λN} with λ2 ≤ λ3 ≤ · · · ≤ λN being nonzero
eigenvalues of Lw(0). Let ¯x(t) = (cid:16)U T ⊗ Id(cid:17) x(t) = h ¯xT
3 (t), · · · ,
N(t)]T , then multiagent system (3) can be transformed into
¯xT
1 (t), ηT (t)iT
0T
0 ∆w(0) # ,
(3)
(4)
with η(t) = [ ¯xT
2 (t), ¯xT
¯x1(t) = A ¯x1(t),
(5)
η(t) = (cid:16)IN−1 ⊗ A − ¯U T Lw(t) ¯U ⊗ BKu(cid:17) η(t).
Let ei (i = 1, 2, · · · , N) denote N -dimensional column vectors with the ith element 1 and
0 elsewhere, then we define
(6)
x¯c(t)
∆
=
N
Xi=2
Uei ⊗ ¯xi(t),
xc(t)
∆
= Ue1 ⊗ ¯x1(t) =
1
√N
1N ⊗ ¯x1(t).
Due to
N
Xi=2
ei ⊗ ¯xi(t) = h0T , ηT (t)iT
,
it can be shown by (7) that
x¯c(t) = (U ⊗ Id)h0T , ηT (t)iT
From (8), one can see that
.
xc(t) = (U ⊗ Id)h ¯xT
1 (t), 0TiT
.
6
(7)
(8)
(9)
(10)
Because U⊗ Id is nonsingular, x¯c(t) and xc(t) are linearly independent by (9) and (10). Due
to (cid:16)U T ⊗ Id(cid:17) x(t) = h ¯xT
, one can obtain that x(t) = x¯c(t) + xc(t). By the structure
of xc(t) in (8), multiagent system (1) achieves consensus if and only if limt→+∞η(t) = 0;
that is, subsystems (5) and (6) describe the consensus motion and relative state motion of
multiagent system (1), respectively.
1 (t), ηT (t)iT
The following theorem presents a sufficient condition for leaderless adaptive guaranteed-
performance consensualization, which can realize completely distributed guaranteed-per-
formance consensus design; that is, the design approach of gain matrices in protocol (2) is
independent of the time-varying interaction topology and its eigenvalues.
Theorem 1. For any given translation factor γ > 0, multiagent system (1) is adaptively
guaranteed-performance consensualizable by protocol (2) if there exists a matrix PT =
P > 0 such that PA + AT P − γPBBT P + 2Q ≤ 0. In this case, Ku = BT P, Kw = PBBT P
and the guaranteed-performance cost satisfies that
1
N
1N1T
N! ⊗ P! x(0) + γZ +∞
0
J∗r = xT (0) IN −
Proof. First of all, we design Ku and Kw such that limt→+∞η(t) = 0. Construct the follow-
ing new Lyapunov function candidate
N! ⊗ PBBT P! x(t)dt.
xT (t) IN −
1N1T
1
N
V(t) = ηT (t) (IN−1 ⊗ P) η(t) +
N
Xi=1 Xk∈Ni
(wik(t) − wik(0))2
2
+
γ
2N
N
N
Xi=1
Xk=1,k,i
(γik − wik(t)).
Due to PT = P > 0 and γik ≥ wik(t), one has V(t) ≥ 0. Since it is assumed that the
interaction topology is undirected, one has Lw(t) = LT
w(t) . Let Ku = BT P, then the time
derivative of V(t) along the solution of subsystem (6) is
V(t) = ηT (t)(cid:16)IN−1 ⊗(cid:16)PA + AT P(cid:17) − 2 ¯U T Lw(t) ¯U ⊗ PBBT P(cid:17) η(t)
N
N
N
+
Xi=1 Xk∈Ni
(wik(t) − wik(0)) wik(t) −
γ
2N
Xi=1
Xk=1,k,i
wik(t).
(11)
Due to UU T = IN, one can show that ¯U ¯U T = LN, where LN is the Laplacian matrix of a
complete graph with the weights of all the edges 1/N. Thus, it can be obtained by (2) and
(4) that
N
Xi=1 Xk∈Ni
(wik(t) − wik(0)) wik(t) −
γ
2N
N
N
Xi=1
Xk=1,k,i
wik(t)
7
= xT (t)(cid:0)(cid:0)2Lw(t) − 2Lw(0) − γLN(cid:1) ⊗ Kw(cid:1) x(t)
= ηT (t)(cid:16)(cid:16)2 ¯U T Lw(t) ¯U − 2∆w(0) − γIN−1(cid:17) ⊗ Kw(cid:17) η(t).
Let Kw = PBBT P, then one can derive from (11) and (12) that
V(t) ≤
N
Xi=2
¯xT
i (t)(cid:16)PA + AT P − (2λi + γ) PBBT P(cid:17) ¯xi(t).
(12)
Due to γ > 0 and λi > 0 (i = 2, 3, · · · , N), one can obtain that if PA + AT P − γPBBT P < 0,
then
V(t) ≤ −ε
N
Xi=2
i (t) ¯xi(t) = −εkη(t)k2
¯xT
for some positive real constant ε. Therefore, η(t) converges to 0 asymptotically, which
means that multiagent system (1) achieves leaderless adaptive consensus.
In the following, we determine the guaranteed-performance cost. It can be shown that
N
N
Xk=1
(xk(t) − xi(t))T Q (xk(t) − xi(t)) = xT (t) (2NLN ⊗ Q) x(t).
Xi=1
Due to xT (t) (LN ⊗ I) x(t) = ηT (t)η(t), it can be derived that
xT (t) (LN ⊗ Q) x(t) ≤
N
Xi=2
¯xT
i (t)Q ¯xi(t).
Let h > 0, then one can obtain by (13) and (14) that
N
Xi=2 Z h
0
∆
=
Jh
r
1
N
N
Xi=1
N
Xk=1 Z h
0
(xk(t) − xi(t))T Q (xk(t) − xi(t)) dt ≤
Moreover, one can show that
0
V(t)dt − V(h) + V(0) = 0.
Z h
Due to limt→+∞ (wik(t) − γik) = 0, one has
N
N
Xi=1
Xk=1,k,i
lim
h→+∞
(γik − wik(h)) = 0.
8
(13)
(14)
2 ¯xT
i (t)Q ¯xi(t)dt.
(15)
(16)
(17)
If PA + AT P − γPBBT P + 2Q ≤ 0, then one can obtain from (15) to (17) that
lim
h→+∞
r ≤ ηT (0) (IN−1 ⊗ P) η(0) +
Jh
γ
2N
N
N
Xi=1
Xk=1,k,i
(γik − wik(0)).
Since η(t) = (cid:2)0(N−1)d×d, I(N−1)d(cid:3)(cid:16)U T ⊗ Id(cid:17) x(t) and ¯U ¯U T = LN, one has
ηT (0) (IN−1 ⊗ P) η(0) = xT (0) IN −
Due to limt→+∞ (wik(t) − γik) = 0, one can show that
Xi=1
wik(t)dt = 2NZ +∞
N! ⊗ P! x(0).
(γik − wik(0)) =
N
Xk=1,k,iZ +∞
0
Xk=1,k,i
1N1T
N
Xi=1
1
N
N
N
xT (t) (LN ⊗ Kw) x(t)dt.
0
(18)
(19)
(20)
From (18) to (20), the conclusion of Theorem 1 can be obtained.(cid:3)
In [46], it was shown that the Riccati equation PA+AT P−γPBBT P+2Q = 0 has a unique
and positive definite solution P for any given γ > 0 if (A, B) is stabilizable. Moreover, from
the proof of Theorem 1, we eliminate the impacts of the nonzero eigenvalues of Lw(0) by
introducing a positive constant γ to construct the term (2λi + γ) PBBT P (i = 2, 3, · · · , N).
Thus, γ can be regarded as the rightward translated quantity of the nonzero eigenvalues of
2Lw(0) and can be given previously.
Furthermore, a large γ may regulate the consensus control gain by Theorem 1, so we
can choose some proper γ and P to regulate the consensus control gain. We introduce a
gain factor δ > 0 such that P ≤ δI, where δ can also be regarded as an upper bound of the
eigenvalue of P. Thus, one can show that PBBT P ≤ δ2BBT if the maximum eigenvalue of
BBT is not larger than 1. Based on LMI techniques, by Schur complement lemma in [45],
an adaptive guaranteed-performance consensualization criterion with a given gain factor
is proposed as follows.
Corollary 1. For any given gain factor δ > 0, multiagent system (1) is leaderless adap-
tively guaranteed-performance consensualizable by protocol (2) if λmax(cid:16)BBT(cid:17) ≤ 1 and
there exist γ > 0 and PT = P ≥ δ−1I such that
Ξ = " A P + PAT − γBBT 2 PQ
−2Q # < 0.
In this case, Ku = BT P−1, Kw = P−1BBT P−1 and the guaranteed-performance cost satisfies
that
J∗r =
N
Xi=2
δk ¯xi(0)k2 + γδ2Z +∞
0
dt!.
(cid:13)(cid:13)(cid:13)
2
BT ¯xi(t)(cid:13)(cid:13)(cid:13)
9
In the following, we give an approach to determine the consensus motion. Due to
1 U T = 1T
eT
N. √N, one has
1 ⊗ Id(cid:17)(cid:16)U T ⊗ Id(cid:17) x(0) =
¯x1(0) = (cid:16)eT
1
√N
N
Xi=1
xi(0).
Because subsystem (5) describes the consensus motion of multiagent system (1), the fol-
lowing corollary can be obtained by (8).
Corollary 2. If multiagent system (1) achieves leaderless adaptive guaranteed-performance
consensus, then
lim
t→+∞
xi(t) − eAt
1
N
N
Xi=1
xi(0)
= 0 (i = 1, 2, · · · , N) .
Remark 2. In [39–42], guaranteed-performance consensus criteria are associated with
the Laplacian matrix of the interaction topology or its minimum and maximum nonzero
eigenvalues, which means that global information of the whole system is required es-
sentially and their approaches are not completely distributed. A novel dilation adaptive
consensus strategy was proposed to realize completely distributed consensus control in
[43] , where guaranteed-performance constraints were not considered and the dilation
factor is inversely proportional to the minimum nonzero eigenvalue. When guaranteed-
performance constraints are considered, the precise value of the dilation factor is required.
However, to obtain the precise value of the dilation factor, it is necessary to determine the
minimum nonzero eigenvalue. In Theorem 1, the translation-adaptive strategy is proposed
to deal with guaranteed-performance constraints without using information of the Lapla-
cian matrix. Actually, we realize the eigenvalue translation by the special property of the
Laplacian matrix of a complete graph; that is, all its eigenvalues are identical.
3. Leader-follower adaptive guaranteed-performance consensualization
This section focuses on leader-follower adaptive guaranteed-performance consensus,
where multiagent systems consist of a leader and N − 1 followers. Sufficient conditions
for leader-follower adaptive guaranteed-performance consensualization are presented on
the basis of the Ricatti inequality and LMIs, respectively, and the guaranteed-performance
cost is determined.
3.1. Problem description for leader-follower cases
Without loss of generality, we set that agent 1 is the leader and the other N − 1 agents
are followers. Thus, the dynamics of multiagent systems with the leader-follower structure
can be described by
10
xi(t) = Axi(t) + Bui(t),
( x1(t) = Ax1(t),
where i = 2, 3, · · · , N, A ∈ Rd×d, B ∈ Rd×p, x1(t) is the state of the leader, and xi(t) and ui(t)
are the state and the control input of the ith follower, respectively. The leader does not
receive any information from followers and there at least exists an undirected path from
the leader to each follower. Furthermore, it is assumed that the subgraph among followers
is undirected but can be unconnected.
(21)
A new consensus protocol is proposed to realize adaptive guaranteed-performance track-
ing as follows
ui(t) = li1wi1(t)Ku (x1(t) − xi(t)) + Ku Pk∈Ni,k,1
wi1(t) = (x1(t) − xi(t))T Kw (x1(t) − xi(t)) ,
Jl = J f l + J f f ,
(xk(t) − xi(t)),
(22)
where i = 2, 3, · · · , N, Ku ∈ Rp×d and Kw ∈ Rd×d are gain matrices with KT
w = Kw ≥ 0, Ni
denotes the neighbor set of agent i, l11 = 0, li1 = 1 if agent i can receive the information
of the leader and li1 ≡ 0 otherwise, li1wi1(t) with wi1(0) = 1 is the adaptively regulated
interaction strength from the leader to the ith follower, it is supposed that the upper bound
of wi1(t) is γi1, and for QT = Q > 0,
N
0
J f l =
J f f =
Xi=2 Z +∞
Xi=2
N − 1
1
N
li1(x1(t) − xi(t))T Q (x1(t) − xi(t)) dt,
Xk=2 Z +∞
(xk(t) − xi(t))T Q (xk(t) − xi(t)) dt.
N
0
From protocol (22), interaction strengths from the leader to followers are adaptively time-
varying, but interaction strengths among followers are time-invariant and are equal to 1.
Now, we give the definition of the leader-follower adaptive guaranteed-performance con-
sensualization as follows.
Definition 2. Multiagent system (21) is said to be leader-follower adaptively guaranteed-
performance consensualizable by protocol (22) if there exist Ku and Kw such that limt→+∞
(xi(t) − x1(t)) = 0 (i = 2, 3, · · · , N) and Jl ≤ J∗l for any bounded initial states xi(0)(i = 1, 2,
· · · , N), where J∗l is said to be the guaranteed-performance cost.
In the following, an approach is given to design gain matrices Ku and Kw such that
multiagent system (21) with protocol (22) achieves leader-follower adaptive guaranteed-
performance consensus and the guaranteed-performance cost J∗l is determined. Further-
more, it is revealed that the consensus control gain can be regulated by introducing a gain
factor.
11
3.2. Adaptive guaranteed-performance consensus design for leader-follower cases
2 (t), ξT
, then one can
N(t)iT
3 (t), · · · , ξT
obtain by (21) and (22) that
Let ξi(t) = xi(t)− x1(t) (i = 2, 3, · · · , N) and ξ(t) = hξT
ξ(t) = (cid:16)IN−1 ⊗ A −(cid:16)L f f + Λw(t), f l(cid:17) ⊗ BKu(cid:17) ξ(t),
where L f f is the Laplacian matrix of the interaction topology among followers and Λw(t), f l =
diag{l21w21(t), l31w31(t), · · · , lN1wN1(t)} denotes the interaction from the leader to followers.
If limt→+∞ξ(t) = 0, then multiagent (21) with protocol (22) achieves leader-follower con-
sensus. Let x(t) = hxT
, then the following theorem presents a leader-
follower adaptive guaranteed-performance consensualization criterion and determines the
guaranteed-performance cost.
2 (t), · · · , xT
N(t)iT
1 (t), xT
(23)
Theorem 2. For any given translation factor γl > 0, multiagent system (21) is leader-
follower adaptively guaranteed-performance consensualizable by protocol (22) if there
exists a matrix RT = R > 0 such that RA + AT R − γlRBBT R + 3Q ≤ 0. In this case,
Ku = BT R, Kw = RBBT R and the guaranteed-performance cost satisfies that
J∗l = xT (0) " N − 1 −1T
N−1
IN−1
−1N−1
#⊗R! x(0)+γlZ +∞
0
xT (t) " N − 1 −1T
N−1
IN−1
−1N−1
#⊗RBBT R! x(t)dt.
Proof. Construct a new Lyapunov function candidate as follows
V(t) = ξT (t) (IN−1 ⊗ R) ξ(t) +
N
Xi=2
li1(wi1(t) − wik(0))2 + γl
N
Xi=2
(γi1 − wi1(t)).
Due to RT = R > 0 and γi1 ≥ wi1(t)(i = 2, 3, · · · , N), one can obtain that V(t) ≥ 0. Let
Ku = BT R and Kw = RBBT R, then the derivative of V(t) with respect to time t can be given
by (23) as follows
V(t) = ξT (t)(cid:16)IN−1 ⊗(cid:16)RA + AT R(cid:17) − 2(cid:16)L f f + Λw(t), f l(cid:17) ⊗ RBBT R(cid:17) ξ(t)
N
N
+
Xi=2
2li1 (wi1(t) − wi1(0)) wi1(t) − γl
wi1(t).
Xi=2
By (22), then one can show that
N
Xi=2
wi1(t) = ξT (t)(cid:16)IN−1 ⊗ RBBT R(cid:17) ξ(t),
12
(24)
(25)
N
li1 (wi1(t) − wi1(0)) wi1(t) = ξT (t)(cid:16)Λw(t), f l − Λw(0), f l(cid:17) ξ(t).
Xi=2
Let ξ(t) = (cid:16) U ⊗ IN−1(cid:17) ξ(t), where U T (cid:16)L f f + Λw(0), f l(cid:17) U = diagn λ2, λ3, · · · λNo, then it can
be found from (24) to (26) that
(26)
V(t) ≤
N
Xi=2
ξT
i (t)(cid:16)RA + AT R − (2 λi + γl)RBBT R(cid:17) ξi(t).
Due to L f f + Λw(0), f l are positive, one has λi > 0 (i = 2, 3, · · · , N). Hence, if RA + AT R −
γlRBBT R < 0, then
N
V(t) ≤ −κ
Xi=2 (cid:13)(cid:13)(cid:13)
ξi(t)(cid:13)(cid:13)(cid:13)
2
= −κkξ(t)k2
for some κ > 0. Hence, ξ(t) converges to 0 asymptotically; that is, multiagent system (21)
with protocol (22) achieves leader-follower adaptive consensus.
In the following, the guaranteed-performance cost is determined. Since the interaction
topology among followers is undirected, it can be derived for h > 0 that
N
Xi=2
N
Xk=2 Z h
0
(xk(t) − xi(t))T Q (xk(t) − xi(t)) dt
Jh
f f
∆
=
1
N − 1
= Z h
0
ξT (t) (2LN−1 ⊗ Q) ξ(t)dt.
Furthermore, due to wi1(0) = 1, one can show that
li1(x1(t) − xi(t))T Q (x1(t) − xi(t)) dt
Jh
f l
N
0
∆
=
Xi=2 Z h
= Z h
0
ξT (t)(cid:16)Λw(0), f l ⊗ Q(cid:17) ξ(t)dt.
(27)
(28)
Since all the eigenvalues of the Laplacian matrix LN−1 of the completed graph with weights
N − 1 are 1, all the eigenvalues of 2LN−1 + Λw(0), f l are positive and less than 3, so it can be
derived by (27) and (28) that
Jh
l
∆
= Jh
f f + Jh
f l ≤ Z h
0
ξT (t) (3IN−1 ⊗ Q) ξ(t)dt.
13
(29)
Moreover, it can be shown that
V(t)dt − V(h) + V(0) = 0.
Z h
Due to limt→+∞ (wi1(t) − γi1) = 0 (i = 2, 3, · · · , N), one can show that
0
N
Xi=2
(γi1 − wi1(h)) = 0.
lim
h→+∞
If RA + AT R − γlRBBT R + 3Q ≤ 0, then it can be derived from (29) to (31) that
N
lim
h→+∞
Jh
l ≤ ξT (0) (IN−1 ⊗ R) ξ(0) + γl
Xi=2
(γi1 − wi1(0)).
It can be shown that
ξT (0) (IN−1 ⊗ R) ξ(0) = xT (0) " N − 1 −1T
N−1
−1N−1
IN−1
wi1(t)dt = Z +∞
xT (t) " N − 1 −1T
N−1
IN−1
From (32) to (34), the conclusion of Theorem 2 can be obtained.(cid:3)
# ⊗ R! x(0),
Xi=2 Z +∞
(γi1 − wi1(0)) =
−1N−1
N
N
Xi=2
0
0
# ⊗ Kw! x(t)dt.
(30)
(31)
(32)
(33)
(34)
If (A, B) is stabilizable, then RA + AT R − γlRBBT R + 3Q = 0 has a unique and positive
definite solution. Hence, it is not difficult to find that a necessary and sufficient condition
for leader-follower adaptive guaranteed-performance consensualizability is that (A, B) is
stabilizable. Furthermore, γl can be regarded as the rightward translated quantity of the
nonzero eigenvalues of 2(L f f + Λw(0), f l). Moreover, similar to the analysis of Corollary
1, the following corollary gives an approach to regulate the consensus control gain by
introducing a gain factor δl > 0 such that R ≤ δlI.
Corollary 3. For any given gain factor δl > 0, multiagent system (21) is leader-follower
adaptively guaranteed-performance consensualizable by protocol (22) if λmax(cid:16)BBT(cid:17) ≤ 1
and there exist γl > 0 and RT = R ≥ δ−1
Θ = " A R + RAT − γlBBT 3 RQ
−3Q # < 0.
l I such that
In this case, Ku = BT R−1, Kw = R−1BBT R−1 and the guaranteed-performance cost satisfies
that
J∗l =
N
Xi=2
δlkξi(0)k2 + γlδ2
l Z +∞
0
dt!.
(cid:13)(cid:13)(cid:13)
2
BT ξi(t)(cid:13)(cid:13)(cid:13)
14
In the above main results, the Riccati inequality method and the variable changing ap-
proach are used to determine gain matrices of consensus protocols. The variable changing
approach does not introduce any conservatism because it is an equivalent transformation.
However, the Riccati inequality method may bring in some conservatism due to the scala-
bility of the Lyapunov function. It was shown that the Riccati inequality method is exten-
sively used in optimization control and usually has less conservatism in [46]. Moreover,
there are two critical difficulties in obtaining Theorems 1 and 2. The first one is to design
a proper Lyapunov function which can translate the nonzero eigenvalues of the Laplacian
matrix of the interaction topology rightward. The second one is to construct the relation-
ship between the linear quadratic index and the Laplacian matrix of the interaction topol-
ogy, as shown in (13) and (14) for leaderless cases and (27) and (28) for leader-follower
cases.
Remark 3. According to the leader-follower structure feature, it is designed that interac-
tion strengths from the leader to followers are adaptively adjusted in protocol (22), but
interaction strengths among followers are fixed. For leaderless cases, it is required that in-
teraction strengths among all neighboring agents are adaptively time-varying as shown in
protocol (2). It can be found that fixed interaction weights among followers can simplify
the analysis and design of the whole system. Furthermore, it should be pointed that con-
sensus performance among no neighboring agents for both leaderless and leader-follower
cases can be adjusted by gain matrices given in Theorems 1 and 2, which reveals that mul-
tiagent systems can regulate global performance by local and indirect interactions. More-
over, compared with the guaranteed-performance cost in [39–42], both J∗r and J∗l involve
integral terms, which are introduced by adaptively time-varying interaction weights.
Remark 4. Both the guaranteed-performance cost J∗r and J∗l are associated with the initial
states of all agents, and the key differences between them are the structures of coupling ma-
trices. It can be found that the coupling matrix IN − N−11N1T
N in J∗r is the Laplacian matrix
of a complete graph with edge weights N−1, which means that the guaranteed-performance
cost J∗r is jointly determined by state errors of all agents. However, the coupling matrix
# in J∗l is the Laplacian matrix of a star graph with edge weights 1, where
" N − 1 −1T
N−1
−1N−1
IN−1
the leader is the central node and there do not exist interactions among followers. In this
case, the guaranteed-performance cost J∗l is decided by state errors between the leader and
all followers, but it is independent of state errors among followers. Actually, the two cou-
pling matrices intrinsically reflect the impacts of topology structures of multiagent systems
on the guaranteed-performance cost.
15
4. Numerical simulations
This section presents two numerical simulation examples to demonstrate theoretical re-
sults for both leaderless multiagent systems and leader-follower multiagent systems.
4.1. Numerical simulation for leaderless cases
6
5
1
4
2
3
Fig. 1: Interaction topology of the leaderless multiagent supporting system.
Ma et al.
[47] proposed the model of the multiagent supporting system which has
potential applications in earthquake damage-preventing buildings, water-floating plants
and large-diameter parabolic antennae, and gave a centralized control approach. Xi et
et al. [48] presented a distributed approach to control the multiagent supporting system,
where the dynamics of each agent with the self feedback matrix [−88.6792, 4.7642] can
be modeled as (1) with
A = "
1
0
1 # ,
−100 0 # , B = " 0
and the interaction topology of this leaderless multiagent supporting system is shown in
Fig. 1 with all edge weights equal to 1. Let
0 2 # ,
Q = " 1 0
then one can obtain from Theorem 1 that
2.1307 # ,
P = " 223.5978 3.9324
3.9324
Ku = [3.9324, 2.1307] ,
16
4.5399 # ,
Kw = " 15.4638 8.3788
8.3788
and the guaranteed-performance cost is J∗r = 1694.6.
)
t
(
1
x
2
1.5
1
0.5
0
−0.5
−1
−1.5
−2
0
10
5
0
−5
)
t
(
2
x
0.5
1
1.5
t / s
2
2.5
3
−10
0
0.5
1
2
2.5
3
1.5
t / s
Fig. 2: State trajectories of the leaderless multiagent supporting system.
250
200
150
100
50
0
0
0
0.5
0.5
1
1
J∗r
Jr
1.5
1.5
t / s
2500
2000
1500
1000
500
2
2
2.5
2.5
0
3
3
Fig. 3: Performance function of the leaderless multiagent supporting system.
Fig. 2 depicts the state trajectories of the leaderless multiagent supporting system, where
the trajectories marked by circles are the curves of the consensus function given in Corol-
lary 2 and the trajectories of states of six agents are represented by full curves. The tra-
jectory of the performance function Jh
r with h = 3 is shown in Fig. 3. One can see that
state trajectories of all agents converge to the ones marked by circles and the performance
function Jh
r < J∗r ; that is, this multiagent supporting
r converges to a finite value with Jh
17
system achieves leaderless adaptive guaranteed-performance consensus without using the
global information of the interaction topology. However, the distributed approach in [48]
required the precise value of the minimum eigenvalue of the interaction topology; that is,
the completely distributed control cannot be realized.
4.2. Numerical simulation for leader-follower cases
6
5
1
4
2
3
Fig. 4: Interaction topology of the leader-follower multiagent system.
Consider a leader-follower multiagent system with one leader and five followers, where
the interaction topology is shown in Fig. 4 with all edge weights equal to 1 and the dy-
namics of each agent is modeled by (21) with
1
1
−30 −12.5
0.5
0
16
0
0
0
30
0
0
1
−16 0
0.30 0.30 0.20 0.10
0.30 0.50 0.10 0.10
0.20 0.10 0.50 0.15
0.10 0.10 0.15 0.10
1
19
0
0
.
, B =
,
Let
A =
Q =
R =
then one can obtain by Theorem 2 that
7.7420 −0.1280 −6.0953 0.6304
0.0148
−0.1280
−6.0953
0.1259
0.7516
0.6304
0.1680
7.0299
0.1259
0.0404
0.1680
0.0148
,
18
Ku = [5.3100, 0.6396, − 2.9033, 0.9116] ,
28.1961
3.3963 −15.4165
−1.8570
3.3963
0.4091
8.4292
−15.4165 −1.8570
0.5831
4.8406
−2.6466
Kw =
and the guaranteed-performance cost is J∗l = 872.5.
4.8406
0.5831
−2.6466
0.8310
,
)
t
(
1
x
)
t
(
3
x
4
3
2
1
0
−1
−2
−3
−4
0
4
3
2
1
0
−1
−2
−3
−4
0
)
t
(
2
x
)
t
(
4
x
30
20
10
0
−10
−20
0
10
5
0
−5
−10
−15
0
0.5
1
1.5
t / s
2
2.5
3
0.5
1
1.5
t / s
2
2.5
3
0.5
1
1.5
t / s
2
2.5
3
0.5
1
1.5
t / s
2
2.5
3
Fig. 5: State trajectories of the leader-follower multiagent system.
The state trajectories of the leader-follower multiagent system are shown in Fig. 5,
where the trajectories marked by circles denote the states of the leader and the trajectories
of states of five followers are represented by full curves. In Fig. 6, the trajectory of the
performance function Jh
l with h = 3 is shown. It can be found that state trajectories of all
followers converge to the ones of the leader and the performance function Jh
l converges
to a finite value with Jh
l < J∗l , which means that this multiagent system achieves leader-
follower adaptive guaranteed-performance consensus without using the global information
19
1200
900
600
300
0
0
0.5
1
J∗l
Jl
1.5
t / s
2
2.5
3
Fig. 6: Performance function of the leader-follower multiagent system.
of the interaction topology. Moreover, it should be pointed out that the computational
complexity does not increase as the number of agents increases since the main results of
the current paper are completely distributed.
5. Conclusions
A new adaptive consensus scheme with a quadratic cost function was proposed to re-
alize the completely distributed guaranteed-performance consensus control. The adap-
tive guaranteed-performance consensualization criterion for the leaderless case was given
by regulating the interaction weights among all neighboring agents, and the adaptive
guaranteed-performance consensualization criterion for the leader-follower case was pre-
sented by regulating the interaction weights from the leader to its followers. Furthermore,
for leaderless and leader-follower multiagent systems, the regulation approaches of the
consensus control gain were proposed, respectively, which adjust control gains by choos-
ing the different translation factors. Moreover, explicit expressions of the guaranteed-
performance costs were given, where the coupling matrix associated with the leaderless
case is the Laplacian matrix of a complete graph and the coupling matrix associated with
the leader-follower case is the Laplacian matrix of a star graph with the leader being the
central node.
Furthermore, the future research directions can focus on two aspects. The first one is to
study the practical applications of multiagent systems combining main results in the cur-
rent paper with structure characteristics of practical multiagent systems, such as network
congestion control systems and single-link manipulator systems with a flexible joint, et al.
The other one is to investigate the impacts of directed topologies, time-varying delays and
given cost budgets on adaptive guaranteed-performance consensus of multiagent systems.
20
References
[1] Y. Ma & J. Zhao (2018). Distributed integral-based event-triggered scheme for co-
operative output regulation of switched multi-agent systems. Information Sciences,
doi: 10.1016/j.ins.2018.02.021.
[2] S. Zuo, Y. Song, F. L. Lewis, & A. Davoudi (2017). Output containment control
of linear heterogeneous multi-agent systems using internal model principle. IEEE
Transactions on Cybernetics, 47(8), 2099-2109.
[3] Z. Ji, & H. Yu (2017). A new perspective to graphical characterization of multi-agent
controllability. IEEE Transactions on Cybernetics, 47(6), 1471-1483.
[4] H. Du, G. Wen, Y. Cheng, Y. He, & R. Jia (2017). Distributed finite-time cooperative
control of multiple high-order nonholonomic mobile robots. IEEE Transactions on
Neural Networks and Learning Systems, 28(12), 2998-3006.
[5] X. Liu & Z. Ji (2017). Controllability of multi-agent systems based on path and cycle
graphs. International Journal of Robust and Nonlinear Control, 28(1), 296-309.
[6] H. Du & S. Li (2016). Attitude synchronization for flexible spacecraft with commu-
nication delays. IEEE Transactions on Automatic Control, 61(11), 3625-3630.
[7] Z. Zhang, G. Duan, & M. Hou (2018). An improved adaptive dynamic surface control
approach for uncertain nonlinear systems. International Journal of Adaptive Control
and Signal Processing, DOI: 10.1002/acs.2870.
[8] N. Cai, M. He, Q. Wu, & M. J. Khan (2018). On almost controllability of dynamical
complex networks with noises. Journal of Systems Science and Complexity, DOI:
10.1007/s11424-017-6273-7.
[9] H. Paggi, J.Soriano, & J. A. Lara (2018). A multi-agent system for minimizing infor-
mation indeterminacy within information fusion scenarios in peer-to-peer networks
with limited resources. Information Sciences, doi: 10.1016/j.ins.2018.04.019.
[10] D. Wang, Z. Wang, M. Chen, & W. Wang (2018). Distributed optimization for multi-
agent systems with constraints set and communication time-delay over a directed
graph. Information Sciences, 438, 1-14.
[11] C. Chen, Z. Liu, Y. Zhang, C. L. P. Chen, & S. Xie (2016). Saturated nussbaum
nunction based approach for robotic systems with unknown actuator dynamics. Au-
tomatica, 46(10), 2311-2322.
[12] X. Wu, K. Zhang, & M. Cheng (2017). Computational method for optimal control
of switched systems with input and state constraints. Nonlinear Analysis: Hybrid
Systems, 26, 1-18.
[13] S. Zuo, A. Davoudi, Y. Song, & F. L. Lewis (2016). Distributed finite-time voltage
and frequency restoration in islanded AC microgrids. IEEE Transactions on Indus-
trial Electronics, 63(10), 5988-5997.
21
[14] D. Wang, D. Wang, W. Wang, Y. Liu, & F. E. Alsaadi (2018). A modified dis-
tributed optimization method for both continuous-time and discrete-time multi-agent
systems. Neurocomputing, 275, 725-732.
[15] J. Xi, C. Wang, H. Liu, & Z. Wang (2018). Dynamic output feedback guaranteed-
cost synchronization for multiagent networks with given cost budgets. IEEE Access,
DOI: 10.1109/ACCESS.2018.2819989.
[16] B. Xu & F. Sun (2018). Composite intelligent learning control of strict-feedback
systems with disturbance. IEEE Transactions on Cybernetics, 48(2), 730-741.
[17] B. Xu, D. Wang, Y. Zhang, & Z. Shi (2017). DOB based neural control of flexible
hypersonic flight vehicle considering wind effects. IEEE Transactions on Industrial
Electronics, 64(11), 8676-8685.
[18] N. Cai, C. Diao, & M. J. Khan (2017). A new clustering method based on quasi-
consensus motions of dynamical multi-agent systems. Complexity, 4978613.
[19] R. Olfati-Saber, & R. M. Murray (2004). Consensus problems in networks of agents
with switching topology and time-delays. IEEE Transactions on Automatic Control,
49(9), 1520-1533.
[20] J. Xi, M. He, H. Liu, & J. Zheng (2016). Admissible output consensualization control
for singular multi-agent systems with time delays. Journal of the Franklin Institute,
353(16), 4074-4090.
[21] J. Lian & J. Wang (2015). Passivity of switched recurrent neural networks with time-
varying delays. IEEE Transactions on Neural Networks and Learning Systems, 26(2),
357-366.
[22] Y. Zheng, J. Ma, & L. Wang (2017). Consensus of hybrid multi-agent systems. IEEE
Transactions on Neural Networks and Learning Systems, DOI: 10.1109/TNNL-
S.2017.2651402.
[23] X. Lin & Y. Zheng (2017). Finite-time consensus of switched multiagent systems.
IEEE Transactions on Systems, Man, and Cybernetics: Systems, 47(7), 1535-1545.
[24] J. Xi, Z. Shi, & Y. Zhong (2012). Output consensus analysis and design for high-
order linear swarm systems: Partial stability method. Automatica, 48(9), 2335-2343.
[25] Z. Wu, Y. Xu, R. Lu, Y. Wu, & T. Huang (2018). Event-triggered control for consen-
sus of multi-agent systems with fixed/switching topologies. IEEE Transactions on
Systems, Man and Cybernetics: Systems, DOI: 10.1109/TSMC.2017.2744671.
[26] C. Chen, C. Wen, Z. Liu, K. Xie, Y. Zhang, & C. L. P. Chen (2017). Adaptive con-
sensus of nonlinear multi-agent systems with non-identical partially unknown control
directions and bounded modelling errors. IEEE Transactions on Automatic Control,
62(9), 4654-4659.
[27] Z. Wu, Y. Xu, Y. Pan, P. Shi, & Q. Wang (2018). Event-triggered pinning control for
consensus of multi-agent systems with quantized information. IEEE Transactions on
Systems, Man and Cybernetics: Systems, DOI: 10.1109/TSMC.2017.2773634.
22
[28] C. Chen, C. Wen, Z. Liu, K. Xie, Y. Zhang, & C. L. P. Chen (2017). Adaptive asymp-
totic control of multivariable systems based on a one-parameter estimation approach.
Automatica, 83, 124-132.
[29] Z. Wu, Y. Xu, Y. Pan, H. Su, & Y. Tang (2018). Event-triggered control for consen-
sus problem in multi-agent systems with quantized relative state measurement and
external disturbance. IEEE Transactions on Circuit and Systems I: Regular paper,
DOI: 10.1109/TCSI.2017.2777504.
[30] Z. Liu, X. Yu, Z. Guan, B. Hu, & C. Li (2017). Pulse-modulated intermittent con-
trol in consensus of multi-agent systems. IEEE on Systems, Man, and Cybernetics:
Systems, 47(5), 783-793.
[31] Z. Liu, Z. Guan, X. Shen, & G. Feng (2012). Consensus of multi-agent networks
with aperiodic sampled communication via impulsive algorithm using position-only
measurements. IEEE Transactions on Automatic Control, 57(10), 2639-2643.
[32] J. Xi, C. Wang, H. Liu, & L. Wang (2018). Completely distributed guaranteed-
performance consensualization for high-order multiagent systems with switching
topologies. IEEE Transactions on Systems, Man, and Cybernetics: Systems, In press.
[33] W. Liu & J. Huang (2017). Adaptive leader-following consensus for a class of higher-
order nonlinear multi-agent systems with directed switching networks. Automatica,
79, 84-92.
[34] B. Hu, Z. H. Guan, X. W. Jiang, M. Chi, & L. Yu (2016). On consensus perfor-
mance of nonlinear multi-agent systems with hybrid control. Journal of the Franklin
Institute, 353(13), 3133-3150.
[35] E. Semsar-Kazerooni & K. Khorasani (2009). An optimal cooperation in a team of
agents subject to partial information. International Journal of Control, 82(3), 571-
583.
[36] K. G. Vamvoudakis, F. L. Lewis, & G. R. Hudas (2012). Multi-agent differential
graphical games: Online adaptive learning solution for synchronization with opti-
mality. Automatica, 48(8), 1598-1611.
[37] Y. C. Cao & W. Ren (2010). Optimal linear-consensus algorithms: An LQR per-
spective. IEEE Transactions on Systems, Man, and Cybernetics-Part B: Cybernetics,
40(3), 819-829.
[38] Z. H. Guan, B. Hu, M. Chi, D. X. He, & X. M. Cheng (2014). Guaranteed perfor-
mance consensus in second-order multi-agent systems with hybrid impulsive control.
Automatica, 50(9), 2415-2418.
[39] J. Xi, Z. Fan, H. Liu, & T. Zheng (2018). Guaranteed-cost consensus for multiagent
networks with Lipschitz nonlinear dynamics and switching topologies. International
Journal of Robust and Nonlinear Control, 28(7), 2841-2852.
[40] X. Zhou, P. Shi, C. Lim, C. Yang, & W. Gui (2015). Event based guaranteed-cost con-
sensus for distributed multi-agent systems. Journal of the Franklin Institute, 352(9),
23
3546-3563.
[41] Y. Zhao & W. Zhang (2017). Guaranteed cost consensus protocol design for linear
multi-agent systems with sampled-data information: An input delay approach. ISA
Transactions, 67, 87-97.
[42] J. Xu, G. Zhang, J. Zeng, J. Xi, & B. Du (2017). Robust guaranteed cost consensus
for high-order discrete-time multi-agent systems with parameter uncertainties and
time-varying delays. IET Control Theory & Applications, 11(5), 647-667.
[43] Z. Li, G. Wen, Z. Duan, & W. Ren (2015). Designing fully distributed consensus
protocols for linear multi-agent systems with directed graphs. IEEE Transactions on
Automatic Control, 60(4), 1152-1157.
[44] C. Godsil & G. Royal (2001). Algebraic Graph Theory, New York: Springer-Verlag.
[45] S. Boyd, L. E. Ghaoui, E. Feron, & V. Balakrishnan (1994). Linear Matrix Inequali-
ties in System and Control Theory, SIAM, Philadelphia, PA.
[46] B. D. O. Anderson & J. B. Moore (1989). Optimal Control: Linear Quadratic Meth-
ods, Prentice-Hall International, Inc.
[47] S. Ma, S. Hackwood, & G. Beni (1994). Multi-agent supporting systems (MASS):
control with centralized estimator of disturbance. in Proceeding IEEE/RSJ Interna-
tional Conference Intelligent Robots and Systems, 679-686.
[48] J. Xi, N. Cai, & Y. Zhong (2010). Consensus problems for high-order linear time-
invariant swarm systems. Physica A: Statistical Mechanics and its Applications,
389(24): 5619-5627.
24
|
1709.09569 | 1 | 1709 | 2017-09-27T15:03:48 | Traffic Optimization For a Mixture of Self-interested and Compliant Agents | [
"cs.MA",
"cs.AI"
] | This paper focuses on two commonly used path assignment policies for agents traversing a congested network: self-interested routing, and system-optimum routing. In the self-interested routing policy each agent selects a path that optimizes its own utility, while the system-optimum routing agents are assigned paths with the goal of maximizing system performance. This paper considers a scenario where a centralized network manager wishes to optimize utilities over all agents, i.e., implement a system-optimum routing policy. In many real-life scenarios, however, the system manager is unable to influence the route assignment of all agents due to limited influence on route choice decisions. Motivated by such scenarios, a computationally tractable method is presented that computes the minimal amount of agents that the system manager needs to influence (compliant agents) in order to achieve system optimal performance. Moreover, this methodology can also determine whether a given set of compliant agents is sufficient to achieve system optimum and compute the optimal route assignment for the compliant agents to do so. Experimental results are presented showing that in several large-scale, realistic traffic networks optimal flow can be achieved with as low as 13% of the agent being compliant and up to 54%. | cs.MA | cs | Traffic Optimization For a Mixture of Self-interested and Compliant Agents
Guni Sharon1∗, Michael Albert2∗, Tarun Rambha3∗, Stephen Boyles1, Peter Stone1
1University of Texas at Austin
2Duke University
3Cornell University
[email protected], [email protected], [email protected], [email protected], [email protected]
7
1
0
2
p
e
S
7
2
]
A
M
.
s
c
[
1
v
9
6
5
9
0
.
9
0
7
1
:
v
i
X
r
a
Abstract
This paper focuses on two commonly used path assign-
ment policies for agents traversing a congested network: self-
interested routing, and system-optimum routing. In the self-
interested routing policy each agent selects a path that op-
timizes its own utility, while the system-optimum routing
agents are assigned paths with the goal of maximizing system
performance. This paper considers a scenario where a cen-
tralized network manager wishes to optimize utilities over all
agents, i.e., implement a system-optimum routing policy. In
many real-life scenarios, however, the system manager is un-
able to influence the route assignment of all agents due to lim-
ited influence on route choice decisions. Motivated by such
scenarios, a computationally tractable method is presented
that computes the minimal amount of agents that the sys-
tem manager needs to influence (compliant agents) in order to
achieve system optimal performance. Moreover, this method-
ology can also determine whether a given set of compliant
agents is sufficient to achieve system optimum and compute
the optimal route assignment for the compliant agents to do
so. Experimental results are presented showing that in sev-
eral large-scale, realistic traffic networks optimal flow can be
achieved with as low as 13% of the agent being compliant
and up to 54%.
Introduction
In multiagent systems, there are generally two paradigms of
interaction. Centralized control paradigms assume that a sin-
gle decision making entity is able to dictate the actions of all
the agents, thus leading them to a coordinated social opti-
mum. Decentralized control paradigms, on the other hand,
assume that each agent selects its own actions, and while it
is in principle possible for them to act altruistically, they are
generally assumed to be self-interested.
In this paper, we consider a routing scenario in which a
subset of agents are controlled centrally (compliant agents),
while the remaining are self-interested agents. We model
the system as a Stackelberg routing game (Yang, Zhang,
and Meng 2007) in which the decision maker for the cen-
trally controlled agents is the leader, and the self-interested
agents are the followers. In this paper, we provide a compu-
tationally tractable methodology for 1) determining whether
a given subset of centrally controlled agents are sufficient
∗These authors contributed equally.
to achieve system optimum (SO), 2) determining the maxi-
mum number of agent that may be self-interested such that
the centrally controlled agents can be deployed in order to
induce SO, and 3) computing the actions the leader should
prescribe to a sufficient set of compliant agents in order to
achieve SO.
It has been known for nearly a century in routing games
that agents seeking to minimize their private latency need
not minimize the total system's latency (Pigou 1920; Rough-
garden and Tardos 2002). That is, self-interested agents may
reach a user equilibrium (UE) that is not optimal from
a system perspective. However, if all agents are assigned
paths with minimum system marginal cost then the system
will achieve optimal performance (Pigou 1920; Beckmann,
McGuire, and Winsten 1956; Dietrich 1969).
Therefore, from a system manager perspective, it is de-
sirable that all agents traversing a network would strictly
utilize minimal marginal cost paths, even if the path is
not a minimum latency path for an individual agent. How-
ever, in many important scenarios, it will not be possible
to enforce path assignment on all agents, but it may be
possible to affect the behavior of a subset (the compliant
agents). As a motivating example, consider an opt-in tolling
system where drivers are given positive incentives to en-
roll but, in exchange, they will be subject to tolls that af-
fect their route choice (Sharon et al. 2017). Another rele-
vant example is virtual private network (VPN) path allo-
cation. While each packet within the VPN might be self-
interested, a pro-social network manager might allocate vir-
tual paths that are different from those preferred by the
self-interested packets (Fingerhut, Suri, and Turner 1997;
Duffield et al. 1999).
However, we show that, in the general case, computing the
optimal assignment of compliant agents is NP-hard. There-
fore, we focus on the specific scenario where the portion
of compliant agents is sufficiently large to achieve SO. We
present a novel linear program (LP ) representation for com-
puting the maximal portion of self-interested agents that al-
low the system to achieve SO and to determine whether a
given set of compliant agents is sufficient to achieve SO.
Furthermore, we provide a method to tractably compute the
flow assignment for the compliant agents such that SO per-
formance is guaranteed.
We demonstrate, using a standard traffic simulator over a
(cid:80)
fp = R(s, t).
paths of which e is a part. Define the system flow vector as
f = vect{fp}. f is said to be feasible if for all s, t ∈ V ,
p∈Ps,t
Each link e ∈ E has a latency function le(fe) which,
given a flow volume (fe), returns the latency (travel time)
on e. Following Roughgarden and Tardos (2002) we make
the following assumption:
Assumption 1. The latency function le(fe) is non-negative,
differentiable, and non-decreasing for each link e ∈ E.
defined as lp(f ) = (cid:80)
d
e∈p c(cid:48)
p(f ) = (cid:80)
The latency of a simple path p for a given flow f, is
e∈p le(fe). A feasible flow f is de-
fined as a user equilibrium (UE) if for every s, t ∈ V and
pa, pb ∈ Ps,t with fpa > 0 it holds that lpa (f ) ≤ lpb (f ) (see
Lemma 2.2 in (Roughgarden and Tardos 2002)). In other
words, at UE, no amount of flow can be rerouted to a path
with lower latency when the rest of the flow is fixed.
le(fe)fe, the cost of a path p as cp(f ) = (cid:80)
the cost of a flow f as c(f ) =(cid:80)
Define the system cost associated with link e as ce(fe) =
e∈p ce(fe) and
e∈E ce(fe). Define c(cid:48)
e(x) =
dx ce(x) and c(cid:48)
e(fe). A feasible flow f is
defined as a system optimum (SO) flow if for every s, t ∈ V
and pa, pb ∈ Ps,t with fpa > 0 it holds that c(cid:48)
pa (f ) ≤
c(cid:48)
(f ) (see Lemma 2.5 in (Roughgarden and Tardos 2002)).
pb
In other words, at SO, the benefit from reducing the flow
along any path is always less than or equal to the cost of
by adding the same amount of flow to a parallel, alternative
path. We follow Roughgarden and Tardos (2002), and make
the following assumption:
Assumption 2. The cost function ce(fe) is convex for each
link e ∈ E.
objective is to minimize c(f ) =(cid:80)
Assumptions 1 and 2 imply that the set of SO flows corre-
spond to the set of solutions of a convex program where the
e∈E ce(fe) (see Rough-
wide range of road networks, that the number of compliant
agents necessary to achieve system optimum is a relatively
small percentage of total flow (between 13% and 53%).
Motivation
Recent advances in GPS based tolling technology (Numrich,
Ruja, and Voss 2012) open the possibility of implementing
micro-tolling systems in which specific tolls are charged for
the use of every link within a road network. Setting tolls
appropriately can influence self-interested drivers to prefer
paths with minimum system marginal cost and thus, lead to
improved system performance (Sharon et al. 2017).
Unfortunately, political factors deter public officials from
allowing such a micro-tolling scheme to be realized. Road
pricing is known to cause a great deal of public unrest and is
thus opposed by governmental institutions (Schaller 2010).
To tackle this issue and avoid public unrest, it would be ben-
eficial to have an opt-in micro-tolling system where, given
some initial monetary sign-up incentive, drivers choose to
opt-in to the system and be charged for each journey they
take based on their chosen route. The vehicles belonging to
such drivers would need to be equipped with a GPS device as
well as a computerized navigation system. Given the toll val-
ues and driver's value of time, the navigation system would
suggest a minimal cost route where the cost is a function of
the travel time and tolls.
While addressing the issue of political acceptance, an
opt-in system would result in traffic that is composed of
a mixture of self-interested and compliant agents (compli-
ant in the sense that the system manager can influence their
route choice). Such a scenario raises some practical ques-
tions which are the focus of this paper, namely, what portion
of self-interested agents can the system tolerate while still
reaching optimum performance? The answer to this ques-
tion can help practitioners to determine both the level and
the targeting of incentives in an opt-in system.
garden and Tardos (2002) Corollary 2.7).
Problem definition and terminology
The terminology in this paper follows that of Roughgarden
and Tardos (2002). We review the relevant concepts and no-
tation in this section.
The flow model
The flow model in this work is composed of a directed graph
G(V, E), and a demand function R(s, t) → R+ mapping
a pair of vertices s, t ∈ V to a non-negative real number
representing the required amount of flow between source, s,
and target, t.1 An instance of the flow model is a {G, R}
pair.Ps,t denotes the set of acyclic paths from s to t. Define P
as the collection of all Ps,t (i.e., ∪s,t∈V Ps,t). The variable
fp represents the flow volume assigned to path p. Similarly,
fe is the flow volume assigned to link e. By definition, the
flow on each link (fe) equals the summation of flows on all
Problem Definition
The focus of this paper is a scenario where the demand is
partitioned into self-interested and compliant agents. We de-
fine two types of controllers that assign paths to all of the
agents. These controllers are viewed as players in a Stackel-
berg game (Yang, Zhang, and Meng 2007).
• SO-controller - Stackelberg leader, the SO-controller as-
pires to assign paths to the compliant subset of agents that,
taking into account the self-interested agents' reaction,
optimizes the systems performance (i.e. minimizes total
latency). SO-controller assigned flow will be referred to
as compliant flow.
• UE-controller - Stackelberg follower, considering the
compliant agents' path assignment as fixed, the UE-
controller assigns paths to the self-interested agents, the
UE flow, such that a state of user equilibrium (as defined
above) is achieved.2
1The demand between any source and target, R(s, t), can be
viewed as an infinitely divisible set of agents (also known as a non-
atomic flow (Rosenthal 1973)).
2The UE enforced by the UE-controller applies only for the
self-interested subset of agents. That is, no self-interested agent
can benefit from unilaterally deviating from its assigned path.
The problems addressed in this paper are:
1. Given an instance of the flow model {G, R}, what is the
maximum amount of self-interested agents that can be as-
signed to the UE controller and still permit the SO con-
troller to achieve system optimum?
2. Given a set of compliant agents and an instance of the
flow model {G, R}, can the SO controller assign paths to
them in such a way that the system achieves SO?
3. If SO is achievable, how should the SO-controller as-
sign the compliant flow? Equivalently, what is the optimal
Stackelberg equilibrium?
To the best of our knowledge, this work is the first to answer
these questions in a general setting.
Related Work
Previous work examined mixed equilibrium scenarios where
traffic is composed of: UE and Cournot-Nash (CN) con-
trollers. A CN-controller assigns flows to a given subset of
the demand with the aim of minimizing the total travel time
only for that subset. For instance, a logistic company with
many trucks can be viewed as a CN-controller.
It was shown that the equilibrium for a mixed UE, CN
scenario is unique and can be computed using a convex pro-
gram (Haurie and Marcotte 1985; Yang and Zhang 2008).
On the other hand, no tractable algorithm is known for com-
puting the optimal Stackelberg equilibrium for scenarios that
also include a SO-controller.
Korilis et. al. (1997) examined mixed equilibrium scenar-
ios that do include a SO-controller. In their work, a tech-
nique for computing the a solution for the above questions
#1 and #3 (see problem definition) was suggested for spe-
cific types of flow models. Their technique was proven to
work for networks with a common source and a common
target with any number of parallel links. Moreover, the la-
tency functions were assumed to be of a very specific form
(linear function with a capacity bound). As a result, their so-
lution is not applicable when general networks with arbitrary
latency functions are considered.
Other work (Roughgarden 2004; Immorlica et al. 2009)
studied a variant of the scheduling problem where infinites-
imal jobs must be assigned to a set of shared machines
each of which is affiliated with a non-negative, differen-
tiable, and non-decreasing latency function that, given the
machine load, specify the amount of time needed to com-
plete a job. When considering a scenario where part of the
jobs are assigned to machines by a UE-controller while the
rest are assigned by a SO-controller, they show it is NP-hard
to compute the optimal Stackelberg equilibrium (Roughgar-
den 2004). Their problem can be viewed as a special case of
our problem, specifically a network with a single source and
target with multiple parallel links between them. Given that
in this more restrictive setting computing the optimal Stack-
elberg equilibrium is intractable, the general question in our
setting will also be computationally intractable.
Computing the Maximal UE Flow
We define r∗
Given that finding the optimal Stackelberg equilibrium is
NP-hard for an arbitrary number of compliant agents, this
work focuses on scenarios where the number of compliant
agents is sufficient to achieve SO. As we will show, finding
the optimal Stackelberg equilibrium can be done in polyno-
mial time for such cases. In this section, we will present a
computationally tractable method to compute the maximal
UE flow given an instance of a flow model {G, R}, and we
will provide a method to check, for a given level of compli-
ant flow, whether SO is achievable.
s,t r∗
s,t.
UE as the maximal amount of demand com-
prised of self-interested agents that the system can toler-
ate and still achieve SO. Additionally, we define r∗
s,t as the
amount of demand from source s to target t that is assigned
to the UE-controller. That is, computing r∗
UE is equivalent
to maximizing(cid:80)
We can cast the problem of maximizing(cid:80)
s,t as an
optimization problem, specifically a linear program (LP ).
Assigning values to all variables of type r∗
s,t must follow
some constraints. Specifically, the flow from each origin
to each destination must be both a subflow of some SO
flow, and it must be an acceptable path for all self-interested
agents in the flow, given the compliant flow.
Definition 1 (Subflow of flow f). For a directed graph
G(V, E) and demand function R, a flow f∗ is a subflow of
flow f if for all links e ∈ E, 0 ≤ f∗
e ≤ fe and for each pair
of nodes s, t ∈ V , there exists 0 ≤ rs,t ≤ R(s, t) such that
s,t r∗
(cid:88)
(cid:88)
e∈out(s)
f∗
e − (cid:88)
e − (cid:88)
e∈in(s)
e∈out(t)
f∗
e∈in(t)
(cid:88)
(cid:88)
t
s
f∗
e =
rs,t
f∗
e =
rs,t.
and
A path will only be acceptable for a self-interested agent if
it is the lowest latency path from the origin to the destination
given the compliant flow, and in order for the path to be part
of a valid SO flow, it must be a minimum marginal cost path.
Therefore, a path p, leading from vertex s to vertex t, will
be said to be zero reduced cost if there is no other path, p(cid:48),
leading from s to t with lower latency or lower marginal
cost.
Definition 2 (Zero reduced cost path). For a flow model
{G, R}, a zero reduced cost path with regard to flow assign-
ment f is a path p ∈ Ps,t such that ∀ p(cid:48) ∈ Ps,t : lp(f ) ≤
lp(cid:48)(f ) and c(cid:48)
p(cid:48)(f ). A link, e, is defined as a zero
reduced cost link with respect to source s if it is part of any
zero reduced cost path originating from s and terminating
at t for some origin-destination pair (s, t) ∈ V 2. We denote
the set of zero reduced cost links with respect to source s as
Es
RC
We require that the UE flow (flow routed by the UE-
controller) is routed solely via zero reduced cost links/paths.
This is because the UE controller can only assign flow
to minimal latency paths (otherwise self-interested agents
would deviate). However, the need to constrain UE flow to
p(f ) ≤ c(cid:48)
links/paths with minimal marginal cost (c(cid:48)) is less intuitive;
this constraint will be justified later on. Note that it is suf-
ficient to only consider whether or not a link e is part of a
reduced cost path from the origin s to some destination t (not
a specific t) because either link e is along a reduced cost path
from (s, t), or there is no path only along links in Es
RC that
includes e. Moreover, we can efficiently compute the set of
zero reduced cost links for any origin destination pair (s, t)
by applying uniform cost search from s to t and marking all
links that are part of optimal paths, once with regard to mini-
mal total latency (arg minp∈Ps,t (lp(f SO)), and second with
regard to minimal marginal cost (arg minp∈Ps,t (c(cid:48)
p(f SO)).
Let the constant f SO denote the flow vector at a SO so-
lution.3 The SO flow is not unique when latency functions
are non-decreasing, and the maximal amount of UE flow
permitted may, in general, depend on the specific SO flow.
Therefore, we must efficiently search over the space of SO
flows. This is possible due to the following lemmas.
Lemma 1. For any two flows that achieve SO, f SO and
f SO, le(f SO
) = le( f SO
).
e
e
e
(cid:54)= f SO
Proof. Given Assumption 2, a SO flow is the solution to a
convex program (Roughgarden and Tardos 2002). The so-
lutions to a convex program form a convex set. Suppose
that there are two flows that both achieve SO, but for which
. Then ce(fe) = le(fe)fe must be a linear func-
f SO
e
(to see this, note that any con-
tion between f SO
vex combination of f SO and f SO is also an SO solution,
but if ce(fe) is not linear, then the total system travel time
would be strictly less, a contradiction). Since le(fe) is a non-
decreasing function, the only way for ce(fe) to be linear is
for le(fe) to be constant between f SO
and f SO
and f SO
.
e
e
e
e
Lemma 2. The set of zero reduced cost paths is identical for
all SO solutions.
Proof. By Lemma 1, all SO flows have the same latency on
each link, so the SO solutions can differ by at most flows
along a set of links with constant latency over the range of
which the two flows differ on those links. Since we assume
that the latency functions are differentiable, the derivatives
of the latency function are zero over the range at which they
are constant. Therefore, c(cid:48)
e(fe) is con-
stant over the range as well. This implies that any path that
is reduced cost in one flow is also reduced cost in the other
flow, since the latency functions and c(cid:48)
e(fe) are constant for
every link e.
e(fe) = le(fe) + fel(cid:48)
e
Define the constant ¯f SO
e = sup{f : le(f ) = le(f SO
)},
i.e. ¯f SO
is the largest flow value such that the latency on link
e is equal to the latency at an SO solution. Note that if le is
strictly increasing at f SO
. However, if le
is constant at f SO
, then ¯f SO
e > f SO
e = f SO
e
Given that the zero reduced cost paths are the same for
all SO flows (Lemma 2), and any SO flow has the same
, then ¯f SO
.
e
e
e
e
3A SO flow can be efficiently computed as a solution to a con-
vex program (Roughgarden and Tardos 2002; Dial 2006).
e
For each vertex, s, and link, e, define variable xs
latency on all links (Lemma 1), it will be sufficient to only
on each link e ∈ E.
search over flows that are less than ¯f SO
e denoting
the amount of UE flow originating from source s that is as-
signed to link e. Let in(v) denote the set of links for which
v is the tail vertex and out(v) the set of links for which v is
the head vertex.
Definition 3. For a given flow model {G, R}, the UE linear
program is:
(cid:88)
s,t∈V
r∗
s,t
∀ s, t ∈ V
∀ s ∈ V
e = r∗
xs
s,v ∀ s, t ∈ V
max
r∗
s,t,xs
e
r∗
s,t
t∈V
e∈out(s)
(cid:88)
e − (cid:88)
subject to
(cid:88)
s,t ≤ R(s, t)
r∗
(cid:88)
xs
e =
(cid:88)
s
e ≥ 0, r∗
xs
xs
e = 0
e ≤ ¯f SO
xs
s,t ≥ 0
e∈in(t)
xs
e
e∈out(t)
e = (cid:80)
(1)
(2)
(3)
(4)
(5)
(6)
∀ e ∈ E, s ∈ V
∀ s, t ∈ V, e ∈ E
∀ s ∈ V, e ∈ E \ Es
RC
(7)
v xv
The flow f UE
e defined by a feasible solution to
the UE linear program (given constraints (2)-(7)) is a UE
subflow. The flow defined by an optimal solution to the UE
linear program is an optimal UE subflow.
Note that the number of variables is {∀s ∈ V, ∀t ∈
e} = O(V 2 +V E), and the number
V, ∀e ∈ E : r∗
of constraints is also O(V 2 +V E). Therefore, since the
number of variables and constraints are polynomial in the
flow model, the optimal solution to the UE linear program
can be computed in polynomial time (Karmarkar 1984).
s,t, xs
Revisiting the definition of a zero reduced cost path/link
(Definition 2), the need to constrain UE flow to minimal
marginal cost paths (in addition to minimal latency paths)
is explained by the fact that, at SO, the UE flow must be a
subflow of f SO. Consider the problem instance depicted in
Figure 1. The latency function, flow at SO, travel time at SO,
and marginal cost at SO are all listed above each link. The
double line links have zero travel time regardless of the vol-
ume, their only purpose is to limit the possible path assign-
ment between vertices 2 and 4. Notice that, for flow traveling
from vertex 2 to 4, the dotted path is of minimal latency but
not minimal marginal cost. Running the above LP on this
instance would result in r∗
rect value. However, if the dotted path is considered to be of
zero reduced cost with respect to source 2 (despite having
a non minimal marginal cost) then running the LP would
result in r∗
Theorem 1. A UE subflow, f UE, defined by a feasible solu-
tion to the UE linear program is a subflow of a SO flow.
UE = 2 ·(cid:112)2/3 which is the cor-
UE = 1 + 2 ·(cid:112)2/3.
constructed an SO flow, f, for which for all links e ∈ E,
e ≤ fe, a contradiction.
f UE
Lemma 3. For a network {G, R}, let f∗ be a subflow of a
feasible flow f. Then the flow f(cid:48) such that f(cid:48)
e is
also a subflow of f.
Proof. First, 0 ≤ f(cid:48)
flow. Now set r(cid:48)
e∈out(s) f(cid:48)
e ≤ fe, by the definition of a sub-
s,t. Then for all s, t ∈
e∈in(s) f(cid:48)
s,t) =
V , (cid:80)
(cid:80)
s,t, and similarly for(cid:80)
t r(cid:48)
s,t = R(s, t) − r∗
e∈in(t) f(cid:48)
e = (cid:80)
e −(cid:80)
t(R(s, t) − r∗
e∈out(t) f(cid:48)
e −(cid:80)
e = fe − f∗
e
Theorem 2. The optimal value of the UE linear program
for a network instance {G, R} is the maximum amount of
UE agents that the network can support and achieve SO.
Proof. First, by Theorem 1, there exists an SO flow such
that the optimal UE subflow, f UE, is a subflow of the SO
flow, and by Lemma 3, there exists a subflow of compliant
agents that can achieve the SO solution. Moreover, by the
definition of the UE linear program and Lemma 2, the UE
flow is only along zero reduced cost paths. By the definition
of zero reduced cost paths, all UE agents are willing to take
the assigned paths. Therefore, the SO solution is achievable
with the UE flow, and there is some volume of UE flow that
is equal to the objective of the UE linear program.
Now, suppose that there was another UE flow assignment,
f(cid:48), for which compliant flow could be assigned in such a
way that the SO total system travel time was achieved and
the total UE flow volume was larger than the value returned
by the UE linear program. Note that this flow assignment
(f(cid:48)) must be a subflow of some SO flow, f. Moreover, by
the definition of UE flow and the fact that all paths in a SO
solution are minimum marginal cost paths, all paths assigned
with a UE flow greater than zero must be a zero reduced
cost path. Therefore, the flow f(cid:48) satisfies the equations (2)-
(6), and since the UE linear program returns the optimal UE
flow assignment, this is a contradiction.
While we've demonstrated that we can compute the maxi-
mal UE flow that permits an SO solution given the appropri-
ate assignment of the compliant flow, it is likely that a more
common problem would be to determine, for a given set of
compliant agents, whether or not it is possible to achieve SO
with that set. Our methodology also provides an answer to
this question, as the following Corollary demonstrates.
Corollary 1. For a given network instance {G, R} and
given a set of compliant demand, rC
s,t, from each origin des-
tination pair s, t ∈ V , there exists a compliant flow f C such
that the network achieves SO if and only if there exists an
e for all s ∈ V and e ∈ E such that rUE
s,t = R(s, t) − rC
xs
s,t
and xs
e are a solution to the UE linear program.
Proof. By Theorem 1, any solution to the UE linear pro-
gram defines a subflow of an SO flow. Therefore, if rU E
s,t
and xs
e is a solution, there exists an assignment of the com-
pliant flow that achieves SO.
Moreover, if there exists an assignment of the com-
plaint flow, f C, such that a UE subflow with demands rU E
s,t
achieves system optimum, then the UE flow is only along
Figure 1: A problem instance where the minimal marginal
cost condition is required. R(1, 3) = R(3, 5) = 1 +
(cid:112)2/3, R(2, 4) = 1. Above each link are listed (in top to
down order): v, the flow at SO (in green), t, the travel time
at SO (in blue), c(cid:48), the marginal cost at SO (in red), and
l(x), the latency function (in a bounding box).
e
Proof. First, note that by equations (2)–(4), the subflow UE
subflow, f UE
, satisfies flow conservation constraints. Equa-
tion (2) states that the flow along all reduced cost paths from
origin s to destination t must be less then total demand for
(s, t). Then equations (3) and (4) state that the flow out of
node v must either be due to the demand generated by node
v or the flow into it, minus the flow that reaches v as a desti-
nation. Therefore, f UE
is a subflow of a feasible flow.
e
e > f SO
e
e ≤ f SO
What must be shown is that there must exist a SO flow,
for all e. If e is such that le
f SO, such that f UE
is strictly increasing at an SO solution, and therefore will
be strictly increasing at all SO solutions by Lemma 1, then
and constraint (5) guarantees the claim. Let E(cid:48)
e = ¯f SO
f SO
be the set of links such that the latency function is constant
at a SO flow. Therefore, it only needs to be shown that there
e ≤ f SO
exists a SO solution, f, such that for e ∈ E(cid:48), f UE
.
Suppose that there existed a set of links e ∈ E(cid:48) such
. Let f SO be an
that for all SO flows f SO, f UE
SO flow. Then there must exist an origin destination pair
(s, t) such that there are two sets of paths P>,P< ⊂ Ps,t
for which for all p ∈ P>, f UE
, and for all
p(cid:48) ∈ P<, f UE
and all paths only differ by links
in E(cid:48). This is because the total flow between any origin-
destination is larger in the SO flow by equation (2). More-
) since
the flow along non-constant latency links constrains the total
) units of flow from paths in
set P> to paths in set P< in the SO flow f SO. Denote the
new flow by f(cid:48). The total travel time for f(cid:48) cannot increase
because the flow has only increased on constant latency
links, and the new flow does not exceed ¯f SO
on any link.
The total travel time also cannot have decreased because
f SO was an SO flow, so f(cid:48) is also an SO flow. Continue
this procedure until there does not exist a link e ∈ E(cid:48) for
which f UE
exceeds the transformed SO flow. Then we have
over,(cid:80)
flow. Move(cid:80)
) ≤(cid:80)
p − f SO
(f UE
p − f UE
( f SO
p − f SO
(f UE
p(cid:48) < f SO
p(cid:48)
> f SO
p
p
p∈P>
p
p
p∈P>
p
p(cid:48)∈P<
e
e
e
e
e
Figure 2: Three representative network topologies: I - Sioux Falls, SD, II - Eastern Massachusetts (Ellipsoids represent different
zones), III - Anaheim, CA.
zero reduced cost paths by definition of UE flow and SO,
and the UE subflow is feasible. Therefore, the decomposed
UE flow satisfies the constraints of the linear program.
Flow Assignment for Compliant Agents
Given that we can now determine both the maximal amount
of UE flow that a system can tolerate and achieve system
optimum and, for a given set of compliant agents, whether
or not a system can achieve optimum, we are only left with
assigning the compliant flow to paths. This section tackles
the question of how to assign paths to a, sufficiently large,
set of compliant agents such that SO is achieved.
The methodology from the previous section immediately
suggests a solution. Given a network instance {G, R}, sup-
s,t for all
pose that we have compliant demand equal to rC
s, t ∈ V . Then we must find a SO flow, f SO, such that rC
s,t
s,t permit subflows of the SO solution.
and rUE
Such a SO flow must exist by Theorem 1 and Corollary 1.
The first step is to compute the UE subflow, f UE, given
UE demand. From the previous section: this exists and is
computationally tractable. Any feasible subflow, f C, with
s,t such that the total flow along link e satisfies
demand rC
e ≤ ¯fe
SO has latency equal to the SO solution,
f C
e + f UE
and the flow f C
, by Lemma 1, is an SO solution.
We can compute f C with the following linear program:
s,t = R(s, t)−rC
e + f UE
e
subject to(cid:88)
(cid:88)
e∈out(v)
f C
e − (cid:88)
e − (cid:88)
f C
e∈in(v)
e∈in(v)
0 ≤ f C
e∈out(v)
e − f UE
e
e ≤ ¯f SO
max
f C
e
f C
e =
1
(cid:88)
(cid:88)
t
s
(rC
v,t)
f C
e =
(rC
s,v)
∀ v ∈ V
∀ v ∈ V
∀ e ∈ E
We know that a solution to the above linear program exists
and it can be computed tractably.
The final step is to decompose the compliant flow, f C,
into a per path assignment for each origin-destination pair
(s, t) in order to assign individual agents to a path. This can
be done in time O(V E) using standard flow decomposi-
tion algorithms (see Section 3.5 of Ahuja, Magnanti, et. al.
(1993) for a discussion).
Experimental Results
We are interested in the viability of opt-in micro-tolling
schemes to more efficiently utilize road networks. As such,
we haven undertaken an empirical study to investigate the
minimal amount of compliant flow required for SO (r∗
UE) in
six realistic traffic scenarios over actual road networks.
Scenarios
Each traffic scenario is defined by the following attributes:
1. The road network, G(V, E), specifying the set of vertices
and links where each link is affiliated with a length, ca-
pacity and speed limit. Networks are, following standard
practice, partitioned into traffic analysis zones (TAZs) and
each zone contains a node belonging to V called the cen-
troid. All traffic originating and terminating within the
zone is assumed to enter and leave the network at cen-
troids.
2. A trip table which specifies the traffic demand between
pairs of centroids. The demand function R between nodes
other than centroids is set to zero.
The following benchmark scenarios were chosen both
for their diversity of topology and traffic volume and
their widespread use within the traffic literature: Sioux
Falls, Eastern Massachusetts, Anaheim, Chicago Sketch,
Philadelphia, and Chicago-regional. All traffic scenarios
are available at: https://github.com/bstabler/
Scenario
Sioux Falls
Eastern MA
Anaheim
Chicago S
Philadelphia
Chicago R
Vertices
24
74
416
933
13389
12982
Links
76
258
914
2950
40003
39018
Zones
24
74
38
387
1525
1790
Total Flow
360,600
65,576
104,694
1,260,907
18,503,872
1,360,427
UE TTT
7,480,225
28,181
1,419,913
18,377,329
335,647,106
33,656,964
SO TTT
7,194,256
27,323
1,395,015
17,953,267
324,268,465
31,942,956
% Improve
3.82
3.04
1.75
2.31
3.39
5.09
Threshold % compliant
13.04
6.19E-11
19.73
3.04E-13
19.76
8.05E-11
27.29
9.14E-10
49.59
4.20E-09
53.34
4.14E-07
Table 1: Required fraction of compliant agents given as "% compliant" for different scenarios along with network specifications
for each scenario: number of vertices, links and zones followed by the Total Travel Time (TTT) at UE (0% compliant agents)
and SO (100% compliant agents). The percentage of improvement of the SO TTT over the UE TTT is given as "% improve".
TransportationNetworks. Figure 2 depicts three
representative network topologies (the three smallest net-
works).
The Traffic Model
A macroscopic model was used in order to evaluate traffic
formation. Macroscopic models calculate the UE in a given
scenario using algorithm B (Dial 2006). For all scenarios,
the model assumed that travel times follow the Bureau of
Public Roads (BPR) function (Moses and Mtoi 2017) with
the commonly used parameters β = 4, α = 0.15. The SO
solution is computed by replacing the latency functions with
c(cid:48)
e(x) and using algorithm B to obtain the equilibrium solu-
tion (Dial 1999). Since solving for the UE and SO solutions
requires solving a convex program (Dial 2006), we only
solve them to a certain precision. To measure convergence,
given an assignment of agents to paths, we define the av-
erage excess cost (AEC) as the average difference between
the travel times on paths taken by the agents and their short-
est alternative path. The algorithm terminates when the AEC
is less than 1E-12 minutes (except for Chicago-regional for
which 1E-10 was used due to the size of the network). There-
fore, a minimum marginal cost path is only a minimum up
to a threshold.
A link e is defined to be zero reduced cost with respect to
s if it carries flow originating at s in the SO solution (i.e.,
the link belongs to a minimum marginal cost path) and if the
difference between the least latency path that include e and
the least latency unrestricted path, both leading from s to the
head vertex of e, is less than a threshold T .
The threshold T is defined as follows. for each origin s
and link e we calculate the least marginal cost path (c(cid:48)) lead-
ing from s to the head vertex of e at the SO solution. We
do this once while restricting the path to include e and once
without such restriction. The difference between these two
values is stored and T is set to be the maximum of these
difference across all the links and origins in the network.
Results
Table 1 presents the percentage of flow that must be com-
pliant in order to guarantee an SO solution for six different
traffic scenarios. Each scenario is affiliated with the number
of vertices, links, and zones comprising the affiliated road
network as well as the number of trips that make up the af-
filiated demand.
The columns "UE TTT" and "SO TTT" represent the to-
tal travel time (in minutes) over all agents for the case where
100% of the agents are controlled by the UE controller (UE
solution) and when 100% of the agents are controlled by the
SO controller (SO solution) respectively. The percentage of
improvement in total travel time between UE TTT and SO
TTT is also shown under "% improve".
The percentage of required compliant flow (formally
r∗
s,t R(s, t)) as computed by the UE
linear program (Definition 3) is presented for each scenario
under "% compliant".4
UE/R where R =(cid:80)
The results suggest that as the size of the network (i.e.,
the number of nodes and vertices) increases, a greater frac-
tion of compliant travelers are needed to ensure the network
achieves system optimum. This appears to be due to an in-
creasing number of used paths at the SO solution as the net-
work size increases. As the number of paths grow, the set of
zero reduced cost paths grows more slowly, and, therefore, a
higher percentage of compliant agents is required.
Summary
This paper discussed a scenario where a set of agents tra-
verse a congested network, and a centralized network man-
ager optimizes the flow (minimizes total latency) using a
set of compliant agents. A methodology was presented for
computing the minimal volume of traffic flow that needs
to be compliant in order to reach a state of optimal traffic
flow. Moreover, the methodology extends to inferring which
agents should be compliant and how exactly the compli-
ant agents should be assigned to paths. Experimental results
demonstrate that the required percentage of agents that are
compliant is relatively small (between 13% and 53%) for
several realistic road networks.
Going forward, it would be worthwhile to explore the pos-
sibility of approximation algorithms for assigning compliant
flow when the UE demand is too large to achieve system
optimum. Given that the optimal solution to this problem is
known to be NP-hard, an efficient approximation algorithm
would be a useful tool as opt-in network routing systems are
implemented. Further, in order to limit the necessary opt-in
incentives, there is work needed to develop systems that tar-
get particularly influential users to opt-in to these systems.
Acknowledgements
The authors would like to thank Josiah Hanna and Michael
Levin for contributing useful comments and discussions in
4Statistical analysis for Table 1 is not presented, as the macro-
scopic model is deterministic.
[Moses and Mtoi 2017] Moses, R., and Mtoi, E. T. 2017.
Calibration and evaluation of link congestion functions.
Journal of Transportation Technologies 4(2).
[Numrich, Ruja, and Voss 2012] Numrich, J.; Ruja, S.; and
Voss, S. 2012. Global navigation satellite system based
tolling: state-of-the-art. NETNOMICS: Economic Research
and Electronic Networking 13(2):93.
[Pigou 1920] Pigou, A. C. 1920. The Economics of Welfare.
Palgrave Macmillan.
[Rosenthal 1973] Rosenthal, R. W. 1973. A class of games
International
possessing pure-strategy Nash equilibria.
Journal of Game Theory 2(1):65–67.
[Roughgarden and Tardos 2002] Roughgarden, T., and Tar-
dos, ´E. 2002. How bad is selfish routing? Journal of the
ACM (JACM) 49(2):236–259.
[Roughgarden 2004] Roughgarden, T.
Stackel-
berg scheduling strategies. SIAM Journal on Computing
33(2):332–350.
[Schaller 2010] Schaller, B. 2010. New York Citys conges-
tion pricing experience and implications for road pricing ac-
ceptance in the United States. Transport Policy 17(4):266–
273.
[Sharon et al. 2017] Sharon, G.; Hanna, J. P.; Levin, M. W.;
Rambha, T.; Albert, M.; Boyles, S. D.; and Stone, P. 2017.
Real-time adaptive tolling scheme for optimized social wel-
In Proceedings of the 16th Inter-
fare in traffic networks.
national Conference on Autonomous Agents and Multiagent
Systems (AAMAS-17).
[Yang and Zhang 2008] Yang, H., and Zhang, X. 2008. Ex-
istence of anonymous link tolls for system optimum on net-
works with mixed equilibrium behaviors. Transportation
Research Part B: Methodological 42(2):99–112.
[Yang, Zhang, and Meng 2007] Yang, H.; Zhang, X.; and
Meng, Q. 2007. Stackelberg games and multiple equilib-
rium behaviors on networks. Transportation Research Part
B: Methodological 41(8):841–861.
2004.
the course of this research.
A portion of this work has taken place in the Learn-
ing Agents Research Group (LARG) at the Artificial In-
telligence Laboratory, The University of Texas at Austin.
LARG research is supported in part by grants from the
National Science Foundation (CNS-1305287, IIS-1637736,
IIS-1651089,
IIS-1724157), The Texas Department of
Transporation, Intel, Raytheon, and Lockheed Martin. Peter
Stone serves on the Board of Directors of Cogitai, Inc. The
terms of this arrangement have been reviewed and approved
by the University of Texas at Austin in accordance with its
policy on objectivity in research.
The authors would also like to acknowledge the support of
the Data-Supported Transportation Operations & Planning
Center and the National Science Foundation under Grant
No. 1254921.
References
[Ahuja, Magnanti, and Orlin 1993] Ahuja, R. K.; Magnanti,
T. L.; and Orlin, J. B. 1993. Network Flows: Theory, Algo-
rithms, and Applications. Prentice Hall, 1 edition.
[Beckmann, McGuire, and Winsten 1956] Beckmann, M.;
McGuire, C. B.; and Winsten, C. B. 1956. Studies in the
Economics of Transportation. Yale University Press.
[Dial 1999] Dial, R. B. 1999. Network-optimized road pric-
ing: Part i: a parable and a model. Informs 47.
[Dial 2006] Dial, R. B. 2006. A path-based user-equilibrium
traffic assignment algorithm that obviates path storage and
enumeration. Transportation Research Part B: Methodolog-
ical 40(10):917–936.
Uber ein Paradoxon aus
[Dietrich 1969] Dietrich, B. 1969.
der Verkehrsplanung. Unternehmensforschung 12:258–268.
[Duffield et al. 1999] Duffield, N. G.; Goyal, P.; Greenberg,
A.; Mishra, P.; Ramakrishnan, K. K.; and van der Merive,
J. E. 1999. A flexible model for resource management
in virtual private networks. In ACM SIGCOMM Computer
Communication Review, volume 29, 95–108. ACM.
[Fingerhut, Suri, and Turner 1997] Fingerhut, J. A.; Suri, S.;
and Turner, J. S. 1997. Designing least-cost nonblocking
broadband networks. Journal of Algorithms 24(2):287–309.
[Haurie and Marcotte 1985] Haurie, A., and Marcotte, P.
1985. On the relationship between Nash-Cournot and
Wardrop equilibria. Networks 15(3):295–308.
[Immorlica et al. 2009] Immorlica, N.; Li, L. E.; Mirrokni,
V. S.; and Schulz, A. S.
2009. Coordination mecha-
nisms for selfish scheduling. Theoretical Computer Science
410(17):1589–1598.
[Karmarkar 1984] Karmarkar, N. 1984. A new polynomial-
time algorithm for linear programming. In Proceedings of
the Sixteenth Annual ACM Symposium on Theory of Com-
puting, STOC '84, 302–311. New York, NY, USA: ACM.
[Korilis, Lazar, and Orda 1997] Korilis, Y. A.; Lazar, A. A.;
and Orda, A. 1997. Achieving network optima using stack-
elberg routing strategies. IEEE/ACM Transactions on Net-
working (TON) 5(1):161–173.
|
1211.2476 | 1 | 1211 | 2012-11-11T23:09:02 | Random Utility Theory for Social Choice | [
"cs.MA",
"cs.LG",
"stat.ML"
] | Random utility theory models an agent's preferences on alternatives by drawing a real-valued score on each alternative (typically independently) from a parameterized distribution, and then ranking the alternatives according to scores. A special case that has received significant attention is the Plackett-Luce model, for which fast inference methods for maximum likelihood estimators are available. This paper develops conditions on general random utility models that enable fast inference within a Bayesian framework through MC-EM, providing concave loglikelihood functions and bounded sets of global maxima solutions. Results on both real-world and simulated data provide support for the scalability of the approach and capability for model selection among general random utility models including Plackett-Luce. | cs.MA | cs |
Random Utility Theory for Social Choice
Hossein Azari Soufiani
SEAS, Harvard University
[email protected]
David C. Parkes
SEAS, Harvard University
[email protected]
Abstract
Lirong Xia
SEAS, Harvard University
[email protected]
Random utility theory models an agent's preferences on alternatives by drawing
a real-valued score on each alternative (typically independently) from a param-
eterized distribution, and then ranking the alternatives according to scores. A
special case that has received significant attention is the Plackett-Luce model, for
which fast inference methods for maximum likelihood estimators are available.
This paper develops conditions on general random utility models that enable fast
inference within a Bayesian framework through MC-EM, providing concave log-
likelihood functions and bounded sets of global maxima solutions. Results on
both real-world and simulated data provide support for the scalability of the ap-
proach and capability for model selection among general random utility models
including Plackett-Luce.
1
Introduction
Problems of learning with rank-based error metrics [16] and the adoption of learning for the purpose
of rank aggregation in social choice [7, 8, 23, 25, 29, 30] are gaining in prominence in recent years.
In part, this is due to the explosion of socio-economic platforms, where opinions of users need to be
aggregated; e.g., judges in crowd-sourcing contests, ranking of movies or user-generated content.
In the problem of social choice, users submit ordinal preferences consisting of partial or total ranks
on the alternatives and a single rank order must be selected to be representative of the reports.
Since Condorcet [6], one approach to this problem is to formulate social choice as the problem
of estimating a true underlying world state (e.g., a true quality ranking of alternatives), where the
individual reports are viewed as noisy data in regard to the true state. In this way, social choice can
be framed as a problem of inference.
In particular, Condorcet assumed the existence of a true ranking over alternatives, with a voter's pref-
erence between any pair of alternatives a, b generated to agree with the true ranking with probability
p > 1/2 and disagree otherwise. Condorcet proposed to choose as the outcome of social choice the
ranking that maximizes the likelihood of observing the voters' preferences. Later, Kemeny's rule
was shown to provide the maximum likelihood estimator (MLE) for this model [32].
But Condorcet's probabilistic model assumes identical and independent distributions on pairwise
comparisons. This ignores the strength in agents' preferences (the same probability p is adopted
for all pairwise comparisons), and allows for cyclic preferences. In addition, computing the winner
through the Kemeny rule is ΘP
To overcome the first criticism, a more recent literature adopts the random utility model (RUM)
from economics [26]. Consider C = {c1, .., cm} alternatives.
In RUM, there is a ground truth
utility (or score) associated with each alternative. These are real-valued parameters, denoted by
(cid:126)θ = (θ1, . . . , θm). Given this, an agent independently samples a random utility (Xj) for each
alternative cj with conditional distribution µj(·θj).
2 -complete [13].
1
Usually θj is the mean of µj(·θj).1 Let π denote a permutation of {1, . . . , m}, which naturally
corresponds to a linear order: [cπ(1) (cid:31) cπ(2) (cid:31) ··· (cid:31) cπ(m)]. Slightly abusing notation, we also use
π to denote this linear order. Random utility (X1, . . . , Xm) generates a distribution on preference
orders, as
Pr(π (cid:126)θ) = Pr(Xπ(1) > Xπ(2) > . . . > Xπ(m))
(1)
The generative process is illustrated in Figure 1.
Figure 1: The generative process for RUMs.
Adopting RUMs rules out cyclic preferences, because each agent's outcome corresponds to an order
on real numbers, and it also captures the strength of preference, and thus overcomes the second
criticism, by assigning a different parameter (θj) to each alternative.
A popular RUM is Plackett-Luce (P-L) [18, 21], where the random utility terms are generated ac-
cording to Gumbel distributions with fixed shape parameter [2,31]. For P-L, the likelihood function
has a simple analytical solution, making MLE inference tractable. P-L has been extensively applied
in econometrics [1, 19], and more recently in machine learning and information retrieval (see [16]
for an overview). Efficient methods of EM inference [5, 14], and more recently expectation propa-
gation [12], have been developed for P-L and its variants.
In application to social choice, the P-L model has been used to analyze political elections [10]. EM
algorithm has also been used to learn the Mallows model, which is closely related to the Condorcet's
probabilistic model [17].
Although P-L overcomes the two difficulties of the Condorcet-Kemeny approach, it is still quite
restricted, by assuming that the random utility terms are distributed as Gumbel, with each alternative
is characterized by one parameter, which is the mean of its corresponding distribution. In fact, little
is known about inference in RUMs beyond P-L. Specifically, we are not aware of either an analytical
solution or an efficient algorithm for MLE inference for one of the most natural models proposed by
Thurstone [26], where each Xj is normally distributed.
1.1 Our Contributions
In this paper we focus on RUMs in which the random utilities are independently generated with
respect to distributions in the exponential family (EF) [20]. This extends the P-L model, since
the Gumbel distribution with fixed shape parameters belonging to the EF. Our main theoretical
contributions are Theorem 1 and Theorem 2, which propose conditions such that the log-likelihood
function is concave and the set of global maxima solutions is bounded for the location family, which
are RUMs where the shape of each distribution µj is fixed and the only latent variables are the
locations, i.e., the means of µj's. These results hold for existing special cases, such as the P-L
model, and many other RUMs, for example the ones where each µj is chosen from Normal, Gumbel,
Laplace and Cauchy.
1µj(·θj) might be parameterized by other parameters, for example variance.
2
We also propose a novel application of MC-EM. We treat the random utilities ( (cid:126)X) as latent variables,
and adopt the Expectation Maximization (EM) method to estimate parameters (cid:126)θ. The E-step for
this problem is not analytically tractable, and for this we adopt a Monte Carlo approximation. We
establish through experiments that the Monte-Carlo error in the E-step is controllable and does not
affect inference, as long as numerical parameterizations are chosen carefully. In addition, for the E-
step we suggest a parallelization over the agents and alternatives and a Rao-Blackwellized method,
which further increases the scalability of our method.
We generally assume that the data provides total orders on alternatives from voters, but comment on
how to extend the method and theory to the case where the input preferences are partial orders.
We evaluate our approach on synthetic data as well as two real-world datasets, a public election
dataset and one involving rank preferences on sushi. The experimental results suggest that the
approach is scalable despite providing significantly improved modeling flexibility over existing ap-
proaches.
For the two real-world datasets we have studied, we compare RUMs with normal distributions and
P-L in terms of four criteria: log-likelihood, predictive log-likelihood, Akaike information criterion
(AIC), and Bayesian information criterion (BIC). We observe that when the amount of data is not
too small, RUMs with normal distributions fit better than P-L. Specifically, for the log-likelihood,
predictive log-likelihood, and AIC criteria, RUMs with normal distributions outperform P-L with
95% confidence in both datasets.
2 RUMs and Exponential Families
In social choice, each agent i ∈ {1, . . . , n} has a strict preference order on alternatives. This
provides the data for an inferential approach to social choice. In particular. let L(C) denote the set
of all linear orders on C. Then, a preference-profile, D, is a set of n preference orders, one from
each agent, so that D ∈ L(C)n.
A voting rule r is a mapping that assigns to each preference-profile a set of winning rankings,
r : L(C)n (cid:55)→ (2L(C) \ ∅). In particular, in the case of ties the set of winning rankings may include
more than a singleton ranking. In the maximum likelihood (MLE) approach to social choice, the
preference profile is viewed as data, D = {π1, . . . , πn}.
Given this, the probability (likelihood) of the data given ground truth (cid:126)θ (and for a particular (cid:126)µ) is
µπ(n)(xπ(n))..µπ(1)(xπ(1))dxπ(1)dxπ(2)..dxπ(n) (2)
xπ(1)=xπ(2)
The MLE approach to social choice selects as the winning ranking that which corresponds to the (cid:126)θ
that maximizes Pr(D (cid:126)θ). In the case of multiple parameters that maximize the likelihood then the
MLE approach returns a set of rankings, one ranking corresponding to each parameterization.
In this paper, we focus on probabilistic models where each µj belongs to the exponential family
(EF). The density function for each µ in EF has the following format:
Pr(X = x) = µ(x) = eη(θ)T (x)−A(θ)+B(x),
(3)
where η(·) and A(·) are functions of θ, B(·) is a function of x, and T (x) denotes the sufficient
statistics for x, which could be multidimensional.
Example 1 (Plackett-Luce as an RUM [2]) In the RUM, let µj's be Gumbel distributions. That
is, for alternative j ∈ {1, . . . , m} we have µj(xjθj) = e−(xj−θj )e−e
−(xj−θj ). Then, we have:
, where η(θj) = λj = eθj ,
Pr(π (cid:126)λ) = Pr(xπ(1) > xπ(2) > .. > xπ(m)) = (cid:81)m
(cid:80)m
T (xj) = −e−xj , B(xj) = −xj and A(θj) = −θj.This gives us the Plackett-Luce model.
j=1
λπ(j)
j(cid:48)=j λπ(j(cid:48) )
3 Global Optimality and Log-Concavity
In this section, we provide a condition on distributions that guarantees that the likelihood function (2)
is log-concave in parameters (cid:126)θ. We also provide a condition under which the set of MLE solutions
3
Pr(D (cid:126)θ) =(cid:81)n
(cid:90) ∞
P (π(cid:126)θ) =
(cid:90) ∞
i=1 Pr(πi (cid:126)θ), where,
(cid:90) ∞
xπ(n)=−∞
..
xπ(n−1)=xπ(n)
is bounded when any one latent parameter is fixed. Together, this guarantees the convergence of our
MC-EM approach to a global mode with an accurate enough E-step. We focus on the location family,
which is a subset of RUMs where the shapes of all µj's are fixed, and the only parameters are the
means of the distributions. For the location family, we can write Xj = θj +ζj, where Xj ∼ µj(·θj)
and ζj = Xj − θj is a random variable whose mean is 0 and models an agent's subjective noise.
The random variables ζj's do not need to be identically distributed for all alternatives j; e.g., they
can be normal with different fixed variances.
We focus on computing solutions ((cid:126)θ) to maximize the log-likelihood function,
n(cid:88)
l((cid:126)θ; D) =
log Pr(πi (cid:126)θ)
(4)
i=1
Theorem 1 For the location family, if for every j ≤ m the probability density function for ζj is
log-concave, then l((cid:126)θ; D) is concave.
Proof sketch: The theorem is proved by applying the following lemma, which is Theorem 9 in [22].
Lemma 1 Suppose g1((cid:126)θ, (cid:126)ζ), ..., gR((cid:126)θ, (cid:126)ζ) are concave functions in R2m where (cid:126)θ is the vector of m
parameters and (cid:126)ζ is a vector of m real numbers that are generated according to a distribution whose
pdf is logarithmic concave in Rm. Then the following function is log-concave in Rm.
Li((cid:126)θ, G) = Pr(g1((cid:126)θ, (cid:126)ζ) ≥ 0, ..., gR((cid:126)θ, (cid:126)ζ) ≥ 0), (cid:126)θ ∈ Rm
(5)
r((cid:126)θ, (cid:126)ζ) = θπi(r) + ζ i
πi(r) − θπi(r+1) − ζ i
πi(r+1)
Li((cid:126)θ, πi) = Li((cid:126)θ, Gi) = Pr(gi
R((cid:126)θ, (cid:126)ζ) ≥ 0), (cid:126)θ ∈ Rm
To apply Lemma 1, we define a set Gi of function gi's that is equivalent to an order πi in the sense of
inequalities implied by RUM for πi and Gi (the joint probability in (5) for Gi to be the same as the
probity of πi in RUM with parameters (cid:126)θ). Suppose gi
for r = 1, .., m − 1.
Then considering that the length of order πi is R + 1, we have:
1((cid:126)θ, (cid:126)ζ) ≥ 0, ..., gi
(6)
r((cid:126)θ, (cid:126)ζ) ≥ 0 is equivalent to that in πi alternative πi(r) is preferred to alternative
πi(r+1) for r = 1, .., mi − 1 and gi
πi(r) − θπi(r+1) − ζ i
πi(r+1) for r = mi, .., m − 1. Considering that gi
This is because gi
πi(r + 1) in the RUM sense.
To see how this extends to the case where preferences are specified as partial orders, we consider
in particular an interpretation where an agent's report for the ranking of mi alternatives implies that
r((cid:126)θ, (cid:126)ζ) =
all other alternatives are worse for the agent, in some undefined order. Given this, define gi
πi(mi) −
r((cid:126)θ, (cid:126)ζ) = θπi(mi) + ζ i
θπi(r) + ζ i
r(·)s are linear (hence, concave) and
θπi(r+1) − ζ i
using log concavity of the distributions of (cid:126)ζ i = (ζ i
m)'s, we can apply Lemma 1 and prove
(cid:3)
log-concavity of the likelihood function.
It is not hard to verify that pdfs for normal and Gumbel are log-concave under reasonable conditions
for their parameters, made explicit in the following corollary.
Corollary 1 For the location family where each ζj is a normal distribution with mean zero and
with fixed variance, or Gumbel distribution with mean zeros and fixed shape parameter, l((cid:126)θ; D) is
concave. Specifically, the log-likelihood function for P-L is concave.
1, ζ i
2, .., ζ i
The concavity of log-likelihood of P-L has been proved [9] using a different technique. Using Fact
3.5. in [24], the set of global maxima solutions to the likelihood function, denoted by SD, is convex
since the likelihood function is log-concave. However, we also need that SD is bounded, and would
further like that it provides one unique order as the estimation for the ground truth.
For P-L, Ford, Jr. [9] proposed the following necessary and sufficient condition for the set of global
maxima solutions to be bounded (more precisely, unique) when(cid:80)m
Condition 1 Given the data D, in every partition of the alternatives C into two nonempty subsets
C1 ∪C2, there exists c1 ∈ C1 and c2 ∈ C2 such that there is at least one ranking in D where c1 (cid:31) c2.
j=1 eθj = 1.
4
We next show that Condition 1 is also a necessary and sufficient condition for the set of global
maxima solutions SD to be bounded in location families, when we set one of the values θj to be 0
(w.l.o.g., let θ1 = 0). If we do not bound any parameter, then SD is unbounded, because for any (cid:126)θ,
any D, and any number s ∈ R, l((cid:126)θ; D) = l((cid:126)θ + s; D).
Theorem 2 Suppose we fix θ1 = 0. Then, the set SD of global maxima solutions to l(θ; D) is
bounded if and only if the data D satisfies Condition 1.
Proof sketch:
If Condition 1 does not hold, then SD is unbounded because the parameters for all alternatives in
C1 can be increased simultaneously to improve the log-likelihood. For sufficiency, we first present
the following lemma.
Lemma 2 If alternative j is preferred to alternative j(cid:48) in at least in one ranking then the difference
of their mean parameters θj(cid:48) − θj is bounded from above (∃Q where θj(cid:48) − θj < Q) for all the (cid:126)θ
that maximize the likelihood function.
Proof: Suppose that j (cid:31) j(cid:48) in rank i, then for any (cid:126)θ ∈ Rm:
Li((cid:126)θ, πi) = Li((cid:126)θ, Gi) = Pr(g1((cid:126)θ, (cid:126)ζ) ≥ 0, ..., gR((cid:126)θ, (cid:126)ζ) ≥ 0)
≤ Pr(gπi(r)((cid:126)θ, (cid:126)ζ) ≥ 0, gπi(r+1)((cid:126)θ, (cid:126)ζ) ≥ 0, . . . , gπi(r(cid:48))((cid:126)θ, (cid:126)ζ) ≥ 0) ≤ Pr(ζj − ζj(cid:48) ≥ θj(cid:48) − θj), (7)
where j = πi(r) and j(cid:48) = πi(r(cid:48)).
Let K = l((cid:126)0; D). Since the log-likelihood is always smaller than 0, it follows that for any (cid:126)θ ∈ SD
and any i ≤ n, Li((cid:126)θ; πi) ≥ K.
Hence, Pr(ζj − ζj(cid:48) ≥ θj(cid:48) − θj) ≥ K.
Therefore, there exists K(cid:48) such that θj(cid:48) − θj < K(cid:48), where K(cid:48) depends on the fixed ζj(cid:48) and ζj. (cid:3)
Now consider a directed graph GD, where the nodes are the alternatives, and there is an edge be-
tween cj to cj(cid:48) if in at least one ranking cj (cid:31) cj(cid:48). By Condition 1, for any pair j (cid:54)= j(cid:48), there is a path
from cj to cj(cid:48) (and conversely, a path from cj(cid:48) to cj). To see this, consider building a path between
j and j(cid:48) by starting from a partition with C1 = {j} and following an edge from j to j1 in the graph
where j1 is an alternatives in C2 for which there must be such an edge, by Condition 1. Consider the
partition with C1 = {j, j1}, and repeat until an edge can be followed to vertex j(cid:48) ∈ C2. It follows
from Lemma 2 that for any (cid:126)θ ∈ SD we have θj − θj(cid:48) < Qm, using the telescopic sum of bounded
values of the difference of mean parameters along the edges of the path, since the length of the path
is no more than m (and tracing the path from j to j(cid:48) and j(cid:48) to j), meaning that SD is bounded. (cid:3)
Now that we have the log concavity and bounded property, we need to declare conditions under
which the bounded convex space of estimated parameters corresponds to a unique order. The next
theorem provides a necessary and sufficient condition for all global maxima to correspond to the
same order on alternatives. Suppose that we order the alternatives based on estimated θ's (meaning
that cj is ranked higher than cj(cid:48) iff θj > θj(cid:48)).
Theorem 3 The order over parameters is strict and is the same across all (cid:126)θ ∈ SD if, for all (cid:126)θ ∈ SD
and all alternatives j (cid:54)= j(cid:48), θj (cid:54)= θj(cid:48).
Proof: Suppose for the sake of contradiction there exist two maxima, (cid:126)θ, (cid:126)θ∗ ∈ SD and a pair of
alternatives j (cid:54)= j(cid:48) such that θj > θj(cid:48) and θ∗
j . Then, there exists an α < 1 such that the jth
and j(cid:48)th components of α(cid:126)θ + (1 − α)(cid:126)θ∗ are equal, which contradicts the assumption.
(cid:3)
Hence, if there is never a tie in the scores in any (cid:126)θ ∈ SD, then any vector in SD will reveal the
unique order.
j(cid:48) > θ∗
4 Monte Carlo EM for Parameter Estimation
In this section, we propose an MC-EM algorithm for MLE inference for RUMs where every µj
belongs to the EF.2
2Our algorithm can be naturally extended to compute a maximum a posteriori probability (MAP) estimate,
when we have a prior over the parameters (cid:126)θ. Still, it seems hard to motivate the imposition of a prior on
parameters in many social choice domains.
5
E-Step : Q((cid:126)θ, (cid:126)θt) = E (cid:126)X
M-step : (cid:126)θt+1 ∈ arg max
(cid:126)θ
4.1 Monte Carlo E-step by Gibbs sampler
The E-step can be simplified using (3) as follows:
n(cid:89)
i=1
E (cid:126)X{log
m(cid:88)
EX i
j
n(cid:88)
=
Pr((cid:126)xi, πi (cid:126)θ) D, (cid:126)θt} = E (cid:126)X{log
m(cid:88)
n(cid:88)
jθj) πi, (cid:126)θt} =
{log µj(xi
n(cid:89)
i=1
Pr((cid:126)xi (cid:126)θ) Pr(πi(cid:126)xi) D, (cid:126)θt}
The EM algorithm determines the MLE parameters (cid:126)θ iteratively, and proceeds as follows. In each
iteration t + 1, given parameters (cid:126)θt from the previous iteration, the algorithm is composed of an
E-step and an M-step. For the E-step, for any given (cid:126)θ = (θ1, . . . , θm), we compute the conditional
expectation of the complete-data log-likelihood (latent variables (cid:126)x and data D), where the latent
variables (cid:126)x are distributed according to data D and parameters (cid:126)θt from the last iteration.
For the M-step, we optimize (cid:126)θ to maximize the expected log-likelihood computed in the E-step, and
use it as the input (cid:126)θt+1 for the next iteration:
(cid:41)
Pr((cid:126)xi, πi (cid:126)θ) D, (cid:126)θt
(cid:40)
n(cid:89)
log
i=1
Q((cid:126)θ, (cid:126)θt)
(η(θj)EX i
j
{T (xi
j) πi, (cid:126)θt} − A(θj) + W,
i=1
j=1
i=1
j=1
j
j
{T (xi
{B(xi
j) πi, (cid:126)θt} where T (xi
j) πi, (cid:126)θt} only depends on (cid:126)θt and D (not on (cid:126)θ), which means that it can
where W = EX i
be treated as a constant in the M-step. Hence, in the E-step we only need to compute Si,t+1
=
j) is the sufficient statistic for the parameter θj in the model. We
EX i
j) πi, (cid:126)θt}. However, we can use a Monte
are not aware of an analytical solution for EX i
Carlo approximation, which involves sampling (cid:126)xi from the distribution Pr((cid:126)xi πi, (cid:126)θt) using a Gibbs
sampler, and then approximates Si,t+1
j ) where N is the number of samples in
the Gibbs sampler.
In each step of our Gibbs sampler for voter i, we randomly choose a position j in πi and
πi(j) according to a TruncatedEF distribution Pr(· xπi(−j), (cid:126)θt, πi), where xπi(−j) =
sample xi
( xπi(1), . . . , xπi(j−1), xπi(j+1), . . . , xπi(m)). The TruncatedEF is obtained by truncating the tails
of µπi(j)(·θt
πi(j)) at xπi(j−1) and xπi(j+1), respectively. For example, a truncated normal distribu-
tion is illustrated in Figure 2.
k=1 T (xi,k
(cid:80)N
{T (xi
by 1
N
j
j
j
for xi,k
j
To further
Rao-Blackwellized:
improve the
Gibbs sampler, we use Rao-Blackwellized [4]
xi,k−j, πi, (cid:126)θt}
using E{T (xi,k
estimation
j )
, where xi,k−j is all
instead of the sample xi,k
j
Finally, we esti-
of (cid:126)xi,k except
.
j ) xi,k−j, πi, (cid:126)θt} in each step
mate E{T (xi,k
(cid:80)N
the Gibbs sampler using M samples as
of
(cid:80)M
j ) xk−j, πi, (cid:126)θt} (cid:39)
k=1 E{T (xi,k
Si,t+1
j
∼
l=1 T (xil,k
1
xi,k−j, πi, (cid:126)θ).
Rao-Blackwellization
the variance of
the estimator be-
conditioning and expectation in
j ) xi,k−j, πi, (cid:126)θt}.
NM
Pr(xil,k
reduces
cause of
E{T (xi,k
), where xil,k
(cid:80)N
(cid:39) 1
k=1
N
j
j
j
6
Figure 2: A truncated normal distribution.
4.2 M-step
imize(cid:80)n
η(θj)(cid:80)n
compute θt+1
j
i=1
In the E-step we have (approximately) computed Si,t+1
j) πi, (cid:126)θt} − A(θj) + EX i
j
{T (xi
j=1(η(θj)EX i
(cid:80)m
for each j ≤ m separately to maximize(cid:80)n
(cid:80)n
− nA(θj).
j
i=1 Si,t+1
j
For the case of the normal distribution with fixed variance, where η(θj) = 2θj and A(θj) = (θj)2,
we have θt+1
. The algorithm is illustrated in Figure 3.
j = 1
n
i=1 Si,t+1
j
. In the M-step we compute (cid:126)θt+1 to max-
j) πi, (cid:126)θt}). Equivalently, we
j) πi, (cid:126)θt}−A(θj)} =
{T (xi
{B(xi
i=1{η(θj)EX i
j
j
Figure 3: The MC-EM algorithm for normal distribution.
4.3 Convergence
In the last section we showed that if the RUM satisfies the premise in Theorem 1 and Theorem 2 the
data satisfies Condition 1, then the log-likelihood function is concave, and the set of global maxima
solutions is bounded. This guarantee the convergence of MC-EM for an exact E-step.
In general, MC-EM methods do not have the uniform convergence property of EM methods. In
order to control the error of approximation in the MC-E step we can increase the number of samples
with the iterations [28]. However, in our application, we are not concerned with the exact estimation
of (cid:126)θ, as we are only interested in their orders relative to each-other. Therefore, as long as the
approximation error remains relatively small, such that the differences of θjs are much larger than
the error, we are safe to stop.
A known problem with Gibbs sampling is that it can introduce correlation among samples. To
address this, we sub-sample the samples to reduce the correlation, and call the ratio of sub-sampling
the thinning factor (0 < F ≤ 1). A suitable thinning ratio can be set using empirical results from
the sampler.
With an approach similar to [3], we can derive a relationship between the variance of error in (cid:126)θt+1
and the Monte-Carlo error in the E-step approximation:
Var (θj
t+1) =
1
n2
Var (Si,t+1
j
) =
1
M N n2
i=1
Var (xi
j) ≤ F V
M N n
,
(8)
where N is number of samples in Gibbs sampler, M is the number of samples for Rao-
Blackwellization, n is number of agents, F is the thinning factor and V = maxj(Var x∼µj (x)),
t+1)
and samples xi
arbitrarily small by increasing M N.
j are assumed to be independent. Given, T , V and n, we can make Var (θj
5 Experimental Results
We evaluate the proposed MC-EM algorithm on synthetic data as well as two real world data sets,
namely an election data set and a dataset representing preference orders on sushi. For simulated data
7
n(cid:88)
i=1
n(cid:88)
we use the Kendall correlation [11] between two rank orders (typically between the true order and
the method's result) as a measure of performance.
5.1 Experiments for Synthetic Data
We first generate data from Normal models for the random utility terms, with means θj = j and
equal variance for all terms, for different choices of variance (Var = 2, 4). We evaluate the perfor-
mance of the method as the number of agents n varies. The results show that a limited number of
iterations in the EM algorithm (at most 3), and samples M N = 4000 (M=5, N=800) are sufficient
for inferring the order in most cases. The performance in terms of Kendall correlation for recovering
ground truth improves for larger number of agents, which corresponds to more data. See Figure 4,
which shows the asymptotic behavior of the maximum likelihood estimator in recovering the true
parameters. Figure 4 left and middle panels show that the more the size of dataset the better the
performance of the method.
Moreover, for large variances in data generation, due to increasing noise in the data, the rate that
performance gets better is slower than that for the case for smaller variances. Notice that the scales
on the y-axis are different in the left and middle panels.
Figure 4: Left and middle panel: Performance for different number of agents n on synthetic data for m = 5, 10
and Var = 2, 4, with specifications M N = 4000, EM iterations = 3. Right panel: Performance given
access to sub-samples of the data in the public election dataset, x-axis: size of sub-samples, y-axis: Kendall
Correlation with the order obtained from the full data-set. Dashed lines are the 95% confidence intervals.
5.2 Experiments for Model Robustness
We apply our method to a public election dataset collected by Nicolaus Tideman [27], where the
voters provided partial orders on candidates. A partial order includes comparisons among a subset
of alternative, and the non-mentioned alternatives in the partial order are considered to be ranked
lower than the lowest ranked alternative among mentioned alternatives.
The total number of votes are n = 280 and the number of alternatives m = 15. For the purpose of
our experiments, we adopt the order on alternatives obtained by applying our method on the entire
dataset as an assumed ground truth, since no ground truth is given as part of the data. After finding
the ground truth by using all 280 votes (and adopting a normal model), we compare the performance
of our approach as we vary the amount of data available. We evaluate the performance for sub-
samples consisting of 10, 20, . . . , 280 of samples randomly chosen from the full dataset. For each
sub-sample size, the experiment is repeated 200 times and we report the average performance and
the variance. See the right panel in Figure 4. This experiment shows the robustness of the method,
in the sense that the result of inference on a subset of the dataset shows consistent behavior with the
case that the result on the full dataset. For example, the ranking obtained by using half of the data
can still achieve a fair estimate to the results with full data, with an average Kendall correlation of
greater than 0.4.
5.3 Experiments for Model Fitness
In addition to a public election dataset, we have tested our algorithm on a sushi dataset, where 5000
users give rankings over 10 different kinds of sushi [15]. For each experiment we randomly choose
8
n ∈ {10, 20, 30, 40, 50} rankings, apply our MC-EM for RUMs with normal distributions where
variances are also parameters.
In the former experiments, both the synthetic data generation and the model for election data, the
variances were fixed to 1 and hence we had the theoretical guarantees for the convergence to global
optimal solutions by Theorem 1 and Theorem 2. When we let the variances to be part of parametriza-
tion we lose the theoretical guarantees. However, the EM algorithm can still be applied, and since
the variances are now parameters (rather than being fixed to 1), the model fits better in terms of
log-likelihood.
For this reason, we adopt RUMs with normal distributions in which the variance is a parameter that
is fit by EM along with the mean. We call this model a normal model. We compute the difference
between the normal model and P-L in terms of four criteria: log-likelihood (LL), predictive log-
likelihood (predictive LL), AIC, and BIC. For (predictive) log-likelihood, a positive value means
that normal model fits better than P-L, whereas for AIC and BIC, a negative number means that
normal model fits better than P-L. Predictive likelihood is different from likelihood in the sense
that we compute the likelihood of the estimated parameters for a part of the data that is not used for
parameter estimation.3 In particular, we compute predictive likelihood for a randomly chosen subset
of 100 votes. The results and standard deviations for n = 10, 50 are summarized in Table 1.
n = 10
n = 50
LL
Pred. LL
AIC
BIC
LL
Dataset
Sushi
22.6(6.3) 40.1(5.1) -35.2(12.6) -6.1(12.6)
Election 9.4(10.6) 91.3(103.8) -8.8(21.2) 4.2(21.2) 44.8(15.8) 87.4(30.5) -79.6(31.6) -50.5(31.6)
8.8(4.2) -56.1(89.5) -7.6(8.4) 5.4(8.4)
Pred. LL
AIC
BIC
Table 1: Model selection for the sushi dataset and election dataset. Cases where the normal model fits better
than P-L statistically with 95% confidence are in bold.
When n is small (n = 10), the variance is high and we are unable to obtain statistically significant
results in comparing fitness. When n is not too small (n = 50), RUMs with normal distributions
fit better than P-L. Specifically, for log-likelihood, predictive log-likelihood, and AIC, RUMs with
normal distributions outperform P-L with 95% confidence in both datasets.
5.4
Implementation and Run Time
The running time for our MC-EM algorithm scales linearly with number of agents on real world
data (Election Data) with slope 13.3 second per agent on an Intel i5 2.70GHz PC. This is for 100
iterations of EM algorithm with Gibbs sampling number increasing with iterations as 2000 + 300 ∗
iteration steps.
Acknowledgments
This work is supported in part by NSF Grant No. CCF- 0915016. Lirong Xia is supported by NSF
under Grant #1136996 to the Computing Research Association for the CIFellows Project. We thank
Craig Boutilier, Jonathan Huang, Tyler Lu, Nicolaus Tideman, Paolo Viappiani, and anonymous
NIPS-12 reviewers for helpful comments and suggestions, or help on the datasets.
References
[1] Steven Berry, James Levinsohn, and Ariel Pakes. Automobile prices in market equilibrium. Economet-
rica, 63(4):841 -- 890, 1995.
[2] Henry David Block and Jacob Marschak. Random orderings and stochastic theories of responses.
Contributions to Probability and Statistics, pages 97 -- 132, 1960.
In
[3] James G. Booth and James P. Hobert. Maximizing Generalized Linear Mixed Model Likelihoods with an
Automated Monte Carlo EM Algorithm. JRSS. Series B, 61(1):265 -- 285, 1999.
[4] Steve Brooks, Andrew Gelman, Galin Jones, and Xiao-Li Meng, editors. Handbook of Markov Chain
Monte Carlo. Chapman and Hall/CRC, 2011.
3The use of predictive likelihood allows us to evaluate the performance of the estimated parameters on the
rest of the data, and is similar in this sense to cross validation for supervised learning.
9
[5] Francois Caron and Arnaud Doucet. Efficient Bayesian Inference for Generalized Bradley-Terry Models.
Journal of Computational and Graphical Statistics, 21(1):174 -- 196, 2012.
[6] Marquis de Condorcet. Essai sur l'application de l'analyse `a la probabilit´e des d´ecisions rendues `a la
pluralit´e des voix. Paris: L'Imprimerie Royale, 1785.
[7] Vincent Conitzer, Matthew Rognlie, and Lirong Xia. Preference functions that score rankings and maxi-
mum likelihood estimation. In Proc. IJCAI, pages 109 -- 115, 2009.
[8] Vincent Conitzer and Tuomas Sandholm. Common voting rules as maximum likelihood estimators. In
Proc. UAI, pages 145 -- 152, 2005.
[9] Lester R. Ford, Jr. Solution of a ranking problem from binary comparisons. The American Mathematical
Monthly, 64(8):28 -- 33, 1957.
[10] Isobel Claire Gormley and Thomas Brendan Murphy. A grade of membership model for rank data.
Bayesian Analysis, 4(2):265 -- 296, 2009.
[11] Przemyslaw Grzegorzewski. Kendall's correlation coefficient for vague preferences. Soft Computing,
13(11):1055 -- 1061, 2009.
[12] John Guiver and Edward Snelson. Bayesian inference for Plackett-Luce ranking models. In Proc. ICML,
pages 377 -- 384, 2009.
[13] Edith Hemaspaandra, Holger Spakowski, and Jorg Vogel. The complexity of Kemeny elections. Theoret-
ical Computer Science, 349(3):382 -- 391, December 2005.
[14] David R. Hunter. MM algorithms for generalized Bradley-Terry models.
volume 32, pages 384 -- 406, 2004.
In The Annals of Statistics,
[15] Toshihiro Kamishima. Nantonac collaborative filtering: Recommendation based on order responses. In
Proc. KDD, pages 583 -- 588, 2003.
[16] Tie-Yan Liu. Learning to Rank for Information Retrieval. Springer, 2011.
[17] Tyler Lu and Craig Boutilier. Learning mallows models with pairwise preferences. In Proc. ICML, pages
145 -- 152, 2011.
[18] R. Duncan Luce. Individual Choice Behavior: A Theoretical Analysis. Wiley, 1959.
[19] Daniel McFadden. Conditional logit analysis of qualitative choice behavior. In Frontiers of Econometrics,
pages 105 -- 142, New York, NY, 1974. Academic Press.
[20] Carl N. Morris. Natural Exponential Families with Quadratic Variance Functions. Annals of Statistics,
10(1):65 -- 80, 1982.
[21] R. L. Plackett. The analysis of permutations. JRSS. Series C, 24(2):193 -- 202, 1975.
[22] Andr´s Pr´ekopa. Logarithmic concave measures and related topics. In Stochastic Programming, pages
63 -- 82. Academic Press, 1980.
[23] Ariel D. Procaccia, Sashank J. Reddi, and Nisarg Shah. A maximum likelihood approach for selecting
sets of alternatives. In Proc. UAI, 2012.
[24] Frank Proschan and Yung L. Tong. Chapter 29. log-concavity property of probability measures. FSU
techinical report Number M-805, pages 57 -- 68, 1989.
[25] Magnus Roos, Jorg Rothe, and Bjorn Scheuermann. How to calibrate the scores of biased reviewers by
quadratic programming. In Proc. AAAI, pages 255 -- 260, 2011.
[26] Louis Leon Thurstone. A law of comparative judgement. Psychological Review, 34(4):273 -- 286, 1927.
[27] Nicolaus Tideman. Collective Decisions and Voting: The Potential for Public Choice. Ashgate Publishing,
2006.
[28] Greg C. G. Wei and Martin A. Tanner. A Monte Carlo Implementation of the EM Algorithm and the Poor
Man's Data Augmentation Algorithms. JASA, 85(411):699 -- 704, 1990.
[29] Lirong Xia and Vincent Conitzer. A maximum likelihood approach towards aggregating partial orders. In
Proc. IJCAI, pages 446 -- 451, 2011.
[30] Lirong Xia, Vincent Conitzer, and J´erome Lang. Aggregating preferences in multi-issue domains by using
maximum likelihood estimators. In Proc. AAMAS, pages 399 -- 406, 2010.
[31] John I. Jr. Yellott. The relationship between Luce's Choice Axiom, Thurstone's Theory of Comparative
Judgment, and the double exponential distribution. J. of Mathematical Psychology, 15(2):109 -- 144, 1977.
[32] H. Peyton Young. Optimal voting rules. Journal of Economic Perspectives, 9(1):51 -- 64, 1995.
10
|
1811.05053 | 1 | 1811 | 2018-11-13T00:13:24 | Distributed Cooperative Spectrum Sharing in UAV Networks Using Multi-Agent Reinforcement Learning | [
"cs.MA"
] | In this paper, we develop a distributed mechanism for spectrum sharing among a network of unmanned aerial vehicles (UAV) and licensed terrestrial networks. This method can provide a practical solution for situations where the UAV network may need external spectrum when dealing with congested spectrum or need to change its operating frequency due to security threats. Here we study a scenario where the UAV network performs a remote sensing mission. In this model, the UAVs are categorized into two clusters of relaying and sensing UAVs. The relay UAVs provide a relaying service for a licensed network to obtain spectrum access for the rest of UAVs that perform the sensing task. We develop a distributed mechanism in which the UAVs locally decide whether they need to participate in relaying or sensing considering the fact that communications among UAVs may not be feasible or reliable. The UAVs learn the optimal task allocation using a distributed reinforcement learning algorithm. Convergence of the algorithm is discussed and simulation results are presented for different scenarios to verify the convergence. | cs.MA | cs |
Distributed Cooperative Spectrum Sharing in UAV Networks Using
Multi-Agent Reinforcement Learning
Alireza Shamsoshoaraa, Mehrdad Khaledia, Fatemeh Afghaha, Abolfazl Razia, Jonathan
Ashdownb
aSchool of Informatics, Computing and Cyber Systems, Northern Arizona University, Flagstaff, AZ, USA
bComputer Information Systems Department, SUNY Polytechnic Institute, Utica, NY, USA
Abstract
In this paper, we develop a distributed mechanism for spectrum sharing among a network of
unmanned aerial vehicles (UAV) and licensed terrestrial networks. This method can provide
a practical solution for situations where the UAV network may need external spectrum when
dealing with congested spectrum or need to change its operational frequency due to security
threats. Here we study a scenario where the UAV network performs a remote sensing mission.
In this model, the UAVs are categorized to two clusters of relaying and sensing UAVs. The
relay UAVs provide a relaying service for a licensed network to obtain spectrum access
for the rest of UAVs that perform the sensing task. We develop a distributed mechanism
in which the UAVs locally decide whether they need to participate in relaying or sensing
considering the fact that communications among UAVs may not be feasible or reliable. The
UAVs learn the optimal task allocation using a distributed reinforcement learning algorithm.
Convergence of the algorithm is discussed and simulation results are presented for different
scenarios to verify the convergence1.
Keywords: Spectrum Sharing, multi-Agent Learning, UAV Networks, reinforcement
learning.
1. Introduction
Unmanned Aerial Vehicles (UAVs) have been recently used in many civilian, commercial
and military applications [23, 22, 15, 4, 8, 2]. With recent advances in design and production
of UAVs, the global market revenue of UAVs is expected to reach $11.2 billion by 2020 [28].
Spectrum management is one of the key challenges in UAV networks, since spectrum
shortage can impede the operation of these networks. In particular, in applications involving
1This material is based upon the work supported by the National Science Foundation under Grant No.
1755984.
Accepted in IEEE Consumer Communications & Networking Conference - CCNC 2019
a low-latency video streaming, the UAVs may require additional spectrum to complete their
mission. The conventional spectrum sharing mechanism such as spectrum sensing may not
be very practical in UAV systems noting the considerable required energy for spectrum
sensing or the fact that they cannot guarantee a continuous spectrum access. The property-
right spectrum sharing techniques operate based on an agreement between the licensed and
unlicensed users where the spectrum owners lease their spectrum to the unlicensed ones in
exchange for certain services such as cooperative relaying or energy harvesting.
In this paper, we studied the problem of limited spectrum in UAV networks and consid-
ered a relay-based cooperative spectrum leasing scenario in which a group of UAVs in the
network cooperatively forward data packet for a ground primary user (PU) in exchange for
spectrum access. The rest of the UAVs in the network utilize the obtained spectrum for
transmission and completion of the remote sensing operation. Thus, the main problem is to
partition the UAV network into two task groups in a distributed way.
It is worth noting that cooperative spectrum sharing has been studied previously in the
context of cognitive radio networks [29, 1, 9, 21]. The existing models are mostly centralized
and the set of relay nodes is typically chosen by the PU. Such solutions, however, are
not applicable to UAV networks, due to their distributed infrastructure and autonomous
functionality.
To tackle this problem, we utilize multi-agent reinforcement learning [7, 13, 17, 6, 18,
16, 19], which is an effective tool for designing algorithms in distributed systems, where the
environment is unknown and a reliable communication among agents is not guaranteed. The
main problems in distributed multi-agent reinforcement learning include dealing with state
space complexity and the lack of complete information about other agents. There have been
proposals in the literature to address these issues through message passing or simplifying
assumptions. For instance, [7] assumes that the decision of an agent depends only on a
limited group of other agents, which decomposes the state space and simplifies the problem.
In another work [6], a Bayesian setting is proposed where each agent has some distributional
knowledge about other agents' decisions. Such simplifications, however, are not applicable
to the distributed UAV network environment.
In this paper, we propose a distributed multi-agent reinforcement learning algorithm for
task allocation among UAVs. Each UAV either joins a relaying group to provide relaying
service for the PU or performs data transmission to the UAV fusion center. In this approach,
each UAV maintains a local table about the respective rewards for its actions in different
states. The tables are updated locally based on a feedback from PU receiver and the UAV
fusion node. We define utilities for both the PU and the UAV network, and the objective is
to maximize the total utility of the system (sum utility of the PU and the UAV network).
2
We discuss the convergence of our learning algorithm and we present simulation results to
verify the convergence to the optimal solution.
The remainder of this paper is organized as follows. In Section 2, the system model and
the assumptions of the proposed model are described. In Section 3, we propose a distributed
multi-agent learning algorithm to solve the spectrum sharing problem.
In Section 4, we
present simulation results and discuss the performance of our distributed learning algorithm.
Finally, we make concluding remarks in Section 5.
2. System Model
We consider a licensed primary user (PU) who is willing to share a part of its spectrum
with a network of UAVs, in exchange for receiving a cooperative relaying service. The UAV
network consists of N UAVs which can be partitioned into two sets depending on the task
of the UAV. In fact, UAVs either relay for the PU or utilize the spectrum to transmit their
own packets to the fusion center. Let K be the number of nodes who perform the relaying
task and N − K denote the number of UAVs that transmit packets to their fusion center.
In this paper, we assume that both the PU's transmitter and receiver are terrestrial, while
UAVs are operating in high elevation. Also, we assume no reliable direct link exists between
the PU's transmitter and receiver. Moreover, there is zero chance for direct transmission
between the UAVs' source and fusion, due to their distances from the fusion center. Fig.
1 illustrates a sample scenario with 6 total nodes, where the nodes are partitioned into a
set of 4 relay nodes for the fusion center, and 2 other nodes relay information for the PU
receiver on the ground.
The PU's transmitter intends to send its packet to a designated receiver, which is far
away from its location. Hence, a single or a number of UAVs are required to deliver its
information to the receiver. In addition, we assume that the UAVs' spectrum is congested
or unreliable, therefore the UAVs are required to lease additional spectrum from the PU
to communicate with their fusion. By delivering the PU's packet, the UAVs gain spectrum
access to send their own packets. All the UAVs transmitters and receivers are assumed
to be equipped with a single antenna. Also, we assume that the channels between UAVs,
source, fusion, and PU transmitter and receiver are slow Rayleigh fading with a constant
coefficient over one time slot. The channel coefficients are defined as follows: i) hP T,Ui refers
to the channel parameters between the PU's transmitter and ith UAV; ii) hUi,P R denotes
the parameters between the ith UAV and the PU's receiver; iii) hS,Ui and hUi,F , respectively
denote the channel coefficients between the Source and the ith UAV, and between the ith
UAV and the fusion center. For the sake of simplicity, the instant Channel State Information
3
Figure 1: System Model: A sample Scenario with 6 UAVs, where four UAVs handle packets relaying between
the Source and Fusion Center and two UAVs relay packets for the Primary User.
Figure 2: Communication channels for a single relay
(CSI) are assumed to be available for all UAVs following similar works in [30, 32, 5, 3, 26]
The source of the noise at the receivers is considered as a symmetric normally distributed
random variable, denoted by z ∼ CN (0, σ2). Many works such as [24, 20, 26] optimized the
power consumption and nodes' lifetime in this area. On the other hand, power optimization
is not the purpose of this work, hence we assume constant powers during the transmissions.
However, the transmission power for the Source and the PU transmitter is less than those
of the UAVs. Half-duplex strategy is utilized in this work. Without loss of generality, time-
division notations are characterized in order to ensure the half-duplex operations. After
these assumptions, the channel and system model for a single relay is shown in Fig. 2. In
this model, all UAVs and terminals utilize a single antenna for transmission.
In the first half of a transmission cycle, the source transmits its packet and the relay
4
UAVs receive the information. The channel model for the first half is presented as follows:
yU [n] = hS,U xS[n] + zr[n],
(1)
where xS is the source's transmitted signal and yU is the UAV's received signal. Then, in
the second half of the transmission, the UAV sends the received packet in the previous time
slot. We can write the second half as another model for the received signal as follow:
yF [n] = hU,F xU [n] + zF [n],
(2)
where xU is the UAV's transmitted signal and yF is the destination's received signal.
In equations (1) and (2), the CSI parameters hij represent the effects of the path loss
and likewise zj represents the effect of noise and interference terms at the receiver, where
i ∈ {Source, PU-Transmitter, UAV} and j ∈ {Fusion, PU-R, UAV}. In our scenario, hij is
calculated by the proper receiver.
Based on equations (1) and (2), the throughput capacity of the non-degraded discrete mem-
oryless broadcast channel is expressed in (3) [33]:
CT hroughput = max
w→x→yd
{I(w; yd)},
(3)
where d ∈ {Fusion, PU-Receiver}, w is the message word and x is the codeword which has
been assigned to each message by the encoder. Preferably, equation (3) should be solved
for the optimal joint distribution of both w and x. However, as discussed in [25], we can
achieve the suboptimal throughput rate in (4), with the aid of assumption x = w. Also,
p(x) denotes the probability mass function (pmf) for the codeword.
RT hroughput = max
p(x)
{I(x; yd)}
(4)
In scenarios, where users can exploit the existence of UAVs, different cooperation protocols
such as Decode and Forward (DF) and Amplify and Forward (AF) can be used [12]. The
idea behind the concept of cooperative relaying is that a set of relay nodes decode, amplify
and collectively "beam-form" the signal received from the source node (potentially with
help of source node itself) towards a designated destination in order to exploit transmission
diversity and increase the overall throughput of the system [14].
Considering an AF cooperation, each UAV first amplifies the signals from the source and then
cooperates with source to send its information to the fusion center or to the PU-Receiver.
According to [11], the mutual information for i) the first set of source, UAV, fusion and ii)
the second set of PU-T, UAV, PU-R can be written as equations (5) and (6), respectively.
5
In these equations, PS denotes the transmitter power from the source of the UAV network
and i specifies the index for the UAV.
ISFAF = log2(1 + PShSF2
PShS,Ui2 PUihUi,F2
+
1 + PShS,Ui2 + PUihUi,F2 )
IP U (T R)AF = log2(1 + PP ThP T,P R2
PP ThP T,Ui2 PUihUi,P R2
1 + PP ThP T,Ui2 + PUihUi,P R2 )
+
(5)
(6)
We denote the throughput rate for both primary users and source-fusion users as (7) and
(8), respectively.
RP U = IP U (T R)AF
RSF = ISFAF
(7)
(8)
It is noteworthy that these equations are valid only for cooperation with a single Relay
or UAV. However, the objective of this paper is dealing with Multi-UAV or Multi-Agent
relays. Fig. 3 demonstrates the distribution of N UAVs into two groups including K UAVs
facilitating the air source-to-fusion communication and N − K UAVs providing relaying
service for a ground-based primary transmitter-receiver pair. Hence, the equations for multi-
UAV should be changed to (9) and (10). In (9), i defines the lower bound for the first UAV
in the source-fusion pair and i + N − K denotes the upper bound.
RSF (Multi-UAV) = log2(1 + PShSF2
i+N−K(cid:88)
PShS,Uj2 PUjhUj ,F2
1 + PShS,Uj2 + PUjhUj ,F2 )
+
j=i
(9)
Here, RSF (Multi-UAV) is the achievable rate for the fusion center. This rate is achieved
with the help of (N − K) UAVs. PS and PUi are transmission powers for the source and
the ith UAV, respectively. Also, hSF denotes the channel coefficient for the pair of source-
fusion center, hS,Uj stands for the channel between the source and jth UAV, and finally hUj ,F
denotes CSI for the jth UAV and the fusion. In (10), m and m + K define the lower and
upper bound for the first and last UAV in the source-fusion pair respectively.
RP U (Multi-UAV) = log2(1 + PP ThP T,P R2
m+K(cid:88)
PP ThP T,Ul2 PUlhUl,P R2
1 + PP ThP T,Ul2 + PUlhUl,P R2 )
+
l=m
6
(10)
Figure 3: System Model: Dividing UAVs into K and N − K groups, for cooperating in two sets of Source-
Fusion and Primary Transmitter-Receiver.
In (10), RP U (Multi-UAV) is the achievable rate for the primary transmitter-receiver pair
with the aid of K UAVs. PP T and PUl are transmission power for the primary user and
the ith UAV, respectively. Moreover, hP T,P R, denotes the channel coefficients for primary
transmitter and receiver. hP T,Ul stands for the primary transmitter and lth UAV. Finally
hUl,P R is CSI parameters for the lth UAV and the primary receiver. Based on the assumption
of long distance between the source and the fusion center and also the long distance between
the primary transmitter and receiver, we can assume that hSF and hP T,P R are negligible.
Time is slotted and at the end of each time slot, the fusion center and the primary
receiver send feedback to the UAVs informing them about the achieved accumulated rates.
This information is used by each UAV to decide on joining a task group. The goal is to find
the optimal task allocation for UAVs in a fully distributed way such that the total utility of
the system (i.e. sum utility of UAV network (9) and the PU (10)) is maximized. We assume
7
that the UAVs decide locally with no information exchange among themselves.
It is noteworthy that in some cases, the maximum throughput is achieved when all
UAVs join the same set and deliver packets only for one set, which is not consistent with
the proposed model. If all UAVs are distributed in the set of source-fusion, then the total
throughput rate is zero because there is no available spectrum for UAVs to utilize for their
transmission. Also, if all UAVs are partitioned in the primary set, then the sum throughput
rate is equal to the rate of the primary user.
In this case the proposed method handles
this issue by considering the Jain fairness index [10]. Based on the fact that we only have
two sets and based on the Jain index definition, (11) describes the fairness for the proposed
method in our system model.
× ((cid:80)
i xi)2(cid:80)
i x2
i
J(x) =
1
n
,
(11)
Here, n is equal to 2 and i ∈ {0, 1} which indicates the set of source-fusion or Primary Users.
We assume that x0 and x1 are equal to the number of UAVs in the Fusion-Source set and
the Primary Users set, respectively. Therefore, we can define the fairness as (12).
F ariness =
1
2
× (#UF + #UP )2
(#UF )2 + (#UP )2
(12)
Now, if all UAVs are distributed in one set, then the fairness will be minimum (0.5), and if
the UAVs are partitioned equally among two sets, then the fairness will be maximum (1).
Based on these definitions, we define (13), as the gain value for each time slot which
indicates the efficiency and performance for the distributed UAVs in two sets.
Gain = γ1 × ∆(RateF usion)
+ γ2 × ∆(RateP rimary) + γ3 × (F airness)
(13)
In (13), ∆(RateFusion) is the difference between the rate at time t and the average of
previous rates for the fusion center and ∆(RatePrimary) is the difference between the rate at
time t and the average of previous rates for the primary user. Also, γ1, γ2, and γ3 are defined
to control the gain value. Then, we use this gain in our proposed method as described in
section 3.
3. The Distributed Learning Algorithm for Task Allocation
The proposed method is a general form of the Q-learning algorithm [31] for a distributed
multi-agent environment.
8
t , a2
Let ai
t ,··· , aN
t denote the action chosen by UAV i at time t, and let Ai denote the set of all
possible actions for UAV i. We consider two possible actions for a UAV that correspond
to either joining the relaying task group or the fusion task partition. Therefore, the set of
possible actions are identical across UAVs. We denote the action vector of UAVs at time t
t ), and we refer to the set of all possible action vectors by U. There
by ut = (a1
is a finite set of states S, where state s ∈ S corresponds to the current task partition. A
deterministic transition rule δ governs the transition between states, i.e. δ : S × U → S.
The reward function r maps the current state and action vector to a real value, that is
r : S × U → R. At the beginning of each time step, the UAVs observe the current state
(this information is obtained by the feedback from the previous step). Then, each UAV
independently decides on its action (i.e. which task group to join) without knowing any
information about actions of the other agents. The rewards associated with the UAVs'
actions are computed by the PU receiver and the UAV fusion. The reward is basically the
gain obtained from the task partitioning, taking into account the utilities of the PU and the
UAV network. After the reward is calculated, a feedback message from the PU receiver and
the UAV fusion is broadcasted to the UAVs. This feedback message contains the reward
and the current task partitions.
The feedback information is used to update and maintain local Q-tables at each UAV. A
Q-table basically represents the quality of different actions for a given state. For instance,
qi
t(s, a) denotes the quality of action a at state s for UAV i at time t. Individual Q-tables are
updated as follows. At first, the tables are initialized with qi
0(s, a) = 0. Then, the following
equation is used to update the Q-tables:
qi
t+1(s, a) =
qi
t(s, a),
if s (cid:54)= st or a (cid:54)= ai
t,
(1 − α) qi
α ·(cid:0)rt + β · maxa(cid:48)∈Ai qi
t(s, a)+
t(δ(st, ut), a(cid:48))(cid:1),
otherwise,
(14)
where 0 ≤ α < 1 is the learning rate, rt is the reward or the gain obtained at time t, as
defined in the system model, and 0 ≤ β < 1 is the discount factor to control the weight of
future rewards in the current decisions.
The main idea is that in our distributed environment, the UAVs are unable to keep a
global Q-table, corresponding to the current action vectors, i.e. Q : S × U → R. Instead,
each UAV i keeps a local (and considerably smaller) Q-table which cares about its own
9
current action, i.e. qi : S × Ai → R. This approach significantly reduces the complexity
of the algorithm and eliminates the need for coordination (or sharing information) with
other UAVs at the time of decision making. However, we need a projection method that
compresses the information of the global Q-table into the local small tables.
The results in [13] prove that in a deterministic multi-agent Markov decision process
and for the same sequence of states and actions, if every independent learner chooses locally
optimal actions, the result would be the same as choosing the optimal action from a global
table. We utilize this result and consider an optimistic projection method that assumes each
UAV chooses the maximum quality action from its local table. This reasonable assumption
is a necessary condition for the optimality of the learning algorithm. It is worth noting that
the existence of a unique optimal solution is the sufficient condition for the optimality of
this algorithm. It means that there should be a unique task partition, which results in the
maximum total utility. If multiple task partitions yield the maximum utility, it is possible
that the UAVs act optimally and choose the optimal actions in their local Q-tables, but
the combination of their actions may not be optimal. In this case, message passing among
UAVs is needed as they need to coordinate decisions at every step.
It should also be noted that in learning algorithms we need a balance between exploring
new actions and exploiting the previously learned quality of actions. Therefore, a greedy
strategy that always exploits the Q-table and chooses the optimal action from the Q-table
may not provide enough exploration for the UAV to guarantee an optimal performance. A
very common approach is to add some randomness to the policy [27]. We use -greedy with
a decaying exploration, in which a UAV chooses a random exploratory action at state s with
probability (s) = c/n(s), where 0 < c < 1 and n(s) is the number of times the state s
has been observed so far. The UAV exploits greedily from its Q-table with probability of
1− (s). In this approach, the probability of exploration decays over time as the UAVs learn
more.
Similar to the original Q-learning for a single agent environments, the proposed learning
algorithm converges if the state-action pairs are observed infinitely many times. Also, the
time complexity of the algorithm is in the order of O(S×Ai), where S is the size of the
state space, and Ai is the size of action space for UAV i. Since there are only two possible
actions in our application, the complexity can be expressed as O(S). In terms of space
complexity, each UAV i needs to keep a table of size S × Ai.
10
Figure 4: Topology for 2 UAVs in a 100 x 100 mission area.
4. Simulation Results
In this section, we present the simulation results to evaluate the performance of the
proposed method. We simulate our system model for a ground-based primary transmitter-
receiver pair along with the pair of source and fusion for the UAV network. The location
of primary users, source and fusion are fixed during the simulation. However, the UAVs are
distributed randomly in the environment. The channels between nodes i and j are obtained
from hi,j ∼ CN (0, d−2
i,j ), where di,j is the distance between nodes i and j. The duration of
one time slot, T , is assumed to be equal to 1. The values of γ1 , γ2 and γ3 are set to 2, 2
and 0.4, respectively.
Scenario I: 2 UAVs
In the first scenario, we consider two UAVs to be partitioned into two task groups. The
network topology for this scenario is demonstrated in Fig. 4. Since in this scenario we
only have 2 nodes, the possible states for task allocation is equal to 22 = 4. Hence, the
Q-tables will be learned after a few iterations. Fig. 5 illustrates the summation of the
obtained throughput. The convergence to the optimal task allocation occurs after the 35th
iteration, since the number of states is relatively small. The matrix below shows the final
task allocation values for these UAVs.
(cid:104)
(cid:105)
0 1
11
Figure 5: Sum Rate for 2 UAVs for 100 iterations
In this notation, 0 corresponds to the set of source-fusion and 1 means the set of the primary
users. UAV1 who has a lower relative distance to the source-fusion, is allocated to the fusion
set, while UAV2 is allocated to the another set to relay for the primary network.
Scenario II: 6 UAVs
In this scenario, we consider 6 UAVs to show that the convergence of the proposed method
is achieved after more iterations compared to the case of 2 UAVs in the first scenario, since
the number of states with 6 nodes is equal to 26 = 64. This means, at least 64 iterations are
required for the algorithm to just test all the states.
Fig. 6 demonstrates the network topology with these 6 UAVs for the primary user and
the fusion. As we can see in Fig. 7, the convergence to the best task allocation occurred
after 240 iterations. This implies that the more UAVs are added to the model, the more
iterations will be taken to the convergence epoch. Moreover, Fig. 8 shows the number
task partitions) in this scenario. After the 240th
of UAVs switching their actions (i.e.
iteration, when the convergence happens, we see that no UAV changes its task partition,
and the number of switches stays at zero.
Also, task matrix shown below denotes the final task allocation for the 6 UAVs.
(cid:104)
(cid:105)
1 0 1 0 0 1
12
Figure 6: Topology for 6 UAVs in 100 x 100 simulation field
Based on this matrix, UAVi; i ∈ {2, 4, 5} are considered for the set of source-fusion and the
rest of UAVs are assigned to the relay task group for the primary network. This allocation
makes sense considering the location of UAVs and their relative distances.
5. Conclusion
In this paper, we studied the task allocation problem for spectrum management in UAV
networks. We considered a cooperative relay system in which a group of UAVs provide
relaying service for a ground-based primary user in exchange for spectrum access. The
borrowed spectrum is not necessarily used by the relay UAV, rather is used by other UAVs
to transmit their own information to a fusion center. This makes a win-win situation for
both networks. We defined utilities for both the UAV network and the ground-based primary
network based on the achieved rates. Next, we proposed a distributed learning algorithm
by which the UAVs take proper decisions by joining the relaying or fusion task groups
without the need for information exchange or knowledge about other UAV's decisions. The
algorithm converges to the optimal task partitioning that maximizes the total utility of the
system. Simulation results were presented in different scenarios to verify the convergence of
the proposed algorithm.
13
Figure 7: Sum Rate for 6 UAVs for 1000 Iterations
References
[1] Afghah, F., Costa, M., Razi, A., Abedi, A., Ephremides, A., Dec 2013. A reputation-based stackelberg
game approach for spectrum sharing with cognitive cooperation. In: Decision and Control (CDC), 2013
IEEE 52nd Annual Conference on. pp. 3287 -- 3292.
[2] Afghah, F., Razi, A., May 2014. Cooperative spectrum leasing in cognitive radio networks. In: 2014
National Wireless Research Collaboration Symposium. pp. 106 -- 111.
[3] Afghah, F., Shamsoshoara, A., Njilla, L., Kamhoua, C., 2018. A reputation-based stackelberg game
model to enhance secrecy rate in spectrum leasing to selfish iot devices. arXiv preprint arXiv:1802.05832.
[4] Afghah, F., Zaeri-Amirani, M., Razi, A., Chakareski, J., Bentley, E., June 2018. A coalition formation
approach to coordinated task allocation in heterogeneous uav networks. In: 2018 Annual American
Control Conference (ACC). pp. 5968 -- 5975.
[5] Al-Talabani, A., Deng, Y., Nallanathan, A., Nguyen, H. X., 2016. Enhancing secrecy rate in cognitive
radio networks via stackelberg game. IEEE Transactions on Communications 64 (11), 4764 -- 4775.
[6] Chalkiadakis, G., Boutilier, C., 2003. Coordination in multiagent reinforcement learning: A bayesian
approach. In: Proceedings of the Second International Joint Conference on Autonomous Agents and
Multiagent Systems. AAMAS '03. ACM, New York, NY, USA, pp. 709 -- 716.
URL http://doi.acm.org/10.1145/860575.860689
[7] Guestrin, C., Lagoudakis, M. G., Parr, R., 2002. Coordinated reinforcement learning. In: Proceedings of
the Nineteenth International Conference on Machine Learning. ICML '02. Morgan Kaufmann Publishers
Inc., San Francisco, CA, USA, pp. 227 -- 234.
[8] Khaledi, M., Rovira-Sugranes, A., Afghah, F., Razi, A., June 2018. On greedy routing in dynamic
14
Figure 8: Number of Switching UAVs for 1000 iterations
uav networks. In: 2018 IEEE International Conference on Sensing, Communication and Networking
(SECON Workshops). pp. 1 -- 5.
[9] Korenda, A., Amirani, M. Z., Afghah, F., March 2017. A hierarchical stackelberg-coalition formation
game theoretic framework for cooperative spectrum leasing. In: 51th Annual Conference on Information
Systems and Sciences (CISS'17). pp. 1 -- 6.
[10] Lan, T., Kao, D., Chiang, M., Sabharwal, A., 2010. An axiomatic theory of fairness in network resource
allocation. IEEE.
[11] Laneman, J. N., Tse, D. N., Wornell, G. W., 2004. Cooperative diversity in wireless networks: Efficient
protocols and outage behavior. IEEE Transactions on Information theory 50 (12), 3062 -- 3080.
[12] Laneman, J. N., Wornell, G. W., Tse, D. N., 2001. An efficient protocol for realizing cooperative
diversity in wireless networks. In: Information Theory, 2001. Proceedings. 2001 IEEE International
Symposium on. IEEE, p. 294.
[13] Lauer, M., Riedmiller, M. A., 2000. An algorithm for distributed reinforcement learning in cooperative
multi-agent systems. In: Proceedings of the Seventeenth International Conference on Machine Learning.
ICML '00. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, pp. 535 -- 542.
URL http://dl.acm.org/citation.cfm?id=645529.658113
[14] Levin, G., Loyka, S., 2012. Amplify-and-forward versus decode-and-forward relaying: which is better?
In: 22th International Zurich Seminar on Communications (IZS). Eidgenossische Technische Hochschule
Zurich, pp. 123 -- 126.
[15] Mousavi, S., Afghah, F., Ashdown, J. D., Turck, K., April 2018. Leader-follower based coalition forma-
tion in large-scale uav networks, a quantum evolutionary approach. In: IEEE INFOCOM 2018 - IEEE
Conference on Computer Communications Workshops (INFOCOM WKSHPS). pp. 882 -- 887.
15
[16] Mousavi, S., Schukat, M., Howley, E., Borji, A., Mozayani, N., 2016. Learning to predict where to look
in interactive environments using deep recurrent q-learning. arXiv preprint arXiv:1612.05753.
[17] Mousavi, S. S., Schukat, M., Howley, E., 2016. Deep reinforcement learning: an overview. In: Proceed-
ings of SAI Intelligent Systems Conference. Springer, pp. 426 -- 440.
[18] Mousavi, S. S., Schukat, M., Howley, E., 2017. Traffic light control using deep policy-gradient and
value-function-based reinforcement learning. IET Intelligent Transport Systems 11 (7), 417 -- 423.
[19] Mousavi, S. S., Schukat, M., Howley, E., Mannion, P., 2017. Applying q (λ)-learning in deep reinforce-
ment learning to play atari games. AAMAS Adaptive Learning Agents (ALA) Workshop.
[20] Mozaffari, M., Saad, W., Bennis, M., Debbah, M., 2016. Optimal transport theory for power-efficient
deployment of unmanned aerial vehicles. In: Communications (ICC), 2016 IEEE International Confer-
ence on. IEEE, pp. 1 -- 6.
[21] Namvar, N., Afghah, F., March 2015. Spectrum sharing in cooperative cognitive radio networks: A
matching game framework. In: 2015 49th Annual Conference on Information Sciences and Systems
(CISS). pp. 1 -- 5.
[22] Peng, H., Razi, A., Afghah, F., Ashdown, J. D., 2018. A unified framework for joint mobility prediction
and object profiling of drones in uav networks. CoRR abs/1808.00058.
[23] Razi, A., Afghah, F., Chakareski, J., Oct 2017. Optimal measurement policy for predicting uav network
topology. In: 2017 51st Asilomar Conference on Signals, Systems, and Computers. pp. 1374 -- 1378.
[24] Roberge, V., Tarbouchi, M., Labont´e, G., 2013. Comparison of parallel genetic algorithm and particle
swarm optimization for real-time uav path planning. IEEE Transactions on Industrial Informatics 9 (1),
132 -- 141.
[25] Shafiee, S., Ulukus, S., 2007. Achievable rates in gaussian miso channels with secrecy constraints. In:
Information Theory, 2007. ISIT 2007. IEEE International Symposium on. IEEE, pp. 2466 -- 2470.
[26] Shamsoshoara, A., Darmani, Y., 2015. Enhanced multi-route ad hoc on-demand distance vector routing.
In: Electrical Engineering (ICEE), 2015 23rd Iranian Conference on. IEEE, pp. 578 -- 583.
[27] Singh, S., Jaakkola, T., Littman, M. L., Szepesv´ari, C., March 2000. Convergence results for single-step
on-policy reinforcement-learning algorithms. Machine Learning. 38 (3), 287 -- 308.
[28] STAMFORD, C., Feb 2017. Gartner says almost 3 million personal and commercial drones will be
shipped in 2017.
URL http://www.gartner.com/newsroom/id/3602317
[29] Stanojev, I., Simeone, O., Bar-Ness, Y., Yu, T., May 2008. Spectrum leasing via distributed cooperation
in cognitive radio. In: 2008 IEEE International Conference on Communications. pp. 3427 -- 3431.
[30] Stanojev, I., Yener, A., 2013. Improving secrecy rate via spectrum leasing for friendly jamming. IEEE
Transactions on Wireless Communications 12 (1), 134 -- 145.
[31] Watkins, C. J., Dayan, P., 1992. Q-learning. Machine learning 8 (3-4), 279 -- 292.
[32] Wu, Y., Liu, K. R., 2011. An information secrecy game in cognitive radio networks. IEEE Transactions
on Information Forensics and Security 6 (3), 831 -- 842.
[33] Zhang, P., Yuan, J., Chen, J., Wang, J., Yang, J., April 2009. Analyzing amplify-and-forward and
decode-and-forward cooperative strategies in wyner's channel model. In: 2009 IEEE Wireless Commu-
nications and Networking Conference. pp. 1 -- 5.
16
|
1909.01239 | 1 | 1909 | 2019-09-03T15:11:00 | Threshold Greedy Based Task Allocation for Multiple Robot Operations | [
"cs.MA"
] | This paper deals with large-scale decentralised task allocation problems for multiple heterogeneous robots with monotone submodular objective functions. One of the significant challenges with the large-scale decentralised task allocation problem is the NP-hardness for computation and communication. This paper proposes a decentralised Decreasing Threshold Task Allocation (DTTA) algorithm that enables parallel allocation by leveraging a decreasing threshold to handle the NP-hardness. Then DTTA is upgraded to a more practical version Lazy Decreasing Threshold Task Allocation (LDTTA) by combining a variant of Lazy strategy. DTTA and LDTTA can release both computational and communicating burden for multiple robots in a decentralised network while providing an optimality bound of solution quality. To examine the performance of the proposed algorithms, this paper models a multi-target surveillance scenario and conducts Monte-Carlo simulations. Simulation results reveal that the proposed algorithms achieve similar function values but consume much less running time and consensus steps compared with benchmark decentralised task allocation algorithms. | cs.MA | cs | Threshold Greedy Based Task Allocation for Multiple Robot
Operations (cid:63)
Teng Li a, Hyo-Sang Shin* a, Antonios Tsourdos a
aCranfield University, Cranfield MK43 0AL, UK
9
1
0
2
p
e
S
3
]
A
M
.
s
c
[
1
v
9
3
2
1
0
.
9
0
9
1
:
v
i
X
r
a
Abstract
This paper deals with large-scale decentralised task allocation problems for multiple heterogeneous robots with monotone
submodular objective functions. One of the significant challenges with the large-scale decentralised task allocation problem
is the NP-hardness for computation and communication. This paper proposes a decentralised Decreasing Threshold Task
Allocation (DTTA) algorithm that enables parallel allocation by leveraging a decreasing threshold to handle the NP-hardness.
Then DTTA is upgraded to a more practical version Lazy Decreasing Threshold Task Allocation (LDTTA) by combining a
variant of Lazy strategy. DTTA and LDTTA can release both computational and communicating burden for multiple robots in
a decentralised network while providing an optimality bound of solution quality. To examine the performance of the proposed
algorithms, this paper models a multi-target surveillance scenario and conducts Monte-Carlo simulations. Simulation results
reveal that the proposed algorithms achieve similar function values but consume much less running time and consensus steps
compared with benchmark decentralised task allocation algorithms.
Key words: Multi-robot systems, decentralised task allocation, submodular welfare maximisation, lazy decreasing threshold.
1 Introduction
Multi-Robot Systems (MRS) have been gaining increas-
ing attention thanks to their ability to coordinate simul-
taneous or co-operate to achieve common goals. MRS
provides some fundamental strengths that could not be
achieved with single-agent systems, e.g., increased flex-
ibility, enhanced reliability and resilience, simultaneous
broad area coverage or capability to operate outside
the communication range of base stations. Specific ap-
plications under consideration include search and res-
cue [25, 38], precision agriculture [5, 28], and large area
surveillance [1, 8, 19, 27].
Efficient cooperation of MRS is a vital part for their suc-
cessful operations and effective assignment, termed as
task allocation, of the available resources is the key en-
abler of such cooperation. This is because the strength
of MRS hinges on the distributed nature of the sensing
resources available, making the successful assignment of
(cid:63)
* Corresponding author.
Email addresses: [email protected] (Teng Li),
[email protected] (Hyo-Sang Shin*),
[email protected] (Antonios Tsourdos).
these resources key to maximising its operational advan-
tages.
The main issue with the task allocation problem is that
it has been proven to be NP-hard in general [41]. This
means that task allocation problems require exponential
time to be solved optimally, thus require a careful craft of
approximation strategies. A key trade-off that must be
performed concerns the optimality of the solution versus
the computational complexity.
In terms of handling the NP-hardness of computation,
extensive approaches have been developed, e.g., the
heuristic approach and the approximation approach.
Heuristic approach involves genetic algorithm [3,22,46],
ant colony optimisation [6, 20, 35], market-based al-
gorithms [9, 15, 26, 30, 34] etc.. Heuristic algorithms
can achieve feasible solutions with certain convergence
speed. However, these algorithms cannot provide any
approximation guarantee for the solution quality. Ap-
proximation approach solves task allocation problems
efficiently while providing an optimality bound of solu-
tion quality [10,11,13,16,18,24,33,36,37,39,40,42,44,45],
if the problem meets certain conditions, e.g., submod-
ularity. Submodularity is a ubiquitous feature in the
optimisation problems where the marginal gain of one
element will diminish as more elements have already
Preprint submitted to Automatica
4 September 2019
been selected [4]. The detailed definition of submodu-
larity will be given in Section 2. It is well known that
the Sequential Greedy Algorithm (SGA) can provide
an optimality bound of 1/2 for maximising monotone
submodular objective functions subject to a partition
matroid constraint [31].
This paper focuses on task allocation algorithms that
utilise the approximation approach and can be lever-
aged for decentralised task allocation. Note that in a
centralised architecture, all the agents and environment
data is communicated to a centralised entity. This may
not be possible in some realistic scenarios because re-
lying on a central entity could remove resilience or the
bandwidth to communicate all the information may not
be available. Capability to be decentralised can relax
such issues and thus enabling decentralisation provides
an option that could be beneficial for extending MRS
operations in practice.
The decentralised approximation approach has been fre-
quently applied in the task allocation of MRS applica-
tions. The work [36] presented a distributed SGA for
a large group of earth-observing satellites to automat-
ically assign themselves locations based on their local
information and communication. Williams et al. [44] in-
vestigated the surveillance mission in an urban environ-
ment applying decentralised SGA with the consideration
of the intersection of multiple matroid constraints. Sun
et al. [42] solved the multi-agent coverage problems us-
ing distributed SGA in an environment with obstacles
and provided a tighter optimality bound by analysing
the curvature of submodularity. The decentralised ap-
proximation approach was also applied in search and lo-
calisation [13] and sensor network [10, 24] to get a task
allocation solution efficiently.
The communicating complexity is another aspect of the
NP-hardness in the decentralised task allocation prob-
lems. In decentralised networks, robots need to commu-
nicate with each other frequently and make a consensus
before they can find an effective task allocation solution.
Too many consensus steps could impact the efficiency of
the task allocation process and put a heavy burden on
the onboard communicating units.
In terms of relaxing the communicating complexity, sev-
eral works of the decentralised approximation approach
have been studied to reduce the reliance on the global in-
formation consensus. The benchmark decentralised task
allocation algorithm CBBA [9] can provide an undimin-
ished optimality bound of 1/2 and reduce communicat-
ing complexity by building task bundles. These task bun-
dles enable parallel task allocation hence reducing the
number of global information consensus steps. However,
as mentioned in [9], robots need to rebuild their bundles
continually once any constraint conflict appears. The is-
sue with the bundle rebuilding is that evaluating objec-
tive functions is usually time-consuming, which incurs
observably increased computational complexity. Ghare-
sifard et al. [16] provided lower bounds on the perfor-
mance of the distributed SGA according to the network
topology where each agent only had limited information
of other agents' selections. Grimsman et al. [18] extended
the work of Gharesifard et al. [16] and suggested the
best system topology design. Although these two works
reduced the communicating complexity of distributed
SGA, the quality of solutions without global informa-
tion consensus was also discounted. To handle this is-
sue, Qu et al. [37] proposed an assumption of admissible
task sets for a group of observing satellites. With the
help of this assumption, distributed SGA could achieve
an undiminished optimality bound even though only lo-
cal information and communication were accessible. The
assumption is reasonable for homogeneous and spatially
static agents, but difficult to be applied for general het-
erogeneous moving MRS.
Previous works indicate that the goals of reducing com-
putational complexity, reducing communicating com-
plexity, and achieving good solution quality are usually
conflicting in the large-scale task allocation problems,
unless resorting to particular assumptions [11, 37]. One
intuitive research question would be how to balance
these goals in general large-scale task allocation prob-
lems.
This paper aims to develop decentralised algorithms
that can reduce both computational and communicat-
ing complexity while retaining a good optimality bound
of the solution for general heterogeneous MRS in large-
scale task allocation problems. To this end, we propose
an efficient decentralised task allocation algorithm,
which is named as Decreasing Threshold Task Alloca-
tion (DTTA), by leveraging a decreasing threshold [2, 7]
to the decentralised SGA.
In each iteration of the original decentralised SGA, each
robot is required to evaluate all the remaining tasks
then make a consensus with other robots to find a task-
robot pair that can provide the largest marginal value.
Hence, only one task can be allocated during each con-
sensus step with SGA. The exploitation of the decreas-
ing threshold concept relaxes this limitation. Hence, our
proposition allows DTTA to consume fewer numbers of
objective function evaluations and consensus steps and
hence to reduce computational and communicating com-
plexity.
This paper also extends the proposed DTTA algorithm
to a more practical version, Lazy Decreasing Thresh-
old Task Allocation (LDTTA), by combining a vari-
ant of the Lazy Greedy [29]. In LDTTA, the remaining
tasks are sorted in descending order according to their
marginal values. During each iteration of LDTTA, we
propose each robot to reevaluate the first element from
the sorted remaining tasks and to compare it with the
current threshold instead of the second element which
2
is used in the original Lazy Greedy algorithm [29]. The
Lazy Greedy concept developed could further relax the
computational and communicating complexity.
The proposed DTTA and LDTTA algorithms enable
parallel allocation with the help of a decreasing thresh-
old instead of building task bundles [9] or making partic-
ular assumptions [11, 37]. To the best of our knowledge,
this paper is the first work that can enable parallel allo-
cation without incurring increased computational com-
plexity or resorting to additional specific assumptions.
Theoretical analysis reveals that DTTA and LDTTA
achieve an optimality bound of at least (1/2 − ) with
computational complexity of O( r
) for each robot,
where r = T is the number of tasks. Here, is a thresh-
old decreasing parameter for the trade-off of solution
quality versus computational and communicating com-
plexity. As expected, LDTTA outperforms DTTA, espe-
cially in terms of computational complexity.
ln r
The performances of DTTA and LDTTA are validated
via numerical simulations. For the simulation-based val-
idation, a scenario of multi-target surveillance mission
using multiple heterogeneous UAVs is modelled. Monte-
Carlo simulations are conducted to compare the per-
formances of DTTA and LDTTA with those of SGA
and CBBA. Simulation results demonstrate that the
proposed algorithms achieve almost the same objective
function values, but consume much less running time and
consensus steps, compared with benchmark algorithms.
The simulation results also confirm that the trade-off of
solution quality, running time, and consensus steps can
be obtained by adjusting the threshold decreasing pa-
rameter .
The rest of the paper is organised as follows. Section 2
provides basic concepts and definitions. The proposed
algorithm is described in detail and analysed in Section
3. Section 4 models a surveillance scenario and compares
the performances of DTTA and LDTTA with those of
benchmark algorithms through numerical simulations. A
summary of contributions and future research directions
are offered in Section 5.
2 Preliminaries
This section presents the necessary definitions and fun-
damental concepts that are related to the proposed de-
centralised task allocation algorithms.
Definition 1 (Task Allocation) The task allocation
problem is to allocate a set of tasks T to a set of robots
A so as to maximise the total value measured as
fa(Ta),
(1)
3
f (T ,A) =
(cid:88)
a∈A
where Ta is the subset of tasks allocated to robot a, fa :
2N → R≥0 is the objective function for robot a. The
notations introduced in this definition are also used in
the following part of this paper.
Definition 2 (Submodularity [14]) A set function f :
2N → R is submodular, if ∀ X, Y ⊆ N ,
f (X) + f (Y ) ≥ f (X ∩ Y ) + f (X ∪ Y ),
where N is a finite set. Equivalently, ∀ A ⊆ B ⊆ N and
u ∈ N\B,
f (A ∪ {u}) − f (A) ≥ f (B ∪ {u}) − f (B).
(2)
Eqn. (2) is known as the diminishing returns, which is
an important property of submodular functions. Specif-
ically, the marginal gain of a given element u will never
increase as more elements have already been selected.
Definition 3 (Marginal gain [23]) For a set function
f : 2N → R, a set S ⊆ N , and an element u ∈ N , define
the marginal gain of u given S as
∆f (uS) := f (S ∪ {u}) − f (S).
Definition 4 (Monotonicity [23]) A set function f :
2N → R is monotone, if ∀A ⊆ B ⊆ N , f (A) ≤ f (B). f
is non-monotone if it is not monotone.
This paper only considers monotone normalised (i.e.
f (∅) = 0) non-negative (i.e. f (S) ≥ 0, ∀S ⊆ N ) sub-
modular objective functions.
Definition 5 (Matroid [2]) A matroid is a pair M =
(N ,I) where N is a finite set, and I ⊆ 2N is a collection
of independent sets, satisfying:
• ∅ ∈ I
• if A ⊆ B, B ∈ I, then A ∈ I
• if A, B ∈ I,A < B, then ∃ u ∈ B\A such that A ∪
{u} ∈ I.
Specifically, matroid constraints include uniform ma-
troid constraint and partition matroid constraint. The
uniform matroid constraint is also called cardinality con-
straint which is a special case of matroid constraint
where any subset S ⊆ N satisfying S ≤ k is indepen-
dent. Partition matroid constraint means that a subset
S can contain at most a certain number of elements from
each of the disjoint partitions.
In this paper, one task can only be allocated to at most
one robot, but one robot can take more than one task.
Therefore, the task allocation problem can be solved as
submodular maximisation subject to a partition matroid
constraint, which is also known as Submodular Welfare
Maximisation (SWM) [39, 43].
3 Algorithms and Analysis
This section describes the proposed task allocation algo-
rithms and analyses their theoretical performances in de-
tails. First, the basic decentralised task allocation algo-
rithm DTTA is introduced in Algorithm 1. Then, DTTA
is extended to a more practical version (i.e., LDTTA)
as presented in Algorithm 2 by developing and combin-
ing a variant of the Lazy Greedy strategy [29]. Finally,
the theoretical performances of the proposed algorithms
are analysed through an equivalent reformed decreasing
threshold greedy algorithm for submodular maximisa-
tion subject to a partition matroid constraint in Algo-
rithm 3.
3.1 Algorithms
Let us discuss the details of the DTTA and LDTTA al-
gorithms. DTTA and LDTTA consist of two phases: ini-
tialisation phase (Algorithm 1, lines 1∼4; Algorithm 2,
lines 1∼10) and task assignment phase (Algorithm 1,
lines 5∼21; Algorithm 2, lines 11∼35). Na is the set that
contains the remaining tasks for robot a. A is an auxil-
iary set that contains the ids of robots who are assigned
tasks during each iteration. A = ∅ means that no robot
can find a qualified task under the current threshold,
which is a trigger of decreasing the threshold (Algorithm
1, line 6; Algorithm 2, line 12). A qualified task is a task
whose marginal value for robot a given the selected task
list Ta is no less than the current threshold θ. J is an
auxiliary set that contains the corresponding task ids
that are assigned to robots in A during each iteration.
Wa is an auxiliary set that contains the marginal values
of tasks from Na in Algorithm 2.
a
In the initialisation phase, robot a conducts preparatory
work. Robot a evaluates all tasks from Na and finds the
largest marginal value ωmax
(Algorithm 1, line 2; Algo-
rithm 2, lines 2∼8). After getting its ωmax
, robot a makes
a consensus with other robots through the M axCons
function to find the globally largest marginal value ωmax
(Algorithm 1, line 3; Algorithm 2, line 9). The value of
ωmax
a∗ will be set as the initial threshold value d (Algo-
rithm 1, line 4; Algorithm 2, line 10).
a∗
a
a and ω∗
In the task assignment phase, robot a searches quali-
fied tasks then negotiates with other robots to assign
tasks and solve conflicts. First, robot a tries to find one
qualified task from the remaining tasks in Na. The no-
tations j∗
a are used to store the qualified task
and its marginal value (Algorithm 1, lines 7∼9; Algo-
rithm 2, lines 14∼17). Robot a will stop searching once
it finds one qualified task which is not necessarily the
best one. There is no need to reevaluate all the remain-
ing tasks in Na. Note that, in the decentralised SGA,
each robot needs to reevaluate all the remaining tasks to
find the task that can provide the largest marginal value.
If robot a cannot find a qualified task, the values of ω∗
a
4
and j∗
a will be set as 0 and none, respectively (Algorithm
1, lines 11∼12; Algorithm 2, lines 28∼29). Then, robot
a coordinates with other robots through the M axCoor
function to assign tasks and solve conflicts (Algorithm
1, lines 14∼18; Algorithm 2, lines 18∼23, lines 30∼32).
If no robot can provide a qualified task, then the auxil-
iary set A will be empty, and all robots will reduce the
threshold to the next value θ(1 − ) and reevaluate the
remaining tasks until reaching the stop condition (Algo-
rithm 1, line 5; Algorithm 2, line 11).
fa(jaTa)
while A (cid:54)= ∅ do
r d; θ ← θ(1 − )) do
Algorithm 1 decentralised DTTA for Robot a
Input: fa : 2T → R≥0,T ,A,
Output: A set Ta ⊆ T
1: Ta ← ∅, Na ← T
a ← max
2: ωmax
ja∈Na
a∗ ← M axCons(a, ωmax
3: a∗, ωmax
, r ← T , A ← {a∗}
4: d ← ωmax
a∗
5: for (θ = d; θ ≥
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20: end for
21: return Ta
a ← ∆fa(jaTa)
ω∗
a ← ja
j∗
else
a ← 0
ω∗
a ← none
j∗
end if
A, J ← M axCoor(a, j∗
if a ∈ A then
a}
Ta ← Ta ∪ {j∗
end if
Na ← Na\J
end while
if ∃ ja ∈ Na such that ∆fa(jaTa) ≥ θ then
)
a
a, ω∗
a)
a
The function M axCons and M axCoor in Algorithms
1 and 2 represent the communication and negotiation
among all robots. These two functions are described in
details as follows.
• M axCons is the Maximum Consensus function. Robot
a sends its locally best marginal value ωmax
together
with the corresponding robot id a to all its neighbour-
ing robots who have direct communication connections
with robot a. At the same time, robot a also receives the
same kind of information from its neighbouring robots.
After receiving the information from its neighbouring
robots, robot a sends the updated information to all its
neighbouring robots again until no information is up-
dated. In this case, robot a can pass the information for
its neighbouring robots who have no direct connection
between each other. All robots can make a consensus
and finally find the globally largest marginal value ωmax
.
It is proven that the Maximum Consensus function can
reach the convergence in finite time [12, 17, 21, 32]. Note
that this function is only utilised for one time by robot
a in the initialisation phase of the proposed algorithms.
a∗
• M axCoor is defined as the Maximum Coordination
function. Robot a sends the selected task id j∗
a and its
marginal value ω∗
a to all neighbouring robots and re-
ceives such information from its neighbouring robots.
Similar to the function M axCons, M axCoor also con-
ducts a consensus on the information of all robots. Then,
M axCoor returns the auxiliary robot set A and auxil-
iary task set J according to the following criteria: if there
is no conflict between robots (i.e., no task is selected by
more than one robot), then every robot and its qualified
task will be added into A and J, respectively; if there is
any conflict (i.e., more than one robot selects the same
task j∗
a), then the robot who has the largest marginal
value for the task j∗
a will be added into A, and the task
j∗
a will be added into J; if the input variable ω∗
a from
all robots are zeroes, then both A and J will be empty
which triggers the decrease of the threshold θ. With this
strategy, more than one task could be allocated during
one iteration, which can reduce the number of task as-
signment iterations and consensus steps.
In the original Lazy greedy [29], the updated marginal
value of the first element from the sorted remaining ele-
ment list is compared with the formerly second marginal
value. Instead, the updated first marginal value is com-
pared with the current threshold in the variant of lazy
greedy. The proposed new lazy greedy strategy is also
applicable for the general decreasing threshold greedy
algorithm [2].
Note that algorithm 2 could help robot a to find qualified
tasks in a more efficient way than Algorithm 1 in prac-
tice. In the initialisation phase of Algorithm 2, robot a
evaluates the tasks in Na and sorts them in descending
order according to their marginal values (Algorithm 2,
lines 2∼7). During the task assignment phase of Algo-
rithm 2, robot a checks whether the first marginal value
ωa1 from the sorted marginal value set Wa is no less
than the current threshold θ (Algorithm 2, line 13). If
ωa1 < θ, then there is no need for robot a to continue
searching for qualified tasks under the current thresh-
old because submodularity guarantees that the marginal
value of a task will never increase. Otherwise, robot a
reevaluates the first task ja1 from the sorted task list
Na given the current allocated task set Ta, and updates
the first marginal value ωa1 from the sorted marginal
value set Wa (Algorithm 2, line 14). If this updated ωa1
is still no less than θ, it implies that the first task ja1
from the sorted task list Na is a qualified task for robot
a. Robot a will use this task and its marginal value to
negotiate with other robots (Algorithm 2, lines 15∼23).
Otherwise, robot a re-sorts Wa and Na in decreasing or-
der (Algorithm 2, lines 24∼26). If robot a is unable to
find a qualified task, it still needs to communicate with
other robots and remove the tasks from Na that are se-
lected by other robots to prevent potential conflicts (Al-
gorithm 2, lines 28∼32).
a ← ωa1
)
a
Algorithm 2 decentralised LDTTA for Robot a
Input: fa : 2T → R≥0,T ,A,
Output: A set Ta ⊆ T
1: Ta ← ∅, Na ← T , Wa ← ∅
2: for ja ∈ Na do
ωaj ← ∆fa(jaTa)
Wa ← Wa ∪ {ωaj}
while A (cid:54)= ∅ do
while ωa1 ≥ θ do
ωa1 ← ∆fa(ja1Ta)
if ωa1 ≥ θ then
r d; θ ← θ(1 − )) do
a ← ωa1
ω∗
a ← ja1
j∗
A, J ← M axCoor(a, j∗
if a ∈ A then
a}
Ta ← Ta ∪ {j∗
3:
4:
5: end for
6: Sort Wa in descending order
7: Sort Na according to sorted Wa
8: ωmax
a∗ ← M axCons(a, ωmax
9: a∗, ωmax
10: d ← ωmax
, r ← T , A ← {a∗}
a∗
11: for (θ = d; θ ≥
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34: end for
35: return Ta
// ωa1 is the first element of the sorted marginal value
set Wa; ja1 is the first element of the sorted task set Na;
Wa(J) is the set of marginal values that are correspond-
ing to the tasks in J
end if
end while
a ← 0
ω∗
a ← none
j∗
A, J ← M axCoor(a, j∗
Na ← Na\J
Wa ← Wa\Wa(J)
end if
Na ← Na\J
Wa ← Wa\Wa(J)
Re-sort Wa and Na
else
end while
a, ω∗
a)
a, ω∗
a)
3.2 Analysis
Algorithms 1 and 2 have the same theoretical perfor-
mance in terms of both optimality bound and compu-
tational complexity. In each iteration, both algorithms
help robot a to find a qualified task whose marginal
value is no less than the current threshold θ. There-
fore, DTTA and LDTTA achieve the same theoretical
optimality bound. In LDTTA, the remaining tasks from
Na are sorted in descending order according to their
marginal values. Robot a is more likely to find a qual-
ified task from the front positions of Na, requiring less
number of objective function evaluations. This is the rea-
son why LDTTA could be more efficient than DTTA in
practice. However, in the worst case, DTTA and LDTTA
5
if S ∪ {u} /∈ I then
r d; θ ← θ(1 − )) do
Algorithm 3 equivalent Decreasing Threshold Greedy
subject to matroid constraints
Input: f : 2N → R≥0,N ,I, r,
Output: A set S ∈ I
1: S ← ∅, R ← N , Q ← OP T
2: d ← max
u∈N f (u)
3: for (θ = d; θ ≥
for u ∈ R do
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
c ← u
Sc ← S
S ← S ∪ {c}
Q ← Q ∪ {c}
Let Kc ⊆ Q\S be the smallest set such
Q ← Q\Kc
R ← R\{c}
if ∆f (uS) <
R ← R\{u}
R ← R\{u}
if ∆f (uS) ≥ θ then
that Q\Kc ∈ I
r d then
else
else
14:
15:
16:
17:
18:
19:
20:
21:
22:
23: end for
24: return S
end if
end for
end if
end if
need to evaluate all the remaining tasks to find a quali-
fied task. Therefore, DTTA and LDTTA have the same
computational complexity in theory.
For the convenience of analysing the theoretical per-
formance, DTTA and LDTTA are transformed into an
equivalent centralised version, i.e., Algorithm 3. We con-
sider the ground set as a set of task-agent pairs (N :=
T × A) and each task-agent pair as an element of the
ground set (uj,a := j × a ∀j ∈ T , a ∈ A). Then, only one
task-agent pair can be selected from the task-agent pairs
that are corresponding to the same specific task. All the
combinations of conflict-free task-agent pairs constitute
I, which is denoted as the collection of all independent
sets. Additionally, according to the definition of Task Al-
location, the objective function of one agent has no im-
pact on the objective function of another agent as long
as their selected task sets are disjoint. Therefore, the
task allocation problem can be considered as SWM. We
adapt the analysing strategy from [14] and transform
Algorithms 1 and 2 to Algorithm 3.
In Algorithm 3, Sc, Q and Kc have no impact on the
final solution S. They appear only for analysis: Sc is a set
containing the elements that have already been selected
before the new element c is added into S; Q is a set
that starts as the optimal solution OP T and changes
over time meanwhile keeps containing S; Kc is a set that
contains the element to be removed from Q in order to
keep Q independent.
The theoretical performance of the proposed task allo-
cation algorithms DTTA and LDTTA is summarised in
Theorem 1.
Theorem 1 Both DTTA and LDTTA achieve an opti-
mality bound of at least ( 1
submodular objective functions. The computational com-
plexity for each robot is O( r
number of all tasks, is the threshold decreasing param-
eter.
2 − ) for maximising monotone
), where r = T is the
ln r
The computational complexity can be easily proven as
follows.
PROOF. Assume that there are totally x number of
iterations, then
(1 − )x =
r
.
Solving the above equation yields
x =
ln r
ln 1
1−
≤ 1
ln
r
.
(3)
(4)
For each robot, there are at most r number of tasks to
be evaluated during each iteration. Therefore, the com-
putational complexity for each robot is O( r
ln r
).
In the following, the optimality bound of DTTA and
LDTTA is analysed through Algorithm 3 since they
have precisely the same theoretical approximation per-
formance.
Lemma 1 f (S) > 1
1+ f (Q).
PROOF. At the end of each iteration in Algorithm 3,
Q is always independent and contains S, i.e. S ⊆ Q ∈
I. The property of independent systems implies that
S ∪{q} ∈ I ∀q ∈ Q\S. At the termination of Algorithm
r d ∀q ∈ Q\S and f (S) ≥ d. Clearly,
3, ∆f (qS) <
Q\S ≤ r, thus(cid:88)
∆f (qS) <
d
r
(cid:88)
q∈Q\S
≤ · Q\S
r
≤ · f (S).
f (S)
q∈Q\S
6
Let Q\S = {q1, q2,··· , qQ\S}, then
f (Q) − f (S) =
≤
Q\S(cid:88)
Q\S(cid:88)
(cid:88)
i=1
i=1
∆f (qiS)
∆f (qS)
q∈Q\S
=
< · f (S).
∆f (qiS ∪ {q1,··· , qi−1})
(submodularity)
Rearranging the above inequation yields
{Kc}c∈S and {Kc\S}c∈S are disjoint. According to the
evolution of Q, the set Q can be rewritten as
Q = (OP T\∪c∈S Kc)∪S = (S∪OP T )\∪c∈S (Kc\S).
(9)
Denote S as {c1, c2,··· , cS}. It is clear that Sci ⊆ S ⊆
(S ∪ OP T )\ ∪c∈S (Kc\S). Using Eqn. (9), we have
f (Q) = f ((S ∪ OP T )\ ∪c∈S (Kc\S))
= f (S ∪ OP T )
= f (S ∪ OP T )
− ∆f (∪c∈S(Kc\S)(S ∪ OP T )\ ∪c∈S (Kc\S))
f (S) >
1
1 +
f (Q).
−
∆f ((Kci\S)(S ∪ OP T )\ ∪1≤j≤i (Kcj\S))
S(cid:88)
S(cid:88)
i=1
≥ f (OP T )
i=1
≥ f (OP T ) −
S(cid:88)
≥ f (OP T ) −(cid:88)
> f (OP T ) −(cid:88)
c∈S
i=1
Lemma 1 indicates that, at the termination of Algorithm
3, f (S) gets a close value to f (Q) if the value of is
small enough. This implies the elements in Q\S, whose
marginal values are less than
r d, have very limited con-
tribution to f (S). Likewise, if the marginal value of a
task for a robot is less than
r d, then we consider this
task negligible for this robot. This is the reason why we
choose
r d as the terminal threshold. The task allocation
algorithm terminates when all marginal values of the re-
maining tasks are less than
Lemma 2 f (Q) > f (OP T ) − 1
r d.
1− f (S).
PROOF. According to the property of matroid, Kc
contains at most one element. Therefore,
Kc\S ≤ 1.
(5)
In the iteration where the element c is selected, it implies
that the marginal value of c is no less than the current
threshold θ, i.e.
∆f (cSc) ≥ θ.
(6)
If an element q ∈ Kc\S was not selected before this
iteration, then it implies that
∆f (qSc) < θ/(1 − ).
(7)
Combining Eqns. (6) and (7), we get
∆f (cSc) > (1 − )∆f (qSc) ∀q ∈ Kc\S.
(8)
Additionally, in the iteration when Kc (cid:54)= ∅, the element
contained in Kc in this iteration will never appear in
the Kc again in other iterations. This implies that both
7
−
∆f ((Kci\S)(S ∪ OP T )\ ∪1≤j≤i (Kcj\S))
∆f ((Kci\S)Sci)
(cid:88)
∆f (qSc)
(cid:88)
q∈Kc\S
∆f (cSc)
(submodularity)
(submodularity)
c∈S
q∈Kc\S
1
1 −
1
1 −
· f (S).
· f (S)
= f (OP T ) − Kc\S ·
≥ f (OP T ) − 1
1 −
(Eqn. (8))
(Eqn. (5))
The proof of Theorem 1 is completed by combining
Lemma 1 and Lemma 2,
f (S) >
1 −
2 − 2 · f (OP T ) > (
1
2
− )f (OP T ).
According to the analytical results, the trade-off between
the approximation ratio and computational complexity
can be obtained by adjusting the threshold decreasing
parameter .
4 Numerical Simulations
This section verifies the proposed algorithms through
Monte-Carlo simulations. The simulation scenario is
modelled as a multi-target surveillance mission using a
group of heterogeneous UAVs.
4.1 Surveillance Scenario Modelling
For validation, we develop a simple model of the surveil-
lance scenario. In the model, it is assumed that there are
a set of heterogeneous tasks (T ) that are randomly lo-
cated on a L×L 2-D space. A set of heterogeneous UAVs
(A) equipped with different sensors is sent to carry out
these tasks automatically. The task allocation mission
aims to maximise the overall objective function value
while using as less running time and consensus steps as
possible. The constraint is that one task can only be al-
located to one UAV, but one UAV can carry out multiple
tasks.
The objective function of the surveillance mission for
UAV a is modelled as:
fa(Ta) =
majvjλτ (pj
a)
d
λσ(pj
a)
n
.
(10)
Ta(cid:88)
j=1
A(cid:88)
Ta(cid:88)
The overall objective function of the task allocation for
the surveillance mission can be obtained by combining
Eqns. (1) and (10):
f (T ,A) =
majvjλτ (pj
a)
d
λσ(pj
a)
n
(11)
a=1
j=1
where Ta ⊆ T is the task list containing all the tasks
that are selected by UAV a in sequence, pa is the path of
UAV a generated according to the order that the tasks
appear in Ta, pj
a represents the part of pa from the initial
position of UAV a to the position of task j, τ (pj
a) is the
length of the path fragment pj
a) is the number
of tasks within the path fragment pj
a.
a, and σ(pj
Four factors are considered in the objective function of
the surveillance mission.
• Task importance factor vj: Different tasks have differ-
ent values that are marked with an importance factor
vj ∈ (0, 1]. UAVs tend to give priority to carrying out
the tasks that are more valuable than others.
• Task-UAV fitness factor maj: Tasks with different
properties require different sensors to be detected effec-
tively. Therefore, the fitness factor maj ∈ (0, 1] reflects
the match fitness between the task j and UAV a.
• Distance discount factor λd: The mission should be
completed as soon as possible. With the help of the dis-
tance discount factor λd ∈ (0, 1], UAVs intend to carry
out the nearest tasks to them firstly, and the lengths of
all UAVs' paths should be balanced.
• Task number discount factor λn: It is usually risky
to allocate a large number of tasks to one UAV, but
8
only a few to others. The task number discount factor
λn ∈ (0, 1] helps to balance the numbers of allocated
tasks among all UAVs.
Note that, Eqn. (11) is proven to be monotone, non-
negative, and submodular in the appendix.
4.2 Simulation Results
The performances of the proposed algorithms are com-
pared with the benchmark decentralised task allocation
algorithms SGA and CBBA. We run 100 rounds of
Monte-Carlo simulations then get the mean values of
objective function value, running time, and consensus
steps, respectively. The running time is measured as
the number of objective function evaluations, which is
independent on the computer status. The running time
for each UAV approximately equals to the total running
time divided by the number of UAVs. The performances
of these algorithms are also compared through ratios
with SGA as a baseline.
In the simulations, assume that there are totally 200
tasks to be carried out by a fleet of varying numbers of
UAVs from 10 to 50 denoted as Na, and let L = 10km.
Set the importance factor of each task as a uniformly ran-
dom number vj ∈ [0.6, 1.0], and the match fitness factor
of each task-UAV pair as a uniformly random number
maj ∈ [0.5, 1.0]. Set the distance discount factor λd =
0.95, and the task number discount factor λn = 0.98,
respectively. In algorithms DTTA and LDTTA, set the
threshold decreasing parameter as = 0.05.
The performances of DTTA and LDTTA are compared
with benchmark task allocation algorithms SGA and
CBBA in Fig. 1. The advantage of CBBA is that it can
reduce consensus steps with the help of building task
bundles while achieving the same solution quality com-
pared with SGA. Note that, the number of consensus
steps required by SGA is equal to the number of tasks,
i.e., 200 in this case. As shown in Fig. 1 (c), (e), and (f),
CBBA consumes 45.9% of consensus steps of SGA when
there are 10 UAVs. The percentage goes up to 86.5% as
the number of UAVs increases to 50. As mentioned in [9],
agents need to continually rebuild their bundles to solve
the conflict during the task allocation process. The is-
sue with the bundle rebuilding is that when there are a
large number of tasks, evaluating the objective function
is quite time-consuming. This is the reason why CBBA
performs poorly in terms of running time, as shown in
Fig. 1 (b), (e), and (f). By contrast, DTTA and LDTTA
perform much better in terms of both consensus steps
and running time while obtaining almost the same func-
tion values compared with SGA and CBBA. As shown in
Fig. 1 (f), when Na = 50, LDTTA consumes only 1.2%
of running time and 14.0% of consensus steps with SGA
as a baseline. Fig. 1 (c) indicates that as the number of
UAVs increases, unlike CBBA, the numbers of consen-
sus steps for DTTA and LDTTA decrease because more
(a) Function Value
(b) Running Time
(c) Consensus Steps
(d) Running Time
(e) Ratio Comparison Na = 10
(f) Ratio Comparison Na = 50
Fig. 1. Performance comparison of the proposed algorithms and benchmark algorithms
tasks can be allocated within one consensus step when
there are more UAVs. This is an extremely beneficial
feature in the large-scale task allocation problems where
scalability is one of the key concerns.
According to Fig. 1 (d), (e), and (f), LDTTA consumes
much less running time and slightly fewer consensus
steps with slightly higher function values compared with
DTTA under the same . In LDTTA, the remaining
tasks are sorted in descending order according to their
marginal values. On average, LDTTA is more likely
to find a qualified task with a higher marginal value,
thereby achieving better solution quality than DTTA
does. Meanwhile, since UAVs and tasks are heteroge-
neous and located at different positions, UAVs are less
likely to have a conflict with each other using LDTTA
9
(a) Function Value
(b) Running Time
(c) Consensus Steps
(d) Ratio Comparison Na = 30
Fig. 2. Trade-off of function value, running time and consensus steps
than using DTTA. Thus, LDTTA consumes fewer con-
sensus steps than DTTA. Overall, LDTTA has the best
performance among all these algorithms.
Fig. 2 demonstrates the trade-off performance of
LDTTA where the threshold decreasing parameter is
set as 0.1, 0.2, and 0.3 respectively. The results show
that as increases, the function value, running time, and
consensus steps decrease, which verifies the trade-off.
Fig. 2 (d) compares the ratio of function value, running
time, and consensus steps with = 0.1 as a baseline. It
indicates that as the value of increases from 0.1 to 0.3,
the ratios of function value and running time decrease
slowly at a similar pace, while the ratio of consensus
steps drops rapidly.
5 Conclusions
This paper proposed two efficient decentralised al-
gorithms, i.e., Decreasing Threshold Task Allocation
(DTTA) and an upgraded version Lazy Decreasing
Threshold Task Allocation (LDTTA), for Multi-Robot
Systems (MRS) in large-scale task allocation problems.
DTTA and LDTTA enabled parallel allocation with the
help of a decreasing threshold hence releasing the com-
putational and communicating burden for MRS. The
performance of the proposed algorithms was analysed
theoretically and verified through Monte-Carlo simu-
lations of a multi-target surveillance mission using a
group of heterogeneous UAVs. Simulation results indi-
cated that the proposed algorithms consumed much less
running time and consensus steps while achieving al-
most the same objective function values compared with
the benchmark task allocation algorithms. The results
of the proposed algorithms are expected instrumental
in the large-scale physical applications where MRS usu-
ally suffer from the computational and communicating
burden.
One future work would be further analysing the theo-
retical performance of the proposed algorithms by com-
bining the curvature of the submodular objective func-
tion to provide a tighter optimality bound of the solu-
tion quality. Another future work would be verifying the
performance of the proposed algorithms through physi-
cal experiments of a surveillance mission in a large area
using multiple UAVs.
10
References
[1] Gustavo SC Avellar, Guilherme AS Pereira, Luciano CA
Pimenta, and Paulo Iscold. Multi-uav routing for area
coverage and remote sensing with minimum time. Sensors,
15(11):27783 -- 27803, 2015.
[2] Ashwinkumar Badanidiyuru and Jan Vondr´ak.
Fast
In
algorithms
Proceedings
annual ACM-SIAM
symposium on Discrete algorithms, pages 1497 -- 1514. SIAM,
2014.
for maximizing submodular
twenty-fifth
functions.
of
the
[3] Xiaoshan Bai, Weisheng Yan, Shuzhi Sam Ge, and Ming Cao.
An integrated multi-population genetic algorithm for multi-
vehicle task assignment in a drift field. Information Sciences,
453:227 -- 238, 2018.
[4] Maria-Florina Balcan. Learning submodular functions with
In AAMAS, page 3,
applications to multi-agent systems.
2015.
[5] Antonio Barrientos, Julian Colorado, Jaime del Cerro,
Alexander Martinez, Claudio Rossi, David Sanz, and Joao
Valente. Aerial remote sensing in agriculture: A practical
approach to area coverage and path planning for fleets of
mini aerial robots. Journal of Field Robotics, 28(5):667 -- 689,
2011.
[6] Hamid Reza Boveiri. A novel aco-based static task scheduling
approach for multiprocessor environments.
International
Journal of Computational Intelligence Systems, 9(5):800 --
811, 2016.
[7] Niv Buchbinder, Moran Feldman, and Roy Schwartz.
Comparing
trade-off in
submodular maximization. Mathematics of Operations
Research, 42(2):308 -- 329, 2016.
and oranges: Query
apples
[8] Jes´us Capitan, Luis Merino, and An´ıbal Ollero. Cooperative
decision-making
for multi-target
surveillance with multiples uavs. Journal of Intelligent &
Robotic Systems, 84(1-4):371 -- 386, 2016.
uncertainties
under
[9] Han-Lim Choi, Luc Brunet, and Jonathan P How. Consensus-
based decentralized auctions for robust task allocation. IEEE
transactions on robotics, 25(4):912 -- 926, 2009.
[10] Micah Corah and Nathan Michael. Distributed submodular
maximization on partition matroids for planning on large
sensor networks. In 2018 IEEE Conference on Decision and
Control (CDC), pages 6792 -- 6799. IEEE, 2018.
[11] Micah Corah and Nathan Michael. Distributed matroid-
constrained submodular maximization for multi-robot
exploration: Theory and practice. Autonomous Robots,
43(2):485 -- 501, 2019.
[12] Jorge Cort´es. Distributed algorithms for reaching consensus
on general functions. Automatica, 44(3):726 -- 737, 2008.
[13] Huanyu Ding and David Castan´on. Multi-agent discrete
search with limited visibility.
In 2017 IEEE 56th Annual
Conference on Decision and Control (CDC), pages 108 -- 113.
IEEE, 2017.
[14] Moran Feldman, Christopher Harshaw, and Amin Karbasi.
Greed is good: Near-optimal submodular maximization via
greedy optimization. arXiv preprint arXiv:1704.01652, 2017.
[15] Brian P Gerkey and Maja J Mataric. Sold!: Auction methods
for multirobot coordination. IEEE transactions on robotics
and automation, 18(5):758 -- 768, 2002.
[17] Silvia Giannini, Antonio Petitti, Donato Di Paola, and
Alessandro Rizzo. Asynchronous max-consensus protocol
with time delays: Convergence results and applications. IEEE
Transactions on Circuits and Systems I: Regular Papers,
63(2):256 -- 264, 2016.
[18] David Grimsman, Mohd Shabbir Ali, Joao P Hespanha, and
Jason R Marden. The impact of information in greedy
submodular maximization. IEEE Transactions on Control
of Network Systems, 2018.
[19] Jingjing Gu, Tao Su, Qiuhong Wang, Xiaojiang Du, and
Mohsen Guizani. Multiple moving targets surveillance
based on a cooperative network for multi-uav.
IEEE
Communications Magazine, 56(4):82 -- 89, 2018.
[20] Sorin Ilie and Costin Badica. Multi-agent approach to
distributed ant colony optimization. Science of Computer
Programming, 78(6):762 -- 774, 2013.
[21] Franck Iutzeler, Philippe Ciblat, and J´er´emie Jakubowicz.
Analysis of max-consensus algorithms in wireless channels.
IEEE Transactions on Signal Processing, 60(11):6103 -- 6107,
2012.
[22] Kelin Jose and Dilip Kumar Pratihar. Task allocation
and collision-free path planning of centralized multi-robots
system for
inspection using heuristic
methods. Robotics and Autonomous Systems, 80:34 -- 42, 2016.
industrial plant
[23] Andreas Krause and Daniel Golovin. Submodular function
maximization., 2014.
[24] Rajiv Ranjan Kumar, Pradeep Varakantham, and Akshat
Kumar. Decentralized planning in stochastic environments
with submodular rewards. In AAAI, pages 3021 -- 3028, 2017.
[25] Heba Kurdi, Jonathon How, and Guillermo Bautista. Bio-
inspired algorithm for task allocation in multi-uav search and
rescue missions. In AIAA Guidance, Navigation, and Control
Conference, page 1377, 2016.
[26] Dong-Hyun Lee. Resource-based task allocation for multi-
robot systems. Robotics and Autonomous Systems, 103:151 --
161, 2018.
[27] Pei Li and Haibin Duan. A potential game approach to
multiple uav cooperative search and surveillance. Aerospace
Science and Technology, 68:403 -- 415, 2017.
[28] G´abor Milics. Application of uavs in precision agriculture.
In International Climate Protection, pages 93 -- 97. Springer,
2019.
[29] Michel Minoux.
Accelerated greedy algorithms
maximizing submodular set functions.
techniques, pages 234 -- 243. Springer, 1978.
for
In Optimization
[30] Daniel Morgan, Giri P Subramanian, Soon-Jo Chung,
and Fred Y Hadaegh. Swarm assignment and trajectory
optimization using variable-swarm, distributed auction
assignment and sequential convex programming.
The
International Journal of Robotics Research, 35(10):1261 --
1285, 2016.
[31] George L Nemhauser, Laurence A Wolsey, and Marshall L
Fisher. An analysis of approximations for maximizing
submodular set functionsi. Mathematical programming,
14(1):265 -- 294, 1978.
[32] Reza Olfati-Saber and Richard 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] Bahman Gharesifard and Stephen L Smith. Distributed
submodular maximization with limited information. IEEE
transactions on control of network systems, 5(4):1635 -- 1645,
2017.
[33] Jakub Ondr´acek, Ondrej Vanek, and Michal Pechoucek.
Solving infrastructure monitoring problems with multiple
heterogeneous unmanned aerial vehicles. In Proceedings of
the 2015 International Conference on Autonomous Agents
11
and Multiagent Systems, pages 1597 -- 1605. International
Foundation for Autonomous Agents and Multiagent Systems,
2015.
[34] Michael Otte, Michael J Kuhlman, and Donald Sofge.
Auctions for multi-robot task allocation in communication
limited environments. Autonomous Robots, pages 1 -- 38, 2019.
[35] Parag C Pendharkar. An ant colony optimization heuristic
Journal of
for constrained task allocation problem.
Computational Science, 7:37 -- 47, 2015.
[36] Guannan Qu, Dave Brown, and Na Li. Distributed greedy
algorithm for satellite assignment problem with submodular
utility function. IFAC-PapersOnLine, 48(22):258 -- 263, 2015.
[37] Guannan Qu, Dave Brown, and Na Li. Distributed greedy
algorithm for multi-agent task assignment problem with
submodular utility functions. Automatica, 105:206 -- 215, 2019.
[38] Jurgen Scherer, Saeed Yahyanejad, Samira Hayat, Evsen
Yanmaz, Torsten Andre, Asif Khan, Vladimir Vukadinovic,
Christian Bettstetter, Hermann Hellwagner, and Bernhard
Rinner. An autonomous multi-uav system for search and
rescue. In Proceedings of the First Workshop on Micro Aerial
Vehicle Networks, Systems, and Applications for Civilian
Use, pages 33 -- 38. ACM, 2015.
[39] Pau Segui-Gasco, Hyo-Sang Shin, Antonios Tsourdos, and
VJ Segui.
Decentralised submodular multi-robot task
allocation. In Intelligent Robots and Systems (IROS), 2015
IEEE/RSJ International Conference on, pages 2829 -- 2834.
IEEE, 2015.
[40] Min-Guk Seo, Hyo-Sang Shin, and Antonios Tsourdos. Task
allocation in agricultural remote sensing applications using
submodular maximization algorithm. In 2018 37th Chinese
Control Conference (CCC), pages 6860 -- 6865. IEEE, 2018.
[41] Hyo-Sang Shin and Pau Segui-Gasco. Uav swarms: Decision-
making paradigms. Unmanned Aircraft Systems, page 397,
2016.
[42] Xinmiao Sun, Christos G Cassandras, and Xiangyu Meng.
Exploiting submodularity to quantify near-optimality in
multi-agent coverage problems. Automatica, 100:349 -- 359,
2019.
[43] Jan Vondr´ak. Optimal approximation for the submodular
welfare problem in the value oracle model. In Proceedings of
the fortieth annual ACM symposium on Theory of computing,
pages 67 -- 74. ACM, 2008.
[44] Ryan K Williams, Andrea Gasparri, and Giovanni Ulivi.
Decentralized matroid optimization for topology constraints
in multi-robot allocation problems.
In Robotics and
Automation (ICRA), 2017 IEEE International Conference
on, pages 293 -- 300. IEEE, 2017.
[45] Lifeng Zhou and Pratap Tokekar.
Sensor assignment
algorithms to improve observability while tracking targets.
IEEE Transactions on Robotics, 2019.
[46] WANG Zhu, LIU Li, LONG Teng, and WEN Yonglu. Multi-
uav reconnaissance task allocation for heterogeneous targets
using an opposition-based genetic algorithm with double-
chromosome encoding. Chinese Journal of Aeronautics,
31(2):339 -- 350, 2018.
A Analysis of the Objective Function
In the appendix, we prove that the objective function
Eqn. (11) is monotonically increasing, non-negative, and
submodular. We start the proof from analysing Eqn.
(10).
For a random UAV a, assume that there are two task
selection sets T1 and T2 satisfying T1 ⊆ T2 ⊆ T . A new
task j ∈ T \T2 is going to be considered by this UAV.
pj
1 represents the path from initial position of UAV a to
the task j along the task set T1 and pj
2 is defined in the
same way with T2.
Lemma 3 The objective function Eqn. (10) is non-
negative and increasing.
PROOF. By definition, if Ta = ∅, then fa(Ta) = 0 and
these four factors maj, vj, λd, λn ∈ (0, 1] are positive.
Then the marginal value of task j is always positive, i.e.
majvjλτ (pj
a)
d
λσ(pj
a)
n
> 0.
Eqn. (10) is the sum of all marginal values, hence it is
non-negative and increasing.
Lemma 4 The objective function Eqn. (10) is submod-
ular.
PROOF. Since T1 ⊆ T2, the length of the path pj
no greater than that of the path pj
tasks along the path pj
path pj
1 is
2, and the number of
1 is no greater than that of the
2, i.e.
0 < τ (pj
1) ≤ τ (pj
2),
0 < σ(pj
1) ≤ σ(pj
2).
(A.1)
The marginal value of task j given T1 is
∆fa(jT1) = majvjλτ (pj
1)
d
λσ(pj
1)
n
.
The marginal value of task j given T2 is
∆fa(jT2) = majvjλτ (pj
2)
d
λσ(pj
2)
n
.
According to Eqn. (A.1) and since λd, λn ∈ (0, 1], we
have
∆fa(jT1) ≥ ∆fa(jT2).
Therefore, Eqn. (10) is submodular.
The task sets for all UAVs are disjoint since one task can
be allocated to no more than one UAV. According to the
property of submodularity, the sum of submodular func-
tions is submodular. Therefore, the objective function
Eqn. (11) is monotonically increasing, non-negative, and
submodular.
12
|
1203.6035 | 1 | 1203 | 2012-03-27T17:31:24 | A Multi-Agent Prediction Market based on Partially Observable Stochastic Game | [
"cs.MA"
] | We present a novel, game theoretic representation of a multi-agent prediction market using a partially observable stochastic game with information (POSGI). We then describe a correlated equilibrium (CE)-based solution strategy for this game which enables each agent to dynamically calculate the prices at which it should trade a security in the prediction market. We have extended our results to risk averse traders and shown that a Pareto optimal correlated equilibrium strategy can be used to incentively truthful revelations from risk averse agents. Simulation results comparing our CE strategy with five other strategies commonly used in similar markets, with both risk neutral and risk averse agents, show that the CE strategy improves price predictions and provides higher utilities to the agents as compared to other existing strategies. | cs.MA | cs |
A Multi-Agent Prediction Market based on
Partially Observable Stochastic Game
Janyl Jumadinova (1), Prithviraj Dasgupta (1)
((1) University of Nebraska at Omaha,
Computer Science Department, Omaha NE)
July 5, 2018
We present a novel, game theoretic representation of a multi-agent prediction
market using a partially observable stochastic game with information (POSGI).
We then describe a correlated equilibrium (CE)-based solution strategy for this
game which enables each agent to dynamically calculate the prices at which it
should trade a security in the prediction market. We have extended our results
to risk averse traders and shown that a Pareto optimal correlated equilibrium
strategy can be used to incentively truthful revelations from risk averse agents.
Simulation results comparing our CE strategy with five other strategies com-
monly used in similar markets, with both risk neutral and risk averse agents,
show that the CE strategy improves price predictions and provides higher util-
ities to the agents as compared to other existing strategies.
1
Introduction
Forecasting the outcome of events that will happen in the future is a frequently
indulged and important task for humans. It is encountered in various domains
such as forecasting the outcome of geo-political events, betting on the outcome
of sports events, forecasting the prices of financial instruments such as stocks,
and casual predictions of entertainment events. Despite the ubiquity of such
forecasts, predicting the outcome of future events is a challenging task for hu-
mans or even computers - it requires extremely complex calculations involving
a reasonable amount of domain knowledge, significant amounts of information
processing and accurate reasoning. Recently, a market-based paradigm called
prediction markets has shown ample success to solve this problem by using the
aggregated 'wisdom of the crowds' to predict the outcome of future events. This
is evidenced from the successful predictions of actual events done by prediction
markets run by the Iowa Electronic Marketplace(IDEM), Tradesports, Holly-
wood Stock Exchange, the Gates-Hillman market [26], and by companies such
as Hewlett Packard [29], Google [7] and Yahoo's Yootles [34]. A prediction mar-
ket for a real-life event (e.g., "Will Obama win the 2008 Democratic Presidential
nomination?") is run for several days before the event happens. The event has
a binary outcome (yes/no or 1/0). On each day, humans, called traders, that
are interested in the outcome of the event express their belief on the possible
outcome of the event using the available information related to the event. The
1
information available to the different traders is asymmetric, meaning that dif-
ferent traders can possess bits and pieces of the whole information related to the
event. The belief values of the traders about the outcome of the event are ex-
pressed as probabilities. A special entity called the market maker aggregates the
probabilities from all the traders into a single probability value that represents
the possible outcome of the event. The main idea behind the prediction mar-
ket paradigm is that the collective, aggregated opinions of humans on a future
event represents the probability of occurrence of the event more accurately than
corresponding surveys and opinion polls. Despite their overwhelming success,
many aspects of prediction markets such as a formal representation of the mar-
ket model, the strategic behavior of the market's participants and the impact of
information from external sources on their decision making have not been ana-
lyzed extensively for a better understanding. We attempt to address this deficit
in this paper by developing a game theoretic representation of the traders' inter-
action and determining their strategic behavior using the equilibrium outcome
of the game. We have developed a correlated equilibrium (CE)-based solution
strategy for a partially observable stochastic game representation of the predic-
tion market. We have also empirically compared our CE strategy with five other
strategies commonly used in similar markets, with both risk neutral and risk
averse agents, and showed that the CE strategy improves price predictions and
provides higher utilities to the agents as compared to other existing strategies.
2 Related Work
Prediction markets were started in 1988 at the Iowa Electronic Marketplace
[19] to investigate whether betting on the outcome of gee-political events (e.g.
outcome of presidential elections, possible outcome of international political or
military crises, etc.) using real money could elicit more accurate information
about the event's outcome than regular polls. Following the success of prediction
markets in eliciting information about events' outcomes, several other predic-
tion markets have been started that trade on events using either real or virtual
money. Prediction markets have been used in various scenarios such as pre-
dicting the outcome of gee-political events such as U.S. presidential elections,
determining the outcome of sporting events, predicting the box office perfor-
mance of Hollywood movies, etc. Companies such as Google, Microsoft, Yahoo
and Best Buy have all used prediction markets internally to tap the collective in-
telligence of people within their organizations. The seminal work on prediction
market analysis [13, 32, 33] has shown that the mean belief values of individual
traders about the outcome of a future event corresponds to the event's market
price. Since then researchers have studied prediction markets from different per-
spectives. Some researchers have studied traders' behavior by modeling their
interactions within a game theoretic framework. For example, in [3, 11] the
authors have used a Shapley-Shubik game that involves behavioral assumptions
on the agents such as myopic behavior and truthful revelations and theoretically
analyzed the aggregation function and convergence in prediction markets. Chen
et. al. [4] characterized the uncertainty of market participants' private informa-
tion by incorporating aggregate uncertainty in their market model. However,
both these models consider traders that are risk-neutral, myopic and truthful.
Subsequently, in [6] the authors have relaxed some of these assumptions within
2
a Bayesian game setting and investigated the conditions under which the play-
ers reveal their beliefs truthfully. Dimitrov and Sami [9] have also studied the
effect of non-myopic revelations by trading agents in a prediction market and
concluded that myopic strategies are almost never optimal in the market with
the non-myopic traders and there is a need for discounting. Other researchers
have focused on designing rules that a market maker can use to combine the
opinions (beliefs) from different traders. Hanson [15] developed a market scor-
ing rule that is used to reward traders for making and improving a prediction
about the outcome of an event. He further showed how any proper scoring rule
can serve as an automated market maker. Das [8] studied the effect of spe-
cialized agents called market-makers which behave as intermediaries to absorb
price shocks in the market. Das empirically studied different market-making
strategies and concludes that a heuristic strategy that adds a random value to
zero-profit market-makers improves the profits in the markets.
The main contribution of our paper, while building on these previous direc-
tions, is to use a partially observable stochastic game (POSG) [14] that can be
used by each agent to reason about its actions. Within this POSG model, we
calculate the correlated equilibrium strategy for each agent using the aggregated
price from the market maker as a recommendation signal. We have also con-
sidered the risk preferences of trading agents in prediction markets and shown
that a Pareto optimal correlated equilibrium solution can incentively truthful
revelation from risk averse agents. We have compared the POSG/correlated
equilibrium based pricing strategy with five different pricing strategies used in
similar markets with pricing data obtained from real prediction prediction mar-
ket events. Our results show that the agents using the correlated equilibrium
strategy profile are able to predict prices that are closer to the actual prices that
occurred in real markets and these traders also obtain 35 − 127% higher profits.
3 Preliminaries
Prediction Market. Our prediction market consists of N traders, with each
trader being represented by a software trading agent that performs actions on
behalf of the human trader. The market also has a set of future events whose
outcome has not yet been determined. The outcome of each event is considered
as a binary variable with the outcome being 1 if the event happens and the
outcome being 0 if it does not. Each outcome has a security associated with
it. A security is a contract that yields payments based on the outcome of an
uncertain future event. Securities can be purchased or sold by trading agents
at any time during the lifetime of the security's event. A single event can have
multiple securities associated with it. Trading agents can purchase or sell one
or more of the securities for each event at a time. A security expires when the
event e associated with it happens at the end of the event's duration. At this
point the outcome of the event has just been determined and all trading agents
are notified of the event's outcome. The trading agents that had purchased the
security during the lifetime of the event then get paid $1 if the event happens
with an outcome of 1, or, they do not get paid anything and lose the money
they had spent on buying the security if the event happens with an outcome
of 0. On each day, a trading agent makes a decision of whether to buy some
securities related to ongoing events in the market, or whether to sell or hold
3
some securities it has already purchased.
Figure 1: Essential operations performed by a trading agent in a prediction
market and the portion of the prediction market where the POSG representation
is applied.
Market Maker. Trading in a prediction market is similar to the continuous
double auction (CDA) protocol. However, using the CDA directly as the trading
protocol for a prediction market leads to some problems such as the thin market
problem (traders not finding a trading partner immediately) and traders poten-
tially losing profit because of revealing their willingness to trade beforehand to
other traders. To address these problems, prediction markets use a trading in-
termediary called the market maker. The market maker aggregates the buying
and selling prices of a security reported in the trades from the trading agents
(they can be identical) and 'posts' these prices in the market. Trading agents
interact with the market maker to buy and sell securities, so that they do not
have to wait for another trading agent to arrive before they can trade. A market
maker uses a market scoring rule (MSR) to calculate the aggregated price of a
security. Recently, there has been considerable interest in analyzing the MSRs
[5, 15, 27] and an MSR called the logarithmic MSR (LMSR) has been shown to
guarantee truthful revelation of beliefs by the trading agents [15]. LMSR allows
a security's price, and payoffs to agents buying/selling the security, to be ex-
pressed in terms of its purchased or outstanding quantity. Therefore, the 'state'
of the securities in the market can be captured only using their purchased quan-
tities. The market-maker in our prediction market uses LMSR to update the
market price and to calculate the payoffs to trading agents. We briefly describe
the basic mechanism under LMSR. Let Ξ be the set of securities in a prediction
market and q = (q1, q2...qΞ) denote the vector specifying the number of units
of each security held by the different trading agents at time t. The LMSR first
calculates a cost function to reflect the total money wagered in the prediction
4
market by the trading agents as: C(q) = b·ln(PΞ
j=1 eqj /b). It then calculates the
aggregated market price π for the security ξ ∈ Ξ as: πξ = e(qξ/b)/ PΞ
j=1(e(qj /b))
[5] 1. Trading agents inform the quantity of a security they wish to buy or
sell to the market maker. If a trading agent purchases δ units of a security, the
market maker determines the payment the agent has to make as C(q +δ)−C(q).
Correspondingly, if the agent sells δ quantity of the security, it receives a payoff
of C(q) − C(q − δ) from the market maker. Figure 1 shows the operations per-
formed by a trader (trading agent) in a prediction market and the portions of
the prediction market that are affected by the POSG representation described
in the next section.
4 Partially Observable Stochastic Games for Trad-
ing Agent Interaction
For simplicity of explanation, we consider a prediction market where a single
security is being traded over a certain duration. This duration is divided into
trading periods, with each trading period corresponding to a day in a real pre-
diction market. The 'state' of the market is expressed as the quantity of the
purchased units of the security in the market. At the end of each trading period,
each trading agent receives information about the state of the market from the
market maker. With this prior information, the task of a trading agent is to
determine a suitable quantity to trade for the next trading period, so that its
utility is maximized. In this scenario, the environment of the agent is partially
observable because other agents' actions and payoffs are not known directly, but
available through their aggregated beliefs. Agents interact with each other in
stages (trading periods), and in each stage the state of the market is determined
stochastically based on the actions of the agents and the previous state. This
scenario directly corresponds to the setting of a partially observable stochastic
game [12, 14]. A POSG model offers several attractive features such as struc-
tured behavior by the agents by using best response strategies, stability of the
outcome based on equilibrium concepts, lookahead capability of the agent to plan
their actions based on future expected outcomes, ability to represent the tem-
poral characteristics of the interactions between the agents, and, enabling all
computations locally on the agents so that the system is robust and scalable.
Previous research has shown that information related parameters in a predic-
tion market such as information availability, information reliability, information
penetration, etc., have a considerable effect on the belief (price) estimation by
trading agents. Based on these findings, we posit that a component to model
the impact of information related to an event should be added to the POSG
framework. With this feature in mind, we propose an interaction model called
a partially observable stochastic game with information (POSGI) for capturing
the strategic decision making by trading agents. A POSGI is defined as:
Γ = (N, S, (Ai)i∈N , (Ri)i∈N , T, (Oi)i∈N , Ω, (Ii)i∈N ), where N is a finite set of
agents, S is a finite, non-empty set of states - each state corresponding to certain
1Parameter b (determined by the market maker) controls the monetary risk of the market
maker as well as the quantity of shares that trading agents can trade at or near the current
price without causing massive price swings. Larger values for b allows trading agents to trade
more frequently but also increases the market maker's chances to lose money.
5
quantity of the security being held (purchased) by the trading agents. Ai is a
finite non-empty action space of agent i s.t. ak = (a1,k, ..., aN ,k) is the joint ac-
tion of the agents and ai,k is the action that agent i takes in state k. In terms of
the prediction market, a trading agent's action corresponds to certain quantity
of security it buys or sells, while the joint action corresponds to changing the
purchased quantity for a security and taking the market to a new state. Ri,k is
the reward or payoff for agent i in state k which is calculated using the LMSR
market maker. T : T (s, a, s′) = P (s′s, a) is the transition probability of moving
from state s to state s′ after joint action a has been performed by the agents.
Oi is a finite non-empty set of observations for agent i that consists of the mar-
ket price and the information signal, and oi,k ∈ Oi is the observation agent i
receives in state k. Ω : Ω(sk, Ii,k, oi,k) = P (oi,ksk, Ii,k) is the observation prob-
ability for agent i of receiving observation oi,k in state sk when the information
signal is Ii,k. Finally, Ii is the information set received by agent i for an event
Ii = Sk Ii,k where Ii,k ∈ {−1, 0, +1} is the information received by agent i in
state k. The complete information arriving to the market I = Si∈N Ii is tempo-
rally distributed over the duration of the event, and, following the information
arrival patterns observed in stock markets [23], we assume that new information
arrives following a Poisson distribution. Information that improves the proba-
bility of the positive outcome of the event is considered positive(Ii,k = +1) and
vice-versa, while information that does not affect the probability is considered
to have no effect (Ii,k = 0). For example, for a security related to the event
"Obama wins 2008 presidential elections", information about Oprah Winfrey
endorsing Obama would be considered high impact positive information and
information about Obama losing the New Hampshire Primary would be consid-
ered negative information.
Based on the POSGI formulation of the prediction market, the interaction of
an agent with the environment (prediction market) and the information source
can be represented by the transition diagram shown in Figure 22. The environ-
ment (prediction market) goes through a set of states S = {s1, ..., sH } : S ∈ S,
where H is the duration of the event in the prediction market and sh represents
the state of the market during trading period h. This state of the market is not
visible to any agent. Instead, each agent i has its own internal belief state Bi,h
corresponding to its belief about the actual state sh. Bi,h gives a probability
distribution over the set of states S, where Bi,h = (b1,h, ..., bS,h). Consider
trading period h − 1 when the agents perform the joint action ah−1. Because of
this joint action of the agents the environment stochastically changes to a new
state sh, defined by the state transition function T (sh−1, ah−1, sh). There is also
an external information state, Ih, that transitions to the next state Ih+1 given
by the Poisson distribution [23] from which the information signal is sampled.
The agent i doesn't directly see the environment state, but instead receives an
observation oi,Sh = (πsh , Ii,sh ), that includes the market price πsh correspond-
ing to the state sh as informed by the market maker, and the information signal
Ii,sh . The agent i then uses a belief update function to update its beliefs. Fi-
nally, agent i selects an action using an action selection strategy and receives a
reward Ri,sh .
2We have only shown one agent i to keep the diagram legible, but the same representation
is valid for every agent in the prediction market. The dotted lines represent that the reward
and environment state is determined by the joint action of all agents.
6
the length of the bars indicate the strength of the information's impact
event
announced
high impact information
low impact information
event
ended
Hidden Environment (Prediction Market)
T(s2,a2,s3)
s2
s3
. . .
T(s1,a1,s2)
sH
s1
T(s1,a1,s2)
prediction
market's state
Ri,1
reward for
agent i
ai,1
action by
agent i
Ri,2
ai,2
Ri,3
ai,3
Ri,H
ai,H
b(b1,o1,a1)
Bi,1
belief state
oi,1
observation
of agent i
I1
information state
Bi,2
oi,2
I2
b(b2,o2,a2)
Bi,3
b(bH+1,oH+1,aH+1)
. . .
Bi,H
Agent i
oi,3
I3
oi,H
Information
. . .
IH
Transitions using Poisson distribution
Figure 2: An agent interactions with the hidden environment (prediction mar-
ket) and an external information source.
4.1 Trading agent belief update function
Recall from Section 4 that a belief state of a trading agent is a probability
vector that gives a distribution over the set of states S in the prediction market,
i.e. Bi,h = (b1,h, ..., bS,h). A trading agent uses its belief update function
b : ℜS × Ai × Oi → ℜS to update its belief state based on its past action
ai,h−1, past belief state Bi,h−1 and the observation oi,Sh . The calculation of
the belief update function for each element of the belief state, bs′,h, s′ ∈ S, is
described below:
bs′,h = P (s′ai,h−1, oi) =
P (s′, ai,h−1, oi, )
P (ai,h−1, oi)
=
P (ois′, ai,h−1) · P (s′, ai,h−1)
P (oiai,h−1)P (ai,h−1)
(1)
7
Because ai,h−1 is conditionally independent given s′ and oi is conditionally in-
dependent given ai,h−1, we can rewrite Equation 1 as:
bs′,h =
P (ois′) · P (s′, ai,h−1)
P (oi)P (ai,h−1)
=
Pι∈I P (ι)P (ois′, ι) Ps∈S P (s)P (s′s, ai,h−1)P (ai,h−1)
P (oi)P (ai,h−1)
= Pι∈I P (ι)P (ois′, ι) Ps∈S P (s)P (s′s, ai,h−1)
P (oi)
= Pι∈I P (ι)Ω(s′, ι, oi) Ps∈S T (s, ai,h−1, s′)bs,h−1
P (oi)
(2)
All the terms in the r.h.s. of the Equation 2 can be calculated by an agent:
P (ι) is the probability of receiving information signal ι and is available from
the Poisson distribution [23] for the information arrival, Ω(s′, ι, oi) is the prob-
ability of receiving observation oi in state s′ when the information signal is ι,
T (s, ai,h−1, s′) is the probability that the state s transitions to state s′ after
agent i takes action ai,h−1, bs,h−1 is the past belief of agent i about state s,
P (oi) is the probability of receiving observation oi, which is constant and can
be viewed as a normalizing constant.
...
q+4
+1,+1
+1,0
0,+1
...
q+3
+1,+1
0, 0
+1,-1
-1,+1
q+2
+1,+1
-1,-1
-1,0
0,-1
+1,0
0,+1
-1,0
0,-1
+1,0
0,+1
q+1
0,0
-1,+1
+1,-1
0, 0
+1,-1
-1,+1
q
-1,-1
+1,+1
-1,-1
q-4
...
q-2
0,0
-1,+1
+1,-1
-1,-1
+1,+1
0,+1
+1,0 -1,0
0,-1
-1,0
0,-1
0,+1
+1,0
q-1
-1,-1
0,0
-1,+1
+1,-1
q-3
...
Figure 3: Finite state automata of the environment represented by the number
of outstanding units of the security, q, in the prediction market.
4.2 Trading agent action selection strategy
The objective of a trading agent in a prediction market is to select an action
at each stage so that the expected reward that it receives is maximized. To
understand this action selection process, we consider the decision problem facing
each trading agent. Consider two agents whose available actions during each
time step are to buy (=+1) or sell(=-1) only one unit of the security or not
8
Player 2
Sell
Hold
Buy
Sell
C(q) - C(q-1),
C(q) - C(q-1)
C(q) - C(q-1),
0
C(q) - C(q-1),
C(q+1) - C(q)
Player 1
Hold
0,
C(q) - C(q-1)
0,
0
0,
C(q+1) - C(q)
Buy
C(q+1) - C(q),
C(q) - C(q-1)
C(q+1) - C(q),
0
C(q+1) - C(q),
C(q+1) - C(q)
Figure 4: Two player normal form game in a state for which the number of
outstanding shares is q.
do anything (=0) by holding the security. Let the market state be denoted by
q, the number of purchased units of the security. Based on the set of actions
available to each agent, the state can transition to one of the following states
q + 2 (both agents buy), q + 1 (only one agent buys), q (both agents hold, or, one
agent buys while the other agent sells, resulting in no transition), q + 1 (only
agent sells), and q + 2 (both agents sell), as shown in Figure 3. We can expand
this state space further by adding more states and transitions, but the number
of states remains finite because the set of states S of the POSGI is finite. Also,
since the number of units of a security is finite, the number of transitions from a
state is guaranteed to be bounded. We can construct a normal form game from
this representation to capture the decision problem for each agent. The payoff
matrix for this 2-player game at state q is shown in Figure 4. From this payoff
matrix, we can use game-theoretic analysis to determine equilibrium strategies
for each trading agent and find the actions that maximize its expected reward,
as described below.
4.3 Correlated Equilibrium (CE) calculation
In the POSGI, the aggregated price information received by a trading agent
from the market maker can be treated as a recommendation signal for selecting
the agent's strategy. This situation lends itself to a correlated equilibrium (CE)
[1] where a trusted external agent privately recommends a strategy to play to
each player. A CE is more preferred to the Nash or Bayesian Nash equilibrium
because it can lead to improved payoffs, and it can be calculated using a linear
program in time polynomial in the number of agents and number of strategies.
Each agent i has a finite set of strategy profiles, Φi defined over its ac-
tion space Ai. The joint strategy space is given by Φ = QN
i=1 Φi and let
Φ−i = Qj6=i Φj. Let φ ∈ Φ denote a strategy profile and φi denote player
i's component in φ. The utility received by each agent i corresponds to its
payoff or reward and is given by the agent i's utility function ui(φ) = Ri
(for legibility we have dropped state k, but the same calculation applies at
9
every state). A correlated equilibrium is a distribution p on Φ such that for
all agents i and all strategies φi, φ′
i if all agents follow a strategy profile φ
that recommends player i to choose strategy φi, agent i has no incentive to
play another strategy φ′
i instead. This implies that the following expression
i ∈ Φi and where
holds: Pφ−i∈Φ−i
ui(φ′
i while
all the other agents keep their strategies fixed at φ−i and p(φ) is the probability
of realizing a given strategy profile φ.
i, φ−i) is the utility that agent i gets when it changes its strategy to φ′
i, φ−i)) ≥ 0, ∀i ∈ N , ∀φi, φ′
p(φ)(ui(φ) − ui(φ′
Theorem 1. A correlated equilibrium exists in our POSGI-based prediction
market representation at each stage (trading period).
Proof. At each stage in our prediction market, we can specify the correlated
equilibrium by means of linear constraints as given below:
X
φ−i∈Φ−i
p(φ)(ui(φ) − ui(φ′
i, φ−i)) ≥ 0, ∀i ∈ N, ∀φi, φ′
i ∈ Φi
p(φ) = 1,
X
φ∈Φ
p(φ) ≥ 0
(3)
(4)
(5)
Equation 3 states that when agent i is recommended to select strategy φi, it
must get no less utility from selecting strategy φi as it would from selecting any
other strategy φ′
i. Constraints 4 and 5 guarantee that p is a valid probability
distribution. We can rewrite the linear program specification of the correlated
equilibrium above by adding an objective function to it.
max X
φ∈Φ
p(φ), or min − X
φ∈Φ
p(φ) s.t.
X
p(φ)(ui(φ) − ui(φ′
i, φ−i)) ≥ 0,
φ∈Φ,φ−i∈Φ−i
φ ≥ 0
(6)
(7)
(8)
Program 7 is either trivial with a maximum of 0 or unbounded. Hoense-
laar [16] proved that there is a correlated equilibrium if and only if Program
7 is unbounded. To prove the unboundedness we consider the dual problem of
Equation 7 given in Equation 10.
max 0, s.t.
X
p(φ)[(ui(φ) − ui(φ′
i, φ−i)]T ≤ −1
φ∈Φ,φ−i∈Φ−i
φ ≥ 0
(9)
(10)
(11)
where for every p(φ) there is p(φ) such that p(φ)[(ui(φ) − ui(φ′
i, φ−i)]p(φ) = 0.
Also in [28], the authors showed that the problem given in Equation 10
is always infeasible. From operations research we know that when the dual
problem is infeasible the primal problem is feasible and unbounded. This means
that the primal problem from Equation 7 is always unbounded. We can then
conclude that there is at least one correlated equilibrium in every trading period
of the prediction market.
10
The CE calculation algorithm used by the trading agents in our agent-based
prediction market is shown in Algorithm 1. The calculation of the matrix values
of the U matrix must be done once for each of the N agents. Using the ellipsoid
algorithm [28] the computation of the utility difference ui(φ) − ui(φi, φ−i) for
each agent i can be done in Φi2 time. Therefore, the time complexity of the
CECalc algorithm during each trading period comes to N × Φi2.
CECalc(D, Φ)
Input: D, Φ //D is the duration of the market, Φ is the set of strategies
Output: p //correlated equilibrium
foreach t ← 0 to D do
//do this in each trading period
Let U be the matrix consisting of the values of (ui(φ) − ui(φi, φ−i)),
∀i ∈ N, φ ∈ Φ, φ−i ∈ Φi
p′
t ← getDualDistribution(Φ, U );
pt ← solve for pt s.t. ptU T p′
t = 0;
return pt;
end
GetDualDistribution(Φ, U )
Input: Φ, U
Output: ∆
l = 0;
p′
l ∈ [0, 1];
∆ = {};
while U T · p′
l ≤ −1 is feasible do
∆ = ∆ + p′
l;
l+1 = pl + ǫN ; //increase all elements of p′ by some small amounts
p′
from vector ǫN
l + +;
end
return ∆;
Algorithm 1: Correlated Equilibrium Algorithm
4.4 Correlated Equilibrium with Trading Agents' Risk Pref-
erences
Incorporating the risk preferences of the trading agents is an important factor
in prediction markets. For example, the erroneous result related to the non-
correlation between the trader beliefs and market prices in a prediction market
in [24] was because the risk preferences of the traders were not accounted for,
as noted in [13]. This problem is particularly relevant for risk averse traders
because the beliefs(prices) and risk preferences of traders have been reported
to be directly correlated [10, 20]. In this section, we examine CE that provides
truthful revelation incentives in the prediction market where the agent popula-
tion has different risk preferences. The risk preference of an agent i is modeled
through a utility function called the constant relative risk aversion (CRRA).
We use CRRA utility function to model risk averse agents because it allows to
model the effect of different levels of risk aversion and it has been shown to be a
11
better model than alternative families of risk modeling utilities [31]. It has been
widely used for modeling risk aversion in various domains including economic
domain [17], psychology [21] and in the health domain [2]. The CRRA utility
function is given below:
ui(Ri) =
i
R1−θi
1 − θi
, if θi 6= 1
= ln(Ri), if θi = 1
Here, −1 < θi < 1 is called the risk preference factor of agent i and Ri is the
payoff or reward to agent i calculated using the LMSR [5]. When Ri < 0, we
may get a utility in the form of a complex number. In that case, we convert
the complex number to a real number by calling an existing function that uses
magnitude and the angle of the complex number for conversion. For a risk
neutral agent, θi = 0, which makes ui = Ri, as we have assumed for our CE
analysis of the risk neutral agent in Section 4.3. For risk averse agents, θi > 0.
This makes the CRRA utility function concave. But the concave structure of
the utility ui for a risk averse agent does not affect the existence of at least one
correlated equilibrium because the unboundedness of Equation 7 is not affected
by the concave structure of ui. However, the equilibrium obtained with the
risk averse utility function can be different from the one obtained with the risk
neutral utility function - because the best response p(φ) to p(φ−i) calculated
with the risk neutral utility function might not remain the best response when
the agents are risk averse and use the risk averse utility function. To find
a correlated equilibrium p(φ) in the market with risk averse agents, we first
characterize the set of all Pareto optimal strategy profiles. A strategy profile
φP is Pareto optimal if there does not exist another strategy profile φ′ such that
ui(φ′) ≥ ui(φP ) ∀i ∈ N with at least one inequality strict. In other words, a
Pareto optimal strategy profile is one such that no trader could be made better
off without making someone else worse off. A Pareto optimal strategy profile
can be found by maximizing weighted utilities
maxφ
N
X
i=1
λiui(φ) for some λi
(12)
Setting λi = 1 for all i ∈ N gives a utilitarian social welfare function. The max-
imization problem in Equation 12 can be solved using the Lagrangian method.
We get the following system of N equations:
N
X
i=1
λi
ui(φ)
φi
= 0 , ∀j = 1, ..., N
(13)
that must hold at φP . Each of these equations is obtained by taking a partial
derivative of the respective agent's weighted utility with respect to respective
agent's strategy profile, thus solving the maximization problem given in Equa-
tion 12. By solving the system of equations 13 we get the set of Pareto optimal
strategy profiles. To determine if p(φ) is a correlated equilibrium we simply
check whether it satisfies correlated equilibrium constraints for Pareto optimal
strategy profile given below:
p(φ)(ui(φ) − ui(φ′
Pφ−i∈Φ−i
i, φ−i)) ≥ 0, ∀i ∈ N, ∀φi, φ′
i ∈ Φi
12
Name
No. of days
Duration of 1 day
No. of agents
No. of events
Max. allowed buy/sell quantity of security
Price of security
Value
50
1sec
2
1
1
[0, 1]
Table 1: Parameters used for our simulation experiments
Proposition 1. If p is a correlated equilibrium and φP is a Pareto optimal
strategy profile calculated by p in a prediction market with risk averse agents,
then the strategy profile φP is incentive compatible, that is each agent is best off
reporting truthfully.
Proof. We prove by contradiction. Suppose that φP is not an incentive com-
patible strategy, that is, there is some other φ′ for which
ui(φ′) ≥ ui(φP )
(14)
Equation 14 violates two properties of φP . First of all, since φP is Pareto
optimal, we know that Equation 14 is not true, since ui(φP ) ≥ ui(φ′) by the
definition of Pareto optimal strategy profile. Secondly, if we rewrite Equation 14
as ui(φP ) − ui(φ′) ≤ 0 and multiply both sides by p(φP ), we get p(φP )[ui(φP ) −
ui(φ′)] ≤ 0. Since p is a correlated equilibrium this inequality can not hold,
otherwise it would violate the definition of the correlated equilibrium.
5 Experimental Results
We have conducted several simulations using our POSGI prediction market.
The main objective of our simulations is to test whether there is a benefit to the
agents to follow the correlated equilibrium strategy. We do this by analyzing
the utilities of the agents and the market price. The default values for the
statistical distributions for market related parameters such as the number of
days over which the market runs representing the duration of an event and the
information signal arrival rate were taken from data obtained from the Iowa
Electronic Marketplace(IEM) movie market for the event Monsters Inc.
[19]
and are shown in Table 1.
For all of our simulations, we assume there is one event in the market with
two outcomes (positive or negative); each outcome has one security associated
with it. We consider events that are disjoint (non-combinatorial). This allows us
to compare our proposed strategy empirically with other existing strategies while
using real data collected from the IEM, which also considers non-combinatorial
events. Since we consider disjoint events, having one event vs. multiple events
does not change the strategic behavior of the agent or the results of each agent3.
We report the market price for the security corresponding to the outcome of the
event being 1 (event occurs).
3We have verified positively that our system can scale effectively with the number of events
and agents.
13
We compare the trading agents' and market's behavior under various strate-
gies employed by the agents. We use the following five well-known strategies for
comparison [22] which are distributed uniformly over the agents in the experi-
ments.
1. ZI (Zero Intelligence) - each agent submits randomly calculated quantity
to buy or sell.
2. ZIP (Zero Intelligence Plus) - each agent selects a quantity to buy or sell
that satisfies a particular level of profit by adopting its profit margin based
on past prices.
3. CP (by Preist and Tol) - each agent adjusts its quantity to buy or sell based
on past prices and tries to choose that quantity so that it is competitive
among other agents.
4. GD (by Gjerstad and Dickhaut) - each agent maintains a history of past
transactions and chooses the quantity to buy or sell that maximizes its
expected utility.
5. DP (Dynamic Programming solution for POSG game) - each agent uses
dynamic programming solution to find the best quantity to buy or sell
that maximizes its expected utility given past prices, past utility, past
belief and the information signal [14].
6. CE (Correlated equilibrium solution) - each agent uses Algorithm 1 from
Section 4.3 to determine a correlated equilibrium strategy that gives a
recommended quantity to buy or sell.
y
g
e
t
a
r
t
s
E
C
w
o
l
l
o
f
e
m
i
t
e
h
t
f
o
e
g
a
t
n
e
c
r
e
P
100
90
80
70
60
50
40
30
20
10
0
ZI
ZI
ZIP
CP
GD
DP
CE
1
0.9
0.8
0.7
0.6
0.5
e
c
i
r
p
t
e
k
r
a
M
DP
0.4
0
10
ZIP
CP
GD
Trading Strategies
y
t
i
l
i
t
u
e
v
i
t
a
u
m
u
C
l
10
8
6
4
2
0
−2
−4
−6
−8
0
ZI
ZIP
CP
GD
DP
CE
10
40
50
20
30
40
50
Number of days
20
30
Number of days
a
b
c
Figure 5: (a) Percentage of the number of days each compared trading strategy
follows a correlated equilibrium strategy, (b) Market Prices, and, (c) Utilities of
the risk neutral agent while using different trading strategies.
5.1 Risk-neutral Agents
For our first group of experiments we assume that the agents are risk neutral.
First, we attempt to understand the correspondence between our CE strategy
and each of the other compared strategies. For comparing the CE strategy with
14
each strategy, we ran the simulations with identical settings, once with agents
using the CE strategy for making trading decisions and then with the same
agents using the compared strategy for making trading decisions. Figure 5(a)
shows the number of times(days) an action recommended to a trading agent by
the CE strategy is the same as the action recommended to a trading agent by
one of the compared strategies, expressed as a percentage. We observe that the
action recommended to a trading agent employing the ZI strategy is the same
as the action recommended to a trading agent using CE strategy only 11% of
the time, while for the trading agent using the more refined DP strategy it is the
same 68% of the time. The higher percentage of adoption of the CE strategy by
the more refined strategies indicates that the DP strategy is more sophisticated
than simpler strategies.
In our next set of experiments, we compare the market prices and utili-
ties obtained by a trading agent over time (days) while using different trading
strategies. In each simulation run, the trading agent uses one of the compared
strategies for determining its action at each time step (day). We compare the
cumulative market prices and utilities for each such simulation. We report a
scenario where the event in the prediction market happens and show the market
prices and utilities for the security corresponding to the positive outcome of the
event. In this scenario, the market price should approach $1 (event happens)
as the prediction market's duration nears end. Figure 5(b) shows the market
prices of the orders placed by the trading agents during the duration of the event
for different strategies. We observe that agents using the CE strategy perform
better since they are able to trade at prices that are closer to $1, indicating that
agents using the CE strategy are able to respond to other agents' strategies
and predict the aggregated price of the security more efficiently. This efficiency
is further supported by the graph in Figure 5(c) that shows the utility of the
agents while using different strategies. We see that the agents using the CE
strategy are finally able to obtain 38% more utility than the agents following
the next best performing strategy (DP), and 85% more utility than the agents
following the second worst performing strategy (ZIP).
1
Market price
e
c
i
r
p
t
e
k
r
a
M
0.5
Cumulative utility using CE
Cumulative utility using ZIP
Market price
cumulative utility using CE
cumulative utility using CP
20
1
e
c
i
r
p
t
e
k
r
a
M
0.5
y
t
i
l
i
t
u
e
v
i
t
a
l
u
m
u
C
0
0
5
10
15
a
25
20
Number of days
30
35
40
45
50
0
5
10
15
b
20
Number of days
25
30
20
y
t
i
l
i
t
u
e
v
l
i
t
a
u
m
u
C
0
35
40
45
50
Figure 6: Comparison of the cumulative market prices and utilities obtained by:
(a) CE and ZIP strategies, and, (b) CE and CP strategies.
For our following set of experiments, we further compare the CE strategy
15
e
c
i
r
p
t
e
k
r
a
M
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Market price
20
Cumulative utility using CE
Cumulative utility using GD
0
10
y
t
i
l
i
t
u
e
v
l
i
t
a
u
m
u
C
1
0.8
0.6
0.4
0.2
e
c
i
r
p
t
e
k
r
a
M
Market price
Cumulative utility using CE
Cumulative utility using DP
25
20
Number of days
30
5
10
15
a
35
40
45
50
5
10
15
b
20
Number of days
25
30
15
y
t
i
l
i
t
u
e
v
10
l
i
t
a
u
m
u
C
5
0
35
40
45
−5
50
Figure 7: Comparison of the cumulative market prices and utilities obtained by:
(a) CE and GD strategies, and, (b) CE and DP strategies.
with other strategies by using two agents in the same simulation run - one
agent using CE strategy and the other agent using the compared strategy. The
results are reported in Figure 6 and 7. We can see that in each scenario the CE
strategy outperforms other strategies. From Figures 6(a) and 7(b) respectively,
we observe that the trading agent using CE strategy gets 95% more utility than
the trading agent using ZIP strategy and 41% more utility than the trading
agent using DP strategy. Overall, we can say that using the POSGI model
allows the agent to avoid myopically predicting prices and use the correlated
equilibrium to calculate prices more accurately and obtain higher utilities.
5.2 Risk averse Agents
y
g
e
t
a
r
t
s
E
C
w
o
l
l
o
f
e
m
i
t
e
h
t
f
o
e
g
a
t
n
e
c
r
e
P
100
90
80
70
60
50
40
30
20
10
0
ZI
ZI
ZIP
CP
GD
DP
CE
1
0.9
0.8
0.7
0.6
0.5
0.4
e
c
i
r
p
t
e
k
r
a
M
DP
0.3
0
10
ZIP
CP
GD
Trading Strategies
20
30
Number of days
y
t
i
l
i
t
u
e
v
i
t
a
u
m
u
C
l
8
6
4
2
0
−2
−4
−6
−8
0
ZI
ZIP
CP
GD
DP
CE
10
40
50
20
30
40
50
Number of days
a
b
c
Figure 8: (a) Percentage of the number of days each compared trading strategy
follows a correlated equilibrium strategy, (b) Market Prices, and, (c) Utilities of
the risk averse agent while using different trading strategies.
Next, we conduct similar simulations for the market while using risk averse
agents that use the risk averse utility function given in Section 4.4. Since in
real world most people are risk averse, we ran several experiments with different
16
Strat. Risk Neutral Agent
Risk Averse Agent
%FCE Price Utility %FCE Price Utility
32% 132%
11%
26% 89%
26%
41%
20% 83%
16% 58%
47%
68%
10% 41%
36% 123%
25% 85%
23% 77%
19% 52%
9%
38%
11%
22%
35%
39%
60%
ZI
ZIP
CP
GD
DP
Table 2: Percentage of the number of days each compared strategy follows the
CE strategy (%F CE in cols. 2 and 5), the percentage of the difference between
each compared strategy and the CE strategy for the market price (cols. 3 and
6) and utility (cols. 4 and 7). Results are shown for both risk neutral and risk
averse trading agents.
values of the risk preference θi > 0. We report the results for θi = 0.8 (strongly
risk averse trading agent) to see a better effect of the risk averse preference of
trading agents on the market prices and their own utilities. To compare the CE
strategy with each of the other strategies we ran the simulation with identical
settings, once with risk averse agents using the CE strategy for making trading
decision and then with the same risk averse agents using the compared strategy
for making trading decisions. Figure 8(a) shows the number of times an action
recommended to a risk averse trading agent by the CE strategy is the same as
the action recommended to a risk averse trading agent by one of the compared
strategies, expressed as a percentage. We can see that the risk averse agents
employing any of the five compared strategies end up adopting the action that
it the same as the action proposed by the CE strategy only 11 − 60% of the
time. We notice that the ZI agents are not affected by the risk averse behavior
of the agents (they follow the CE strategy 11% of the time when the agents
are either risk averse or risk neutral) because they do not use their past utility
to determine future prices. However, other compared agent strategies use their
past utilities to predict future prices. The concave nature of the CRRA utility
function with θi > 0 lowers the utilities that these agents receive, and, this
results in these agents following the CE strategy less often. The effect of the
lowered utility due to the risk averse (concave) utility function is also seen in
Figures 8(b) and (c). Figures 8(b) shows the market prices for an event that
has a final outcome = 1. We again observe that the agents using CE strategy
are able to predict the aggregated price of the security more efficiently during
the duration of the event. Figure 8(c) that shows the utility of the risk averse
agents while using different strategies. We see that the agents using the CE
strategy are able to obtain 42% more utility than the agents following the next
best performing strategy (DP) and 127% more utility than the agents following
the worst performing strategy (ZI).
The summary of all of our results is given in Table 2. For each type of
agent, risk neutral or risk averse, the first column, %FCE, indicates percentage
of the number of days the actions of the trading agents using each compared
strategy are the same as the CE strategy. The second and the third column
show the percentage of the difference between each compared strategy and the
CE strategy for the market price and utility, respectively.
In summary, the
17
POSGI model and the CE strategy result in better price tracking and higher
utilities because they provide each agent with a strategic behavior while taking
into account the observations of the prediction market and the new information
of the events.
6 Conclusions
In this paper, we have described an agent-based POSGI prediction market with
an LMSR market maker and empirically compared different agent behavior
strategies in the prediction market. We proved the existence of a correlated equi-
librium in our POSGI prediction market with risk neutral agents, and showed
how correlated equilibrium can be obtained in the prediction market with risk
averse agents. We have also empirically verified that when the agents follow
a correlated equilibrium strategy suggested by the market maker they obtain
higher utilities and the market prices are more accurate.
In the future we are interested in conducting experiments in an n-player sce-
nario for the POSGI formulation given in Section 3 using richer commercially
available prediction market data sets. We also plan to analyze the performance
of the CE strategy in the scenarios where the outcome of the event does not
correspond to the predicted outcome of the event by the market price, i.e. when
prediction market fails. We also plan to investigate the dynamics evolving from
multiple prediction markets that interact with each other. Finally, we are in-
terested in exploring truthful revelation mechanisms that can be used to limit
untruthful bidding in prediction markets.
Acknowledgments
This research has been sponsored as part of the COMRADES project funded
by the Office of Naval Research, grant number N 000140911174.
References
[1] R. J. Aumann. Subjectivity and correlation in randomized strategies. Jour-
nal of Math- ematical Economics, 1(1):67-96, 1974.
[2] H. Bleichrodt, J. van Rijn, M. Johannesson. Probability weighting and
utility curvature in QALY-based decision making. Journal of Mathematical
Psychology, 43:238-260, 1999.
[3] Y. Chen. Predicting Uncertain Outcomes Using Information Markets:
Trader Behavior and Information Aggregation. New Mathematics and Nat-
ural Computation, 2(3):1-17, 2006.
[4] Y. Chen, T. Mullen, C. Chu. An In-Depth Analysis of Information Markets
with Aggregate Uncertainty. Electronic Commerce Research 2006; 6(2):201-
22.
[5] Y. Chen, D. Pennock. Utility Framework for Bounded-Loss Market Maker.
Proc. of the 23rd Conference on Uncertainty in Articifial Intelligence (UAI
2007), pages 49-56, 2007.
18
[6] Y. Chen, S. Dimitrov, R. Sami, D.M. Reeves, D.M. Pennock, R.D. Hanson,
L. Fortnow, R. Gonen. Gaming Prediction Markets: Equilibrium Strategies
with a Market Maker. Algorithmica, 2009.
[7] B. Cowgill, J. Wolfers, E. Zitzewitz. Using Prediction Markets to Track
Information Flows: Evidence from Google. Dartmouth College (2008);
http://www.bocowgill.com/GooglePredictionMarketPaper.pdf.
[8] S. Das. The Effects of Market-making on Price Dynamics. Proceedings of
AAMAS 2008; 887-894.
[9] S. Dimitrov and R. Sami. Nonmyopic Strategies in Prediction Markets.
Proc. of the 9th ACM Conference on Electronic Commerece (EC), pages
200-209, 2008.
[10] S. Dimitrov. Information Procurement and Delivery: Robustness in Predic-
tion Markets and Network Routing. PhD Thesis, University of Michigan,
2010.
[11] J. Feigenbaum, L. Fortnow, D. Pennock and R. Sami. Computation in a
Distributed Information Market. Theor. Comp. Science, 343(1-2):114-132,
2005.
[12] D. Fudenberg and J. Tirole. Game Theory. MIT Press, 1991.
[13] S. Gjerstad. Risks, Aversions and Beliefs in Predictions Markets. mimeo,
U. of Arizona, 2004.
[14] E. Hansen, D. Bernstein, S. Zilberstein. Dynamic programming for partially
observable stochastic games. In Proceedings of the 19th National Conference
on Artificial Intelligence, pages 709-715, 2004.
[15] R. Hanson. Logarithmic Market scoring rules for Modular Combinatorial
Information Aggregation. Journal of Prediction Markets, 1(1):3-15, 2007.
[16] A. Hoenselaar. Equilibria
in Multi-Player Gamesequilibria. 2007.
http://www.cs.brown.edu/courses/cs244/notes/andreas.pdf.
[17] C.A. Holt, S.K. Laury. Risk aversion and incentive effects. American Eco-
nomic Review, 92:1644-1655, 2002.
[18] K. Iyer, R. Johari, C. Moallemi. Information Aggregation in Smooth Mar-
kets. Proceedings of the 11th ACM conference on Electronic commerce
(EC'10), pages 199-206, 2010.
[19] Iowa Electronic Marketplace. URL: www.biz.uiowa.edu/iem/
[20] J.B. Kadane and R.L.Winkler. Separating probability elicitation from utili-
ties. Journal of the American Statistical Association, 83(402):357-363, 1988.
[21] R.D. Luce, C.L. Krumhansi. Measurement, scaling, and psychophysics. In
Stevens Handbook of Experimental Psychology, Wiley: New York, 1:3-74,
1988.
19
[22] H. Ma, H. Leung. Bidding Strategies in Agent-Based Continuous Dou-
ble Auctions. Whitestein Series In Software Agent Technologies And Auto-
nomic Computing, 2008.
[23] J. Maheu, T. McCurdy. New Arrival Jump Dynamics, and Volatility Com-
ponents for Individual Stock Returns. The Journal of Finance, 59(2):755-
793, 2004.
[24] C. Manski. Interpreting the Predictions of Prediction Markets. Economic
Letters, 91(3):425-429, 2006.
[25] M. Ostrovsky. Information Aggregation in Dynamic Markets with Strate-
gic Traders. Proc. of the 10th ACM conference on Electronic commerce
(EC'09), pages 253-254, 2009.
[26] A. Othman and T. Sandholm. Automated Market Making in the Large:
The Gates Hillman Prediction Market. of the 11th ACM conference on
Electronic commerce (EC'10), pages 367-376, 2010.
[27] A. Othman, T. Sandholm, D. Reeves and D. Pennock. A Practical
Liquidity-Sensitive Automated Market Maker. of the 11th ACM conference
on Electronic commerce (EC'10), pages 377-386, 2010.
[28] C. Papadimitriou, T. Roughgarden. Computing correlated equilibria in
multi-player games. Journal of ACM, 55(3):1-29, 2008.
[29] C. Plott and K. Chen. Information Aggregation Mechanisms: Concept,
Design and Implementation for a Sales Forecasting Problem. Social Science
working paper no. 1131, California Institute of Technology, 2002.
[30] A. Tversky and D. Kahneman. The framing of decisions and the psychology
of choice. Science, 211:453-458, 1981.
[31] P.P. Wakker. Explaining the Characteristics of the Power (CRRA) Utility
Family, Health Economics, 17:1329-1344, 2008.
[32] J. Wolfers and E. Zitzewitz. Prediction Markets. Journal of Economic Per-
spectives, 18(2):107-126, 2004.
[33] J. Wolfers and E. Zitzewitz. Interpreting Prediction Markets as Probabili-
ties. NBER Working Paper No. 12200, May 2006.
[34] Yootles. URL: http://www.yootles.com
20
|
1304.6360 | 1 | 1304 | 2013-04-23T17:40:18 | Assessment of Path Reservation in Distributed Real-Time Vehicle Guidance | [
"cs.MA"
] | In this paper we assess the impact of path reservation as an additional feature in our distributed real-time vehicle guidance protocol BeeJamA. Through our microscopic simulations we show that na\"{\i}ve reservation of links without any further measurements is only an improvement in case of complete market penetration, otherwise it even reduces the performance of our approach based on real-time link loads. Moreover, we modified the reservation process to incorporate current travel times and show that this improves the results in our simulations when at least 40% market penetration is possible. | cs.MA | cs | Assessment of Path Reservation
in Distributed Real-Time Vehicle Guidance
Department of Computer Science / TU Dortmund, Germany
Sebastian Senge
3
1
0
2
r
p
A
3
2
]
A
M
.
s
c
[
1
v
0
6
3
6
.
4
0
3
1
:
v
i
X
r
a
Abstract -- In this paper we assess the impact of path reser-
vation as an additional feature in our distributed real-time
vehicle guidance protocol BeeJamA. Through our microscopic
simulations we show that naıve reservation of links without any
further measurements is only an improvement in case of complete
market penetration, otherwise it even reduces the performance
of our approach based on real-time link loads. Moreover, we
modified the reservation process to incorporate current travel
times and show that this improves the results in our simulations
when at least 40% market penetration is possible.
Index Terms -- vehicle route guidance, distributed system,
swarm intelligence
I. INTRODUCTION
It is widely acknowledged that traffic congestions in urban
road networks have enormous negative impacts with respect
to economical as well as ecological aspects [1]. Hence, there
has been extensive research in the field of traffic flow opti-
mization [2] and (on-line) vehicle guidance recently. In [3]
we proposed a distributed, swarm intelligence-based vehicle
guidance approach termed BeeJamA. In contrast
to state-
of-the-art commercially available systems, based on central-
ized shortest path algorithms, distributed routing protocols
in general have potentially a higher scalability in terms of
road segment (link) load update frequency and covered area.
The BeeJamA approach borrows from the foraging behavior
of honey bees and utilizes a multi-layer hierarichal concept
to disseminate routing relevant link load information every
second but in a limited vicinity only, similar to the scout-
forager behavior of its natural counterpart.
In [4], [5] two distributed ant-based vehicle guidance ap-
proaches have been described. The latter also includes a dy-
namic path reservation approach, however, it is only evaluated
with a complete, or 100%, (market) penetration. Obviously,
this is a factious assumption as different vehicle guidance
concepts are and will be used in realistic environments. As
the contribution of this paper, we add distributed path reser-
vation to our BeeJamA approach and evaluate the resulting
protocol under varying penetration rates on a complex road
network with a microscopic traffic simulator. We will show
that a simple and naıve link reservation even decreases the
protocol's performance with respect to average travel times if
the penetration is not complete. Additionally, we combine the
strength of the BeeJamA protocol, fast link load updates, with
path reservations and show that this hybrid approach performs
better than the naıve one. In our simulations reservations
increase the performance of BeeJamA if the penetration is
at least 40%.
The remainder of this paper is structured as follows. In the
next section, the basic path reservation concept is described.
In Sec. III, the plain BeeJamA protocol is briefly explained
and the additional reservation procedure is introduced. Sec. IV
presents the simulation setup as well as the simulation results.
Finally, we conclude the findings in Sec. V.
II. PATH RESERVATION
Path reservation refers to the registration of vehicles' ex-
pected arrival times on the links of their anticipated path. In
subsequent guidance decisions these link reservations can be
incorporated to help predicting future link loads. The plain
BeeJamA protocol so far intentionally avoids incorporating
any kind of traffic load prediction, but tries to offer routing
recommendations based on up-to-date traffic information be-
fore each intersection in due time. Path reservation, however,
requires some temporal lookahead, since the link travel time
(the link weight from the perspective of routing protocols)
depends on the number of reservations. Hence, one must infer
from the number of reservations of a link the expected travel
time. The basic tools to express this relation are empirical
density / travel time diagrams or formal link performance
functions (LPF). Empirical density / travel time relationships
may be learned offline or online by means of machine learning
techniques like artificial neural networks. LPFs are equations
typically derived from empirical findings as the well known
BPR function (Bureau of Public Roads, superseded by the Fed-
eral Highway Administration of the United States Department
of Transportation, cf. [6]), given by:
(cid:32)
(cid:19)β(cid:33)
(cid:18) fe
ce
te(fe) = t0
e
1 + α
,
(1)
where te(fe) is the resulting travel time for flow fe attempting
to use link e, when the free flow travel time is t0
e and ce the
capacity (in vehicles per hour) the link has been design for.
Different values of α and β are used to adjust the function to
different road types and can be found in [6].
We derived a LPF under the assumption of equidistant
headways of the queue-based macroscopic MATSim traffic
simulator we used in our simulations from the source code and
checked it empirically in a simple test setup where a single
link is on the only path from source to destination for all
vehicles. The function is given by:
te(fe) =
(fe − 1) 3600
ce
fe < ce
else,
,
(2)
(cid:40)t0
e,
t0
e +
(cid:16)
(cid:17)
Fig. 1: Link performance function
Fig. 2: Positive effect of reservation on travel time distribution
where ce is maximum capacity of the link e, a value the
simulator computes on startup from static properties of a
link like type and number of lanes of the road. In general,
the headways must not be not equidistant, thus a pile-up in
the incoming queue of a link occurs and the second case
(fe ≥ ce) describes the travel time by its own. Fig. 1 depicts
the MATSim LPF as well as three BPR instances (with
α = 0.15, β = 4 (default), α = 0.88, β = 9.8 (highway),
α = 1, β = 5.4 (multilane), cf. [6] for more details).
Essentially, path reservation then works as follows. With
each link a reservation log is associated, which consists of
slots (in this paper with a length of 1 minute). Now suppose
a vehicle v is currently at node n0 at time t0 and a path
of nodes (n0, n1, . . . , nm) is calculated, where nm is the
vehicle's destination. First, the expected travel time t0,1of link
(n0, n1) is calculated using the link's LPF. Then, the vehicle
is registered at the link, i.e. each reservation slot for the
link in the interval [t0, t0,1] is incremented and the process
is repeated with link (n1, n2) until the last link is reached.
Thus, subsequent path calculations of other vehicles respect
the path choice of v, since the former registrations lead to
higher LPF values. If a vehicle re-calculates a path during its
travel, it has to update its previous path reservation, i.e. it
removes, advances or postpones the link reservations made so
far and/or adds reservations on other links depending on the
newly calculated path to reflect changes.
Fig. 2 may serve as a general motivation for the idea of
reserving links on a path. There, we compared a dynamic
version of centralized shortest path routing protocol with
(DynResSP) and without (DynSP) path reservation, that has
access to refreshed averaged travel times over the links of the
network every ten minutes. In contrast to the non-reserving
version, the DynResSP protocol reserves each time a new
path is calculated (in other words each 10min at earliest)
the changed path, i.e. the link is annotated such that it is
added when the reserving vehicle will arrive at the beginning
of the link and how long it will likely stay on the link.
(Detailed elucidations of the simulation setup can be found in
Sec. IV.) The average travel time dropped from about 100min
to 42min. As one can see from the boxplots of the travel
time distribution, the lower, medium and upper quartile are
considerably better with path reservation, although there are
a lot of outliers in case of the path reservation protocol and
the maximum values do not differ as desirable. Additionally,
the penetration is 100% and such perfect conditions cannot be
achieved in realistic environments. However, this simplistic
setup makes clear that path reservation may have enormous
potential to reduce travel times.
III. THE BEEJAMA PROTOCOL
The BeeJamA vehicle guidance approach is a distributed
routing protocol based on swarm intelligence and is opti-
mized to disseminate least cost path information in large road
networks as fast as possible. A comprehensive description
of the protocol and additional literature references can be
found in [3]. The following sections outline the basic ideas
of BeeJamA, which are then extended by path reservation.
A. Basic Concepts
BeeJamA is a distance vector protocol and continuously
floods bee agents from each node through the network in the
opposite direction of traffic, so-called upstream scouts, (see
Fig. 3(a)). Once a node in the network receives a scout, it
updates its routing table and floods the agent once again to
all predecessors (except the node the bee arrived from). On
their travel a scout accumulates the travel time from the links
and hence disseminates the cumulated travel time upstream
through the network until the hop limit of a scout is reached.
A scout's hop limit depends on the hierarchy level of its origin
node. An hierarchy of overlapping zones is established by
means of a leader election protocol that creates clusters of
adjacent nodes and assigns (at least one) leader node as cluster
representative per cluster. The leaders may start a new leader
election process among them and thereby create cluster on a
higher logical layer level. The highest layer (with the clusters
0500010000150002000001000020000300004000050000Vehicles / hourTravel Time (in seconds)MATSimBPR DefaultBPR HighwayBPR MultilanelllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllDynSP 10minDynResSP 10min0100200300400500Travel Time (in min) of largest extent) floods bees with infinite hop limit, lower
layers have finite and decreasing hop limits (depending on
the configuration which is a tradeoff between accuracy and
communication overhead). As a result, it is finally guaranteed
that at each node at least one routing table entry leads to at
least one cluster representative the destination node is part
of. Then, vehicles are always forwarded on the current least
cost path to cluster representative of the lowest layer of which
scouts have arrived so far at the current node. Finally, the
last cluster is the one-element cluster of the destination node
itself. The protocol
is based on a vehicle-to-infrastructure
(V2I) architecture, where so-called navigators act in lieu of the
nodes and links in their limited area of influence. These areas
are a node partition of the road network and the navigators
maintain routing tables for each of these nodes as well as
handle the outgoing and incoming bee agents.
B. Integration of Path Reservation in BeeJamA
The basic idea to implement path reservation in BeeJamA is
that the upstream scouts additionally disseminate the predicted
future travel times. Unfortunately, an upstream scout does
not know when vehicles (moving downstream) will arrive at
the node the upstream scout currently passes. Thus, it is not
known for which time in future the LPF should be evaluated.
To circumvent this problem, we introduce downstream scouts
which traverse links in the original direction of traffic flow and
work otherwise as their upstream counterparts: cumulating the
travel time and updating tables. However, these table entries
are not used to forward vehicles but to tell the upstream
scouts how long a vehicle currently would need to travel from
the downstream scout's origin to the current node. Fig. 3(b)
illustrates this concept. There, node a floods a downstream
agent and a situation of the asynchronous process is depicted,
where the scout on the left (and least cost) path to e has already
reached e (suppose that is the vehicle's destination). Therefore,
node e has learned that a vehicle starting now at a would need
8 minutes to arrive at e. The next generation of upstream scouts
starting at e can now evaluate the LPF for the correct time, i.e.
in 5min from now on at node d for link (d, e) (see Fig. 3(c)).
The next necessary difference compared to the plain Bee-
is the use of foragers, which are used to
JamA protocol
increment the number of registered vehicles on a link (see
Fig. 3(d)). This agent is called a forager, since it announces a
consumption of available resources, as a forager bee in reality
reduces the amount of available nectar. After the navigator
selected the next hop for the requesting vehicle, a forager is
emitted on the link towards the selected next hop on the least
cost path, i.e. node b in the example. After arriving there, the
same navigators computes the next hop for the forager as if
it was a vehicle and forwards the agent accordingly. Thus,
the forager agent traverses the same path immediately (only
delayed by the transmission in the communication network),
the slower vehicle would take later if the least cost path does
not change in the meantime. When the forager arrives at the
destination, the whole path is transmitted back to the vehicle.
Subsequent foragers started for this vehicle would possibly
result in different paths due to road network load changes,
which the vehicle could test for equality and de-register at
previously chosen links if the path changed.
By virtue of this approach, a completely distributed and dy-
namic path reservation protocol emerges. The main drawback
of this simple and naive approach (ResBeeJamA-N) is that link
travel time is solely derived from the reservations on a link. In
case of 100% penetration this may be and, as the simulation
results show, is beneficial in terms of average travel times. For
the common situation with incomplete penetration of a single
approach, better ways to predict the expected travel time must
be used. There exist a lot of approaches for short and long time
traffic forecasting [7] often requiring expensive computation
or long learning phases, though. Neither is advantageous in
highly dynamic traffic situations, which is why we present a
simple hybrid technique of incorporating a LPF and the current
mean travel time in the next section.
C. Hybrid Path Reservation
Our dynamic hybrid approach (ResBeeJamA-DH) evaluates
the prediction precision per link and weights the link per-
formance prediction and current load accordingly. The better
the prediction over time is, the higher is the impact of the
link performance function. The basic notion of this approach
is a simple heuristical rule of thumb: if the prediction for
a link in the past was precise, the predictions in the future
will be precise as well. To assess the past prediction accuracy
we compare the difference of past predictions and actual
observed travel time on a link. Formally, we apply a regression
analysis technique, the Pearson correlation coefficient. Thus,
a correlation of time (last hour) and the prediction error
(difference of actual and predicted travel time in the last hour)
is calculated and serves as a basis for weighting the ratio of
the LPF and the current mean travel time.
The Pearson product-moment correlation coefficient for
random variables X, Y is defined as
ρX,Y =
cov(X, Y )
σX σY
,
(3)
where cov(X, Y ) denotes the covariance of X and Y and
σX , σY the variances of X and Y , respectively. Using mea-
sures x1, . . . , xn and y1, . . . , yn the empirical version of Eq. 3
is:
,
(4)
(cid:80)n
i=1(xi − x)2(cid:112)(cid:80)n
(cid:112)(cid:80)n
i=1(xi − x)(yi − y)
i=1(yi − y)2
r =
where x, y are the sample means. During the simulation we
take every minute, at time i, an observation of the current
cur as experienced by the vehicles on each
mean travel time te,i
link e and calculate as well a prediction te,i
lpf according the
current reservations on e. The prediction error err is then
cur − te,i
simply denoted by the difference erre,i = te,i
lpf . The
sample xi expresses the time of an observation relative to
the beginning of the current observation time window (at
least one minute, {x1 = 1} and one hour at a maximum
{x1 = 1, . . . , x60 = 60}). Obviously, these samples are always
(a)
(c)
Fig. 3: BeeJamA agents' concepts
(b)
(d)
Fig. 4: Interpretation of prediction error correlation
te =
cur,
cur,
(5)
positive correlation r will be found and vice versa. In case
of an increasing error on the link, we conclude that
the
predictions are less reliable and the current travel time may
be a better estimation of future travel times (cf. Fig. 4). We
express this intuition by
rte
lpf + (1 − r)te
lpf + rte
tlpf ,
(1 − r)te
IV. SIMULATION STUDIES
r ∈ [−1, 0)
r = 0
r ∈ (0, 1]
strictly increasing, i.e. xi < xi+1. The sample yi
is then
the corresponding prediction error. If the prediction error is
increasing over time (and there is a linear dependence), a
The guidance protocols has been evaluated using the MAT-
Sim microscopic traffic simulator on a road network extracted
from Open Street Map. It is a part of the German Ruhr District
debcgfaUpstream ScoutVehicleUpstream ScoutdebcgfaDownstreamScoutDownstreamScoutVehicle+ 2min+ 3min+ 5min+ 5min+ 8min2min1min2min3min3minLink Travel TimeEst. Arrival TimeLeast Cost PathdebcgfaUpstream ScoutVehicleUpstream Scout+ 2min+ 3min+ 5minDisseminates current travel timeDisseminates current and expected travel times based on est. upstream arrival times and reservations Est. Arrival TimedebcgfaForagerVehicleReserved PathIncrement reservations during given period[+0min,+2min][+2min, +3min][+3min, +5min][+5min, +8min]Prediction error (in seconds)Last hour of simulation time Prediction gets better,weight LPF travel times higher(positive correlation) Prediction gets worse,weight current travel time higher (negative correlation) Fig. 5: Ruhr District
Fig. 7: Comparison for a 50% penetration rate
Fig. 6: BeeJamA vs. ResBeeJamA-N
(see Fig. 5) with a size of about 40 × 20km2. The road
network exhibits 5616 nodes and 10185 links. Throughout all
simulations an identical setup of the same 300.000 vehicles is
reused, where in each of the first five simulation hours 60.000
vehicles start.
We compared two routing protocols: a dynamic centralized
shortest path protocol (DynSP) and our own distributed real-
time protocol BeeJamA. The DynSP protocol is powered by a
dynamic version of the A* algorithm, i.e. the link weights are
updated regularly and the vehicle recalculates the least cost
path with the same frequency during its travel. For example
if a DynSP 30min protocol is deployed, every 30min the
link weights as estimated by the current mean travel time is
refreshed. Moreover, all vehicles guided by a DynSP 30min
protocol, requests an initial route and afterwards each 30min.
In the best case a vehicle recalculates its least cost path right
after the link weights were updated, in the worst case the
link weights are nearly 30min old and thus maybe outdated.
The reserving modifications of these protocols are denoted
as DynResSP and ResBeeJamA and a reserve slot length of
1min is used. The BeeJamA protocol floods each second a
generation of scouts, the areas are obtained by a grid partition
(with a box side length of 1500m).
In the following simulation studies we will also evaluate
Fig. 8: BeeJamA vs. ResBeeJamA-DH
the influence of the protocol's penetration, i.e. the (market)
share of a given protocol. In this paper, remaining vehicles
are guided by a DynSP 30min protocol. For example,
if
ResBeeJamA has a penetration of 30%, 90.000 vehicles are
guided by ResBeeJamA and the remaining 210.000 vehicles
by DynSP 30min.
In Sec. II, the Fig. 2 was used to motivate the concept
the boxplots
of reservation. With the above elucidations,
belong to the travel time distributions of a DynSP 10min
and DynResSP 10min, respectively. As stated in Sec. II, it
is obvious that in this example reservation is beneficial. Next,
we will evaluate if BeeJamA as a real-time protocol could gain
an improvement through reservation.
Fig. 6 depicts the average travel time with an increasing
penetration of BeeJamA and ResBeeJamA-N. Obviously, the
naıve reservation is not able to improve the performance of
BeeJamA, except for the unlikely situation of a complete
penetration, where the average travel
time dropped from
29.17min to 27.68min (see Tab. I for details). Next, in Fig. 7
the cumulative number of arrived vehicles over time under
lllllllllllPenetration Rate (in %)Avg. Travel Time (in min)501001502002500102030405060708090100lBeeJamAResBeeJamA−NSimulation time (in min)Cummulated number of arrived vehicles (in 1000 vehicles)BeeJamAResBeeJamA−NResBeeJamA−SResBeeJamA−DH050100200300400500600700050100150200250300lllllllllllPenetration Rate (in %)Avg. Travel Time (in min)501001502002500102030405060708090100lBeeJamAResBeeJamA−DH TABLE I: BeeJamA travel times (in minutes)
Penetration
BeeJamA
ResBeeJamA
DH
N
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
257.74
155.32
91.02
58.44
45.00
32.78
30.00
28.02
28.96
29.18
29.17
257.74
211.91
159.10
134.29
123.15
94.44
81.77
78.78
76.95
72.23
27.68
257.74
160.46
100.49
69.03
40.91
31.65
29.04
26.91
26.12
25.92
26.53
again both protocols behave similar, however, in the case of
the lower penetration the plain BeeJamA protocol performs
better (resulting in a remarkable difference in the mean travel
time of about 8min) and vice versa in the other case (with a
difference of about 3min in favour of path reservation).
V. CONCLUSION
We presented a thorough simulative analysis of a distributed,
on-line vehicle guidance protocol with a path reservation
extension under varying penetrations. We showed that a naıve
implementation, where the reservations are the only source
to predict future travel times, does not improve the results
in terms of average travel times of our BeeJamA protocol,
except in the unlikely case of a 100% penetration. A hybrid
implementation where the current mean travel time over links
are incorporated additionally improved the results significantly,
though. In future work, will extend the proposed approach to
include marginal cost pricing strategies.
REFERENCES
[1] H. van Essen, A. Schroten, M. Otten, D. Sutter, C. Schreyer, R.
Zandonella, M. Maibach, C. Doll. (2011, Nov.). "External costs of
transport in Europe - Update study for 2008", UIC, [Online]. Avail-
able: http://www.uic.org/IMG/pdf/external costs of transport in europe-
update study for 2008-2.pdf
[2] S. K. Zegeye, B. De Schutter. J. Hellendoorn, E.A. Breunesse, and A.
Hegyi. "A Predictive Traffic Controller for Sustainable Mobility Using
Parameterized Control Policies", In: Trans. on ITS, Vol. 13/3, 2012
[3] S. Senge and H. F. Wedde, "2-Way Evaluation of the Distributed
BeeJamA Vehicle Routing Approach," in Proc. IEEE Intell. Veh. Symp.,
2012, pp. 205-210
[4] B. Tatomir, L.J.M. Rothkrantz, "H-ABC: A scalable dynamic routing
algorithm", In: Recent Advances in Artificial Life, World Scientific
Publishing, Singapore, 2005
[5] R. Claes, T. Holvoet, D. Weyns. "A Decentralized Approach for Antici-
patory Vehicle Routing Using Delegate Multiagent Systems", In: Trans.
on ITS, Vol. 12/2, 2011
[6] A. J. Horrowitz. (1991). "Delay/Volume Relations for Travel Forecasting
Based upon the 1985 Highway Capacity Manual", Federal Highway
Administration, US Department of Transportation, [Online]. Available:
https://pantherfile.uwm.edu/horowitz/www/
SpeedVolume1985HCMReport.pdf
[7] A. Hofleitner, R. Herring, P. Abbeel, and A. Bayen. "Learning the
Dynamics of Arterial Traffic From Probe Data Using a Dynamic Bayesian
Network", In: Trans. on ITS, Vol. 13/4, 2012
Fig. 9: 30% penetration: BeeJamA vs. ResBeeJamA-DH
Fig. 10: 90% penetration: BeeJamA vs. ResBeeJamA-DH
different kinds of BeeJamA for a penetration of 50% is shown.
Beside ResBeeJamA-N and -DH, we also evaluated a simple
static approach (-S), where the weight of the current and LPF
travel time is proportional to the penetration rate, i.e. for the
50% case, the link weight is 2 × tlpf . The intuitive notion of
this weighting is that under a normal distribution assumption,
the number of vehicles actually on a link is twice the registered
vehicles if the penetration is 50%.
The naıve path reservation as well as the static version per-
forms worse, it takes roughly 700min for all vehicles to arrive
compared to about 450min for the plain BeeJamA and the
ResBeeJamA-DH protocol. Interestingly, the DH reservation
is even slightly better and hence we focused in the following
simulations on this variant. Fig. 8 shows the average travel
time over time for the BeeJamA and ResBeeJamA-DH. Even
though, both curves are similar, with at least 40% penetration
the path reservation promises benefits for this certain simula-
tion setup. Fig. 9 and Fig. 10 depict the cumulative number of
arrived vehicles broken down by (Res)BeeJamA and DynSP
protocol for a penetration of 30% and 90%, respectively. Once
lllllllllllllllllllllllllllllllllllllllllllllllllllllllSimulation time (in min)Cummulated number of arrived vehicles (in 1000 vehicles)0501001502002503000100200300400500600lBeeJamA / BeeJamA guidedBeeJamA / DynSP guidedResBeeJamA−DH / ResBeeJamA−DH guidedResBeeJamA−DH / DynSP guidedllllllllllllllllllllllllllllllllllllllllSimulation time (in min)Cummulated number of arrived vehicles (in 1000 vehicles)0501001502002503000100200300400lBeeJamA / BeeJamA guidedBeeJamA / DynSP guidedResBeeJamA−DH / ResBeeJamA−DH guidedResBeeJamA−DH / DynSP guided |
1807.02870 | 2 | 1807 | 2018-09-16T02:35:47 | QDDS: A Novel Quantum Swarm Algorithm Inspired by a Double Dirac Delta Potential | [
"cs.MA"
] | In this paper a novel Quantum Double Delta Swarm (QDDS) algorithm modeled after the mechanism of convergence to the center of attractive potential field generated within a single well in a double Dirac delta well setup has been put forward and the preliminaries discussed. Theoretical foundations and experimental illustrations have been incorporated to provide a first basis for further development, specifically in refinement of solutions and applicability to problems in high dimensional spaces. Simulations are carried out over varying dimensionality on four benchmark functions, viz. Rosenbrock, Rastrigrin, Griewank and Sphere as well as the multidimensional Finite Impulse Response (FIR) Filter design problem with different population sizes. Test results illustrate the algorithm yields superior results to some related reports in the literature while reinforcing the need of substantial future work to deliver near-optimal results consistently, especially if dimensionality scales up. | cs.MA | cs | QDDS: A Novel Quantum Swarm Algorithm
Inspired by a Double Dirac Delta Potential
Saptarshi Sengupta, Sanchita Basak, Richard Alan Peters II
Department of Electrical Engineering and Computer Science
Vanderbilt University, Nashville, TN, USA
E-mail: {saptarshi.sengupta, sanchita.basak, alan.peters}@vanderbilt.edu
Abstract -- In this paper a novel Quantum Double Delta Swarm
(QDDS) algorithm modeled after the mechanism of convergence
to the center of attractive potential field generated within a single
well in a double Dirac delta well setup has been put forward and
the preliminaries discussed. Theoretical
foundations and
experimental illustrations have been incorporated to provide a
first basis for further development, specifically in refinement of
solutions and applicability to problems in high dimensional spaces.
Simulations are carried out over varying dimensionality on four
benchmark functions, viz. Rosenbrock, Rastrigrin, Griewank and
Sphere as well as the multidimensional Finite Impulse Response
(FIR) Filter design problem with different population sizes. Test
results illustrate the algorithm yields superior results to some
related reports in the literature while reinforcing the need of
substantial
to deliver near-optimal results
consistently, especially if dimensionality scales up.
future work
Keywords -- Quantum Particle Swarm; Swarm Intelligence; Double
Delta Potential Well; Quantum Mechanics; Optimization
I. INTRODUCTION
The traditional Quantum-behaved Particle Swarm Optimization
(QPSO) algorithms [1-3] (both Types I and II) extend the
classical Newtonian dynamics of agent propagation in the
canonical Particle Swarm Optimization (PSO) [4-5] [13] to a
quantum framework. The convergence of particles to promising
regions in the solution space in QPSO is driven by an attractive
potential field directed towards the center of a singular Dirac
Delta well. This is followed by the collapse of the wavefunction
indicative of the particles' states using recursive Monte Carlo.
The traditional types of QPSO are quite efficient and
inexpensive candidates suitable for highly non-linear and non-
convex optimization leveraging the quantum nature of the
particles and the corresponding wavefunctions that sample a
larger region of the solution space as compared to their classical
counterpart employing a binary strategy: a particle is either
present at a unique location or not. However, the intuitively
simple rendering of the binding force exerted by a single delta
well on a particular particle has been the subject of scrutiny and
further research as demonstrated by [6]. One line of thought
contends that the attractive coupling offered by a multi-well
attractor is stronger than that offered by a singular one therefore
facilitating a stronger stable equilibrium criterion [7]. Xie et al.
[6] have recently proposed a quantum-behaved PSO based on a
double delta model which assimilates the following three
components: a) the global best (gbest) position, b) an agent's
location with respect to the gbest position and c) an agent's
location with respect to the mean of individual agents' best
positions. The authors chose to model the personal and global
best positions as centers of two singular delta potential wells,
thereby arriving at a multi-scale representation of a double delta
potential well. Simulations on widely used test functions such
as Rosenbrock, Rastrigrin, Griewank and Sphere using varying
population sizes for problem dimensionality 10 through 30 at a
step size of 10 indicate the effective outcomes obtained using
the algorithm. Further, the authors note that using two attractors
in place of one increased the global search capability and
convergence accuracy. In our work however, we are concerned
more about finding the state equations of particles based on two
spatially co-located delta potential wells. To this effect, an
isolated system of double Dirac-delta potential wells and
convergence to its centers are considered. The resulting
iterative state updates mimic the trajectory of a bound particle
(E<0) as it moves towards the lowest energy configuration viz.
the point where the attractive potential is the least i.e. the center
of the dominant well. By identifying constraints on the motion
of the particles guided by their probabilistic nature of existence,
an iterative scheme of convergence is put forward. The
resulting algorithm (QDDS) is tested on a suite of benchmark
functions which have many local minima, are bowl-shaped or
valley-shaped as well as on the design optimization goal of a
low-pass Finite
filter. Trial
evaluations indicate the efficiency of the algorithm in finding
solutions of acceptable quality with room for improvement both
in terms of computational expense and finetuning of solutions.
The organization of the paper is as follows: Section II discusses
the physics governing the quantum mechanical model and sets
up algorithmic foundations, Section III walks through the
pseudocode with implementation details and Section IV reports
test results. Section V briefly analyzes the outcomes followed
by an analysis of the QDDS mechanism in Section VI with
concluding remarks in Section VII.
Impulse Response
(FIR)
II. SWARM PROPAGATION USING A DOUBLE-DELTA
POTENTIAL WELL
We start from the time-independent Schrodinger's wave
equation which is stated as:
[−
ћ2
2𝑚
𝛻2 + 𝑉(𝑟)]𝜓(𝑟) = 𝐸 𝜓(𝑟) (1)
ψ(r), V(r), m, E and ћ represent the wave function, the potential
function, the reduced mass, the energy of the particle and
Planck's constant respectively. Let us consider a particle in a
double delta well, whose potential can be expressed as:
𝑉(𝑟) = −𝛼{𝛿(𝑟 + 𝑎) + 𝛿(𝑟 − 𝑎)}
α expresses the strength of the well and {-a, a} are the centers
of the two wells. Considering the even solutions of the time-
independent Schrodinger's equation (Eq. 1) and assuming V= 0
at regions away from the centers of the two wells we get:
(2)
−
ћ2
2𝑚
𝑑2
𝑑𝑟2 𝜓(𝑟) = 𝐸𝜓(𝑟) (3)
The intention is to find solutions of the wave function ψ in a
double delta well setup for E<0 (bound states) in regions
ℝ1: 𝑟 ∈ (−∞, 𝑎), ℝ2: 𝑟 ∈ (−𝑎, 𝑎) 𝑎𝑛𝑑 ℝ3: 𝑟 ∈ (𝑎, ∞).
Assuming k to be equal to (√2𝑚𝐸 ћ⁄ ), the even solutions can
be expressed as [8]:
𝜓𝑒(𝑥) = {
𝐴𝑒−𝑘𝑟 𝑟 > 𝑎
𝐵𝑒−𝑘𝑟 + 𝐶𝑒𝑘𝑟 0 < 𝑟 < 𝑎
𝐵𝑒𝑘𝑟 + 𝐶𝑒−𝑘𝑟 − 𝑎 < 𝑟 < 0
𝐴𝑒𝑘𝑟 𝑟 < −𝑎
(4)
To solve for the constants described in the above equation, we
solve for the continuity of the wave function 𝜓𝑒 at 𝑟 = 𝑎 and
𝑟 = −𝑎 and for the continuity of the derivative of the wave
function at 𝑟 = 0. Hence, we arrive at the equations for 𝜓𝑒
stated below [8]:
𝜓𝑒(𝑟) = {
𝐵(1 + 𝑒2𝑘𝑎)𝑒−𝑘𝑟 𝑟 > 𝑎
𝐵(𝑒−𝑘𝑟 + 𝑒𝑘𝑟) − 𝑎 < 𝑟 < 𝑎
𝐵(1 + 𝑒2𝑘𝑎)𝑒𝑘𝑟 𝑟 < −𝑎
(5)
We are not considering the solution of the odd wave function
𝜓𝑜 here as the existence of a solution is not guaranteed [8]. It is
also interesting to note that the bound state energy in a double
delta potential well is lower than that compared to a single delta
potential well by approximately by a factor of (1.11)2 ≈ 1.2321
[7]. Next, if we consider the behavioral dynamics of a particle
to be compliant with the Schrodinger wave equation i.e. Eq. (1),
then we need to find the particle's probability density function
for its behavioral characterization, which is given by the square
of the magnitude of the wave function described in Eq. (5).
Thus, we have to find 𝜓(𝑟)2. In order to say that there is a
greater than 50% chance of finding a particle in the vicinity of
the center of any of the potential wells, the following criterion
must be satisfied [1]:
𝑟
∫ 𝜓(𝑟)2 𝑑𝑟
−𝑟
> 0.5 (6)
= 0.5𝑔 (1 < 𝑔 < 2) (7)
where −𝑟 and 𝑟 denote the two boundaries of the vicinity.
Although we are trying to confine the particle in any one of the
two potential wells, the wave function considered here is
different from that considered in the case of a single potential
well setup (as in traditional QPSO), This is due to the influence
of the other well which is taken into consideration when
deriving conditions for the confinement of a particle in a single
well. Eq. (6) can also be written as:
𝑟
∫ 𝜓(𝑟)2 𝑑𝑟
−𝑟
For ease of computation, we now consider that one of the wells
is centered at 0. Solving for conditions of confinement of the
particle in that well and computing ∫ (𝜓(𝑟)2 𝑑𝑟
for regions
ℝ20−: 𝑟′ ∈ (−𝑟, 0) by applying the second condition of Eq. (5)
and ℝ20+: 𝑟′ ∈ (0, 𝑟) by applying the first condition of Eq. (5),
we arrive at the equation below:
𝐵2 =
We replace the denominator of the R.H.S. (𝑒2𝑘𝑟 − 5𝑒−2𝑘𝑟 +
4𝑘𝑟 + 4) as δ. Thus, we get:
δ = 𝑒2𝑘𝑟 − 5𝑒−2𝑘𝑟 + 4𝑘𝑟 + 4 (9)
Equating 𝐵2 in L.H.S. of equation (8) for any two consecutive
iterations (assuming it is a constant over iterations as it not a
function of time) we get Eq. (10), Eq. (11) and Eq. (12):
𝑒2𝑘𝑟− 5𝑒−2𝑘𝑟+4𝑘𝑟+4
(8)
𝑟
−𝑟
𝑘𝑔
𝑔𝑖𝑡𝑒𝑟−1
𝑒2𝑘𝑟𝑖𝑡𝑒𝑟−1− 5𝑒−2𝑘𝑟𝑖𝑡𝑒𝑟−1+4𝑘𝑟𝑖𝑡𝑒𝑟−1+4
⇒
=
𝑔𝑖𝑡𝑒𝑟−1
δ𝑖𝑡𝑒𝑟−1
𝑔𝑖𝑡𝑒𝑟
δ𝑖𝑡𝑒𝑟
=
𝑒2𝑘𝑟𝑖𝑡𝑒𝑟− 5𝑒−2𝑘𝑟𝑖𝑡𝑒𝑟+4𝑘𝑟𝑖𝑡𝑒𝑟+4
𝑔𝑖𝑡𝑒𝑟
(10)
(11)
⁄
⇒ δ𝑖𝑡𝑒𝑟 = 𝐺. δ𝑖𝑡𝑒𝑟−1 (0.5 < 𝐺 < 2) (12)
G is the ratio (𝑔𝑖𝑡𝑒𝑟 𝑔𝑖𝑡𝑒𝑟−1
) and can vary from 0.5 to 2 since
(1<g<2). To keep a particle moving towards the center of a
potential well we find 𝛿𝑖𝑡𝑒𝑟 (0.5 𝛿𝑖𝑡𝑒𝑟−1 < 𝛿𝑖𝑡𝑒𝑟 < 2 𝛿𝑖𝑡𝑒𝑟−1).
Thus, we find 𝛿𝑖𝑡𝑒𝑟 at the current iteration based on 𝛿𝑖𝑡𝑒𝑟−1
(found in the previous iteration) by adding or subtracting the
gradient of 𝛿𝑖𝑡𝑒𝑟−1 multiplied by a learning rate. The governing
conditions of finding 𝛿𝑖𝑡𝑒𝑟 depend on the relation of 𝛿𝑖𝑡𝑒𝑟−1
with its version in the previous iteration, i.e., δ𝑖𝑡𝑒𝑟−2 as well as
the sign of the gradient of 𝛿𝑖𝑡𝑒𝑟−1 as described in Algorithm 1.
The learning rate θ is designed as:
𝜃 = (1 − є ) (
𝑚𝑎𝑥. 𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠 − 𝑐𝑢𝑟𝑟𝑒𝑛𝑡 𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛
𝑚𝑎𝑥. 𝑖𝑡𝑒𝑟𝑎𝑡𝑖𝑜𝑛𝑠
) + є (13)
є is a small fraction between [0,1] set by the user. The learning
rate θ decreases linearly from 1 to є with the passage of
iterations. Once we obtain 𝛿𝑖𝑡𝑒𝑟, we back-solve Eq. (9) to
retrieve 𝑟𝑖𝑡𝑒𝑟 which denotes a particle's position as well as the
potential solution for that particular iteration. We let 𝑟𝑖𝑡𝑒𝑟, i.e. a
particle's position in the current iteration maintain a component
towards the best position found so far (gbest) along with its
current solution obtained from 𝛿𝑖𝑡𝑒𝑟. Subsequently, a cost
function is computed with the solution 𝑟𝑖𝑡𝑒𝑟 and if it is better
than the best cost found thus far, the cost and the corresponding
solution are stored in memory. This process is repeated over the
total number of evaluations to find the overall best cost and best
solution for the swarm. The complete procedure is described in
Section III.
III. QUANTUM DOUBLE DELTA SWARM ALGORITHM (QDDS)
In this section, we present the pseudocode of the Quantum
Double Delta Swarm (QDDS) Algorithm.
for each dimension
Initialize positions 𝒓𝟏 and 𝒓𝟐 for iterations 1 and 2
Algorithm 1. Quantum Double Delta Swarm Algorithm
Initialization Phase
1: Initialize k
2: Initialize a small constant λ randomly
3: Initialize maximum no. of iterations as maxiter
4: Initialize bestcost
5: for each particle
6:
7:
8: end for
9: end for
10: Generate 𝜹𝟏 and 𝜹𝟐 from 𝒓𝟏 and 𝒓𝟐 according to Eq. (9)
11: Set iteration count iter=3
Optimization Phase
12: while (iter<maxiter) and
{(𝜹𝒊𝒕𝒆𝒓−𝟏 < 𝟎. 𝟓 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐) or (𝜹𝒊𝒕𝒆𝒓−𝟏 > 𝟐 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐)}
13: Find learning rate θ using eq. (13)
14: Select a particle randomly
15: for each dimension
16:
17:
18:
19:
20:
21:
22:
23:
24: end if
25: end for
26: Solve 𝒓𝒊𝒕𝒆𝒓 from 𝜹𝒊𝒕𝒆𝒓
27: Generate a random number 𝝆 between 0 and 1
28: 𝒓𝒊𝒕𝒆𝒓=𝝆 ∗ 𝒓𝒊𝒕𝒆𝒓 + (𝟏 − 𝝆) ∗ 𝒓𝒈𝒃𝒆𝒔𝒕
29: Compute cost using 𝒓𝒊𝒕𝒆𝒓
30: if 𝒄𝒐𝒔𝒕𝒊𝒕𝒆𝒓<bestcost
31:
32:
33: end if
34: iter = iter + 1
35: end while
if (𝜹𝒊𝒕𝒆𝒓−𝟏 > 𝟐 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐) and 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏>0
𝜹𝒊𝒕𝒆𝒓 = 𝜹𝒊𝒕𝒆𝒓−𝟏 − 𝜽 ∗ 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏 ∗ 𝝀
elseif (𝜹𝒊𝒕𝒆𝒓−𝟏 > 𝟐 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐) and 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏<0
𝜹𝒊𝒕𝒆𝒓 = 𝜹𝒊𝒕𝒆𝒓−𝟏 + 𝜽 ∗ 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏 ∗ 𝝀
elseif (𝜹𝒊𝒕𝒆𝒓−𝟏 < 𝟎. 𝟓 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐) and 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏<0
𝜹𝒊𝒕𝒆𝒓 = 𝜹𝒊𝒕𝒆𝒓−𝟏 − 𝜽 ∗ 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏 ∗ 𝝀
elseif (𝜹𝒊𝒕𝒆𝒓−𝟏 < 𝟎. 𝟓 ∗ 𝜹𝒊𝒕𝒆𝒓−𝟐) and 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏>0
𝜹𝒊𝒕𝒆𝒓 = 𝜹𝒊𝒕𝒆𝒓−𝟏 + 𝜽 ∗ 𝜵𝜹𝒊𝒕𝒆𝒓−𝟏 ∗ 𝝀
bestcost = 𝒄𝒐𝒔𝒕𝒊𝒕𝒆𝒓
best_solution = 𝒓𝒊𝒕𝒆𝒓
IV. EXPERIMENTS AND RESULTS
A. Benchmark Functions
To evaluate the performance of the proposed algorithm,
benchmark functions such as Rastrigrin, Rosenbrock, Sphere
and Griewank have been considered. In addition to this, we test
the algorithm on the multidimensional low pass FIR filter
design problem using filter orders 10 and 20.
Table 1. Benchmark Functions Considered for Testing
Function
Rastrigrin
Rosenbrock
Sphere
Expression
𝑓(𝑥) = 𝐴𝑛 + ∑ [𝑥𝑖
𝑛
𝑖=1
2 − 𝐴𝑐𝑜𝑠(2𝜋𝑥𝑖)]
, A=10
𝑛−1
𝑓(𝑥) = ∑[100(𝑥𝑖+1 − 𝑥𝑖
2)2 + (1 − 𝑥𝑖)2]
𝑖=1
𝑛
2
𝑓(𝑥) = ∑ 𝑥𝑖
𝑖=1
Griewank
𝑓(𝑥) = 1 +
1
4000
𝑛
∑ 𝑥𝑖
𝑖=1
𝑛
𝑖=1
2 − ∏ cos (
𝑥𝑖
√𝑖
)
Min
0
0
0
0
B. Example Application: Design of High Dimensional
Finite Impulse Response (FIR) Filters
This subsection outlines
the design procedure of a
multidimensional low-pass Finite Impulse Response (FIR)
filter as proposed in [9-10]. The ideal filter response 𝐻𝑑(𝑒𝑗𝜔)
and system transfer function 𝐻(𝑧) governing the design of the
filter are given by Eqs. (14) and (15) respectively:
𝐻𝑑(𝑒𝑗𝜔) = 1 0 ≤ 𝜔 ≤ 𝜔𝑝
= 0 𝜔𝑠 ≤ 𝜔 ≤ 𝜋 (14)
𝐻(𝑧) = ∑
Here h(n) denotes the filter's impulse response and N is the order
of the filter having N+1 coefficients. The normalized passband
and stopband edge frequencies are 𝜔𝑝 and 𝜔𝑠 respectively and
Ep and Es are errors in pass band and stop band given by Eqs.
(16) and (17).
𝑛 = 0,1, … . 𝑁 (15)
ℎ(𝑛)𝑧−𝑛
𝑁
𝑛=0
𝐸𝑝 =
1
𝜋
𝜔𝑝
∫ (1 − 𝐻(𝜔))2𝑑𝜔
0
𝐸𝑠 =
1
𝜋
𝜋
∫ 𝐻(𝜔)2𝑑𝜔
𝜔𝑠
(16)
(17)
A cost function 𝛾 of choice used in [9] is re-used for the
minimization objective:
𝛾 = 𝜂𝐸𝑝 + (1 − 𝜂)𝐸𝑠 0 ≤ 𝜂 ≤ 1 (18)
The main focus of the optimization routine is to find a balanced
response that seeks to minimize the cost function 𝛾, thereby
minimizing the passband and stopband errors. The trade-off
between reducing 𝐸𝑝 and 𝐸𝑠 is controlled by selection of a
weight 𝜂 ∈ [0,1] as per the user's design condition.
C. Simulation Results on the Benchmark Functions
Table 2. Experimental Results for the Rosenbrock Function using 10 Independent Trials of QDDS
P
Dim
Iter
20
10
1000
20
30
1500
2000
40
10
1000
20
30
1500
2000
80
10
1000
20
30
1500
2000
PSO [3]
Mean
± St. Dev.
94.1276
±194.3648
204.337
±293.4544
313.734
±547.2635
71.0239
±174.1108
179.291
±377.4305
289.593
±478.6273
37.3747
±57.4734
83.6931
±137.2637
202.672
±289.9728
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
QPSO [3]
Mean
± St. Dev.
51.9761
±0.4737
136.8782
±0.6417
157.4707
±0.8287
17.3177
±0.1515
54.0411
±0.4210
81.1382
±0.0319
7.5755
±0.2708
32.9970
±0.2068
53.6422
±0.2616
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
WQPSO [3]
Mean
± St. Dev.
35.8436
±0.2843
62.7696
±0.4860
70.9525
±0.4283
16.9583
±0.1336
54.2439
±0.3752
57.0883
±0.3437
10.1650
±0.2345
47.0275
±0.3507
51.8299
±0.3103
Iter
250
375
500
250
375
500
250
375
500
QDDS
Best
Worst
8.8224
8.9700
18.9071
19.2355
29.2827
29.5926
8.7483
9.0642
18.9883
19.2278
29.1769
29.5359
8.8569
9.0454
18.8227
19.2097
29.2291
29.5685
Mean
± St. Dev
8.8912
±0.0529
19.0593
±0.0991
29.4457
±0.0883
8.9030
±0.0875
19.1421
±0.0745
29.3855
±0.1321
8.9213
±0.0670
19.0663
±0.1339
29.4015
±0.1079
Table 3. Experimental Results for the Rastrigrin Function using 10 Independent Trials of QDDS
P
Dim
Iter
20
10
1000
20
30
1500
2000
40
10
1000
20
30
1500
2000
80
10
1000
20
30
1500
2000
PSO [3]
Mean
± St. Dev.
5.5382
±3.0477
23.1544
±10.4739
47.4168
±17.1595
3.5778
±2.1384
16.4337
±5.4811
37.2796
±14.2838
2.5646
±1.5728
13.3826
±8.5137
28.6293
±10.3431
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
QPSO [3]
Mean
± St. Dev.
4.8274
±0.0015
16.0519
±0.0414
33.7218
±0.0114
3.1794
±6.033e-04
10.8824
±0.0496
21.4530
±0.0949
2.2962
±0.0130
7.8544
±0.0011
15.9474
±0.0198
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
WQPSO [3]
Mean
± St. Dev.
4.0567
±0.0094
12.1102
±0.0287
23.5593
±0.0713
2.8163
±0.0083
9.7992
±0.0628
17.4436
±0.0034
1.8857
±0.0118
7.2855
±0.0032
15.0255
±0.0294
Iter
250
375
500
250
375
500
250
375
500
QDDS
Best
Worst
0.0425
0.2394
0.4765
0.7564
0.7499
1.5330
0.0459
0.2628
0.3117
0.6847
0.6379
1.4135
0.0543
0.2145
0.3059
0.7633
0.9302
1.7447
Mean
± St. Dev
0.1214
±0.0592
0.5776
±0.0967
1.1709
±0.2528
0.1163
±0.0577
0.5539
±0.1300
1.1440
±0.2768
0.1397
±0.0486
0.5312
±0.1331
1.3041
±0.2578
Table 4. Experimental Results for the Sphere Function using 10 Independent Trials of QDDS
P
Dim
Iter
20
10
1000
20
30
1500
2000
40
10
1000
20
30
1500
2000
80
10
1000
20
30
1500
2000
PSO [3]
Mean
± St. Dev.
3.16e-20
±6.23e-20
5.29e-11
±1.56e-10
2.45e-06
±7.72e-06
3.12e-23
±8.01e-23
4.16e-14
±9.73e-14
2.26e-10
±5.10e-10
6.15e-28
±2.63e-27
2.68e-17
±5.24e-17
2.47e-12
±7.16e-12
QPSO [3]
WQPSO [3]
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
Mean
± St. Dev.
1.3909e-41
±1.4049e-43
3.5103e-22
±3.5452e-24
5.3183e-14
±5.3623e-16
2.5875e-71
±2.6137e-73
3.7125e-42
±3.7500e-44
4.2369e-30
±1.7009e-33
8.5047e-102
±7.5974e-104
1.1542e-68
±1.1585e-70
2.2866e-49
±2.3070e-51
Mean
± St. Dev.
2.2922e-056
±1.5365e-58
2.9451e-40
±2.8717e-42
3.9664e-33
±3.8435e-35
5.5806e-80
±5.6370e-82
8.8186e-055
±7.1785e-57
5.4389e-44
±2.4132e-45
4.7144e-106
±4.7620e-108
2.5982e-74
±2.6243e-76
2.3070e-51
±1.9125e-62
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
Iter Mean
QDDS
Best
Worst
250
375
500
250
375
500
250
375
500
± St. Dev
6.2437e-04
±3.0752e-04
0.0027
±6.9011e-04
0.0067
±0.0013
7.3478e-04
±2.3036e-04
0.0028
±8.3546e-04
0.0060
±8.4108e-04
5.1027e-04
±1.4758e-04
0.0026
±6.0137e-04
0.0055
±8.8175e-04
3.0163e-04
0.0014
0.0019
0.0039
0.0046
0.0085
4.9032e-04
0.0012
0.0016
0.0041
0.0048
0.0072
3.1457e-04
7.5794e-04
0.0020
0.0040
0.0043
0.0069
Table 5. Experimental Results for the Griewank Function using 10 Independent Trials of QDDS
QPSO [3]
WQPSO [3]
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
Mean
± St. Dev.
5.5093e-04
±0.0657
1.0402e-04
±0.0211
1.2425e-04
±0.0110
1.6026e-04
±0.0496
1.7127e-04
±0.0167
3.9088e-05
±0.0085
3.3744e-04
±0.0327
4.1701e-04
±0.0168
1.3793e-05
±0.0106
Iter
1000
1500
2000
1000
1500
2000
1000
1500
2000
Mean
± St. Dev.
5.6353e-04
±5.5093e-04
2.1318e-04
±1.0402e-04
2.1286e-04
±1.2425e-04
0.0020
±1.6026e-04
1.6861e-04
±1.7127e-04
3.6762e-05
±3.9088e-05
1.5281e-04
±3.3744e-04
3.2549e-04
±4.1701e-04
4.2231e-05
±1.3793e-05
Iter
250
375
500
250
375
500
250
375
500
Mean
± St. Dev
8.0851e-05
±3.3375e-05
1.9821e-04
±6.2856e-05
2.5607e-04
±6.4991e-05
6.9932e-05
±3.3276e-05
1.9180e-04
±4.7197e-05
2.3775e-04
±5.3165e-05
7.4205e-05
±3.2774e-05
1.8714e-04
±5.6483e-05
2.8736e-04
±4.6883e-05
QDDS
Best
Worst
2.3539e-05
1.2727e-04
1.2262e-04
2.9547e-04
1.6388e-04
3.5948e-04
3.2751e-05
1.3565e-04
1.1890e-04
2.7165e-04
1.5328e-04
3.1541e-04
2.9087e-05
1.4314e-04
1.1359e-04
2.5826e-04
2.0340e-04
3.6768e-04
P
Dim
Iter
20
10
1000
20
30
1500
2000
40
10
1000
20
30
1500
2000
80
10
1000
20
30
1500
2000
PSO [3]
Mean
± St. Dev.
0.09217
±0.0833
0.03002
±0.03255
0.01811
±0.02477
0.08496
±0.0726
0.02719
±0.02517
0.01267
±0.01479
0.07484
±0.07107
0.02854
±0.0268
0.01258
±0.01396
D. Parameter Settings
We choose the constant k to be 5 and λ to be the product of a
random number drawn from a normal distribution with 𝜇 = 0
and 𝜎 = 0.5 and a factor of the order of 10−3. 𝜌 is a random
number drawn between 0 to 1. The learning rate θ decreases
linearly with iterations from 1 to 0.3. All experiments are
carried out on two Intel(R) Core(TM) i7-5500U CPU @
2.40GHz with 8GB RAM and one Intel(R) Core(TM) i7-2600U
CPU @ 3.40GHz with 16GB RAM using MATLAB R2017a.
10 trials are carried out and results reported without any use of
GPUs.
Tables 2 through 5 report performance of the QDDS algorithm
on the Rosenbrock, Rastrigrin, Sphere and Griewank functions
as well as compare and contrast with performances of PSO,
QPSO and Weighted Mean Best QPSO (WQPSO) on the same
benchmarks as reported by Xi et al. in [3]. All mean values and
standard deviations listed in Tables 2,3,4 and 5 for the
algorithms PSO, QPSO and WQPSO have been obtained from
the work of Xi et al [3]. These have been used for a comparison
of the performance of our algorithm (QDDS) on one-fourth the
number of iterations with all other conditions for population
and dimension remaining the same.
P and Dim represent the number of particles and the
dimensionality of the functions. Figures 1 through 12 plot the
convergence profiling of the above experiments on the stated
benchmarks of orders 10, 20 and 30 over 10 independent trials.
For purposes of brevity, results of simulations using only
population size 20 are reported in this section. A detailed
analysis can be found in Section V.
Figure 1: Conv. Profiling of Rastrigrin using QDDS
(dimension=10, population=20)
Figure 2: Conv. Profiling of Rastrigrin using QDDS
(dimension=20, population=20)
Figure 3: Conv. Profiling of Rastrigrin using QDDS
(dimension=30, population=20)
Figure 8: Conv. Profiling of Griewank using QDDS
(dimension=20, population=20)
Figure 9: Conv. Profiling of Griewank using QDDS
(dimension=30, population=20)
Figure 10: Conv. Profiling of Sphere using QDDS
(dimension=10, population=20)
Figure 4: Conv. Profiling of Rosenbrock using QDDS
(dimension=10, population=20)
Figure 5: Conv. Profiling of Rosenbrock using QDDS
(dimension=20, population=20)
Figure 6: Conv. Profiling of Rosenbrock using QDDS
(dimension=30, population=20)
Figure 7: Conv. Profiling of Griewank using QDDS
Figure 11: Conv. Profiling of Sphere using QDDS
(dimension=10, population=20)
(dimension=20, population=20)
Figure 12: Conv. Profiling of Sphere using QDDS
(dimension=30, population=20)
E. Simulation Results for the Finite Impulse Response
(FIR) Filter Design Problem
illustrates
following subsection
The
the design of a
multidimensional low-pass Finite Impulse Response (FIR)
filter using a population size of 1000 and an iteration count of
250 and 500 for filter orders 10 and 20 respectively. The
passband and stopband edges are set at 0.3π and 0.6π.
Table 6. Simulations for the 10-Dimensional FIR Filter
Design Problem using 10 Independent Trials of QDDS
Mean Cost
St. Dev
Best Cost Worst Cost
1.4817e-05
1.6762e-05
5.7632e-08
4.5367e-05
Δ
(dB)
-13.6466
Figure 13: Response of the 10-Dimensional FIR Filter
Table 7. Best 10-Dimensional Filter Coefficients (10 Trials)
Filter
h(1) = h(10)
Coefficients
h(2) = h(9)
h(3) = h(8)
h(4) = h(7)
h(5) = h(6)
0.070824792496751651
-0.063184376757871669
-0.038806613903081974
0.013227497402604124
0.39889122413816075
Table 8. Simulations for the 20-Dimensional FIR Filter
Design Problem using 10 Independent Trials of QDDS
Δ
(dB)
Mean Cost
St. Dev
Best Cost Worst Cost
-17.7398
7.1306e-05
9.9875e-05
1.6055e-06
3.1458e-04
Figure 14: Response of the 20-Dimensional FIR Filter
Table 9. Best 20-Dimensional Filter Coefficients (10 Trials)
Filter
Coefficients
h(1) = h(20)
h(2) = h(19)
h(3) = h(18)
h(4) = h(17)
h(5) = h(16)
h(6) = h(15)
h(7) = h(14)
h(8) = h(13)
h(9) = h(12)
h(10) = h(11)
0.011566963779404912
0.0077331878563942523
-0.0094736298940968737
-0.0068424142182682956
0.024047530227972496
0.04099248691610477
0.14983102243854188
0.0057626071427242216
-0.0038505536917844913
0.28023279944300716
Tables 6 and 8 list the maximum stopband attenuation and the
mean, standard deviation, best and worst cost values when
using QDDS for 10 trials. Tables 7 and 9 report the set of best
filter coefficients for orders 10 and 20 whereas Figures 13 and
14 plot the corresponding filter responses.
V. ANALYSIS OF EXPERIMENTAL RESULTS
the Rosenbrock
From Table 2 it is observed that the QDDS algorithm performs
significantly better on
function of
dimensionality 10, 20 and 30 as compared to PSO, QPSO and
WQPSO. In Table 3 the QDDS algorithm generates mean
values which are at least 11.521 times smaller than WQPSO or
even smaller in case of QPSO and PSO on the Rastrigrin
function of dimensionality 10, 20 and 30. However, the
standard deviation values obtained using WQPSO and QPSO
are clearly superior to those found using QDDS. The results
from Table 4 using the Sphere function indicate the sub-par
performance of QDDS with respect to the competitor
algorithms. Table 5 reports somewhat comparable results in
terms of mean cost using WQPSO, QPSO and QDDS, while it
is to be noted that QDDS has a standard deviation at least ~160
times smaller than that of QPSO. The convergence profiles in
Figures 1 through 12 point out that QDDS is fairly consistent in
its ability to converge to local optima of acceptable quality. It
is obvious that the solutions to the problems discussed in the
paper are in fact local optima, however the solution qualities
corresponding to some of these local optima obtained using
QDDS are evidently superior to some related reports in the
literature [3,6,11,12]. One way to improve the performance of
QDDS may be to not use gradient descent but a problem-
independent optima seeking mechanism in the 𝛿 update step of
the algorithm in Section III. The FIR filter responses record a
maximum stopband attenuation of -13.6466 dB and -17.7398
dB on dimensions 10 and 20 for a QDDS implementation using
gradient descent.
VI. ANALYSIS OF THE QDDS MECHANISM
The Quantum Double Delta Swarm (QDDS) Algorithm is
based on a quantum double Dirac delta potential well model
and is an extension of the conventional QPSO (Type I and Type
II) which are modeled after a singular Dirac delta potential well.
The intuitively simple iterative updates of QDDS lead the
swarm towards fitter regions of the search space in conjunction
with reaching for regions of lower energy for a particle under
the influence of a spatially co-located attractive double delta
potential. The current form of the algorithm, however is prone
to delivering suboptimal results because of the use of a gradient
descent scheme in the 𝜹 update phase. In addition to this, the
time complexity of the algorithm is markedly high due to the
computationally heavy numerical approximation of 𝑟𝑖𝑡𝑒𝑟 from
𝛿𝑖𝑡𝑒𝑟 in the transcendental Eq. (9), as also outlined in Algorithm
1. The effects of using different social/cognition attractors as
well as multi-scale particle topologies remain to be investigated
and a thorough characterization of initialization schemes versus
numerical accuracy is a logical follow-up. Overall, despite the
high computational overhead QDDS produces acceptable
solutions for some problems (Tables 2-3, 5) and not for some
others, as seen in Table 4. However, the fact that QDDS is based
on a legitimate optimization phenomenon in quantum physics
and that it produces good quality solutions on some classical
benchmarks warrants some resource expenditure in exploring
better how it works.
VII. CONCLUDING REMARKS
In this paper, a new swarming mechanism derived from the
quantum mechanics of a double delta potential well is proposed
and its inner workings explored. The mechanism (QDDS) is
formally derived and a computable form is put forward
followed by experiments to determine its accuracy in finding
global optima. This is achieved by approaching some classic
benchmark functions such as Rosenbrock, Rastrigrin, Sphere
and Griewank as well as the multidimensional FIR filter design
problem. Experimental results provide
the
performance of the proposed approach on Rosenbrock,
Rastrigrin and Griewank functions on which it appears to
perform better, while not performing as well in the Sphere
function. In addition to this, successful implementation of FIR
filters of orders 10 and 20 are observed. While the present
version of the algorithm is computationally expensive and
stagnates occasionally, there is room for improvement in both
areas. Future studies would aim at addressing these issues in
addition to performing statistical significance tests to quantify
the performance of the approach over an extensive suite of
separable and non-separable functions of much higher
dimensionality. However, QDDS captures a
relatively
unexplored approach in blending swarm intelligence and
insight
into
optimization and extends the literature on quantum-inspired
computational intelligence.
ACKNOWLEDGEMENTS
This work was made possible by the financial and computing
support by the Vanderbilt University Department of EECS. The
authors would also like to thank the anonymous reviewers for
their suggestions on improving the paper.
CONFLICT OF INTEREST
The authors declare that there is no conflict of interest regarding
the publication of this paper.
REFERENCES
[1] Sun, J., Feng, B., Xu, W.B., "Particle swarm optimization with particles
having quantum behavior.", IEEE Proceedings of Congress on
Evolutionary Computation, pp. 325 -- 331, 2004.
[2] Sun, J., Xu,W.B., Feng, B., "A global search strategy of quantum behaved
particle swarm optimization.", Cybernetics and Intelligent Systems
Proceedings of the 2004 IEEE Conference, pp. 111 -- 116, 2004.
[3] Xi, M., Sun, J., Xu, W., "An improved quantum-behaved particle swarm
optimization algorithm with weighted mean best position", Applied
Mathematics and Computation, Volume 205, Issue 2, 2008, pp. 751-759.
[4] Kennedy, J., Eberhart, R., "Particle swarm optimization.", Proc. IEEE
Int. Conf. Neural Network, 1995.
[5] Eberhart, R., Shi, Y., "Comparison between genetic algorithms and
particle swarm optimization.", Proc. 7th Ann. Conf. Evolutionary
Computation, San Diego, 2000.
[6] Xie, Z., Liu, Q., Xu, L., "A New Quantum-Behaved PSO: Based on
Double δ-Potential Wells Model", Proc. of 2016 Chinese Intelligent
Systems Conference, CISC 2016, Lecture Notes in Electrical
Engineering, vol 404. Springer, Singapore, 2016, pp 211-219.
[7] Basak, S., Lecture Notes, P303 (PE03) Quantum Mechanics I, National
India,
Institute
http://www.niser.ac.in/~sbasak/p303_2010/06.09.pdf.
Education
Science
of
and
Research,
[8] Griffiths, David J. (2005), Introduction to Quantum Mechanics, 2nd
Edition; Pearson Education - Problem 2.27.
[9] Dhabal, S. and Sengupta, S., "Efficient design of high pass FIR
filter using quantum-behaved particle swarm optimization with
weighted mean best position," Proceedings of
the 2015 Third
International Conference on Computer, Communication, Control
and Information Technology (C3IT), Hooghly, 2015, pp. 1-6.
[10] Sengupta, S. and Basak, S., "Computationally efficient low-pass
FIR filter design using Cuckoo Search with adaptive Levy step
size," 2016 International Conference on Global Trends in Signal
Computing
Processing,
Communication
(ICGTSPICC), Jalgaon, 2016, pp. 324-329.
Information
and
[11] Han,P., Yuan, S., Wang, D., "Thermal System Identification Based on
Double Qantum Particle Swarm Optimization", Intelligent Computing in
Smart Grid and Electrical Vehicles, ICSEE 2014, LSMS 2014,
Communications in Computer and Information Science, vol 463.
Springer, Berlin, Heidelberg, 2014, pp 125-137.
[12] Jia, P., Duan, S., Yan, J., "An enhanced quantum-behaved particle swarm
optimizatin based on a novel computing way of local attractor",
Information 2015, 6, 633 -- 649.
[13] Sengupta, S. et al., "Particle Swarm Optimization: A Survey of Historical
and Recent Developments with Hybridization Perspectives", ArXiv
Preprint ArXiv:1804.05319, 2018.
|
1203.5399 | 1 | 1203 | 2012-03-24T10:12:05 | Agent-time Epistemics and Coordination | [
"cs.MA",
"cs.DC",
"cs.LO"
] | A minor change to the standard epistemic logical language, replacing $K_{i}$ with $K_{\node{i,t}}$ where $t$ is a time instance, gives rise to a generalized and more expressive form of knowledge and common knowledge operators. We investigate the communication structures that are necessary for such generalized epistemic states to arise, and the inter-agent coordination tasks that require such knowledge. Previous work has established a relation between linear event ordering and nested knowledge, and between simultaneous event occurrences and common knowledge. In the new, extended, formalism, epistemic necessity is decoupled from temporal necessity. Nested knowledge and event ordering are shown to be related even when the nesting order does not match the temporal order of occurrence. The generalized form of common knowledge does {\em not} correspond to simultaneity. Rather, it corresponds to a notion of tight coordination, of which simultaneity is an instance. | cs.MA | cs |
Agent-time Epistemics and Coordination
Ido Ben-Zvi
Department of Electrical Engineering, Technion
[email protected]
Yoram Moses
Department of Electrical Engineering, Technion
[email protected]
November 4, 2018
Abstract
A minor change to the standard epistemic logical language, replac-
ing Ki with K(cid:104)i,t(cid:105) where t is a time instance, gives rise to a generalized
and more expressive form of knowledge and common knowledge oper-
ators. We investigate the communication structures that are necessary
for such generalized epistemic states to arise, and the inter-agent coor-
dination tasks that require such knowledge. Previous work has estab-
lished a relation between linear event ordering and nested knowledge,
and between simultaneous event occurrences and common knowledge.
In the new, extended, formalism, epistemic necessity is decoupled from
temporal necessity. Nested knowledge and event ordering are shown to
be related even when the nesting order does not match the temporal
order of occurrence. The generalized form of common knowledge does
not correspond to simultaneity. Rather, it corresponds to a notion of
tight coordination, of which simultaneity is an instance.
1
Introduction
We have recently embarked on an in-depth inquiry concerning the relation
between knowledge, coordination and communication in multi agent sys-
tems [5, 6, 4]. This study uncovered new structural connections between
the three in systems where agents have accurate clocks, and there are (com-
monly known) bounds on the time it may take messages to be delivered,
between any two neighboring agents. We call this the synchronous model.
In such a setting, one often reasons about what agents know at particular
1
time points in the past or future, as well as, in particular, what they know
about what other agents will know (or have known) at various other times,
etc.
The emphasis on information regarding the times at which facts are
known has lead us to consider a formalization in which epistemic operators
are indexed by a pair consisting of an agent and a time -- thus, K(cid:104)i,t(cid:105) refers
to what i knows at time t -- rather than the more traditional epistemic op-
erator Ki, in which knowledge is associated with an agent, and the formulas
are evaluated with respect to a particular time. An agent-time pair (cid:104)i, t(cid:105) is
called a node, and we distinguish the new node-based (or nb-) language from
the traditional agent-based one.
In this paper we formulate an expressive language with nb-knowledge
operators, and use it to extend and strictly strengthen the previously estab-
lished relations between knowledge, coordination and communication. Our
earlier inquiries reduced coordination tasks to states of knowledge, and then
analyzed the communication requirements required to obtain these states
of knowledge. Together these provide a crisp structural characterization
of necessary and sufficient conditions for particular coordination tasks, in
terms of the communication required.
We specify coordination tasks in terms of orchestrating a pattern of re-
sponses in reaction to a spontaneous external input initiated by the environ-
ment. Such a spontaneous event is considered as the trigger for its responses.
As shown in [5], ensuring a linearly ordered sequence of responses to such
a triggering event requires a state of deeply nested knowledge, in which the
last responder to know that the next-to-last responder knows, . . . , that the
agent performing the first response knows that the trigger event occurred.
In asynchronous systems without clocks, such nested knowledge can only
be obtained via a message chain from the trigger passing through the re-
sponders in their order in the sequence. The situation in the synchronous
setting, in the presence of timing information, is much more delicate and in-
teresting. In that case, such nested knowledge requires a particular pattern
of communication called a centipede. While nested knowledge is captured
by a centipede, common knowledge corresponds to a more restrictive and
instructive communication pattern called a broom 1). Using the connection
between common knowledge and simultaneity, brooms are shown to underly
natural forms of simultaneous coordination.
The node-based operators give rise to natural strict generalizations of
nested and common knowledge. Of particular interest is node-based com-
1A structure called centibroom in [5] has since been renamed to be a broom.
2
mon knowledge, which is represented by an epistemic operator CA, in which
A is a set of agent-time pairs (cid:104)i, t(cid:105). So C{(cid:104)i,t(cid:105),(cid:104)j,t+3(cid:105)}ϕ indicates, among other
things, that agent i at time t knows that agent j at time t + 3 will know that
i at time t knew that ϕ holds. While nb-common knowledge strictly gener-
alizes classical common knowledge, it shares most of the typical attributes
associated with common knowledge. However, whereas (traditional) com-
mon knowledge is intimately related to simultaneity, nb-common knowledge
is not! We will show that, instead, nb-common knowledge precisely captures
a form of coordination we call tightly-timed response (TTR ). Of course, si-
multaneity is a particular form of tightly-timed coordination.
In general, we follow the pattern of investigation that we established in
earlier papers: Given an epistemic state (such as nb-nested or nb-common
knowledge) we examine, on the one hand, what form of communication
among the agents is necessary for achieving such a state, and on the other
we characterize the type of coordination tasks that require such a state.
As nb-knowledge operators generalize the standard agent-based ones, the
results we obtain here generalize and extend the ones of [5, 4]. Relating
knowledge to communication, we present two "knowledge-gain" theorems
(formulated in the spirit of Chandy and Misra's result for systems without
clocks [7]). Nb-nested and nb-common knowledge are shown to require more
flexible variants of the centipede and the broom communication structures
(respectively), in order to arise.
Perhaps more instructive is the characterization of the coordinative tasks
related to these epistemic states. As mentioned above, nested knowledge
characterizes agents that are engaged in responding in a sequential order to
a triggering event. It turns out that Nb-nested knowledge is similarly neces-
sary when the ordering task predefines not only the sequence of responses,
but also a bound on the elapsed time between each response and its suc-
cessor. We define this task as the Weakly Timed Response problem (WTR).
Interestingly, minute differences in the coordinative task specification result
in dramatic changes to the required epistemic state. As for common knowl-
edge, as hinted above it will be seen that the node-based version relaxes the
typical simultaneous responses requirement into tightly-timed ones.
The main contributions of this paper are:
• Node-based epistemic operators are defined. Under a natural semantic
definition, both nb-knowledge and nb-common knowledge are S5 oper-
ators, while nb-common knowledge satisfies fixed-point and induction
properties analogous to those of standard common knowledge.
• The theory of coordination and its relation to epistemic logic are ex-
3
tended, significantly generalizing previous results. Nb-semantics helps
to decouple epistemic and temporal necessity. Namely, there are cases
in which knowledge about (the guarantee of) someone's knowledge at
a future time, typically based on known communication, is required in
order to perform an earlier action.
• The new epistemic operators are used as a formal tool for the study of
generalized forms of coordination and the communication structures
they require. On the one hand, natural coordination tasks that require
nb-nested and nb-common knowledge are identified. On the other
hand, the communication structures necessary for attaining these nb-
epistemic states are established. Combining the two types of results
yields new structural connection between communication and coordi-
nation tasks.
• The well-known strong connection between common knowledge and
simultaneity established in [12, 11, 10] is shown to be a particular in-
stance of a more general phenomenon: We prove that tightly-timed
synchronization is very closely related to node-based common knowl-
edge. The form of tight synchronization corresponding to agent-based
common knowledge is precisely simultaneous coordination.
This paper is structured as follows. In the next section we define the
syntax and semantics of two languages: One is a traditional agent-based
logic of knowledge, and the other a node-based logic. An embedding of
agent-based formulas in the nb-language is established, and basic properties
of nb-knowledge and nb-ck are discussed. Section 3 presents the synchronous
model within which we will study knowledge and coordination. In Section 4
a review of the notions underlying the earlier analysis, and its main results,
is presented. Our new analysis is presented in Sections 5 and 6. Section 5
defines "uneven" centipedes and relates them to nested nb-knowledge. It
then defines a coordination task called weakly-timed response and shows
that it is captured by nested nb-knowledge. Analogously, Section 6 defines
uneven brooms and relates them to nested nb-common knowledge. It then
defines tightly-timed response and shows that it is captured by nested nb-
ck. Finally, a short discussion and concluding remarks and presented in
Section 7. Proofs of the claims in Sections 5 and 6 are presented in the
Appendix.
4
2 Agent and Node-based Semantics
We consider both the standard, agent-based, epistemic language and its
extension into the node-based variant within the interpreted systems frame-
work of [11]. In this framework, the multi agent system is viewed as consist-
ing of a set P = {1, . . . , n} of agents, connected by a communication network
which serves as the exclusive means by which the agents interact with each
other.
We assume that, at any given point in time, each agent in the system is in
some local state. A global state is just a tuple g = (cid:104)(cid:96)e, (cid:96)1, . . . , (cid:96)n(cid:105) consisting of
local states of the agents, together with the state (cid:96)e of the environment. The
environment's state accounts for everything that is relevant to the system
that is not contained in the state of the agents. A run r is a function
from time to global states. Intuitively, a run is a complete description of
what happens over time in one possible execution of the system. We use
ri(t) to denote agent i's local state (cid:96)i at time t in run r, for i = 1, . . . , n.
For simplicity, time here is taken to range over the natural numbers rather
than the reals (so that time is viewed as discrete, rather than dense or
continuous). Round t in run r occurs between time t − 1 and t. A system
R is an exhaustive set of all possible runs, given the agents' protocol and
the context, where the latter determines underlying characteristics of the
environment as a whole.
To reason about the knowledge states of agents, a simple logical lan-
guage is introduced. Since we are focusing on coordination tasks in which
actions are triggered by spontaneous events, the only primitive propositions
we consider are ones that state that an event has occurred. The standard,
agent-based, variant of this language is L0. It uses the following grammar,
with the usual abbreviations for ∨ and ⇒, and with e, i and G used as terms
for an event, an agent and a group of agents, respectively.
ϕ
::=
occurred(e) ϕ ∧ ϕ ¬ϕ Kiϕ EGϕ CGϕ
The semantics is as follows. Propositional connectives, omitted from the
list, are given their usual semantics.
Definition 1 (L0 semantics). The truth of a formula ϕ ∈ L0 is defined with
respect to a triple (R, r, t).
• (R, r, t) (cid:15) occurred(e) iff event e has occurred in r by time t.
• (R, r, t) (cid:15) Kiϕ iff (R, r(cid:48), t) (cid:15) ϕ for all r ∈ R s.t. ri(t) = r(cid:48)i(t).
• (R, r, t) (cid:15) EGϕ iff (R, r, t) (cid:15) Kiϕ for every i ∈ G.
5
• (R, r, t) (cid:15) CGϕ iff (R, r, t) (cid:15) (EG)kϕ for every k ≥ 1.
Note that while non-epistemic formulas require only a run r and time t
in order to be evaluated, the epistemic operators also require the complete
system of runs R (which provides the analogue of the set of states, or possible
worlds in standard modal logic). The second clause, giving semantics for the
knowledge operator Ki, has built-in an assumption that time is common
knowledge, since knowledge at time t in r depends only on truth in other
runs, that are indistinguishable for agent i from r at the same time t. This
is appropriate for our intended analysis, since in our model (to be presented
in Section 3) time is assumed to be common knowledge. We used L0 to
formalize the analysis in [5].
It will be convenient to
We next define the node-based language L1.
define the set V = {(cid:104)i, t(cid:105) : i ∈ P, t is a time} of all possible nodes. In the
grammar we now use e, α and A as terms for an event, an agent-time node
and a group of such nodes, respectively.
ϕ
::=
occurredt(e) ϕ ∧ ϕ ¬ϕ Kαϕ EAϕ CAϕ
Notice that all statements in this language are "time-stamped:" they
refer to explicit times at which the stated facts hold. As a result, they
are actually time-invariant, and state facts about the run, rather than facts
whose truth depends on the time of evaluation. Therefore, semantics for
formulas of L1 are given with respect to a system R and a run r ∈ R. The
semantics is as follows (again. ommiting propositional clauses).
Definition 2 (L1 semantics). The truth of a formula ϕ ∈ L1 is defined with
respect to a pair (R, r).
• (R, r) (cid:15) occurredt(e) iff event e has occurred in r by time t
• (R, r) (cid:15) K(cid:104)i,t(cid:105)ϕ iff (R, r(cid:48)) (cid:15) ϕ for all runs r(cid:48) s.t. ri(t) = r(cid:48)i(t).
• (R, r) (cid:15) EAϕ iff (R, r) (cid:15) Kαϕ for every α ∈ A.
• (R, r) (cid:15) CAϕ iff (R, r) (cid:15) (EA)kϕ for every k ≥ 1.
In principle, the node-based semantics, as proposed here, can be seen
as a simplified version of a real-time temporal logic with an explicit clock
variable [1], and more generally of a hybrid logic [2]. We are not aware of
instances in which such logics have been combined with epistemic operators.
Nevertheless, in this paper we aim to utilize the formalism, rather than
explore it. Hence, issues of expressibility, completeness and tractability are
left unattended, to be explored at a future date.
6
We now demonstrate that, in a precise sense, the traditional language
L0 can be embedded in the language L1, in a meaning-preserving manner.
We do this by way of defining a "timestamping" operation ts transforming
a formula ϕ ∈ L0 and a time t to a formula ts(ϕ, t) = ϕt ∈ L1 that is
timestamped by t. We then prove the following lemma showing that the
timestamping is sound.
Lemma 1. There exists a function ts : L0 × T ime → L1 such that for every
ϕ ∈ L0 , time t, and ϕt = ts(ϕ, t):
(R, r, t) (cid:15) ϕ
(R, r) (cid:15) ϕt.
iff
Lemma 1 shows that L1 is at least as expressive as L0. Yet L1 is not
equivalent in expressive power to L0, since it allows for multiple temporal
reference points where epistemic operators are involved. For example, the
formula K(cid:104)i,t(cid:105)K(cid:104)j,t(cid:48)
(cid:105)ϕ cannot be translated into an equivalent L0 formula,
since L0 does not allow for the temporal reference point to be shifted when
switching from the outer knowledge operator to the inner one.
Compare a typical agent-based nested knowledge formula, such as ψ =
KiKjϕ, with a node based counterpart such as ψ(cid:48) = K(cid:104)i,t(cid:105)K(cid:104)j,t+3(cid:105)ϕ. In ψ,
gaining knowledge about the epistemic state of other agents means that
the knowledge of the referred other agent has already been gained. By
contrast, in ψ(cid:48) agent i has epistemic certainty concerning the knowledge
state of another agent j, at a particular time in the future. To sum up, the
node-based formalism allows us to differentiate between epistemic necessity,
and temporal tense. Formally trivial, this decoupling is quite elusive as we
tend to mix one with the other. The rest of this paper follows up on the
implications of this point.
It is well-known that standard agent-based common knowledge is closely
Indeed, both CGϕ ⇒ EGCGϕ and
related to simultaneity [12, 11, 10].
KiCGϕ ⇒ CGϕ are valid formulas. (Recall that a formula ψ ∈ L0 is valid
if (R, r, t) (cid:15) ψ for all choices of R, run r ∈ R and time t.) Thus, the first
instant at which CGϕ holds must involve a simultaneous change in the local
states of all members of G. In contrast, simultaneity is not an intrinsic prop-
erty of node-based common knowledge. As an example, consider the node set
A = {(cid:104)i, t(cid:105),(cid:104)j, t + 10(cid:105)}. Although we still have that both K(cid:104)i,t(cid:105)CAϕ ⇒ CAϕ
and CAϕ ⇒ K(cid:104)j,t+10(cid:105)CAϕ are valid formulas, if the current time is t and
i knows that CAϕ holds, this does not mean that j currently knows this
too. The bond with simultaneity has been broken. As we shall see in Sec-
tion 6, however, a notion of tight temporal coordination that generalizes
simultaneity is intrinsic to node-based common knowledge.
7
In the spirit of the treatment in [11], for a formula ψ ∈ L1 we write
R (cid:15) ψ and say that ψ is valid in (the system) R if (R, r) (cid:15) ψ for all r ∈ R.
Node-based common knowledge manifests many of the logical properties
shown by the standard notion of common knowledge. Proof of the following
lemma, and of all new lemmas in this paper, can be found in the Appendix.
Lemma 2.
• Both K(cid:104)i,t(cid:105) and CA are S5 modalities.
• CAϕ ⇒ EA(ϕ ∧ CAϕ) is valid.
• If R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ) then R (cid:15) ϕ ⇒ CAψ
The second clause of Lemma 2 corresponds to the so-called "fixed-point"
axiom of common knowledge, while the third clause corresponds to the "in-
duction (inference) rule" [11].
3 The Synchronous Model
As mentioned, we focus on synchronous systems where the clocks of the in-
dividual agents are all synchronized, and there are commonly known bounds
on message delivery times. Generally, in order to define a system of runs
that conforms to the required setting of interest, we define the system of
runs R as a function R = R(P, γ) of the protocol P followed by the agents,
and the underlying context of use γ.
We identify a protocol for an agent i with a function from local states
of i to nonempty sets of actions. (In this paper we assume a deterministic
protocol, in which a local state is mapped to a singleton set of actions. Such
a protocol essentially maps local states to actions.) A joint protocol is just
a sequence of protocols P = (P1, . . . , Pn), one for each agent.
In this paper we will assume a specified protocol for the agents, namely
the fip, or full information protocol. In this protocol, every agent sends out
its complete history to each of its neighbors, on every round. In order to be
able to do that, the agent must be able to recall its own history. We therefore
also assume this capability, called perfect recall, for the agents. Note that
using a pre-specified protocol is not in general necessary, and is done for
deductive purposes only. Our findings can be applied to all protocols under
slight modifications.
In order for a well defined system of runs R to emerge, the context γ
needs to be rigorously defined as well. We also assume a specific context
8
γmax, within which the agents are operating. Most notably, γmax specifies
that (a) agents share a universal notion of time, (b) the communication
network has upper bounds on delivery times, and (c) all nondeterminism is
deferred to the environment agent.
A word regarding nondeterminism, that plays a crucial role in the anal-
ysis of knowledge gain. Formal analysis allows us to escape the more diffi-
cult questions associated with this concept. For us, intuitively, spontaneous
events are ones that cannot be foretold by the agents in the system. These
could stand for a power shortage, but also for some user input that is com-
municated to an agent via a console. Formally, since all of the agents are
following a deterministic protocol, nondeterminism is only introduced into
the system by the environment.
For brevity, we must omit the formal details2, and make do with describ-
ing the properties of the outcome system R = R(fip, γmax):
• Global clock and global network - The current time is always common
knowledge. The network, which can be encoded as a weighted graph, is
also common knowledge. For each (i, j) connected by a communication
channel, maxij, the weight on the corresponding network edge, denotes
the maximal transmission times for messages sent along this channel.
• Events - There are four kinds of events: message send and receive
events, internal calculations, and external inputs. Events occur at a
single agent, within a single round. All are self explanatory except for
the last type. External input events occur when a signal is received by
an agent from "outside" the system. This could be user input, fate,
and also - less dramatically - it could be used to signify the initial
values for internal variables at the beginning of a run.
• Environment protocol - The environment is responsible for the occur-
rence of two of the event types: message deliveries and external inputs.
(a) message deliveries - the protocol dictates that message deliveries
will occur only for messages that have been sent and are still en route.
Apart from that, the environment nondeterministically chooses when
to deliver sent messages, subject only to the constraint that for every
communication channel (i, j), transmission on the channel does not
take longer than maxij rounds. (b) external inputs - in each round
the environment nondeterministically chooses a (possibly empty) set
of agents at which external inputs will occur, and the kind of events
2The interested reader can find complete accounts of the context we assume in [5, 4]
9
that will occur there. The choice is entirely unconstrained by previous
or simultaneous occurrences.
4 Previous Findings
Our approach is based on the findings of Lamport [13] and of Chandy and
Misra [7]. In his seminal analysis, Lamport defines potential causality, a for-
malization of message chains. Two events are related by potential causality
when the coordinates marking their occurrences (the time and agent at
which they occur) are related by an unbroken sequence of messages. We
rephrase the relation as one holding between pairs of agent-time nodes.
Definition 3 (Potential causality [13]). Fix r ∈ R. The potential causality
relation (cid:32) over nodes of r is the smallest relation satisfying the following
three conditions:
1. If t ≤ t(cid:48) then (cid:104)i, t(cid:105) (cid:32) (cid:104)i, t(cid:48)(cid:105);
2. If some message is sent at (cid:104)i, t(cid:105) and received at (cid:104)j, t(cid:48)(cid:105) then (cid:104)i, t(cid:105) (cid:32)
3. If (cid:104)i, t(cid:105) (cid:32) (cid:104)h, t(cid:48)(cid:48)(cid:105) and (cid:104)h, t(cid:48)(cid:48)(cid:105) (cid:32) (cid:104)j, t(cid:48)(cid:105), then (cid:104)i, t(cid:105) (cid:32) (cid:104)j, t(cid:48)(cid:105).
(cid:104)j, t(cid:48)(cid:105); and
Lamport used the relation as a basis for his logical clocks mechanism that
allows distributed protocol designers in asynchronous contexts to temporally
order events despite the lack of synchronization. Chandy and Misra later
gave an epistemic analysis, showing that for agent j at time t(cid:48) to know of
an occurrence at agent i's at time t, it must be that (cid:104)i, t(cid:105) (cid:32) (cid:104)j, t(cid:48)(cid:105) holds.
In [5, 6, 4] we applied similar methodology to analyze knowledge gain and
temporal ordering of events in synchronous systems, like the one described
here. Note that temporally ordering events in a system with global clock is
quite easy if exact timing is prearranged as a part of the protocol:
Charlie will deposit the money at 3pm, you will sign the contract at 4,
and I will deliver the merchandise at 5.
Coordination becomes more challenging once the time of occurrence of
the triggering, initial, action is nondeterministic:
Charlie will deposit the money, err... whenever, then you will sign the
contract, and only then will I deliver the merchandise.
To approach the issue of event ordering we define the Ordered Response
coordination challenge. Such problems are defined based on a triggering
event es, which is an event of type external input, and a set of responses
10
α1, . . . , αk. Each response αh is a pair (cid:104)ih, ah(cid:105) indicating an action ah that
agent ih is required to carry out. When a response αh gets carried out in
a specific run, we denote with αh = (cid:104)ih, th(cid:105) the specific agent-time node at
which action ah occurs.
Definition 4 (Ordered Response [5]). Let es be an external input nondeter-
ministic event. A protocol P solves the instance OR = (cid:104)es, α1, . . . , αk(cid:105) of
the Ordered Response problem if it guarantees that
1. every response αh, for h = 1, . . . , k, occurs in a run iff the trigger
event es occurs in that run.
2. for h = 1, . . . , k − 1, response αh occurs no later than response αh+1
(i.e. th ≤ th+1).
We can show that solving the challenge for me, you and Charlie requires
that you will know that Charlie has deposited before you sign the contract,
and that I will know that you know that Charlie has made the deposit,
before I deliver the merchandise. Theorem 1 below proves this link between
coordination and knowledge. Like all of the results quoted from previous
works in this section, it is expressed using the language L0.
Theorem 1 ([4]). Let OR =(cid:104)es, α1, . . . , αk(cid:105) be an instance of OR , and
assume that OR is solved in the system R. Let r ∈ R be a run in which es
occurs, let 1 ≤ h ≤ k, and let th be the time at which ih performs action ah
in r. Then
(R, r, th) (cid:15) KihKih−1 ··· Ki1occurred(es).
Given that nested knowledge can be shown to be a necessary prerequi-
site for ordering actions, what patterns of communication will provide such
knowledge gain? One way to ensure event ordering is to insist on a message
chain, linking Charlie to you and then to me. In fact, as Chandy and Misra
show, it is the only way to ensure this, in an asynchronous system. But the
synchronous setting allows for more flexibility in ensuring event ordering.
For example, we could have:
Charlie sends messages to both you and me, alerting us of the deposit.
Given that transmission times are bounded from above, upon getting a mes-
sage from Charlie I calculate how long before the message sent to you is
guaranteed to arrive, and deliver the merchandise only then.
The extra flexibility, when compared to asynchronous systems, is based
on the availability of guarantees on message delivery times. We formalize
these guarantees in the following way:
11
Definition 5 (Bound guarantee [5]). Fix r ∈ R. The bound guarantee
relation (cid:57)(cid:57)(cid:75) over nodes of r is the smallest relation satisfying the following
three conditions:
1. If t ≤ t(cid:48) then (cid:104)i, t(cid:105) (cid:57)(cid:57)(cid:75) (cid:104)i, t(cid:48)(cid:105);
2. If the network contains a channel (i, j) with weight maxij then (cid:104)i, t(cid:105) (cid:57)(cid:57)(cid:75)
3. If (cid:104)i, t(cid:105) (cid:57)(cid:57)(cid:75) (cid:104)h, t(cid:48)(cid:48)(cid:105) and (cid:104)h, t(cid:48)(cid:48)(cid:105) (cid:57)(cid:57)(cid:75) (cid:104)j, t(cid:48)(cid:105), then (cid:104)i, t(cid:105) (cid:57)(cid:57)(cid:75) (cid:104)j, t(cid:48)(cid:105).
(cid:104)j, t + maxij(cid:105); and
With most of the building blocks in place, we are almost ready to de-
scribe the Knowledge Gain Theorem, the technical result at the core of
our [5] paper. The theorem characterizes the communication pattern that is
necessary for nested knowledge gain in synchronous systems. Intuitively, it
shows that message chains still play an important part in information flow,
but the synchronous equivalent of a message chain is much more flexible -
since here agents can also use bound guarantees in order to ensure that a
message arrives at its destination. The resultant message chain abstraction,
called the centipede, is defined below.
Figure 1: A centipede
Definition 6 (Centipede [5]). Let r ∈ R, let ih ∈ P for 0 ≤ h ≤ k and let
t ≤ t(cid:48). A centipede for (cid:104)i0, . . . , ik(cid:105) in the interval [t, t(cid:48)] in r is a sequence
(cid:104)θ0, θ1, . . . , θk(cid:105) of nodes such that θ0 = (cid:104)i0, t(cid:105), θk = (cid:104)ik, t(cid:48)(cid:105), θ0 (cid:32) θ1 (cid:32) ··· (cid:32)
θk, and θh (cid:57)(cid:57)(cid:75) (ih, t(cid:48)) holds for h = 1, . . . , k − 1.
A centipede for (cid:104)i0, . . . , ik(cid:105) is depicted in Figure 1. It shows a message
chain connecting (i0, t) and (ik, t(cid:48)), and along this chain a sequence of "route
splitting" nodes θ1, θ2, etc. such that each θh can guarantee the arrival of
a message to ih by time t(cid:48). Such a message can serve to inform ih of the
12
hi0,tihi1,t0ihi2,t0ihik,t0ihik−1,t0iθ1θ2θk−1 occurrence of a trigger event at (cid:104)i0, t(cid:105), and as the set of previously made
guarantees gets shuffled on to the next splitting node, the last agent ik can
be confident that by time t(cid:48) all previous agents are already informed of the
occurrence.
Theorem 2 (Knowledge Gain Theorem [5]). Let r ∈ R. Assume that e is
an external event occurring at (cid:104)i0, t(cid:105) in r.
If (R, r, t(cid:48)) (cid:15) Kik Kik−1··· Ki0occurred(e), then there is a centipede for (cid:104)i0, . . . , ik(cid:105)
in the interval [t, t(cid:48)] in r.
Theorem 1 states that solving the ordering problem requires nested
knowledge of the occurrence of the trigger event to have been gained. Theo-
rem 2 then shows that such knowledge gain can only take place if the agents
are related by the centipede communication pattern.
Although stated in terms of our system R, where agents are assumed
to be following fip, under some further generalization both theorems can
be made to apply for all systems based on the synchronous context γmax,
regardless of the specific protocol.3
The ordering problem, nested knowledge and the centipede define a "ver-
tical stack", going from coordination, to knowledge, to communication and
centered about nested knowledge gain. We now examine another such ver-
tical stack, this time defined based on common knowledge. Common knowl-
edge has been associated with simultaneous action already in [11, 12]. We
touch upon this relation by defining the Simultaneous Response problem
and then using Theorem 3 below.
Definition 7 (Simultaneous Response [5]). Let es be an external input non-
deterministic event. A protocol P solves the instance SR = (cid:104)es, α1, . . . , αk(cid:105)
of the Simultaneous Response problem if it guarantees that
1. every response αh, for h = 1, . . . , k, occurs in a run iff the trigger
event es occurs in that run.
2. all of the responses α1, . . . , αk are performed simultaneously (i.e. t1 =
t2 = ··· = tk).
Theorem 3 ([5]). Let SR = (cid:104)es, α1, . . . , αk(cid:105), and assume that SR is solved
in R. Moreover, let G = {i1, . . . , ik} be the set of processes appearing in
the response set of SR. Finally, let r ∈ R be a run in which es occurs, and
let t be the time at which the response actions are performed in r. Then
(R, r, t) (cid:15) CGoccurred(es).
3 In order to generalize the theorems in this way we need to extend potential causality
into a relation called syncausality. See [5] for more.
13
Relating common knowledge to a necessary communication pattern is
more difficult, especially if you consider the formally-convenient classifi-
cation of common knowledge as an infinite conjunction of nested mutual
knowledge operators. Intuitively, borrowing the centipede from the nested
knowledge results quotes above, this would mean that an infinite centipede
is required in order for common knowledge to arise. However, it turns out
that the necessary communication pattern is quite finite, and more in line
with a fixed point view: In order for the group of agents G to gain common
knowledge that es has occurred, there must exist a single agent-time node θ
that is message chain related to the site of occurrence, and that can guaran-
tee that forwarded messages with information regarding es to all members of
G will arrive by t(cid:48). We call this pattern the broom. The following definition,
and Figure 2 below, should make it apparent why.
Figure 2: A broom
all
Definition 8 (Broom [5]). Let t ≤ t(cid:48) and G ⊆ P. Node θ is a broom for
(cid:104)i0, G(cid:105) in the interval [t, t(cid:48)] in r if (cid:104)i0, t(cid:105) (cid:32) θ and θ (cid:57)(cid:57)(cid:75) (cid:104)ih, t(cid:48)(cid:105) holds for
ih ∈ G.
Theorem 4 states the necessity connection between common knowledge
gain and the broom structure. As before, when we were dealing with nested
knowledge, Theorems 3 and 4 can be made to hold for all synchronous (γmax
based) systems, regardless of protocol.
Theorem 4 (Common Knowledge Gain [5]). Let G ⊆ P, and let r ∈
R. Assume that e is an external input event at (cid:104)i0, t(cid:105) in r. If (R, r, t(cid:48)) (cid:15)
CG(occurred(e)), then there is a broom θ for (cid:104)i0, G(cid:105) in interval [t, t(cid:48)] in r.
5 Nested Knowledge and Weak Bounds
We now re-approach the so-called "vertical stack" centered about nested
knowledge. This time we replace standard nested formulas such as KiKjKkϕ
14
hi0,tiθhi1,t0ihi2,t0ihi3,t0ihik,t0i with nb-formulations such as K(cid:104)i,t(cid:105)K(cid:104)j,t−4(cid:105)K(cid:104)k,t+8(cid:105)ϕ.
What communication pattern is required in order to attain such nested
knowledge? The following generalization of the centipede echos the above
mentioned break between temporal precedence and necessity.
Definition 9 (Uneven centipede). Let r ∈ R, let A = (cid:104)α0, α1, . . . , αk(cid:105)
be a sequence of nodes. An (uneven) centipede for A in r is a sequence
(cid:104)θ0, θ1, . . . , θk(cid:105) of nodes such that θ0 = α0, θk = αk, θ0 (cid:32) θ1 (cid:32) ··· (cid:32) θk,
and θh (cid:57)(cid:57)(cid:75) αh holds for h = 1, . . . , k − 1.
Figure 3 shows such an uneven centipede. It is termed uneven because
the "legs" of the centipede end at nodes at a variety of different times,
whereas in the traditional centipede all legs ended at nodes of time tk. More
interestingly, note that the node α2 temporally precedes α1 in Fig. 3. In-
tuitively though, this does not seem to concur with the epistemic status
of the two nodes, because information about the occurrence of the trigger
event at α0 flows through θ1 and θ2 to α2, along with θ1's guarantee that
by t1 agent i1 will have received word of the occurrence as well. Thus,
Kα2Kα1occurredt0(es) is attained by information flowing from a witness for
α1 to the node α2. Thus, in a precise sense, while α1 occurs temporally later
than α2, it "epistemically precedes" α2 in this run.
Figure 3: A uneven centipede
Theorem 5 shows that indeed the uneven centipede is necessary for nb-
nested knowledge gain. Interestingly, the proof argument is identical to that
which was used to prove the original Theorem 2. These proofs, originating
in [5], can be found in the Appendix. The only thing changed is the formal
language, and the proposed communication patterns. The same goes for
Theorem 7 below, which utilizes the same proof as Theorem 4, with an
update from L0 to L1. It is a case where our understanding of what is going
on has been limited purely by the expressivity of the formal apparatus of
which we made use.
15
α0θ1θ2θk−1α1α2αk−1αk Theorem 5. Let r ∈ R. Assume that es is an external event occurring at
α0 = (cid:104)i0, t0(cid:105) in r.
If (R, r) (cid:15) Kαk Kαk−1··· Kα0occurredt0(es), then there is an uneven centipede
for (cid:104)α0, . . . , αk(cid:105) in r.
We now turn to explore the implications of nb-nested knowledge for the
coordinated Ordered Response problem. In line with our above discussion
concerning nb-semntics and the decoupling of epistemic necessity and past
tense, we expect that nb-nested knowledge will allow for greater flexibility
in the timing of performed responses, solving tasks such as the following:
Once Charlie deposits the money, you will sign the contract, and I will
deliver the merchandise no more than 5 days after you sign.
This suggests the OR variant (cid:104)deposit,(cid:104)Y ou, sign(cid:105),(cid:104)I, deliver(cid:105)(cid:105), where
the required timing is such that tdeliver ≤ tsign+5. Assume that Charlie, You
and I act at nodes αCh, αY and αI respectively. By modifying Theorem 1,
we should be able to show that KαI KαY occurredtY (deposit) holds. Once we
have shown this, Theorem 5 will show that the required communication is
as shown in Figure 4a.
(a)
(b)
Figure 4
Perhaps surprisingly, it turns out that the required epistemic state for
ensuring the correct execution of the problem as defined, is actually
KαY KαI occurredtI (deposit)! Which in turn requires the inverted centipede
shown in Figure 4b. To see why, consider that the requirement tdeliver ≤
tsign +5 again. Seemingly an upper bound constraint upon me to deliver the
merchandise no more than 5 days after contract sign, it can also be read as a
lower bound constraint upon You to sign the contract no sooner than 5 days
before delivery takes place... The point is, however, that I cannot promise
to deliver the merchandise within the set time, since word of the signing
16
αChθYαYαItI≤tY+5αChθIαIαYtY≥tI−5 may arrive at my site a week after it occurs. Whereas You, no matter how
late (or how soon) you hear of my having delivered the merchandise, can
always wait a few days to ensure that the 5 days have passed.
The discussion concerning this timed coordination task points out a valu-
able insight. The knowledge that each agent must possess as it responds,
concerns those responses whose time of occurrence is already bounded from
above. These imply a lower bound for the agent's own action. In the first
proposed (informal) definition, it is actually the case that You have knowl-
edge of an upper limit upon my response. This brings about the reversal of
the required nested knowledge, in comparison to what we expected it to be
like. A task definition that exposes the implied upper bounds is in place.
We now provide a formal definition for the kind of coordination task
that is characterized by nb-nested knowledge. The definition is phrased
in accordance with the above discussion, and as Theorem 6 below shows,
solutions to the problem indeed require that the intuitively proper nested
knowledge holds. The required coordination is considered weakly timed,
reflecting the added existence of timing constraints, which only bound from
below. As we will see in the next section, the tightly timed task is similar,
but contains stronger bounds on relative timing of responses.
Definition 10 (Weakly Timed Response). Let es be an external input non-
deterministic event. A protocol P solves the instance
WTR = (cid:104)es, α1 : δ1, α2 : δ2, . . . , αk−1 : δk−1, αk(cid:105) of the Weakly Timed Re-
sponse problem if it guarantees that
1. every response αh, for h = 1, . . . , k, occurs in a run iff the trigger
event es occurs in that run.
2. in a run where response αh occurs at αh = (cid:104)ih, th(cid:105) for all h ≤ k, for
every such h we have that th+1 ≥ th + δh.
Before stating the theorem relating the WTR problem to nb-nested knowl-
edge, another nuance should be observed. The problem definition specifies
that even though agent ik may not know the exact time at which responses
are performed by other agents, it can work out an upper bound on the time
of responses carried out by agents i1 to ik−1. For example, response αk−1
gets carried out at tk−1 which no later than tk − δk−1. Response αk−2 is
then bounded with respect to αk by tk−2 ≤ tk−1 − δk−2 ≤ tk − δk−1 − δk−2,
etc. We will use
h = (cid:104)ih, tk
βk
h(cid:105)
where tk
h = tk −
17
k−1(cid:88)
j=h
δj
to denote this upper limit: the latest possible node at which response αh
gets carried out, given that response αk is performed at time tk. Note that
th ≤ tk
h is an upper bound on th. For the same reason
we also have that tk
, since
h since by definition tk
h ≤ tk+1
k−1(cid:88)
tk+1
h = tk+1 − δk −
h
j=h
δj = tk+1
k −
δj ≥ tk −
k−1(cid:88)
j=h
δj = tk
h.
k−1(cid:88)
j=h
k−1(cid:88)
j=h
βk
h = (cid:104)ih, tk −
δj(cid:105).
Theorem 6. Let WTR = (cid:104)es, α1 : δ1, α2 : δ2, . . . , αk−1 : δk−1, αk(cid:105) be an in-
stance of WTR, and assume that WTR is solved in the system R. Let r ∈ R
be a run in which es occurs. For each h ≤ k, let αh = (cid:104)ih, th(cid:105) be the node at
which response αh gets performed, and let
Then
(R, r) (cid:15) Kαk Kβk
k−1 ··· Kβk
1
occurredt1(es).
6 Common Knowledge and Tight Bounds
In this section we examine nb-common knowledge and its relation to com-
munication and coordination. Just as nb-nested knowledge requires an ex-
tension to the centipede structure, nb-common knowledge can only arise if
the uneven broom communication structure, seen in Figure 5 and defined
below, takes place in the run. ("Uneven" again comes from the broom's
uneven legs.)
Definition 11 (Uneven broom). Let A = {α1,...,αk} be a sequence of nodes
and let α0 be a node. Node θ is an (uneven) broom for (cid:104)α0, A(cid:105) in r if α0 (cid:32) θ
and θ (cid:57)(cid:57)(cid:75) αh holds for all h = 1, .., k.
Theorem 7 proves that the uneven broom characterizes the necessary
communication for nb-common knowledge to arise.
Theorem 7. Let A ⊆ V with (cid:104)ik, tk(cid:105) being the latest node (tk ≥ th for all
(cid:104)ih, th(cid:105) ∈ A) , and let r ∈ R. Assume that es is an external input event at
α0 in r. If (R, r) (cid:15) CA(occurredtk (es)), then there is a uneven broom θ for
(cid:104)α0, A(cid:105) in r.
18
Figure 5: A uneven broom
Recall the Simultaneous Response problem, that shows standard com-
mon knowledge as the necessary requirement when a group of agents need
to act simultaneously. Yet, as discussed in the introduction, node based
common knowledge does away with simultaneity while retaining other prop-
erties of common knowledge. In accordance, we expect it to serve as the
epistemic requirement for some weakened generalization of the simultane-
ity requirement. The following coordination task seems like a promising
candidate.
Definition 12 (Tightly Timed Response). Let es be an external input non-
deterministic event. A protocol P solves the instance
TTR = (cid:104)es, α1 : δ1, α2 : δ2, . . . , αk : δk(cid:105)
of the Tightly Timed Response problem if it guarantees that
1. every response αh, for h = 1, . . . , k, occurs in a run iff the trigger
event es occurs in that run.
2. For every h, g ≤ k the relative timing of the responses is exactly the
difference in the associated delta values: th − tg = δh − δg
Note how the new TTR problem definition generalizes the previous SR
definition: A simultaneous response problem SR = (cid:104)es, α1, . . . , αk(cid:105) is simply
an extremely tight TTR problem, where δ1, .., δk are all set to 0.
Intuitively, we expect that an agent i participating in a solution to a TTR
task by performing response αi at time ti will know that agent j will perform
its own response αj at precisely tj = δj − δi + ti. We also expect agent j
at tj to know that i is carrying out αi at ti. Theorem 8 below shows that
these individually specified knowledge states, and others that are derived
from the TTR definition, add up to a node based common knowledge gain
requirement.
19
α0θα1α2α3αk Theorem 8. Let TTR = TTR = (cid:104)es, α1 : δ1, α2 : δ2, . . . , αk : δk(cid:105), and as-
sume that TTR is solved in R. Let r ∈ R be a run in which es occurs, and
let A = {α1, . . . , αk} be the set of nodes at which the responses are carried
out in the run r (α1 occurs at node α1, etc.). WLOG, let αh = (cid:104)ih, th(cid:105) be
the earliest node in A. Then (R, r) (cid:15) CAoccurredth(es).
Summing up, the TTR problem requires agents to be fully informed with
respect to each other's response time - without the extra requirement for
simultaneity. Since each of the agents, as it responds, knows the response
times of all other agents - we end up with node based common knowledge,
which in turn still requires that a broom node exist that can guarantee that
all of its messages to the responding agents are delivered by the time these
agents are set to respond.
7 Conclusions
This paper explores the implications of a new formalism for epistemic state-
ments upon the notions of nested and common knowledge. It checks how
these altered concepts interact with communication on the one hand, and
coordination on the other - along the lines of [5, 4]. The new formalism
is seen to allow for a decoupling of epistemic necessity and the past tense,
in the sense that knowing that an event must occur or that an agent gains
knowledge of a fact no longer entails that the occurrence, or the knowledge
gain, have happened in the past. This, in turn, allows for a wider range
of coordination tasks to be characterized in terms of the epistemic states
that they necessitate. We define two such coordination tasks. The Weakly
Timed and the Tightly Timed Response problems, extending our previous
definitions of Ordered and Simultaneous Response [5, 4].
Our analysis of the relations between knowledge and coordination yields
two valuable insights. First, when agents must respond within interrelated
time bounds, as seen in the WTR problem, the crucial knowledge that they
must gain before applying their assigned actions concerns those responses
whose response times are bounded from above with respect to their own
response. This allows them, if necessary, to delay their own action in order
to conform with the problem requirements.
The enquiry into the node-based extension to common knowledge is even
more rewarding. Much has been written about the relation between com-
mon knowledge and simultaneity. Formal analysis of this relation is given
in [12, 11]. Other analyses implicitly rely upon simultaneity in accounting for
common knowledge gain, using such concepts as public announcements [3]
20
or the copresence heuristics [9]. The analysis presented here sheds light on
this issue by generalizing common knowledge and showing that the more
general form corresponds to a temporally tight form of coordination. The
previously established connection between common knowledge and simul-
taneity is, in fact, a particular instance of this more general connection.
Recall by Lemma 1 that there is an embedding of the standard common
knowledge CGoccurred(es) at time t into the formula CAoccurredt(es) ∈ L1
where all nodes in A are timed to t. The particular form of coordination
corresponding to the embedded formula is tight coordination with all deltas
set to 0 -- namely, simultaneity at time t.
We consider this paper as a point of departure for several lines of poten-
tial further research. First, we have only touched the surface of the required
logical analysis for node based semantics. Completeness and tractability is-
sues remain unknown, as well as possible variations on the node based theme
that will chime in with ongoing research in the temporal logic and model
checking communities. Second, although we have mainly been concerned
with temporally oriented coordination, the analysis may also be instructive
where other aspects of coordination are concerned. Chwe's [8] analysis of
the communication network required in order to bring about a revolt in a
social setting is a case in point for further research. Finally, we have yet to
study the implications of the WTR and TTR problems defined here in the
context of distributed computing tasks, where timed coordination is often
essential.
References
[1] R. Alur and T. Henzinger. Logics and models of real-time: A survey.
In Real Time: Theory in Practice, volume 600, pages 74 -- 106. Springer-
Verlag, 1992.
[2] C. Areces and B. ten Cate. Hybrid logics. In P. Blackburn, F. Wolter,
and J. van Benthem, editors, Handbook of Modal Logics. Elsevier, 2006.
[3] A. Baltag, L. S. Moss, and S. Solecki. The logic of public announce-
ments, common knowledge, and private suspicions. In Proceedings of
the 7th conference on Theoretical aspects of rationality and knowledge,
TARK '98, San Francisco, CA, USA, 1998. Morgan Kaufmann Publish-
ers Inc.
[4] I. Ben-Zvi. Causality, Knowledge and Coordination in Distributed Sys-
tems. PhD thesis, Technion, Israel, 2011.
21
[5] I. Ben-Zvi and Y. Moses. Beyond Lamport's happened-before: On the
role of time bounds in synchronous systems. In DISC 2010, pages 421 --
436, 2010.
[6] I. Ben-Zvi and Y. Moses. On interactive knowledge with bounded com-
munication (extended abstract). In LOFT 2010, 2010.
[7] K. M. Chandy and J. Misra. How processes learn. Distributed Com-
puting, 1(1):40 -- 52, 1986.
[8] M. S.-Y. Chwe. Communication and coordination in social networks.
Review of Economic Studies, 67(1), 2000.
[9] H. H. Clark and C. R. Marshall. Definite reference and mutual knowl-
In A. K. Joshi, B. L. Webber, and I. Sag, editors, Elements
edge.
of Discourse Understanding, pages 10 -- 63. Cambridge University Press,
Cambridge, 1981.
[10] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Common knowl-
edge revisited. In Y. Shoham, editor, Theoretical Aspects of Rationality
and Knowledge: Proc. Sixth Conference, pages 283 -- 298. Morgan Kauf-
mann, San Francisco, Calif., 1996.
[11] R. Fagin, J. Y. Halpern, Y. Moses, and M. Y. Vardi. Reasoning about
Knowledge. MIT Press, Cambridge, Mass., 2003.
[12] J. Y. Halpern and Y. Moses. Knowledge and common knowledge in a
distributed environment. Journal of the ACM, 37(3):549 -- 587, 1990. A
preliminary version appeared in Proc. 3rd ACM Symposium on Princi-
ples of Distributed Computing, 1984.
[13] L. Lamport. Time, clocks, and the ordering of events in a distributed
system. Communications of the ACM, 21(7):558 -- 565, 1978.
22
A Proofs
This appendix contains all lemmas and major proofs necessary to support
the claims of the paper. Note that the formal work leading up to Theorems 5
and 7 is non-trivial, and the interested reader would be better off reading
the more detailed account presented in [5, 6, 4].
Lemma 1. There exists a function ts : L0 × T ime → L1 such that for every
ϕ ∈ L0 , time t, and ϕt = ts(ϕ, t):
(R, r, t) (cid:15) ϕ
(R, r) (cid:15) ϕt.
iff
Proof Proof is by structural induction on ϕ. We go over the clauses. Each
clause is used to define ts, but it is also easy to see that based on the clause
definition (R, r, t) (cid:15) ϕ iff (R, r) (cid:15) ϕt.
• ts(occurred(e), t) = occurredt(e).
• ts(Kiψ, t) = K(cid:104)i,t(cid:105)ψ.
• ts(EGψ, t) = EAψ where A = {(cid:104)j, t(cid:105) : j ∈ G}.
• ts(CGψ, t) = CAψ where A = {(cid:104)j, t(cid:105) : j ∈ G}.
(cid:4)
Lemma 2.
• Both K(cid:104)i,t(cid:105) and CA are S5 modalities.
• CAϕ ⇒ EA(ϕ ∧ CAϕ) is valid.
• If R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ) then R (cid:15) ϕ ⇒ CAψ
Proof
For S5 modality the K(cid:104)i,t(cid:105) case is immediate. We focus on showing for CA:
K (CAϕ ∧ CA(ϕ ⇒ ψ) ⇒ CAψ): Suppose that (R, r) (cid:15) CAϕ∧CA(ϕ ⇒
ψ) ∧ ¬CAψ for some r, ϕ, ψ and A. Then there is a sequence
α1, α2, .., αn ∈ A such that (R, r) (cid:50) KαnKαn−1 ··· Kα1ψ. From
(R, r) (cid:15) CAϕ and (R, r) (cid:15) CA(ϕ ⇒ ψ) we obtain (R, r) (cid:15)
KαnKαn−1 ··· Kα1ϕ and (R, r) (cid:15) KαnKαn−1 ··· Kα1(ϕ ⇒ ψ) re-
and trivial induction we get that (R, r) (cid:15) KαhKαh−1 ··· Kα1ψ for
all h ≤ n, and in particular for h = n. This contradicts the
assumption that (R, r) (cid:15) CAϕ ∧ CA(ϕ ⇒ ψ) ∧ ¬CAψ.
spectively, and using the K axiom for the knowledge operator
23
(R, r) (cid:50) Kαn ··· Kα1Kα(cid:48)
ϕ.
m ··· Kα(cid:48)
1
T (CAϕ ⇒ ϕ): Suppose that (R, r) (cid:15) CAϕ for some r, ϕ and A. Fix
some (cid:104)i, t(cid:105) ∈ A. From (R, r) (cid:15) CAϕ we get (R, r) (cid:15) K(cid:104)i,t(cid:105)ϕ, and
by definition of (cid:15) we conclude that (R, r(cid:48)) (cid:15) ϕ for all r(cid:48) such that
r(cid:48)i(t) = ri(t), and in particular for r(cid:48) = r. Thus (R, r) (cid:15) ϕ.
4 (CAϕ ⇒ CACAϕ): Suppose not. Then there exist r, A, ϕ such that
(R, r) (cid:15) CAϕ and a sequence α1, α2, .., αn ∈ A such that (R, r) (cid:50)
KαnKαn−1 ··· Kα1CAϕ. Once again, there must be a sequence
α(cid:48)1, α(cid:48)2, .., α(cid:48)m ∈ A such that
Yet this contradicts (R, r) (cid:15) CAϕ.
5 (¬CAϕ ⇒ CA¬CAϕ): Suppose not. Then there exist r, A, ϕ such
that (R, r) (cid:15) ¬CAϕ and a sequence α1, α2, .., αn ∈ A such that
(R, r) (cid:50) KαnKαn−1 ··· Kα1¬CAϕ. By definition of (cid:15) there exists
r(cid:48) s.t. rn
in(tn) = rin(tn) and where
(R, rn) (cid:50) Kαn−1Kαn−1 ··· Kα1¬CAϕ.
i1(t1) = r2
i1(t1), r2
we have (i) r1
Similar argumentation provides us with runs rh for all 1 ≤ h < n
such that (R, rh) (cid:50) Kαh−1Kαh−2 ··· Kα1¬CAϕ. In particular for
h = 1 we get that (R, r1) (cid:50) ¬CAϕ, and hence (R, r1) (cid:15) CAϕ,
and finally that (R, r1) (cid:15) KαnKαn−1 ··· Kα1CAϕ. Yet by build
i2(t2) = r3
i2(t2), etc., giving us
(R, r) (cid:15) CAϕ, contradicting our assumption.
N (If (cid:15) ϕ then (cid:15) CAϕ): We show by induction on h that for any h
length sequence α1, α2, .., αh ∈ A we have (cid:15) KαhKαh−1 ··· Kα1ϕ.
For h = 0 this is immediate from the assumption. Assume for h.
Fix run r. Note that for every run r(cid:48) s.t. rih(th) = r(cid:48)ih(th) we have
(R, r(cid:48)) (cid:15) Kαh−1 ··· Kα1ϕ by the inductive hypothesis. Hence we
get that (R, r) (cid:15) KαhKαh−1 ··· Kα1ϕ. As this is true for any r we
conclude that (cid:15) KαhKαh−1 ··· Kα1ϕ as required. By definition of
CAϕ we get that (cid:15) CAϕ.
CAϕ ⇒ EA(ϕ ∧ CAϕ): Fix r, ϕ, A. From (R, r) (cid:15) CAϕ we get by definition
of CA that (i) (R, r) (cid:15) Kαϕ and (ii) (R, r) (cid:15) KαCAϕ for all α ∈ A.
Hence we also get (iii) (R, r) (cid:15) EAϕ and (iv) (R, r) (cid:15) EACAϕ. Putting
(iii) and (iv) together we get (R, r) (cid:15) EA(ϕ ∧ CAϕ).
24
If R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ) then R (cid:15) ϕ ⇒ CAψ: Suppose R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ).
Fix a sequence α1, α2, .., αk ∈ A. We will now show by induction
on h ≤ kthat R (cid:15) ϕ ⇒ Kαk Kαk−1 ··· Kα1ψ. For h = 1 we have by
assumption that R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ). We replace EA by Kα1, re-
calling that α1 ∈ A. We get R (cid:15) ϕ ⇒ Kα1ψ. Assume for h and
look at the case of h + 1. By assumption R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ)
and hence R (cid:15) ϕ ⇒ Kαh+1(ϕ ∧ ψ). Fix run r.
If (R, r) (cid:50) ϕ then
(R, r) (cid:15) ϕ ⇒ Kαh+1Kαh ··· Kα1ψ and we are done. Else, we have
that (R, r) (cid:15) EA(ϕ ∧ ψ), and so (R, r) (cid:15) Kαh+1(ϕ ∧ ψ). Thus, we
get that (R, r(cid:48)) (cid:15) ϕ for every r(cid:48) such that rih+1(th+1) = r(cid:48)ih+1(th+1).
By the inductive hypothesis we get that (R, r(cid:48)) (cid:15) Kαh ··· Kα1ψ. By
definition of (cid:15) we get that (R, r) (cid:15) Kαh+1Kαh ··· Kα1ψ. We conclude
that R (cid:15) ϕ ⇒ Kαk Kαk−1 ··· Kα1ψ. As this holds for any sequence of
nodes in A we get that R (cid:15) ϕ ⇒ CAψ as required.
(cid:4)
Definition 13 (Past cone). Fix r ∈ R and θ ∈ V. The past cone of θ in r,
past(r, θ), is the set {ψ : ψ (cid:32) θ}.
Definition 14 (Agree upon). We say that two runs r and r(cid:48) agree on the
node (i, t) ∈ V if
1. ri(t) = r(cid:48)i(t),
2. the same external inputs and messages arrive at (i, t) in both runs,
and
3. the same actions are performed by i at time t.
Lemma 3. Let r ∈ R and let θ ∈ V. Then there is a run r(cid:48) ∈ R such that
1. past(r(cid:48), θ) = past(r, θ),
2. r(cid:48) and r agree on all the nodes of past(r(cid:48), θ).
3. the only nondeterministic events in r(cid:48) occur at nodes of past(r(cid:48), θ).
Lemma 4. If (cid:104)i, t(cid:105) (cid:32) (cid:104)j, t(cid:48)(cid:105) then t ≤ t(cid:48), with t = t(cid:48) holding only if i = j.
Definition 15 (Early delivery). When a message sent at (cid:104)i, t(cid:105) arrives at
(cid:104)j, t(cid:48)(cid:105) prior to the maximal allowed delay, i.e. when t(cid:48) < t + maxij, then we
say that an early delivery nondeterministic event has occurred.
25
Lemma 5. Fix a run r, and let θ (cid:32) θ(cid:48). If θ (cid:54)(cid:57)(cid:57)(cid:75) θ(cid:48) then there is a node β
such that θ (cid:32) β (cid:57)(cid:57)(cid:75) θ(cid:48) and an early delivery occurs at β in r.
Lemma 6. Let e be the delivery of an external input that occurs at the node
α0 in r ∈ R. If (R, r) (cid:15) Kα1occurredt1(e) then α0 (cid:32) α1.
Theorem 5 (Knowledge Gain). Let r ∈ R. Assume that e is an external
input event occurring at α0 in r.
If (R, r) (cid:15) Kαk Kαk−1··· Kα1occurredt1(e), then there is an uneven centipede
for (cid:104)α0, . . . , αk(cid:105) in r.
Proof First note that k ≥ 1 and (R, r) (cid:15) Kαk Kαk−1 ··· Kα1occurredt1(e)
imply by the Knowledge Axiom that (R, r) (cid:15) Kαk occurredt1(e). It follows
that α0 (cid:32) αk in r by Lemma 6. We prove the claim by induction on k ≥ 1:
k = 1 As argued above, α0 (cid:32) αk. Thus, α0 (cid:32) α1 since k = 1, and so
(cid:104)α0, αk(cid:105) is a (trivial) centipede for (cid:104)α0, αk(cid:105) in r.
k ≥ 2 Assume inductively that the claim holds for k − 1. Moreover, as-
sume that (R, r) (cid:15) Kαk Kαk−1 ··· Kα1occurredt1(e). Let r(cid:48) be the run
guaranteed by Lemma 3 to exist with respect to r, ik and tk. Re-
call from Lemma 3(1) that past(r(cid:48), αk) = past(r, αk). Thus, α0 (cid:32) αk
in r implies that α0 (cid:32) αk in r(cid:48) too. Moreover, by Lemma 3(2) we
have that r and r(cid:48) agree on the nodes of past(r, αk), so in particular
r(cid:48)ik (tk) = rik (tk). Since (R, r) (cid:15) Kαk Kαk−1 ··· Kα1occurredt1(e) and
r(cid:48)ik (tk) = rik (tk), we have that
(R, r(cid:48), t(cid:48)) (cid:15) Kαk−1 ··· Kα1occurredt1(e).
By the inductive hypothesis there exists a centipede (cid:104)α0, θ1, . . . , θk−1(cid:105)
for (cid:104)α0, . . . , αk−1(cid:105) in r(cid:48). Let c ≥ 0 be the minimal index for which
θc (cid:57)(cid:57)(cid:75) θh for all h = c + 1, . . . , k − 1. Clearly c ≤ k − 1, since θk−1 (cid:57)(cid:57)(cid:75)
θk−1.
• If c = 0 then α0 (cid:57)(cid:57)(cid:75) θh (cid:57)(cid:57)(cid:75) αh, and thus also α0 (cid:57)(cid:57)(cid:75) αh, for
h = 1, . . . , k − 1. Since α0 (cid:32) αk in r, it follows that the tuple
(cid:104)(α0)d−1, αk(cid:105) (in which α0 plays the role of the first k − 1 nodes)
is a centipede for (cid:104)α0, . . . , αk(cid:105) in r.
• Otherwise, c > 0 and θc−1 (cid:32) θc while θc−1 (cid:54)(cid:57)(cid:57)(cid:75) θc. By Lemma 5 it
follows that there exists a node β such that θc−1 (cid:32) β (cid:57)(cid:57)(cid:75) θc, and
β is the site of an early receive in the run r(cid:48). By construction of r(cid:48),
early receives can arrive only at nodes in past(r(cid:48), αk). It follows
26
that β (cid:32) αk in r(cid:48), and since past(r, αk) = past(r(cid:48), αk), we have
that β (cid:32) αk in r too and hence also that past(r, β) = past(r(cid:48), β).
It follows that in r
α0 (cid:32) θ1 (cid:32) ··· (cid:32) θc−1 (cid:32) β (cid:32) αk.
Recall that (cid:57)(cid:57)(cid:75) depends only on the weighted communication net-
work, which is the same in both r and r(cid:48). Thus, θj (cid:57)(cid:57)(cid:75) αj for all
0 < j ≤ c − 1. Moreover, β (cid:57)(cid:57)(cid:75) θh (cid:57)(cid:57)(cid:75) αh and so β (cid:57)(cid:57)(cid:75) αh for all
h = c, c + 1, . . . , k− 1. It follows that (cid:104)α0, θ1, . . . , θc−1, (β)k−c, αk(cid:105)
is a centipede for (cid:104)α0, . . . , αk(cid:105) in r.
It follows that a centipede for (cid:104)α0, . . . , αk(cid:105) in r is guaranteed to exist
in all cases, as claimed.
2T heorem 5
Theorem 6. Let WTR=(cid:104)es, α1 : δ1, . . . , αk−1 : δk−1, αk(cid:105) be an instance of
WTR, and assume that WTR is solved in the system R. Let r ∈ R be a run
in which es occurs. For each h ≤ k, let αh = (cid:104)ih, th(cid:105) be the node at which
response αh gets performed, and let
k−1(cid:88)
j=h
βk
h = (cid:104)ih, tk −
δj(cid:105).
Then
(R, r) (cid:15) Kαk Kβk
k−1 ··· Kβk
1
occurredtk
1
(es).
Proof We prove the theorem by induction on h ≤ k.
h = 1 : By definition of r, α1 gets performed at α1. Since performing
a local action is written in the agent's local state (and hence known
to the agent), and since it is always performed no sooner than the
then we are done. Else, as tk
occurredtk
triggering event es, we have (R, r) (cid:15) Kα1occurredt1(es). If h = k = 1
1 ≥ t1 and as agents have perfect recall,
we get (R, r) (cid:15) Kβk
(es).
h > 1 : Assume that (R, r) (cid:15) Kβk
sponse αh gets performed at αh in r we have (R, r) (cid:15) Kαhoccurredth
(ah).
Since WTR is solved in the system R we have that in every runr(cid:48) such
that rih(th) = r(cid:48)ih(th) response αh−1 gets performed at some th−1 such
(es). As re-
h−1 ··· Kβk
1
occurredtk
1
h
1
1
27
that th−1 ≤ th
h−1. Note that every system that solves WTR also solves
the sub problem WTR'=(cid:104)es, α1 : δ1, . . . , αh−2 : δh−2, αh−1(cid:105). Based on
the inductive hypothesis we obtain that
(R, r(cid:48)) (cid:15) Kαh−1Kβk
h−2 ··· Kβk
1
occurredtk
1
(es).
From tk
h−1 ≥ th−1, and based on perfect recall, we derive that
(R, r(cid:48)) (cid:15) Kβk
h−1
Kβk
h−2 ··· Kβk
1
occurredtk
1
(es).
By our choice of runs r(cid:48) and the definition of (cid:15) this gives us
(R, r(cid:48)) (cid:15) KαhKβk
h−1
Kβk
h−2 ··· Kβk
1
occurredtk
1
(es).
If h = k then we are done. Else, once again deploying tk
conclude that
h ≥ th we
(R, r(cid:48)) (cid:15) Kβk
h
Kβk
h−1
and we are done.
Kβk
h−2 ··· Kβk
1
occurredtk
1
(es),
2T heorem 6
Theorem 7. Let A ⊆ V with (cid:104)jk, t(cid:48)k(cid:105) being the earliest node (t(cid:48)k ≤ t(cid:48)h for
all (cid:104)jh, t(cid:48)h(cid:105) ∈ A) , and let r ∈ R. Assume that e is an external input event
at α0 in r. If (R, r) (cid:15) CA(occurredt(cid:48)
(e)), then there is a uneven broom θ for
(cid:104)α0, A(cid:105) in r.
Proof Assume the notations and conditions of the theorem. Denote A =
{α1, . . . , αk} and d = t(cid:48)k − t0, the time difference between the occurrence of
e and the latest node in A . Since (R, r) (cid:15) CA(occurredt(cid:48)
(e)) we have by
definition of common knowledge that
(R, r) (cid:15) Ek(d+1)
(e). In particular, this implies that
A
k
k
k
occurredt(cid:48)
(R, r) (cid:15) (Kαk ··· Kα1)d+1occurredt(cid:48)
k
(e),
where (Kαk ··· Kα1)d+1 stands for d + 1 consecutive copies of Kαk ··· Kα1.
By the Knowledge Gain Theorem 5, there is a corresponding centipede σ =
(cid:104)θ0, θ1, . . . , θk(d+1)(cid:105) in r. Denote θh = (ih, th) for all 0 ≤ h ≤ k·(d+1). Recall
that, by definition, θh (cid:32) θh+1 holds for all h < k· (d + 1). By Lemma 4
we obtain that if θh (cid:54)= θh+1 then th < th+1. It follows that there can be at
28
(cid:32) θ(cid:48)2
(cid:32) ··· (cid:32) θ(cid:48)(cid:96) in σ. Every θ(cid:48)h represents a
most d + 1 distinct nodes θ(cid:48)1
segment θx, . . . , θx+s of the nodes in σ. By the pigeonhole principle, one of
the θ(cid:48) nodes must represent a segment consisting of at least k of the original
θs in σ. Denoting this node by θ, we obtain that θ (cid:57)(cid:57)(cid:75) αh for every αh ∈ A.
Moreover, by definition of the centipede and transitivity of (cid:32) we have that
α0 (cid:32) θ. It follows that θ is a centibroom for (cid:104)α0, G(cid:105) in r.
2T heorem 7
Theorem 8. Let TTR = (cid:104)es, α1 : δ1, . . . , αk : δk(cid:105), and assume that TTR is
solved in R. Let r ∈ R be a run in which es occurs, and let A = {α1, . . . , αk}
be the set of nodes at which the responses are carried out in the run r (α1
occurs at node α1 = (cid:104)i1, t1(cid:105), etc.). Let α(cid:48) = (cid:104)i(cid:48), t(cid:48)(cid:105) be the earliest node in A.
Then (R, r) (cid:15) CAoccurredt(cid:48)(es).
Proof Fix h, g ∈ {1..k}. We first show that
R (cid:15) occurredth(ah) ⇒ EA(occurredth(ah) ∧ occurredt(cid:48)(es)).
Choose r(cid:48) such that (R, r(cid:48)) (cid:15) occurredth(ah). Note that since TTR is solved
in R and since th− tg = δh− δg in every triggered run by problem definition,
we get (R, r(cid:48)) (cid:15) occurredth(ah) ↔ occurredtg (ag). Since performing a local
action is written, at least for the current round, in the agent's local state
(and hence known to the agent), we obtain that (R, r(cid:48)) (cid:15) Kαg occurredtg (ag),
and hence also (R, r(cid:48)) (cid:15) Kαg occurredth(ah). Since h is arbitrarily chosen in
{1..k}, node αg knows this for all responding nodes, and in particular for the
earliest responding node α(cid:48) = (cid:104)i(cid:48), t(cid:48)(cid:105). As responses always occur no sooner
than the trigger, we get that (R, r(cid:48)) (cid:15) Kαg occurredt(cid:48)(es). Putting these re-
sults together we conclude that (R, r(cid:48)) (cid:15) Kig (occurredth(ah)∧occurredt(cid:48)(es)).
Since g is arbitrarily chosen in {1..k}, we get (R, r(cid:48)) (cid:15) EA(occurredth(ah) ∧
occurredt(cid:48)(es)), from which it follows that
(R, r(cid:48)) (cid:15) occurredth(ah) ⇒ EA(occurredth(ah) ∧ occurredt(cid:48)(es))
by our choice of r(cid:48). As false antecedents imply anything, we also get
(R, r(cid:48)(cid:48)) (cid:15) occurredth(ah) ⇒ EA(occurredth(ah) ∧ occurredt(cid:48)(es))
in runs r(cid:48)(cid:48) where response αh does not occur (non triggered runs) or where it
occurs at a later time than th. We thus conclude that R (cid:15) occurredth(ah) ⇒
EA(occurredth(ah) ∧ occurredt(cid:48)(es)).
R (cid:15) ϕ ⇒ CAψ from R (cid:15) ϕ ⇒ EA(ϕ ∧ ψ). Setting ϕ = occurredth(ah) and
Recall the Knowledge Induction Rule in Lemma 2, that provides us with
29
ψ = occurredt(cid:48)(es) we apply the rule, and based on the above result obtain
R (cid:15) occurredth(ah) ⇒ CAoccurredt(cid:48)(es). We conclude by taking notice that
(R, r) (cid:15) occurredth(ah) by assumption, and hence also (R, r) (cid:15) CAoccurredt(cid:48)(es).
2T heorem 8
30
|
1907.10384 | 2 | 1907 | 2019-07-31T09:31:16 | Towards automatic estimation of conversation floors within F-formations | [
"cs.MA",
"cs.CV",
"cs.CY"
] | The detection of free-standing conversing groups has received significant attention in recent years. In the absence of a formal definition, most studies operationalize the notion of a conversation group either through a spatial or a temporal lens. Spatially, the most commonly used representation is the F-formation, defined by social scientists as the configuration in which people arrange themselves to sustain an interaction. However, the use of this representation is often accompanied with the simplifying assumption that a single conversation occurs within an F-formation. Temporally, various categories have been used to organize conversational units; these include, among others, turn, topic, and floor. Some of these concepts are hard to define objectively by themselves. The present work constitutes an initial exploration into unifying these perspectives by primarily posing the question: can we use the observation of simultaneous speaker turns to infer whether multiple conversation floors exist within an F-formation? We motivate a metric for the existence of distinct conversation floors based on simultaneous speaker turns, and provide an analysis using this metric to characterize conversations across F-formations of varying cardinality. We contribute two key findings: firstly, at the average speaking turn duration of about two seconds for humans, there is evidence for the existence of multiple floors within an F-formation; and secondly, an increase in the cardinality of an F-formation correlates with a decrease in duration of simultaneous speaking turns. | cs.MA | cs | 2019 8th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW)
Towards automatic estimation of conversation floors
within F-formations
Chirag Raman
Delft University of Technology
[email protected]
Hayley Hung
Delft University of Technology
[email protected]
9
1
0
2
l
u
J
1
3
]
A
M
.
s
c
[
2
v
4
8
3
0
1
.
7
0
9
1
:
v
i
X
r
a
Abstract -- The detection of free-standing conversing groups has
received significant attention in recent years. In the absence of
a formal definition, most studies operationalize the notion of
a conversation group either through a spatial or a temporal
lens. Spatially, the most commonly used representation is the
F-formation, defined by social scientists as the configuration
in which people arrange themselves to sustain an interaction.
However, the use of this representation is often accompanied
with the simplifying assumption that a single conversation occurs
within an F-formation. Temporally, various categories have been
used to organize conversational units; these include, among
others, turn, topic, and floor. Some of these concepts are hard to
define objectively by themselves. The present work constitutes an
initial exploration into unifying these perspectives by primarily
posing the question: can we use the observation of simultaneous
speaker turns to infer whether multiple conversation floors exist
within an F-formation? We motivate a metric for the existence
of distinct conversation floors based on simultaneous speaker
turns, and provide an analysis using this metric to characterize
conversations across F-formations of varying cardinality. We
contribute two key findings: firstly, at the average speaking turn
duration of about two seconds for humans, there is evidence
for the existence of multiple floors within an F-formation;
and secondly, an increase in the cardinality of an F-formation
correlates with a decrease in duration of simultaneous speaking
turns.
Index Terms -- free-standing conversational groups, conversa-
tion floors, speaking turns
I. INTRODUCTION
Imagine a social scenario like a mingling or networking
event. Interactions in such a setting involve multiple dynamic
conversations which are a medley of ever evolving topics and
partners. And yet, humans can instinctively navigate the com-
plexities of such encounters. How do we do this? We regulate
our exchanges both spatially and temporally using implicit
social norms or explicit behavioural signals [1]. Furthermore,
these cues could be either verbal or non-verbal, expressed
visually, vocally, or verbally through spoken language.
A deeper understanding of these group dynamics constitutes
a natural objective towards the realisation of machines with so-
cial skills. For instance, consider a social robot approaching a
group of people in a public space, or the use-case of evaluating
attendee experience at a conference poster session. In these
and other cases, having an understanding of the dynamics,
and where channels of social influence lie, would enable the
This research was partially funded by the Netherlands Organization for
Scientific Research (NWO) under project number 639.022.606.
978-1-7281-3891-6/19/$31.00 ©2019 IEEE
Fig. 1. Depiction of a single F-formation with multiple conversation floors.
The darker green regions within dotted lines represent distinct simultaneous
conversation floors. Most works representing a conversing group as an F-
formation make the simplifying assumption that a single conversation occurs
within an F-formation with a joint focus of attention for all members.
artificial agent to develop increasingly sophisticated policies
for interaction or inference. Conversation groups have been of
importance in the application domains of social robotics [2] --
[5], activity recognition [6], [7], social surveillance [8] -- [10],
and social signal processing [11], [12].
Fundamental to the study of such conversations is defining
the notion of a free-standing conversational group (FCG).
While it is easier to objectively conceptualize an FCG in
spatial terms in a scene of multiple interacting groups, delin-
eating the boundary of conversations poses a greater technical
challenge. We could think of separating conversations on
the basis of topics, but this is challenging if audio data is
unavailable due to privacy concerns. We could operationalize
a conversation as a set of participating members, but this
membership is challenging to infer visually for non-speaking
participants. This often leads to the simplifying assumption
in some literature that
the focus of an FCG is a single
conversation. As we illustrate in Fig. 1, and discuss in the
following sections, this may not always be the case.
In the present work, we dive beyond the geometric bounds
of an FCG to gain a deeper understanding of the conversations
occuring within it. In this initial approach, we focus specifi-
cally on speaking participants as the most decisive indicator
of the existence of a conversation. Concretely, we pose the
following broad research questions:
RQ 1. Can we use observed speaker turns to infer the con-
versation floors within an F-formation?
RQ 2. How does the cardinality of an F-formation affect the
conversation floors developed within it?
The ground truth for speaker turns in this work comes from
manual annotations of video data, mimicking use-cases where
audio data might be unavailable due to privacy concerns.
Concretely, our contributions are as follows: conceptually, we
provide an indicator of distinct conversation floors that uses
speaking turns alone, and situate this indicator in schisming
literature [13] -- [15]; analytically, we provide evidence that
multiple conversation floors exist within an F-formation, and
show that the cardinality of an F-formation correlates nega-
tively with turn duration of simultaneous speakers.
The rest of this paper is organized as follows. We describe
some of the spatial and temporal perspectives used to study
FCGs in Section II. In Section III we provide a review of
literature involving the use of these spatial or temporal notions,
motivating the need to consider both of these aspects in unison.
In Section IV, we propose an operationalization of an indicator
of distinct conversation floors, building upon the concepts of
conversation schisming. The dataset we use and the exper-
iments performed for answering the research questions are
described in Section V and Section VI respectively. Finally,
Section VII summarizes our findings and concludes the paper.
II. BACKGROUND
Spatial Factors. One of the most common proxemic notions
to describe an FCG is Adam Kendon's Facing Formation, or
F-formation, originally defined as:
An F-formation arises whenever two or more people
sustain a spatial and orientational relationship in
which the space between them is one to which they
have equal, direct, and exclusive access. [16, p. 210]
Kendon argues that activity is always located, and denotes
the space in front of a person that is used for the activity
as the person's transactional segment. When two or more
people come together to perform some activity, they are liable
to arrange themselves such that their individual transactional
segments overlap to create a joint transactional space. This
joint space between the interactants is called an o-space. As
we discuss in the next section, many computational works
involving the automatic detection of FCGs from video focus
on the detection of F-formations, often assuming that
the
transaction involves a single conversation.
Temporal Factors. The conversation of focus in an FCG,
however, is dynamic in nature. If conversations change over
time, what are the temporal units that describe their underlying
structure? Some of the terms used in early literature to orga-
nize conversational units include turn, topic, gap, and floor.
Edelsky provides an excellent review of these concepts in [17],
stating that most of these units were defined on the basis
of some technical or mechanical structure such as signals of
speakers or auditors, ignoring the intention of the participant.
Using inferred participants' meanings rather than technical
definitions, Edelsky defines turns and floors as follows:
The floor is defined as the acknowledged what's-
going-on within a psychological time/space. What's
going on can be the development of a topic or
a function (teasing, soliciting a response, etc.) or
an interaction of the two. It can be developed or
controlled by one person at a time or by several
simultaneously or in quick succession.[17, p. 405]
III. RELATED WORK
Detecting Conversational Groups. In most works, a con-
versational group is operationalized as an F-formation. Early
work on the task of detecting FCGs in video data developed
concurrently from two perspectives: those that estimate the
location of the o-space using a Hough-voting strategy [8],
[18]; or those that view an F-formation as a set with individ-
uals being assigned exclusive membership [12], [19]. There
has also been considerable work focused on incorporating
temporal
information for the same task of detecting con-
versational groups [7], [20] -- [22]. Notably, these approaches
utilise the head pose as a proxy for Visual Focus of Attention
(VFoA) [9] in addition to the body pose to model F-formation
membership, and assume a single conversation within an F-
formation. The assumption that members in a group have a
single joint focus of attention is seen in other works as well.
Hung, Jayagopi, Ba, et al. [23] model a single joint focus
of visual attention of participants to estimate dominance in
groups.Vazquez, Steinfeld, and Hudson Vazquez, Steinfeld,
and Hudson also assume a single conversation within an F-
formation while developing a policy for a robot to be aware
of a single focus of attention of the conversation.
Estimating involvement. In a conversation, the floor is typ-
ically held by a single participant at a time [13]. What then
characterizes the silent participants in a conversation group?
The following works demonstrate that the task of estimating
participant involvement is subjective in nature, and that gaze
behaviour and turn-taking patterns can be informative.
Zhang and Hung [24], [25] study the task of detecting
associates of an F-formation; members that are attached to
an F-formation but do not have full status [16]. They argue
that the labeling of conversation groups is not an objective
task. Collecting multiple annotations of perceived associates,
they demonstrate how detecting them can improve initial esti-
mates of full-members of an F-formation. Oertel, Funes Mora,
Gustafson, and Odobez [26] characterize silent particiants
into multiple categories (attentive listener, side participant,
bystander) from audiovisual cues. Oertel and Salvi [27] also
show that it is possible to estimate individual engagement and
group involvement in a multiparty corpus by analysing the
participants' eye-gaze patterns. Bohus and Horvitz [28] pro-
pose a self-supervised method for forecasting disengagement
with an interactive robot using a conservative heuristic. The
heuristic is constructed by leveraging features that capture how
close the participant is, whether a participant is stationary or
moving, and whether a participant is attending to the robot.
shown promising results in estimating the presence of voice
activity from automated gestural analysis or accelerometer
data [33] -- [35]. It therefore seems feasible that speaker turns
can be automatically estimated without audio data. Combined
with the observation that the co-existence of two turn-taking
systems is the most decisive characteristic of schisming, we
argue that it is in turn reasonable to explore the inference of
schisms without audio data through speaking turns.
it
is useful
Linking schisming to floors and F-formations. While Egbert
does explicitly use the term floor to describe the conversations
resulting from a schism,
to observe how this
relates back to Edelsky's view of floors. Edelsky defined
floors in terms of the acknowledged what's-going-on within a
psychological time space. The object of focus here could either
be a topic or some other function. To borrow Goffman's terms,
a schism effectively changes a gathering into a multifocused
one, where each object of focus can be viewed to correspond
to a floor in Edelsky's definition. However, if the participant's
lower bodies remain configured such that their transactional
segments overlap to produce a common o-space, they would
still remain in the same F-formation even if the conversation
has undergone a schism into two or more distinct floors. Fig. 1
depicts this situation conceptually.
Combining these two broad observations, we argue that it is
feasible to explore the existence of distinct conversation floors
within an F-formation without audio data, whilst capturing
speaker turns from visual observations. We propose to start
with the following metric. Given a sliding window w of speak-
ing duration d, we consider a speaker to be a participant who
speaks for the entire duration d. The number of simultaneous
speakers thus defined corresponds to the number of distinct
conversation floors at that position of w, since they correspond
to speaking turns in distinct floors.
Of course, the metric is inextricably tied to the duration d
being considered; too short a duration, and the concurrent turns
might capture either backchannels or the overlapping speech
within the same floor as described in Egbert's work. However,
a reasonably long duration would capture the speaking turns of
participants holding distinct floors. This leads to the question:
what qualifies as a reasonable choice for d to differentiate
overlaps within a floor from turns in distinct floors? In our
Some works also used turn-taking features to estimate
some notion of involvement. Pentland, Madan, and Gips [29]
measured engagement by the z-scored influence each person
has on the other's turn-taking for a pair of participants. Hung
and Gatica-Perez [30] found that the pause duration between
an individual's turns, aggregated at group level,
is highly
predictive of cohesion in small group meetings.
Schisming. In a conversation with at least four participants,
the conversation sometimes splits up into two or more conver-
sations. This transformation is referred to as a schism [13] or
schisming. One of the earliest allusions to the phenomenon of
schisming based on anecdotal evidence occurs in the work of
Goffman, who suggested that a gathering of two participants
exhausts an encounter and forms a fully-focused gathering [31,
p. 91]. With more than two participants, there may be persons
officially present in the situation who are not themselves so
engaged. These bystanders change the gathering into a partly-
focused one. If more than three persons are present, there may
be more than one encounter carried on in the same situation,
resulting in a multifocused gathering.
In subsequent work, Sacks, Schegloff, and Jefferson [13]
and Goodwin [14] both indicated that the co-existence of two
turn-taking systems is the most decisive characteristic of schis-
ming. This view was supported by Egbert, who demonstrated
that although schisming is a participation framework with
two simultaneous conversations, each with its own turn-taking
system, there is an interface between them during schisming
[15]. She also makes a systematic differentiation between
overlap and simultaneous talk during schisming. In overlap,
simultaneous speakers compete for the floor, an event usually
resolved by returning to one-speaker-at-a-time. In schisming
by contrast, simultaneous speakers orient to one of two distinct
floors, an event which if resolved successfully, results in the
establishment of two floors [15, p. 43]. Overlapping speech
is therefore expected to occur throughout the lifespan of all
conversation floors within an F-formation.
IV. METHODOLOGY
In this section we build upon the previously discussed
concepts to propose using simultaneous speakers in an F-
formation as an initial conservative indicator of the existence
of distinct conversation floors.
A common concern with observing groups of conversing
people is the potential violation of privacy. In our experience
with collecting group interaction datasets, participants often
regard having their microphone data recorded and transcribed
as being more invasive than being captured on video. In these
situations, the lack of verbal information makes it extremely
challenging to infer the topics being discussed. How can we
then investigate the existence of distinct conversations? Two
observations could prove useful:
Inferring schisms without audio data. The relationship
between body movements such as gestures and speech has
been long established in literature [32]. Some works have
Fig. 2.
Illustration of gaps, within-overlaps, and between-overlaps for two
speakers (S1 and S2) within the same floor. The scheme was originally
proposed by Heldner and Edlund[36] and adopted by Levinson and Torreira
in their analysis[37].
experiments, we set the lower bound of d at one second. Here
we provide evidence from literature to justify this choice.
Choice of speaking window duration. In a study of gaps
and overlaps in conversations, Heldner and Edlund report that
on average 40% of the speaker transitions in their corpora
involved overlaps (including any overlap of over 10 ms) [36].
These represent overlaps for competing for the floor. As for the
duration of these overlaps, their histogram makes clear that the
duration follows a mode of 50 ms in the Spoken Dutch Corpus,
with a mean of 610 ms, and median of 470 ms, all under one
second. In a follow-up detailed statistical analysis, Levinson
and Torreira differentiate between types of overlaps: between-
overlaps, that refer to overlaps where the floor was transfered
without a silent gap between speakers; and within-overlaps,
where overlapping speech occured in between a speaking
turn and did not result in a transfer of floor [37]. Fig. 2
illustrates these types of overlaps. They used the Switchboard
Corpus of English telephone conversations for their analysis,
and found that only 3.8% of the signal corresponded to
simultaneous speech of both speakers. This fits well with
Sacks and colleagues' observations that "overwhelmingly, one
party speakes at a time" [13, p. 700], for physically situated
embodied social interactions. As for the duration, between-
overlaps had a modal duration of 96 ms, a median of 205 ms, a
mean of 275 ms. On the other hand, within-overlaps exhibited
an estimated modal duration of 350 ms, a median of 389 ms,
a mean of 447 ms. Further, of all the overlaps annotated, 73%
involved a backchannel. These statistics indicate that choosing
a lower bound for d would reasonably capture simultaneous
speech that does not belong to the same floor.
As for the upper bound, a reasonable value should be at
least greater than the average turn duration of a speaker. Using
the same operationalization proposed in [36], Levinson and
Torreira report that contiguous speech delimited by a silent
interval of at least 180 ms had a mean duration of 1680 ms,
and a median of 1227 ms.
Fig. 3. Snapshots of the mingling session (Cameras 1-5) in MatchNMingle.
mingling interaction were recorded for each of the three days;
56 minutes on the first, 50 minutes on the second, and 45
minutes on the third, respectively.
Annotations. The dataset provides of annotations for both F-
formations and a variety of social actions. The F-formations
were annotated directly from a video of the interacting par-
ticipants captured from overhead cameras. The annotations
were made for every second for an interval of 10 minutes
per day. Each F-formation annotation provides the participant
IDs for its members and the start and end times delimiting
the lifetime of the F-formation. In all, 174 F-formations
were annotated across 30 minutes. Of these, we filtered out
those with cardinality less than four, and those for which a
participant was found to leave the field of view of the cameras.
This left us with 34 F-formations for our experiments.
Of the social actions annotated, we only use the Speaking
Status -- defined as whether or not a person is speaking. The
social actions were annotated for a 30 minute segment for each
day, by eight annotators hired for the task and trained by an
expert. The annotations were made at the frame level using
a tool that allowed for interpolation across frames. In all, 20
annotations per second for each social action are provided.
Further, the speaking status is estimated from video alone, by
observing lip movements or inferring from the participants'
head and body gestures.
V. DATASET
VI. EXPERIMENTS
For this study, we use the publicly available MatchNMingle
dataset [38] that records in-the-wild interactions of 92 people
during speed-dates followed by a cocktail party. Three sessions
of speed-dates and mingling were recorded in all across three
days. We specifically focus on the cocktail party recordings
that capture free standing conversations between participants.
Fig.3 shows the video recordings from five cameras on the last
day of data collection. The participants were not given a script
to follow and were free to choose the participants they wished
to interact with. This allows us to study naturally evolving
F-formations and conversation floors in an in-the-wild setting.
Dataset Statistics. The dataset consists of a total of 92 single,
heterosexual participants (46 women: 19-27 years with a mean
age of 21.6 years and standard deviation of 1.9 years; and
46 men: 18-30 years with a mean age of 22.6 years and
standard deviation of 2.6 years). Over 45 minutes of free
We perform two sets of experiments: first we identify the
number of simultaneous speakers in an F-formation using
the methodology described in Section IV, and then evaluate
whether the number of members in an F-formation (cardinal-
ity) affects the speaking duration of simultaneous speakers.
Simultaneous Speakers in an F-formation. The purpose of
this experiment is to evaluate the following -- can we infer the
existence of distinct conversation floors within an F-formation
from simultaneous speaker turns? To recap, this intuition build
upon early work on schisming indicating that the co-existence
of two turn-taking systems is the most decisive characteristic
of distinct conversation floors [13], [14]. Here we consider F-
foramtions of cardinality four and above, since the possibility
of distinct conversations occurs only for those F-formations.
We slide a window w of duration d across the lifetime of the
F-formation in steps of one second. For every position of w,
Fig. 4. Plotting the effect of varying the speaking duration threshold d on the number of simultaneous speakers per cardinality of F-formation. To aggreagate
the data from each F-formation, the maximum of the number of simultaneous speakers is considered over all the sliding window positions across the lifetime
of the F-formation. The y-axis plots the Mean (Maximum number of simultaneous speakers over window positions) over F-formations.
we count the number of participants with a positive speaking
status for the entire duration d. We plot the maximum number
of simultaneous speakers over all positions of w. Following
the formulation described in Section IV, this represents the
maximum number of distinct conversation floors that were
observed during the life-time of the F-formation. We vary d
from 1-20 seconds to guard against the possibility that the
smaller values of d might capture co-narration or overlaps
within the same floor. The upper bound of 20 seconds was
chosen as sanity check; we expected to see very few speakers
have a speaking turn that long.
The max operator was chosen to aggregate the number of
simultaneous speakers across all window positions into the
most conservative measure for what this experiment seeks to
evaluate. A value of one for the maximum number speakers
over all positions of w would indicate that only a single
conversation floor existed within the F-formation. Therefore,
observing values greater than one for the max metric would
indicate the presence of distinct floors with more certainty than
other choices of summarizing statistics.
Fig. 4 plots the mean number of distinct conversation floors
per F-formation against varying values of d, per cardinality of
F-formation. Cardinality here refers to the number of members
in an F-formation. As a sanity check, we would expect the
numbers upper-bounded by the number of people in the F-
formation; at worst, every person in the F-formation speaks
simultaneously to compete for the floor they are a part of.
On the same note, we observe that the starting mean values
all seem reasonable: about 2 for cardinalities four and five,
about 3 for cardinality six, and about 4 for cardinality seven.
Assuming that it is common for speakers to have at least one
conversing partner, we would expect about half the number
of simultaneous speakers as members in an F-formation. Our
minimum choice of d was chosen to be greater than the modal
duration of overlaps found in previous work [37], so it is less
likely that the lower turn durations capture competing overlaps
for the same floor. Moreover, at the average turn length of
about two seconds observed by Levinson and Torreira [37], we
observe that the maximum number of simultaneous speakers is
greater than one at all cardinalities considered. This suggests
that the simplifying assumption from previous research of a
single conversation within an F-formation is insufficient.
We also observe a decreasing trend for the curves in Fig. 4.
This seems intuitive, as it is much less likely that participants
would speak for the entire duration of a window as d increases.
Interestingly, there is a single example of a speaker speaking
for 20 seconds in an F-formation of cardinality seven. On
closer inspection, this turned out to be an error in speaking
status annotation, and we manually fixed this error for subse-
quent analysis.
Effect of cardinality on turn duration of simultaneous
speakers. Sacks, Schegloff, and Jefferson observed that there
is a "pressure for minimization of turn size, distinctively
operative with three or more parties" [13, p. 713]. They
note that the possibility of a schism introduced by the fourth
participant may influence the turn-taking system by 'spreading
the turns around' if there is an interest in retaining participants
in the conversation. However, they concede that this effect
is equivocal, since turn distribution can also be used for
encouraging schisming. In this experiment, we explore this
effect and pose the question as follows: for a given speaking
turn duration d, do we observe a decrease in the maximum
number of conversation floors observed over an F-formation's
lifetime with an increase in the cardinality of an F-formation?
Qualitatively, this corresponds to the steepness of fall-off
of the curves in Fig. 4. It seems that
the the curves for
cardinality six and seven falloff more steeply than those for
cardinalities four and five. To quantitatively test if cardinality
has an effect, we fit a Generalized Linear Model (GLM) to the
same data as in the previous experiment with an interaction
factor between cardinality and the speaking turn duration d.
Specifically, we assume the maximum number of simultaneous
speakers observed over the lifetime of each F-formation, yi to
GENERALIZED LINEAR MODEL REGRESSION RESULTS
TABLE I
Intercept
Turn-duration
Cardinality
Turn-duration:Cardinality
Coef (β)
0.0626
0.0057
0.1869
-0.0025
Std Err
0.339
0.002
0.072
0.001
z
0.184
2.296
2.603
-4.543
P>z
0.854
0.022
0.009
0.000006
NOMINAL P-VALUES FOR SIX POST-HOC GLM REGRESSION
TABLE II
COMPARISONS
Cardinality Pairs
4-5
4-6
4-7
5-6
5-7
6-7
Intercept (β0)
0.196
0.364
0.697
0.079
0.434
0.275
d (β1)
0.855
0.0007
0.428
0.0008
0.413
0.006
c (β2)
0.794
0.010
0.030
0.016
0.043
0.657
d:c (β3)
0.403
0.00002*
0.009
0.00016*
0.052
0.024
d = turn-duration, c = cardinality, d:c = interaction-factor. βs
denote the corresponding regression coefficients. * denotes sig-
nificance at a threshold of 0.001 after Bonferroni correction for
six tests.
be realizations of independent Poisson random variables, with
Yi ∼ P (µi) and model µi as follows:
log(µi) = β0 + β1 ∗ di + β2 ∗ ci + β3 ∗ di ∗ ci
(1)
where di refers to the duration of the speaking window,
and ci refers to the cardinality for the ith observation. The βs
refer to the regression coefficients. The GLM was fit using
the statsmodels python package. The results of the GLM
regression test are provided in Table I. We conclude that
cardinality and the two-way interaction between cardinality
and turn duration are statistically significant at a significance
level of 0.01. Turn duration is itself significant at a significance
level of 0.05.
tells us that
While the previous test
turn duration and
cardinality are significant, we still need to perform post-hoc
comparisons to ascertain the differences between the cardinali-
ties. We fit multiple GLMs to each possible pair of cardinalities
being considered and correct the corresponding p-values using
the Bonferroni correction for multipe testing. Table II provides
the corrected p-values for the post-hoc comparisons. From the
last column, we find that cardinality and its interaction with
turn-duration are significant between the cardinalities {4, 6},
and {5, 6} at a significance level of 0.001.
One potential limitation of this analysis is the imbalance
in the number of F-formations of different cardinalities. F-
formations of cardinality four were the most common in the
data, with reasonable number of samples to infer a pattern.
We believe that the intuition of cardinality and its interaction
with speaking turn duration being significant is still a sound
intuition, although the statistical significance should perhaps
be viewed within the context of the number of F-formations
we see in the data. Fig. 5 plots the number of observations
that contributed to the graphs in Fig. 4.
VII. CONCLUSION
In this study, we presented an initial exploration into
unifying the spatial and temporal perspectives of a free-
standing conversing group. Specifically, we proposed using
simultaneous speaking turns as an indicator for the existence
of distinct conversation floors. In the absence of audio data to
identify the topics being discussed, our proposed metric can
Fig. 5. Number of F-formations at different speaking turn durations.
be used to gain a deeper understanding of the conversation
dynamics within an F-formation, since speaking turns can be
inferred from visual or wearable-sensor data. Our experiments
demonstrate that at an average turn duration of two seconds for
humans [37], there is evidence of multiple conversation floors
within a single F-formation. Further, we found that an increase
in cardinality of an F-formation correlates with a decrease in
turn duration of simultaneous speakers, specifically between
F-formations of sizes {4,6}, and {5,6} in our data. A deeper
analysis would be required to identify whether the differences
in F-formations of cardinality six hold across datasets, with
preferably more examples of F-fomrations of size six and
greater. In this initial approach to the problem, our study does
not account for the behaviour of the silent participants, or the
evolution of turn taking dynamics within a floor. These remain
promising avenues to explore for future works.
ACKNOWLEDGMENT
Chirag Raman thanks Stavros Makrodimitris, Madhumita
Sushil, Giovanni Cassani, Erik B. van den Akker, and Yesh-
wanth Napolean for their time and thoughtfulness.
REFERENCES
[1] A. Vinciarelli, M. Pantic, and H. Bourlard, "Social signal processing:
Survey of an emerging domain," Image and Vision Computing, Nov. 1,
2009.
[2] H. Huettenrauch, K. S. Eklundh, A. Green, and E. A. Topp, "In-
vestigating spatial relationships in human-robot interaction," in 2006
IEEE/RSJ International Conference on Intelligent Robots and Sys-
tems, Oct. 2006.
[3] M. Vazquez, A. Steinfeld, and S. E. Hudson, "Parallel detection of
conversational groups of free-standing people and tracking of their
lower-body orientation," IEEE, Sep. 2015.
[4] -- -- , "Maintaining awareness of the focus of attention of a con-
versation: A robot-centric reinforcement learning approach," in 2016
25th IEEE International Symposium on Robot and Human Interactive
Communication (RO-MAN), IEEE, Aug. 2016.
in Proceedings of the 2015 ACM on International Conference on
Multimodal Interaction - ICMI '15, ACM Press, 2015.
[27] C. Oertel and G. Salvi, "A gaze-based method for relating group
involvement to individual engagement in multimodal multiparty dia-
logue," in Proceedings of the 15th ACM on International conference
on multimodal interaction - ICMI '13, ACM Press, 2013.
[28] D. Bohus and E. Horvitz, "Managing human-robot engagement with
forecasts and... um ... hesitations," in Proceedings of the 16th Interna-
tional Conference on Multimodal Interaction - ICMI '14, ACM Press,
2014.
[29] A. Pentland, A. Madan, and J. Gips, "Perception of social interest,"
in Proceedings of the 5th international conference on devlopment
and learning ICDL 2006, Department of Psychological and Brain
Sciences, Indiana University, 2006.
[30] H. Hung and D. Gatica-Perez, "Estimating cohesion in small groups
using audio-visual nonverbal behavior," IEEE Transactions on Multi-
media, Oct. 2010.
E. Goffman, Behavior in public places: Notes on the social organi-
zation of gatherings, 1. paperback ed., 24. printing. New York, NY:
The Free Press, 1966.
[32] D. McNeill, Language and gesture. Cambridge University Press,
[31]
2000.
[34]
[33] H. Hung and S. O. Ba, "Speech/non-speech detection in meetings
from automatically extracted low resolution visual features," in 2010
IEEE International Conference on Acoustics, Speech and Signal
Processing, IEEE, 2010.
E. Gedik and H. Hung, "Speaking status detection from body
movements using transductive parameter transfer," in Proceedings
of the 2016 ACM International Joint Conference on Pervasive and
Ubiquitous Computing: Adjunct, ACM, 2016.
[35] -- -- , "Personalised models for speech detection from body move-
ments using transductive parameter transfer," Personal and Ubiquitous
Computing, Aug. 1, 2017.
[36] M. Heldner and J. Edlund, "Pauses, gaps and overlaps in conversa-
[37]
[38]
tions," Journal of Phonetics, Oct. 1, 2010.
S. C. Levinson and F. Torreira, "Timing in turn-taking and its impli-
cations for processing models of language," Frontiers in Psychology,
Jun. 12, 2015.
L. Cabrera-Quiros, A. Demetriou, E. Gedik, L. v. d. Meij, and H.
Hung, "The MatchNMingle dataset: A novel multi-sensor resource
for the analysis of social interactions and group dynamics in-the-wild
during free-standing conversations and speed dates," IEEE Transac-
tions on Affective Computing, 2018.
[5] M. Vazquez, E. J. Carter, B. McDorman, J. Forlizzi, A. Steinfeld, and
S. E. Hudson, "Towards robot autonomy in group conversations: Un-
derstanding the effects of body orientation and gaze," in Proceedings
of the 2017 ACM/IEEE International Conference on Human-Robot
Interaction - HRI '17, ACM Press, 2017.
[6] W. Choi and S. Savarese, "A unified framework for multi-target
tracking and collective activity recognition," in European Conference
on Computer Vision, Springer Berlin Heidelberg, 2012.
[7] K. Tran, A. Gala, I. Kakadiaris, and S. Shah, "Activity analysis in
crowded environments using social cues for group discovery and
human interaction modeling," Pattern Recognition Letters, Jul. 2014.
[8] M. Cristani, L. Bazzani, G. Paggetti, A. Fossati, D. Tosato, A. D. Bue,
G. Menegaz, and V. Murino, "Social interaction discovery by statis-
tical analysis of f-formations," British Machine Vision Association,
2011.
[9] M. Farenzena, A. Tavano, L. Bazzani, D. Tosato, G. Paggetti, G.
Menegaz, V. Murino, and M. Cristani, "Social interactions by visual
focus of attention in a three-dimensional environment," Expert Sys-
tems, May 2013.
L. Bazzani, M. Zanotto, M. Cristani, and V. Murino, "Joint individual-
group modeling for tracking," IEEE Transactions on Pattern Analysis
and Machine Intelligence, Apr. 1, 2015.
[11] G. Groh, A. Lehmann, J. Reimers, M. R. Friess, and L. Schwarz,
"Detecting social situations from interaction geometry," in 2010 IEEE
Second International Conference on Social Computing, Aug. 2010.
[12] H. Hung and B. Krse, "Detecting f-formations as dominant sets,"
ACM Press, Proceedings of the 13th international conference on
multimodal interfaces, 2011.
[13] H. Sacks, E. A. Schegloff, and G. Jefferson, "A simplest systematics
for the organization of turn-taking for conversation," Studies in the
organization of conversational interaction, 1974.
[14] C. Goodwin, "Forgetfulness as an interactive resource," Social Psy-
[10]
chology Quarterly, 1987.
[15] M. M. Egbert, "Schisming: The collaborative transformation from a
single conversation to multiple conversations," Research on Language
& Social Interaction, Jan. 1997.
[16] A. Kendon, Conducting interaction: Patterns of behavior in focused
encounters. Cambridge University Press, 1990.
[22]
[21]
[19]
[17] C. Edelsky, "Who's got the floor?" Language in Society, 1981.
[18]
F. Setti, O. Lanz, R. Ferrario, V. Murino, and M. Cristani, "Multi-
scale f-formation discovery for group detection," IEEE International
Conference on Image Processing, Sep. 2013, pp. 3547 -- 3551, ISBN:
978-1-4799-2341-0.
F. Setti, C. Russell, C. Bassetti, and M. Cristani, "F-formation de-
tection: Individuating free-standing conversational groups in images,"
PLOS ONE, May 21, 2015.
[20] X. Alameda-Pineda, Y. Yan, E. Ricci, O. Lanz, and N. Sebe, "Ana-
lyzing free-standing conversational groups: A multimodal approach,"
ACM Press, Proceedings of the 23rd ACM international conference
on Multimedia, 2015.
S. Vascon, E. Z. Mequanint, M. Cristani, H. Hung, M. Pelillo, and
V. Murino, "Detecting conversational groups in images and sequences:
A robust game-theoretic approach," Computer Vision and Image
Understanding, Feb. 2016.
E. Ricci, J. Varadarajan, R. Subramanian, S. R. Bulo, N. Ahuja, and
O. Lanz, "Uncovering interactions and interactors: Joint estimation
of head, body orientation and f-formations from surveillance videos,"
in 2015 IEEE International Conference on Computer Vision (ICCV),
IEEE, Dec. 2015.
[23] H. Hung, D. B. Jayagopi, S. Ba, J.-M. Odobez, and D. Gatica-Perez,
"Investigating automatic dominance estimation in groups from visual
attention and speaking activity," ACM Press, International conference
on Multimodal interfaces, 2008.
L. Zhang and H. Hung, "Beyond f-formations: Determining social
involvement in free standing conversing groups from static images,"
IEEE Proceedings of the Conference on Computer Vision and Pattern
Recognition, Jun. 2016.
[25] -- -- , "On social involvement in mingling scenarios: Detecting asso-
ciates of f-formations in still images," IEEE Transactions on Affective
Computing, 2018.
[26] C. Oertel, K. A. Funes Mora, J. Gustafson, and J.-M. Odobez,
"Deciphering the silent participant: On the use of audio-visual cues
for the classification of listener categories in group discussions,"
[24]
|
1812.01132 | 1 | 1812 | 2018-12-03T23:20:52 | A Unified Approach to Dynamic Decision Problems with Asymmetric Information - Part II: Strategic Agents | [
"cs.MA",
"cs.GT",
"math.OC"
] | We study a general class of dynamic games with asymmetric information where agents' beliefs are strategy dependent, i.e. signaling occurs. We show that the notion of sufficient information, introduced in the companion paper team, can be used to effectively compress the agents' information in a mutually consistent manner that is sufficient for decision-making purposes. We present instances of dynamic games with asymmetric information where we can characterize a time-invariant information state for each agent. Based on the notion of sufficient information, we define a class of equilibria for dynamic games called Sufficient Information Based Perfect Bayesian Equilibrium (SIB-PBE). Utilizing the notion of SIB-PBE, we provide a sequential decomposition of dynamic games with asymmetric information over time; this decomposition leads to a dynamic program that determines SIB-PBE of dynamic games. Furthermore, we provide conditions under which we can guarantee the existence of SIB-PBE. | cs.MA | cs | A Unified Approach to Dynamic Decision
Problems with Asymmetric Information -
Part II: Strategic Agents
Hamidreza Tavafoghi, Yi Ouyang, and Demosthenis Teneketzis
Abstract
We study a general class of dynamic games with asymmetric information where agents' beliefs are
strategy dependent, i.e. signaling occurs. We show that the notion of sufficient information, introduced
in the companion paper [2], can be used to effectively compress the agents' information in a mutually
consistent manner that is sufficient for decision-making purposes. We present instances of dynamic
games with asymmetric information where we can characterize a time-invariant information state for
each agent. Based on the notion of sufficient information, we define a class of equilibria for dynamic
games called Sufficient Information Based Perfect Bayesian Equilibrium (SIB-PBE). Utilizing the notion
of SIB-PBE, we provide a sequential decomposition of dynamic games with asymmetric information
over time; this decomposition leads to a dynamic program that determines SIB-PBE of dynamic games.
Furthermore, we provide conditions under which we can guarantee the existence of SIB-PBE.
I. INTRODUCTION
We study a general class of stochastic dynamic games with asymmetric information. We
consider a setting where the underlying system has Markovian dynamics controlled by the agents'
joint actions at every time. The instantaneous utility of each agent depends on the agents' joint
8
1
0
2
c
e
D
3
]
A
M
.
s
c
[
1
v
2
3
1
1
0
.
2
1
8
1
:
v
i
X
r
a
A preliminary version of this paper will appear in the Proceeding of the 57th IEEE Conference on Decision and Control
is with the Department of Mechanical Engineering at
(CDC), Miami Beach, FL, December 2018 [1].
H. Tavafoghi
the University of California, Berkeley (e-mail:
[email protected]). Y. Ouyang is with Preferred Networks America, Inc. (e-mail: [email protected]). D.
Teneketzis is with the Department of Electrical Engineering and Computer Science at the University of Michigan, Ann Arbor
(e-mail: [email protected])
This work was supported in part by the NSF grants CNS-1238962, CCF-1111061, ARO-MURI grant W911NF-13-1-0421, and
ARO grant W911NF-17-1-0232.
November 23, 2018
DRAFT
2
actions and the system state. At every time, each agent makes a private noisy observation that
depends on the current system state and the agents' past actions. Therefore, at every time agents
have asymmetric and imperfect information about the history of the game. Furthermore, the
information that an agent possesses about the history of the game at each time instant depends
on the other agents' past actions and strategies; this phenomenon is known as signaling among the
agents. Moreover, at each time, each agent's strategy depend on his information about the current
system state and the other agents' strategies. Therefore, the agents' decisions and information
are coupled and interdependent over time.
There are three main challenges in the study of dynamic games with asymmetric information.
First, since the agents' decisions and information are interdependent and coupled over time, we
need to determine the agents' strategies simultaneously for all times. Second, the agents' strategy
domains grow over time as they acquire more information. Third, in contrast to dynamic teams
where agents coordinate their strategies, in dynamic games each agent's strategy is his private
information as he chooses it individually so as to maximize his utility. Therefore, in dynamic
games each agent needs to form a belief about other agents' strategies as well as about the game
history.
In this paper, we propose a general approach for the study of a dynamic games with asymmetric
information that addresses the stated-above challenges. We build our approach based on notion
of sufficient information, introduced in the companion paper [2], and define a class of sufficient
information based assessments, where strategic agents compress their information in a mutually
consistent manner that is sufficient for decision-making purposes. Accordingly, we propose the
notion of Sufficient Information Based Perfect Bayesian Equilibrium (SIB-PBE) for dynamic
games that characterizes a set of equilibrium outcomes. Using the notion of SIB-PBE, we provide
a sequential decomposition of the game over time, and formulate a dynamic program that enables
us to compute the set of SIB-PBEs via backward induction. We discover specific instances of
dynamic games where we can determine a set of information states for the agents that have
time-invariant domain. We determine conditions that guarantee the existence of SIB-PBEs. We
discuss the relation between the class of SIB-PBE and PBE in dynamic games, and argue that
the class of SIB-PBE provides a simpler and more robust set of equilibria than PBE that are
consistent with agents' rationality.
The notion of SIB-PBE we introduce in this paper provides a generalization/extension of
Markov Perfect Equilibrium (MPE) to dynamic games with asymmetric information. The authors
November 23, 2018
DRAFT
3
in [3] introduce the notion of Markov Perfect Equilibrium that characterizes a subset of Subgame
Perfect Equilibria (SPE) for dynamic games with symmetric information and provide a sequential
decomposition of the game over time. Moreover, our results, along with those in the companion
paper [2], provide a unified approach to the study of dynamic decision problems with asymmetric
information and strategic or nonstrategic agents.
A. Related Literature
Dynamic games with asymmetric information have been investigated extensively in literature
in the context of repeated games; see [4] -- [7] and the references therein. The key feature of
these games is the absence of a dynamic system. Moreover, the works on repeated games study
primarily their asymptotic properties when the horizon is infinite and agents are sufficiently
patient (i.e. the discount factor is close one). In repeated games, agents play a stage (static)
game repeatedly over time. As a result, the decision making problem that each agent faces is
very simple. The main objective of this strand of literature is to explore situations where agents
can form self-enforcing punishment/reward mechanisms so as to create additional equilibria that
improve upon the payoffs agents can get by simply playing an equilibrium of the stage game over
time. Recent works (see [8] -- [10]) adopt approaches similar to those used in repeated games to
study infinite horizon dynamic games with asymmetric information when there is an underlying
dynamic Markovian system. Under certain conditions on the system dynamics and information
structure, the authors of [8] -- [10] characterize a set of asymptotic equilibria when the agents are
sufficiently patient.
The problem we study in this paper is different from the ones in [4] -- [10] in two aspects. First,
we consider a class of dynamic games where the underlying system has a general Markovian
dynamics and a general information structure, and we do not restrict attention to asymptotic
behaviors when the horizon is infinite and the agents are sufficiently patient. Second, we study
situations where the decision problem that each agent faces, in the absence of strategic interac-
tions with other agents, is a Partially Observed Markov Decision Process (POMDP), which is
a complex problem to solve by itself. Therefore, reaching (and computing) a set of equilibrium
strategies, which take into account the strategic interactions among the agents, is a very chal-
lenging task. As a result, it is not very plausible for the agents to seek reaching an equilibria
that is generated by the formation of self-enforcing punishment/reward mechanisms similar to
those used in infinitely repeated games (see Section VII for more discussion). We believe that
November 23, 2018
DRAFT
4
our results provide new insight into the behavior of strategic agents in complex and dynamic
environments, and complement the existing results in the repeated games literature with simple
and (mostly) static environments.
The works in [11] -- [14] consider dynamic zero-sum games with asymmetric information. The
authors of [11], [12] study zero-sum games with Markovian dynamics and lack of information
on one side (i.e. one informed player and one uninformed player). The authors of [13], [14]
study zero-sum games with Markovian dynamics and lack of information on both sides. The
problem that we study in this paper is different from the ones in [11] -- [14] in three aspects.
First, we study a general class of dynamic games that include dynamic zero-sum games with
asymmetric information as a special case. Second, we consider a general Markovian dynamics for
the underlying system whereas the authors of [11] -- [14] consider a specific Markovian dynamics
where each agent observes perfectly a local state that evolves independently of the other local
states conditioned on the agents' observable actions. Third, we consider a general information
structure that allows us to capture scenarios with unobservable actions and imperfect observations
that are not captured in [11] -- [14].
The problems investigated in [15] -- [20] are the most closely related to our problem. The authors
of [15], [16] study a class of dynamic games where the agents' common information based belief
(defined in [15]) is independent of their strategies; that is, there is no signaling among them.
This property allows them to apply ideas from the common information approach developed
in [21], [22], and define an equivalent dynamic game with symmetric information among the
fictitious agents. Consequently, they characterize a class of equilibria for dynamic games called
Common Information based Markov Perfect Equilibrium. Our results are different from those in
[15], [16] in two aspects. First, we consider a general class of dynamic games where the agents'
CIB beliefs are strategy-dependent, thus, signaling is present. Second, the proposed approach
in [15], [16] requires the agents to keep track of all of their private information over time. We
propose an approach to effectively compress the agents' private information, and consequently,
reduce the number of variables on which the agents need to form CIB beliefs.
The authors of [17] -- [20] study a class of dynamic games with asymmetric information
where signaling occurs. When the horizon in finite, the authors of [17], [18] introduce the
notion of Common Information Based Perfect Bayesian Equilibrium, and provide a sequential
decomposition of the game over time. The authors of [19], [20] extend the results of [17], [18]
to finite horizon Linear-Quadratic-Gaussian (LQG) dynamic games and infinite horizon dynamic
November 23, 2018
DRAFT
5
games, respectively. The class of dynamic games studied in [17] -- [20] satisfies the following
assumptions: (i) agents' actions are observable (ii) each agent has a perfect observation of his
own local states/type (iii) conditioned on the agents' actions, the evolution of the local states
are independent.
We relax assumptions (i)-(iii) of [17] -- [20], and study a general class of dynamic games with
asymmetric information, hidden actions, imperfect observations, and controlled and coupled
dynamics. As a result, each agent needs to form a belief about the other agents' past actions
and private (imperfect) observations. Moreover, in contrast to [17] -- [20], an agent's, say agent
i's, belief about the system state and the other agents' private information is his own private
information and is different from the CIB belief. In this paper, we extend the methodology
developed in [17], [18] for dynamic games, and generalize the notion of CIB-PBE. Furthermore,
we propose an approach to effectively compress the agents' private information and obtain the
results of [17] -- [20] as special cases.
B. Contribution
We develop a general methodology for the study and analysis of dynamic games with asym-
metric information, where the information structure is non-classical; that is, signaling occurs.
We propose an approach to characterize a set of information states that effectively compress
the agents' private and common information in a mutually consistent manner. We characterize a
subclass of Perfect Bayesian Equilibria, called SIB-PBE, and provide a sequential decomposition
of these games over time. This decomposition provides a backward induction algorithm to
determine the set of SIB-PBEs. We discover special instances of dynamic games where we
can identify a set of information states with time-invariant domain. We provide conditions that
guarantee the existence of SIB-PBEs in dynamic games with asymmetric information. We show
that the methodology developed in this paper generalizes the existing results on dynamic games
with non-classical information structure.
C. Organization
The rest of the paper is organized as follows. In Section II, we describe our model. In Section
III, we discuss the main issues that arise in the study of dynamic games with asymmetric
information. We provide the formal definition of Perfect Bayesian Equilibrium in Section IV. In
Section V, we describe the sufficient information approach to dynamic games with asymmetric
information and introduce the notion of Sufficient Information Based (SIB) assessment and
November 23, 2018
DRAFT
6
SIB-PBE. In Section VI, we present our main results and provide a sequential decomposition of
dynamic games over time. We discuss our results in Section VII, and compare the notion of SIB-
PBE with other equilibrium concepts..In Section VIII, we determine conditions that guarantee the
existence of SIB-PBE in dynamic games with asymmetric information. We conclude in Section
IX. The proofs of all the theorems and lemmas appear in the Appendix.
Remark 1. Section I-D on notation and Section V-A on the definition of sufficient private
information are similar to the ones appearing in the companion paper [23]; moreover, the
model presented in Section II with strategic agents is similar to that of the companion paper
[23] with non-strategic agents. All these sections are included in this paper for ease of reading
and to make the paper self-contained.
D. Notation
t , . . . ,X N
(resp. functions f 1
t , . . . ,X N
t
t (·), . . . ,f N
t ) (resp. ft(·) := (f 1
Random variables are denoted by upper case letters, their realizations by the corresponding
lower case letters. In general, subscripts are used as time index while superscripts are used to
index agents. For t1≤ t2, Xt1:t2 (resp. ft1:t2(·)) is the short hand notation for the random variables
(Xt1,Xt1+1, ...,Xt2) (resp. functions (ft1(·), . . . ,ft2(·))). When we consider a sequence of random
variables (resp. functions) for all time, we drop the subscript and use X to denote X1:T (resp.
f (·) to denote f1:T (·)). For random variables X 1
t (·)),
t (·))) to denote the vector of the set
we use Xt := (X 1
of random variables (resp. functions) at t, and X−n
t ) (resp.
:= (X 1
t (·))) to denote all random variables (resp. functions)
f−n
at t except that of the agent indexed by n. P{·} and E{·} denote the probability and expectation
of an event and a random variable, respectively. For a set X , ∆(X ) denotes the set of all
beliefs/distributions on X . For random variables X, Y with realizations x,y, P{xy} := P{X =
xY = y} and E{Xy} :=E{XY = y}. For a strategy g and a belief (probability distribution) π,
π{·}) to indicate that the probability (resp. expectation) depends on the
we use Pg
choice of g and π. We use 1{X=x} to denote the indicator function for event X = x. For sets A
and B we use A\B to denote all elements in set A that are not in set B.
(·) := (f 1
t (·), . . . ,f n−1
π{·} (resp. Eg
(·),f n+1
(·), . . . ,f N
t (·), . . . ,f N
t , . . . ,X n−1
t
,X n+1
, . . . ,X N
t
t
t
t
t
1) System dynamics: There are N strategic agents who live in a dynamic Markovian world
over horizon T :={1, 2, ..., T}, T <∞. Let Xt∈Xt denote the state of the world at t∈T . At time
II. MODEL
November 23, 2018
DRAFT
7
t, each agent, indexed by i ∈ N := {1, 2, ..., N}, chooses an action ai
the set of available actions to him at t. Given the collective action profile At := (A1
state of the world evolves according to the following stochastic dynamic equation,
t ∈ Ai
t, where Ai
t denotes
t ), the
t , ..., AN
Xt+1 = ft(Xt, At, W x
t ),
(1)
1:T−1 is a sequence of independent random variables. The initial state X1 is a random
where W x
variable that has a probability distribution η ∈ ∆(X1) with full support.
At every time t ∈ T , before taking an action, agent i receives a noisy private observation
t ∈ Y i
Y i
t of the current state of the world Xt and the action profile At−1, given by
t = Oi
Y i
t(Xt, At−1, W i
t ),
(2)
1:T , i ∈ N , are sequences of independent random variables. Moreover, at every t ∈ T ,
where W i
all agents receive a common observation Zt ∈ Zt of the current state of the world Xt and the
action profile At−1, given by
Zt = Oc
t (Xt, At−1, W c
t ),
(3)
1:T , and W i
1:T , i ∈ N are mutually independent.
where W c
1:T , is a sequence of independent random variables. We note that the agents' actions
At−1 is commonly observable at t if At−1 ⊆ Zt. We assume that the random variables X1,
W x
agents' aggregate information is given by Ht :=(cid:81)
1:T−1, W c
2) Information structure: Let Ht ∈ Ht denote the aggregate information of all agents at time
1:t−1}, i.e. Ht denotes
t. Assuming that agents have perfect recall, we have Ht = {Z1:t, Y 1:N
the set of all agents' past observations and actions. The set of all possible realizations of the
τ <t Ai
τ.
At time t∈T , the aggregate information Ht is not fully known to all agents. Let Ct :={Z1:t}∈
Ct denote the agents' common information about Ht and P i
t denote
t and Ct denote the set of all possible realizations
agent i's private information about Ht, where P i
of agent i's private and common information at time t, respectively. In Section II-A, we discuss
several instances of information structures that can be captured as special cases of our model.
τ≤t Zτ ×(cid:81)
1:t−1}\Ct ∈ P i
τ ×(cid:81)
i∈N(cid:81)
i∈N(cid:81)
t := {Y i
τ≤t Y i
1:t, Ai
, A1:N
1:t
3) Strategies and Utilities: Let H i
i at t, where Hi
behavioral strategy gi
t denote the information available to agent
t denote the set of all possible realizations of agent i's information at t. Agent i's
t), t ∈ T ,
t, t ∈ T , is defined as a sequence of mappings gi
t → ∆(Ai
t : Hi
t := {Ct, P i
t} ∈ Hi
November 23, 2018
DRAFT
that determine agent i's action Ai
t for every realization hi
t ∈ Hi
t of the history at t ∈ T .
8
Agent i's instantaneous utility at t depends on the system state Xt and the collective action
1:T so as to maximize his
t(Xt,At). Agent i chooses his strategy gi
profile At, and is given by ui
total (expected) utility over horizon T , given by,
(cid:88)
t∈T
U i(X1:T , A1:T ) =
ui
t(Xt, At).
(4)
To avoid measure-theoretic technical difficulties and for clarity and convenience of exposition,
we assume that all the random variables take values in finite sets.
Assumption 1. (Finite game) The sets Xt, Zt, Y i
t, Ai
t, N , and T are finite.
Moreover, we assume that given any sequence of actions a1:t−1 up to time t−1, every possible
realization xt∈Xt of the system state at t has a strictly positive probability of realization.
Assumption 2. (Strictly positive transition matrix) For all t∈T , xt∈Xt and a1:t−1∈A1:t−1, we
have P{xta1:t−1} > 0.
Furthermore, we assume that for any sequence of actions {a1:T}, all possible realizations of
1:T } have positive probability. That is, no agent can infer perfectly another
private observations {y1:N
agent's action based only on his private observations.
Assumption 3. (Imperfect private monitoring) For all t∈T , y1:t ∈Y1:t, and a1:t−1 ∈A1:t−1, we
have P{y1:ta1:t−1} > 0.
Remark 2. We can relax Assumptions 2 and 3 under certain conditions and obtain results
similar to those appearing in this paper; for instance, when agents actions are observable we
can relax Assumptions 2 and 3. Broadly, the crucial assumption that underlies our results is
that every deviation that can be detected by agent i at any time t must be also detectable by all
agents at the same time t based only on the common information Ct. Due to space limitation
we do not include the discussion of Assumptions 2 and 3 and the extension of our results when
we relax them; we refer an interested reader to [24].
A. Special Cases
We discuss several instances of dynamic games with asymmetric information that are special
cases of the general model described above.
November 23, 2018
DRAFT
9
1) Nested information structure: Consider a two-player game with one informed player and one
uninformed player and a general Markovian dynamics. At every time t∈T , the informed player
makes a private perfect observation of the state Xt, i.e. Y 1
t = Xt. The uninformed player does
not have any observation of the state Xt. Both the informed and uninformed players observe
each others' actions, i.e. Zt = {At−1}. Therefore, we have P 1
t = ∅, and Ct =
1:t−1} for all t ∈ T . The above nested information structure corresponds to dynamic
{A1
games considered in [11], [25], [26], where in [25], [26] the state Xt is static.
t = {X1:t}, P 2
1:t−1,A2
t ,X 1
t ,X 2
t ,...,X N
2) Independent dynamics with observable actions: Consider an N-player game where the
t ) has N components. The agents' actions At are observable
state Xt := (X 0
by all agents, i.e. At−1 ⊂ Zt for all t ∈ T . At every time t ∈ T , agent i makes a perfect
observation of its local state X i
t . Moreover, at time t all agents
t ), i ∈ N .
make a common imperfect observation of state X i
t evolves independently over time as
Conditioned on the agents' collective action At, each X i
, i∈N ,
X i
t+1 = ft(X i
1:t ,A1:t−1}.
t ∈ T are mutually independent. Therefore, we have P i
The above environment includes the dynamic game considered in [17], [18] as special cases.
t ) for all i∈N and t∈T . We assume that X1, W c
t , t∈T , and W x,i
t
1:t,Z 1:N
t as well as a global state X 0
1:t} and Ct = {X 0
t given by Z i
t,At−1,W x,i
t,At−1,W c,i
t = {X i
t = Oc
t (X i
t ,X 2
t ,...,X N
t ) has N components. Agent i, i ∈ N , perfectly controls X i
3) Perfectly controlled dynamics with hidden actions: Consider a N-player game where the
t, i.e.
t, t∈T , is not observable by all other agents −i. Every agent
t ) of the system state at t∈T . Therefore,
state Xt := (X 1
t+1 = Ai
X i
i, i∈N , makes a noisy private observation Y t
we have P i
t. Agent i's actions Ai
i (Xt, W i
t :={A1:t, Y i
1:t}, Ct =∅.
III. APPRAISALS AND ASSESSMENTS
In this section we provide an overview of the notions of appraisals, assessments, and an
equilibrium solution concept for dynamic games with asymmetric information. We argue that
an equilibrium solution concept must consist of a pair of a strategy profile and a belief system
(to be defined below), and discuss the importance of off-equilibrium path beliefs in dynamic
games.2
In a dynamic game with asymmetric information agents have private information about the
evolution of the game, and they do not observe the complete history of the game given by
2We refer the interested reader to the papers by Battigalli [27], Myerson and Remy [28], and Watson [29] for more discussion.
November 23, 2018
DRAFT
10
{Ht,Xt}. Therefore, at every time t ∈ T , each agent, say agent i ∈ N, needs to form (i) an
appraisal about the current state of the system Xt and the other agents' information H−i
(appraisal
about the history), and (ii) an appraisal about how other agents will play in the future, so as
to evaluate the performance of his strategy choices (appraisal about the future). Given the other
t at t∈T , along with (i) other
agents' strategies g−i, agent i can utilize his own information H i
agents' past strategies g−i
t:T to form these appraisals
about the history and future of the game, respectively.
1:t−1 and (ii) other agents' future strategies g−i
t
In contrast to dynamic teams where agents have a common objective and coordinate their
strategies, in dynamic games each agent has his own objective and chooses his strategy gi so
as to maximize his objective. Thus, unlike dynamic teams, in dynamic games strategy gi is
agent i's private information and not known to other agents. Therefore, in dynamic games, each
agent needs to form a prediction about the other agents' strategies. We denote this prediction by
g∗1:N
1:T that is actually being played by the agents.
Following Nash's idea, we assume that agents share a common prediction g∗ about the actual
strategy g. We would like to emphasize that the prediction g∗ does not necessarily coincide
with the actual strategy g. As we point out later, one requirement of an equilibrium is that for
every agent i∈N , the prediction g∗i must be an optimal strategy for him given the other agents
prediction strategies g∗−i.
to distinguish it from the strategy profile g1:N
1:T
Since an agent's actual strategy, say agent i's strategy gi, is his own private information, it
is possible that gi is different from the prediction g∗i. Below we discuss the implication of
an agent's deviation from the prediction strategy profile g∗. For that matter, we first consider
an agent who may want to deviate from g∗, and then we consider an agent who faces such a
deviation and his response.
In dynamic games, when agent i ∈ N chooses his strategy gi, he needs to know how other
agents will play for any choice of gi which can be different from the prediction g∗i. Therefore,
the prediction g∗ has to be defined at all possible information realizations (i.e. information sets)
of every agent, those that have positive probability under g∗ as well as those that have zero
probability under g∗.3 Using the prediction g∗, any agent, say agent i, can form an appraisal
about the future of the game for any strategy choice gi, and evaluate the performance of gi.
3This is not an issue in dynamic teams since agents coordinate in advance their choice of strategy profile g, and no agent has
an incentive to (privately) deviate from it. Hence, the agents' strategy profile g is needed to be defined only on information sets
of positive probability under g.
November 23, 2018
DRAFT
11
By the same rationale, when agent i chooses gi he needs to determine his strategy for all
of his information sets, even those that have zero probability under g∗−i. This is because it is
possible that some agent j∈N may deviate from g∗j and play a strategy gj that is different from
the prediction g∗j. Agent i must foresee these possible deviations by other agents and determine
his response to these deviations.
To determine his optimal strategy gi at any information set, agent i needs to first form an
appraisal about the history of the game at t as well as an appraisal about the future of the game
using the strategy prediction g∗−i. For an information set hi
t that is compatible with the prediction
g∗−i given his strategy gi at t∈T (i.e. hi
t has positive probability of being realized under g∗),
agent i can use Bayes' rule to derive the appraisal about the history of the game at t. However, for
t that has zero probability under the prediction g∗−i given gi, agent i cannot
an information set hi
anymore rely on the prediction g∗ and use Bayes' rule to form his appraisal about the history
1:t−1 is
of the game at t. The realization of history hi
not (completely) correct, thus, he needs to revise his original prediction g∗−i
1:t−1 and to form a
revised appraisal about the history of the game at t. Therefore, agent i must determine how to
t, t∈T , that
form/revise his appraisal about the history of the game for every realization hi
has zero probability under g∗−i. We note that upon reaching an information set of measure zero,
agent i only revises his prediction g∗−i
1:t−1 about other agents' past strategies, but does not change
his prediction g∗−i
t:T about their future strategies. This is because at equilibrium, the prediction
g∗−i
t:T specifies a set of strategies for other agents that are optimal in the continuation game that
takes place after the realization of the information set hi
t tells agent i that his original prediction g∗−i
t of zero probability under g∗
t∈Hi
1:t−1.4
We describe below how one can formalize the above issues we need to consider in the study of
dynamic games with asymmetric information. Following the game theory literature [30], agents'
appraisals about the history and future of the game can be captured by an assessment that all
agents commonly hold about the game. We define an assessment as a pair of mappings (g∗,µ),
where g∗:={g∗i
t) denotes a prediction about agent i's strategy at
t, and µ := {µi
t → ∆(Xt×H−i
t ) denotes agent i's belief about the
t ,i∈N, t∈T }, g∗i
:Hi
t,i ∈ N,t ∈ T }, where µi
t → ∆(Ai
t : Hi
t
4In dynamic teams, agents only need to determine their optimal strategy g for information sets that have positive probability
under g. As a result, a collective choice of strategy g is optimal at every information set with positive probability if and only
if it maximizes the (expected) utility of the team from t = 1 up to T . However, in dynamic games agents need to determine
their strategies for all information sets irrespective of whether they have zero or positive probability under g∗. Therefore, if a
choice of strategy gi maximizes agent i's (expected) utility from t = 1 to T = 1, it does not imply that it is also optimal at
all information sets that have zero probability under {g∗−i, gi}. Consequently, a choice of agent i's strategy must be optimal
for all continuation games that follow after a realization of an information set hi
t irrespective of whether it has zero or positive
probability.
November 23, 2018
DRAFT
12
t
given his information H i
t) denotes agent i's belief about the history {Xt,H−i
t,i∈N,t∈T } is called a belief system. For every i∈N , t∈T , and hi
system state Xt and agents −i's information H−i
t. The collection of
t ∈Hi
mappings µ :={µi
t,
t:T denotes agent
µi
t(hi
i's prediction about all other agents' continuation strategy from t onward. We note that µi
t(hi
t)
determines agent i's appraisal about the history of the game when hi
t has either positive or
zero probability under g∗. Therefore, using an assessment (g∗,µ) each agent can fully construct
appraisals about the history and future of the game at any t∈T .
t } of the game, and g∗−i
t(hi
t) at information set hi
Using the definition of an assessment, we can extend the idea of Nash equilibrium to dynamic
games with asymmetric information. An equilibrium of the dynamic game is defined as a common
assessment (g∗, µ) among the agents that satisfies the following conditions under the assumption
that the agents are rational. (i) Agent i ∈ N chooses his strategy gi
1:T so as to maximize his
total expected utility (4) in all continuation games given the assessment (g∗,µ) about the game.
Therefore, the prediction g∗i
1:T that other agents hold about agent i's strategy must be a maximizer
of agent i's total expected utility under the assessment (g∗,µ). (ii) For all t∈T , agent i's, i∈N ,
t that has positive probability of realization under g∗,
belief µi
t } conditioned on the realization hi
must be equal to the probability distribution of {Xt,H−i
(determined via Bayes' rule) assuming that agents −i play according to g∗−i
t has zero
probability under the assessment g∗, the belief µi
t) cannot be determined via Bayes' rule and
must be revised. The revised belief must satisfy a certain set of "reasonable" conditions so as
to be compatible with agent i's rationality. Various sets of conditions have been proposed in
the literature (see [30], [31]) to capture the notion of "reasonable" beliefs that are compatible
with the agents' rationality. Different sets of conditions for off-equilibrium beliefs µi
t) result
in the different equilibrium concepts that are proposed for dynamic games with asymmetric
information.
1:t . When hi
t ∈ Hi
t(hi
t
t(hi
In this paper, we consider Perfect Bayesian Equilibrium (PBE) as the equilibrium solution
concept. In the next section we provide the formal definition of PBE.
IV. PERFECT BAYESIAN EQUILIBRIUM
The formal definition of Perfect Bayesian Equilibrium (PBE) for dynamic games in extensive
form can be found in [31]. In this paper we use a state space representation for dynamic games
instead of an extensive game form representation, therefore, we need to adapt the definition of
PBE to this representation. A PBE is defined as an assessment (g∗,µ) that satisfies the sequential
November 23, 2018
DRAFT
rationality and consistency conditions. The sequential rationality condition requires that for all
i ∈ N , the prediction g∗i is optimal for agent i given the assessment (g∗,µ). The consistency
condition requires that for all i∈N , t∈T , and hi
t) must be compatible
with prediction g∗. We formally define these conditions below.
t, agent i's belief µ(hi
t∈Hi
13
Let P(g∗−i
t:T ,g∗i
t:T )
µi
t
{·hi
t} denote the probability measure induced by the stochastic process that starts
t,ct}, where random variables {Xt,P −i
t } are distributed
,pi
t), assuming that agents i and −i take actions according
µ{·}
t:T , respectively. In the sequel, to save some notation, we write Pg∗
at time t with initial condition {Xt,P −i
t
according to probability distribution µi
t(hi
to strategies g∗i
instead of P(g∗−i
t:T and g∗−i
t:T ,g∗i
t:T )
µi
t
{·} whenever there is no confusion.
Definition 1 (Sequential rationality). We say that an assessment (g∗,µ) is sequentially rational
if ∀i∈N , t∈T , and hi
t, the strategy prediction g∗i
t:T is a solution to
t∈Hi
(cid:40) T(cid:88)
τ =t
E(g∗−i
t:T ,gi
µi
t
t:T )
sup
gi
t:T
t(Xt, At)hi
ui
t
(cid:41)
(5)
t
The sequential rationality condition (5) requires that, given the assessment (g∗,µ), the pre-
diction strategy gi∗
for agent i is an optimal strategy for him for all continuation games after
t ∈ H i, irrespective of whether hi
history realization hi
t has positive or zero probability under
(g∗,µ). That is, the common prediction g∗i about agent i's strategy must be an optimal strategy
choice for him since it is common knowledge that he is a rational agent. We note that the
sequential rationality condition defined above is more restrictive than the optimality condition
for Bayesian Nash Equilibrium (BNE) which only requires (5) to hold at t = 1. By the sequential
rationality condition, we require the optimality of prediction g∗ even along off-equilibrium paths,
and thus, we rule out the possibility of non-credible threats (see [30] for more discussion).
The sequential rationality condition results in a set of constraints that the strategy prediction
g∗ must satisfy given a belief system µ. As we argued in Section III, the belief system µ must
be also compatible with the strategy prediction g∗. The following consistency condition captures
such compatibility between the belief system µ and the prediction g∗.
Definition 2 (Consistency). We say that an assessment (g∗,µ) is consistent if
i) for all i ∈ N, t ∈ T \{1}, hi
thi
t such that Pg∗
t−1 ∈ Hi
t−1, and hi
t ∈ Hi
µ{hi
t−1} > 0, the belief
November 23, 2018
DRAFT
µi
t(hi
t) must satisfy Bayes' rule, i.e.
ii) for all i∈N , t∈T \{1}, hi
µ{hi
thi
t−1} = 0, we have
t)(x1:t, p−i
t ) =
t(hi
µi
t−1∈Hi
Pg∗{hi
t
hi
t−1}
t−1}
t, xt, p−i
thi
Pg∗{hi
t such that Pg∗
.
t−1, and hi
t∈Hi
t)(x1:t, p−i
µi
t(hi
t ) > 0
only if there exists an open loop strategy (A−i
1:t−1 = ai
1:t−1) such that
P(A−i
1:t−1=a−i
1:t−1,Ai
1:t−1=ai
1:t−1 = a−i
1:t−1,Ai
1:t−1){xt, p−i
t } > 0.
14
(6)
(7)
t(hi
{hi
µi
1
{hi
t} > 0, belief µi
First, consider an information set hi
t indicates that some agent has deviated from prediction g∗−i
The above consistency condition places a restriction on the belief system µ so that it is
compatible with the strategy prediction g∗. For information sets along equilibrium paths, i.e.
Pg∗
t) must be updated according to (6) via Bayes' rule since agent i's
µi
1
observations are consistent with the prediction g∗. For information sets along off-equilibrium
t} = 0, agent i needs to revise his belief about the strategy of agents −i as
paths, i.e. Pg∗
the realization of hi
1:t . As pointed
t) must be "reasonable". Definition 2 provides a set of such
out before, the revised belief µi(hi
"reasonable" conditions captured by (6) and (7) that we discuss further below.
t along an off-equilibrium path such that Pg∗
t−1} > 0.
t−1 at t−1, hi
t has a positive probability under
That is, conditioned on reaching information set hi
t−1} and Pg∗
t} = 0, we have
{hi
thi
t−1}Pg∗
{hi
the prediction strategy g∗. Since Pg∗
t−1} = 0. Therefore, hi
{hi
Pg∗
t−1 is also an information set along an off-equilibrium path, and
µi
1
t−1) is a revised belief that agent i holds at t−1. Note that if the assessment (g∗,µ) satisfies
µi(hi
the sequential rationality condition, g∗ is a best response for all agents in all continuation games
that follow the realization of every information set of positive or zero probability. Moreover,
since Pg∗
t−1 is consistent with
the strategy prediction g∗
t−1. Therefore, agent i does not have any reason to further revise his
belief about agents −i's strategy beyond the revision that results in µi
t−1). Thus, agent i
t−1) at t−1 and updating it via Bayes'
determines his belief µi
t) by utilizing his belief µi
rule assuming that agents −i' play according to g∗−i
t−1} > 0, the realization of hi
t conditioned on reaching hi
t−1 (see part (i), eq. (6)).
t} = Pg∗
t−1(hi
t−1(hi
{hi
thi
{hi
thi
t−1
µi
t−1
{hi
{hi
t(hi
µi
1
µi
t−1
µi
t−1
µi
1
µi
1
Next, consider an information set hi
t along an off-equilibrium path such that Pg∗
t−1 at t− 1, hi
t−1} = 0.
t has a zero probability of
thi
{hi
µi
t
That is, conditioned on reaching information set hi
November 23, 2018
DRAFT
15
t
t(hi
t(hi
t indicates that agents −i
realization under the prediction g∗. In this case, the realization of hi
have deviated from prediction g∗−i
1:t−1, and this deviation has not been detected by agent i before.
Therefore, agent i needs to form a new belief about agents −i's private information P −i
and the
t). Part (ii) of the consistency condition concerns such belief revisions
state Xt by revising µi
and requires that the support of agent i's revised belief µi
t) includes only the states and private
information that are feasible under the system and information dynamics (1) and (2), that is,
they are reachable under some open-loop control strategy (A−i
1:t−1). We
note that since we are using a state representation of the dynamic game, we need to impose
such a requirement, whereas in the equivalent extensive form representation of the game such a
requirement is satisfied by the construction of the game-tree.
1:t−1 = a−i
1:t−1 = ai
1:t−1,Ai
t } > 0 for all (A1:t−1 =
Remark 3. Under Assumptions 2 and 3, we have P(A1:t−1=a1:t−1)
a1:t−1). Therefore part (ii) of the consistency conditions is trivially satisfied. In the rest of the
paper, we ignore part (ii) and only consider part (i) of the definition of consistency. In [24], we
discuss the case when we relax Assumptions 2 and 3.
{x1:t,p−i
µ1
We can now provide the formal definition of PBE for the dynamic game of Section II.
Definition 3. An assessment (g∗,µ) is called a PBE if it satisfies the sequential rationality and
consistency conditions.
The definition of Perfect Bayesian equilibrium provides a general formalization of outcomes
that are rationalizable (i.e. consistent with agents' rationality) under some strategy profile and
belief system. However, as we argue further in Section VII, there are computational and philo-
sophical reasons that motivate us to define a sub class of PBEs that provide a simpler and
more tractable approach to characterizing the outcomes of dynamic games with asymmetric
information.
There are two major challenges in computing a PBE (g∗,µ). First, there is an inter-temporal
coupling between the agents' strategy prediction g∗ and belief system µ. According to the
consistency requirement, the belief system µ has to satisfy a set of conditions given a strategy
prediction g∗. On the other hand, by sequential rationality, a strategy prediction g∗ must satisfy
a set of optimality condition given belief system µ. Therefore, there is a circular dependency
between a prediction strategy g∗ and a belief system µ over time. For instance, by sequential
rationality, agent i's strategy gi∗
t:T and on
t at time t depends on the agents' future strategies g∗
November 23, 2018
DRAFT
16
1:t−1 indirectly through the consistency condition for µi
the agents' past strategies g∗
t. As a result,
one needs to determine the strategy prediction g∗ and belief system µ simultaneously for the
whole time horizon so as to satisfy the sequential rationality and consistency conditions, and
thus, cannot sequentially decompose the computation of PBE over time. Second, the agents'
t, i ∈ N , has a growing domain over time. Hence, the agents' strategies have
information hi
growing domains over time, and this feature further complicates the computation of PBEs of
dynamic games with asymmetric information.
The definition of PBE requires an agent to keep track of all observations he acquires over
time and to form beliefs about the private information of all other agents. As we show next,
agents do not need to keep track of all of their past observations to reach an equilibrium. They
can take into account fewer variables for decision making and ignore part of their information
that is not relevant to the continuation game. As we argue in Section VII, the class of simpler
strategies proposed in this paper characterize a more plausible prediction about the outcome of
the interaction among agents when the underlying system is highly dynamic and there exists
considerable information asymmetry among them.
V. THE SUFFICIENT INFORMATION APPROACH
We characterize a class of PBEs that utilize strategy choices that are simpler than general
behavioral strategies as they require agents to keep track of only a compressed version of their
information over time. We proceed as follows. In Section V-A we provide sufficient conditions for
the subset of private information an agent needs to keep track of over time for decision making
purposes. In Section V-D, we introduce the sufficient information based belief as a compressed
version of the agents' common information that is sufficient for decision-making purposes. Based
on these compressions of the agents' private and common information, we introduce the notion
of sufficient information based assessments and Sufficient Information Based-Perfect Bayesian
Equilibrium (SIB-PBE) in Sections V-D and V-E, respectively.
A. Sufficient Private Information
The key ideas for compressing an agent's private information appear in Definitions 4 below;
We refer an interested reader to the companion paper [2] for discussion on the rationale behind
Definition 4.
November 23, 2018
DRAFT
Definition 4 (Sufficient private information). We say Si
sufficient private information for the agents if,
t = ζ i
t (P i
t , Ct; g∗
1:t−1), i ∈ N , t ∈ T , is
(i) it can be updated recursively as
Si
t = φi
t(Si
t−1, H i
t\H i
t−1; g∗
1:t−1) if t ∈ T \{1},
(8)
17
(ii) for any strategy profile g∗ and for all realizations {ct, pt, pt+1, zt+1, at} ∈ Ct ×Pt ×Pt+1 ×
Zt+1 of positive probability,
Pg∗
1:t
=Pg∗
1:t
st+1,zt+1
,
(9)
(cid:110)
(cid:12)(cid:12)(cid:12)st,ct,at
(cid:111)
where s1:N
τ = ζ 1:N
τ
(p1:N
τ
,cτ ;g∗
(iii) for every strategy profile g∗ of the form g∗ := {g∗i
: S i
(cid:110)
(cid:12)(cid:12)(cid:12)pt,ct,at
(cid:111)
st+1,zt+1
1:τ−1) for τ ∈ T ;
(cid:111)
(cid:12)(cid:12)(cid:12)ct,pi
t,at
ui
t(Xt,At)
t}∈Ct×P i
t), i ∈ N,t ∈ T } and
t × Ct → ∆(Ai
(cid:111)
t
(cid:110)
(cid:12)(cid:12)(cid:12)ct,si
(10)
=E g∗
1:t−1
ui
t(Xt,At)
t,at
,
t of positive probability where s1:N
τ = ζ 1:N
τ
(p1:N
τ
,cτ ; g∗
1:τ−1)
at∈At, t∈T ;
E g∗
1:t−1
(cid:110)
for all realizations {ct,pi
for τ ∈ T ;
i∈N , and t∈T ,
for all realizations {ct,pi
for τ ∈ T .
(iv) given an arbitrary strategy profile g∗ of the form g∗ := {gi
t×Ct → ∆(Ai
t), i∈N , t∈T },
t : S i
(cid:111)
t,ct
,
(cid:12)(cid:12)(cid:12)si
(cid:110)
=Pg∗
1:t−1
s−i
t
t,ct
(cid:111)
(cid:12)(cid:12)(cid:12)pi
Pg∗
(cid:110)
t}∈Ct×P i
1:t−1
s−i
t
t of positive probability where s1:N
τ = ζ 1:N
τ
(p1:N
τ
(11)
,cτ ; g∗
1:τ−1)
We note that the conditions of Definition 4 is written in terms of strategy prediction profile g∗
for dynamic games. This is because, as we discussed before, the agents' actual strategy profile
g is their private information. Therefore, each agent i, i ∈ N , evaluates the sufficiency of a
compression of his private information using the strategy prediction he holds about other agents.
B. Sufficient Common Information
Based on the characterization of sufficient private information, we present a statistic (com-
pressed version) of the common information Ct that agents need to keep track of over time for
decision making purposes.
November 23, 2018
DRAFT
18
t, and St :=(cid:81)N
, t∈T . Define S i
t
i=1S i
Consider the sufficient private information S1:N
t to be the set of all possible
t. Let γt :Ct→ ∆(Xt×St) denote a mapping that determines
realizations of Si
a conditional probability distribution over the system state Xt and the agents' sufficient private
information St given the common information Ct at time t. We call the collection of mappings
γ := {γt,t ∈ T } a Sufficient Information Based belief system (SIB belief system). Note that
γt is only a function of the common information Ct, and thus, it is computable by all agents.
Let Πγ
t := γt(Ct) denote the (random) sufficient information based belief that agents hold under
belief system γ at t. We can interpret Πγ
t as the common belief that each agent holds about the
system state Xt and all the agents' (including himself) sufficient private information St at time
t. We call the SIB belief Πt a sufficient common information for the agents. In the rest of the
paper, we write Πt and drop the superscript γ whenever such a simplification in notation is clear.
Moreover, we use the terms sufficient common information and SIB belief interchangeably.
C. Special Cases:
We consider the special classes described in Section II and identify the sufficient information
1:T and SIB belief for each of them.
S1:N
1) Nested information structure: The uninformed agent (agent 2) has no private information,
t =∅. Thus, S2
t = Xt. Consequently, we
P 2
is the same as
can set S1
SIB belief Πt =P{XtA1
t =∅, thus, the uninformed agent's belief about P 1
1:t−1}.
t =∅. For the informed agent (agent 1) consider P 1,pr
t = Xt. Note that P 2
1:t−1,A2
2) Independent dynamics with observable actions: Consider Si
t , j ∈ N
have independent dynamics given the collective action At that is commonly observable by all
agents. Therefore, agent i's belief about Xj, j (cid:54)= i, is the same as SIB belief, Πt =Pg{X j
t Ct} =
Pg{X j
t. Note that X j
t = X i
3) Perfectly controlled dynamics with hidden actions: Since agent i, i ∈ N , perfectly controls
t over time t∈T , we set Si
X i
t } and Πt =∅.
t ={Ai
t−1, Y i
t P i
t , Ct}.
t
D. Sufficient Information based Assessment
As we discussed in Section IV, to form a prediction about the game we need to determine
an assessment about the game that is sequentially rational and consistent. We show below that
using the sufficient private information S1:N
and sufficient common information (the SIB belief)
Πt, we can form a sufficient information based assessment about the game. We prove that such
a sufficient information based assessment is rich enough to capture a subset of PBE.
t
November 23, 2018
DRAFT
19
Consider a class of strategies that utilize the information given by (Πt,Si
t) for agent i∈N at
t → ∆(Ai
t : ∆(Xt×St)×S i
t) a Sufficient Information Based (SIB)
time t. We call the mapping σi
strategy for agent i at time t. A SIB strategy σi
t determines a probability distribution for agent i's
t). A SIB strategy is a behavioral strategy where
t at time t given his information (Πt,Si
action Ai
agents only use the SIB belief Πt = γt(Ct) (instead of complete common information Ct), and
the sufficient private information Si
1:t−1) (instead of complete private information
t = ζ i
t ). A collection of SIB strategies {σ1
1:T} is called a SIB strategy profile σ. The set of
P i
SIB strategies is a subset of behavioral strategies, defined in Section II, as we can define,
t ,Ct; g∗
1:T,...,σN
t (P i
g(σ,γ),i
t
(hi
t) := σi
t(πγ
t , si
t).
In Section IV, we defined a consistency condition between strategy prediction g∗ and a
belief system µ. Below, we provide an analogous consistency condition between a SIB strategy
prediction σ∗ and a SIB belief system γ.
Definition 5. A pair (σ∗,γ) of a SIB strategy prediction profile σ∗ and belief system γ satisfies
the consistency condition if
(i) for all t∈T \{1}5, zt ∈Zt, πt−1 = γt−1(ct−1), and πt = γt({ct−1,zt}) such that Pσ∗
πt−1{zt} > 0,
t
πt must satisfy Bayes' rule, i.e.,
πt(xt, st) =
t
Pσ∗
πt−1{xt, st, zt}
Pσ∗
πt−1{zt}
t
∀xt ∈ Xt,∀st ∈ St,
,
(12)
(ii) for all t ∈ T \{1}, ct−1 ∈ Ct−1, πt−1 = γt−1(ct−1), zt ∈ Zt, and πt = γt({ct−1,zt}) such that
Pσ∗
πt−1{zt} = 0, we have
t
πt(xt, st) > 0,
∀xt ∈ Xt,∀st ∈ St,
only if there exists an open-loop strategy (A1:t−1 = a1:t−1) such that P(A1:t−1=a1:t−1)
0, and
π1
{ct−1,zt} >
P(A1:t−1=a1:t−1)
π1
{xt, st} > 0,
(13)
(iii) for all t ∈ T \{1}, ct−1 ∈ Ct−1, πt−1 = γt−1(ct−1), zt ∈ Zt, and πt = γt({ct−1,zt}) such that
5For t=1, Π1 is given by the conditional probability at t=1 as Π1(x1,s1):=
(cid:80)
P{s1x1,z1}
P{z1x1}η(x1) .
x1∈X1
November 23, 2018
DRAFT
Pσ∗
πt−1{zt} = 0, we have
t
(cid:88)
xt∈Xt
πt(xt, st) > 0,
∀st ∈ St
20
if there exists an open-loop strategy (A1:t−1 = a1:t−1) such that P(A1:t−1=a1:t−1)
and
π1
{ct−1,zt} > 0,
(cid:88)
xt∈Xt
P(A1:t−1=a1:t−1)
π1
{xt, st} > 0.
(14)
Parts (i) and (ii) of Definition 5 follow from rationales similar to their analogues in Definition
2, and require a SIB belief system to satisfy a sets of constraints with respect to a SIB strategy
profile that are similar to those for an assessment (g∗,µ). Definition 5 requires an additional
condition described by part (iii). By (14), we require a SIB belief system γ consistent with the
SIB strategy profile σ∗ to assign a positive probability to every realization st of the agents'
sufficient private information St that is "plausible" given the common information realization
ct = {ct−1,zt}; plausibility of st given ct means that there exists an open-loop strategy profile
(A1:t−1 = a1:t−1) consistent with the realization ct that leads to the realization of st with positive
probability. Therefore, part (iii) ensures that there exists no possible incompatibility between the
SIB belief Πt and the agents' sufficient private information St+1. As we show later (Section
V-E), such a compatibility condition allows each agent to refine the SIB belief Πt using his own
private sufficient information Si
t, and to form his private belief about the game.
Remark 4. Assumptions 2 and 3 imply that (13) holds for all (A1:t−1 = a1:t−1) such that
P(A1:t−1=a1:t−1) {ct} > 0. Therefore, in the rest of the paper, we ignore part (ii) of the consistency
condition for SIB belief systems. Moreover, under Assumptions 2 and 3, condition (14) is
πt(xt,pt) > 0 whenever
always satisfied. Therefore, condition (iii) is equivalent to having(cid:80)
πt−1{zt} = 0.
Pσ∗
Given a SIB strategy profile prediction σ∗, a consistent SIB belief must satisfy (12), which
determines the SIB belief Πt at t in terms of the SIB belief Πt−1 at t − 1 and the new common
information Zt at t. We define a SIB belief update rule as a mapping ψt : ∆(Xt−1×St−1)×Zt →
xt∈Xt
November 23, 2018
DRAFT
∆(Xt ×St), t∈T that determines recursively the SIB belief
Πψ
t
:= ψt(Πψ
t−1, Zt),
21
(15)
t indicates that the SIB belief Πψ
t−1 at t−1.6 The superscript
as a function of new common observation Zt at t and the SIB belief Πψ
ψ in Πψ
t is generated using the SIB update rule ψ. Let γψ denote
the common belief system that is equivalent to the SIB update rule ψ. We call a SIB belief update
rule ψ consistent with a SIB strategy profile σ∗ if the equivalent SIB belief system γψ is consistent
with σ∗ (Definition 5).
Define a SIB assessment (σ∗,γ) as a pair of SIB strategy profile σ∗ and a SIB belief system γ.
Below, we show that a consistent SIB assessment (σ∗,γ) is equivalent to a consistent assessment
(g∗,µ) as defined in Section IV (Definition 2).
Lemma 1. For any given SIB assessment (σ∗,γ), there exists an equivalent assessment (g∗,µ)
of a behavioral strategy prediction g∗ and belief system µ such that:
i) the behavioral strategy g∗ is defined by
ii) the belief system µ is consistent with g∗ and satisfies
g∗i
t (hi
t) := σ∗i
t (πγ
t , si
t);
Pg∗(cid:8)s−i
t
hi
t
(cid:9) ,
(16)
(17)
(cid:9) = P(cid:8)s−i
t
πt, si
t
for all i ∈ N , t ∈ T ,hi
t ∈ Hi
t, and s−i
t ∈ S−i
t
.
Lemma 1 shows that the set of consistent SIB assessment (σ∗,γ) is equivalent to a subset of
consistent assessments (g∗,µ). That is, using the SIB belief system γ and SIB strategy profile σ∗,
agents can form a consistent assessment about the evolution of the game. Moreover, condition
t capture all the
(17) implies that the SIB belief Πt along with agent i's sufficient information Si
information in H i
t that is relevant to agent i's belief about S−i
.
t
6Upon reaching an information set of measure zero (parts (ii) and (iii) of Definition 5), the revised SIB belief could be
a function of Ct = {Ct−1,Zt}, rather than only Πt−1(Ct−1) and Zt. Therefore, the set of SIB belief systems that can be
generated from SIB update rules is a subset of all consistent SIB belief systems given by Definition 5. However, we argue that
upon reaching an information set of measure zero, it is more plausible to revise the SIB belief only as a function of relevant
information Πt−1(Ct−1) and Zt; Ct is irrelevant given Πt−1(Ct−1) and Zt.
November 23, 2018
DRAFT
22
E. Sufficient Information based PBE
Using the result of Lemma 1, we can define a class of PBE, called Sufficient Information
based PBE (SIB-PBE), as a set of equilibria for dynamic games with asymmetric information
that can be expressed as SIB assessments.
Definition 6. A SIB assessment (σ∗,γ) is called a SIB-PBE if γ is consistent with σ∗ (Definition
5), and the equivalent consistent assessment (g∗, µ), given by Lemma 1, is a PBE.
In the sequel, we also call a consistent pair (σ∗,ψ) of a SIB strategy prediction profile σ∗ and
SIB belief update rule ψ a SIB-PBE if (σ∗,γψ) is a SIB-PBE.
Throughout Section V, we assumed that agents play according to the strategy predictions g∗
(or SIB strategy predictions σ∗). However, an agent's, say agent i ∈ N 's, actual strategy gi is
his private information and could be different from g∗ if such a deviation is profitable for him.
The proposed class of SIB assessments imposes two restrictions on agents' strategies and beliefs
compared to the general class of assessment presented in Section IV. First, it requires that each
agent i, i∈N , must play a SIB strategy σ∗i instead of a general behavioral strategy g∗i. Second,
it requires that each agent i, i∈N , must form a belief about the status of the game using only
t (instead of a general belief
the SIB belief Πt along with his sufficient private information Si
t). A strategic agent i∈N does not restrict his choice of strategy to SIB strategies, and may
µi
deviate from σ∗i to a non-SIB strategy gi if it is profitable to him. Moreover, a strategic agent
i does not limit himself to form belief about the current status of the game only based on Πt
t, and may instead use a general belief µi if it enables him to improve his expected utility.
and Si
In the next section, we address these strategic concerns, and show that no agent i∈N wants to
deviate from (Π,σ∗) and play a non-SIB strategy gi when all other agents are playing according
to SIB assessment (Π,σ∗). This result allows us to focus on the class of SIB assessments, and
develop a methodology to sequentially decompose the dynamic game over time.
VI. MAIN RESULTS
In this section, we show that the class of SIB assessments is rich enough to capture the agents'
strategic interactions.We first show that when agents −i play according to a SIB assessment
(σ∗,γ), agent i, i∈N , cannot mislead these agents by playing a strategy gi different from σ∗i,
thus, creating dual beliefs, one belief that is based on the SIB assessment (σ∗,γ) the functional
form of which is known to all agents, and another belief that is based on his private strategy
November 23, 2018
DRAFT
23
gi that is only known to him (Theorem 1). Then, we show that given that agents −i play SIB
strategy σ∗−i, agent i has a response that is a SIB strategy (Theorem 2).
The result of Theorems 1 (resp. 2) for agent i ∈ N assumes that all other agents −i are playing
according to strategy prediction g∗−i (resp. σ∗−1). The same results hold for every continuation
game that starts at any time t ∈ T along an off-equilibrium path; they can be proved by relabeling
time t as time 1, and using the SIB belief πt = γt(ct) and the corresponding belief µi
t), defined
by Lemma 1, as the initial common belief for the continuation game.
t(hi
Using the results of Theorems 1 and 2, we present a methodology to determine the set
of SIB-PBEs of stochastic dynamic games with asymmetric information (Theorem 3). The
proposed methodology leads to a sequential decomposition of stochastic dynamic games with
asymmetric information. This decomposition gives rise to a dynamic program that can be utilized
to compute SIB-PBEs via backward induction. We proceed by stating the following results from
the companion paper [2, Theorem 1].
Theorem 1 (Policy-independence belief property [2]).
(i) Consider a general strategy prediction profile g∗. If agents −i play according to strategy
predictions g∗−i, then for every strategy gi that agent i plays,
Pg∗,g−i(cid:110)
xt, p−i
t
(cid:111)
(cid:12)(cid:12)(cid:12)hi
t
= Pg∗−i(cid:110)
xt, p−i
t
(cid:111)
(cid:12)(cid:12)(cid:12)hi
t
.
(18)
(ii) Consider a SIB strategy prediction profile σ∗ along with the associated consistent update
rule ψ. If agents −i play according to SIB strategy predictions σ∗−i, then for every general
strategy gi that agent i plays,
Pσ∗−i,gi
ψ
xt, p−i
t
=Pσ∗−i
ψ
xt, p−i
t
.
(19)
(cid:110)
(cid:12)(cid:12)(cid:12)hi
(cid:111)
t
(cid:110)
(cid:12)(cid:12)(cid:12)hi
(cid:111)
t
Part (i) of Theorem 1 states that under perfect recall agent i's belief is independent of his
actual strategy gi. Therefore, agent i cannot mislead agents −i by deviating from the SIB strategy
prediction g∗i to a behavioral strategy gi so as to create dual beliefs (described above) that he
can use to his advantage. Part (ii) of Theorem 1 concerns situations where all agents except
agent i play SIB strategies σ∗−i. Given SIB update rule ψ that is consistent with σ∗, it states
that agent i's beliefs independent of his actual strategy gi. We note when agents −i play SIB
strategies the consistency condition for SIB update rule ψ depends on strategy prediction σ∗i
they have about agent i.
November 23, 2018
DRAFT
24
Using the result of Theorem 1, we show that agent i∈N does not gain by playing a non-SIB
strategy gi when all other agents −i are playing SIB strategies σ∗−i.
Theorem 2 (Closedness of SIB strategies). Consider a consistent SIB assessment (σ∗,γψ) where
ψ is a SIB update rule consistent with σ∗. If every agent j∈N , j(cid:54)= i plays the SIB strategy σ∗j,
then, there exists a SIB strategy σi for agent i that is a best response to σ∗−i.
The results of Theorems 1 and 2 address the two restrictions (discussed above) imposed in
SIB assessments on the agents' beliefs and strategies, respectively.
Based on these results, we restrict attention to SIB assessments, and provide a sequential
decomposition of dynamic games with asymmetric information. A SIB-PBE is SIB assessment
that is a fixed point under the best response map for all agents. Below, we formulate a dynamic
program that enables us to compute SIB-PBEs of dynamic games with asymmetric information.
Consider a dynamic program over time horizon T ∪{T + 1} with information state {Πt,St},
t ∈ T . Let Vt := {V i
: ∆(Xt × St) × St → R, i ∈ N} denote the value function that captures
the continuation payoffs for all agents, for all realizations of the SIB belief Πt and the agents'
T +1 = 0 for all i ∈ N . For each stage t ∈ T of
private sufficient information St, t ∈ T . Set V i
the dynamic program consider the following static game.
t
Stage game Gt(πt,Vt+1,ψt+1): Given the value function Vt+1 and SIB update rule ψt+1, we
define the stage game Gt(πt,Vt+1,ψt) as a static game of asymmetric information among agents
for every realization πt. Each agent i∈N has private information Si
t that is distributed according
to πt, which is common knowledge among the agents. Given a realization at of the agents'
collective action profile and a realization st of the agents' sufficient private information, agent
i's utility is given by
(cid:110)
(cid:12)(cid:12)(cid:12)πt,st,at
(cid:111)
t(at,st,πt,Vt+1,ψt+1) := Eπt
¯U i
ui
t(Xt,at)+V i
t+1 (ψt+1(πt,Zt+1),St+1)
.
(20)
BNE correspondence: We define the correspondence BNEt (Vt+1,ψt+1), t∈T , as the corre-
spondence mapping that characterizes the set of BNEs of the stage game Gt(πt,Vt+1,ψt+1) for
every realization of πt; this correspondence is given by
BNEt (Vt+1,ψt+1) :=(cid:8)σ∗
t :∀πt∈ ∆(Xt×St), σ(πt,·) is a BNE of Gt(πt,Vt+1,ψt+1)(cid:9).
(21)
We say σ∗(πt,·) is a BNE of the stage game Gt(πt,Vt+1,ψt) if for all agents i ∈ N , and for all
November 23, 2018
DRAFT
t∈S i
t,
si
σ∗i(πt,si
t)∈ arg max
α∈∆(Ai
t)
(cid:110)¯U i
Eπt
t((α,σ∗−i(πt,S−i
t )),St,πt,Vt+1,ψt+1)
(cid:12)(cid:12)(cid:12)πt,si
(cid:111)
t
.
25
(22)
(23)
(24)
(25)
(26)
Below, we provide a sequential decomposition of dynamic games with asymmetric information
using the stage game and the BNE correspondence defined above.
Theorem 3 (Sequential decomposition). A pair (σ∗,ψ) of a SIB strategy profile σ∗ and a SIB
update rule ψ (equivalently, a SIB assessment (σ∗,γψ)) is a SIB-PBE if (σ∗, ψ) solves the
following dynamic program:
for t = T +1:
for t ∈ T :
T +1(·) := 0
V i
∀ i ∈ N ;
t ∈ BNEt (Vt+1, ψt+1) ,
σ∗
ψt+1 is consistent with σ∗,
t (πt,st) :=Eσ∗
V i
(cid:110)¯U i
t
ψt+1
t((σ∗(πt,St)),St,πt,Vt+1,ψt+1)
, ∀i∈N.
(cid:12)(cid:12)(cid:12)πt,si
(cid:111)
t
VII. DISCUSSION
We showed that SIB assessments proposed in this chapter are rich enough to capture a set
of PBE. However, we would like to point out that the concept of SIB-PBE does not capture all
PBEs of a dynamic game in general. We elaborate on the relation between the of SIB-PBE and
PBE below. We argue that the set of SIB-PBEs are more plausible to arise as the information
asymmetry among the agents increases and the underlying system is dynamic.
We presented an approach to compress the agents' private and common information by
providing conditions sufficient to characterize the information that is relevant for decision making
purposes. Such information compression means that the agents do not incorporate into their
decision making processes their observations that are irrelevant to the continuation game. As
we show in [2], this information compression is without loss of generality for dynamic team
problems. However, this is not the case in dynamic games. In general, the set of SIB-PBEs
of a dynamic game is a subset of all PBEs of that game. This is because in a dynamic game
agents can incorporate their past irrelevant observations into their future decisions so as to create
November 23, 2018
DRAFT
26
rewards (resp. punishments) that incentivize the agents to play (resp. not play) specific actions
over time. By compressing the agents' private and common information in SIB assessments, we
do not capture such punishment/reward schemes that are based on past irrelevant observations.
Below, we present an example where there exists a PBE that cannot be captured as a SIB-PBE.
Consider a two-agent repeated game with T = 2 and a payoff matrix given in Table I. At
each stage, agent 1 chooses from {U,D}, and agent 2 chooses from {L,M,R}. We assume
that agents' actions are observable. Therefore, the agents have no private information, and the
sufficient private information and SIB belief are trivial. The stage game has two equilibria in
pure strategies given by (D,M ) and (U,R). Using the results of Theorem 3, we can characterize
four SIB-PBEs of the repeated game that correspond to the different combinations of the two
equilibria of the stage game as follows: (DD,MM ), (U U,RR), (DU,MR), and (UD,RM ).
However, there exists an another PBE of the repeated game that cannot be captured as a SIB-
PBE. Consider the following equilibrium: Play (U,L) at t = 1. If agent 2 plays L at t = 1 then
play (U,R); otherwise, play (D,M ) at t = 2. Note that agent 1's decision at t = 2 depends on
agent 2's action at t = 1, which is a payoff-irrelevant information since the two stages of the
game are independent. Nevertheless, agent 1 utilizes agent 2's action at t = 1, and punishes him
(agent 2) by playing D at t = 2 if he deviates from playing L at t = 1.
L
R
M
(0,2)
(1,2)
(2,10)
U (8,3)
D (0,1)
(0,0)
TABLE I: Payoff matrix
We would like to point out that there are instances of dynamic games with asymmetric
information, such as zero-sum dynamic games [32], where the equilibrium payoffs for the agents
are unique. In these games it is not possible to incorporate pay-off irrelevant information so as to
construct additional equilibria where the agents' payoffs are different from the ones corresponding
to SIB-PBEs; this is clearly the case for zero-sum games since the agents do not cooperate on
creating punishment/reward schemes due to the zero-sum nature of the game. In Section VIII,
we utilize this fact and prove the existence of SIB-PBEs in zero-sum games. Furthermore, we
show that SIB-PBEs are equivalent to PBEs, in terms of the agents' equilibrium payoffs, in these
games.
While it is true that in general, the set of PBEs of a dynamic game is larger than the set of
SIB-PBEs of that game, in the remainder of this section, we provide three reasons on why in
November 23, 2018
DRAFT
27
a highly dynamic environment with information asymmetry among agents, SIB-PBEs are more
plausible to arise as an outcome of a game.
First, we argue that in the face of a highly dynamic environment, an agent with partial
observations of the environment should not behave fundamentally different whether he interacts
in a strategic or cooperative environment. From the single-agent decision making point of view
(i.e. control theory), SIB strategies are the natural choice of an agent for decision making purposes
(See Thm. 2 in the companion paper [2]).7
Second, we argue that in a highly dynamic environment with information asymmetry among
the agents, the formation of punishment/reward schemes that utilize the agents' payoff-irrelevant
information requires prior complex agreements among the agents; these complex agreements are
sensitive to the parameters of the model and are not very plausible to arise in practice when
the decision making problem for each agent is itself a complex task. We note that the set of
PBEs that cannot be captured as SIB-PBEs are the ones that utilize payoff-irrelevant information
to create punishment/reward schemes in the continuation game as in the example above. How-
ever, such punishment/reward schemes require the agents to form a common agreement among
themselves on how to utilize such payoff-irrelevant information and how to implement such
punishment/reward schemes. The formation of such a common agreement among the agents
is more likely in games where the underlying system is not highly dynamic (as in repeated
games [7]) and there is no much information asymmetry among agents. However, in a highly
dynamic environment with information asymmetry among agents the formation of such common
agreement becomes less likely for the following reasons. First, in those environments each agent's
individual decision making process is described by a complex POMDP; thus, strategic agents
are less likely to form a prior common agreement (that depend on the solution of the individual
POMDPs) in addition to solving their individual POMDPs. Second, as the information asymmetry
increases among agents, punishment/reward schemes that utilize payoff-irrelevant information
require a complex agreement among the agents that is sensitive and not robust to changes in
the assumptions on the information structure of the game. For instance, consider the example
described above, but assume that agents observe imperfectly each others' actions at each stage
(Assumption 3). Let 1 − , ∈ (0,1), denote the probability that agents observe each others'
7The SIB approach proposed in this paper for dynamic games along with the SIB approach to dynamic teams proposed in the
companion paper [2] provide a unified approach to the study of agents' behavior in a dynamic environment with information
asymmetry among them.
November 23, 2018
DRAFT
28
actions perfectly, and denote the probability that their observation is different from the true
action of the other agent. Then, the described non-SIB strategy profile above remains as a PBE of
the game only if ≤ 1
5. The author of [33] provides a general result on the robustness of above-
mentioned punishment/reward schemes in repeated games; he shows that the set of equilibria
that are robust to changes in information structure that affect only payoff-irrelevant signals does
not include the set of equilibria that utilize punishment/reward schemes described above.
Third, the proposed notion of SIB-PBE can be viewed as a generalization of Markov Perfect
Equilibrium [3] to dynamic games with asymmetric information. Therefore, a similar set of
rationales that support the notion of MPE also applies to the notion of SIB-PBE as follows. First,
the set of SIB assessments describe the simplest form of strategies capturing the agents' behavior
that is consistent with the agents' rationality. Second, the class of SIB assessments captures the
notion that "bygones are bygones", which also underlies the requirement of subgame perfection
in equilibrium concepts for dynamic games. That is, the agents' strategies in two continuation
games that only differ in the agents' information about payoff-irrelevant events must be identical.
Third, the class of SIB assessments embodies the principle that "minor changes in the past should
have minor effects". This implies that if there exists a small perturbation in the specifications of
the game or the agents' past strategies that are irrelevant to the continuation game, the outcome
of the continuation game should not change drastically. The two-step example above presents
one such situation, where one equilibrium that is not SIB-PBE disappears suddenly as → 1
5.
VIII. EXISTENCE OF EQUILIBRIA
As we discussed in Section VII, there exist PBEs that cannot be described as SIB-PBEs in
general. Therefore, the standard results that guarantee the existence of a PBE for dynamic games
with asymmetric information [31, Proposition. 249.1] cannot be used to guarantee the existence
of a SIB-PBE in these games. In this Section, we discuss the existence of SIB-PBEs for dynamic
games with asymmetric information. We provide conditions that are sufficient to guarantee the
existence of SIB-PBEs (Lemmas 2 and 3). Using the result of Lemma 2, we prove the existence
of SIB-PBEs for zero-sum dynamic games with asymmetric information (Theorem 4). Using
the result of Lemma 3, we identify instances of non-zero-sum dynamic games with asymmetric
information where we can guarantee the existence of SIB-PBEs.
Lemma 2. The dynamic program given by (24)-(26) has at least one solution at stage t if the
value function Vt+1 is continuous in πt+1.
November 23, 2018
DRAFT
29
We note that the condition of Lemma 2 is always satisfied for t = T by definition of VT +1; see
(20) and (23). However, for t < T , it is not straightforward, in general, to prove the continuity
of the value function Vt in πt. Given Vt+1 is continuous in πt+1, the result of [34, Theorem 2]
implies that the set of equilibrium payoffs for the state game at t is upper-hemicontinuous in
πt. Therefore, if the stage game Gt(πt,Vt+1,ψt+1) has a unique equilibrium payoff for every πt,
we can show that Vt is continuous in πt for t < T . Using this approach, we prove the existence
of SIB-PBEs for zero-sum games below.
Theorem 4. For every dynamic zero-sum game with asymmetric information there exists a SIB-
PBE that is a solution to the dynamic program given by (24)-(26).
For dynamic non-zero-sum games, it is harder to establish that Vt is continuous in πt for t < T
since the set of equilibrium payoffs is not a singleton in general. However, we conjecture that
for every dynamic game with asymmetric information described in Section II, at every stage of
the corresponding dynamic program, it is possible to select a BNE for every realization of πt
so that the resulting Vt is continuous in πt.
In addition to the results of Lemma 2 and Theorem 4, we provide another condition below
that guarantees the existence of SIB-PBEs in some instances of dynamic games with asymmetric
information.
Lemma 3. A dynamic game with asymmetric information described in Section II has at least
such that the SIB update rule ψ1:T is
one SIB-PBE if there exits sufficient information S1:N
1:t
independent of σ∗.
The independence of SIB update rule from σ∗ is a condition that is not satisfied for all dynamic
games with asymmetric information. Nevertheless, we present below special instances where this
condition is satisfied.
1) Nested information structure with one controller [11]: Consider the nested information
structure case described in Section II. Assume that the evolution of the system is controlled only
by the uniformed player and is given by Xt+1 = ft(Xt,A2
t = 0, it is
easy to check that Pσ∗{πt+1πt,at} =P{πt+1πt,at} for all t∈T .
t = Xt and S2
t,Wt). For S1
2) Independent dynamics with observable actions and no private valuation [18]: Consider
the model with independent dynamics and observable actions described in Section II. Assume
that agent i's, i∈N , instantaneous utility is given by ui
t ) (no private valuation); that is,
t(At,X−i
November 23, 2018
DRAFT
agent i's utility at t does not depend on X i
information for agent i. Hence, the condition of Lemma 3 is trivially satisfied.
t. It is easy to verify that Si
t = ∅ is sufficient private
30
3) Delayed sharing information structure with d = 1 [35], [36]: Consider a delayed sharing
information structure where agents' actions and observations are revealed publicly with d time
step delay. When delay d = 1, we have P i
t . Then, it is easy to verify
that the condition of Lemma 3 is satisfied.
t}. Let Si
t ={Y i
t = Y i
t = P i
4) Uncontrolled state process with hidden actions: Consider an N-player game with
uncontrolled dynamics given by Xt+1 = ft(Xt, Wt), t ∈ T . At every time t ∈ T , agent i,
i ∈ N , receives a noisy observation Y i
t). The agents' actions are hidden. Thus,
1:t−1} and Ct = ∅. Hence, the condition of Lemma 3 is trivially satisfied. We note
t ={Y i
P i
that in the case where a subset of the agents' observations is revealed to all agents' with some
delay, i.e. Ct⊆{Y1:t}, the condition of Lemma 3 is also satisfied.
t = Oi
1:t, Ai
t(X i
t , Z i
IX. CONCLUSION
We proposed a general approach to study dynamic games with asymmetric information. We
presented a set of conditions sufficient to characterize an information state for each agent that
effectively compresses his common and private information in a mutually consistent manner.
Along with the results in the companion paper [2], we showed that the characterized information
state provides a sufficient statistic for decision making purposes in strategic and non-strategic
settings. We introduced the notion of Sufficient Information based Perfect Bayesian Equilibrium
that characterizes a set of outcomes in a dynamic game. We provided a sequential decomposition
of the dynamic game over time, which leads to a dynamic program for the computation of the
set of SIB-PBEs of the dynamic game. We determined conditions that guarantee the existence of
SIB-PBEs. Using these conditions, we proved the existence of SIB-PBE for dynamic zero-sum
games and for special instances of dynamic non-zero-sum games.
REFERENCES
[1] H. Tavafoghi, Y. Ouyang, and D. Teneketzis, "A sufficient information approach to decentralized decision making," in 57th
IEEE Conference on Decision and Control (CDC), 2018.
[2] H. Tavafoghi, Y. Ouyang, and D. Teneketzis, "A unified approach to dynamic multi-agent decision problems with
asymmetric information - part i: Non-strategic agents," working paper, 2018.
[3] E. Maskin and J. Tirole, "Markov perfect equilibrium: I. observable actions," Journal of Econonomic Theory, vol. 100,
no. 2, pp. 191 -- 219, 2001.
[4] S. Zamir, "Repeated games of incomplete information: Zero-sum," Handbook of Game Theory, vol. 1, pp. 109 -- 154, 1992.
November 23, 2018
DRAFT
31
[5] F. Forges, "Repeated games of incomplete information: non-zero-sum," Handbook of Game Theory, vol. 1, pp. 109 -- 154,
1992.
[6] R. Aumann, M. Maschler, and R. Stearns, Repeated games with incomplete information. MIT press, 1995.
[7] G. Mailath and L. Samuelson, Repeated Games and Reputations. Oxford university press Oxford, 2006.
[8] J. Horner, T. Sugaya, S. Takahashi, and N. Vieille, "Recursive methods in discounted stochastic games: An algorithm for
δ → 1 and a folk theorem," Econometrica, vol. 79, no. 4, pp. 1277 -- 1318, 2011.
[9] J. Escobar and J. Toikka, "Efficiency in games with Markovian private information," Econometrica, vol. 81, no. 5, pp. 1887 --
1934, 2013.
[10] T. Sugaya, "Efficiency in Markov games with incomplete and private information," working paper, 2012.
[11] J. Renault, "The value of Markov chain games with lack of information on one side," Math. Oper. Res., vol. 31, no. 3,
pp. 490 -- 512, 2006.
[12] P. Cardaliaguet, C. Rainer, D. Rosenberg, and N. Vieille, "Markov games with frequent actions and incomplete information-
the limit case," Mathematics of Operations Research, 2015.
[13] F. Gensbittel and J. Renault, "The value of Markov chain games with incomplete information on both sides," Mathematics
of Operations Research, vol. 40, no. 4, pp. 820 -- 841, 2015.
[14] L. Li, C. Langbort, and J. Shamma, "Solving two-player zero-sum repeated Bayesian games," arXiv preprint
arXiv:1703.01957, 2017.
[15] A. Nayyar, A. Gupta, C. Langbort, and T. Bas¸ar, "Common information based Markov perfect equilibria for stochastic
games with asymmetric information: Finite games," IEEE Transactions on Automatic Control, vol. 59, pp. 555 -- 570, March
2014.
[16] A. Gupta, A. Nayyar, C. Langbort, and T. Bas¸ar, "Common information based Markov perfect equilibria for linear-Gaussian
games with asymmetric information," SIAM J. Control Optim., vol. 52, no. 5, pp. 3228 -- 3260, 2014.
[17] Y. Ouyang, H. Tavafoghi, and D. Teneketzis, "Dynamic oligopoly games with private Markovian dynamics," in 54th IEEE
Conference on Decision and Control (CDC), 2015.
[18] Y. Ouyang, H. Tavafoghi, and D. Teneketzis, "Dynamic games with asymmetric information: Common information based
perfect Bayesian equilibria and sequential decomposition," IEEE Transactions on Automatic Control, 2017.
[19] D. Vasal and A. Anastasopoulos, "Signaling equilibria for dynamic LQG games with asymmetric information," in 55th
IEEE Conference on Decision and Control (CDC), pp. 6901 -- 6908, 2016.
[20] A. Sinha and A. Anastasopoulos, "Structured perfect Bayesian equilibrium in infinite horizon dynamic games with
asymmetric information," American Control Conference, 2016.
[21] A. Nayyar, A. Mahajan, and D. Teneketzis, "Optimal control strategies in delayed sharing information structures," IEEE
Transactions on Automatic Control, vol. 56, no. 7, pp. 1606 -- 1620, 2011.
[22] A. Nayyar, A. Mahajan, and D. Teneketzis, "Decentralized stochastic control with partial history sharing: A common
information approach," IEEE Transactions on Automatic Control, vol. 58, no. 7, pp. 1644 -- 1658, 2013.
[23] H. Tavafoghi and D. Teneketzis, "Dynamic market mechanisms for wind energy," arXiv:1608.04143, 2016.
[24] H. Tavafoghi, On Design and Analysis of Cyber-Physical Systems with Strategic Agents. PhD thesis, University of Michigan,
2017.
[25] J. Renault, "The value of repeated games with an informed controller," Mathematics of Operations Research, vol. 37,
no. 1, pp. 154 -- 179, 2012.
[26] L. Li and J. Shamma, "Efficient strategy computation in zero-sum asymmetric repeated games," arXiv preprint
arXiv:1703.01952, 2017.
November 23, 2018
DRAFT
32
[27] P. Battigalli, "Strategic independence and perfect Bayesian equilibria," Journal of Economic Theory, vol. 70, no. 1, pp. 201 --
234, 1996.
[28] R. Myerson and P. Reny, "Open sequential equilibria of multi-stage games with infinite sets of types and actions," working
paper, 2015.
[29] J. Watson, "Perfect Bayesian equilibrium: general definitions and illustrations," working paper, 2016.
[30] D. Fudenberg and J. Tirole, Game theory. 1991. Cambridge, Massachusetts, 1991.
[31] M. J. Osborne and A. Rubinstein, A Course in Game Theory. MIT press, 1994.
[32] S. Sorin, A First Course on Zero-sum Repeated Games, vol. 37. Springer Science & Business Media, 2002.
[33] D. Miller, "Robust collusion with private information," The Review of Economic Studies, vol. 79, no. 2, pp. 778 -- 811, 2012.
[34] P. Milgrom and R. Weber, "Distributional strategies for games with incomplete information," Mathematics of Operations
Research, vol. 10, no. 4, pp. 619 -- 632, 1985.
[35] T. Bas¸ar, "Two-criteria LQG decision problems with one-step delay observation sharing pattern," Information and Control,
vol. 38, no. 1, pp. 21 -- 50, 1978.
[36] T. Bas¸ar, "Decentralized multicriteria optimization of linear stochastic systems," IEEE Transactions on Automatic Control,
vol. 23, no. 2, pp. 233 -- 243, 1978.
[37] E. Hendon, H. Jacobsen, and B. Sloth, "The one-shot-deviation principle for sequential rationality," Games and Economic
Behavior, vol. 12, no. 2, pp. 274 -- 282, 1996.
[38] A. Nedich, "Optimization i," Lecture Notes, 2008.
[39] E. Ok, Real Analysis with Economic Applications, vol. 10. Princeton University Press, 2007.
[40] K. Border, Fixed point theorems with applications to economics and game theory. Cambridge university press, 1989.
[41] E. Einy, O. Haimanko, and B. Tumendemberel, "Continuity of the value and optimal strategies when common priors
change.," International Journal of Game Theory, vol. 41, no. 4, 2012.
APPENDIX
Proof of Lemma 1. For any given consistent SIB assessment (σ∗,γ), let g∗ denote the behavioral
strategy profile constructed according to (16). In the following, we construct recursively a belief
system µ that is consistent with g∗ and satisfies (17).
For t = 1, we have P i
1 = Y i
1 and C1 = Z1. Define
µi
t(hi
1)(x1, p−i
1 ) :=
.
(27)
(cid:80)
P{y1, z1x1}η(x1)
x1∈X1
1, z1x1}η(x1)
P{yi
For t > 1, if Pg∗
µi
t−1
{hi
thi
t−1} > 0 (i.e. no deviation from g∗
t−1 at t−1), define µi
t recursively by
Bayes' rule,
µi
t(hi
t)(xt, p−i
t ) :=
{hi
Pg∗
µi
t−1
Pg∗
µi
t−1
t, xt, p−i
{hi
thi
hi
t−1}
t−1}
t
.
(28)
November 23, 2018
DRAFT
For t > 1, if Pg∗
µi
t−1
{hi
thi
t−1} = 0 (i.e. there is a deviation from g∗
t−1 at t−1), define µi
t as,
µi
t(hi
t)(xt, p−i
t ) :=
Pt
St
(cid:80)
33
(29)
γt(ct)(xt, st)
γ(ct)(xt, s−i
t
,
, si
t)
s−i
t ∈S−i
t
where sj
t = ζ j
t (pj
t , ct; g∗
1:t−1) for all j ∈ N .
At t = 1, (17) holds by construction from (27). For t > 1,
Pg∗{S−i
Pg∗{si
t, ct} = Pg∗{S−i
pi
t} = Pg∗{S−i
hi
t, ct} =
si
Pg∗{S−i
t
t
t
t
tct}
, si
tct} =
(cid:80)
πt(S−i
s−i
t ∈S−i
t
t
, si
t)
πt(s−i
t
, si
t)
= P{S−i
t
t, πt}
si
where the second equality follows from (11). Therefore, (17) holds for all t ∈ T .
Lemma 4. [2, Lemma 2] Given a SIB strategy profile σ∗ and update rule ψ consistent with σ∗,
ψ {St+1,Πt+1pt, ct, at} =Pσ∗
Pσ∗
ψ {St+1,Πt+1st, πt, at}.
(30)
for all st, πt, at.
Proof of Theorem 2. Consider a "super dynamic system" as the collection of the original dy-
namic system along with agents −i who play according to SIB assessment (σ∗, γψ). This
superdynamic system captures the system agent i "sees". We establish the claim of Theorem 2
in two steps: (i) we show that from agent i's viewpoint the super dynamic system is a POMDP,
t} is an information state for agent i when he faces the super
and (ii) we show that {Πt, Si
1:T (·,·). Therefore, without loss of optimality, agent
dynamic system with the original utility ui
i can choose his optimal decision strategy (best response) from the class of strategies that are
functions of the information state {Πt, Si
t}, i.e. the class of SIB strategies.
To establish step (i), consider X :={Xt,Πt,St, Πt−1,St−1} as the system state at t for the super
t ,Zt}. To show that the
dynamic system. Agent i's observation at time t is given by Y i
super dynamic system is a POMDP, we need to show that it satisfies the following properties:
(a) it has a controlled Markovian dynamics, that is,
t := {Y i
ψ {xt+1x1:t, ai
Pσ∗
1:t} = Pσ∗−i
ψ
{xt+1xt, ai
t}, ∀t ∈ T ,
(31)
1:t, yi
(b) agent i's observation Y i
t is a function of system state Xt along with the previous action Ai
t−1,
November 23, 2018
DRAFT
that is,
ψ {yi
Pσ∗
tx1:t,ai
1:t−1, yi
1:t−1} = Pσ∗−i{yi
txt,ai
t−1}, ∀t ∈ T ,
34
(32)
(c) agent i's instantaneous utility at t ∈ T can be written as a function ut(xt,ai
Xt along with his action Ai
t) of system state
(cid:8)ui
t, that is,
t(Xt,At)x1:t,ai
Eσ∗
ψ
1:t, yi
1:t
(cid:9) = ut(xt,ai
t), ∀t ∈ T .
(33)
Property (a) is true because,
ψ {xt+1x1:t, ai
Pσ∗
(cid:88)
(cid:104)Pσ∗
(cid:88)
a−i
t ,zt+1,yt+1
a−i
t ,zt+1,yt+1
(cid:88)
(cid:104)P
1:t} = Pσ∗
ψ {xt+1,πt+1,st+1,πt,stx1:t,π1:t,s1:t,yi
1:t, yi
1:t,z1:t,ai
1:t}
=
ψ {xt+1,πt+1,st+1,a−i
Pσ∗
t ,zt+1,yt+1x1:t,π1:t,s1:t,yi
1:t}
1:t,z1:t,ai
by system dynamics (1) and (2)
=
ψ {πt+1,st+1x1:t,π1:t,s1:t,yi
1:t,z1:t,ai
1:t,a−i
P{zt+1,yt+1xt+1,at}P{xt+1xt,at}σ∗−i
Define Z := {zt+1 : πt+1 = ψt+1(πt, zt+1)} &
t+1, zt+1, aj
Y(zt+1) := {yt+1 : sj
t,{yj
t
t}),∀j}
t+1 = φj
t+1(sj
=
t ,xt+1,zt+1,yt+1}
(cid:105)
(πt,s−i
t )(a−i
t )
(cid:105)
ψ{πt+1,st+1st,πt,yt+1,xt+1,zt+1,at}P{yt+1,zt+1xt+1,at}P{xt+1xt,at}σ∗−i
t
(πt,s−i
t )(a−i
t )
a−i
t , zt+1 ∈ Z,
yt+1 ∈ Y(zt+1)
=
ψ {xt+1,πt+1,st+1xt,πt,st,ai
Pσ∗−i
t} = Pσ∗−i
ψ {xt+1xt,ai
t}.
November 23, 2018
DRAFT
Property (b) is true because
35
ψ {yi
Pσ∗
1:t−1, yi
1:t−1} = Pσ∗
ψ {yi
t,ztx1:t,π1:t,s1:t,yi
1:t−1,z1:t−1,ai
1:t−1}
tx1:t,ai
(cid:88)
a−i
t−1
(cid:104)Pσ∗
(cid:88)
ψ {yi
a−i
t−1
=
ψ {yi
Pσ∗
t,zt,a−i
t−1x1:t,π1:t,s1:t,yi
1:t−1,z1:t−1,ai
1:t−1}
=
1:t−1,z1:t−1,ai
1:t−1,a−i
t−1}σ∗−i
t
(πt−1,st−1)(a−i
t−1)
(cid:105)
t,ztx1:t,π1:t,s1:t,yi
(cid:88)
P{yi
t,ztxt,ai
a−i
t−1
by system dynamics (2)
=
t−1}σ∗−i
t−1,a−i
t
(πt−1,st−1)(a−i
t−1)
Property (c) is true because
ψ {yi
txt,ai
t−1}.
=
t,ztxt,πt−1,st−1,ai
t−1} = Pσ∗
(cid:9)
t(Xt,At)x1:t,ai
1:t,yi
1:t
1:t,yi
1:t
t(Xt,At)xt,πt,st,x1:t−1,ai
t(Xt,(ai
t,σ∗−i(πt,s−i
t,σ∗−i(πt,s−i
t ))) := ut(xt,ai
t)
Pσ∗−i{yi
ψ
(cid:8)ui
(cid:8)ui
(cid:8)ui
Eσ∗
=Eσ∗
=Eσ∗−i
ψ
ψ
= ui
t(xt,(ai
(cid:9)
t )))xt,πt,st,x1:t−1,ai
1:t,yi
1:t
(cid:9)
To establish step (ii), that is, to show that {Πt,Si
t} is an information state for agent i when
he interacts with the superdynamic system defined above based on SIB assessment (σ∗, γψ), we
need to prove: (1) it can be updated recursively at t, i.e. it can be determined using {Πt−1,Si
t−1}
and {Y i
t} ={Y i
t} is
t ,Zt,Ai
t,Ai
independent of H i
t and his actual strategy gi; and (3) it is sufficient to evaluate the agent i's
instantaneous utility at t for every action ai
t}; (2) agent i's belief about {Πt+1,Si
t+1} conditioned on {Πt,Si
t, for all t∈T .
t∈Ai
t ,Ai
Condition (1) is satisfied since Πt = ψt(Πt−1,Zt) and Si
t = φt(Si
t−1,{Y i
t ,Zt,Ai
t}) for t ∈ T \{1};
see part (i) of Definition 4 and (15).
November 23, 2018
DRAFT
To prove condition (2), let
t (ct)),
for all j ∈ N and t ∈ T . Then condition (2) is satisfied since
t ) = σ
∗j
t (hj
g
∗j
t (l(hj
t ), γψ
(cid:88)
t,ai
t} =
h−i
t ,a−i
t
ψ{si
Pσ∗
t+1,πt+1,h−i
t ,a−i
t hi
t}
t,ai
by Theorem 1 and (34)
=
t+1,πt+1ht,at}Pg∗−i{h−i
t hi
t}g∗−i
t
(h−i
t )(a−i
t )
ψ {si
t+1,πt+1hi
Pσ∗
(cid:88)
(cid:88)
ψ{si
Pσ∗
h−i
t ,a−i
t
ψ{si
Pσ∗
t ,s−i
t+1
h−i
t ,a−i
h−i
t ,a−i
(cid:88)
(cid:88)
(cid:88)
s−i
t ,a−i
=
t+1,πt+1,s−i
t+1ht,at}Pg∗−i{h−i
t hi
t}g∗−i
t
(h−i
t )(a−i
t )
by Lemma 4 and s−i
t = ζ−i
; g∗
1:t−1) (see Definition 4)
t (h−i
=
t
ψ{st+1,πt+1st,πt,at}Pg∗−i{s−i
Pσ∗
t ,s−i
t+1
t hi
t}g∗−i
t
(h−i
t )(a−i
t )
by part (ii) of Lemma 1
=
ψ{st+1,πt+1st,πt,at}P{s−i
t si
Pσ∗
t ,s−i
t+1
t,πt}g∗−i
t
(h−i
t )(a−i
t )
by (34)
=
ψ{si
Pσ∗
t ,s−i
t+1
h−i
t ,a−i
t+1,πt+1,s−i
t+1ht,at}P{s−i
t si
t,πt}σ∗−i
t
(πt,s−i
t )(a−i
t )
=
ψ{si
Pσ∗
t+1,πt+1si
t}.
t,πt,ai
.
To prove condition (3), we need to show that for all ai
Eg∗−i{ut
i(Xt,A−i
t ,ai
t)hi
t} =Eg∗−i{ut
t},
t)πt,si
t ∈ Ai
t,
i(Xt,A−i
t ,ai
for all hi
t, πt, si
t, t ∈ T .
November 23, 2018
36
(34)
(35)
DRAFT
By Lemma 1,
Pg∗−i{s−i
t
t} = P{s−i
hi
t
t}.
πt, si
Setting πt = γψ(ct), and A−i
ψ {ui
Eσ∗
t
ψ
t ,ai
t)hi
t = σ∗−i(πt, S−i
t ),
t} =Eσ∗
ψ {ui
t(Xt,σ∗−i
t(Xt,σ∗−i
t(Xt,σ∗−i
t(Xt,σ∗−i
(πt,S−i
(cid:110)Eσ∗−i{ui
(cid:110)Eσ∗−i{ui
(cid:110)Eσ∗−i{ui
(cid:110)Eσ∗−i{ui
t(Xt,A−i
= Eσ∗
= Eσ∗
= Eσ∗−i
= Eσ∗−i
= Eσ∗−i
t(Xt,σ∗−i
ψ {ui
ψ
ψ
ψ
t
t
t
t
t
t}(cid:12)(cid:12)(cid:12)hi
(cid:111)
t,ct}(cid:12)(cid:12)(cid:12)hi
(cid:111)
t,ct}(cid:12)(cid:12)(cid:12)hi
(cid:111)
t}(cid:12)(cid:12)(cid:12)hi
(cid:111)
t
t
t
t ),ai
t)hi
t}
i,hi
t ,πt,st
t
t(Xt,σ∗−i
(πt,S−i
t)S−i
(πt,S−i
t ),ai
t)S−i
(πt,S−i
t ),ai
t ,πt,si
t)S−i
(πt,S−i
t)S−i
(πt,S−i
t ),ai
t)πt,si
t}.
(πt,S−i
t = σ∗−i
t ),ai
t ),ai
t ,πt,si
t ,πt,si
37
(36)
(37)
The first equality above is by substituting A−i
t ). The second equality follows
from the smoothing property of conditional expectation. The third equality holds by condition
(iii) of Definition 4. The fourth equality follows from Theorem 1 since S−i
is a function of H−i
.
The fifth equality holds since for every xt, st, πt, ct,
t
t
t
P{xtst,πt,ct} =
P{xt,stπt,ct}
P{st,πt,ct} =
πt(xt,st)
πt(xt,st)
xt
= P{xtst,πt}.
(cid:80)
The last equality is true by (36). By (37) we prove condition (3) for {Πt,Si
state, and thus establish the result of Theorem 2.
t} to be an information
Proof of Theorem 3. Let (σ∗,ψ) denote a solution of the dynamic program. We note that the SIB
update rule ψ is consistent with σ∗ by requirement (25). Therefore, we only need to show that
the SIB assessment (σ∗,ψ) is sequentially rational. To prove it, we use the one-shot deviation
principle for dynamic games with asymmetric information [37]. To state the one-shot deviation,
we need the following definitions.
Definition 7 (One-shot deviation). We say gi is a one-shot deviation from g∗i if there exists a
unique hi
Definition 8 (Profitable one-shot deviation). Consider an assessment (g∗,µ). We say gi is a
profitable one-shot deviation for agent i if gi is a one-shot deviation from g∗i at hi
t such that
t∈ Hi such that gi
τ) for all hi
τ)(cid:54)= g∗i
τ ∈ Hi.
t)(cid:54)= g∗i
t), and gi
τ (cid:54)= hi
t, hi
t (hi
τ (hi
τ(hi
t(hi
November 23, 2018
DRAFT
t)(cid:54)= g∗i
gi
t(hi
t (hi
t), and
(cid:41)
(cid:12)(cid:12)(cid:12)hi
t
ui
τ(Xτ,Aτ)
(cid:40) T(cid:88)
>E(g∗−i,g∗i)
µ
τ =t
38
(cid:41)
(cid:12)(cid:12)(cid:12)hi
t
ui
τ(Xτ,Aτ)
(cid:40) T(cid:88)
τ =t
E(g∗−i,gi)
µ
One-shot deviation principle [37]: A consistent assessment (g∗,µ) is a PBE if and only if
there exists no agent that has a profitable one-shot deviation.
Below, we show that the consistent SIB assessment (σ∗,ψ) satisfies the sequential rationality
condition using the one-shot deviation principle.
Consider an arbitrary agent i∈ N , time t∈ T , and history realization hi
t. Agent i has a
profitable one-shot deviation at hi
t only if
σ∗i
t (πt,si
t) /∈ arg max
Eσ∗
t)∈∆(Ai
t)
gi
t(hi
π
t((σ∗(πt,St)),St,πt,Vt+1,ψt+1)
Given (πt,Vt+1,ψt+1,σ∗
t ), the expected value of the function ¯U i
t, agent i's belief about S−i
given hi
function of si
about S−i
maximization problem above can be written as a function of πt and si
t(πt,si
σi
t is only a function of si
t conditioned on hi
t is only a
t). Agent i's belief
t and πt (see (17)). Therefore, any solution to the
t, that is, it is a SIB strategy
t) for agent i. Consequently, agent i has a profitable one-shot deviation only if
, as well as agent i's strategy gi
t(hi
t
t
σ∗i
t (πt,si
t) /∈ arg max
Eσ∗
t)∈∆(Ai
t)
σi
t(πt,si
π
t((σ∗(πt,St)),St,πt,Vt+1,ψt+1)
By (24), σ∗
t is BNE of the stage game Gt(πt, Vt+1, ψt+1) at t (see also (21)), i.e.
σ∗i
t (πt,si
t)∈ arg max
Eσ∗
t)∈∆(Ai
t)
σi
t(πt,si
π
t((σ∗(πt,St)),St,πt,Vt+1,ψt+1)
t (πt,si
t) at hi
Consequently, there exists no profitable deviation from σ∗i
t. Therefore, there exists
no agent that has a profitable one-shot deviation. Hence, by one-shot deviation principle, the
consistent SIB assessment (σ∗,ψ) is sequentially rational, and thus, it is a SIB-PBE.
Proof of Lemma 2. We prove below that if Vt+1(·,st+1) is continuous in πt+1, then the dynamic
program has a solution at stage t, t ∈ T ; that is, there exists at least one σ∗
t ∈
BNEt(Vt+1,ψt+1), where ψt+1 is consistent with σ∗
t .
For every πt, define a perturbation of the stage game Gt(πt,Vt+1,ψt+1) by restricting the set of
strategies of each agent to mixed strategies that assign probability of at least > 0 to every action
t ∈ Ai
t of agent i ∈ N ; for every agent i ∈ N we denote this class of -restricted strategies by
ai
t such that σ∗
November 23, 2018
DRAFT
(cid:110)¯U i
(cid:110)¯U i
(cid:110)¯U i
t∈ Hi
(cid:12)(cid:12)(cid:12)hi
(cid:111)
.
t
t
(cid:12)(cid:12)(cid:12)πt,si
(cid:111)
(cid:12)(cid:12)(cid:12)πt,si
(cid:111)
t
.
.
update rule ψ
t ∈ Σ
∗,
σ
(see (12))
(cid:40)
t := Σ1,
t and Σ
Σi,
perturbed stage game has an equilibrium σ
t ×. . .×ΣN,
t
. In the following we prove that, for every > 0, the corresponding
∗,
t
along with a consistent update rule ψ
t+1.
39
We note that when the agents' equilibrium strategies are perfectly mixed strategies, then the
t+1 is completely determined via Bayes' rule. Therefore, for every strategy profile
t+1 := βt+1(σ∗,), where βt+1(σ∗,) is Bayes' rule where σ∗, is utilized
t we can write ψ
For every agent i ∈ N , define a best response correspondence BRi,
: Σ
t
t
⇒ Σi,
BRi,
t (σ
∗,
t ) :=
t ∈ Σi,
σi
t : σi
t(πt,si
Eσ
∗−i,
t
,σi,
t{¯U i
t(At,St,πt,Vt+1,βt+1(σ
t ))si
∗,
t)∈ arg max
t ∈Σi,
σi,
t
t as
(cid:41)
t,πt},∀πt,si
t
, (38)
which determines the set of all agent i's best responses within the class of -restricted strategies
assuming that agents −i are playing σ
and the update rule ψ
t+1 = βt+1(σ
∗−i,
t
∗,
t ).
For every i∈N and σ∗,∈ Σ
t, we prove below that BRi,
t (σ
∗,
t ) is non-empty, convex, closed,
and upper hemicontinuous.
We note that
t(At,St,πt,Vt+1,βt+1(σ
∗,
t ,σi,
t {¯U i
Eσ
∗−i,
t
,Ai
t=ai
t{¯U i
t ))si
∗,
t,πt}
t ))si
∗,
t(At,St,πt,Vt+1,βt+1(σ
t,πt}σi,
t (πt,si
t)(ai
t)
Eσ
(cid:88)
(cid:88)
ai
t
=
=
U
∗,
πt,si
t
(ai
t)σi,
t (πt,si
t)(ai
t),
ai
t
where
Therefore, for every πt,si
t, we have
U σ∗,
πt,si
t
t) :=Eσ
(ai
∗−i,
t
,Ai
t=ai
t{¯U i
t(At,St,πt,Vt+1,βt+1(σ
t ))si
∗,
t,πt}.
(cid:88)
ai
t
U σ∗,
πt,si
t
(ai
t)α(ai
t).
σi,
t (πt,si
arg max
α∈∆(Ai
t):α(ai
t)≥,∀ai
t
t) ∈
(cid:80)
U σ∗,
πt,si
t
t
ai
t
t):α(ai
t)α(ai
t)≥,∀ai
We note that maxα∈∆(Ai
(ai
of [38], the set of agent i's best responses BRi,
in πt+1 then Vt+1 is continuous in agent i's strategy σi
is continuous in agent i's strategy σi
strategy σi
t) is a linear program, thus, by Theorem 16
∗,
t ) is closed and convex. If Vt+1 is continuous
t. Moreover, the instantaneous utility ui
t , given by (20), is continuous in agent i'
t. Therefore, by the maximum theorem [39] the set of i's best responses in upper
t. Therefore, ¯U i
t (σ
t
November 23, 2018
DRAFT
∗,
hemicontinuous in σ
t
and non-empty.
Consequently, we establish that for every i ∈ N , BRi,
40
∗,
t ) is closed, convex, upper hemi-
t :=×i∈N BRi,
t where× denotes
t (σ
t. Define BR
t(σ
t (σ
t ∈ Σ
∗,
continuous, and non-empty for every σ
the Cartesian product. The correspondence BR
t since BRi,
and non-empty for every σ
and non-empty for every σ
[40, Corollary 15.3], the correspondence BR
game has an equilibrium σ
t ∈ Σ
∗,
t ∈ Σi,
∗i,
t
∗,
t
∗,
t ) is closed, convex, upper hemicontinuous,
∗,
t ) is closed, convex, upper hemicontinuous,
for all i∈N . Therefore, by Kakutani's fixed-point theorem
t has a fixed point. Therefore, every perturbed stage
along with a consistent update rule ψt+1 = βt+1(σ
∗,
t ).
Now consider the sequence of these perturbed games when → 0. Since the set of agents'
strategies is compact, there exists a subsequence of these perturbed games whose equilibrium
∗,
strategies converge, say to σ∗
t ). We
∗,
note that ψ∗
. We
t
show below that for every agent i ∈ N , σ∗i
t+1 when he
chooses his strategy from the unconstrained class of SIB strategies.
t(σ∗
t+1 denote the convergence point of βt+1(σ
∗,
t ) (i.e. Bayes' rule) is continuous in σ
As we proved above, the set of agent i's best responses BRi
is a best response for him given Vt+1,ψ∗
t+1 is consistent with σ∗
t . Similarly let ψ∗
t since βt+1(σ
t
t ) is upper hemicontinuous
t (πt,·) is also a best response for agent i in the stage game
t ∈ BNEt(Vt+1, ψ∗
t+1 is consistent with σ∗
t .
t+1) where ψ∗
and closed given ψ∗
Gt(πt,Vt+1,ψ∗
t+1. Therefore, σ∗
t+1). Consequently, σ∗
Proof of Theorem 4. We have a Bayesian zero-sum game with finite state and action spaces.
By [41, Theorem 1] the equilibrium payoff is a continuous function of the agents' common
prior/belief. Using this result, we prove, by backward induction, that every stage of the dynamic
program described by (24)-(26), has a solution and Vt is continuous in πt for all t.
For t = T +1 the dynamic program has a solution trivially since the agents have utility for
time less than or equal to T . Moreover, VT +1(.,.) = 0 is trivially continuous in πT +1.
For t≤ T , assume that Vt+1 is continuous in πt+1. Then, by Lemma 2 the dynamic program
has a solution at t. We note that the continuation game from t to T is a dynamic zero-sum
game with finite state and actions spaces. Therefore, as we argued above, by [41, Theorem 1]
the agents' equilibrium payoff at t (i.e. Vt) is unique and is continuous in the agents' common
prior given by πt.
Therefore, by induction we establish the assertion of Theorem 4.
Proof of Lemma 3. Assume that ψ1:T is independent of σ. Then, the evolution of Πt is in-
dependent of σ∗ and known a priori. As a result, we can ignore the consistency condition
November 23, 2018
DRAFT
41
(25) in the dynamic program. Given ψt+1, the stage game Gt(πt,Vt+1,ψt+1) is a static game of
incomplete information with finite actions (given by A1:N
) for
every πt. Therefore, by the standard existence results for finite games [30, Theorem 1.1], there
exists an equilibrium for the stage game BNEt(Vt+1,ψt+1). Consequently, the correspondence
BNEt(Vt+1,ψt+1) is non-empty for every t ∈ T , thus, the dynamic programming given by (24-
26) has a solution.
) and finite types (given by S 1:N
t
t
November 23, 2018
DRAFT
|
1011.5914 | 1 | 1011 | 2010-11-26T21:17:31 | Static and Expanding Grid Coverage with Ant Robots : Complexity Results | [
"cs.MA"
] | In this paper we study the strengths and limitations of collaborative teams of simple agents. In particular, we discuss the efficient use of "ant robots" for covering a connected region on the Z^{2} grid, whose area is unknown in advance, and which expands at a given rate, where $n$ is the initial size of the connected region.
We show that regardless of the algorithm used, and the robots' hardware and software specifications, the minimal number of robots required in order for such coverage to be possible is \Omega({\sqrt{n}}).
In addition, we show that when the region expands at a sufficiently slow rate, a team of \Theta(\sqrt{n}) robots could cover it in at most O(n^{2} \ln n) time.
This completion time can even be achieved by myopic robots, with no ability to directly communicate with each other, and where each robot is equipped with a memory of size O(1) bits w.r.t the size of the region (therefore, the robots cannot maintain maps of the terrain, nor plan complete paths).
Regarding the coverage of non-expanding regions in the grid, we improve the current best known result of O(n^{2}) by demonstrating an algorithm that guarantees such a coverage with completion time of O(\frac{1}{k} n^{1.5} + n) in the worst case, and faster for shapes of perimeter length which is shorter than O(n). | cs.MA | cs |
Static and Expanding Grid Coverage with Ant Robots :
Complexity Results
Yaniv Altshulera,b,∗, Alfred M. Brucksteina
aTechnion -- Israel Institute of Technology, Technion City, Haifa 32000, Israel
bDeutsche Telekom Laboratories, Ben Gurion University of the Negev, POB 653, Beer
Sheva 84105, Israel
Abstract
In this paper we study the strengths and limitations of collaborative teams
of simple agents. In particular, we discuss the efficient use of "ant robots"
for covering a connected region on the Z2 grid, whose area is unknown in
advance, and which expands at a given rate, where n is the initial size of the
connected region. We show that regardless of the algorithm used, and the
robots' hardware and software specifications, the minimal number of robots
required in order for such coverage to be possible is Ω(√n).
we show that when the region expands at a sufficiently slow rate, a team
of Θ(√n) robots could cover it in at most O(n2 ln n) time. This completion
time can even be achieved by myopic robots, with no ability to directly
In addition,
communicate with each other, and where each robot is equipped with a
memory of size O(1) bits w.r.t the size of the region (therefore, the robots
cannot maintain maps of the terrain, nor plan complete paths). Regarding
∗Corresponding Author (Tel : +972.544.950093 Fax : +972.153.544.950093)
Email addresses: [email protected] (Yaniv Altshuler),
[email protected] (Alfred M. Bruckstein)
URL: www.cs.technion.ac.il/~yanival (Yaniv Altshuler),
www.cs.technion.ac.il/~freddy (Alfred M. Bruckstein)
Preprint submitted to Theory of Computer Science
September 27, 2018
the coverage of non-expanding regions in the grid, we improve the current
best known result of O(n2) by demonstrating an algorithm that guarantees
such a coverage with completion time of O( 1
k n1.5 + n) in the worst case, and
faster for shapes of perimeter length which is shorter than O(n).
Keywords: Collaborative Cleaning, Collaborative Search, Decentralized
Systems, Grid Search, Expanding Domains
1. Introduction
Motivation. In nature, ants, bees or birds often cooperate to achieve com-
mon goals and exhibit amazing feats of swarming behavior and collaborative
problem solving. It seems that these animals are "programmed" to interact
locally in such a way that the desired global behavior will emerge even if
some individuals of the colony die or fail to carry out their task for some
other reasons. It is suggested to consider a similar approach to coordinate
a group of robots without a central supervisor, by using only local interac-
tions between the robots. When this decentralized approach is used much of
the communication overhead (characteristic to centralized systems) is saved,
the hardware of the robots can be fairly simple, and better modularity is
achieved. A properly designed system should be readily scalable, achieving
reliability and robustness through redundancy.
Multi-Agent Robotics and Swarm Robotics. Significant research effort
has been invested during the last few years in design and simulation of multi-
agent robotics and intelligent swarm systems, e.g. (1 -- 3).
Swarm based robotic systems can generally be defined as highly decentral-
ized collectives, i.e. groups of extremely simple robotic agents, with limited
2
communication, computation and sensing abilities, designed to be deployed
together in order to accomplish various tasks.
Tasks that have been of particular interest to researchers in recent years
include synergetic mission planning (4, 5), patrolling (6, 7), fault-tolerant
cooperation (8, 9), network security (10), swarm control (11, 12), design of
mission plans (13, 14), role assignment (15 -- 17), multi-robot path planning
(7, 18, 19), traffic control (20, 21), formation generation (22 -- 24), formation
keeping (25, 26), exploration and mapping (27 -- 29), target tracking (30, 31)
and distributed search, intruder detection and surveillance (32, 33).
Unfortunately, the mathematical / geometrical theory of such multi agent
systems is far from being satisfactory, as pointed out in (34 -- 37) and many
other papers.
Multi Robotics in Dynamic Environments. The vast majority of the
works mentioned above discuss challenges involving a multi agent system
operating on static domains. Such models, however, are often too limited
to capture "real world" problems which, in many cases, involve external ele-
ment, which may influence their environment, activities and goals. Designing
robotic agents that can operate in such environments presents a variety of
mathematical challenges.
The main difference between algorithms designed for static environments
and algorithms designed to work in dynamic environments is the fact that the
agents' knowledge base (either central or decentralized) becomes unreliable,
due to the changes that take place in the environment. Task allocation,
cellular decomposition, domain learning and other approaches often used by
multi agents systems -- all become impractical, at least to some extent.
3
Hence, the agents' behavior must ensure that the agents generate a desired
effect, regardless the changing environment.
One example is the use of multi agents for distributed search. While many
works discuss search after "idle targets", recent works considered dynamic
targets, meaning targets which while being searched for by the searching
robots, respond by performing various evasive maneuvers intended to pre-
vent their interception. This problem, dating back to World War II opera-
tions research (see e.g. (38, 39)), requires the robotic agents to cope with
a search area that expands while scanned. The first documented example
for search in dynamic domains discussed a planar search problem, consider-
ing the scanning of a corridor between parallel borders. This problem was
solved in (40) in order to determine optimal strategies for aircraft searching
for moving ships in a channel.
A similar problem was presented in (41), where a system consisting of a
swarm of UAVs (Unmanned Air Vehicles) was designed to search for one or
more "smart targets" (representing for example enemy units, or alternatively
a lost friendly unit which should be found and rescued). In this problem the
objective of the UAVs is to find the targets in the shortest time possible.
While the swarm comprises relatively simple UAVs, lacking prior knowledge
of the initial positions of the targets, the targets are assumed to be adversar-
ial and equipped with strong sensors, capable of telling the locations of the
UAVs from very long distances. The search strategy suggested in (41) defines
flying patterns for the UAVs to follow, designed for scanning the (rectangu-
lar) area in such a way that the targets cannot re-enter areas which were
already scanned by the swarm without being detected. This problem was
4
further discussed in (42), where an improved decentralized search strategy
was discussed, demonstrating nearly optimal completion time, compared to
the theoretical optimum achievable by any search algorithm.
Collaborative Coverage of Expanding Domains. In this paper we shall
examine a problem in which a group of ant-like robotic agents must cover an
unknown region in the grid, that possibly expands over time. This problem
is also strongly related to the problem of distributed search after mobile
and evading target(s) (42 -- 44) or the problems discussed under the names of
"Cops and Robbers" or "Lions and Men" pursuits (45 -- 48).
We analyze such issues using the results presented in (49 -- 51), concerning
the Cooperative Cleaners problem, a problem that assumes a regular grid
of connected 'pixels' / 'tiles' / 'squares' / 'rooms', part of which are 'dirty',
the 'dirty' pixels forming a connected region of the grid. On this dirty grid
region several agents move, each having the ability to 'clean' the place (the
'room', 'tile', 'pixel' or 'square') it is located in.
In the dynamic variant
of this problem a deterministic evolution of the environment in assumed,
simulating a spreading contamination (or spreading fire). In the spirit of (52)
we consider simple robots with only a bounded amount of memory (i.e. finite-
state-machines).
First, we discuss the collaborative coverage of static grids. We demon-
strate that the best completion time known to date (O(n2), achievable for
example using the LRTA* search algorithm (53)) can be improved to guar-
antee grid coverage in O( 1
k n1.5 + n).
Later, we discuss the problem of covering an expanding domain, namely --
a region in which "covered" tiles that are adjacent to "uncovered" tiles be-
5
come "uncovered" every once in a while. Note that the grid is infinite, namely
although initial size of the region is n, it can become much greater over time.
We show that using any conceivable algorithm, and using as sophisticated
and potent robotic agents as possible, the minimal number of robots below
which covering such a region is impossible equals Ω(√n). We then show
that when the region expands sufficiently slow, specifically -- every O( c0
)
γ1
time steps (where c0 is the circumference of the region and where γ1 is a
geometric property of the region, which ranges between O(1) and O(ln n)),
a group of Θ(√n) robots can successfully cover the region. Furthermore, we
demonstrate that in this case a cover time of O(n2 ln n) can be guaranteed.
These results are the first analytic results ever concerning the complexity
of the number of robots required to cover an expanding grid, as well as for
the time such a coverage requires.
2. Related Work
In general, most of the techniques used for the task of a distributed cov-
erage use some sort of cellular decomposition. For example, in (54) the area
to be covered is divided between the agents based on their relative locations.
In (55) a different decomposition method is being used, which is analytically
shown to guarantee a complete coverage of the area. Another interesting
work is presented in (56), discussing two methods for cooperative coverage
(one probabilistic and the other based on an exact cellular decomposition).
All of the works mentioned above, however, rely on the assumption that
the cellular decomposition of the area is possible. This in turn, requires the
use of memory resources, used for storing the dynamic map generated, the
6
boundaries of the cells, etc'. As the initial size and geometric features of the
area are generally not assumed to be known in advance, agents equipped with
merely a constant amount of memory will not be able to use such algorithms.
Surprisingly, while some existing works concerning distributed (and de-
centralized) coverage present analytic proofs for the ability of the system
to guarantee the completion of the task (for example, in (55 -- 57)), most of
them lack analytic bounds for the coverage time (although in many cases an
extensive amount of empirical results of this nature are made available by
extensive simulations).
An interesting work discussing a decentralized coverage of terrains is pre-
sented in (58). This work examines domains with non-uniform traversability.
Completion times are given for the proposed algorithm, which is a general-
ization of the forest search algorithm (59). In this work, though, the region
to be searched is assumed to be known in advance -- a crucial assumption
for the search algorithm, which relies on a cell-decomposition procedure.
A search for analytic results concerning the completion time of ant-robots
covering an area in the grid revealed only a handful of works. The main result
in this regard is that of (60, 61), where a swarm of ant-like robots is used
for repeatedly covering an unknown area, using a real time search method
called node counting. By using this method, the robots are shown to be able
to efficiently perform such a coverage mission (using integer markers that are
placed on the graph's nodes), and analytic bounds for the coverage time are
discussed. Based on a more general result for strongly connected undirected
graphs shown in (62, 63), the cover time of teams of ant robots (of a given
size) that use node counting is shown to be tk(n) = O(n√n), when tk(n) the
7
cover time of a region of size n using k robots. It should be mentioned though,
that in (60) the authors clearly state that it is their belief that the coverage
time for robots using node counting in grids is much smaller. This evaluation
is also demonstrated experimentally. However, no analytic evidence for this
was available thus far.
Another algorithm to be mentioned in this scope is the LRTA* search
algorithm. This algorithm was first introduced in (53) and its multi-robotics
variant is shown in (62) to guarantee cover time of undirected connected
graphs in polynomial time. Specifically, on grids this algorithm is shown to
guarantee coverage in O(n2) time (again, using integer markers).
Vertex-Ant-Walk, a variant of the node counting algorithm is presented in
(64), is shown to achieve a coverage time of O(nδG), where δG is the graph's
diameter. Specifically, the cover time of regions in the grid is expected to be
O(n2) (however for various "round" regions, a cover time of approximately
O(n1.5) can be achieved). This work is based on a previous work in which a
cover time of O(n2δG) was demonstrated (65).
Another work called Exploration as Graph Construction, provides a cov-
erage of degree bounded graphs in O(n2) time, can be found in (66). Here a
group of limited ant robots explore an unknown graph using special "mark-
ers".
Interestingly, a similar performance can be obtained by using the simplest
algorithm for multi robots navigation, namely -- random walk. Although in
general undirected graphs a group of k random walking robots may require
up to O(n3) time, in degree bounded undirected graphs such robots would
achieve a much faster covering, and more precisely, O(cid:16)E2log3n
k2
(cid:17) (67). For
8
regular graphs or degree bounded planar graphs a coverage time of O(n2)
can be achieved (68), although in such case there is also a lower bound for
the coverage time, which equals Ω(n(log n)2) (69).
We next show that the problem of collaborative coverage in static grid do-
mains can be completed in O( 1
k n1.5 + n) time and that collaborative coverage
of dynamic grid domains can be achieved in O(n2 ln n).
3. The Dynamic Cooperative Cleaners Problem
Following is a short summary of the Cooperative Cleaners problem, as
appears in (49) (static variant) and (50, 51) (dynamic variant).
We shall assume that the time is discrete. Let the undirected graph
G(V, E) denote a two dimensional integer grid Z2, whose vertices (or "tiles")
have a binary property called 'contamination'. Let contt(v) state the con-
tamination state of the tile v at time t, taking either the value "on" or
"off ". Let Ft be the contaminated sub-graph of G at time t, i.e. : Ft =
{v ∈ G contt(v) = on}. We assume that F0 is a single connected compo-
nent. Our algorithm will preserve this property along its evolution.
Let a group of k robots that can move on the grid G (moving from a
tile to its neighbor in one time step) be placed at time t0 on F0, at point
p0 ∈ Ft. Each robot is equipped with a sensor capable of telling the con-
tamination status of all tiles in the digital sphere of diameter 7, surrounding
the robot (namely, in all the tiles that their Manhattan distance from the
robot is equal or smaller than 3. See an illustration in Figure 1). A robot
is also aware of other robots which are located in these tiles, and all the
robots agree on a common direction. Each tile may contain any number of
9
robots simultaneously. Each robot is equipped with a memory of size O(1)
bits1. When a robot moves to a tile v, it has the possibility of cleaning this
tile (i.e. causing cont(v) to become off. The robots do not have any prior
knowledge of the shape or size of the sub-graph F0 except that it is a single
and simply connected component.
X
Figure 1: An illustration of a digital sphere of diameter 7, placed around a robot.
The contaminated region Ft is assumed to be surrounded at its bound-
ary by a rubber-like elastic barrier, dynamically reshaping itself to fit the
evolution of the contaminated region over time. This barrier is intended to
guarantee the preservation of the simple connectivity of Ft, crucial for the
operation of the robots, due to their limited memory. When a robot cleans
a contaminated tile, the barrier retreats, in order to fit the void previously
occupied by the cleaned tile. Every d time steps, the contamination spreads.
That is, if t = nd for some positive integer n, then :
∀v ∈ Ft ∀u ∈ 4 − N eighbors(v) , contt+1(u) = on
1For counting purposes the agents must be equipped with counters that can store the
number of agents in their immediate vicinity. This can of course be implemented using
O(log k) memory. However, throughout the proof of Lemma 5 in (49) it is shown that the
maximal number of agents that may simultaneously reside in the same tile at any given
moment is upper bounded by O(1). Therefore, counting the agents in the immediate
vicinity can be done using counters of O(1) bits.
10
Here, the term 4 − N eighbors(v) simply means the four tiles adjacent to
tile v (namely, the tiles whose Manhattan distance from v equals 1). While
the contamination spreads, the elastic barrier stretches while preserving the
simple-connectivity of the region, as demonstrated in Figure 2. For the robots
who travel along the tiles of F , the barrier signals the boundary of the con-
taminated region.
Figure 2: A demonstration of the barrier expansion process as a result of a contamination
spread.
The robots' goal is to clean G by eliminating the contamination entirely.
It is important to note that no central control is allowed, and that the
system is fully decentralized (i.e. all robots are identical and no explicit com-
munication between the robots is allowed). An important advantage of this
approach, in addition to the simplicity of the robots, is fault-tolerance -- even
if almost all the robots die and evaporate before completion, the remaining
ones will eventually complete the mission, if possible.
A Survey of Previous Results. The cooperative cleaners problem was
previously studied in (49) (static version) and (50, 51) (dynamic version). A
cleaning algorithm called SWEEP was proposed (used by a decentralized
group of simple mobile robots, for exploring and cleaning an unknown "con-
11
taminated" sub-grid F , expanding every d time steps) and its performance
analyzed.
The SWEEP algorithm is based in a constant traversal of the contami-
nated region, preserving the connectivity of the region while cleaning all non
critical points -- points which when cleaned disconnect the contaminated
region. Using this algorithm the agents are guaranteed to stop only upon
completing their mission. The algorithm can be implemented using only local
knowledge, and local interactions by immediately adjacent agents. At each
time step, each agent cleans its current location (assuming it is not a critical
point), and moves according to a local movement rule, creating the effect
of a clockwise "sweeping" traversal along the boundary of the contaminated
region. As a result, the agents "peel" layers from the region, while preserving
its connectivity, until the region is cleaned entirely. An illustration of two
agents working according to the protocol can be seen in Figure 3.
s
✻
✻
❅❅ ❅❅
❅❅❅❅❅❅❅❅
Figure 3: An example of two agents using the SWEEP protocol, at time step 40 (with
contamination spreading speed d > 40). All the tiles presented were contaminated at time
0. The black dot denotes the starting point of the agents. The X's mark the critical points
which are not cleaned. The black tiles are the tiles cleaned by the first agent. The second
layer of marked tiles represent the tiles cleaned by the second agent.
12
In order to formally describe the SWEEP algorithm, we should first
define several additional terms. Let τ (t) =(cid:0)τ1(t), τ2(t), . . . , τk(t)(cid:1) denote the
locations of the k agents at time t. In addition, let τi(t) denote the "previous
location" of agent i. Namely, the last tile that agent i had been at, which is
different than τi(t). This is formally defined as :
τi(t) , τi(x) s.t. x = max{j ∈ N j < t and τi(j) 6= τi(t)}
The term ∂F denotes the boundary of F , defined via :
∂F = {v v ∈ F ∧ 8 − N eighbors(v) ∩ (G \ F ) 6= ∅}
The term 'rightmost' can now be defined as follows :
• If t = 0 then select the tile as instructed in Figure 4.
• If τi(t) ∈ ∂Ft then starting from τi(t) (namely, the previous boundary
tile that the agent had been in) scan the four neighbors of τi(t) in a
clockwise order until a boundary tile (excluding τi(t)) is found.
• If not τi(t) ∈ ∂Ft then starting from τi(t) scan the four neighbors of
τi(t) in a clockwise order until the second boundary tile is found.
A schematic flowchart of the protocol, describing its major components
and procedures is presented in Figure 5. The complete pseudo-code of the
protocol and its sub-routines appears in Figures 6 and 7. Upon initialization
of the system, the System Initialization procedure is called (defined in Fig-
ure 6). This procedure sets various initial values of the agents, and calls the
protocol's main procedure -- SWEEP (defined in Figure 7). This procedure
in turn, uses various sub-routines and functions, all defined in Figure 6. The
13
s ❝
s ❝
s ❝
❝
s
❝
s
❝
s
❝
s
❝
s
❝
s
s
❝
❝
s
❝
s
❝
s
❝
s
s❝
❝
s
❝
s
❝
s
❝
s
❝
s
❝
s
❝
s
❝
s
s
❝
❝
s
❝
s
s ❝
s ❝
s ❝
❝
s
❝
s
s ❝
Figure 4: When t = 0 the first movement of an agent located in (x, y) should be decided
according to initial contamination status of the neighbors of (x, y), as appears in these
charts -- the agent's initial location is marked with a filled circle while the destination is
marked with an empty one. All configurations which do not appear in these charts can be
obtained by using rotations. This definition is needed in order to initialize the traversal
behavior of the agents in the correct direction.
SWEEP procedure is comprised of a loop which is executed continuously,
until detecting one of two possible break conditions. The first, implemented
in the Check Completion of Mission procedure, is in charge of detecting
cases where all the contaminated tiles have been cleaned. The second con-
dition, implemented in the Check Near Completion of Mission procedure, is
in charge of detecting scenarios in which every contaminated tile contains
at least a single agent. In this case, the next operation would be a simul-
taneous cleaning of the entire contaminated tiles. Until these conditions are
met, each agent goes through the following sequence of commands. First
each agent calculates its desired destination at the current turn. Then, each
agent calculated whether it should give a priority to another agent located
at the same tile, and wishes to move to the same destination. When two or
14
more agents are located at the same tile, and wish to move towards the same
direction, the agent who had entered the tile first gets to leave the tile, while
the other agents wait. In case several agents had entered the tile at the same
time, the priority is determined using the Priority function. Before actually
moving, each agent who had obtained a permission to move, must now locally
synchronize its movement with its neighbors, in order to avoid simultaneous
movements which may damage the connectivity of the region. This is done
using the waiting dependencies mechanism, which is implemented by each
agent via an internal positioning of itself in a local ordering of his neighbor-
ing agents. When an agent is not delayed by any other agent, it executes its
desired movement. It is important to notice that at any given time, waiting
or resting agents may become active again, if the conditions which made
them become inactive in the first place, had changed.
Following are several results that we will later use. While using these
results, we note that completely cleaning a region is at least as strong as cov-
ering it, as the number of "uncovered" tiles at any given time can be modeled
by the number of "contaminated" tiles, since the number of uncovered tiles
in the original region to be explored is clearly upper bounded at all times by
the number of remaining contaminated tiles that belong to this region.
Result 1. (Cleaning a Non-Expanding Contamination) The time it
takes for a group of K robots using the SWEEP algorithm to clean a region
F of the grid is at most:
tstatic , 8(∂F0 − 1) · (W (F0) + k)
k
+ 2k
Here W (F ) denotes the depth of the region F (the shortest path from
some internal point in F to its boundary, for the internal point whose shortest
15
Figure 5: A schematic flow chart of the SWEEP protocol. The smooth lines represent
the basic flow of the protocol while the dashed lines represent cases in which the flow is
interrupted. Such interruptions occur when an agent calculates that it must not move
until other agents do so (either as a result of waiting or resting dependencies -- see lines 6
and 14 of SWEEP for more details).
16
Arbitrarily choose a pivot point p0 in ∂F0, and mark it as critical point
Place all the agents on p0
For (i = 1; i ≤ k; i + +) do
Call Agent Reset for agent i
1: System Initialization
2:
3:
4:
5:
6:
7:
8:
9: End procedure
End for
Call SWEEP for agent i
Wait two time steps
10: Agent Reset
11:
resting ← false
12:
dest ← null /* destination */
13:
near completion ← false
14:
saturated perimeter ← false
15:
waiting ← ∅
16: End procedure
17: Priority
18:
19:
20: End procedure
/* Assuming the agent moved from (x0, y0) to (x1, y1) */
priority ← 2(x1 − x0) + (y1 − y0)
If ((x, y) = p0) and (x, y) has no contaminated neighbors then
21: Check Completion of Mission
22:
23:
24:
25:
26: End procedure
Clean (x, y)
STOP
If (x, y) is contaminated then
/* Cases where every tile in Ft contains at least a single agent */
near completion ← false
If each of the contaminated neighbors of (x,y) contains at least one agent then
If each of the contaminated neighbors of (x,y) satisfies near completion then
27: Check "Near Completion" of Mission
28:
29:
30:
31:
32:
33:
34:
35:
36:
near completion ← true
Clean (x, y) and STOP
/* Cases where every non-critical tile in ∂Ft contains at least 2 agents */
saturated perimeter ← false
If ((x, y) ∈ ∂Ft) and both (x, y) and all of its non-critical neighbors
in ∂Ft contain at least two agents then
17
37:
38:
saturated perimeter ← true
If ((x, y) ∈ ∂Ft) and both (x, y) and all of its neighbors in ∂Ft has
saturated perimeter = true then
39:
40: End procedure
Ignore resting commands for this time step
Check Completion of Mission
Check "Near Completion" of Mission
dest ← rightmost neighbor of (x,y) /* Calculate destination */
destination signal bits ← dest /* Signaling the desired destination */
/* Calculate resting dependencies (solves agents' clustering problem) */
1: SWEEP Protocol /* Controls agent i after Agent Reset */
2:
3:
4:
5:
6:
7:
8:
9:
K2 : Agents signaling towards dest which entered (x,y) with agent i,
From all agents in (x,y) except agent i we define be the following groups:
K1 : Agents signaling towards dest which entered (x,y) before agent i
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
and with higher priority than this of agent i
resting ← false
If (K1 6= ∅) or (K2 6= ∅) then
resting ← true
If (current time-step T did not end yet) then jump to 4 Else jump to 30
waiting ← ∅ /* Waiting dependencies (agents synchronization) */
Let active agent denote a non-resting agent which didn't move in T yet
If (x-1,y) ∈ Ft contains an active agent then waiting ← waiting ∪ {left}
If (x,y-1) ∈ Ft contains an active agent then waiting ← waiting ∪ {down}
If (x-1,y-1) ∈ Ft contains an active agent then waiting ← waiting ∪ {l-d}
If (x+1,y-1) ∈ Ft contains an active agent then waiting ← waiting ∪ {r-d}
If dest = right and (x+1,y) contains an active agent j, and dest j 6= left, and
there are no other agents delayed by agent i (i.e. (x-1,y) does not contain
active agent l with dest l =right and no active agents in (x,y+1),(x+1,y+1),
(x-1,y+1), and (x+1,y) does not contain active agent n with dest n = left),
then (waiting ← waiting ∪ {right}) and (cid:0)waitingj ← waitingj \ {left}(cid:1)
If dest = up and (x,y+1) contains an active agent j, and dest j 6= down, and
there are no other agents delayed by agent i (i.e. (x,y-1) does not contain
active agent l with dest l =up and no active agents in (x+1,y),(x+1,y+1),
(x-1,y+1), and (x,y+1) does not contain active agent n with dest n = down),
then (waiting ← waiting ∪ {up}) and (cid:0)waitingj ← waitingj \ {down}(cid:1)
If (waiting 6= ∅) then
If (T has not ended yet) then jump to 4 Else jump to 30
/* Decide whether or not (x,y) should be cleaned */
If ¬ ((x,y) ∈ ∂Ft) or ((x,y) ≡ p0) or (x,y) has 2 contaminated tiles in its
4N eighbors which are not connected via a path of contaminated tiles from its
8N eighbors then
(x,y) is an internal point or a critical point and should not be cleaned
Clean (x,y) if and only if it does not still contain other agents
Else
26:
27:
28:
29: Move to dest
30: Wait until T ends.
31:
Return to 2
18
Figure 7: The SWEEP cleaning protocol.
path is the longest) and as defined above, ∂F denotes the boundary of F ,
defined via :
∂F = {v v ∈ F ∧ 8 − N eighbors(v) ∩ (G \ F ) 6= ∅}
The term 8 − N eighbors(v) is used to denote the eight tiles that tile v is
immediately surrounded by.
Result 2. (Universal Lower Bound on Contaminated Area) Using
any cleaning algorithm, the area at time t of a contaminated region that
expands every d time steps can be recursively lower bounded, as follows :
St+d ≥ St − d · k +j2p2 · (St − d · k) − 1k
Here St denotes the area of the contaminated region at time t (such that
S0 = n).
Result 3. (Upper Bound on Cleaning Time for SWEEP on Ex-
panding Domains) For a group of k robot using the SWEEP algorithm
to clean a region F on the grid, that expands every d time steps, the time it
takes the robots to clean F is at most d multiplied by the minimal positive
value of the following two numbers :
(A4 − A1A3) ±p(A1A3 − A4)2 − 4A3(A2 − A1 − A1A4)
2A3
where :
, A3 =
,
, A2 =
c0 + 2 + γ2
8 · γ2
d · k
, γ1 = ψ (1 + A2) − ψ (1 + A1)
c0 + 2 − γ2
γ2 · γ
d
γ2 =p(c0 + 2)2 − 8S0 + 8
d − 2k
∂F0 − 1
8(k + W (F0))
4
−
A1 =
4
A4 = γ1 −
γ =
k
19
Here c0 is the circumference of the initial region F0, and where ψ(x) is
the Digamma function (studied in (70)) -- the logarithmic derivative of the
Gamma function, defined as :
ψ(x) =
d
dx
ln Γ(x) =
Γ′(x)
Γ(x)
or as :
ψ(x) =Z ∞
0 (cid:18) e−t
t −
e−xt
1 − e−t(cid:19) dt
Note that although c0 = O(∂F) the actual length of the perimeter of the
region can be greater than its cardinality, as several tiles may be traversed
more than once. In fact, in (49) it is shown that c0 ≤ 2 · ∂F − 2.
4. Grid Coverage -- Analysis
We note again that when discussing the coverage of regions on the grid,
either static or expanding, it is enough to show that the region can be cleaned
by the team of robots, as clearly the cleaned sites are always a subset of the
visited ones. We first present the cover time of a group of robots operating
in non-expanding domains, using the SWEEP algorithm.
Theorem 1. Given a connected region of S0 = n grid tiles and perimeter c0,
that should be covered by a team of k ant-like robots, the robots can cover it
k S1.5
using O(cid:0) 1
Proof. Since ∂F0 = Θ(c0), and W (F0) = O(√S0), recalling Result 1 we can
0 + S0(cid:1) time.
see that :
tk(n) = tstatic(k) = O(cid:18) 1
kpS0 · c0 + c0 + k(cid:19)
20
As c0 = O(S0) and as for practical reasons we assume that k < n this
would equal in the worse case to :
tk(n) = tstatic(k) = O(cid:18) 1
k
S1.5
0 + S0(cid:19)
We now examine the problem of covering expanding domains. The lower
bound for the number of robots required for completing is as follows.
Theorem 2. Given a region of size S0 ≥ 3 tiles, expanding every d time
robots cannot clean the region, regardless
steps, then a team of less than
√S0
d
of the algorithm used.
Proof. Recalling Result 2 we can see that :
St+d − St ≥j2p2 · (St − d · k) − 1k − d · k
√S0
d we can see that :
By assigning k =
∆St = St+d − St ≥(cid:22)2q2 · (St −pS0) − 1(cid:23) −pS0
For any S0 ≥ 3, we see that ∆S0 > 0. In addition, for every S0 ≥ 3 we
dt > 0 for every t ≥ 0. Therefore, for every S0 ≥ 3 the size of
can see that dSt
the region will be forever growing.
Corollary 1. Given a region of size S0 tiles, expanding every d time steps,
where d = O(1) w.r.t S0, then a team of less than Ω(√S0) robots cannot
clean the region, regardless of the algorithm used.
21
Theorem 3. Given a region F of size S0 tiles, expanding every d time steps,
where R(F ) is the perimeter of the bounding rectangle of the region F , then
a team of k robots that at t = 0 are located at the same tile cannot clean the
region, regardless of the algorithm used, as long as d2k < Ω(R(F )).
Proof. For every v ∈ F let l(v) denote the maximal distance between v and
any of the tiles of F , namely :
l(v) = max{d(v, u)u ∈ F}
Let C(F ) = l(vc) such that vc ∈ F is the tile with minimal value of l(v).
Let vs denote the tile the agents are located in at t = 0. Let vd ∈ F
denote some contaminated tile such that d(vs, vd) = l(vs). Regardless of the
algorithm used by the agents, until some agent reaches vd there will pass at
least l(vs) time steps. Let us assume w.l.o.g that vd is located to the right
(or of the same horizontal coordinate) and to the top (or of the same vertical
coordinate) of vs. Then by the time some agent is able to reach vd there
exists an upper-right quarter of a digital sphere of radius j l(vs)
d k + 1, whose
center is vd.
The number of tiles in such a quarter of digital sphere equals :
1
d (cid:23)2
2(cid:22) l(vs)
+
3
2(cid:22) l(vs)
d (cid:23) + 1 = Θ(cid:18) l(vs)2
d2 (cid:19)
It is obvious that the region cannot be cleaned until vd is cleaned. Let
td denote the time at which the first agent reaches vd. It is easy to see that
td ≥ l(vs). Therefore, regardless of activities of the agents until time step
td, there are now k agents that has to clean a region of at least Θ(cid:16) l(vs)2
d2 (cid:17)
tiles, spreading every d time steps. Using Theorem 2 we know that k agents
22
cannot clean an expanding region of k =
√S0
d
tiles. Namely, at time td k
agents could not clean the contaminated tiles if :
d2k < Ω (l(vs))
As l(vs) ≥ C(F ) we know that k agents could not clean an expanding
contaminated region where : d2k < Ω (C(F )). It is easy to see that for every
region F , if R(F ) is the length of the perimeter of the bounding rectangle of
F then C(F ) = Θ(R(F )).
Lemma 1. For every connected region of size S0 ≥ 3 and perimeter of length
c0 :
1
2
c0 < γ2 < c0
S0 is achieved when the region is arranged in the form of an 8-connected
Proof. let us assume by contradiction that (c0 + 2)2 ≤ (8S0 + 8). This means
c0 ≤ √8S0 + 8 − 2. However, the minimal circumference of a region of size
digital sphere, in which case c0 ≥ 4√S0 − 4, contradicting the assumption
that c0 ≤ √8S0 + 8 − 2 for every S0 > 5. Therefore, γ2 ∈ R.
2 c0. Therefore :
Let us assume by contradiction that γ2 < 1
(c0 + 2)2 − 8S0 + 8 <
1
4
c2
0
which implies :
c0 < −
16
6
+r10
2
3
S0 − 8
8
9
< 3.266pS0 − 2
However, we know that c0 ≥ 4√S0 − 4, which contradicts the assumption
that γ2 < 1
2 c0 for every S0 ≥ 3.
23
Let us assume by contradiction that γ2 > c0. Therefore :
which implies :
(c0 + 2)2 − 8S0 + 8 > c2
0
c0 > 4S0 − 6
However, we know that c0 ≤ 2S0 − 2 (as c0 is maximized when the tiles are
arranged in the form of a straight line), contradicting the assumption that
γ2 > c0 for every S0 ≥ 3.
Lemma 2. For every connected region of size S0 ≥ 3 and perimeter of length
c0 :
Ω(1) < γ1 < O(ln n)
Proof. Let us observe γ1 :
4
c0 + 2 + γ2
γ1 , ψ(cid:18)1 +
(cid:19) − ψ(cid:18)1 +
From Lemma 1 we can see that 1 < (cid:0)1 + c0+2−γ2
−γ where γ is the Euler-Mascheroni constant, defined as :
k! − log(n)# =Z ∞
1
(cid:1) < 1
1 (cid:18) 1
1
n→∞" n
Xk=1
γ = lim
⌊x⌋ −
x(cid:19) dx
c0 + 2 − γ2
4
(cid:19)
4
4 c0. Note that ψ(1) =
which equals approximately 0.57721. In addition, ψ(x) is monotonically in-
creasing for every x > 0. As we also know that ψ(x) is upper bounded by
O(ln x) for large values of x, we see that :
−0.58 < ψ(cid:18)1 +
c0 + 2 − γ2
4
(cid:19) < O(ln n)
(4.1)
24
From Lemma 1 we also see that 1 <(cid:0)1 + c0+2+γ2
(cid:19) = Θ(ln n)
ψ(cid:18)1 +
c0 + 2 + γ2
4
4
(cid:1) < 1.5
Combining equations 4.1 and 4.2 we see that :
Ω(1) < γ1 < O(ln n)
4 c0 meaning that :
(4.2)
(4.3)
Theorem 4. Result 3 returns a positive real number for the covering time
), and where γ1 shifts from O(1) to O(ln S0)
of a region of S0 tiles that expands every d time steps, when the number of
robots is Θ(√S0) and d = Ω( c0
as c0 grows from O(√S0) to O(S0), defined as :
(cid:19) − ψ(cid:18)1 +
γ2 =p(c0 + 2)2 − 8S0 + 8
γ1 = ψ(cid:18)1 +
Proof. Following are the requirements that must hold in order for Result 3
c0 + 2 − γ2
c0 + 2 + γ2
(cid:19)
4
γ1
4
to yield a real number :
• d · k 6= 0
• ∂F > 1
• A3 6= 0
• (c0 + 2)2 > 8S0 − 8
• (A1A3 − A4)2 ≥ 4A3(A2 − A1 − A1A4)
25
The first and second requirements hold for every non trivial scenario. The
third requirement is implied by the fourth. The fourth assumption is a direct
result of Lemma 1.
As for the last requirement, we ask that :
A2
1A2
3 + A2
4 ≥ 4A2A3 − 4A1A3 − 2A1A3A4
which subsequently means that we must have :
2
γ 2
d2k2 (c2
+γ2
0 + γ2
2 γ 2
2 − c0γ2)
d − γγ1γ2
d
1 + γ 2
γ2
dk
≥ 4
c0
γ2γ
4γ2 − c0γ1+
d − 2γ1 + 2 γ2γ
d +
γ1γ2 − γ 2
2 γ
d
Using Lemma 1 and Lemma 2 we should make sure that :
γ2
2
dk2
c2
0 + γ2
1 d + γ2
2 γ2 − γγ1γ2 ≥ O(cid:18) c0γ2γ1
k
+
c0γ2
2 γ
dk (cid:19)
Using W (F ) = O(√S0) and Ω(√S0) = ∂F = O(S0) and dividing by γ2
2
(which we know to be larger than zero), we can write the above as follows :
k ln S0
c2
0
dk2 +
k2 + d ln2 S0
c2
0
ln S0
+ 1 ≥
c0√S0
dk2 +
c2
0
dk2 . In addition, 1 ≥ 1
+
+
k
c0
dk
+
1
d(cid:19)
dk2 ≥ c0√S0
0
d and also 1 ≥ ln S0
(as Eq. 4.6 shows that k ≥ ln S0). In order to have also 1 ≥ c0
c0
and
dk we
O(cid:18)ln S0
As c0 ≥ √S0 then c2
c0
+
1 ≥ ln S0
must have :
k
d · k = Ω(c0)
26
(4.4)
In addition, we should also require that the result µ would be positive (as
it denotes the coverage time). Namely, that :
A4 +p(A1A3 − A4)2 − 4A3(A2 − A1 − A1A4) > A1A3
For this to hold we shall merely require that:
A2 − A1 − A1A4 ≤ 0
(as A3 is known to be positive). Assigning the values of A1, A2, A4, this
translates to :
c0 + c2
0
γ
d ≤ O(c0γ1)
Dividing by c0 we can write :
1 +
1 + c0
√S0
k − d
c0
d
+ k
c0
≤ O(γ1)
Namely :
c0√S0
k
c0 +
+ k ≤ dO(γ1)
As c0 is the dominant element of the left side of the inequation, we see
that :
Assigning this lower bound for d we can now see that :
d = Ω(cid:18) c0
γ1(cid:19)
Ω(pS0) ≤ k ≤ O(c0)
Therefore, we shall select the value of k such that :
(4.5)
(4.6)
This also satisfies Equation 4.4.
k = Θ(pS0)
27
Theorem 5. The time it takes a group of k = Θ(√S0) robots using the
SWEEP algorithm to cover a connected region of size S0 tiles, that expands
every d = Ω( c0
γ1
) time steps, is upper bounded as follows :
where γ1 shifts from O(1) to O(ln S0) as c0 grows from O(√S0) to O(S0),
tSU CCESS = O(cid:0)S2
0 ln S0(cid:1)
defined as :
γ1 = ψ(cid:18)1 +
4
c0 + 2 + γ2
(cid:19) − ψ(cid:18)1 +
γ2 =p(c0 + 2)2 − 8S0 + 8
Proof. Recalling Result 1 we know that if :
c0 + 2 − γ2
4
(cid:19)
8(∂F0 − 1) · (W (F0) + k)
k
+ 2k < d
then the robots could clean the region before it expands even once. In this
case, the cleaning time would be O( 1
k√S0·c0+c0) as was shown in Theorem 1.
Therefore, we shall assume that :
8(∂F0 − 1) · (W (F0) + k)
k
+ 2k ≥ d
(4.7)
Observing Result 3 we see that :
tSU CCESS =
A3
d · O A1 + A4
d · O A1 + A4
A3
A3
+sA2
1 + A1A4 + A1 + A2
√A3pA1A4 + A1 + A2
A3
+
+
A2
4
A2
3! ≤
A3 ! ≤
+ A4
28
A3
d · O A1 + A4
d · O
√k
√A3
+ pA1A4 + √A1 + √A2
! =
√c0 + γ2
+q kd
+ kγ+
c0 + γ2 + dk γ1
γ2
√γ2
γ2
√c0+γ2√dγ1+γ2·γ
Using the fact that γ2 = Θ(c0) (Lemma 1) we can rewrite this expression
as :
d · O(cid:18)c0 + dk
γ1
c0
+ kγ + √kpdγ1 + c0γ + √kd(cid:19)
Recalling Equation 4.7, and as W (F0) = O(√S0), we can see that :
(4.8)
d = O(cid:18)√S0 · c0
γ = O(cid:18)√S0
Therefore, γ can now be written as :
+pS0 +
k
k
+ c0 + k(cid:19)
k
√S0
+ 1(cid:19)
In addition, remembering that O(√S0) ≤ c0 ≤ O(S0) we can rewrite the
expression of Equation 4.8 as follows :
Using Lemma 2 we see that :
c0
+
γ1 +
+
√k
c0 + dk γ1
c0
+ k√S0 + k2
c0 + dk γ1
√S0
c0
k + √S0 + k√S0
√S0
+ k√S0 + k2
√S0
k + 4√S0 +
d · O
+ √kd
=
√kc0q d
d · O
4√S0(cid:17) + √kd
√kc0(cid:16)√γ1 + 4√S0p γ1
pc0 ln S0√S0 +pc0k√S0 + k 4√S0 + √kd
d · O
=
d · O
c0 + k√S0 ln S0 + k2
√S0
pc0 ln S0√S0 +pc0k√S0
+ k√S0 + k2
√S0
c0 + dk ln S0
c0
+
+
29
(4.9)
Assuming that k > O(ln S0) and as c0 = O(S0) we can now write :
0
S 2.5
k + S2
k3
√S0
O
Using Equation 4.6 we can see that this translates to :
0 ln S0 + S1.5
+ S 2.25
0√k
0 k ln S0 + k2√S0 ln S0+
0 √k + S0.75
+ S1.75
0 k1.5
O(cid:0)S1.5
0 k ln S0(cid:1)
(4.10)
(4.11)
5. Conclusions
In this paper we have discussed the problem of cleaning or covering a
connected region on the grid using a collaborate team of simple, finite-state-
automata robotic agents. We have shown that when the regions are static,
this can be done in O( 1
k n1.5 + n) time
in the worst case, thus improving the previous results for this problem. In
k√n· c0 + c0 + k) time which equals O( 1
addition, we have shown that when the region is expanding in a constant rate
(which is "slow enough"), a team of Θ(√n) robots can still be guaranteed to
clean or cover it, in O(n2 ln n) time.
In addition, we have shown that teams of less than Ω(√n) robots can
never cover a region that expands every O(1) time steps, regardless of their
sensing capabilities, communications and memory resources employed, or
the algorithm used. As to regions that expand slower than every O(1) time
steps, we have shown the following impossibility results. First, a region of
n tiles that expands every d time steps cannot be covered by a group of k
agents if dk ≤ O(√n). Using Theorem 4 we can guarantee a coverage when
30
dk = Ω( n1.5
ln n ), or even for dk = Ω(n)) (when the region's perimeter c0 equals
O(n)).
Second, a spreading region cannot be covered when d2k is smaller than
the order of the perimeter of the bounding rectangle of the region (which
equals O(n) in the worse case and O(√n) for shapes with small perimeters).
Using Theorem 4 we can guarantee a coverage when d2k = Ω( n2.5
ln2 n ), or for
d2k = Ω(n1.5) (when the region's perimeter c0 equals O(n)).
We believe that these results can be easily applied to various other prob-
lems in which a team of agents are required to operate in an expanding grid
domain. For example, this result can show that a team of Θ(√n) cops can
always catch a robber (or for that matter -- a group of robbers), moving
(slowly) in an unbounded grid. Alternatively, robbers can be guaranteed to
escape a team of less than O(√n) cops, if the area they are located in is
unbounded.
6. Vitae
References
[1] S. Mastellone, D. Stipanovi, C. Graunke, K. Intlekofer, M. Spong, For-
mation control and collision avoidance for multi-agent non-holonomic
systems: Theory and experiments, The International Journal of
Robotics Research 27 (1) (2008) 107 -- 126.
[2] S. DeLoach, M. Kumar, Intelligence Integration in Distributed Knowl-
edge Management, Idea Group Inc (IGI), 2008, Ch. Multi-Agent Sys-
tems Engineering: An Overview and Case Study, pp. 207 -- 224.
31
Dr. Altshuler received his PhD in Computer Science
from the Technion -- Israel Institute of Technology
(2009), and is now a post-doc researcher at the Deutsche
Telekom Research Lab in Ben Gurion University.
Prof. Bruckstein is a professor at the Computer Science
department at the Technion -- Israel Institute of Tech-
nology, received his PhD in Electrical Engineering from
Stanford University (1984).
32
[3] I. Wagner, A. Bruckstein, From ants to a(ge)nts: A special issue on ant --
robotics, Annals of Mathematics and Artificial Intelligence, Special Issue
on Ant Robotics 31 (1 -- 4) (2001) 1 -- 6.
[4] U. V. et. al, RoboCup 2007: Robot Soccer World Cup XI, Vol. 5001 of
Lecture Notes in Computer Science, Springer Berlin, 2008.
[5] R. Alami, S. Fleury, M. Herrb, F. Ingrand, F. Robert, Multi-robot coop-
eration in the martha project, IEEE Robotics and Automation Magazine
5 (1) (1998) 36 -- 47.
[6] N. Agmon, S. Kraus, G. Kaminka, Multi-robot perimeter patrol in ad-
versarial settings, in: IEEE International Conference on Robotics and
Automation (ICRA 2008), 2008, pp. 2339 -- 2345.
[7] N. Agmon, V. Sadov, G. Kaminka, S. Kraus, The impact of adver-
sarial knowledge on adversarial planning in perimeter patrol, in: AA-
MAS '08: Proceedings of the 7th international joint conference on Au-
tonomous agents and multiagent systems, International Foundation for
Autonomous Agents and Multiagent Systems, Richland, SC, 2008, pp.
55 -- 62.
[8] S. Kraus, O. Shehory, G. Taase, Coalition formation with uncertain
heterogeneous information, in: Proceedings of the second international
joint conference on Autonomous agents and multiagent systems, 2003,
pp. 1 -- 8.
[9] H. Work, E. Chown, T. Hermans, J. Butterfield, Robust team-play in
highly uncertain environments, in: AAMAS '08: Proceedings of the 7th
33
international joint conference on Autonomous agents and multiagent
systems, International Foundation for Autonomous Agents and Multia-
gent Systems, Richland, SC, 2008, pp. 1199 -- 1202.
[10] M. Rehak, M. Pechoucek, P. Celeda, V. Krmicek, M. Grill, K. Bar-
tos, Multi-agent approach to network intrusion detection,
in: AA-
MAS '08: Proceedings of the 7th international joint conference on Au-
tonomous agents and multiagent systems, International Foundation for
Autonomous Agents and Multiagent Systems, Richland, SC, 2008, pp.
1695 -- 1696.
[11] R. Connaughton, P. Schermerhorn, M. Scheutz, Physical parameter
optimization in swarms of ultra-low complexity agents, in: AAMAS
'08: Proceedings of the 7th international
joint conference on Au-
tonomous agents and multiagent systems, International Foundation for
Autonomous Agents and Multiagent Systems, Richland, SC, 2008, pp.
1631 -- 1634.
[12] M. Mataric, Interaction and intelligent behavior, Ph.D. thesis, Mas-
sachusetts Institute of Technology (1994).
[13] E. Manisterski, R. Lin, S. Kraus, Understanding how people design trad-
ing agents over time, in: AAMAS '08: Proceedings of the 7th interna-
tional joint conference on Autonomous agents and multiagent systems,
International Foundation for Autonomous Agents and Multiagent Sys-
tems, Richland, SC, 2008, pp. 1593 -- 1596.
34
[14] D. MacKenzie, R. Arkin, J. Cameron, Multiagent mission specification
and execution, Autonomous Robots 4 (1) (1997) 29 -- 52.
[15] G. Chalkiadakis, C. Boutilier, Sequential decision making in repeated
coalition formation under uncertainty, in: Proceedings of the 7th in-
ternational joint conference on Autonomous agents and multiagent sys-
tems, 2008, pp. 347 -- 354.
[16] X. Zheng, S. Koenig, Reaction functions for task allocation to coopera-
tive agents, in: Proceedings of the 7th international joint conference on
Autonomous agents and multiagent systems, 2008, pp. 559 -- 566.
[17] C. Candea, H. Hu, L. Iocchi, D. Nardi, M. Piaggio, Coordinating in
multi-agent robocup teams, Robotics and Autonomous Systems 36 (2 --
3) (2001) 67 -- 86.
[18] R. Sawhney, K. Krishna, K. Srinathan, M. Mohan, On reduced time
fault tolerant paths for multiple uavs covering a hostile terrain, in: AA-
MAS '08: Proceedings of the 7th international joint conference on Au-
tonomous agents and multiagent systems, International Foundation for
Autonomous Agents and Multiagent Systems, Richland, SC, 2008, pp.
1171 -- 1174.
[19] A. Yamashita, M. Fukuchi, J. Ota, T. Arai, H. Asama, Motion plan-
ning for cooperative transportation of a large object by multiple mobile
robots in a 3d environment, in: In Proceedings of IEEE International
Conference on Robotics and Automation, 2000, pp. 3144 -- 3151.
35
[20] A. Agogino, K. Tumer, Regulating air traffic flow with coupled agents,
in: AAMAS '08: Proceedings of the 7th international joint conference on
Autonomous agents and multiagent systems, International Foundation
for Autonomous Agents and Multiagent Systems, Richland, SC, 2008,
pp. 535 -- 542.
[21] S. Premvuti, S. Yuta, Consideration on the cooperation of multiple au-
tonomous mobile robots,
in: Proceedings of the IEEE International
Workshop of Intelligent Robots and Systems, 1990, pp. 59 -- 63.
[22] R. Bhatt, C. Tang, V. Krovi, Formation optimization for a fleet of
wheeled mobile robots a geometric approach, Robotics and Autonomous
Systems 57 (1) (2009) 102 -- 120.
[23] J. Fredslund, M. Mataric, Robot formations using only local sensing
and control, in: Proceedings of the International Symposium on Com-
putational Intelligence in Robotics and Automation (IEEE CIRA 2001),
2001, pp. 308 -- 313.
[24] N. Gordon, I. Wagner, A. Bruckstein, Discrete bee dance algorithms
for pattern formation on a grid, in: IEEE International Conference on
Intelligent Agent Technology (IAT03), 2003, pp. 545 -- 549.
[25] K. Bendjilali, F. Belkhouche, B. Belkhouche, Robot formation modelling
and control based on the relative kinematics equations, The Interna-
tional Journal of Robotics and Automation 24 (1) (2009) 3220 -- .
[26] T. Balch, R. Arkin, Behavior-based formation control for multi-robot
36
teams, IEEE Transactions on Robotics and Automation 14 (6) (1998)
926 -- 939.
[27] S. Sariel, T. Balch, Real time auction based allocation of tasks for multi-
robot exploration problem in dynamic environments, in: In Proceed-
ings of the AAAI-05 Workshop on Integrating Planning into Scheduling,
2005, pp. 27 -- 33.
[28] M. Pfingsthorn, B. Slamet, A. Visser, RoboCup 2007: Robot Soccer
World Cup XI, Vol. 5001 of Lecture Notes in Computer Science, Springer
Berlin, 2008, Ch. A Scalable Hybrid Multi-robot SLAM Method for
Highly Detailed Maps, pp. 457 -- 464.
[29] I. Rekleitis, G. Dudek, E. Milios, Experiments in free-space triangu-
lation using cooperative localization, in: IEEE/RSJ/GI International
Conference on Intelligent Robots and Systems (IROS), 2003.
[30] I. Harmatia, K. Skrzypczykb, Robot team coordination for target track-
ing using fuzzy logic controller in game theoretic framework, Robotics
and Autonomous Systems 57 (1) (2009) 75 -- 86.
[31] L. Parker, C. Touzet, Multi-robot learning in a cooperative observation
task, Distributed Autonomous Robotic Systems 4 (2000) 391 -- 401.
[32] G. Hollinger, S. Singh, J. Djugash, A. Kehagias, Efficient multi-robot
search for a moving target, The International Journal of Robotics Re-
search 28 (2) (2009) 201 -- 219.
[33] S. LaValle, D. Lin, L. Guibas, J. Latombe, R. Motwani, Finding an
unpredictable target in a workspace with obstacles, in: Proceedings of
37
the 1997 IEEE International Conference on Robotics and Automation
(ICRA-97), 1997, pp. 737 -- 742.
[34] A. Efraim, D.Peleg, Distributed algorithms for partitioning a swarm of
autonomous mobile robots, Structural Information and Communication
Complexity, Lecture Notes in Computer Science 4474 (2007) 180 -- 194.
[35] R. Olfati-Saber, Flocking for multi-agent dynamic systems: Algorithms
and theory, IEEE Transactions on Automatic Control 51 (3) (2006) 401 --
420.
[36] E. Bonabeau, M. Dorigo, G. Theraulaz, Swarm Intelligence: From Nat-
ural to Artificial Systems, Oxford University Press, US, 1999.
[37] G. Beni, J. Wang, Theoretical problems for the realization of distributed
robotic systems, in: IEEE Internal Conference on Robotics and Automa-
tion, 1991, pp. 1914 -- 1920.
[38] A. Thorndike, Summary of antisubmarine warfare operations in world
war ii, Summary report, NDRC Summary Report (1946).
[39] P. Morse, G. Kimball, Methods of operations research, MIT Press and
New York: Wiley, 1951.
[40] B. Koopman, Search and Screening: General Principles with Historical
Applications, Pergamon Press, 1980.
[41] P. Vincent, I. Rubin, A framework and analysis for cooperative search
using uav swarms, in: ACM Simposium on applied computing, 2004.
38
[42] Y. Altshuler, V. Yanovsky, A. Bruckstein, I. Wagner, Efficient cooper-
ative search of smart targets using uav swarms, ROBOTICA 26 (2008)
551 -- 557.
[43] S. Koenig, M. Likhachev, X. Sun, Speeding up moving-target search,
in: AAMAS '07: Proceedings of the 6th international joint conference
on Autonomous agents and multiagent systems, ACM, New York, NY,
USA, 2007, pp. 1 -- 8.
[44] R. Borie, C. Tovey, S. Koenig, Algorithms and complexity results for
pursuit-evasion problems, In Proceedings of the International Joint Con-
ference on Artificial Intelligence (IJCAI) (2009) 59 -- 66.
[45] R. Isaacs, Differential Games: A Mathematical Theory with Applica-
tions to Warfare and Pursuit, Control and Optimization, John Wiley
and Sons, Reprinted by Dover Publications 1999, 1965.
[46] V. Isler, S. Kannan, S. Khanna, Randomized pursuit-evasion with local
visibility, SIAM Journal of Discrete Mathematics 20 (2006) 26 -- 41.
[47] A. Goldstein, E. Reingold, The complexity of pursuit on a graph, The-
oretical Computer Science 143 (1995) 93 -- 112.
[48] J. Flynn, Lion and man: the general case, SIAM Journal of Control 12
(1974) 581 -- 597.
[49] I. Wagner, Y. Altshuler, V. Yanovski, A. Bruckstein, Cooperative clean-
ers: A study in ant robotics, The International Journal of Robotics
Research (IJRR) 27 (1) (2008) 127 -- 151.
39
[50] Y. Altshuler, A. Bruckstein, I. Wagner, Swarm robotics for a dynamic
cleaning problem, in: IEEE Swarm Intelligence Symposium, 2005, pp.
209 -- 216.
[51] Y. Altshuler, I. Wagner, A. Bruckstein, Swarm ant robotics for a dy-
namic cleaning problem -- upper bounds, The 4th International confer-
ence on Autonomous Robots and Agents (ICARA-2009) (2009) 227 -- 232.
[52] V. Braitenberg, Vehicles, MIT Press, 1984.
[53] R. Korf, Real-time heuristic search, Artificial Intelligence 42 (1990) 189 --
211.
[54] I. Rekleitis, V. Lee-Shuey, A. P. Newz, H. Choset, Limited communica-
tion, multi-robot team based coverage, in: IEEE International Confer-
ence on Robotics and Automation, 2004.
[55] Z. Butler, A. Rizzi, R. Hollis, Distributed coverage of rectilinear environ-
ments, in: Proceedings of the Workshop on the Algorithmic Foundations
of Robotics, 2001.
[56] E. Acar, Y. Zhang, H. Choset, M. Schervish, A. Costa, R. Melamud,
D. Lean, A. Gravelin, Path planning for robotic demining and devel-
opment of a test platform, in: International Conference on Field and
Service Robotics, 2001, pp. 161 -- 168.
[57] M. Batalin, G. Sukhatme, Spreading out: A local approach to multi-
robot coverage, in: 6th International Symposium on Distributed Au-
tonomous Robotics Systems, 2002.
40
[58] X. Zheng, S. Koenig, Robot coverage of terrain with non-uniform
traversability, in: In Proceedings of the IEEE International Conference
on Intelligent Robots and Systems (IROS), 2007, pp. 3757 -- 3764.
[59] X. Zheng, S. Jain, S. Koenig, D. Kempe, Multi-robot forest coverage,
in: In Proc. IROS, press, 2005, pp. 3852 -- 3857.
[60] J. Svennebring, S. Koenig, Building terrain-covering ant robots: A fea-
sibility study, Autonomous Robots 16 (3) (2004) 313 -- 332.
[61] S. Koenig, Y. Liu, Terrain coverage with ant robots: A simulation study,
in: AGENTS'01, 2001.
[62] S. Koenig, B. Szymanski, Y. Liu, Efficient and inefficient ant coverage
methods, Annals of Mathematics and Artificial Intelligence 31 (2001)
41 -- 76.
[63] B. Szymanski, S. Koenig, The complexity of node counting on undirected
graphs, Technical Report, Computer Science Department, Rensselaer
Polytechnic Institute, Troy, NY.
[64] I. Wagner, M. Lindenbaum, A. Bruckstein, Efficiently searching a graph
by a smell-oriented vertex process, Annals of Mathematics and Artificial
Intelligence 24 (1998) 211 -- 223.
[65] S. B. Thrun., Efficient exploration in reinforcement learning -- techni-
cal report cmu-cs-92-102, Technical report, Carnegie Mellon University
(1992).
41
[66] G. Dudek, M. Jenkin, E. Milios, D. Wilkes, Robotic exploration as graph
construction, IEEE Transactions on Robotics and Automation 7 (1991)
859 -- 865.
[67] A. Broder, A. Karlin, P. Raghavan, E. Upfal, Trading space for time
in undirected s − t connectivity, in: ACM Symposium on Theory of
Computing (STOC), 1989, pp. 543 -- 549.
[68] L. Lovasz, Random walks on graphs: A survey, Combinatorics, Paul
Erdos is Eighty 2 (1996) 353 -- 398.
[69] J. Jonasson, O. Schramm, On the cover time of planar graphs, Electron.
Comm. Probab. 5 (2000) 85 -- 90 (electronic).
[70] M. Abramowitz, I. Stegun, Handbook of Mathematical Functions, Ap-
plied Mathematics Series, National Bureau of Standards, 1964, p. 55.
42
|
1905.09177 | 1 | 1905 | 2019-05-22T14:53:39 | Asynchronous Scattering | [
"cs.MA",
"cs.DC"
] | In this paper, we consider the problem of scattering a swarm of mobile oblivious robots in a continuous space. We consider the fully asynchronous setting where robots may base their computation on past observations, or may be observed by other robots while moving.
It turns out that asynchronous scattering is solvable in the most general case when both vision (the ability to see others robots positions) and weak local multiplicity detection are available. In the case of a bidimensional Euclidean space, ASYNC scattering is also solvable with blind robots if moves are rigid. Our approach is constructive and modular, as we present a proof technique for probabilistic robot protocols that is of independent interest and can be reused for other purposes.
On the negative side, we show that when robots are both blind and have no multiplicity detection, the problem is unsolvable, and when only one of those is available, the problem remains unsolvable on the line. | cs.MA | cs | Asynchronous Scattering
Ulysse L´echine1,2, S´ebastien Tixeuil2
1) ´Ecole Normale Sup´erieure de Lyon, Lyon, France
ulysse.lechine [at] ens-lyon.fr
2) Sorbonne Universit´e, CNRS, LIP6, FR-75005, Paris, France
[email protected]
Abstract
In this paper, we consider the problem of scattering a swarm of mobile oblivious robots in a continuous
space. We consider the fully asynchronous setting where robots may base their computation on past
observations, or may be observed by other robots while moving.
It turns out that asynchronous scattering is solvable in the most general case when both vision (the
ability to see others robots positions) and weak local multiplicity detection are available. In the case
of a bidimensional Euclidean space, ASYNC scattering is also solvable with blind robots if moves are
rigid. Our approach is constructive and modular, as we present a proof technique for probabilistic robot
protocols that is of independent interest and can be reused for other purposes.
On the negative side, we show that when robots are both blind and have no multiplicity detection,
the problem is unsolvable, and when only one of those is available, the problem remains unsolvable on
the line.
9
1
0
2
y
a
M
2
2
]
A
M
.
s
c
[
1
v
7
7
1
9
0
.
5
0
9
1
:
v
i
X
r
a
1
1 Introduction
We consider mobile oblivious robots that evolve in a bidimensional Euclidean space [8, 4]. Those robots are
modeled as dimensionless points and repeatedly execute cycles that consist of Look-Compute-Move phases:
they first obtain a snapshot of their environment using visual sensors (Look), then compute their next tenta-
tive location (Compute), and finally move toward their computed destination (Move). After a Look-Compute-
Move cycle is complete, the local memory of a robot r is erased before r starts a new cycle (the robot is said
oblivious). In the semi-synchronous model (SSYNC in the current terminology [8, 4]), a non-empty subset
of robots is chosen by the scheduler (considered as an adversary) to simultaneously and atomically execute a
Look-Compute-Move cycle. An even more restricted model is the fully synchronous model (FSYNC [8, 4]),
where the subset of scheduled robots is the full set of robots. While SSYNC and FSYNC model make it
easier to write and prove robot protocols, they do not represent well the asynchronous behavior of actual
robotic entities. For that purpose, the asynchronous model (ASYNC [5, 4]) has been introduced:
in that
model, each phase of the Look-Compute-Move cycle of a robot may occur at any time. In particular, the
following events may occur: When a robot executes a Look phase, the other robots are not necessarily still;
also, a robot follows a straight path to reach its destination, but not necessarily at a constant speed (though
it cannot move back).
The visual sensors of a robot r output information about the r's environment in an ego-centered coordinate
system (which we denote in the sequel as frame). It is generally assumed that robots' coordinate systems do
not share a common direction (same "North-South axis"), a common orientation (same "North"), a common
chirality (same handedness), or even a common unit distance. In this ego-centered coordinate system, the
output of the visual sensors consists in (possibly partial) information about the positions of the other robots
(the current robot always being at the origin of the coordinate system). A first criterium to classify vision
sensors is its range [5]: how far can it see other robots positions? In this paper we consider two kinds of
robots: perceptive robots can see every location occupied by at least one robot, while blind robots cannot
see any other location than the origin of the coordinate system. A second criterium for vision sensors is its
ability to detect multiple robots placed at the exact same location [7]: strong multiplicity detection returns
the exact number of robots placed on a given location, weak multiplicity detection returns true if a location
hosts at least two robots, and false if it contains exactly one robot, while no multiplicity detection returns
no information about a position other than whether it contains at least one robot. Multiplicity detection is
local if it can only be used on the observing robot's location, and is global if it can be used on every occupied
location within the range of the vision sensor (and this range is greater than 0).
During the move phase, the move of the robot is either rigid or flexible [6]. If robot moves are considered
rigid, a robot always reaches its computed destination (possibly after some unbounded time in the ASYNC
model). If robot moves are flexible, then the scheduler can interrupt a robot r before it reaches its destination.
Then r starts a new Look-Compute-Move cycle. In the classical model [8], the scheduler may not stop r
before it moves at least some distance δ > 0 (unknown to r) unless r target destination is closer than δ (in
that case, r reaches its destination). In this paper, we instead consider a more extreme case where robot
moves can be infinitesimal : each robot r can be interrupted after it moves some distance > 0, for any such
.
The problem we consider is that of scattering: from any initial situation where robots may occupy the
same locations, any execution of a scattering protocol ends up in a situation where all robots occupy distinct
positions. Termination (that is, reaching a fixed global state where all robots remain still) is not required, as
long as robots remain at distinct locations. A deterministic scattering protocol for oblivious mobile robots is
feasible if and only if there does not exists clone robots [8]: two robots are clones if they start at the same
position, are given the same coordinate system, and are always activated simultaneously by the scheduler.
Two deterministic clone robots never separate. By contrast, a robot with no clone can simply move "North"
by one distance unit when activated by the scheduler to separate from another robot initially located at the
same position. As clone robots cannot be avoided in general, most subsequent work studied probabilistic
protocols. The first probabilistic scattering for the SSYNC model is due to Dieudonn´e and Petit [3]: anytime
a robot is activated, it tosses a coin, if the coin is 0, the robot stays in place, otherwise it moves arbitrarily
within its Voronoı cell. Then, Cl´ement et al. [2] proposed an expected constant time scattering protocol
for the SSYNC model: assuming knowledge of n, the number of robots, anytime a robot is activated, it
chooses a destination randomly among n2 ones, still within its Voronoı cell. Recently, Bramas and Tixeuil [1]
considered the random bit complexity of mobile robot scattering in the SSYNC model: anytime a robot
is activated, it chooses randomly among f destinations (f depends on the observed configuration, so no
knowledge of n is assumed) within its Voronoı cell. Bramas and Tixeuil [1] show that their solution is the
best possible in SSYNC (both in terms of coin tossing and round complexity) when n is not known. Overall,
it turns out that the scattering problem for oblivious mobile robots was only investigated in the SSYNC
2
model. The reason is that for all aforementioned protocols, the core proof argument relies on the fact that
once two robots separate, they never occupy the same position ever again, hence the number of robots on a
given multiplicity point always decreases. In the ASYNC model, this property is no longer guaranteed using
previous approaches. For example, consider the protocol of Dieudonn´e and Petit [3], and two stacks of 2 and
3 robots, respectively. First the stack of two robots is activated: one robot chooses to move to destination
d and the other chooses to stay, but the moving robot, called r1, is put on hold. Then the stack of 3 robots
is activated, 2 of those 3 robots, say r2 and r3 decide to move, and the other one decides to stay. Now r2 is
activated again, but its observation of other positions is updated (and so is its Voronoı cell). So, it may turn
out that r2 chooses d as its destination. Now, both r1 and r2 move to d, creating a new multiplicity point.
So, solving the problem in ASYNC requires a fresh approach, at least with respect to the proof techniques.
We investigate the possibility of asynchronous scattering in various settings. In particular, we consider
both perceptive and blind robots, both weak local multiplicity detection and no multiplicity detection, and
both rigid and flexible moves. It turns out that ASYNC scattering is solvable in the most general case when
both vision and weak local multiplicity detection are available.
In the case of a bidimensional Euclidean
space, ASYNC scattering is also solvable with blind robots if moves are rigid. Our approach is constructive
and modular, as we present a proof technique for probabilistic robot protocols that is of independent interest
and can be reused for other purposes. On the negative side, we show that when robots are both blind and
have no multiplicity detection, the problem is unsolvable, and when only one of those is available, the problem
remains unsolvable on the line. Unlike previous approaches [3, 2, 1], we consider the fully realistic ASYNC
model, and unlike Clement et al. [2], we do not consider that robots are aware of n. Our positive result do
not assume that robots share the same chirality, while impossibility results remain valid if robots share a
common chirality.
The remaining of the paper is structured as follows. Section 2 details the mathematical framework we
work with. Section 3 features a general result, which we believe is of independent interest, that is later used
to prove our results about scattering in Section 4 and 5. Section 6 presents preliminary complexity results,
while Section 7 takes a computability perspective. Finally, Section 8 provides some concluding remarks.
2 Framework
2.1 Notations
R denotes the set of real numbers, Q the set of rational numbers, and N the set of natural numbers. The set of
natural numbers except 0 is denoted by N(cid:63). Let f be a function, we can also denote it as f (•). (cid:107)•(cid:107) : R2 (cid:44)→ R
denotes the standard euclidean norm over R2, and (cid:104)•,•(cid:105) denotes the standard scalar product over R2 × R2.
Let E be a set, then E∗ denotes the set of all finite sequences of elements of E, EN denotes the set of all
sequences over E, and the set of infinite sequences E∞ = EN ∪ E∗. Let e ∈ E1 × E2 let p and q be the two
elements such that e = (p, q), then e(1) = p and e(2) = q, we generalize this notation to the case where e is
composed of more than two elements.
2.2 Execution Model
We consider the oblivious Look-Compute-Move model (a.k.a.
setting.
the OBLOT S model) in an asynchronous
The current state of a robot is modelled as a 6-tuple (l, d, r, s, a, c) ∈ S such that:
1. l ∈ R2 is the current location of the robot,
2. d ∈ R2 is the destination of the robot,
3. r ∈ {0; 1} is 1 if the robot is activated and 0 otherwise,
4. s ∈ R, s > 0 is the scale of the referential of the robot (the size of the unit distance),
5. a ∈ [0; π[ is the angle of the referential of the robot (where is the north),
6. c ∈ {0; 1} is 1 if the robot is right-handed and 0 otherwise.
We define the set of configurations C as the set of combinations of the local states of the robots where
at most one robot is activated in each configuration: C = {(s1, . . . , sn) ∈ S n∃i ∈(cid:74)1; n(cid:75)(si(3) = 1 ⇒ ∀j ∈
(cid:74)1; n(cid:75), j (cid:54)= i ⇒ sj(3) = 0)}. We say that the k-th robot is activated if and only if C(k)(3) = 1, otherwise it's
inactivated. Having a single robot activated at any time does not weaken the ASYNC model as it still allows
to simulate simultaneous executions, yet it permits to simplify notations in the sequel.
3
Let APART be the set of configurations such that there is no location occupied by two or more robots,
and every robot is immobile (its destination and its location are equal): APART = {(s1, . . . , sn) ∈ C --
∀i, j ∈(cid:74)0; 1(cid:75)2, si(1) = si(2) ∧ si(1) (cid:54)= sj(1)}.
Let A be the algorithm all robots execute. We consider two settings for visual sensors:
• If vision is assumed A receives as inputs the set of points where there is at least one robot. Said
differently given a point in R2 robots know if there is at least one robot there but not how many. If
vision is not assumed the robot is said to be blind
• If local weak multiplicity is assumed, A receives as input a bit indicating if there is another robot on
the same position as the robot running A .
An execution is an element of C ∞ = C1 → C2 → C3 . . . such that every transition Ci → Ci+1 is valid.
Now, C → D is a valid transition for A if and only if there exists a unique k ∈ N such that C(k) (cid:54)= D(k)
(i.e. only one robot did something) and all of the following properties are satisfied:
• At least one of the following conditions is true:
-- choice of frame C(k) = (p, p, 0, r, θ, b) and D(k) = (p, p, 1, r(cid:48), θ(cid:48), b(cid:48)): the scheduler assigns a frame
to robot k,
-- computation C(k) = (p, p, 1, r, θ, b) and D(k) = (p, q, 0, r, θ, b): the robot k computes its desti-
nation q using A ,
-- ready to move C(k) = (p, q, 0, r, θ, b) where p (cid:54)= q and D(k) = (p, q, 1, r, θ, b): robot k is about
to move,
-- move C(k) = (p, q, 1, r, θ) where p (cid:54)= q and D(k) = (q(cid:48), q, 0, r, θ) where q(cid:48) ∈]p; q]: robot k moves
towards its destination,
-- move and interrupt C(k) = (p, q, 1, r, θ) where p (cid:54)= q and D(k) = (q(cid:48), q(cid:48), 0, r, θ) where q(cid:48) ∈]p; q]:
robot k moves towards its destination, but may be interrupted prematurely by the scheduler.
• If more than one condition is true, then computation, move, and move and interrupt transitions have
priority over the other two.
Notice that transitions are of two types, the ones that inactivate robot k (computation, move, move and
interrupt), and the ones that activate it (choice of frame, ready to move). Due to the higher priority of the
former set of transitions, the property that at most one robot is activated in any configuration is preserved
throughout the execution. Starting from configuration Ci, if A is deterministic, then there is at most one
reachable configuration. If A is a probabilistic algorithm, then there can be more than one reachable con-
figurations. An execution E is fair if all robots perform the computation transition infinitely many times.
Given an algorithm A , we denote by EA the set of executions conforming to A .
Let us now define schedulers. Intuitively the scheduler chooses when robots are activated, by how much
they move when the movement is not rigid, etc. Let A be an algorithm, let σ : C ∗ (cid:44)→ C be a function. If
∀C1, . . . , Ck, Ck → σ(C1, . . . , Ck) is a valid transition for A , then σ is said to be a scheduler for A . This
ensures that all executions conform to A .
σ (i) = σ(EC
σ (0), . . . , EC
Given an initial configuration C, we define execution EC
σ (i − 1)). If A is probabilistic, then EC
σ as the execution starting from C according
to scheduler σ, and Eσ is the set of all possible executions. If A is deterministic, then EC
σ (0) = C, and
∀i ∈ N(cid:63), EC
σ (i) is a random variable defined
the same way as in the deterministic case, and σ is a probabilistic function, that is to say it returns its output
according to a certain distribution law given by the algorithm A for the transitions of the kind computation.
In the deterministic setting, σ is a fair scheduler if σ is a scheduler and Eσ is fair. In the probabilistic
setting, σ is fair if P(Eσ is fair) = 1. Given Algorithm A , the set of all possible schedulers conforming to
A is noted ΣA . If the algorithm is implied, we simply write Σ. If a scheduler never chooses a move and
interrupt transition, the movements are called rigid.
The fine grained approach we used for transitions permits to simulate an arbitrary number of robots
running simultaneously (even when other are moving). Two concurrent robots A and B could first sequentially
execute choice of frame and computation, and at a later time execute (again sequentially) ready to move and
move. A robot A can be seen while moving by robot B interleaving move by A, computation by B, and
ready to move by A. However, our modeling with at most one robot activated in any configuration permits
to simplify proof notations.
4
3 Theorem of Choice
Definition 1 (Safe Zone). Let A be an algorithm, let F ⊂ C , F is called a safe zone for A if valid transistions
only lead to configurations of F . Formally:
Lemma 1. Let F ⊂ C and A be an algorithm. If there exists a positive probability p such that, starting from
any configuration and for any scheduler, every execution reaches a configuration in F , then configurations
in F are reached with probability 1. Formally: ∃p ∈]0; 1],∀(C, σ) ∈ C × ΣA , P(∃i ∈ N, EC
σ (i) ∈ F ) ≥ p ⇒
∀(C, σ) ∈ C × ΣA , P(∃i ∈ N, EC
Proof. Let p be a positive real number. For all (C, σ) ∈ C × Σ, we have that:
∀(σ, C) ∈ ΣA × F,∀i ∈ N, EC
σ (i) ∈ F ⇒ EC
σ (i) ∈ F ) = 1.
σ (i + 1) ∈ F .
from C, following σ, F is reached exactly at the i-th transition
i∈N
P (from C, following σ, F is reached exactly at the i-th transition) ≥ p
(cid:33)
P(∃i ∈ N, EC
σ (i) ∈ F ) = . . .
(cid:32)(cid:91)
(cid:88)
= P
=
i∈N
N(cid:88)
i=0
This implies that for all (C, σ) ∈ C × Σ:
∃N ∈ N,
P (from C, following σ, F is reached exactly at the i-th transition) ≥ p
2
⇒ ∃N ∈ N, P (from C, following σ, F is reached before N -th transition) ≥ p
2
We now prove the lemma. Let C be a configuration and σ a scheduler. With execution EC
P(from C, following σ, F is reached before the N -th transition) ≥ p
2 .
For each configuration C and scheduler σ, we can thus define N (C, σ) as the smallest integer N such that:
σ and i ∈ N,
σ (i) in the sequel.
σ be a random execution starting from C. We define (ni)i∈N and (σi)i∈N as n0 = 0, σ0 = σ, and
σ , ni+1). Now, from C ∈ C , following σ, we run n1
2 . If we did reach F we stop. Otherwise, from Cni, following
2 . If we did reach F , we stop. Otherwise we
we define shif t(σ, EC
∀i ∈ N(cid:63), ni+1 = ni + N (Cni, σi), and σi+1 = shif t(σ, EC
transitions so that F is reached with probability p
σ, we run n2 transitions so that F is reached with probability p
repeat the process. Overall the probability that F is never reached is:
σ (i),•). To simplify notations, Ci denotes EC
σ , i) as σ(EC
σ (0), . . . , EC
Let EC
P(from C, following σ, we never reach F ) = P(∀i, EC
σ (i) /∈ F )
∀j ∈(cid:74)ni, ni+1(cid:74), EC
∀j ∈(cid:74)ni, ni+1(cid:74), E
σ (j) /∈ F )
σi (j) /∈ F )
Cni
= P(
(cid:92)
(cid:92)
≤(cid:89)
= P(
i∈N
i∈N
i∈N
1 − p
2
= 0
The following theorem is a direct corollary of Lemma 1 using the closure property of a safe zone.
Theorem 1 (Choice). Let A be an algorithm, let F ⊂ C be a safe zone for A , if ∃p ∈]0; 1],∀(C, σ) ∈
C × ΣA , P(∃i ∈ N, EC
σ (i) ∈ F ) ≥ p, then ∀(C, σ) ∈ C × ΣA , P(∃N ∈ N,∀i ∈ N, i ≥ N ⇒ EC
σ (i) ∈ F ) = 1
So, if F be a safe zone, if there is a positive probability p, such that for any scheduler σ and any reachable
configuration C, any execution from C that follows σ reaches a configuration in F with probability p, then
F is reached with probability 1. A reachable configuration is a configuration with a positive probability of
occurring in the execution. Note that the converse is also true: If F is reached with probability 1, then
starting from any reachable configuration and following any scheduler, F is reached with probability 1 > 0.
While Theorem 1 is very general, we mostly use it for the purpose of scattering robots: our safe zone F is
a set of configurations where robots remain separated whatever the subsequent execution, and we want that
5
starting from any initial configuration C (possibly with already moving robots) and following any scheduler
σ, algorithm A reaches F with probability 1. Then, using Theorem 1, it becomes sufficient to exhibit a
pattern of random bits used by A so that the robots reach F with probability p > 0, and prove that p is
independent of C and σ.
6
4 Blind Rigid ASYNC Scattering in R2
In this section, we consider the rigid setting when all robots are blind but are endowed with weak local
multiplicity detection.
4.1 Algorithm
Algorithm 1 Blind Rigid ASYNC Scattering in R2
1: if there is no other robot at my position then
2:
3: else
4:
5:
6:
Do nothing
i ← 0
r ← rbit()
while r = 0 do
7:
8:
9:
10:
11: end if
i ← i + 1
r ← rbit()
end while
move to u(i)
The pseudo-code for our algorithm is presented as Algorithm 1. Lines 5 and 8 use the rbit() function that
returns a 1 or a 0 with equal probability. The while loop is used to select a random natural number such
that value n is obtained with probability 2−n. The keyword move to takes as input an element of R2 and
updates the destination of the robot. Finally, u(n) is a computable sequence of points in Q2 such that there
exists a circle containing all u(n), and ∀i, j u(i) (cid:54)= u(j) (such a sequence does exist).
Figure 1: Five positions robot A can go to
4.2 Proof of correctness
Lemma 2. APART is a safe zone for Algorithm 1.
Proof. Starting from a configuration in APART, a robot does nothing when there is no other robot at its
position (Line 2).
Lemma 3. With F = APART, and A =Algorithm 1, the conditions of Lemma 1 are satisfied.
7
Proof. Let σ, C ∈ Σ, C . Step by step, we construct a sequence of positions that reach APART with
probability p > 0, independently of σ and C. We denote by (p1, . . . , pn)t the positions occupied by the
robots at step t, and (q1, . . . , qn)t the positions the robots want to go to at step t. Initially, we define:
• (p1, . . . , pn)0 = (C(1)(1), . . . , C(n)(1))
• (q1, . . . , qn)0 = (C(1)(2), . . . , C(n)(2))
Suppose at step t, the current configuration is D, when the scheduler selects a computation transition from
D to D(cid:48) for a robot k on a multiplicity point, the robot chooses its destination for step t + 1, and we define:
• (p1, . . . , pn)t+1 = (D(1)(1), . . . , D(n)(1)), (= (D(cid:48)(1)(1), . . . , D(cid:48)(n)(1)))
• (q1, . . . , qn)t+1 = (D(cid:48)(1)(2), . . . , D(cid:48)(n)(2))
There is a probability ρ > 0 that ∀i ∈(cid:74)0; n(cid:75), i (cid:54)= k ⇒ q /∈ [pi, qi] ∧ qi /∈ [p, q]. We call this event the (∗) event.
This events means that the robot chose a destination that is not on another robot's path (q /∈ [pi, qi]), and
that does not make the robot cross another robot's destination (qi /∈ [p, q]). As there are at most 2n points
that can be selected by k and do not satisfy these two properties (see Figure 2), there is a probability at least
2−2n of choosing a point that does satisfy those two conditions.
Figure 2: Possible destinations for A so that (∗) occurs.
Assume that at each step (∗) happens then we claim that we do at most n steps. Indeed if we do n+1 steps
then it means that a robot computed twice whilst not alone (from now on we only consider the computation
the robot did "whilst not alone"), let us say it is the k-th robot who is the first robot to compute twice. If the
robot computes twice then it means that at the time of its second computation it is not alone on its position,
the h-th robot is also there. Let us call q the current position of the two robots (notice that q was also the
result of the first computation of k), p the position where k was when it did its first computation, p' the
position where h was when k did its first computation. Let q(cid:48) be the destination of h when k does its second
computation. There are two possibilities : either the robot h had q(cid:48) in mind before the first computation of
k or it had it after.
• if h had it before: then when k did its first computation, it got q ∈ [p(cid:48); q(cid:48)], which is impossible by
assumption (by the condition q /∈ [ph, qh]),
• if h had it after: then h did a computation while at position e and chose q(cid:48) as a destination, with
q ∈ [e, q(cid:48)], which is impossible by assumption (by the condition qk /∈ [e; q(cid:48)]).
8
Since (∗) has probability at least 2−2n of occurring, it has probability at least 2−2n2
of occurring n times
in a row (notice that this probability is independent of σ and C). Now, if it occurs n times in a row, then
APART is reached. Therefore APART has a probability at least 2−2n2
of being reached from C, following
σ.
Theorem 2. Algorithm 1 solves the asynchronous rigid scattering problem in R2.
Proof. The proof is a direct application of theorem of choice to the result of Lemma 3
One remaining interesting question is a necessary and sufficient condition about {u(i)}i∈N such that
Algorithm 1 solves the ASYNC blind rigid scattering problem (see Figure 3). One may think that the
following property is true: "Algorithm 1 solves ASYNC blind rigid scattering if and only if {u(i)}i∈N spans
R2". While it is necessary (Algorithm 1 does not solve scattering when all u(i) are on the same line), it is
not sufficient. Indeed, if ∀n, n = 0[2] ⇒ u(0) = (1, 0) ∧ n = 1[2] ⇒ u(1) = (0, 1), Algorithm 1 does not solve
scattering (see Figure 4), yet {u(i)}i∈N is a base of R2. Instead, we propose the following conjecture:
Conjecture 1. The u(i) permit Algorithm 1 to solve ASYNC blind rigid scattering if and only if there exists
no finite set C ⊂ R2 of points such that from any points of C there is a way for the scheduler to choose a
frame that makes every u(i) coincide with points of C.
Figure 3: For which positions of the countably many crosses (that represent the possible destinations) does
the algorithm solve ASYNC scattering?
9
Figure 4: When E moves, a symmetric situation occur, hence we have cyclic set of configurations.
10
5 Flexible Scattering in ASYNC
In this section, we consider the more difficult case of flexible (a.k.a. non rigid) moves, where the scheduler
may stop the robot before it reaches its destination (but not before the robot moves by some distance δ > 0).
The previous blind solution (Algorithm 1) may no longer be valid as the condition for not ending up in
the same location now requires that robot trajectories do not cross. So, the constant probability (from
any configuration) of not reaching another robot's location is not guaranteed, and so neither is scattering.
As a result, we devise new algorithms for this setting that are not blind, first in the unidimensional case
(Section 5.1), and then in the bidimensional case (Section 5.2).
5.1 Flexible ASYNC scattering in R
Do nothing
Algorithm 2 Flexible ASYNC Scattering in R
1: if there is no other robot at my position then
2:
3: else
4:
5:
6:
if there is only one occupied location then
else
7:
8:
9:
10:
11:
12: end if
r ← rbit()
move by a unit distance to the right if r = 0, to the left otherwise
u ← smallest distance between myself and a robot not on my position
r ← rbit()
move by a distance of u
3 to the right if r = 0, to the left otherwise
end if
The pseudocode for our solution is presented as Algorithm 2. The term "unit distance" found in Line 6
refers to the one given by the scheduler. Figure 5 presents the two possible locations a robot located in A
may chose as a target destination when occupied locations are only A, B, and C.
Figure 5: A(cid:48) and A(cid:48)(cid:48) are the two positions a robot located in A can go to (assuming it's not alone on its
position)
Lemma 4. APART is a safe zone for algorithm 2.
Proof. Starting from a configuration in APART, a robot does nothing when there is no other robot at its
position (Line 2).
Theorem 3. Algorithm 2 solves the ASYNC flexible scattering problem in R.
Proof. Our strategy is to increase by at least one the number of isolated (and hence immobile) robots at each
step of the proof. We first fix a left-right orientation (unknown to the robots), and number the robots from 1
to n. Let σ be a scheduler and C a configuration. For the first step, we devise a strategy to put one robot to
the right of every other robot (and it remains so forever): if a robot located at the rightmost position (called
"rightmost robot") is requested to make a choice, then if no other rightmost robot already goes to the right,
go to the right; otherwise, go to the left. We refer to this event as (∗). We now assume that event (∗) always
happens when it can, i.e. our strategy occurs.
1. This strategy occurs with probability at least 2−2n. Indeed, if a rightmost robot decides to go left once,
it is never considered again as it cannot become a rightmost robot again. In more details, Algorithm 2
requires that for a robot A to cross a robot B to the right, B should go to the left. With our strategy,
at least one rightmost robot goes right or stays still. It follows that the number of times we need to
repeat the strategy is at most 2n, as at most n rightmost robots may be requested to go to the left
(and if a rightmost robot goes to the left it must be that a rightmost robot has chosen to go to the
right). Then, we need each robot on the rightmost position stack to make the right choice between left
11
and right, each one has a probability 1
probability at least 2−2n.
2 of making this right choice. So, the strategy is followed with
2. There is a time after which no robot wants to go further to the right of or at the same position as the
rightmost robot, and the rightmost robot does not want to move to the left. Said differently there is
a time after which the rightmost robot remains the rightmost robot forever (we call this property P ).
This is entailed by two facts: (i), (∗) can occur at most 2n times (see above) and if (∗) occurs finitely
many times, then property P is satisfied. It is easy to be convinced of this claim if one notices that a
non-rightmost robot cannot become a rightmost robot via a computation (unless the rightmost robots
all move left). Next, because we follow the strategy, the rightmost robots cannot go left (except maybe
at the beginning).
We now give the technical proof : Assuming the robots follow our strategy, there is a time after which
no one wants to go further to the right of or at the same position as the rightmost robot and the
rightmost robot does not want to move to the left, as a result the rightmost robot will remain the
rightmost robot forever (we call this property P).
We will prove it by induction on the number of robots who may become the rightmost robot at least
once.
Firstly it is easy to be convinced there is a time tright after which the position of the rightmost robot
can only increase.
Secondly if at a time t > tright a robot is still (its position and its destination is the same) and is not
one of the rightmost robots, it can never become one of the rightmost robot. We call this property S.
Let us now prove property P : by contradiction, let us assume that "who the rightmost robot" is changes
an infinite amount of time, let us place ourselves after (∗) does not happen anymore. Since we assume
(∗) happens whenever it can happen then it means that there is no situation, after some time t > tright,
where one of the rightmost robot is asked to compute.
We say that a robot is in the race for rightmost status if there is a scheduler such that following this
scheduler and our strategy there is a non-zero probability that this robot becomes, at least for some
time, the rightmost robot. In other words a robot is in the race for rightmost status if there is a way
for this robot to become one of the rightmost robots at some point.
We place ourselves at time t, we number from 1 to r the robots in the race for rightmost status. There
is a time t' where the rightmost robot is crossed by the robot k (by assumption). There is a time t"
at which k stops (either because the scheduler interrupted him or because k reached its destination) at
this point k's position and k's destination are the same. If k never moves again it is removed from the
race. If k moves once again then it means there is a time t3 where another robot q which goes on top
of it and k performs a computation. If at time t3 k is not the rightmost robot then it is removed from
the race (by property S). But k is not the rightmost robot at time t3 because t3 > t, and by definition
of t the rightmost robot cannot compute after t. Therefore k is removed from the race. By induction
this proves claim P (P is equivalent to having only one robot in the race for rightmost status).
3. After long enough the rightmost robot does not want to move : by the last point there is a time at
which there the rightmost robot does not move or moves to the right by fairness of the scheduler it
will eventually reach its destination or be interrupted and never be reached by another robot (by the
previous point)
The set of configurations F1 were a single rightmost robot does not want to move, and every other robot
does not want to cross it is safe (by points 2 and 3 above). Moreover from any starting configuration C and
for any scheduler σ, there is a probability at least 2−2n to reach this set. By theorem of choice, F1 is thus
reached with probability 1.
For the next step of the strategy, we assume that there exists some integer k ≥ 1 such that the following
property holds: "the set of configuration Fk where the k rightmost robots are lonely and do not want to
move, and every other robot does not want to cross the leftmost of those k rightmost robots is safe". If a
configuration is in Fk, We call Rk the set of robots minus the k single rightmost robots. Then, we use a
strategy similar to the base case of the induction: if you are a rightmost robot of Rk and you are required to
make a choice, if no other robots at your current location is already going to the right, then go to the right,
otherwise, go to the left. By a similar set of arguments as for the base case, after long enough, a configuration
in Fk+1 is reached, and this configuration is safe.
Step by step, we conclude that we eventually reach a configuration in Fn, and this configuration is safe.
Obviously, Fn satisfies APART.
12
5.2 Flexible ASYNC Scattering in R2
is also known as the convex hull. Cardinal directions (Line 7) are given by the scheduler, while the vector −→ex
The pseudo-code for algorithm scattering is given as Algorithm 3. The smallest enclosing polygon (line 10)
(Line 14) denotes the first axis of the frame given by the scheduler to the robot.
When we write "move alongside θ", where θ is an angle, we mean that the robot should move in the
direction given by the vector −→v , of size one and where angle(−→ex,−→v ) = θ. An illustration of the algorithm is
given in Figure 6.
Algorithm 3 Flexible ASYNC Scattering in R2 for robot i
1: if there is no other robot on my position then
2:
3: else
4:
5:
if there is only one occupied location then
Do nothing
the South if r1 = 1 and r2 = 1, to the West otherwise.
else
r1 ← rbit(), r2 ← rbit()
move by a unit distance to the East if r1 = 0 and r2 = 0, to the North if r1 = 0 and r2 = 1, to
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
center G
end if
16:
17: end if
Let A be my current position
Let P be the smallest enclosing polygon of all robots
Let G be the barycenter of P
For every robot (including myself) r, compute Hr, the homothety of P that passes through r with
u ← smallest distance between Hi and another polygon or G
θ ← angle(−→ex,
r1 ← rbit(), r2 ← rbit()
Move by distance u, alongside θ + π
10 if r1 = 0 and r2 = 0, alongside θ − π
−→
GA), u ← u/3
10 if r1 = 0 and r2 = 1,
alongside −θ + π
10 if r1 = 1 and r2 = 0, alongside −θ − π
10 if r1 = 1 and r2 = 1.
Lemma 5. APART is a safe zone for Algorithm 3.
Proof. Starting from a configuration in APART, a robot does nothing when there is no other robot at its
position (Line 2).
Theorem 4. Algorithm 3 solves the flexible ASYNC scattering problem in R2.
Proof. The proof for the R2 case follows the same pattern as the one for R. However, it is now more ambiguous
to talk about "farthest" robots. We use the smallest enclosing polygon to define the set of farthest robots,
but instead of two directions (left and right) in the case of R, we now have multiple angles to distinguish
between farthest robots. In more details,let us denote by −→ex the first coordinate vector of R2, let C be a finite
set of points in R2, let θ be a number in [0; π[, c ∈ C is one of the θ-furthest points if (cid:104)c, eiθ(cid:105) = max
(cid:104)c(cid:48), eiθ(cid:105)
c(cid:48)∈C
where eiθ is the vector of length 1 and such that (cid:104)−→ex, eiθ(cid:105) = θ (see figure 7)
We number robots from 1 to n. Let σ be a scheduler and C a configuration.
For the first step, we choose an angle θ arbitrarily, and we devise a strategy to put a robot the farthest
along θ (called θ-farthest). The strategy to follow is:
if a robot r is one of the θ-furthest robots and is
required to make a choice, if no other robot on the same location is already going θ-further, r pick a direction
that makes it go θ-further, otherwise r picks a direction that does the opposite (that is, makes r go θ-closer).
This event is called (∗). Note that it is always possible to comply with the instructions because among the
4 directions robot r can go to, at least one is making r go θ-further.
• this strategy happens with a probability at least 1
. Indeed, if at some point r decides to go θ-closer,
r never applies the strategy again because r cannot become a θ-farthest robot again (by construction
of the algorithm and the strategy). Therefore the event (∗) happens at most 2n times, since there are
4 equally likely destinations and that at least one goes θ-farther, we get 1
4
−2n
• After long enough, no robot wants to go θ-further or equally θ-far as the θ-furthest robots (there can
eventually be multiple θ farthest robots) : the proof is similar to the uni-dimensional case. Assuming
.
−2n
4
13
Figure 6: After observing the current configuration, robot F (assumed not to be alone) computes the smallest
enclosing polygon and its homotheties, and selects the four possible destinations F 1, F 2, F 3, and F 4.
Figure 7: The displayed robots are the θ-farthest robot for each θ that falls in a section of their color. E is
the 0◦-farthest robot as well the 1◦-farthest robot, D is the −45◦-farthest robot, etc.
14
the robots follow our strategy there is a time after which the set of θ-farthest robots never change
anymore. We call this property P.
Firstly is it easy to be convinced that there is a time tθ after which the value of the projection of the
position of the θ-farthest robots over the vector eiθ only increases.
Secondly if at a time t > tθ a robot is still (its position and its destination is the same) and is not
one of the θ-farthest robots, it can never become one of the θ-farthest robots. This is because, in
the algorithm, the robot can move a distance which is strictly inferior to the distance to the closest
enclosing polygon. We call this property S.
Let us now prove property P : by contradiction, let us assume that the set of θ-farthest robots changes
an infinite amount of time, let us place ourselves after (∗) does not happen anymore. Since we assume
(∗) happens whenever it can happen then it means that there is no situation, after some time t ¿ tθ,
where one of the θ-farthest robot is asked to compute. We say that a robot is in the race for θ-farthest
status if there is a scheduler such that following this scheduler and our strategy there is a non-zero
probability that this robot becomes, at least for some time, the θ-farthest robot.
We place ourselves at time t, we number from 1 to r the robots in the race for θ-farthest status. Since
the set of θ-farthest robot always change by assumption, there is a time t' where robot k becomes a
θ-farthest robot (and it was not just before that). There is a time t" at which k stops (either because
the scheduler interrupted him or because k reached its destination) at this point k's position and k's
destination are the same. If k never moves again there are two possibilities
-- there is a point after which it is not a part of the set θ-farthest robots and it is removed from the
race (since it will never move)
-- or it will always be a part of the set of θ-farthest robots.
If k moves once again then it means there is a time t3 where another robot q which goes on top of it and
k performs a computation. If at time t3 k is not the θ-farthest robot then it is removed from the race
(by property S). But k cannot be θ-farthest robot at time t3 because t3 > t (and no robot θ-farthest
robot can compute after time t). Therefore robot k is removed from the race. By induction this proves
claim P because each time we either discard a robot from ever being in the set of θ-farthest robots or
we put a robot in this set forever, at some point the set of θ-farthest robot will never change anymore
(which is property P).
• After long enough, the θ-furthest robots do not want to move anymore. Once the situation described
in the previous bullet case is reached, we just wait for the θ furthest robots to reach their destination
(or be stopped by the scheduler, for what matters).
After events listed in the above point have happened, the set of θ-farthest robots remains the same forever.
Next steps of the strategy :
A polygon (or shell) is said stable if all robots on the border of the polygon are standing still and no one
inside this polygon wants to go outside or on the border of the polygon. We successively apply the strategy
we used for a specific θ to all θ(cid:48) ∈ [0; π[. For a given θ it can be we are already at a point where the set of
θ-farthest robots will always stay the same and we have nothing to do. The cases where we do have something
to do occur at most n times. Once the strategy has been applied for all θ(cid:48) ∈ [0; π[ the outer shell (or convex
hull) of the robots is stable. Just like in the uni-dimensional case we then apply the same strategy to the
robots strictly inside the stable shell. We simply change in the reasoning "θ-furthest" by "θ-furthest among
the robots that are not part of a stable shell".
At the end, by application of the theorem of choice we reach APART.
Observe that Algorithm 3 very easily extends to higher dimension.
15
6 Complexity
The notion of complexity can be studied in two variants: the expected number of coin tosses until scattering,
and the expected number of rounds until scattering. A round is a smallest fragment of an execution during
which all robots execute each at least one Look-Compute-Move cycle.
Theorem 5. Algorithm 1 achieves scattering of n robots in expected O(22n2
Proof. Let E be the expected number of rounds it takes to scatter the robots. There is a 2−2n2
that APART is reached in one round (see the proof of Lemma 3). Therefore, E ≤ 1 + (1 − 2−2n2
E ≤ O(22n2
Theorem 6. Algorithm 1 uses expected O(2n ∗ 2n2
) coin tosses.
) rounds.
)
probability
) ∗ E ⇒
Proof. The proof is similar to the last theorem.
By tweaking the algorithm 3, we can get a better bound of E ≤ ( 4π2
probability of picking the i-th direction to ∼ 6
π2i2 instead of 2−i.
6 )n ∗ n2n. For this change, the
We know that in the SSYNC setting, scattering can be achieved in Θ(n log n) coin throws on average. In
the ASYNC setting, assuming a robot throws a coin anytime it chooses a new destination, it turns out that
Ω(n2) coin tosses are needed for flexible scattering in R.
Theorem 7. Assuming robots are not blind and have weak local multiplicity detection, any algorithm that
solves flexible ASYNC scattering in R and makes a robot throw at least one coin anytime it chooses a new
destination requires Ω(n2) coin tosses.
Figure 8: Steps of the proof
Proof. At step 1, suppose every robot is at the same position (at the origin of the coordinate system). At
step 2, every robot computes. There is thus a point p (denoted by a red cross on Figure 8) on the left or on
the right such that at least n/2 robots wish to cross p. Out of these n/2 robots, we disseminate n/4 along the
path to p by prematurely stopping them. The n/4 other robots are put on hold. At step 3, we choose one of
the n/4 robots that were put on hold at step 2, and call it q. We now make q move toward p in an multi-stop
fashion: anytime q crosses one of the disseminated n/4 robots, say q(cid:63), q's movement is suspended, and q(cid:63)
computes a new destination (since q(cid:63) and q occupy the same location at this moment). When q reaches p,
the scheduler stops it.
Each of the n/4 has made a computation transition (and hence a coin toss). The scheduler now moves
and interrupts each of those in such a way that they remain between the origin and p. The process is then
repeated with every remaining n/4−1 robots that were put on hold. As a result, each of the n/4 disseminated
robots computes n/4 times, yielding a Ω(n2) number of coin tosses.
Our assumption that a robot tosses a coin anytime it chooses a new destination is justified by the following
observation:
in our schedule, the scheduler only selects for computation robots that are on a multiplicity
point; now if a robot on a multiplicity point does not perform a coin toss, its behavior may be identical to
that of other robots on the same multiplicity point, delaying scattering.
16
Vision Weak Local
Multiplicity
Yes
Yes
No
No
Yes
No
Yes
No
1D rigid
2D rigid
1D flexible
2D flexible
OK (Th. 2)
OK (Th. 2)
OK (Th.3)
KO (Th. 8)
KO (Th. 9)
KO (Th. 10) KO (Th. 10) KO (Th. 10) KO (Th. 10)
OK (Th. 3)
KO (Th. 8)
KO (Th. 9)
OK (Th. 4)
?
?
?
Table 1: Summary of possibility and impossibility results
7 Necessary and Sufficient conditions for ASYNC scattering
In this section, we consider necessary and sufficient conditions for the solvability of ASYNC scattering de-
pending on system hypotheses: rigidity of moves, dimensionality of the locations space, vision, and availability
of weak local multiplicity detection. The summary of our results is presented in Table 1.
Theorem 8. Rigid ASYNC scattering in R is impossible if the robots are blind, even with weak local multi-
plicity detection.
Proof. Consider three robots A, B, and C that are still and A is on the left of B, while C shares B's position.
Let d ∈ R be a positive distance. Then, for any p ∈]0, 1[, the scheduler can force a robot that computes a
new destination to compute a distance at least d with probability p (indeed, the scheduler can change the
scale of the coordinate system at every Look-Compute-Move cycle). We now construct a schedule such that
there is a positive probability that the robots never reach APART. We define routine(i) as "Assume there
are exactly two robots p and q sharing the same location and at least one of them, p, is still, while the third
robot is called h. We also assume that q is not going toward h. If at least one of the assumptions is false,
the routine aborts, otherwise, the scheduler does the following:
• If q is still, the scheduler makes p compute and ensures it chooses a destination that is farther than
dist(p, h) with probability 1 − 1
. If p chooses to go towards h, the scheduler makes it move to
the exact location of h, from which we execute routine(i+1). Otherwise, the scheduler simply executes
routine(i+1).
(i+1)
2
• If q is not still, the scheduler makes p compute and ensures it chooses a destination that is farther than
max(dist(h, p), dist(dest(q), p) with probability 1 − 1
. If p goes towards h, the scheduler makes
it move to the exact location of h and executes routine(i+1). If p goes farther than q, the scheduler
makes q go to its destination, then makes p move to the location of q, and finally execute routine(i+1).
If q goes further than p, the scheduler makes p go to its destination, then makes q move to the location
of p and execute routine(i+1)."
(i+1)
2
aborts, and the robots never reach APART. The probability that this execution occurs is(cid:81)+∞
Now, with the initial configuration we considered, let the scheduler execute routine(0). As we request the
scheduler to ensure a certain moving distance with a certain probability, for any integer i, the routine(i) never
) ≥
(i+1)
i=1 (1− 1
2
1/2.
Theorem 9. Rigid ASYNC scattering in R is impossible if robots have no multiplicity detection, even with
full vision.
Proof. Let us first consider a network of three robots A, B, and C such that A is on the left, and B and
C occupy the same location on the right. All robots are still. Now, when B is activated, there must be a
positive probability that B computes a new position. Suppose B only selects one direction: opposite to A.
Then, the same holds for C, as the algorithm is uniform for all robots. The scheduler then uses the following
schedule: (i) activate B until B decides to move to the right by some distance dB > 0, then (ii) activate
C until C decides to move to the right by some distance dC > 0, then (iii.a) if dB < dC, move B until
it reaches its destination, move C until it reaches B, activate B until it decides to go right again by some
B > dC − dB, or (iii.b) if dB > dC, move C until it reaches its
distance d(cid:48)
destination, move B until it reaches C, activate C until it decides to go right again by some distance d(cid:48)
C > 0,
C > dB − dC. This schedule yields an execution such that robots B and C
then repeat the process until d(cid:48)
occupy the same position infinitely often, hence scattering is never achieved. This implies that robots B and
C must run an algorithm such that there exists a positive probability to move to the left (that is, toward A).
Now, consider a network of two robots A and B initially separated and still (hence satisfying APART),
where A is on the right. Since B has exactly the same vision as in the previous situation (it has no multiplicity
B > 0, then repeat the process until d(cid:48)
17
detection), there is a positive probability that B goes to the left (toward A). By symmetry, there is also a
positive probability that A goes to the right (that is, toward B). Now, the distance can be oblivious of the
distance between A and B (that is, an absolute value based on the unit distance given by the scheduler) or a
function of the distance between A and B (since the scale is given by the scheduler, the perceived distance can
be always one unit, so this equates to a fraction of the distance). In the case of absolute value, the scheduler
simply chooses the distance between A and B, so that they occupy the same position again, invalidating
APART safety. In the case of a fraction, say 1/3 (see Figure 9, the same number is chosen by both A and
B with probability p > 0 since they have the same view. Then A computes a right target with probability
p but does not move, then B computes and move four times in a row, hence, with probability p5, A and B
occupy the same position again, hence APART is not safe. If the fraction is 1/k, for some integer k, the
overall probability of a meeting of A and B becomes dependent of k, but remains positive.
Figure 9: Case where f = 1/3
One may wonder if it is possible to scatter robots that are blind and have no weak local multiplicity
detection (even in R2). The following theorem show that it is not the case.
Theorem 10. Rigid ASYNC scattering in R2 is impossible if robots are blind and have no weak local mul-
tiplicity detection.
Proof. Since the robots do not have weak local multiplicity detection and are blind, there exists probability
p > 0 that the robots move, otherwise they would never separate if placed on the same location. Let d ∈ R2
be a position, since the robots are blind by adjusting the scale and orientation of their referential the scheduler
can take the robot to position d with probability p > 0. Let us suppose we have n > 1 robots in R2 that
execute a scattering algorithm. Let us also suppose those robots are in a configuration that satisfies APART.
We now show that it is possible that two robots meet with positive probability. Let d be the position of robot
number 1, the scheduler activates robot number 2 such that robot 2 has a probability p of choosing d as a
destination. If robot 2 chooses d then we have reached a configuration not in APART with probability p. If
robot 2 does not choose d, the scheduler makes robot 2 move to its destination and once arrived, activates it
again so that it chooses d with probability p. By doing this repeatedly, we eventually reach a conifguration
not in APART.
Admittedly, a fairness issue may arise, if robot 1 is not activated fairly. However, the arguments remains
the same if robot 1 is activated and moves every k (for some integer k > 0) activations of robot 2, as the
probability p is independent of the position of robot 1.
The first unresolved case in this table is whether it is possible to scatter blind robots in R2 with flexible
moves. We conjecture the answer is yes and that Algorithm 1 is a solution. However, the proof argument is
expected to be more involved than the one we provided. The second unresolved case is whether it is possible
to scatter robots without weak local multiplicity detection in R2 (be it rigid or flexible moves). We also
conjecture the answer to this question is yes, but different algorithmic techniques from the ones we provided
are expected to be necessary.
8 Concluding Remarks
We presented the first solutions to the oblivious mobile robots asynchronous scattering problem. Contrary
to previous work, we do not assume that the configuration observed by the robots is always up to date or
does not features moving robots. It turns out that the problem can be solved in the most general case when
robots are able to see other occupied location, and can determine whether more than one robot occupies
their current location. Our positive results are constructive, and a byproduct of our approach is a new
proof technique that could reveal useful for other oblivious mobile robot problems requiring randomization.
Perhaps surprisingly, it remains possible to solve the problem when robots are blind (they cannot see other
occupied locations) in a bidimensional Euclidean space. Releasing both assumptions (vision and weak local
18
multiplicity detection) yields immediate impossibility, while removing one of them yields impossibility in the
case of line. Some cases remain open and most likely require new proof arguments.
While our approach was mostly driven by a computability perspective, our positive result command a
thorough complexity analysis of the problem. Preliminary results we obtained show that, under a reasonable
assumption, a lower bound of Ω(n2) coin tosses is necessary to scatter n robots in ASYNC. This contrasts
with the weaker SSYNC model where Θ(n log n) coin tosses are tight. We leave this path of research as a
future work.
References
[1] Quentin Bramas and S´ebastien Tixeuil. The random bit complexity of mobile robots scattering. Int. J.
Found. Comput. Sci., 28(2):111 -- 134, 2017.
[2] Julien Cl´ement, Xavier D´efago, Maria Gradinariu Potop-Butucaru, Taisuke Izumi, and St´ephane Messika.
The cost of probabilistic agreement in oblivious robot networks. Inf. Process. Lett., 110(11):431 -- 438, 2010.
[3] Yoann Dieudonn´e and Franck Petit. Scatter of robots. Parallel Processing Letters, 19(1):175 -- 184, 2009.
[4] Paola Flocchini, Giuseppe Prencipe, and Nicola Santoro, editors. Distributed Computing by Mobile Enti-
ties, Current Research in Moving and Computing, volume 11340 of Lecture Notes in Computer Science.
Springer, 2019.
[5] Paola Flocchini, Giuseppe Prencipe, Nicola Santoro, and Peter Widmayer. Gathering of asynchronous
robots with limited visibility. Theor. Comput. Sci., 337(1-3):147 -- 168, 2005.
[6] Paola Flocchini, Nicola Santoro, Giovanni Viglietta, and Masafumi Yamashita. Rendezvous with constant
memory. Theor. Comput. Sci., 621:57 -- 72, 2016.
[7] Taisuke Izumi, Tomoko Izumi, Sayaka Kamei, and Fukuhito Ooshita. Feasibility of polynomial-time
randomized gathering for oblivious mobile robots. IEEE Trans. Parallel Distrib. Syst., 24(4):716 -- 723,
2013.
[8] Ichiro Suzuki and Masafumi Yamashita. Distributed anonymous mobile robots: Formation of geometric
patterns. SIAM J. Comput., 28(4):1347 -- 1363, 1999.
19
|
1909.05377 | 1 | 1909 | 2019-09-11T21:13:49 | Multi-Agent Control Using Coverage Over Time-Varying Domains | [
"cs.MA",
"eess.SY",
"eess.SY"
] | Multi-agent coverage control is used as a mechanism to influence the behavior of a group of robots by introducing time-varying domain. The coverage optimization problem is modified to adopt time-varying domains, and the proposed control law possesses an exponential convergence characteristic. Cumbrous control for many robots is simplified by deploying distribution and behavior of the robot team as a whole. In the proposed approach, the inputs to the multi-agent system, i.e., time-varying density and time-varying domain, are agnostic to the size of the system. Analytic expressions of surface and line integrals present in the control law are obtained under uniform density. The scalability of the proposed control strategy is explained and verified via numerical simulation. Experiments on real robots are used to test the proposed control law. | cs.MA | cs |
Multi-Agent Control Using Coverage Over
Time-Varying Domains
Xiaotian Xu∗ and Yancy Diaz-Mercado
Abstract
Multi-agent coverage control is used as a mechanism to influence the behavior of a group of robots by
introducing time-varying domain. The coverage optimization problem is modified to adopt time-varying
domains, and the proposed control law possesses an exponential convergence characteristic. Cumbrous
control for many robots is simplified by deploying distribution and behavior of the robot team as a
whole. In the proposed approach, the inputs to the multi-agent system, i.e., time-varying density and
time-varying domain, are agnostic to the size of the system. Analytic expressions of surface and line
integrals present in the control law are obtained under uniform density. The scalability of the proposed
control strategy is explained and verified via numerical simulation. Experiments on real robots are used
to test the proposed control law.
1 Introduction
Coverage control of multi-agent system has been drawing attention for a long time due to its wide appli-
cations, such as surveillance or exploration of an interested region. However, most of works consider these
regions are static. We investigate the case that the interested areas are dynamic in this paper and extend
the potential applications of coverage control of multi-agent systems. Typically, coverage control employs a
group of robots to optimally cover an interested area. A classical solution to coverage problems is proposed
by involving proper partitions of the domain[1, 2]. Coverage control based schemes have been previously de-
veloped for multi-agent systems, e.g., broadcast control (BC) scheme [3] and pivot-based collective coverage
algorithm [4]. In [5, 6] the authors introduce a mechanism to influence the collaborative behavior between
robots in coverage control by associating time-varying densities to the domain. An approach to perform
coverage control on moving regions is discussed in [7]. However, the robot team is not able to maintain the
centroidal Voronoi tessellation (CVT) configuration of the moving region, which will be one of the topics
considered in this paper.
In this paper, the coverage control algorithm [5] is taken as a mechanism to affect the behavior of a big
multi-robot team by involving time-varying domains. We define a time-varying subdomain in the workspace,
and command the robots to optimally distribute in the time-varying subdomain. Under this strategy, when
the subdomain moves and the shape or the scale of the subdomain changes, the robot team will act in
a coordinating manner. The proposed control strategy allows the multi-objective control of multi-agent
systems to be simplified into manipulating the subdomain directly, e.g., patrolling along the trajectory, and
changing the size of robot group. Nevertheless, the formation of the robot team can be controlled by adopting
time-varying densities within the subdomain as suggested in [6, 8].
The outline of this study is as follows. In Section 2, we state the coverage problem and recall pertinent
prior work. The coverage control over time-varying domains strategy is proposed in Section 3. Analytic
solutions are presented in Section 4 under the uniform density. The scalability of the proposed algorithm
is presented and validated in Section 5. A robotic implementation of the proposed control scheme together
with results of experiments are illustrated in Section 6. Finally, conclusions are summarized in Section 7.
∗Authors are with the Department of Mechanical Engineering, 2181 Glenn L. Martin Hall, Building 088, University of
Maryland, College Park, MD 20742, USA. Email: [email protected], [email protected].
1
2 Preliminaries
Coverage control will be used as a mechanism to influence the behavior of the multi-robot team, to coordinate
their inter-robot motion, and manipulate their motion in the workspace. In this section, we provide some
preliminary descriptions of the coverage control problem that is addressed with the modifications needed to
account for time-varying domains.
2.1 The Coverage Problem
Let pi ∈ D ⊆ Rd be the position of the ith robot, i ∈ {1, . . . , n}, in the domain of interest D, i.e., the robot
workspace. Further define a convex time-varying subdomain S(t) ⊂ D, such that robot i is said to lie in the
subdomain at time t if pi(t) ∈ S(t). Let ∂S(t) denote the boundary of the subdomain at time t, and let q(t)
be differentiable for almost every q ∈ ∂S.
We will use the locational cost [9] as a metric of the coverage performance in the subdomain S(t) at time
H(p(t), t) =
(cid:107)pi(t) − q(cid:107)2 φ(q, t) dq
(1)
t:
where p(t) = (cid:2)pT
n (t)(cid:3)T
(cid:90)
n(cid:88)
i=1
Vi(p(t),t)
1 (t), . . . , pT
is the configuration of the multi-robot team and φ : S(t) × [0,∞) →
(0,∞) is a density function that captures the relative importance of the points in the subdomain at time t,
differentiable in both arguments. The subdomain is partitioned into regions of dominance, and these form a
proper partition of the subdomain. We utilize a Voronoi tessellation of the domain, given by
Vi(p, t) = {q ∈ S(t) (cid:107)pi − q(cid:107) ≤ (cid:107)pj − q(cid:107) ∀j}
where for ease of notation we have dropped the explicit time dependency on the configuration of the multi-
robot system.
2.2 Centroidal Voronoi Tessellations
A necessary condition for the minimization of the locational cost in (1) is known to be that the agents form
a centroidal Voronoi tessellation (CVT) of the domain [9], i.e.,
∀i
pi(t) = ci(p, t)
where we define ci(t) ∈ Vi(p, t) to be the center of mass of Voronoi cell i at time t, given by
ci(p, t) =
Vi(p,t) qφ(q, t) dq
mi(p, t)
where mi(p, t) is the mass of the corresponding cell,
2.3 Coverage Control Law
Vi(p,t)
mi(p, t) =
φ(q, t) dq.
(cid:82)
(cid:90)
2
(2)
(3)
(4)
In [5], a control law was proposed which was later shown in [6] to achieve exponential converge to a CVT in
the case of time-varying densities. This control law was called TVD-C for time-varying densities, centralized
case, given by
where the tuning parameter κ > 0 controls the exponential convergence rate, and
p =
(cid:19)
I − ∂c
∂p
(cid:19)−1(cid:18)
(cid:18)
c(p, t) =(cid:2)c1(p, t)T , . . . , cn(p, t)T(cid:3)T
κ(c(p, t) − p) +
∂c
∂t
.
When many agents are used, the control law called TVD-D1, which stands for time varying densities,
decentralized case with 1-hop adjacency information, bypasses difficulties with computing the matrix inverse
in (4) by approximating it with the truncated Neumann series,
(cid:18)
(cid:19)(cid:18)
p =
I +
∂c
∂p
κ(c(p, t) − p) +
(cid:19)
∂c
∂t
.
(5)
The matrix ∂c
∂p in (4) and (5) is a block matrix that possess the sparsity structure of the Delaunay graph
associated with the Voronoi tessellation [6]. The ijth block is given by
(cid:21)
(cid:20) ∂c
(cid:20) ∂c
(cid:21)
∂p
∂p
∂ci
∂pj
=
ij
∂ci
∂pi
=
=
ii
= −
(cid:82)
(cid:88)
j∈NVi
mi(p, t)(cid:107)pi − pj(cid:107)
∂Vij (p,t)(q − ci)(q − pj)T φ(q, t) dq
(cid:82)
∂Vij (p,t)(q − ci)(q − pi)T φ(q, t) dq
mi(p, t)(cid:107)pi − pj(cid:107)
(6)
(7)
when i (cid:54)= j, and
where ∂Vij = Vi ∩ Vj, and NVi is the Delaunay graph neighbor set of agent i. These are (d−1)-dimensional
integrals (e.g., line integrals if D ⊆ R2), and are zero if ∂Vij = ∅ (i.e., two cells are not adjacent), or if the
T(cid:105)T
shared boundary has zero (d−1)-dimensional measure (e.g., points in 2D).
T
∂t in (4) is given by ∂c
∂t =
, . . . , ∂cn
∂t
,
(cid:104) ∂c1
∂t
(cid:82)
Assuming that the domain of interest is static, the partial ∂c
Vi(p,t)(q − ci) ∂φ
∂t dq
where
∂ci
∂t
=
mi(p, t)
.
(8)
Although the control law in (4) was derived with consideration to time-varying densities over static
domains, the proof of exponential convergence in [6] is still valid when the domain of interest is time-varying
with almost everywhere differentiable boundary, as long as pi(t0) ∈ S(t0) ∀i. However, additional terms
are needed in (8) to capture the evolution of the subdomain. In the following sections, we will derive the
needed terms such that (4) is able to provide exponential convergence to a CVT configuration, even when
the subdomain of interest is time-varying.
3 Time-Varying Domains
In order to retain the exponential convergence to a CVT property of control law (4), we will need to add
additional terms in (8) to capture the evolution of the time-varying domain. For the subsequent analysis, it
is assumed that dq
dt exists and is known for almost every q ∈ ∂S(t).
Note that the center of mass integrals in (2) are time dependent in both the integral kernel and the
Lemma 1 (Leibniz Integral Rule[10]). Let Ω(p, t) be a region that depends smoothly on t and that has a
domain of integration. Thus, Leibniz integral rule is needed.
well-defined boundary ∂Ω(p, t). If F (p, t) =(cid:82)
(cid:90)
(cid:90)
where n(q) is the unit outward normal for q ∈ ∂Ω(p, t).
Application of Lemma 1 to (2) yields the following result.
Ω(p,t) f (p, t, q) dq for differentiable f , then
∂F
∂t
=
∂f
∂t
dq +
Ω(p,t)
∂Ω(p,t)
f (p, t, q)
T
∂q
∂t
n(q) dq
Proposition 1 (Time-Varying Domains). Let S(t) be the convex subdomain to be covered with boundary
dt exists and is known for almost every q ∈ ∂S(t). Let ∂Si(p, t) = ∂S(t)∩ Vi(p, t). Then
∂S(t). Assume that dq
the partial derivative ∂ci
∂t , which captures the time evolution of the center of mass of the ith Voronoi cell due
to the motion of the subdomain and the change in density function, is given by
∂ci
∂t
=
1
mi(p, t)
Vi(p,t)
(q − ci)
∂φ
∂t
dq +
1
(q − ci)φ(q, t)
T
dq
dt
n dq.
(9)
(cid:90)
(cid:90)
mi(p, t)
∂Si(p,t)
3
Proof. By the product rule, we find
∂
∂t
∂ci
∂t
=
(cid:16)(cid:82)
(cid:17)
Vi(p,t) qφ(q, t) dq
mi(p, t)
− ci(p, t)
mi(p, t)
∂mi
∂t
.
Both terms now contain derivatives of integrals with time-varying integral kernels and domains. Applying
Leibniz integral rule to the first term yields
(cid:32)(cid:90)
∂
∂t
(cid:32)(cid:90)
∂
∂t
(cid:33)
(cid:33)
(cid:90)
(cid:90)
(cid:90)
(cid:90)
qφ(q, t) dq
=
Vi(p,t)
Vi(p,t)
∂Vi(p,t)
q
∂φ
∂t
dq +
qφ(q, t)
T
∂q
∂t
n dq
Then the latter integral can be split into two terms: for q ∈ ∂Si(p, t), and for q ∈ ∂Vi(p, t)\∂Si(p, t). Note
that only the points on the boundary of the subdomain change explicitly with time, i.e., since ∂q
∂t = 0 for
almost every point q /∈ ∂Si(p, t) we get
qφ(q, t) dq
=
Vi(p,t)
Vi(p,t)
q
∂φ
∂t
dq +
qφ(q, t)
T
dq
dt
n dq.
∂Si(p,t)
Similarly, for ∂mi
∂t we find
Collecting like terms, we get
∂ci
∂t
=
as was to be shown.
(cid:90)
Vi(p,t)
∂mi
∂t
=
∂φ
∂t
dq +
(cid:90)
∂Si(p,t)
φ(q, t)
T
dq
dt
n dq.
(cid:82)
Vi(p,t)(q − ci) ∂φ
∂t dq
mi(p, t)
(cid:82)
+
∂Si(p,t)(q − ci)φ(q, t) dq
dt
T
n dq
mi(p, t)
(cid:4)
Corollary 1 (Exponential Convergence). The control law in (4) with updated partial derivative as in Propo-
sition 1 yields exponential convergence to a CVT with exponential decay rate controlled by κ > 0 over smoothly
time-varying domains and densities, as long as pi(t0) ∈ S(t0) ∀i.
Proof. Same as in [6].
(cid:4)
Proposition 1 introduces a new term in the computation of ∂ci
∂t when compared to (8), which allows us to
explicitly take into account the evolution of the subdomain. This additional term can serve as an exogenous
input to the multi-robot team to control their collective position and scale, while the density function can
be chosen to provide the desired shape in multi-robot team formations, as suggested in [11]. These choices
of inputs possess the advantages of being flexible with respect to the size of the system, and as well as being
identity-agnostic, so that the input may be chosen without the need of performing assignments of roles, such
as is needed in leader-follower schemes.
In the next section, we explore the coverage problem over uniform densities in convex 2-polytope subdo-
mains, and analytical expressions are found for the terms in (4) and (5).
4 Coverage over Convex 2D-Polytopes with Uniform Density
We now focus on the class of coverage problems where the subdomain is a convex polytope in R2 with N
vertices using uniform density, i.e., with φ(q, t) = 1 for all q ∈ S(t) and t ≥ 0. As an immediate consequence
of this, the first term in expression provided in Proposition 1 becomes zero. For this class of problems, we
will further see that we can find analytical expressions in terms of time, neighboring agent positions and
boundary conditions in that agent's Voronoi cell. The control law can be computed solely on local neighbor
information and broadcasted domain information. We begin by considering the vertices of a Voronoi cell.
4
4.1 Vertices of the Voronoi Cell
Assuming that agent i has h(cid:48) neighbors, the ith Voronoi cell will be a convex 2-polytope with at least h(cid:48)−1
vertices. To account for vertices due to the intersection of the Voronoi cell with the the subdomain boundary,
we will assume the Voronoi cell consists of M ≥ h(cid:48)−1 vertices. We will denote the vertex due to neighbors
j and k (which can be determined by employing a Delaunay triangulation [12]) as Vijk. The vertex location
Vijk is given by the circumcenter of the triangle formed by connecting the position of these three agents pi,
pj and pk, and can be found analytically in terms of these [13, 14]. We present the equation below for the
sake of completion,
where the agents are assumed to be oriented in a counterclockwise order, where pab = pb − pa, and where
Vijk =
1
2
αipi + αjpj + αkpk
pij2pjk2 − (pT
ijpjk)2
αi = (cid:107)pjk(cid:107)2pT
ijpik, αj = (cid:107)pik(cid:107)2pT
ijpkj, αk = (cid:107)pij(cid:107)2pT
ikpjk.
Denote ViS(cid:96) to be the (cid:96)th vertex of the ith Voronoi cell that is on the subdomain boundary, where without
loss of generality these are assumed to be ordered counterclockwise. These may be domain polytope vertices,
or may be vertices due to the intersection of the Voronoi cell with the subdomain. As the former vertices
are assumed to be known, we provide an analytical expression for the latter.
Any point q on a face of the subdomain may be expressed using the vertices that define the face via the
parameterization q = L(cid:96)(τ ) = ∂S (cid:96) + (∂S ((cid:96)+1) − ∂S (cid:96))τ for certain τ ∈ [0, 1], where without loss of generality
the vertices of the subdomain ∂S (cid:96) are assumed to be ordered counterclockwise, and where ∂S N +1 = ∂S 1.
Let any point q ∈ ∂Vij, the interior Voronoi face due to two agents i and j that intersects polytope edge L(cid:96),
be given by q = Rij(s) = 1
skew symmetric rotation matrix. The sough after vertex ViSj is given by
2 (pi + pj) + tijs for certain s ∈ R, where tij = S(pj − pi) with S =(cid:2) 0 −1
(cid:3) being a
1 0
(cid:16)
∂S ((cid:96)+1) − ∂S (cid:96)(cid:17)
τ∗
ViSj = ∂S (cid:96) +
which exists if
τ∗ =
2 (pi + pj) − ∂S (cid:96))
( 1
(cid:0)∂S ((cid:96)+1) − ∂S (cid:96)(cid:1)T
(pj − pi)
∈ [0, 1]
found from solving for τ in L(cid:96)(τ ) = Rij(s).
The mass and center of mass can now be computed analytically in terms of the vertices of the Voronoi
cell.
4.2 Mass and Center of Mass of a Voronoi Cell
For convenience, denote the vertices of the ith Voronoi cell as Vij(p, t), j ∈ {1, . . . , M}, which are assumed
without loss of generality to be ordered counterclockwise. Then the mass and center of mass may be computed
as [14]
mi(p, t) =
V T
i(j+1)SVij
1
2
M−1(cid:88)
(cid:0)Vi(j+1) + Vij
j=1
(cid:1)(cid:16)
N−1(cid:88)
(cid:17)
.
V T
i(j+1)SVij
ci(p, t) =
1
6mi(p, t)
j=1
4.3 Partial Derivatives of a Centroid
For a convex 2-polytope subdomain, every Voronoi cell is also a convex 2-polytope, and the boundary
integrals in (6), (7), and (9) may be solved along each face analytically by using the parametric line integral
as in the following lemma.
5
Lemma 2 (Parametric Line Integral). Let the line integral of f (q) along L be denoted by (cid:82)
L f (q)dq where
the point q = [q1, q2]T can be parameterized such that if q1 = h(τ ) and q2 = g(τ ) for some τ ∈ [t0, t1], then
(cid:90)
(cid:90) t1
f (q)dq =
L
t0
f ([h(τ ), g(τ )]T )q(cid:48)(τ )dt
(cid:113)
where q(cid:48)(τ ) =
dτ )2.
dτ )2 + ( dg
( dh
Let that ∂Si(p, t) = ∂S(t) ∩ Vi(p, t) = (cid:83)
j ∂S j
(cid:88)
subdomain boundary. Each of the ∂S j
By linearity of the integral, the partial derivative (9) may be computed as
i
is a straight line, connected by two Voronoi cell vertices vj
is the jth Voronoi cell face shared with the
1 and vj
2.
i , where ∂S j
i
(cid:90)
∂ci
∂t
=
νj(t)
mi
j
∂Sj
i
(q − ci) dq
where νj(t) is the velocity of the points in the boundary ∂Sj
projected in the unit outward direction (i.e., νj(t) = ± dq
S vj
(cid:107)vj
of the vertices).
dt
T
i , assumed to be the same for every point,
2−vj
i depending on the orientation
2−vj
1(cid:107)∀q ∈ ∂Sj
1
Lemma 3 (Analytic Partial Derivatives).
Under uniform density and a convex 2-polytope subdomain with ∂S j
(cid:16)(cid:16)
(cid:17)
(cid:16)
(cid:17)(cid:17)(cid:107)vj
i = vj
1vj
2, (9) can be expressed as
2 − vj
1(cid:107)
∂ci
∂t
=
νj(t)
2mi
2 − ci
vj
+
1 − ci
vj
(cid:88)
j
and for ∂Vij = vj
= − (cid:107)vj
1vj
2, (6) may be expressed as
2 − vj
1(cid:107)
mi(cid:107)pi − pj(cid:107)
1 − pj
vj
1 − ci
vj
(cid:17)(cid:16)
(cid:104)(cid:16)
∂ci
∂pj
(cid:17)T
(cid:16)
+
1
2
1 − ci)(vj
(vj
2 − vj
1)T + (vj
1 − pj)T(cid:17)
(cid:17)(cid:16)
2 − vj
vj
1
2 − vj
vj
1
1)(vj
2 − vj
(cid:16)
1
3
+
(cid:17)T(cid:105)
.
Proof. Assume the following parameterization of the points q ∈ ∂Sj
i = vj
1vj
2 ,
It follows that q(cid:48) = (vj
in (9),
2 − vj
1) for all τ , and consequently dq = (cid:107)vj
2 − vj
1(cid:107)dτ . Applying Lemma 2 to the integral
q(τ ) = vj
1(1 − τ ) + vj
(cid:90) 1
(cid:16)
(cid:88)
j
2τ, τ ∈ [0, 1].
(cid:17)(cid:107)vj
∂ci
∂t
=
νj(t)
mi
0
1(1 − τ ) + vj
vj
2τ − ci
2 − vj
1(cid:107) dτ.
The result follows from expanding and integrating the terms. A similar process can be applied to (6) by
(cid:4)
instead using the vertices that make up the boundary for ∂Vij in the parameterization.
5 Scalability of the Algorithm
In this section, we will argue that the proposed algorithm is scalable for a large group of robots (e.g., robot
swarms). As a well-known fact in the field of computational geometry, the number of edges meeting at a
vertex in the Voronoi diagram is not less than three, and each edge connects two vertices; it follows Euler's
relation that the average number of edges of a Voronoi cell equals six[15, 16]. This fact is the key for our
algorithm's significantly reduced computational cost.
6
Table 1: Number of agents (out of 100) categorized by their neighbor sets Ni (i.e., number of neighbors).
Size of Ni
Simulation
2
3
4
5
6
Number of Agents
Trial 1
Trial 2
Trial 3
Trial 4
Trial 5
4
1
2
3
3
Average
2.6
1
4
3
4
3
3
31
29
28
28
27
28.6
10
9
9
8
9
9
47
50
52
51
52
7
7
7
6
6
6
50.4
6.4
5.1 Complexity Analysis of the Algorithm
Here we analyze the time complexity of the proposed algorithm for a single agent pi.
1. There are many well-studied algorithms to construct Voronoi diagrams, such as the Incremental Inser-
tion Method and the Divide and Conquer Method[15]. The quaternary incremental algorithm builds
the Voronoi diagram for n seeds with average complexity O(n) despite its worst-case complexity O(n2),
and the Divide and Conquer Method has an average and worst-case complexity O(n log n)[16]. In our
algorithm, agents only require the density and boundary information within their own Voronoi cells
(which can be computed in a distributed way [14]) and their Delaunay neighbor positions to compute
the quantities in (4). As described in [5], the matrix inverse in TVD-C can be approximated using
the truncated Neumann series expansion so that an agent only use as much information about the
multi-robot team as it has access to. For one agent in our algorithm, only a partial Voronoi diagram,
which is generated by the agent itself, its neighbors (Ni = 6 on average), and its neighbors' neighbors
(to enclose the Voronoi cells of Ni, a total of 12 on average) is needed instead of the overall Voronoi
, which is on average constant time,
diagram. Thus, the time complexity becomes O
by using the quaternary incremental algorithm.
(cid:16)(cid:12)(cid:12)(cid:12)(cid:83)
(cid:12)(cid:12)(cid:12)(cid:17)
Nj ∪ Ni
j∈Ni
2. We can obtain necessary information for ∂c
∂p in the control law (i.e., masses mi and mj, centroids ci and
cj of Voronoi cell Vi and Vj ∈ Ni) in constant time thanks to analytic expressions derived in previous
section. The complexity becomes O(1) on average in light of the fact of average number of neighbors
is independent of the size of the robot team.
3. The complexity of computing ∂ci
∂t
is determined by the number of edges of its Voronoi cell (O(M )).
Similarly to the previous cases, as the number of edges depends on the number of neighbors, the
complexity is constant time on average.
4. Given the above terms, the control law for agent pi can be calculated with time complexity O(1).
5.2 Validation of Scalability
We show the asymptotic nature of the average number of neighbors and the efficiency of the control law for
a large group of robots via numerical simulations. The simulation of the proposed control strategy was run
for five trials using 100 robots with random initial configurations, and the outcomes are collected in Table
1. Around fifty percent of agents have 6 neighbors in the simulations. The majority of the remaining agents
actually had less, with only about 7% having no more than seven neighbors.This percentage increased with
the size of the swarm.
Moreover, simulations of the proposed control law with and without term ∂c
∂t are conducted. The simula-
tion is shown in Fig. 1i, where a team of 100 robots is performing coverage of a subdomain which is tracking
a circular trajectory. The instantaneous aggregated CVT error ea(t) = (cid:107)p(t)−c(p, t)(cid:107) of the proposed control
strategy with and without the term ∂c
∂t is shown in Fig. 1ii. As we can tell from the results, the steady state
7
(i)
(ii)
Figure 1: Simulation of a group of 100 single-integrator robots (i); the center of subdomain (asterisk) moves
along the circular trajectory (dash line) in counterclockwise fashion. Results of the simulation of the control
law with and without the additional feedforward term ∂c
∂t (ii).
mean values of the simulation with and without ∂c
∂t are 0.0562 m and 0.1248 m respectively. The aggregated
CVT error 0.0562 m for 100 robots (i.e., 0.000562 m on average for each robot) suggests that the proposed
control law converges to a CVT. The ∂c
∂t term improves the error performance by 54.97% in this scenario
with a large group of agents. The error is not exactly zero because the distributed form of the control law (5)
was used, which does not perform feedforward on every agent in the team. Thus, the agents on the boundary
can immediately follow the motion of subdomain as they have direct access to the boundary conditions (i.e.,
whose Voronoi cell boundaries have overlapped edges with subdomain's boundaries), but the agents on the
interior must wait for the boundary conditions to propagate through feedback.
6 Multi-Robot Implementation
(i)
(ii)
Figure 2: Multi-robot implementation of control law with (i) and without (ii) term ∂c
∂t . 10 robots travel from
left end of workspace (blue box) towards the right end (green box) while avoiding the obstacles (red boxes).
The subdomain translates and scales over time. An overhead projector is used to visualize the subdomain,
goal location and obstacles in real time.
In section 4, the coverage control laws (4) and (5) are modified to account for time-variations in the
subdomain and the terms needed for the equations are provided analytically in the case of uniform coverage
on a convex 2-polytope domain. The results were simulated on a large team of robots in Section 5. In this
section, we implement the proposed control strategy on a group of real mobile robots. The experiments are
carried out on a workstation with Intel(R) Xeon(R) W-2125 processor, and the algorithms are implemented
in MATLAB. The multi-robot team consists of 10 Khepera IV differential-drive mobile robots. A motion
capture system consisting of 8 Vicon Vantage V8 cameras are used to provide real-time position information
8
Figure 3: Results of experiments. Profiles of aggregated error from optimal configuration with and without
term ∂c
∂t are indicated by blue solid line and red dot line
of the multi-robot team, and an overhead projector is used to visualize a virtual environment and Voronoi
tessellation. The robot workspace is defined to be a 5.182 m × 3.658 m (17 feet × 12 feet) rectangular area.
6.1 Experiment Description
An experiment is carried out to validate the proposed control strategy with a rectangular subdomain ∂S(t) =
[xmin(t), xmax(t)] × [ymin(t), ymax(t)] with uniform density. In the experiment, we synthesize control laws
for the different agents by driving the coverage domain for a group of robots to achieve a motion plan. The
experiment is shown in Fig. 2. To achieve the task, the subdomain simultaneously translates and scales to
go through the narrower passage and avoid the obstacles. The velocities for the subdomain boundaries are
defined as piece-wise constants to achieve the desired scaling and translation. The metric used to determine
performance of the proposed control strategy on the real robots is same as that in simulation. The control
law (5) and analytic expressions presented in Section 4 are used for the experiment with a control gain κ = 2.
As mentioned in Corollary 1, the aggregated error ea is expected to decay exponentially to zero under
control law TVD-C (almost exponential convergence under control law TVD-D1), up to the error introduced
by the mapping of the control law to the differential-drive motion of the robots, wheel saturations, and delays.
For comparison, the experiment is executed with and without the term ∂c
∂t , to evaluate the contribution by
the inclusion of the term.
6.2 Experimental Results
From Fig. 2, we can tell the difference between performances with and without ∂c
∂t . The robots form and
maintain an almost symmetric configuration with ∂c
∂t (Fig. 2i) while the robots lag behind the motion of
subdomain (Fig. 2ii) and could fail to catch up if the subdomain moves faster. Moreover, the profiles of
(cid:107)p(t) − c(p, t)(cid:107) are plotted in Fig. 3 for the experiments of implementing the control law with the inclusion
of the ∂c
∂t term and the absence of it, and they are shown as blue solid line and red dot line respectively. The
aggregated error ea greatly decreases at first, as expected due to the exponential convergence. As the agents
approach their CVT configuration, delays, saturations, and errors in mapping velocities to differential-drive
introduce small disturbances that result in variations along a small constant value. The mean of the steady-
state values of the metric are 0.0323 m (i.e., 0.0032 m per robot -- well within the footprint of the robots
which are 0.14 m in diameter) and 0.057 m for with and without the inclusion of ∂c
∂t . There is a noticeable
improvement in ea of nearly 43.3% due to the presence of the feedforward term.
9
7 Conclusion
An innovative control scheme is developed of synthesizing multi-agent controllers via coverage control on
time-varying domains. The approach offers a capability of controlling the behavior of multi-agent systems
by simply manipulating the domain to be covered, and it has the advantage of being agnostic to the size of
the system. Analytic solutions for the uniform density case is obtained, and scalability of the approach is
demonstrated. The feasibility of proposed control strategy is validated in simulation for a large number of
agents and experimentally on a team of real differential-drive wheeled robots.
References
[1] Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory, 28(2):129 --
137, 1982.
[2] Jorge Cortes, Sonia Martinez, and Francesco Bullo. Spatially-distributed coverage optimization and
control with limited-range interactions. ESAIM: Control, Optimisation and Calculus of Variations,
11(4):691 -- 719, 2005.
[3] Shalini Darmaraju, Md Abdus Samad Kamal, Madhavan Shanmugavel, and Chee Pin Tan. Coverage
control of a mobile multi-agent serving system in dynamical environment. In 2018 Joint 7th International
Conference on Informatics, Electronics & Vision (ICIEV) and 2018 2nd International Conference on
Imaging, Vision & Pattern Recognition (icIVPR), pages 508 -- 513. IEEE, 2018.
[4] Shaocheng Luo, Jun Han Bae, and Byung Cheol Min. Pivot-based collective coverage control with
a multi-robot team. In 2018 IEEE International Conference on Robotics and Biomimetics (ROBIO),
pages 2367 -- 2372. IEEE, 2018.
[5] S. G. Lee, Y. Diaz-Mercado, and M. Egerstedt. Multirobot control using time-varying density functions.
IEEE Transactions on Robotics, 31(2):489 -- 493, April 2015.
[6] Y. Diaz-Mercado, S. G. Lee, and M. Egerstedt. Human-swarm interactions via coverage of time-varying
densities. In Trends in Control and Decision-Making for Human -- Robot Collaboration Systems, chap-
ter 15, pages 357 -- 385. Springer, Cham, Switzerland, 2017.
[7] Sora Nishigaki, Kenta Hoshino, and Jun Yoneyama. Coverage control in moving regions with unmanned
aerial vehicles. In 2019 IEEE/SICE International Symposium on System Integration (SII), pages 301 --
306. IEEE, 2019.
[8] Xiaotian Xu and Yancy Diaz-Mercado. Multi-robot control using coverage over time-varying domains
In The 2nd International Symposium on Multi-Robot and Multi-Agent Systems
extended abstract.
(MRS2019). IEEE, In Press.
[9] J. Cortes, S. Martinez, T. Karatas, and F. Bullo. Coverage control for mobile sensing networks: varia-
tions on a theme. In Mediterranean Conference on Control and Automation, 2002.
[10] Qiang Du, Vance Faber, and Max Gunzburger. Centroidal voronoi tessellations: Applications and
algorithms. SIAM review, 41(4):637 -- 676, 1999.
[11] Y. Diaz-Mercado, S. G. Lee, and M. Egerstedt. Distributed dynamic density coverage for human-swarm
interactions. In 2015 American Control Conference (ACC), pages 353 -- 358, July 2015.
[12] Der Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. Inter-
national Journal of Computer & Information Sciences, 9(3):219 -- 242, 1980.
[13] Dan Pedoe. Circles: a mathematical view. Cambridge University Press, 1995.
[14] J. Cortes, S. Martinez, T. Karatas, and F. Bullo. Coverage control for mobile sensing networks. IEEE
Transactions on Robotics and Automation, 20(2):243 -- 255, April 2004.
10
[15] Franz Aurenhammer. Voronoi diagrams -- a survey of a fundamental geometric data structure. ACM
Computing Surveys (CSUR), 23(3):345 -- 405, 1991.
[16] Takao Ohya, Masao Iri, and Kazuo Murota. Improvements of the incremental method for the voronoi
diagram with computational comparison of various algorithms. Journal of the Operations Research
Society of Japan, 27(4):306 -- 337, 1984.
11
|
0910.3883 | 1 | 0910 | 2009-10-20T16:05:58 | Variance Analysis of Randomized Consensus in Switching Directed Networks | [
"cs.MA",
"cs.DC",
"cs.NI"
] | In this paper, we study the asymptotic properties of distributed consensus algorithms over switching directed random networks. More specifically, we focus on consensus algorithms over independent and identically distributed, directed Erdos-Renyi random graphs, where each agent can communicate with any other agent with some exogenously specified probability $p$. While it is well-known that consensus algorithms over Erdos-Renyi random networks result in an asymptotic agreement over the network, an analytical characterization of the distribution of the asymptotic consensus value is still an open question. In this paper, we provide closed-form expressions for the mean and variance of the asymptotic random consensus value, in terms of the size of the network and the probability of communication $p$. We also provide numerical simulations that illustrate our results. | cs.MA | cs | Variance Analysis of Randomized Consensus in
Switching Directed Networks
Victor M. Preciado, Alireza Tahbaz-Salehi, and Ali Jadbabaie
1
9
0
0
2
t
c
O
0
2
]
A
M
.
s
c
[
1
v
3
8
8
3
.
0
1
9
0
:
v
i
X
r
a
Abstract -- In this paper, we study the asymptotic properties of
distributed consensus algorithms over switching directed random
networks. More specifically, we focus on consensus algorithms
over independent and identically distributed, directed Erdos-
R´enyi random graphs, where each agent can communicate with
any other agent with some exogenously specified probability p.
While it is well-known that consensus algorithms over Erdos-
R´enyi random networks result in an asymptotic agreement over
the network, an analytical characterization of the distribution
of the asymptotic consensus value is still an open question. In
this paper, we provide closed-form expressions for the mean and
variance of the asymptotic random consensus value, in terms of
the size of the network and the probability of communication p.
We also provide numerical simulations that illustrate our results.
I. INTRODUCTION
Due to their wide range of applications, distributed con-
sensus algorithms have attracted a significant amount of at-
tention in the past few years. Besides their applications in
distributed and parallel computation [1], distributed control
[2], and robotics [3], they have also been used as models
of opinion dynamics and belief formation in social networks
[4], [5]. The central focus in this vast body of literature is
to study whether a group of agents in a network, with local
communication capabilities can reach a global agreement,
using simple, deterministic information exchange protocols.1
More recently, there has also been some interest in under-
standing the behavior of consensus algorithms in random set-
tings [7] -- [12]. The randomness can be either due to the choice
of a randomized network communication protocol or simply
caused by the potential unpredictability of the environment
in which the distributed consensus algorithm is implemented
[13]. It is recently shown that consensus algorithms over i.i.d.
random networks lead to a global agreement on a possibly ran-
dom value, as long as the network is connected in expectation
[10].
While different aspects of consensus algorithms over ran-
dom switching networks, such as conditions for convergence
This research is supported in parts by the following grants: DARPA/DSO
SToMP, NSF ECS-0347285, ONR MURI N000140810747, and AFOSR:
Complex Networks Program.
Victor Preciado is with Department of Electrical and Systems Engineer-
ing, University of Pennsylvania, Philadelphia, PA 19104 (e-mail: preci-
[email protected])
Alireza Tahbaz-Salehi is with Department of Economics and Department of
Electrical and Systems Engineering, University of Pennsylvania, Philadelphia,
PA 19104 (e-mail: [email protected]).
Ali Jadbabaie is with the General Robotics, Automation, Sensing and
Perception (GRASP) Laboratory, Department of Electrical and Systems
Engineering, University of Pennsylvania, Philadelphia, PA 19104 (e-mail:
[email protected]).
1For a survey on the most recent works in this area see [6].
[7] -- [10] and the speed of convergence [13], have been widely
studied, a characterization of the distribution of the asymptotic
consensus value has attracted little attention. Two notable
exceptions are Boyd et al. [14], who study the asymptotic
behavior of the random consensus value in the special case of
symmetric networks, and Tahbaz-Salehi and Jadbabaie [15],
who compute the mean and variance of the consensus value
for general i.i.d. graph processes. Nevertheless, a complete
characterization of the distribution of the asymptotic value for
general asymmetric random consensus algorithms remains an
open problem.
In this paper, we study asymptotic properties of consensus
algorithms over a general class of switching, directed random
graphs. More specifically, we derive closed-form expressions
for the mean and variance of the asymptotic consensus value,
when the underlying network evolves according to an i.i.d.
directed Erdos-R´enyi random graph process. In our model,
at each time period, a directed communication link is estab-
lished between two agents with some exogenously specified
probability p. It is well-known that due to the connectivity of
the expected graph, consensus algorithms over Erdos-R´enyi
random graphs result
in asymptotic agreement. However,
due to the potential asymmetry in pairwise communications
between different agents, the asymptotic value of consensus
is not guaranteed to be the average of the initial conditions.
Instead, agents will asymptotically agree on some random
value in the convex hull of the initial conditions. Our closed-
form characterization of the variance provides a quantitative
measure of how dispersed the random agreement point is
around the average of the initial conditions in terms of the
fundamentals of the model, namely, the size of the network
and the exogenous probability of communication p.
The rest of the paper is organized as follows. In the
next section, we describe our model of random consensus
algorithms. In Section III, we derive an explicit expression
for the variance of the limiting consensus value over switching
directed Erdos-R´enyi random graphs in terms of the size of the
network n and the communication probability p. Section IV
contains simulations of our results and Section V concludes
the paper.
II. CONSENSUS OVER SWITCHING RANDOM GRAPHS
Consider the discrete-time linear dynamical system
x (k) = Wk x (k − 1) ,
(1)
where k ∈ {1, 2, . . . } is the discrete time index, x(k) ∈ Rn
is the state vector at time k, and {Wk}∞
k=1 is a sequence of
stochastic matrices. We interpret (1) as a distributed scheme
where a collection of agents, labeled 1 through n, update their
state values as a convex combination of the state values of their
neighbors at the previous time step. Given this interpretation,
xi(k) corresponds to the state value of agent i at time k,
and Wk captures the neighborhood relation between different
agents at time k: the ij element of Wk is positive only if agent
i has access to the state of agent j. For the remainder of the
paper, we assume that the weight matrices Wk are randomly
generated by an independent and identically distributed matrix
process.
We say dynamical system (1) reaches consensus asymptot-
ically on some path {Wk}∞
k=1, if along that path, there exists
x∗ ∈ R such that xi(k) → x∗ for all i as k → ∞. We refer
to x∗ as the consensus value. It is well-known that for i.i.d.
random networks dynamical system (1) reaches consensus on
almost all paths if and only if the graph corresponding to the
communications between agents is connected in expectation.
More precisely, Tahbaz-Salehi and Jadbabaie [10] show that
Wk . . . W2W1 −→ 1dT almost surely − where d is some
random vector − if and only if the second largest eigenvalue
modulus of EWk is subunit. Clearly, under such conditions,
dynamical system (1) reaches consensus with probability one
where the consensus value is a random variable equal to
x∗ = dT x(0), where x(0) is the vector of initial conditions.
A complete characterization of the random consensus value
x∗ is an open problem. However, it is possible to compute its
mean and variance in terms of the first two moments of the
i.i.d. weight matrix process. In [15], the authors prove that the
conditional mean of the random consensus value is given by
the random consensus value are given by
Ex∗ = x(0)T v1(EWk),
and its conditional variance is equal to
var(x∗) = [x(0) ⊗ x(0)]T vec(cov(d))
(2)
v1(E [Wk ⊗ Wk]) − [x(0)T v1(EWk)]2
= [x(0) ⊗ x(0)]T
where v1 (·) denotes the normalized left eigenvector corre-
sponding to the unit eigenvalue, and ⊗ denotes the Kronecker
product. In the following, we shall use (2) to derive an explicit
expression for the mean and variance of the consensus value
over a class of switching, directed random graphs..
III. VARIANCE ANALYSIS FOR FINITE ERD OS-R ´ENYI
RANDOM GRAPHS
A. Directed Erdos-R´enyi Random Graphs
We consider directed graphs G = (V, E) with a fixed set of
vertices V = {1, ..., n} and directed edges. A directed edge
from vertex i to vertex j is representes as an ordered pair
(i, j), with i, j ∈ V . In a directed Erdos-R´enyi (ER) graph
G (n, p), the existence of a directed edge (i, j) , with i 6= j, is
determined randomly and independently of other edges with
a fixed probability p ∈ [0, 1]. The adjacency matrix A = [aij]
associated to G (n, p) is a random matrix with all zeros in the
diagonal, and off-diagonal elements aij = 1 with probability
p, and 0 with probability 1 − p, for i 6= j. Similarly, the out-
degree matrix can be defined from the adjacency matrix as
D =diag{di}, where di =Pj aij.
2
In what follows, we associate a sequence of stochastic
k=1 to a sequence of i.i.d random realizations
k=1. To each random graph
matrices {Wk}∞
of directed ER graphs {Gk (n, p)}∞
realization, we associate the following stochastic matrix:
Wk = (Dk + In)−1 (Ak + In) ,
(3)
where Ak and Dk are the adjacency and out-degree matrices
of the graph realization. Notice that adding the identity matrix
to the adjacency in (3) is equivalent to introduce a self-loops
(an edge that starts and ends at the same vertex) over every
single vertex in V . These self-loops serve to avoid singularities
associated with the presence of isolated nodes in Gk (n, p) (for
which di = 0, and Dk is not invertible).
B. Variance of Consensus Value
In this section, we derive an explicit expression for the
variance of the limiting consensus value for a switching
directed random graph. We base our analysis in studying the
terms in (2), i.e., v1 (EWk) and v1 (E [Wk ⊗ Wk]). In order
to compute v1 (EWk), we first compute the expectation of
each entry in Wk. The expectation of the diagonal entries of
Wk are equal to E [1/ (di + 1)], where di is a random variable
representing the degree of the i-th node. In a random ER graph
with probability of link p (and complement q , 1 − p), the
probability density of di is a Bernoulli distribution with n − 1
trials and parameter p, i.e., f (di) ∼ Ber (n − 1, p). Hence,
Ewii = E(cid:20)
1
di + 1(cid:21) =
n−1
Xk=0
1
k + 1(cid:18)n − 1
k (cid:19)pkqn−k−1
, f1 (p, n) ,
(4)
=
1 − qn
np
where we have defined f1 for future convenience. Furthermore,
the off-diagonal elements of EWk are equal to
Ewij = E(cid:20) aij
di + 1(cid:21) = E(cid:20)
aij = 1(cid:21) P (aij = 1) ,
where we have applied the law of total expectation in the last
equality. Moreover, it is straightforward to show that f (di −
1aij = 1) ∼ Ber(n − 2, p); thus,
1
di + 1(cid:12)(cid:12)(cid:12)(cid:12)
Ewij = p
n−2
Xk=0
1
k + 2(cid:18)n − 2
k (cid:19)pkqn−k−2
=
qn + np − 1
np (n − 1)
=
1 − f1 (p, n)
n − 1
.
(5)
Taking (4) and (5) into account, we can write EWk as follows:
EWk = Ewij 1n1T
n + (Ewii − Ewij ) In
=
1 − f1 (p, n)
n − 1
1n1T
n −
1 − n f1 (p, n)
n − 1
In.
It is easy to verify that EWk is irreducible. Therefore, as
discussed in the previous section, consensus algorithms over
i.i.d. directed Erdos-R´enyi random graph process converge to
consensus with probability one. Moreover, it is straightfor-
ward to show that vector 1n satisfies the eigenvalue equation
1T
n
EWk = 1T
n ; thus,
v1 (EWk) =
1
n
1n.
(6)
3
Therefore, as expected, the mean of the random consensus
value is equal to the average of x (0), i.e.,
Ex∗ =
1
n
n
Xi=1
xi (0) , ¯x (0) .
(7)
The other term in the expression of variance that we need to
compute is v1(E [Wk ⊗ Wk]). In order to compute this vector,
we first compute the entries of matrix E [Wk ⊗ Wk], which are
of the form E(wij wrs), with i, j, r and s ranging from 1 to n.
The entries can be classified into six different cases depending
on the relations between the indices. Below, we present the
expressions for each case. Some of the expressions are in terms
of the hypergeometric function 3F2(1, 1, 1−n; 2, 2; p/(p−1)),
which for convenience we denote by H(p, n), defined as the
power series
H(p, n) =
n−1
Xk=0
1
(k + 1)2(cid:18)n − 1
k (cid:19)(cid:18) p
1 − p(cid:19)k
.
Fig. 1. The pattern of E[Wk ⊗ Wk] for n = 3. The numbers in parentheses
represent the value of each entry in terms of expressions Q1 to Q6 defined
in (8). The entries that are different from the corresponding entries of K ⊗ K
are marked with bold lines.
In the following expressions we assume that all four indices
i, j, r and s are distinct. Detailed computations are provided
in the Appendix.
Lemma 1: The left eigenvector of E [Wk ⊗ Wk] corre-
sponding to its unit eigenvalue is given by
Q1 = E(w2
Q2 = E(wiiwjj ) = f 2
Q3 = E(wiiwis) = E(wij wii) = E(w2
ii) = qn−1H(p, n),
1 (p, n),
ij )
=
f1(p, n) − qn−1H(p, n)
n − 1
,
Q4 = E(wiiwri) = E(wiiwrs) = f1(p, n)
1 − f1(p, n)
Q5 = E(wij wis) =
(n − 1)(n − 2)
n − 1
1 + 2qn−1H(p, n) − 3f1(p, n)
,
Q6 = E(wij wji) = E(wij wjs) = E(wij wri)
n − 1 (cid:19)2
= E(wij wrj) = E(wij wrs) =(cid:18) 1 − f1(p, n)
.
As stated earlier, every entry of E[Wk ⊗ Wk] is equal to
one of the expressions provided in (8). The key observation
is the pattern that such classification of entries induces in the
matrix. In order to clarify this point, we illustrate this pattern
for n = 3, where the numbers in parenthesis correspond to
one of the six cases identified above. As the figure suggests,
the entries of E[Wk ⊗Wk] are identical to the entries of K ⊗K
except for those at the rows 1 + r(n + 1) for r = 0, ..., n − 1,
where matrix K ∈ Rn×n is defined as
K =
1 − f1(p, n)
n − 1
1n1T
n +
nf1(p, n) − 1
n − 1
In,
with f1(p, n) as its diagonal, and [1 − f1(p, n)]/(n − 1) as its
off-diagonal entries. In Fig. 1, the entries of E[Wk ⊗ Wk] that
are different from the entries of K ⊗ K are marked with bold
lines.
We now exploit the identified pattern to explicitly compute
the left eigenvector v1(E[Wk ⊗ Wk]).
(8)
v1(E[Wk ⊗ Wk]) =
1
δ "ρ(1n⊗1n) + (1 − ρ)
n
Xi=1
(ei ⊗ ei)#
(9)
where ρ and δ depend on p and n as follows:
ρ(p, n) ,
p(n − 1)
p(n − 2) + 1 − (1 − p)n ,
δ(p, n) , n + n(n − 1)ρ(p, n).
(10)
(11)
Proof: First of all, notice that E[Wk ⊗Wk] is a stochastic
matrix whose entries are all strictly positive for p > 0.
Therefore, it has a unique left eigenvector corresponding to its
unit eigenvalue, which means that v1(E[Wk ⊗ Wk]) is well-
defined. We now show that the pattern of this left eigenvector
is of the form
v = α(1n⊗1n) + (β − α)
(ei ⊗ ei),
(12)
n
Xi=1
for some positive numbers α and β. Notice that all entries
of this vector are equal to α, except for the ones indexed
1 + r(n + 1) for r = 0, ..., n − 1, which are equal to β. To
show that the eigenvector we are looking for is indeed of the
given pattern, we premultiply E[Wk ⊗ Wk] by v, and verify
that
vT E[Wk ⊗Wk] = α′(1 ⊗ 1)T +(β ′ −α′)
(ei ⊗ei)T . (13)
n
Xi=1
where α′ and β ′ are positive numbers, given by
(cid:20)α′
β ′(cid:21) =(cid:20)A B
C D(cid:21)(cid:20)α
β(cid:21)
(14)
with coefficients A, B, C, and D defined as
and therefore,
A = 1 + [nf1(p, n) + n − 2]
1 − f1(p, n)
(n − 1)2
B =
1 − f1(p, n)
n − 1
C = [nf1(p, n) + n − 2]
D = f1(p, n)
1 − f1(p, n)
n − 1
Equation (13) suggests that the pattern of v1 is preserved when
it is multiplied by E[Wk ⊗ Wk]. Therefore, the vector defined
in (12) is the unique left eigenvector of the matrix if there are
positive numbers α = α′ and β = β ′ that satisfy (14).
Due to the fact BC = (1−A)(1−D), the matrix in (14) has
an eigenvalue equal to one, with eigenvector (cid:2)B 1 − A(cid:3)T ,
implying that such (α, β, α′, β ′) exist. Thus, the proposed
vector in (12) is an eigenvector of E[Wk ⊗ Wk] as long as
α = βB/(1 − A) = ρ(p, n)β, which means that
v1(E[Wk ⊗ Wk]) =
1
δ "ρ(1n⊗1n) + (1 − ρ)
(ei ⊗ ei)# .
n
Xi=1
Note that δ(p, n), defined in (11), is a normalizing factor
guaranteeing that the elements of the vector sum up to one.
Now that we have derived explicit expressions for the
eigenvectors (6) and (9), we can compute a closed-form
expression for the variance of the limiting consensus value
in terms of p and n.
Theorem 2: The variance of the asymptotic consensus value
x∗ of the distributed update defined in (1) over switching
Erdos-R´enyi random graphs with parameter p is given by
var(x∗) =
1 − ρ
δ
n
Xi=1
[xi(0) − ¯x(0)]2 ,
(15)
where ρ(p, n) and δ(p, n) are defined in (10) and (11),
respectively.
Proof: First, from (6), we have that
(cid:2)x(0)T v1(EWk)(cid:3)2
= 1
n
n
Xi=1
xi(0)!2
= ¯x(0)2.
On the other hand, from (9), we have that
[x(0) ⊗ x(0)]T
v1 (E[Wk ⊗ Wk]) =
[x(0) ⊗ x(0)]T "ρ(1n⊗1n) + (1 − ρ)
[x(0)T 1n] ⊗ [x(0)T 1n]
(ei ⊗ ei)#
n
Xi=1
=
=
1
δ
ρ
δ
n
+
1 − ρ
δ
Xi=1(cid:0)[x(0)T ei] ⊗ [x(0)T ei](cid:1) ,
where we have used the fact that (A⊗B)(C⊗D) = AC⊗BD.
Since the Kronecker terms in the last expression are scalars,
we have
[x(0)⊗x(0)]T v1(E[Wk⊗Wk]) =
ρ
δ
n2 ¯x(0)2+
1 − ρ
δ
n
Xi=1
x2
i (0)
4
[xi(0)]2.
var(x∗) =(cid:16) ρ
δ
n2 − 1(cid:17) [¯x(0)]2 +
1 − ρ
δ
n
Xi=1
By adding and subtracting n 1−ρ
the variance can be rewritten as
δ [¯x (0)]2, the expression for
var x∗ =
n(n − 1)ρ + n − δ
[¯x(0)]2
+
1 − ρ
δ
δ
n
Xi=1
(xi(0) − ¯x(0))2 .
Since δ = n + n(n − 1)ρ, as defined in (11), the first term in
the right-hand-side of the above expression is equal to zero.
This proves the theorem.
Expression (15) shows that, given the parameters of the
random graph process p and n, the variance of the limiting
consensus value, x∗, is equal to the empirical variance of the
initial conditions multiplied by the factor n(1 − ρ)/δ, which
only depends on parameters p and n.
IV. NUMERICAL SIMULATIONS
In this subsection, we present several simulations that il-
lustrate the result in Theorem 2. In our first simulation, we
compare the analytical expression for the variance in (15)
with the empirical variance obtained from 100 realizations of
the random consensus algorithm for n in a certain range. In
our simulations, we compute the (analytical and empirical)
variances for a range of network sizes while keeping the
expected out-degree of the random graphs fixed to a constant
value c (i.e., the probability communication in the random
graphs is then p = c/n, for all n). In Fig. 1, we plot both the
analytical and empirical variances when the network sizes n
goes from 5 to 50 nodes and the expected degree is fixed to
be c = 5, for all n. The initial conditions for each network
size is given by xi (0) = i/n, for i = 1, ..., n.
Several comments are in order about
the behavior of
var (x∗) in Fig. 1. First, for n = c = 5, we have that
p = c/n = 1 (every link exist); hence, the random graph
is not random, but a complete graph Kn, and the distributed
consensus algorithm converges to the average of the initial
conditions with zero variance (as one can check in Fig. 1).
Second, when n is slightly over c, the variance increases quite
abruptly with n until it reaches a maximum value. For c = 5,
this maximum is achieved for a network size of 9 nodes. The
location of this interesting point, that we denote by n (c), can
be easily computed using Theorem 2. Furthermore, for n >
n(c) the variance slowly decreases with the network size. One
can prove that this variance tends asymptotically to zero as
n → ∞ at a rate 1/n.
Furthermore, according to (15), given a vector of x (0)
initial condition, the variance of the asymptotic consensus
value is equal to the empirical variance of the entries of x (0)
rescaled by the factor n (1 − ρ) /δ (where ρ and δ depend on
the random graph parameters p and n). In Fig. 2, we plot the
5
While different aspects of consensus algorithms over random
switching networks, such as conditions for convergence and
the speed of convergence, have been widely studied, a charac-
terization of the distribution of the asymptotic consensus for
general asymmetric random consensus algorithms remains an
open problem.
In this paper, we have derived closed-form expressions for
the expectation and variance of the asymptotic consensus value
as functions of the number of nodes n and the probability of
existence of a communication link, p. While the expectation
of the distribution of the consensus value is simply the mean
of the initial conditions over the nodes of the network, the
variance presents an interesting structure. In particular, the
variance of this limiting distribution of the consensus value is
equal to the empirical variance of the set of initial conditions
multiplied by a factor that depends on p and n. We have
derived an explicit expression for this factor and check its
validity with numerical simulations.
APPENDIX
KRONECKER MATRIX ENTRIES
The Appendix contains the detailed computation of the
entries of matrices EWk and E[Wk ⊗ Wk]. We start by
computing the elements of EWk. The diagonal entries of EWk
are given by:
Fig. 2.
variance for n in the interval [5:50] and p = 5/n for all n.
Comparison between the empirical variance and the analytical
Fig. 3. Several analytical variance for c ∈ {5, 6, 7, 8, 9, 10}.
=
n−1
1
1
1 + di(cid:21) =
Ewii = E(cid:20)
Xk=0
k + 1(cid:18)n − 1
k (cid:19)pk(1 − p)n−k−1
Xk=0
P(di = k)
k + 1
n−1
=
1
1 − qn
np
, f1(p, n)
values of the factor n (1 − ρ) /δ for a set of expected degrees
c ∈ {5, 6, 7, 8, 9, 10} while the network size n varies from 5
to 70. We observe that the behavior of the factor n (1 − ρ) /δ
is similar for any given c. Again, for n = c the variances are
zero and the variances grow abruptly until a maximum, n (c),
is reached. For large values of n, the variance slowly decays
towards zero at a rate 1/n for all c.
On the other hand, the non-diagonal entries of EWk result in:
Ewij =
1
n − 1
[1 − Ewii] =
np − 1 + qn
np(n − 1)
=
1 − f1(p, n)
n − 1
We now turn to the computation of the elements of E[Wk ⊗
Wk], which are of the form E(wij wrs). In what follows we
assume that the indices i, j, r, and s are distinct. We first start
with elements with in the diagonal subblocks of E[Wk ⊗ Wk]:
V. CONCLUSIONS AND FUTURE WORK
We have studied the asymptotic properties of the consensus
value in distributed consensus algorithms over switching, di-
rected random graphs. Due to the connectivity of the expected
graph, consensus algorithms over Erdos-R´enyi random graphs
result in asymptotic agreement. However, the asymptotic value
of consensus is not guaranteed to be the average of the initial
conditions. Instead, agents will asymptotically agree on some
random value in the convex hull of the initial conditions.
Ew2
ii = E(cid:20)
Xk=0
n−1
=
1
(di + 1)2(cid:21)
(k + 1)2(cid:18)n − 1
1
= qn−1H(p, n)
k (cid:19)pkqn−k−1
For the rest of entries in the diagonal blocks, it is useful to
note that Ewii = f1(p, n) and Ewij = 1−f1(p,n)
, as proved
n−1
6
[7] Y. Hatano and M. Mesbahi, "Agreement over random networks," IEEE
Transactions on Automatic Control, vol. 50, no. 11, pp. 1867-1872, 2005.
[8] C. W. Wu, "Synchronization and convergence of linear dynamics in
random directed networks," IEEE Transactions on Automatic Control,
vol. 51, no. 7, pp. 1207-1210, Jul. 2006.
[9] M. Porfiri and D. J. Stilwell, "Consensus seeking over random weighted
directed graphs," IEEE Transactions on Automatic Control, vol. 52, no.
9, pp. 1767-1773, sep 2007.
[10] A. Tahbaz-Salehi and A. Jadbabaie, "A necessary and sufficient con-
dition for consensus over random networks," IEEE Transactions on
Automatic Control, vol. 53, no. 3, pp. 791-795, Apr 2008.
[11] G. Picci and T. Taylor, "Almost sure convergence of random gossip
algorithms," in Proceedings of the 46th IEEE Conference on Decision
and Control, New Orleans, LA, dec 2007, pp. 282-287.
[12] D. Acemoglu, A. Ozdaglar, and A. ParandehGheibi, "Spread of
(mis)information in social networks," Massachusetts Institute of Tech-
nology, LIDS Report 2812, May 2009, submitted for Publication.
[13] F. Fagnani and S. Zampieri, "Randomized consensus algorithms over
large scale networks," IEEE Journal on Selected Areas in Communica-
tions, vol. 26, no. 4, pp. 634-649, May 2008.
[14] S. Boyd, A. Gosh, B. Prabhakar, and D. Shah, "Randomized gossip
algorithms," Special issue of IEEE Transactions on Information Theory
and IEEE/ACM Transactions on Networking, vol. 52, no. 6, pp. 2508-
2530, Jun. 2006.
[15] A. Tahbaz-Salehi and A. Jadbabaie, "Consensus over ergodic station-
ary graph processes," IEEE Transactions on Automatic Control, 2009,
Accepted for Publication.
in Subsection III-B:
E(wiiwrr) = Ewii Ewrr = f 2
1 (p, n)
E(wiiwri) = Ewii Ewri = f1(p, n)
E(wiiwrs) = Ewii Ewrs = f1(p, n)
1 − f1(p, n)
n − 1
1 − f1(p, n)
n − 1
E(wiiwis) =
Ewii − Ew2
ii
n − 1
=
f1(p, n) − qn−1H(p, n)
n − 1
Similarly, we have the following results for the off-diagonal
subblocks:
n − 1 (cid:19)2
E(wij wji) = Ewij Ewji =(cid:18) 1 − f1(p, n)
n − 1 (cid:19)2
E(wij wjs) = Ewij Ewjs =(cid:18) 1 − f1(p, n)
n − 1 (cid:19)2
E(wij wri) = Ewij Ewri =(cid:18) 1 − f1(p, n)
n − 1 (cid:19)2
E(wij wrj) = Ewij Ewrj =(cid:18) 1 − f1(p, n)
n − 1 (cid:19)2
E(wij wrs) = Ewij Ewrs =(cid:18) 1 − f1(p, n)
E(wij wii) = E(wiiwis) =
f1(p, n) − qn−1H(p, n)
n − 1
E(w2
ij ) = E"
a2
ij
(di + 1)2# = E(cid:20)
aij
(di + 1)2(cid:21)
n − 1
= E(wiiwij ) =
f1(p, n) − qn−1H(p, n)
E(wij wis) =
=
Ewij − Ew2
ij − E(wij wii)
n − 2
1 + 2qn−1H(p, n) − 3f1(p, n)
(n − 1)(n − 2)
REFERENCES
[1] J. N. Tsitsiklis, Problems in decentralized decision making and com-
putation, Ph.D. dissertation, Massachusetts Institute of Technology,
Cambridge, MA, 1984.
[2] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile
autonomous agents using nearest neighbor rules," IEEE Transactions on
Automatic Control, vol. 48, no. 6, pp. 988-1001, 2003.
[3] J. Cortes, S. Martinez, and F. Bullo, "Analysis and design tools for dis-
tributed motion coordination," in Proceedings of the American Control
Conference, Portland, OR, Jun. 2005, pp. 1680-1685.
[4] M. H. DeGroot, "Reaching a Consensus," Journal of American Statisti-
cal Association, vol. 69, no. 345, pp. 118-121, Mar. 1974.
[5] B. Golub and M. O. Jackson, "Naive learning in social networks:
Convergence, influence, and the wisdom of crowds," 2007, unpublished
Manuscript.
[6] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooper-
ation in networked multi-agent systems," Proceedings of the IEEE, vol.
95, no. 1, Jan. 2007.
|
1511.00053 | 1 | 1511 | 2015-10-31T01:00:35 | CarPed -- A Hybrid and Macroscopic Traffic and Pedestrian Simulator | [
"cs.MA"
] | Dense human flow has been a concern for the safety of public events for a long time. Macroscopic pedestrian models, which are mainly based on fluid dynamics, are often used to simulate huge crowds due to their low computational costs (Columbo & Rosini 2005). Similar approaches are used in the field of traffic simulations (Lighthill & Whitham 1955). A combined macroscopic simulation of vehicles and pedestrians is extremely helpful for all-encompassing traffic control. Therefore, we developed a hybrid model that contains networks for vehicular traffic and human flow. This comprehensive model supports concurrent multi-modal simulations of traffic and pedestrians. | cs.MA | cs | 1
CarPed – A hybrid and macroscopic traffic and pe-
destrian simulator
Daniel H. Biedermann1 und Peter M. Kielar1 und Quirin Aumann1 und Carlos M. Osorio1
und Celeste T. W. Lai1
1Lehrstuhl für Computergestützte Modellierung und Simulation · Technische Universität München ·
Arcisstrasse 21 · 80333 München
[email protected]
Zusammenfassung
Dichte Personenströme sind ein bekanntes Gefährdungsrisiko für öffentliche Grossveran-
staltungen. Makroskopische Personenstrommodelle, die häufig auf fluid-dynamischen Be-
rechnungen basieren, können mit geringem Rechenaufwand simuliert werden. Aus diesem
Grund werden sie oft für die Simulation grosser Menschenmengen verwendet (COLUMBO &
ROSINI 2005). Ähnliche Ansätze existieren im Forschungsbereich der Verkehrssimulation
(LIGHTHILL & WHITHAM 1955). Eine gekoppelte makroskopische Simulation von Fahrzeu-
gen und Fussgängern ist für eine allumfassende Verkehrsüberwachung und Planung sehr
hilfreich. Aus diesem Grund haben wir ein Hybridmodell entwickelt, das Netzwerke für
Fahrzeuge und Fussgänger bereitstellt. Dieses gekoppelte Modell unterstützt die gleichzeiti-
ge Simulation von Fahrzeug- und Personenverkehr.
Abstract
Dense human flow has been a concern for the safety of public events for a long time. Mac-
roscopic pedestrian models, which are mainly based on fluid dynamics, are often used to
simulate huge crowds due to their low computational costs (COLUMBO & ROSINI 2005).
Similar approaches are used in the field of traffic simulations (LIGHTHILL & WHITHAM
1955). A combined macroscopic simulation of vehicles and pedestrians is extremely helpful
for all-encompassing traffic control. Therefore, we developed a hybrid model that contains
networks for vehicular traffic and human flow. This comprehensive model supports concur-
rent multi-modal simulations of traffic and pedestrians.
1
Introduction
1.1 Motivation
Every year, about 2000 people die due to the movement of large human crowds (HUGHES
2003). The growing eventisation of social and public life in the 21st century (HITZLER
2010) increases the probability and frequency of organized events attended by thousands or
millions of people. In the past, fatal accidents have occurred repeatedly at such large events.
Examples include the catastrophe in the Heysel stadium in 1985 (LEWIS 1989) with 39
fatalities or the Loveparade disaster in 2010 with 21 deaths (HITZLER ET AL. 2011). Profes-
sional crowd control increases the safety of public events and lowers the probability of fatal
2
Daniel H. Biedermann, Peter M. Kielar, Quirin Aumann, Carlos M. Osorio, Celeste T.W. Lai
accidents in the context of large human flows. Pedestrian dynamics simulations support
successful crowd management by helping organizers of public events to foresee and pre-
vent dangerous situations. Macroscopic pedestrian dynamic models like the Hartmann-
Sivers model (HARTMANN & SIVERS 2013) are most suitable for simulating large numbers
of people (BIEDERMANN ET AL. 2014A). The same macroscopic model can be applied to
simulate the motion of cars. By combining human and vehicular flow into one model, we
can achieve an all-embracing simulation model, which covers crowd and traffic control at
the same time. Due to the macroscopic nature of these simulation models, we can simulate
very large and dense scenarios faster than real time.
1.2 Current Hybrid Modeling
Two different kinds of hybrid modeling exist in the field of traffic and crowd control. The
first type combines traffic or pedestrian dynamcis models at different scales in order to
reduce the overall computational effort. This means that simulations in the hazardous areas
of the scenario (places with high densities or small bottlenecks) are calculated with very
detailed but computationally costly models. At the same time, less hazardous areas of the
scenario can be simulated by less detailed but more cost-effective models. This provides us
precious results in the interesting parts of the scenario with a reduced total computational
cost. TOLBA ET AL. (2005) as well as MCCRE & MOUTARI (2010) developed such hybrid
models in the field of vehicular flow. Similar models were developed for pedestrian dynam-
ics. A wide range of hybrid modeling exists; from the coupling of very specific pedestrian
simulation models (e.g. NGUYEN ET AL. 2012)
to generic coupling frameworks
(BIEDERMANN ET AL. 2014B). An overview of the current state of the art can be found by
IJAZ ET AL. (2015).
The second type of hybrid modeling does not combine different scales, but different kinds
of simulations. GALEA ET AL. (2008) describe the coupling of pedestrian dynamics and fire
simulations and GÖTTLICH ET AL. (2011) combine an evacuation simulation with the spread
of hazardous gases. Our approach is the connection of vehicular traffic with human flow.
Some studies have already been carried out in this field. PRETTO ET AL. (2011) coupled
agent-based representations of pedestrians and cars. A macroscopic approach was devel-
oped by BORSCHE ET AL. (2014). They used the classic Lighthill-Whitham-Richards model
(LIGHTHILL & WHITHAM 1955; RICHARDS 1956) for the vehicular flow and the model from
HUGHES (2002) for the macroscopic simulation of pedestrians. Their work is important
progress in the field of hybrid macroscopic modeling. However, they considered vehicles
and humans as two interacting, but separate flows. In reality, cars and pedestrians are not
completely separate: drivers exit their cars after finishing driving and become pedestrians
and pedestrians may enter their vehicles to become drivers. Our approach is a hybrid mod-
el, which is capable of converting humans and vehicles into each other to create a realistic
simulation model.
2 The hybrid simulation model CarPed
2.1 The Hybrid approach
It is important to use a single network that combines pedestrian and traffic simulations to
enhance the safety of traveling to public events. Therefore, we introduce an interface which
CarPed – A hybrid and macroscopic traffic and pedestrian simulator
3
links human and vehicular flow. The movement behavior of cars and pedestrians is based
on the macroscopic Hartmann and Sivers model (HARTMANN & SIVERS 2013) and is calcu-
lated on a discrete network. Roads for cars and walking paths for pedestrians are represent-
ed by the edges of this network. Its nodes represent the links between the different edges
and can be considered crossings. Specifically, the nodes have different flow rates. There are
special nodes, which represent parking lots, and serve as a connection between roads and
walkways. Visitors of the event enter and leave their cars there. To put it another way, the
simulated subjects get transformed from vehicles into pedestrians and vice versa. Therefore,
the parking lots serve as a connection of the traffic and pedestrian networks.
In Figure 1, an exemplary network is shown. The darker nodes and edges represent streets
and their respective intersections, whereas the pedestrians’ walkways and intersections are
represented in a lighter gray. The transformation of cars and pedestrians is carried out at
special parking lot nodes. Green numbered entry nodes represent sources, which are the
starting points of the visitors (e.g. their hometown). The red marked exit node is the final
destination of the visitors (e.g. a large public event), where they are deleted from the net-
work.
Figure 1: An exemplary network consisting of nodes and edges
We implemented the Hartmann and Sivers model for pedestrian simulation and extended it
to traffic simulation by using different model parameters. The extension was done by ap-
plying different maximum capacities to the edges and nodes, using different free flow ve-
locities and other model constants for the simulated subjects. We used data obtained from
WEIDMANN (1993) for the pedestrians and data from WACHS ET AL. (2000) for the vehicle
simulation. Additionally, we conducted a large field study to receive sufficient data for the
transformation process (see Section 3).
2.2 The Hartmann-and-Sivers Model
The Hartmann and Sivers model adopts a structured continuum model on a macroscopic
level. The approach basically relies on fundamental diagrams – the relation between fluxes
and local densities – as well as the explicit consideration of individual velocities; thus
showing better agreement with microscopic pedestrian models (HARTMANN AND SIVERS
2013). The free flow velocity vff is an individual property of each pedestrian / vehicle, i.e. a
macroscopic, time independent state variable attached to each individual. The free flow
velocity describes the highest velocity an individual can reach under optimal circumstances.
4
Daniel H. Biedermann, Peter M. Kielar, Quirin Aumann, Carlos M. Osorio, Celeste T.W. Lai
This property is distributed normally among pedestrians or cars. Thus for modelling human
or vehicular flow, the following equation holds
;
tx
),
(
∂
ρ
v
ff
t
∂
(
vv
(
+
∂
x
∂
;
ρ
tx
,(
))
(
ρ
v
;
tx
),
ff
ff
)
=
vf
(
ff
;
tx
),
The total density of subjects is given by the sum of densities for all free flow velocities.
ρ
tx
),(
max
ffv
∫=
0
;(
νρ
dtx
),
ν
The fundamental relation between velocity and density can be adopted according to:
(1)
(2)
(3)
;
ff
)))
1(
−
−
/1
)
ρ
vv
(
v
= ff
exp(
/1(
maxρργ
The maximum density 𝜌𝑚𝑚𝑚 and the factor 𝛾 are unique parameters for cars and pedestri-
ans. The solvation of the differential equations for pedestrians and cars was done by an
upwind-downwind finite volume scheme (HARTMANN & SIVERS 2013).
2.3 Routing Behavior of Pedestrian and Cars
In their original paper, Hartmann and Sivers used fixed probability values to distribute the
pedestrians at each intersection. These values were obtained by comparing the capacities of
the adjacent edges and did not take the current density of those edges into account. It is
possible that the majority of a flow gets directed into an edge which is already relatively
full, and the required time for the flow to pass the system could increase in an unrealistic
way. Therefore, we extend this approach and introduce a routing algorithm which deter-
mines the possible edges the pedestrians or cars will be allocated at.
To find the fastest path through the system, we use the classical Dijkstra’s algorithm
(DIJKSTRA 1959). This algorithm weighs all edges of the graph to compute the shortest way
through the system. We used the Dijkstra’s algorithm to calculate the route with the shortest
factors:
weight =
length
( ρffvv
)
;
travel time. Therefore, we used the length and the current velocity 𝑣(𝑣𝑓𝑓;𝜌) as weighting
With increasing density the velocity 𝑣(𝑣𝑓𝑓;𝜌) decreases according to Equation (2). If the
density of an edge reaches the maximum density 𝜌𝑚𝑚𝑚, the edge is considered as closed.
Closed edges are ignored for the calculation of the routing algorithm.
2.4 Transformation between the two simulation models
Cars get transformed into pedestrians and vice versa in the parking lot nodes. Typically, for
ordinary cars, one to five people get out of each vehicle, with a prescribed random distribu-
tion. The distribution for the transformation was gathered from field data (see Section 3). It
is also possible to transform incoming pedestrians into cars, using the inverse process. The
transformed pedestrians and cars receive their individual parameters, like their free flow
(4)
CarPed – A hybrid and macroscopic traffic and pedestrian simulator
5
velocity or their maximum density, according to WEIDMANN (1993) respectively WACHS ET
AL. (2000).
3 Visitor distribution of incoming cars
Figure 2: The distribution according to the total number of cars counted and the distribu-
tion according to the total number of passengers.
Pedestrians get transformed into cars and vice versa on special parking nodes in the graph
based scenario. To model a realistic transformation of pedestrians and cars, it is necessary
to know the distribution of passengers per car. This means, that we have to know how many
visitors are normally transported by one car. Therefore, we conducted an extensive field
study to obtain the lacking data. We studied a large music festival in Munich over three
consecutive days and counted the amount of visitors per car. We surveyed a total number of
1960 cars (for details see Table 1), which carried an average of 2.21 persons. Over 70% of
the vehicles were occupied by one or two passengers. A negligible amount of cars trans-
ported more than five people, and the maximum observed was eight passengers per car. The
observed distributions from all three days can be seen in Figure 2. This data was used as a
configuration for the transformation process of the hybrid model.
Date
Visitors
29.05.2015
30.05.2015
31.05.2015
1
144
98
210
2
190
263
526
3
53
62
158
4
33
54
98
5
11
15
36
62
Table 1: Number of visitors per car on the music festival “Rockavaria”
otal
979
452
273
185
>5
4
3
2
9
Total
435
495
1030
1960
4 The CarPed-Toolbox
4.1 Structure and usability
6
Daniel H. Biedermann, Peter M. Kielar, Quirin Aumann, Carlos M. Osorio, Celeste T.W. Lai
In cooperation with professional and experienced event managers, the software toolbox
CarPed was developed as a proof of concept. Great attention was payed towards usability.
Therefore, we implemented an intuitive and easy to use graphical user interface (GUI).
The GUI is divided into three major parts, namely the network input, the model and solving
options and the result window. The network, consisting of nodes and edges, can be easily
entered through the network input. All network depending properties, like the length or the
capacity of an edge, can be determined immediately after the input procedure. By using the
model and solving options window, the user can determine which computational method
should be used and how the data should be processed. Finally, the computed results can be
accessed and visualized in the viewer. As all computed data is stored, the density distribu-
tion for each time-step can be looked at by the user. Different coloring of the edges shows
the current densities. Alternatively, the user can click on an edge or node to get more de-
tailed information about this object like the amount of pedestrians or cars on it. An exem-
plary screenshot of the GUI can be seen in Figure 3.
Figure 3: Representation of an exemplary simulation scenario in CarPed
4.2 Workflow
The network can be entered by clicking on the designated area in the GUI. Special nodes,
like entry and exit nodes or parking lots, can be defined by the user. Additionally, the nodes
can be linked to incoming or outgoing edges. If the input to the system is finished, a check-
ing routine searches for errors and discrepancies. For each algorithm, the user can individu-
ally modify the settings. After everything is set up, the solver begins its calculations.
At the beginning of the simulation, the network model reads the information of nodes and
edges from the database previously created. It constructs a network from the infrastructure
and passes the entire network to the numerical solver. Together with the received input
data, the numerical solver starts generating vehicles and pedestrians in the associated entry
nodes. It determines the initial density for each edge that is connected to the entry node and
computes the density distribution. For each time-step, the results are passed from the nu-
merical solver and stored in the network model. If a subject enters a node, it gets removed
CarPed – A hybrid and macroscopic traffic and pedestrian simulator
7
from this incoming edge and is put into an outgoing edge according to the route calculated
by the Dijkstra algorithm. Additionally, the subject gets transformed if the visited node is a
parking lot. The computations are repeated until all pedestrians and cars have reached an
exit node and are dismissed from the system.
During the result analysis, the GUI displays two visions: a parameter set and an animation
of density distributions. All statistical information is provided by the toolbox and can be
easily accessed.
5 Conclusion and Outlook
This paper describes the dedicated hybrid and macroscopic simulation approach CarPed.
This approach uses the macroscopic Hartmann and Sivers model on a network of nodes and
edges. It is able to simulate human and vehicular flow in one hybrid model. The transfor-
mation between pedestrians and cars is carried out at special nodes, which represent park-
ing lots.
The tool is thought to play a fundamental rule in the planning phase of massive public
events. Organizers can use it to identify pathways and roads, which have a higher risk of
congestions. Therefore, the organizers can develop strategies to avoid dangerous situations.
In future studies, the current approach will be extended to a hybrid model for all-
encompassing traffic control. This extended hybrid model will include more traffic sub-
jects, like trains or ships to receive a global and universal traffic simulator.
Additionally, the CarPed simulator needs to be validated by field data. Therefore, we ob-
served a public music festival over two consecutive years (BIEDERMANN ET AL. 2015). The
first results are promising, but need further investigations.
Acknowledgements
We would like to thank Isabella von Sivers for her support and long conversations accord-
ing to this topic. Additionally, we would like to thank Andrea Mayer, Andreas Riedl and
our other student assistants for their help to collect sufficient data.
This work is supported by the Federal Ministry for Education and Research (Bundesminis-
terium für Bildung und Forschung, BMBF), project MultikOSi, under grant FKZ 3N12823.
References
BIEDERMANN, D. H., KIELAR, P. M., HANDEL, O. (2014A) Betrachtung der Modellierung
und Simulation von Fussgängern im Kontext verschiedener Skalen. In: Proc. of the 26th
Forum Bauinformatik, 171-179
BIEDERMANN, D. H., KIELAR, P. M., HANDEL, O., BORRMANN, A. (2014B) Towards Transi-
TUM: A Generic Framework for Multiscale Coupling of Pedestrian Simulation Models
based on Transition Zones. In: Transportation Research Procedia 2, 495-500
8
Daniel H. Biedermann, Peter M. Kielar, Quirin Aumann, Carlos M. Osorio, Celeste T.W. Lai
BIEDERMANN, D. H., DIETRICH, F., HANDEL, O., KIELAR, P. M., SEITZ, M. (2015) Using
Raspberry Pi for scientific video observation of pedestrians during a music festival.
Technische Universität München
BORSCHE, R., KLAR, A., KÜHN, S., MEURER, A. (2014). Coupling traffic flow networks to
pedestrian motion. In: Mathematical Models and Methods in Applied Sciences 24(02),
359-380
COLOMBO, R. & ROSINI, M. (2005) Pedestrian flows and non-classical shocks. In: Mathe-
matical Methods in the Applied Sciences, 28(13), 1–17
DIJKSTRA, E. W. (1959), A note on two problems in connexion with graphs. In: Numerische
mathematik, 1(1), 269-271.
GALEA, E. R., WANG, Z., VEERASWAMY, A., JIA, F., LAWRENCE P. J., EWER, J. (2008) Cou-
pled Fire/Evacuation Analysis of the Station Nightclub Fire. In: Fire Safety Science 9,
465-476
GÖTTLICH, S., KÜHN, S., OHST, J. P., RUZKA S., THIEMANN M. (2011) Evacuation dynamics
influenced by spreading hazardous material. In: Networks and heterogeneous media,
443-464
HARTMANN, D. & VON SIVERS, I. (2013) Structured first order conservation models for
pedestrian dynamics. In: Networks and Heterogeneous Media 8 (4), 985–1007.
HITZLER, R. (2010), Eventisierung: drei Fallstudien zum marketingstrategischen Massen-
spass. Springer-Verlag
261–277. In: Springer.
HITZLER, R., KIRCHNER, B., BETZ, G. (2011) Das Beispiel Loveparade. Urbane Events,
HUGHES, R. L. (2002) A continuum theory for the flow of pedestrians. In: Transportational
Research Part B: Methodological 36 (2002), 507-535
HUGHES, R. L. (2003) The flow of human crowds. In: Annual review of fluid mechanics 35
(1), 169–182.
Psychology
IJAZ, K., SOHAIL, S., HASHISH, S. (2015) A Survey of Latest Approaches for Crowd Simula-
tion and Modeling using Hybrid Techniques
LEWIS, J. (1989), A value-added analysis of the heysel stadium soccer riot. In: Current
LIGHTHILL, M. J. & WHITHAM, G. B. (1955) On Kinematic Waves. II. A Theory of Traffic
Flow on Long Crowded Roads. In: Proceedings of the Royal Society A: Mathematical,
Physical and Engineering Sciences 229 (1178), 317–345.
MCCREA, J. & MOUTARI, S. (2010), A hybrid macroscopic-based model for traffic flow in
road networks. In: European Journal of Operational Research, 676-684
NGUYEN, T. N. A., ZUCKER, J. D., NGUYEN H. D., ALEXIS D., VO D. A. (2012) Hybrid equa-
tion-based and agent-based modeling of crowd evacuation on road network. In. Proc. of
the 8th ICCS conference, 558-570
PRETTO, C. O., CYBIS, H. B. B., JACOBSEN, A. (2011) A Multi-layer Simulation Model For
Vehicle And Pedestrian Interaction. In: Transportation Research Board 90th Annual
Meeting. 2011.
RICHARDS, P. (1956) Shock waves on the highway. In: Operations research 4 (1), 42–51.
TOLBA, C., LEFEBVRE, D., THOMAS, P., EL MOUDNI, A. (2005), Continuous and timed Petri
nets for the macroscopic and microscopic traffic flow modelling. In: Simulation Model-
ing Practice and Theory 13, 407-436
WACHS, M. ET AL. (2000) Highway capacity manual. In: Transportation Research Board
WEIDMANN, U. (1993), Transporttechnik der Fussgänger. In: Schriftenreihe des IVT Nr. 90
|
1902.03185 | 4 | 1902 | 2019-11-28T14:59:18 | Partner Selection for the Emergence of Cooperation in Multi-Agent Systems Using Reinforcement Learning | [
"cs.MA",
"cs.AI"
] | Social dilemmas have been widely studied to explain how humans are able to cooperate in society. Considerable effort has been invested in designing artificial agents for social dilemmas that incorporate explicit agent motivations that are chosen to favor coordinated or cooperative responses. The prevalence of this general approach points towards the importance of achieving an understanding of both an agent's internal design and external environment dynamics that facilitate cooperative behavior. In this paper, we investigate how partner selection can promote cooperative behavior between agents who are trained to maximize a purely selfish objective function. Our experiments reveal that agents trained with this dynamic learn a strategy that retaliates against defectors while promoting cooperation with other agents resulting in a prosocial society. | cs.MA | cs | Partner Selection for the Emergence of Cooperation in Multi-Agent Systems
Using Reinforcement Learning
Nicolas Anastassacos,12 Stephen Hailes,2 Mirco Musolesi123
1The Alan Turing Institute 2University College London
3University of Bologna
9
1
0
2
v
o
N
8
2
]
A
M
.
s
c
[
4
v
5
8
1
3
0
.
2
0
9
1
:
v
i
X
r
a
Abstract
Social dilemmas have been widely studied to explain how hu-
mans are able to cooperate in society. Considerable effort has
been invested in designing artificial agents for social dilem-
mas that incorporate explicit agent motivations that are cho-
sen to favor coordinated or cooperative responses. The preva-
lence of this general approach points towards the importance
of achieving an understanding of both an agent's internal de-
sign and external environment dynamics that facilitate coop-
erative behavior. In this paper, we investigate how partner
selection can promote cooperative behavior between agents
who are trained to maximize a purely selfish objective func-
tion. Our experiments reveal that agents trained with this dy-
namic learn a strategy that retaliates against defectors while
promoting cooperation with other agents resulting in a proso-
cial society.
Introduction
Human history provides many examples of people be-
having more cooperatively than is typically assumed
in behavior models featuring assumptions of rationality
and self-interest. Examples include behaving environmen-
tally responsibly (littering, recycling, etc.) (Cialdini 2003;
Thøgersen 2006) or citizens honestly paying their taxes
when there are financial incentives to do otherwise (Posner
2000). These types of situations are known as social dilem-
mas and are characterized by a trade-off between an individ-
ual's short-term rewards and the collective long-term inter-
ests of the group, community or society as a whole (Dawes
1980). Understanding the mechanisms for the emergence of
cooperation is still an open problem in several disciplines
ranging from evolutionary biology and economics to au-
tonomous systems.
Recently, Reinforcement Learning (RL) has been ap-
plied to the study of general-sum multi-agent games like
social dilemmas (Hughes et al. 2018; Jaques et al. 2019;
Leibo et al. 2017; Perolat et al. 2017), however, it is known
that it is difficult to obtain optimal results due to the non-
stationarity caused by agents learning simultaneously (Bu-
soniu, Babuska, and De Schutter 2008). Solutions to these
problems can involve introducing techniques like modeling
Accepted for Publication in the Proceedings of the Thirty-Fourth
AAAI Conference on Artificial Intelligence (AAAI 2020). Re-
search Track. New York, New York, USA.
opponents or using joint-action distributions over the pop-
ulation (Lowe et al. 2017) though these techniques tend to
be affected by scalability issues. The traditional RL objec-
tive describes a selfish agent that looks to maximize its own
reward. In decentralized settings, it may not be possible to
impose any cooperative reward function on all the agents.
In these cases, it is important for the environment to be
designed in a way that allows even selfish agents to learn
behaviors which do not severely hinder other agents and,
preferably, are beneficial for the society as a whole. We ar-
gue that an answer to training RL agents in decentralized
multi-agent scenarios potentially lies in understanding the
societal dynamics that produces norm-inducing behaviors
and their effects on how and what agents learn.
We investigate a key characteristic of social interaction:
partner selection. The capability for an individual to freely
choose who they want to interact with has been thought
to have a prominent role in determining the structure of a
population and the competitive and collaborative relation-
ships that form between members of society (Santos, Santos,
and Pacheco 2008). In turn, these relationships can cause
a change in the strategies that agents learn that may go on
to impact partner selection in the future. The development
is dynamic and cyclical and has been hypothesized to be
a driving factor in the emergence of cooperative societies
and, potentially, a catalyst for altruistic behavior (Barclay
and Willer 2006; Cuesta et al. 2015).
Reputation and signaling are directly tied to the notion
of partner selection. Individuals prefer cooperative partners
and are inclined to avoid partnerships with known selfish
individuals (Albert et al. 2007; Milinski, Semmann, and
Krambeck 2002). This tendency towards preferring to inter-
act with reputable individuals suggests that the freedom to
choose may lead agents to develop strategies that maximize
the reward, while also improving reputation or signaling in
order to attract the attention of others. Signaling a desire to
coordinate or cooperate, however, may come at the cost of
becoming a target for exploitation by others. The interplay of
these dynamics has made partner selection a key component
of group formation and group inequality (Bolton and Ocken-
fels 2000; Mohtashemi and Mui 2003). Fu et al. developed a
reputation-based partner switching model that predicts a co-
operative response using agents (Fu et al. 2008) and the re-
peated Prisoner's Dilemma; using result from evolutionary
game theory they derive rules to facilitate partner switching
in order to penalize agents with a low reputation and a net-
work framework that determines the amount of information
that is available to each agent. Motivated by these findings,
we propose a model of partner selection that is character-
ized by two components: (1) N-agents are fairly presented
with only the previous interactions of their opponents and
use this information to select a partner, and then (2) each pair
of agents engages in a Dilemma game and learns from ex-
perience. Using this framework and RL agents trained with
Q-learning (Watkins and Dayan 1992), we are able to sim-
ulate the emergence of cooperation in a decentralized soci-
ety using partner selection and agents that learn according
to a selfish objective function. When we contrast our results
with the outcomes of scenarios in which agents are randomly
paired, the impact of partner selection can be clearly seen on
both the overall societal outcome and on the evolution of the
strategies that are played throughout the simulation.
We show that when selfish agents have the option to
choose a partner they are able to learn to cooperate with
"good" individuals and punish "bad" ones, a strategy of
equivalent retaliation known as Tit-for-Tat (TFT) (Axelrod
and Hamilton 1981). In our results, we can see how TFT
agents act to sustain cooperation in a society by playing a
strategy that encourages other agents to play the same strat-
egy akin to a social norm. Q-agents are capable of learn-
ing this strategy in the presence of partner selection and
maintain cooperation in the society, whereas agents trained
with random matching learn to always defect. The contri-
butions of this work are two-fold. Firstly, we demonstrate
the importance of building environments with societal dy-
namics like partner selection, such that agents can learn
norm-inducing behaviors in decentralized multi-agent set-
tings. Secondly, we provide novel insight into how cooper-
ation can emerge between selfish agents that can be con-
trasted with outcomes seen in evolutionary biology and be-
havioral economics as we use familiar framework in the
repeated Prisoner's Dilemma and a bottom-up approach to
partner selection that results in a well-known strategy, TFT.
For this reason, we believe that the results of this work have
implications not only for the design of cooperative artificial
agents but also for our understanding behavior in animal and
human societies.
Social dilemmas provide a powerful platform to study the
emergence of behavior and the development of coopera-
tive solutions using agents (Izquierdo, Izquierdo, and Gotts
2008; Macy and Flache 2002; Sen and Airiau 2007).
They have featured heavily in behavioral economics, psy-
chology and evolutionary biology (Axelrod and Hamilton
1981; Mohtashemi and Mui 2003; Fehr and Fischbacher
2004; Fu et al. 2008; Santos, Santos, and Pacheco 2008).
Since then, other research has looked to explain cooperative
outcomes in more complex environments termed sequential
social dilemmas and investigate innate human-inspired char-
acteristics such as a preference for fairness, pro-sociality,
or social influence while leveraging the capabilities of RL
agents to learn (Hughes et al. 2018; Jaques et al. 2019;
Social Dilemmas
by
SD
C
D
C
D
R, R S, T
T, S
P, P
rPD
C
D
C
3, 3
4, 0
D
0, 4
1, 1
Table 1: Payoff Matrix for Social Dilemmas and repeated
Prisoner's Dilemma. The motivation to defect comes from
fear of an opponent defecting or acting greedily to gain the
maximum reward when one anticipates the opponent might
cooperate. The game is modeled so that T > R > P > S
and 2R > T + S. From a game theoretic perspective, the
optimal strategy in a finite game is to defect. This is undesir-
able as the agents could both achieve greater reward if they
agreed to cooperate.
Leibo et al. 2017; Perolat et al. 2017; Peysakhovich and
Lerer 2018).
We consider a classic repeated Prisoner's Dilemma (PD)
approach, where each iteration of the game can be character-
ized by a payoff matrix. Agents play multiple rounds of the
Prisoner's Dilemma and there is continuity between games
as an agent's previous actions are used to inform the part-
ner selection of the other agents. The resulting dilemma is
straightforward: agents must decide whether they should (1)
maximize their immediate reward by defecting while broad-
casting information to other agents that may deter them from
future interactions or (2) cooperate, forgoing the immediate
reward and risk being exploited, but potentially attract the
attention of other agents to increase future returns. For clar-
ity, we will name the outcomes in Table 1 as follows: (C,
C) is mutual cooperation, (D, C) is exploitation, (C, D) is
deception, and (D, D) is mutual defection.
Q-Learning
RL algorithms learn a policy from experience balancing ex-
ploration of the environment and exploitation.
We train our agents with Q-learning (Watkins and Dayan
1992). We train agents independently using this algorithm in
a multi-agent setting. In Q-learning, the policy of agent i is
represented through a state-action value function Qi(s, a).
The i-th agent stores a function Qi : Si × Ai → R often
parameterized by a neural network when dealing with a high
dimensional state space.
The policy of agent i is an -greedy policy and is defined
(cid:26)arg maxa∈Ai Qi(s, a) with probability 1 −
πi(s) =
U(Ai)
with probability
(1)
where U(Ai) denotes a sample from the uniform distribu-
tion over the action space. Each agent stores a set of trajec-
tories {(s, a, ri, s(cid:48))t : t = 1, ..., T} by interacting with the
environment and then updates its policy according to
Qi(s, a) ← Qi(s, a) + α[ri + γ max
a(cid:48)∈Ai
Qi(s(cid:48), a(cid:48))− Qi(s, a)]
(2)
where s is the current state, a is the current action, ri is the
reward obtained by agent i and s(cid:48) is the next state. As agents
are treated independently, the learning makes the environ-
ment appear non-stationary from any agent's perspective.
Other works have attempted to address the non-stationarity
problem in multi-agent RL using a combination of joint-
action learning and importance sampling methods (Foerster
et al. 2017). However, for large populations of agents this
significantly increases the amount of required computation
and becomes infeasible for large states and action spaces.
In order to ensure that information stored in each agent's
memory buffer is relevant to the current transition dynam-
ics of the environment, the buffers are refreshed after every
episode and the agent only trains only on the most recent
experiences.
Reinforcement Learning In Multi-Agent Social
Dilemmas
RL is a useful tool for understanding social dilemmas. In
contrast with game theoretic models, deep RL does not as-
sume knowledge of the environment dynamics and it is able
to deal with continuous state and action spaces. It can there-
fore be used with more complex environmental dynamics
that are more representative of the real-world such as inves-
tigating resource appropriation in the presence of scarce re-
sources (Perolat et al. 2017) or cooperative hunting to maxi-
mize efficiency (Leibo et al. 2017). Furthermore, it has also
been used to analyze the outcome of agents trained with
to be inequity averse with social dilemmas (Hughes et al.
2018), and of societies that feature agents with prosocial be-
havior (Peysakhovich and Lerer 2018) building on works in
behavioral economics.
The main difficulties in applying RL to multi-agent social
dilemmas are credit-assignment, non-stationarity and incen-
tive misalignment. The actual return of an agent's action is
not contained just in the reward but is reflected in what the
other agents learn. If an agent defects, the consequences of
an action must be present in the trajectory in order for an
agent to assign an appropriate value. It is difficult to capture
this information in a trajectory without knowing when and
how other agents are learning. To help account for this, the
notion of intrinsic rewards or agent preferences is often uti-
lized during training (Eccles et al. 2019). This drives agents
to learn coordinated behavior that emphasizes cooperation.
In (Hughes et al. 2018) agents are required to balance the
external rewards they receive with their internal preference
for more equal outcomes. In (Jaques et al. 2019), the reward
that agents receive is also split into rewards that are received
from the game environment but also rewards agents for tak-
ing actions that are measured to be "highly influential" based
on the change it causes in other agents' behaviors. Other
work has involved trying to predict changes in opponent's
future behavior in response to current events (Foerster et al.
2018). In contrast, in this work agents do not receive any re-
wards during training other than those received directly from
playing the Dilemma.
Modeling Methodology
N agents play rounds of a Dilemma game (Table 1). In our
setup, agents are able to select a partner at the beginning of
every round to play the Dilemma and the goal of the game
is to achieve the highest possible individual reward after T
rounds. This also means that each agent takes at least T ac-
tions. The best collective outcome for society is for all the
agents to cooperate unconditionally as this would guaran-
tee maximum reward for every interaction. However, such
an outcome is unlikely because the immediate rewards as-
sociated with defecting in a highly cooperative society are
also significantly higher. By introducing partner selection,
agents have to balance the rewards of greedy defecting be-
havior with the future cost of being excluded.
An episode of the game is split into two phases: in the
first phase, each agent has a chance to select another agent
to play the Dilemma. During this phase, actions as ∈ Rn−1
involve selecting one of the other n − 1 agents in the envi-
ronment. Agents cannot select themselves and, if selected,
they cannot refuse to play. There are no other restrictions on
which agents can be selected nor how many times any agent
can be selected. In the second phase, the selecting agent, i,
and chosen agent, j, play one round of the Dilemma, taking
d) where ad ∈ [C, D] and receive reward rd.
actions (ai
Actions and rewards are according to the payoff in Table 1.
Agents receive no reward from selecting a partner and only
receive rewards when playing the Dilemma. Their rewards
from playing must therefore inform their selection.
d, aj
During selection, every agents' most recent h actions are
visible to all other agents. Each action is represented as a
one-hot encoding and concatenated together such that for
the playing phase agents see states sd ∈ R2×h and in the
selection phase ss ∈ R2×h×(n−1). We choose to set h = 1
for interpretability. This limits the number of strategies that
can be learned to just four strategies that are recognized in
other literature (see Results).
In order to assess the emergent behavior associated with
the dynamics of partner selection, it is important to be able
to vary characteristics like the number of parameters and the
agents' exploration rates for playing the Dilemma and se-
lection. To give us this flexibility, each agent consists of two
models: the selection model learns a policy πs to take action
as given ss while the dilemma model learns a policy πd to
take action ad given sd. During each episode, a trajectory
τ = {ss, as, rs, sd, ad, rd, ...} was sampled from πs and πd
and then each model was trained on the states and actions
sampled from their respective policies.
We parameterize each model using a neural network with
one hidden layer. We test various network sizes between 32
and 256. We display the results for the case of a network
size equal to 256 which were the most consistent and infor-
mative for understanding the development of strategy. Each
agent employs an -greedy policy with a fixed d = 0.05
when playing the Dilemma and s = 0.1 when selecting a
partner. A discussion on is present in the Results section.
Both models use the standard a discount rate γ = 0.99.
Results and Discussion
In this section, we present results that indicate there is a sig-
nificant impact on the development and emergence of co-
operation as a result of partner selection. Alongside partner
(a) Matching determined by partner selection
(b) Random matching
Figure 1: The mean and standard deviation of the number of times an outcome occurs between two agents as a percentage over
14 runs of the simulation. In our simulation, the dynamic of partner selection is essential for cooperation to emerge as a viable
action which leads to a significantly higher global reward. With partner selection we observe distinct phases where certain
strategies dominate. The start of each phase is marked with a dashed line. Eventually agents learn to predominantly cooperate
after approximately 15,000 episodes. When pairing is random, agents quickly learn to defect and do not learn any cooperative
strategy throughout the simulation.
selection, we look at the roles of exploration and the types
of learned strategies that promote a cooperative outcome.
Emergence of Cooperation through Partner
Selection
Training Q-agents in the two-player repeated Prisoner's
Dilemma without partner selection results in both agents
learning to defect every time. With random selection in a
multi-agent population, agents converge to the same solu-
tion that is seen in the two-player case shown in Fig 1(b).
When agents are given the ability to choose a partner within
a population cooperation begins to emerge and the returns
for the society steadily improve.
(Rise in mutual defection).
We observe four distinct phases characterized by different
strategies of the agents. In particular, we identify four phases
which we mark in Fig 1(a). We plot the number of agents
that have selected a partner who has previously cooperated
versus a partner who has previously defected in Fig 3. In line
with our expectations, at convergence, the partner of choice
is a partner that has previously cooperated. Each of these
four phases also coincides with a development in the strategy
of the agents in the society which are as follows:
• Phase 1: Agents learn to defect without partner selection
• Phase 2: Agents learn to pick a partner who has demon-
strated cooperative behavior and defect against them (Rise
in exploitation). Agents pick a partner who has demon-
strated cooperative behavior and cooperates with them
(Rise in mutual cooperation).
• Phase 3: Agents learn to retaliate (defect) against exploita-
tive partners (Fall in exploitation). Agents learn to coop-
erate when selected by other cooperative agents (Rise in
mutual cooperation).
• Phase 4: We observe the emergence of a stable coopera-
tive society with the majority of agents adopting a behav-
ior that resembles equivalent retaliation, i.e., adoption of
the Tit-for-Tat strategy.
When limiting the information an agent has of its oppo-
nent to just the action taken at the previous timestep we limit
the space of strategies agents can learn to four strategies: (1)
where agents always cooperate (ALL C), (2) where agents
always defect (ALL D), (3) Tit-for-Tat (TFT) where agents
copy the last action of their opponent, and (4) reverse Tit-for-
Tat (revTFT), where agents play the opposite of their oppo-
nent's last action. To best understand the development of the
agent strategies throughout the simulation, we capture both
the outcomes of the agent encounters at every episode (Fig
1), the improvement in selection accuracy in (Fig 3), and a
box-plot of the strategies being used (4).
In the first phase, all agents quickly learn a defecting pol-
icy that causes a decrease in the society's cumulative reward.
We can explain this outcome by observing that the agents
have not yet learned how to select a partner efficiently. If the
partner selection is uninformed, then each interaction can be
treated independently and thus there are no future rewards
associated with cooperating. Consequently, defecting is the
preferred strategy for almost all agents. The partner selec-
tion accuracy of agents steadily improves, which facilitates
the transition to the second phase.
The second phase begins at approximately 2,500 episodes
into the simulation and is characterized by an increase in ex-
ploitation and a steep decrease in mutual defection. Agents
have improved their capability to select cooperative part-
ners, which increases the rewards associated with defecting.
Although the outcome of the society is significantly differ-
ent as a result of partner selection, the strategies used in the
Dilemma are the same: agents primarily defect. What facil-
itates the transition to the next stage is the following: (1)
agents who cooperate are selected to play more frequently
than defecting agents (and, therefore, are given the oppor-
tunity to potentially receive rewards); and (2) with enough
(a) epsilon=0.1
(b) epsilon=0.01
Figure 2: Higher exploration results in the society going through the different phases quicker but agents demonstrate less overall
cooperation. Lower exploration inhibits the transition to a cooperative phase.
in mutual cooperation and the society has mostly stabilized.
This is also where the society achieves its highest cumulative
reward. It is expected that at this point, where the number of
TFT agents is at its highest, is where the number of ALL D
agents is at its lowest. The number of ALL C agents stays
relatively consistent from phase 2 throughout to phase 4, al-
though it increases when the ALL D strategy is no longer
popular.
The impact of TFT agents here can be clearly seen as they
significantly limit the effectiveness of the defecting strategy.
As long as agents use ALL D, the best response strategy is
to employ TFT. Furthermore, when more agents use TFT,
the more rewards TFT agents will receive during an episode
which further pushes agents to employ it as a strategy. We
can see that once agents are able to select cooperative part-
ners, the TFT strategy acts as a norm by inducing other
agents to play the same strategy which produces the most
stable iteration of the society.
Analyzing Examples From Simulation
Tracking the returns of each agent alone is insufficient to
capture the state of the society and the group's behavior.
For a more granular understanding, we further plot exam-
ples from the phases from a simulation to show the number
of times each agent is selected and the strategies they use
(Fig 5), the individual reward they receive (Fig 6), and the
interactions represented as a network (Fig 7). We specifi-
cally consider examples of interactions between the phase
transitions described in the previous section.
In episode 2,500, rewards are evenly spread across agents,
however, selections are heavily skewed in favor of coop-
erative agents. Nevertheless, they receive very similar re-
wards to other agents using TFT or ALL D strategies. ALL
D agents that are picked more often do not receive signif-
icantly more rewards as they are most often being selected
by agents who are also defecting meaning little reward is
gained from those encounters. In episode 7,500, we observe
more agents using a TFT strategy. However, the two most
successful strategies in this phase are ALL D (agent 16) and
ALL C (agent 7) both of whom benefit from more effective
Figure 3: Percentage of agents selected who have cooperated
in their last interaction.
exploration, cooperation can be sufficiently rewarded and
agents can start to learn to punish agents who would try to
exploit them.
This happens at approximately 5,000 episodes where
there is a steep decline in exploitation. This leads to the
third phase where there is a sufficient amount of cooperators
such that they can be easily selected and exploiters begin to
be punished. In episode 7,500 there is a steady increase in
the number of encounters resulting in mutual cooperation.
Tracking the agent strategies (Fig 4) reveals that this can be
attributed to the rise in TFT. With an increase in the num-
ber of TFT agents, defecting agents exploit others less suc-
cessfully as an agent who has previously cooperated is more
likely to be a TFT agent than an ALL C agent resulting in en-
counters leading to mutual defection. At the 10,000 episode
mark, mutual cooperation overtakes exploitation as the most
common outcome of two agents interacting that is observed
in the society. Looking at Fig 3, this also coincides with the
convergence of partner selection as 90% of agents who are
selected to play the Dilemma have cooperated in their last in-
teraction. After 15,000 training episodes, the society enters
a phase of where a significant majority of interactions result
Figure 4: Box plots representing the number of agents that use each strategy during phase transitions. The rise in ALL C and
TFT agents coincides with an improvement in partner selection. Once partner selection converges, the number of TFT agents
rapidly increases which results in a more stable, cooperative society.
partner selection.
In phase 3, surprisingly, we observe the largest amount
of inequality. Due to the growing number of TFT agents,
agents using ALL C strategy receive significantly more re-
ward while agents that use ALL D and receive the smaller
rewards on average. The amount of reward the defecting
agents receive is lower due to a combination of not being
selected by other agents, the number of TFT agents in the
population (making it difficult for them to find cooperative
agents to exploit), as well as cooperative agents being able
to reliably cooperate with one another. With only a single
available history, the ALL D agents cannot distinguish be-
tween an ALL C agent and a TFT agent that has recently co-
operated. Interestingly, when a TFT agent punishes an ALL
D agent (or faces a TFT or ALL C agent who has recently
explored defection), the TFT agent suffers a cost in the next
round as their chance of getting selected is reduced despite
having behaved in a cooperative way by punishing a defec-
tor. This explains why ALL C receives the highest overall
reward. As ALL C always cooperates, it does not face this
penalty but still benefits from the regulatory behavior pro-
vided by other TFT agents. Nevertheless, we can see that
the strategy that is adopted by most agents is surprisingly
not the strategy that necessarily generates the highest reward
in each episode.
The Role Of Exploration
In RL, it is necessary to give agents the opportunity to ex-
plore the environment to find an optimal strategy. In multi-
agent systems when the agents are continuously developing
new strategies, there is a need to explore constantly in or-
der to react to the shift in environment dynamics. However,
agents exploring can impact the learning of other agents. Too
much exploration at the wrong time can inhibit the conver-
gence to certain behaviors. As an example, if a single agent
has a cooperative strategy when all the other agents have a
defecting strategy, then exploration benefits the cooperative
strategy. On the other hand, when all agents behave cooper-
atively, exploration instead becomes costly for cooperators.
In our work we have suggested that, in order to achieve a co-
operative outcome, the agent strategy undergoes four stages.
We can see that increasing the base exploration rate affects
the rate at which these strategies are learned, however, it also
significantly affects the overall society's return. In Fig 2(a)
a higher exploration rate results in an overall less coopera-
tive outcome and, therefore, less cumulative reward. Simi-
larly, in Fig 2(b), if the exploration rate is too low, it makes
it difficult for agents to adjust their strategies and, despite
doubling the number of episodes, the agents are not able to
reach a cooperative solution.
Although more exploration facilitates a quicker transition
between phases, it also jeopardizes the stability and gains of
the final outcome. We should therefore be mindful in how we
determine the exploration strategy that our agents employ in
the presence of other learners.
Inclusion and Exclusion to Promote
Norm-Inducing Behavior
The impact of partner selection happens in two steps that
can be viewed as inclusive acts and exclusive acts. Firstly,
agents need to be able to distinguish between different types
of agents. Cooperators need to be included to allow cooper-
ation to compete with defecting strategies for reward. Sec-
ondly, agents need to be able to learn or adapt their playing
strategy given a selection strategy that leads to the exclusion
of and retaliation against defectors. We view this behavior as
norm-inducing as it forces agents to conform to an existing
strategy that has been adopted by the majority of the agents.
In order for the norm to persist, the TFT agents need to inter-
act with a mix of both ALL C and ALL D agents and, while
it regulates the behavior of the other agents, the emergence
and presence of TFT behavior itself is not sustained by it.
Although agents that use an ALL C strategy tend to receive
the highest rewards, without other agents employing a TFT
strategy, the stability of a cooperative society is threatened.
Our experiments suggest that the emergence of cooperation
is contingent on having a sufficient number of agents learn-
ing to exclude bad behavior even if they share some of the
associated cost.
In Fig 7, we display centrality plots from network analy-
sis to visualize partner selection with nodes positioned ac-
cording to Fruchterman-Reingold force-directed algorithm
(Fruchterman and Reingold 1991). In these plots, two nodes
are close if there are more interactions between them. As
partner selection becomes more effective in terms of reward
with more training episodes, the network gets tighter with
Figure 5: With every agent searching for a partner that indicates that they will cooperate, the agent that is most commonly
selected are ones that use ALL C. Early on in the training, ALL D is quickly adopted by most agents while reverse TFT is
phased out. As agents learn TFT, the distribution of selections flattens. As more agents play TFT, less agents play ALL D.
Figure 6: In the early stages of the simulation and very few cooperative agents, the spread of rewards per episode is fairly even.
Despite being selected the vast majority of the time, ALL C agents receive similar rewards per episode. As other agents learn
TFT, the amount of reward that ALL C agents receive significantly increases, while it decreases for ALL D agents. Eventually,
when the majority of agents play TFT, the average rewards per agent increases.
Figure 7: Degree centrality plots of 20 agents in a society during training. Each node represents an agent, while each directed
edge represents an interaction between two agents. Agents with more interactions between each other have shorter edges. Axes
simply represent a normalized node position centered on the origin.
ALL C nodes having a higher degree centrality and influ-
ence. In the examples above, ALL D agents are prominent
outside the core of the network while agents who cooper-
ate have higher degree centrality. As training progresses,
the connections are distributed more evenly as more agents
adopt equivalent-retaliation (TFT) strategy and become vi-
able partners for cooperation. With more TFT agents, it is
harder for ALL D agents to find ALL C agents to exploit
which decreases the amount of connections between those
two node types. This causes a change in the network struc-
ture where there is less likely to be just one or two highly
influential nodes. Finally, as mutual cooperation overtakes
exploitation and the society stabilizes, the network connec-
tions become more evenly spread with nodes having similar
degree centrality.
Conclusions
We have presented an analysis of the effect of partner se-
lection on the emergence of cooperation. We have shown
that agents can learn to cooperate in the repeated Prisoner's
Dilemma using multi-agent reinforcement learning. We have
outlined four key phases that begin with defection but, as
partner selection improves, the agents achieve a cooperative
outcome. Although subject to certain environment variables,
such as reward and the amount of information available to
each agent, partner selection allows for a cooperative strat-
egy to be sustained early on in the simulation when the ma-
jority of agents defect. This facilitates a development of a
strategy that learns to interact efficiently with both coop-
erative and defective agents akin to Tit-for-Tat. The stabil-
ity of a cooperative outcome is dependent on how success-
fully agents are able to use information about another agent's
history to inform partner selection and their strategy in the
Dilemma. Although TFT agents tend to receive less reward
than fully cooperative agents at certain stages, the presence
of this strategy is required for cooperation to emerge at all
as it regulates the exploitative behavior of defecting agents.
The results of this work demonstrate the importance of un-
derstanding how agents can learn to interact with each other
in an environment beyond internal agent or reward function
design that can help produce optimal cooperative outcomes
and provides insight into the emergence of cooperation in
social dilemmas.
References
1981. The evolution of cooperation.
[Albert et al. 2007] Albert, M.; Guth, W.; Kirchler, E.; and
Maciejovsky, B. 2007. Are we nice(r) to nice(r) people? -- an
experimental analysis. Experimental Economics 10(1):53 --
69.
[Axelrod and Hamilton 1981] Axelrod, R., and Hamilton,
Science
W. D.
211(4489):1390 -- 1396.
[Barclay and Willer 2006] Barclay, P., and Willer, R. 2006.
Partner choice creates competitive altruism in humans.
Proceedings of the Royal Society B: Biological Sciences
274(1610):749 -- 753.
[Bolton and Ockenfels 2000] Bolton, G. E., and Ockenfels,
A. 2000. Erc: A theory of equity, reciprocity, and competi-
tion. American Economic Review 90(1):166 -- 193.
[Busoniu, Babuska, and De Schutter 2008] Busoniu,
L.;
Babuska, R.; and De Schutter, B. 2008. A comprehen-
IEEE
sive survey of multiagent reinforcement learning.
Transactions on Systems, Man, And Cybernetics-Part C:
Applications and Reviews, 38 (2), 2008.
[Cialdini 2003] Cialdini, R. B. 2003. Crafting normative
messages to protect the environment. Current Directions in
Psychological Science 12(4):105 -- 109.
[Cuesta et al. 2015] Cuesta, J. A.; Gracia-L´azaro, C.; Fer-
rer, A.; Moreno, Y.; and S´anchez, A. 2015. Reputation
drives cooperative behaviour and network formation in hu-
man groups. Scientific Reports 5:7843.
[Dawes 1980] Dawes, R. M. 1980. Social dilemmas. Annual
Review of Psychology 31(1):169 -- 193.
[Eccles et al. 2019] Eccles, T.; Hughes, E.; Kram´ar, J.;
Wheelwright, S.; and Leibo, J. Z. 2019. Learning reciprocity
arXiv preprint
in complex sequential social dilemmas.
arXiv:1903.08082.
[Fehr and Fischbacher 2004] Fehr, E., and Fischbacher, U.
2004. Social norms and human cooperation. Trends in Cog-
nitive Sciences 8(4):185 -- 190.
[Foerster et al. 2017] Foerster, J.; Nardelli, N.; Farquhar, G.;
Afouras, T.; Torr, P. H.; Kohli, P.; and Whiteson, S. 2017.
Stabilising experience replay for deep multi-agent reinforce-
ment learning. In Proceedings of the 34th International Con-
ference on Machine Learning, 1146 -- 1155.
[Foerster et al. 2018] Foerster, J.; Chen, R. Y.; Al-Shedivat,
M.; Whiteson, S.; Abbeel, P.; and Mordatch, I. 2018. Learn-
ing with opponent-learning awareness. In Proceedings of the
17th International Conference on Autonomous Agents and
MultiAgent Systems, 122 -- 130. International Foundation for
Autonomous Agents and Multiagent Systems.
[Fruchterman and Reingold 1991] Fruchterman, T. M., and
1991.
Reingold, E. M.
Graph drawing by force-
Software: Practice and Experience
directed placement.
21(11):1129 -- 1164.
[Fu et al. 2008] Fu, F.; Hauert, C.; Nowak, M. A.; and Wang,
L. 2008. Reputation-based partner choice promotes cooper-
ation in social networks. Physical Review E 78(2):026117.
[Hughes et al. 2018] Hughes, E.; Leibo, J. Z.; Philips, M. G.;
Tuyls, K.; Du´enez-Guzm´an, E. A.; Castaneda, A. G.; Dun-
ning, I.; Zhu, T.; McKee, K. R.; Koster, R.; et al. 2018. In-
equity aversion resolves intertemporal social dilemmas. In
Proceedings of the 35th International Conference on Ma-
chine Learning.
[Izquierdo, Izquierdo, and Gotts 2008] Izquierdo,
S.;
Izquierdo, L.; and Gotts, N. 2008. Reinforcement learning
dynamics in social dilemmas. Journal of Artificial Societies
and Social Simulation 11(2):1.
[Jaques et al. 2019] Jaques, N.; Lazaridou, A.; Hughes, E.;
Gulcehre, C.; Ortega, P.; Strouse, D.; Leibo, J. Z.; and
De Freitas, N. 2019. Social influence as intrinsic motiva-
tion for multi-agent deep reinforcement learning. In Inter-
national Conference on Machine Learning, 3040 -- 3049.
[Leibo et al. 2017] Leibo, J.; Zambaldi, V.; Lanctot, M.;
Marecki, J.; and Graepel, T. 2017. Multi-agent reinforce-
ment learning in sequential social dilemmas. In Proceedings
of the 16th Conference on Autonomous Agents and MultiA-
gent Systems, 464 -- 473.
International Foundation for Au-
tonomous Agents and Multiagent Systems.
[Lowe et al. 2017] Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.;
2017. Multi-agent actor-
Abbeel, P.; and Mordatch, I.
critic for mixed cooperative-competitive environments.
In
Advances in Neural Information Processing Systems, 6379 --
6390.
[Macy and Flache 2002] Macy, M., and Flache, A. 2002.
Learning dynamics in social dilemmas. Proceedings of the
National Academy of Sciences 99(3):7229 -- 7236.
[Milinski, Semmann, and Krambeck 2002] Milinski, M.;
Semmann, D.; and Krambeck, H. 2002. Donors to charity
gain in both indirect reciprocity and political reputation.
Proceedings of the Royal Society of London. Series B:
Biological Sciences 269(1494):881 -- 883.
[Mohtashemi and Mui 2003] Mohtashemi, M., and Mui, L.
2003. Evolution of indirect reciprocity by social informa-
tion: the role of trust and reputation in evolution of altruism.
Journal of Theoretical Biology 223(4):523 -- 531.
[Perolat et al. 2017] Perolat, J.; Leibo, J.; Zambaldi, V.;
Beattie, C.; Tuyls, K.; and Graepel, T. 2017. A multi-agent
reinforcement learning model of common-pool resource ap-
propriation. In Advances in Neural Information Processing
Systems, 3643 -- 3652.
[Peysakhovich and Lerer 2018] Peysakhovich, A., and Lerer,
A. 2018. Prosocial learning agents solve generalized stag
In Proceedings of the 17th
hunts better than selfish ones.
International Conference on Autonomous Agents and Mul-
tiAgent Systems, 2043 -- 2044.
International Foundation for
Autonomous Agents and Multiagent Systems.
[Posner 2000] Posner, E. A. 2000. Law and social norms:
The case of tax compliance. Virginia Law Review 86:1781.
[Santos, Santos, and Pacheco 2008] Santos, F. C.; Santos,
M. D.; and Pacheco, J. M. 2008. Social diversity promotes
the emergence of cooperation in public goods games. Nature
454(7201):213.
[Sen and Airiau 2007] Sen, S., and Airiau, S. 2007. Emer-
In International
gence of norms through social learning.
Joint Conferences on Artificial Intelligence, 1507 -- 1512.
[Thøgersen 2006] Thøgersen, J.
2006. Norms for envi-
ronmentally responsible behaviour: An extended taxonomy.
Journal of Environmental Psychology 26(4):247 -- 261.
[Watkins and Dayan 1992] Watkins, C. J., and Dayan, P.
1992. Q-learning. Machine Learning 8(3-4):279 -- 292.
|
1905.08036 | 1 | 1905 | 2019-05-14T03:44:56 | A Reputation System for Multi-Agent Marketplaces | [
"cs.MA",
"cs.SI"
] | We present an exploration of a reputation system based on explicit ratings weighted by the values of corresponding financial transactions from the perspective of its ability to grant "security" to market participants by protecting them from scam and "equity" in terms of having real qualities of the participants correctly assessed. We present a simulation modeling approach based on the selected reputation system and discuss the results of the simulation. | cs.MA | cs | A Reputation System for Multi-Agent Marketplaces
Anton Kolonin, Ben Goertzel, Cassio Pennachin, Deborah Duong, Matt Iklé ,
Nejc Znidar, Marco Argentieri
SingularityNET Foundation, Amsterdam, Netherlands
{anton, ben, cassio}@singularitynet.io
Abstract
transactions from
We present an exploration of a reputation system
based on explicit ratings weighted by the values of
corresponding financial
the
perspective of its ability to grant "security" to
market participants by protecting them from scam
and "equity" in terms of having real qualities of the
participants correctly assessed. We present a
simulation modeling approach based on the selected
reputation system and discuss the results of the
simulation.
1 Introduction
The latest developments in e-Commerce and the emergence
of global commercial online ecosystems with world-wide
connectivity based on the Internet makes it critical to assess
the reliability of the vendors and suppliers in these
ecosystems in reliable way [Zheng and Jin, 2009]. Multiple
design solutions for reputation systems serving the purpose
exist, such as discussed by other authors: [Swamynathan et
al., 2010], [Sänger and Pernul, 2018]. The most critical part
of any system intended for this purpose appears to be the
ability to handle a high degree of anonymity [Androulaki et
al., 2008] of market participants, as it is characteristic of
modern distributed [Gupta et al., 2003] ecosystems,
including ones that are based on public networks [Blömer et
al., 2015] such as blockchains.
In particular, in this work, we rely on a reputation system
design based on the "weighted liquid rank" concept [Kolonin
et al., 2018] and an implementation of the concept [Kolonin
et al., 2019] applied to generic multi-agent marketplaces.
Given the existing implementation of the system itself and
the simplistic simulation described in the latter publication,
in scope of this work we provide a more advanced simulation
modeling of the same system and investigate market
conditions which make it possible for honest market
participants to be granted "security", protecting them from
scam, and "equity", ensuring that their qualities are assessed
fairly in the market.
2 Reputation System Implementation
Most of the "weighted liquid rank" design of the reputation
system per [Kolonin et al., 2018] and its implementation
according to [Kolonin et al., 2019] are used in this work, and
we here we expand upon these earlier findings, that the
reputation system based on explicit ratings weighted by
financial values provides the best combination of "security"
and "equity". In this work we focus primarily on use of only
explicit ratings weighted by financial values. It should be
noted that, as in the latter work, it has been found that a ratio
of 10% suppliers and 90% consumers with no overlap is
typical in generic online marketplaces, and because of this
the reputation of consumers as raters cannot be assessed
accordingly, thus the "liquid" part of the algorithm is not
effectively applicable in such a case and only the "weighted"
is employed.
Algorithm 1 Weighted Liquid Rank (simplified version)
Inputs:
1) Volume of rated transactions each with financial
value of the purchased product or service and rating
value evaluating quality of the product/service,
covering specified period of time;
2) Reputation ranks for every participant at the end of
the previous time period.
Parameters: List of parmeters, affecting computations
- default value, logarithmic ratings, conservatism,
decayed value, etc.
Outputs: Reputation ranks for every participant at the
end of the previous time period.
1: foreach of transactions do
2:
3:
4:
5:
6: end foreach
let rater_value be rank of the rater at the end of
previous period of default value
let rating_value be rating supplied by
trasaction rater (consumer) to ratee (supplier)
let rating_weight be financial value of the
transaction of its logarithm, if logarithmic ratings
parameter is set to true
sum rater_value*rating_value*rating_weight for
every ratee
previous peiod with differential_ranks based on
parameter of conservatism, so that new_ranks =
(old_ranks*conservatism+N*(1-differential_ranks)),
using decayed value if no rating are given to ratee
during the period
7: do normalization of the sum of the muliplications
per ratee to range 0.0-1.0, get differential_ranks
8: do blending of the old_ranks known at the end of
9: do normalization of new_ranks to range 0.0-1.0
10: return new_ranks
We
open-source
implementations of the reputation system based on the
the
referenced design. The first one was based on
Reputationer
project
at
https://github.com/aigents/aigents-java/ with
a Python
adapter for the SingularityNET reputation system prototype
as
the
https://github.com/singnet/reputation/ project. The second
version was based on a native Python implementation as the
reputation_service_api.py script in the latter project.
During the simulation modeling, the impact of the following
parameters of the reputation system were explored:
the aigents_reputation_api.py
in Aigents
alternative
script of
class
Java
have
explored
two
in
⚫ Rd - default initial reputation rank;
⚫ Rc - decayed reputation in range to be approached
by inactive agents eventually;
⚫ C - conservatism as a blending "alpha" factor
between the previous reputation rank recorded at
the beginning of the observed period and the
differential one obtained during the observation
period;
⚫ FullNorm -- when this boolean option is set to True
the reputation system performs a full-scale
normalization of incremental ratings;
⚫ LogRatings - when this boolean option is set to
True the reputation system applies log10(1+value)
to financial values used for weighting explicit
ratings;
⚫ Aggregation - when this boolean option is set to
True the reputation system aggregates all explicit
ratings between each unique combination of two
agents with computes a weighted average of
ratings across the observation period;
⚫ Downrating - when this boolean option is set to
True the reputation system translates original
explicit rating values in range 0.0-0.25 to negative
values in range -1.0 to 0.0 and original values in
range 0.25-1.0 to the interval 0.0-1.0.
⚫ UpdatePeriod -- the number of days to update
reputation state, considered as observation period
for computing incremental reputations.
3 Reputation System Simulation
Based on a market analysis reported on in the referenced
work [Kolonin et al., 2019] where it was affirmed that a ratio
the
code
simulation,
of which
of 9:1 between buyers and sellers (9 times as many buyers
than sellers) was the most realistic, we used this ratio as a
fixed market parameter. With this ratio, we simulated
populations of 1000 agents interacting over 6 months on an
everyday basis. We also implemented feedback, making the
reputation system available to consumers for choosing
suppliers, who then rate them in the same reputation system.
During initial simulations, 80% of the population were
"good" agents with fair market behaviors and the remaining
20% were "bad" agents that commit scams. It was also
assumed that the amount of "bad" agent transactions is higher
than the amount of "good" agent transactions as a result of
the "bad" agents pumping up the reputation of "fake"
suppliers as they pose as consumers. It was assumed that an
average "bad" agent spends substantially less on every
transaction, compared
to an average "good" agent.
Specifically, in the simulations it was assumed that "bad"
agents emit 10 times more transactions a day than the good
ones, but the ratio of the average transaction payment amount
was varied. For the latter ratio, we considered three types of
conditions: an "unhealthy" market with a ratio equal to 10 (so
the "good" agents have just 10 times more costly purchases),
a "semi-healthy" one with a ratio of 20 and a "healthy" one
with a ratio of 100.
The
at
https://github.com/singnet/reputation/tree/master/agency/pyt
hon/src/snsim/reputation, is meant to resemble a market
place, in which there are different product or service
categories that have a variety of prices reflective of the
market. Each supplier agent has an intrinsic ability to satisfy
its customers with its product or service, and ratings of each
supplier are normally distributed around this "expected
goodness" (Rcea as defined further), here drawn from a normal
distribution with mean of 0.75 and standard deviation of 0.15
to N(0.75, 0.0225), bucketed into five ratings of 0.0, 0.25, 0.5
0.75 and 1.0. Scammer agents are consistently given a zero
by good agents. Each agent has different (normally
distributed) needs for every product or service, and shops for
what it needs the most first. There is a normally distributed
time until each product or service is needed again. There is a
limit to how much shopping an agent can do in a day, and the
number of active agents can vary. Agents have differing
(normally distributed) propensities to try out new agents,
different tolerances for staying with the old ones, and
different capacities to forget. Bad agents have a normally
distributed ring size of other bad agents which they
(mutually) rate positively, here drawn from N(8,4)
Algorithm 2 Market Simulation
Input: Consumer and Supplier trade behaviors
Output: Metrics, Agent qualities, transactions, ranks
1: Assign Agents to Behaviors based on Normal
Random Variates
2: Every day for 6 months:
is
Each consumer makes shopping list
Agents drop past suppliers according to satisfaction
If reputation system in use:
If "winner take all" usage:
Agents choose new suppliers with the highest
reputation score
Else if "roulette wheel" usage:
Agents choose new suppliers in proportion to
their reputation scores
Else if "thresholded random" usage:
Agents choose new suppliers randomly over a
reputation score threshold
Agents make purchases and rate suppliers
Else if reputation system not in use:
If agents have no experience with suppliers
Agents choose new suppliers randomly
Agents make purchases and rate suppliers
3: Print metrics
4 Performance Metrics
In this work we extend the range of the metrics used in
referenced research [Kolonin et al., 2019]. The first key
metric measures are the economic parameters of the market
such as volume ratio for market spending by "good" agents
versus "bad" agents (Vg / Vb ), the percentage of funds spent
by "good" agents to "bad" scammers or "loss to scam"
(Cgbg ), and the ratio between the earnings of scammers and
their spendings or "profit from scam" (Cgbb). The latter
financial metrics serve to identify the amount of financial
resources that fair consumers lose in scams, or the relative
amount that scammers would earn by cheating the reputation
system, so changes in these values based on use of the
reputation system can actually provide the business value of
using it.
Further we measure utility, or how happy fair consumers are
with their purchases. It is the average rating given to
purchases, regardless of the kind of purchase. This metric
would shift with different rules for ratings, such as a different
default rating, however, within a scenario this metric has the
same relative value.
where Fij is the sum of all ratings made by good agents and N
is the number of ratings made by all of the good agents, i.e.
all agents which have an expected goodness of Rea = 1, to be
(1)
discussed further on.
We also explored metrics that assess the reputation system's
prediction of the quality of each agent to be either good or
bad, or ability of the reputation system to have a high
accuracy for both assessments, following the referenced work
[Kolonin et al., 2019]. These metrics compare distributions
of expected reputations Rea per agent a against lists of
computed reputations Rca for the same set of agents. In our
work these values are measured on continuous scale, so Rea
is used to assess the reputation system's ability to predict the
level of "goodness" of an agent. The "bad" agents are
expected to have Rea close to 0.0, while the "good" agents are
expected to have Rea substantially greater than 0 and up to 1.0.
The standared Pearson correlation coefficient PCC between
Rca and Rea is used in the referenced work [Kolonin et al.,
2019]. However, it makes it possible to evaluate only overall
quality of prediction so if the PCC value is below 1.0 then
one can't tell if it is because the flawed "security" or missed
"equity". Because of the reason, we have also used a
"weighted Pearson correlation coefficient," weighted so that
agents matter in proportion to how good they are, or how bad
they are, in order to indicate the capacity of the reputation
system to be used for security (by being good at determining
which agents may be scammers) vs. the capacity of the
reputation system to judge the quality of the honest agent's
products fairly, so that they may participate proportionately
in the economy.
where avg(y,w) is a weighted average. For PCCG, the PCC
weighted towards good, w is the (Rcea) vector. For PCCB it
is the (1-Rcea) vector. We report the weighted average of the
market volume of the PCC, PCCG and PCCB per product or
service category, called PCC by category, PCCG by
category, and PCCB by category. This is useful because we
are ultimately looking for a way to compare providers of a
product or service category, and PCC is a good way to
measure the ability to rank.
The standard confusion matrix metrics for recall, precision,
F1 and accuracy have been used to assess the "security" as
effectiveness of the reputation system for the primary
reputation system goal of protecting the public from scams.
That is, reputation system was considered a tool to solve
"classification problem" categorizing agents as either "good"
or "bad", in the sense of agent being a fair one or a scammer,
respectively. Here, a positive result would mean either a
"good" agent is categorized as "good" in the sense that it is
not a scammer when it is scored over a threshold or a "bad"
agent categorized as "bad", respectively. It is a counted as a
true positive if the agent is in fact good, and is a false positive
if the agent is in fact bad. On the other hand, if an agent is
scored under the threshold and is in fact good it is a false
negative, and if the agent scores under the threshold and is in
fact bad, it is a true negative. In our work, we have used the
threshold as 40% reputation rank value on scale between 0%
and 100%, based on search for the optimal threshold value
differentiating the "good" and the "bad" to the greater extent.
5 Simulation Results
We present three scenario experiments on the reputation
system to find the sets of parameters in which the reputation
system performs best as well as the boundary values of those
left we present expected goodness (Rea) and the lower right we
present computed goodness (Rca), both ordered by expected
goodness. In the charts on the upper and lower right, scammers
appear on the left, and honest agents on the right.
Indeed, the Fig.2 presents how the expected goodness can be
evaluated on per-category basis, assuming that evaluation of
is domain-specific [Kolonin et al., 2018],
reputation
according to specific price ranges and sub-markets of specific
kind of goods or services of the greater market. Where there
are more suppliers and thus more market volume, regardless
of whether bad agents are present, the calculated goodness
are more predictive of the expected goodness of the agent.
The predictions are accurate wherever there are more than
three samples. This result shows a consistent clear
demarcation between bad agents on the left of each chart and
the good agents on the right.
parameters. The first is an experiment is a test of the
parameters under the conditions of a market that is close to
equilibrium, that is, a market in which there is neither a
serious oversupply or serious shortages. Next, in order to test
the robustness of the reputation system we look at its usage
and presentation, in the usual case, where agents trade with
suppliers in proportion to their ratings, and additionally
extremes where agents choose the top-rated supplier, as well
as where they simply choose a random supplier over a
threshold. Finally, to see if the reputation is robust with
respect to market volume surges we take the champion
parameters of the equilibrium market and use them to test
another setting of the reputation system: the periodic setting,
where the update period is set longer than a day, in contexts
with and without periodic surges in purchasing.
The simulation was focused on the case where the reputation
system with explicit
financial
transactions is used. For this case we have explored different
combinations of parameters evaluating different metrics to
access the level of "security" and "equity" provided by the
reputation system.
The extent to which system can provide "security" and
"equity" is illustrated by Fig.1 below. We clearly see that we
can identify "bad agents" with loss of 2 "false negatives". At
first glance one would think that this system only separates
the scammers from the honest agents, providing "security"
while not giving fair ratings to the quality of the honest
agents: it seems that it cannot identify the expected goodness
so the "equity" is not granted. However, the next figure
shows that when you take each product or service category
into account separately, it does predict an agents ratings
accurately.
ratings weighted by
Figure 2. These charts from a single sample run show the calculated
goodness (Rca) along the X axis and the expected goodness value
(Rea) on the Y axis, separated by product or service category. They
include the product or service categories of the realistic simulation
that have more than one supplier. Rca is more predictive of Rea the
larger the sample size (the slope approaching 1), and is accurate
wherever the sample size is over three.
Figure 1. This is an illustration of the suppliers from a sample run
of 1000 agents, 10% of which are suppliers. Each bar is a single
supplier. The reputation system parameters are FullNorm=True,
Weighting=True, LogRatings=False, Rd=0.5, C=0.5, Rc=0.0,
Downrating=False. On the upper left we present computed
goodness (Rca) and on the upper right we present expected
goodness(Rea), both ordered by computed goodness. On the lower
Figure 3. The first of three market scenario types to test the
parameterization and robustness of the reputation system. In this
equilibrium scenario, the boundary value tested is the market
volume ratio, and the parameters tested are the conservatism
(conserv) and default parameters. MVR (Market Volume Ratio)
states: healthy: over 200 semi-heathy: 75-200, unhealthy: below 75.
The best performing combination, default 0.5/conservatism 0.5, is
underlined.
to Scam that comsumers suffer is less when the reputation
system is in use than when it is not. In our consideration of
the importance of balance between offerring protection from
scammers and making sure no honest suppliers are rated
fairly, we have underlined the "champion" parameter
combination
that meets both goals, default 0.5 and
conservatism 0.5.
The healthy market champion's precision of 1.0 in Figure 3
shows that the reputation system did not recommend any of
the scammers despite their best efforts at pumping their
ratings, while at the same time rating the non-scammers fairly
as evidenced by the Pearson-good by category score of 0.92.
The Pearson by category scores show that the system ranks
all agents effectively within a category of goods and services.
If default is set low, then we can make the consumers slightly
happier with their purchases as a result of the reduced loss to
scam, but this is at the expense of ranking good agents fairly.
Figure 4 illustrates the effectiveness of the reputation system
under different usages. One possible usage is winner take all,
when reputation system scores are presented to consumers
such that they would probably just take the highest scoring
supplier. Such a usage has increased the happiness that
consumers have in thier products (increased utility by 11%),
however more studies need to be done in other scenarios that
have enterers and leavers, as winner take all dyanamics have
been known to close markets to newcomers and thus
ultimately result in lower utility. The healthy market with
winner take all dynamics was poor in ranking fair suppliers
(Pearson-good by category of 0.71). However, our champion
scenario, which models the case where customers are
presented with the reputation system such that they would
select agents in proportion to their reputation scores, ranked
non-scamming suppliers fairly (Pearson-good by category of
0.9) so they can participate in the economy. Proportionate
selection usage is roulette wheel selection [Bäck Thomas,
1996], which does better than the thresholded random usage
which does not rank quite as well.
Figure 4. In the second market scenario type we test the robustness
of the reputation system against both normal and extreme usages,
where "Roulette Wheel" indicates agent choice in proportion to the
calculated reputation values, "Thresholded Random" indicates that
agents choose suppliers randomly but over a threshold, and "Winner
Take All" indicates that agents choose the top ranked suppliers.
Figure 3 illustrates how, in a healthy market, the reputation
system is effective in reducing the loss of consumers to
scam,ensuring that nearly all of the suppliers it recommends
are fair market participants while at the same time rating them
fairly. At test of multiple simulation runs shows that, in a
healthy market scenario, there is a 99% chance that the Loss
Figure 5. In the third market scenario type, the period that the
reputation system is updated is tested, in scenarios that are periodic
(the default) in that there is a surge of purchasing every seven days
for half of the goods, and scenarios which are. Scenarios with no
reputation system are presented for comparison in all scenario types,
and best performing parameter combinations in both healthy and
semi-healthy markets, is underlined.
Figure 5 illustrates how the reputation system performs well
even in markets that have periodic surges. The champion of
the market is the champion of the entire paper, where the
reputation system which updates only every week , on a
market with weekend purchasing surges, has a 0.98 Pearson-
good by category score. In markets without surges, the same
system performs nearly as well, and still reduces loss to scam
over the reputation system without the update period and over
no reputation system.
Based on studies within the scope of performed simulations,
the following observations were made.
⚫ The best
set of
are:
Weighting=True, LogRatings=False, Rd=0.5, C=0.5,
Rc=0.0, Downrating=False.
reputation parameters
⚫ Parameters that matter to the result are Weighting, Rd, C,
and Rc. FullNorm does not matter to the result, and
Downrating and LogRatings consistently detract from
the result.
⚫ The reputation system cuts the loss to scam (LTS) that
consumers suffer nearly in half. In a healthy market
LTS is cut from 0.06% to 0.03 or 0.04%.
⚫ The reputation system gives an excellent ability to rank
good agents by product or service category and discern
bad agents in a healthy market, with Pearson-good by
category scores over 0.9 and precision of discerning bad
agents of 1.0, and little societal loss to scam (< 0.0003).
The capability of the reputation is retained in the semi-
healthy market, but decreases dramatically in the
unhealthy market (Pearson of 0.61) making the semi-
healthy market
the
reputation system to function.
the borderline condition for
⚫ The reputation system can handle periodic sales (such as
seasonal and weekly surges in purchasing) with the
same ability to rank good agents and discern bad
(Pearson scores >= 0.9 and precision of 1.0). If there is
no periodicity in purchases, but the periodic mode of the
reputation system is on, the performance is comparable
(Pearson scores = 0.89 and precision of 1.0), however,
if there is a periodicity in purchases and periodic mode
is not employed, performance is slightly lower (Pearson
scores = 0.83 and precision of 1.0).
⚫ The reputation system performs best when used by
agents to trade with suppliers in proportion to their
calculated goodness scores (Precision 1.0, Pearson
scores >=0.9) but still works adequately at the extremes
of "Winner Take All" choice of suppliers (Precision 1.0,
Pearson scores >= 70) and random above threshold
choice (Precision 1.0, Pearson scores >= 80)
It may be possible to incentivize agents to trade with
suppliers in proportion to their calculated goodness
scores financially, not only to keep the reputation
system accurate, but in the interest of equity.
⚫
threshold under which an agent
⚫ Further studies need to be done on automatically finding
a
is never
recommended. Dynamic adjustment of the threshold
will make the system even more robust against market
conditions.
6 Conclusion
Based on the presented work, we can make the following
conclusions.
⚫ For the studied market conditions, the use of the
reputation system based on explicit ratings weighted by
financial values of transactions provides both "security"
in the form of protection from scam and "equity" for the
honest participants of the market
⚫ The parameters that we have found may be used for fine-
tuning of the real reputation systems and the metrics that
we define may be used to optimize these parameters in
further studies.
⚫ The options suggested for threshold management, and
for the incentivization of users to choose amongst the
listed suppliers in proportion to their reputation scores
may be considered as an extra application layer on top of
the reputation system.
The following issues will be considered for our further work.
⚫ Make the simulations more realistic, exploring larger
scales of markets above 1000 agents and 6 months,
studying wider ranges of possible market conditions,
with different populations of scamming agents, different
ratios of transactions committed by the agents and
implementing different
the
behaviors of the agents, primarily having different rates
of newcomers and leavers.
temporal patterns of
⚫ Evaluate possibility to improve "security" and "equity"
with use of rating aggregation.
⚫ Explore possibility of implementation of dynamic
reputation threshold management enabling to cut off
"most likely scam" from "most likely honest" and study
its effect on "security" and "equity".
⚫
⚫ Explore possibility of implementation and evaluation of
the incentivization of the honest agents to act as
members of "reputation militia" so the consumers are
choosing suppliers in proportion to their scores.
Implementation and simulation of the staking-based
reputations [Kolonin et al., 2018] and exploration of its
impact on improvement of "equity", especially in the
case with newcomers entering the system and leavers
leaving the one at realistic rates.
Information Processing, vol 36. Springer, Berlin,
Heidelberg, 2009. https://doi.org/10.1007/978-3-642-
03132-8_24
References
[Androulaki et al., 2008] Androulaki E., Choi S.G.,
Bellovin S.M., Malkin T. Reputation Systems for
Anonymous Networks. In: Borisov N., Goldberg I. (eds)
Privacy Enhancing Technologies. PETS 2008. Lecture
Notes in Computer Science, vol 5134. Springer, Berlin,
Heidelberg, 2008.
https://link.springer.com/chapter/10.1007/978-3-540-
70630-4_13
[Bäck Thomas, 1996] Bäck T. Evolutionary Algorithms in
Theory and Practice. Oxford University Press, Oxford,
United Kingdom, 1996.
[Blömer et al., 2015] Blömer J., Juhnke J. & Kolb C.
Anonymous and Publicly Linkable Reputation Systems.
Financial Cryptography and Data Security 2015,
Cryptology ePrint Archive: Report 2014/546, 2015.
https://eprint.iacr.org/2014/546
[Gupta et al., 2003] Gupta M., Judge P., Ammar M. A
Reputation System for Peer-to-Peer Networks.
NOSSDAV'03, June 1 -- 3, 2003, Monterey, California,
USA, ACM 1-58113-694-3/03/0006, 2003).
https://www.cs.indiana.edu/~minaxi/pubs/reputation.pdf
[Kolonin et al., 2018] Kolonin A., Goertzel D., Duong D.,
Ikle M. A Reputation System for Artificial Societies.
arXiv:1806.07342 [cs.AI], 2018.
https://arxiv.org/abs/1806.07342
[Kolonin et al., 2019] Kolonin A., Goertzel D., Pennachin
C., Duong D., Argentieri M., Znidar N. A Reputation
System for Marketplaces - Viability Assessment. arXiv:
1902.03857 [cs.SI], 2019.
https://arxiv.org/abs/1902.03857
[Swamynathan et al., 2010] Swamynathan G., Almeroth K.,
Zhao B. The design of a reliable reputation system.
Electron Commer Res 10: 239 -- 270, pp.239-270, 2010.
DOI 10.1007/s10660-010-9064-y, 31 August 2010
http://people.cs.uchicago.edu/~ravenben/publications/pd
f/reputation-ecrj10.pdf
[Sänger and Pernul, 2018] Sänger, J. & Pernul G.
Interactive Reputation Systems. Business & Information
Systems Engineering, August 2018, Volume 60, Issue 4,
pp 273 -- 287, 2018 60: 273, 2018.
https://doi.org/10.1007/s12599-017-0493-1
[Zheng and Jin, 2009] Zheng W., Jin L. Online Reputation
Systems in Web 2.0 Era. In: Nelson M.L., Shaw M.J.,
Strader T.J. (eds) Value Creation in E-Business
Management. AMCIS 2009. Lecture Notes in Business
|
0911.1346 | 1 | 0911 | 2009-11-06T20:29:10 | Optimal Approximation Algorithms for Multi-agent Combinatorial Problems with Discounted Price Functions | [
"cs.MA",
"cs.DS"
] | Submodular functions are an important class of functions in combinatorial optimization which satisfy the natural properties of decreasing marginal costs. The study of these functions has led to strong structural properties with applications in many areas. Recently, there has been significant interest in extending the theory of algorithms for optimizing combinatorial problems (such as network design problem of spanning tree) over submodular functions. Unfortunately, the lower bounds under the general class of submodular functions are known to be very high for many of the classical problems.
In this paper, we introduce and study an important subclass of submodular functions, which we call discounted price functions. These functions are succinctly representable and generalize linear cost functions. In this paper we study the following fundamental combinatorial optimization problems: Edge Cover, Spanning Tree, Perfect Matching and Shortest Path, and obtain tight upper and lower bounds for these problems.
The main technical contribution of this paper is designing novel adaptive greedy algorithms for the above problems. These algorithms greedily build the solution whist rectifying mistakes made in the previous steps. | cs.MA | cs | Optimal Approximation Algorithms for Multi-agent
Combinatorial Problems with Discounted Price Functions
Gagan Goel, Pushkar Tripathi and Lei Wang∗
Abstract
Submodular functions are an important class of functions in combinatorial optimiza-
tion which satisfy the natural properties of decreasing marginal costs. The study of
these functions has led to strong structural properties with applications in many areas.
Recently, there has been significant interest in extending the theory of algorithms for
optimizing combinatorial problems (such as network design problem of spanning tree)
over submodular functions. Unfortunately, the lower bounds under the general class of
submodular functions are known to be very high for many of the classical problems.
In this paper, we introduce and study an important subclass of submodular func-
tions, which we call discounted price functions. These functions are succinctly rep-
resentable and generalize linear cost functions. In this paper we study the following
fundamental combinatorial optimization problems: Edge Cover, Spanning Tree, Per-
fect Matching and Shortest Path, and obtain tight upper and lower bounds for these
problems.
The main technical contribution of this paper is designing novel adaptive greedy
algorithms for the above problems. These algorithms greedily build the solution whist
rectifying mistakes made in the previous steps.
9
0
0
2
v
o
N
6
]
A
M
.
s
c
[
1
v
6
4
3
1
.
1
1
9
0
:
v
i
X
r
a
∗College of Computing, Georgia Institute of Technology, Atlanta, GA 30332 -- 0280
1 Introduction
Combinatorial optimization problems such as matching, shortest path and spanning tree
have been extensively studied in operations research and computer science. In these prob-
lems we are given a ground set E and a collection Ω of subsets of the ground set that is
usually implicitly defined by some combinatorial propoerty. We are also given the cost of
each element in E and the price of a subset in Ω is defined to be the sum of costs of its con-
stituents. The objective is to find a subset in Ω with the minimum price. This framework
with linear cost functions has served as the simplified model for many practical problems.
However, linear cost functions do not always model the complex dependencies of the
prices in a real-world setting. Often, they only serve as approximations to the original
functions. Hence solutions to these problems are rarely useful in the practical realm where
these simplifying assumptions breakdown.
Another feature that is observed in a real-world scenarios is the presence of multiple
agents, with different price functions, who wish to collaborate to build a single combinatorial
structure. For linear cost functions, it is easy to see that having multiple agents doesnt
change the complexity of the original problem. This is not the case for more general cost
functions.
Submodular functions provide a natural way to model the non-linear dependencies be-
tween prices. These functions are the discrete analog of convexity and exhibit well known
economic properties such as economy of scale and decreasing marginal cost. There has been
some recent work [18, 9, 11] to study the approximability of combinatorial problems under
such functions and tight upper and lower bounds are known for many problems in this very
general framework.
However, the submodular setting suffers from a major drawback. Although in real
world applications an agent's cost function is not necessarily as simple as a linear function,
it is often not as complicated as a general submodular function either.
i.e. submodular
functions seem to tackle the problem in too much generality which often makes problems
harder than required. As a result most of the bounds presented in previous work in this area
are weak and have little practical applicability. A barrier in designing efficient algorithms
for submodular cost functions arises from difficulty in representing them succinctly. Since a
general submodular function is defined over an exponentially sized domain it is customary
to assume that the function can be accessed by an oracle which can be querried polynomial
number of times. This limits our ability to design efficient algorithms for these problems
owing to the incomplete information about the functions, resulting in information theretic
lower bounds.
In this paper we study an important subclass of succinctly representable submodular
functions called discounted price functions in the multi-agent setting. These functions are
motivated by the observation that often in practical settings agents offer discounts based
on the amount of goods being purchased from them i.e. purchasing more goods from a
single agent results in higher discounts. Apart from such practical considerations discount
functions have strong theoretical motivation too. Recent work [10] pertaining to learning
1
submodular functions has lead to succicnt representations for submodular functions. These
representations also behave like discounted price functions.
Discounted Price Model
We define a function d : R+ → R+ to be a discounted price function if it satisfies the
following properties: (1) d(0) = 0; (2) d is increasing; (3) d(x) ≤ x for all x; (4) d is
concave.
We study combinatorial problems in the following general setting. We are given a set
of elements, E, a collection Ω of its subsets. We are also given a set A of k agents where
each agent, a ∈ A specifies a cost function ca : E → R+ where ca(e) indicates her cost for
the element e. Each agent also declares a discounted price function, da. If an agent a is
e∈T ca(e)). This is
called her discounted price. The objective is to select a subset S from Ω and a partition
assigned a set of elements T , then her total price is specified by da((cid:80)
S1, S2, ..., Sk of S, such that(cid:80)
a∈A da((cid:80)
e∈Sa
ca(e)) is minimized.
Our frame work generalizes the classical single-agent linear cost setting. On the other
hand, (4) implies that discounted price functions are satisfy the decreasing marginal cost
property. Moreover, since any concave function can be well approximated by a piecewise
linear function, discounted price functions can be viewed as a class of succinctly represented
submodular functions.
We study the following four problems over an undirected complete graph G = (V, E).
• Discounted Edge Cover: In this problem, Ω is chosen to be the collection of edge
covers.
• Discounted Spanning Tree: In this problem, Ω is the collection of spanning trees
of the graph.
• Discounted Perfect Matching: In this problem, we assume the graph has an even
number of vertices. Ω is chosen to be the collection of perfect matchings of the graph.
• Discounted Shortest Path: In this problem, we are given two fixed vertices s and
t. Ω consists of all paths connecting s and t.
Our Results
We summarize our results in the following theorem:
Theorem 1.1. There is an O(log n) approximate algorithm for each of the following prob-
lems: discounted edge cover,spanning tree,perfect matching and shortest path, where n is
the number of vertices in their instances. Moreover, it is hard to approximate any of them
within factor (1 − o(1)) log n unless NP=DTIME(nO(log log n)).
Our main technical contribution is the introduction of a new algorithmic technique
called adaptive greedy algorithms. In traditional greedy algorithms like set cover, we make
2
the most 'efficient' decision at every step to inch closer to the objective. However, these
decision are irreversible and cannot be modified in subsequent iterations. For example in
the greedy set cover algorithm, if a set is chosen in a particular iteration it will surely be
included in the solution. This approach works well for problems where the solution does
not need to satisfy combinatorial properties, but fails in the combinatorial setting. This is
because, choosing a particular set may preclude the choice of other efficient sets in future
iterations owing to the combinatorial constraints. To rectify this problem we allow our
algorithm to undo some of the decisions taken during previous iterations. We call such
algorithms adaptive greedy algorithms. Another important contribution of the paper is to
link the shortest path problem to the minimum weight perfect matching problem by a factor
preserving reduction.
From the hardness perspective, we consider the discounted reverse auction problem in
which the collection is devoid of any combinatorial constraints. Using a factor preserving
reduction from the set cover problem we prove that it is hard to approximate the discounted
reverse auction within factor (1 − o(1)) log n unless NP=DTIME(nO(log log n)). We use
discounted reverse auction as a starting point to prove the hardness of all the combinatorial
problems considered in this paper.
From the above argument, one would expect that our combinatorially structured prob-
lems might have weaker lower bounds than discounted reverse auction. However, interest-
ingly, we find it is not the case, as shown by the log approximate algorithms in the section
2.
Related work
The classical versions of edge cover, spanning tree, perfect matching and shortest path
are well studied and polynomial time algorithms are known for all these problems, see
[17, 13, 2, 6, 5, 3]. These have served as part of the most fundamental combinatorial
optimization model in the development of computer science and operation research.
Recently, Svitkina and Fleischer [18] generalized the linear cost settings of some com-
binatorial optimization problems such as sparsest cut, load balancing and knapsack to a
submodular cost setting. They gave(cid:112)n/ log n upper and lower bounds for all these prob-
lems. [11] also studied combinatorial optimization problems in the single agent setting.
Multi-agent setting was first introduced in [9]. In their model of combinatorial problems
with multi-agent submodular cost functions they gave Ω(n) lower bounds for the problems
of submodular spanning tree and perfect matching, and gave a Ω(n2/3) lower bound for
submodular shortest path. They also gave the matching upper bounds for all these problems.
We remark that the lower bounds presented in [18, 9] are information theoretic and not
computational.
Similar generalization from linear objective function to the more general submodular
function was applied to maximization problems in [8, 14, 1]. The multi-agent generalization
of maximization problems which corresponds to combinatorial auction has been extensively
studied both in computer science and economics [4, 16] and tight information lower bounds
3
are known for these problems, see [15].
2 Algorithms for Discounted Combinatorial Optimization
In this section we present approximation algorithms for the combinatorial problems defined
earlier. In each of the problems we are given a complete graph G = (V, E) over n vertices.
We are also given a set A, of k agents each of whom specifies a cost ca : E → R+. Here
ca(e) is the cost for building edge e for agent a. Each agent also specifies a discounted price
function given by da : R+ → R+. The objective is to build a prespecified combinatorial
structure using the edges in E, and allocate these edges among the agents such that the
sum of discounted prices for the agents is minimized.
2.1 Discounted Edge Cover
In this section, we study the discounted edge cover problem and establish a log n approxi-
mate algorithm for this problem.
Here is the basic idea behind our algorithm. Given a discounted edge cover instance,
we construct a set cover instance such that: (1) an optimal edge cover corresponds to a set
cover with the same cost and (2) a set cover corresponds to an edge cover with a smaller
price. In the set cover instance, we apply the greedy algorithm from [12] to get a set cover
whose cost is within log n of the optimal cost, then the corresponding edge cover gives an
log n approximation of the optimum edge cover. We remark that we will have exponentially
many sets in the set cover instance that we construct for our problem. To apply the greedy
algorithm, we need to show that in each step, the set with the lowest average cost can be
found in polynomial time.
Now we state our algorithm formally. Consider a set cover instance where we have to
cover the set of vertices, V , with k2n subsets which are indexed by (a, S) ∈ A × 2V . The
cost of the set (a, S), denoted by cost(a, S), is defined as the minimum discounted price of
an edge cover for the vertices in S that can be built by agent a. For the instance of set cover
described above, we apply the greedy algorithm [12] to get a set cover S. Let Ua be the set
of vertices covered by sets of the form (a, S) ∈ S. Let Ca be agent a's optimal discounted
edge cover of the vertices in Ua. We output {Ca : a ∈ A} as our solution.
The correctness of the algorithm follows from the observation that each Ca is a cover of
Sa thus their union must form an edge cover of the vertex set V .
Now we show that the running time of the algorithm is polynomial in k and n. Notice
that when Ua is given, since agent a's discounted price function is increasing, Ca can be
found in polynomial time. Hence we only need to show that the greedy algorithm on our
set cover instance can be done in polynomial time.
Recall that the greedy algorithm from [12] covers the ground set, V , in phases. Let Q
be the set of covered vertices in the beginning of a phase and the average cost of a set (a, S)
is defined as αa(S) = cost(a, S)/S − Q. The algorithm picks the set with the smallest
4
average cost, covers the corresponding vertices in that set and go to the next phase until
all the vertices are covered. To show that this algorithm can be implemented efficiently, we
only need to show the following:
Lemma 2.1. For any Q ⊂ V , we can find min{cost(a, S)/S − Q : (a, S) ∈ A × 2V } in
polynomial time.
Proof. We can iterate over all choices of agent a ∈ A, thus the problem boils down to finding
min{cost(a, S)/S − Q : S ⊆ V } for each a ∈ A.
For each integer d, if we can find min{cost(a, S) : S − Q = d} in polynomial time, then
we are done since then we can just search over all the possible sizes of S− Q. Unfortunately,
it is NP-hard to compute min{cost(a, S) : S − Q = d} for all integer d. We will use claim
1 to circumvent this problem.
Claim 1. For any graph G = (V, E) and Q ⊆ V and for any positive integer d, we can find
the set (a, S) minimizing cost(a, S) such that S − Q is at least d, in polynomial time.
Proof. Refer to Appendix A1.
By claim 1 above we can generate a collection of subsets {Si ⊆ V : 1 ≤ i ≤ n}, such
that (a, Si) has the lowest value of cost(a, S) among all sets S which satisfy S − Q ≥ i.
Claim 2. min{cost(a, S)/S − Q : S ⊆ V } = min{cost(a, Si)/Si − Q : 1 ≤ i ≤ n}.
Proof. Let ¯S be the set that has the minimum average cost with respect to agent a. Suppose
¯S−Q = d. By our choice of Sd, we have Sd−Q ≥ d = ¯S−Q and cost(a, Sd) ≤ cost(a, ¯S).
Therefore we have cost(a, Sd)/Sd − Q ≤ cost(a, ¯S)/ ¯S − Q, hence they must be equal.
By iterating over all a ∈ A, we can find min{cost(a, S)/S − Q : (a, S) ∈ A × 2V } in
polynomial time.
Next we show that the approximation factor of our algorithm is log n. Let OP TEC
and OP TS denote the costs of the optimal solutions for the discounted edge cover instance
and the corresponding set cover instance respectively. Let {Ca : a ∈ A} be the edge cover
reported by our algorithm. For all a ∈ A let Oa be the set of vertices covered by agent
a in the optimal edge cover. The sets {(a, Oa) : a ∈ A} form a solution for the set cover
instance. Therefore OP TS ≤ OP TEC.
Since we use the greedy set cover algorithm to approximate OP TS, we have
(cid:88)
a∈A
da((cid:88)
e∈Ca
ca(e)) ≤ (log n)OP TS ≤ (log n)OP TEC
Thus we have the following theorem.
Theorem 2.2. There is a polynomial time algorithm which finds a log n-approximate so-
lution to the discounted edge cover problem for any graph over n vertices.
5
2.2 Discounted Spanning Tree
In this section, we study the discounted spanning tree problem and establish a log n ap-
proximation algorithm for this problem.
Let us first consider a simple O(log2 n)-approximation algorithm. Observe that a span-
ning tree is an edge cover with the connectivity requirement. If we apply the greedy edge
cover algorithm in section 2.1, there is no guarantee that we will end up with a connected
edge cover. We may get a collection of connected components. We can subsequently con-
tract these components and run the greedy edge cover algorithm again on the contracted
graph. We keep doing this until there is only one connected component. By this method,
we will get a connected edge cover, i.e. a spanning tree.
We now analyze the above algorithm. Let OP TST be the price of the minimum dis-
counted spanning tree. After each execution of the greedy edge cover algorithm, there is no
isolated vertex, hence the contraction decreases the number of vertices by at least a a factor
of half, therefore we will have to run the greedy edge cover algorithm at most O(log n)
times. Let OP T r
EC be the price of the minimum edge cover for the graph obtained after
the rth contraction and let Cr be the edge cover that we produce for this iteration. Using
EC is
theorem 2.2, the price of Cr is at most (log n)OP T r
at most OP TST for every r. Hence the price of Cr is bounded by (log n)OP TST . Since
there are at most O(log n) iterations, the price of the spanning tree produced by the above
algorithm is bounded by O(log2 n)OP TST .
EC. It is easy to see that OP T r
We observe that greedy edge cover and contraction are two main steps in the above
algorithm. Intuitively, they are used to satisfy the covering and connectivity requirements
respectively. The algorithm proceeds by alternately invoking these subroutines. Based on
this observation, our idea to get a log n approximation algorithm is to apply the following
adaptive greedy algorithm: rather than apply contraction after each complete execution of
the greedy edge cover, we interleave contraction with the iterations of the greedy edge cover
algorithm. After each iteration, we modify the graph and motivate our algorithm to get a
connected edge cover at the end.
Now we describe our adaptive greedy algorithm. For every agent a and subset of vertices
S we define cost(a, S) as the cost of the optimal edge cover for S. We define the average cost
a = cost(a, S)/S. The algorithm proceeds in phases and in each phase
of a set (a, S) as αS
we have two steps, search and contraction. In each phase r, during the search step we
find the set (ar, Sr) with the lowest average cost and set the potential of each vertex v ∈ Sr
as p(v) = αSr
ar . The search step is followed by a contraction step, where we modify the graph
by contracting every connected component in the induced subgraph of agent ar's optimal
edge cover for the set Sr. After this we begin the next phase. The algorithm terminates
when we have contracted the original graph to a single vertex. For every agent, we find
the set of all edges assigned to her across all the search steps declare this as her bundle of
assigned edges. Finally remove unnecessary edges from the set of assigned edges to get a
spanning tree.
It is easy to see that we get a connected edge cover at the end of the algorithm, which
proves the correctness of the algorithm. To analyze the running time, we observe that there
6
can be at most n phases and by Lemma 2.1, each phase can be implemented in polynomial
time. Hence the algorithm runs in polynomial time.
Next, we prove that the approximation factor of the algorithm is O(log n). Let OP TST
be the price of the optimal solution of the discounted spanning tree instance and let {Ta :
a ∈ A} be our solution. Let V (cid:48) be the set of contracted vertices we produced during the
algorithm. Number the elements of V and V (cid:48) in the order in which they were covered by
the algorithm, resolving ties arbitrarily. Suppose V = {v1, ..., vn} and V (cid:48) = {z1, ..., zn(cid:48)}.
Obviously, n(cid:48) ≤ n.
It is easy to verify that (cid:80)
a da(Ta) ≤ (cid:80)
i p(vi) +(cid:80)
j p(zj). Therefore we only need to
bound the potentials of the vertices in V ∪ V (cid:48).
Claim 3. p(vi) ≤ OP TST
n−i+1 for any i ∈ {1··· n} and p(zj) ≤ OP TST
n(cid:48)−j
for any j ∈ {1··· n(cid:48)}.
Proof. Refer to Appendix A3.
(cid:88)
da(Ta) ≤ (cid:88)
a
OP TST
n − i + 1
1≤i≤n
From the above claim, we have
+ (cid:88)
1≤j≤n(cid:48)
OP TST
n(cid:48) − j
≤ (log n + log n(cid:48))OP TST ≤ O(log n)OP TST
Therefore we have the following:
Theorem 2.3. There is a polynomial time algorithm which finds an O(log n)-approximate
solution to the discounted spanning tree problem for any graph with n vertices.
2.3 Discounted Perfect Matching
In this section we provide a log n approximate algorithm for the discounted (minimum)
perfect matching problem.
The main difficulty in tackling matching is that unlike spanning tree or edge cover, the
union of perfect matchings for two overlapping subsets of vertices may not contain perfect
matching over their union. So a greedy algorithm which iteratively matches unmatched
vertices cannot be used to solve this problem.
Here we provide an adaptive greedy algorithm for solving this problem. In our algorithm
we successively match unmatched vertices. Once a vertex gets matched, it remains matched
through the course of the algorithm. However unlike traditional greedy algorithms, we
provide the algorithm the ability to change the underlying matching for previously matched
vertices. The algorithm is adaptive in the sense that it alters some of the decisions taken
in previous phases to reduce the price incurred in the current phase.
Let us explain the algorithm in greater detail now. The algorithm runs in phases and
in each phase it augments the set of matched vertices by adding unmatched vertices and
rematching some of the previously matched vertices. At the end of each phase let Z denote
the set of matched vertices and let M be the underlying matching covering Z. For every
7
a ∈ A we define cM
a : E → R+ which attains value 0 for all edges in M and takes the value
ca(e) for all other edges e /∈ M. For any F ⊆ E let V (F ) be the set of vertices in the graph
induced by F . For a given matching M, we define the average cost of a set (a, F ) ∈ A× 2E
with respect to M as αM
a (F )/V (F ) − Z.
a (F ) = cM
In the beginning of every phase we pick the element (a, F ) from A× 2E with the lowest
average cost such that there exists a perfect matching over Z ∪ V (F ) using edges from
M ∪ F . We defer details of this subroutine until the proof of lemma 2.4. For each vertex
v ∈ V (F ) − Z we define its potential p(v) = αM
a (F ). Finally we augment the set Z with
vertices in V (F ) and update the underlying matching covering it. The algorithm terminates
when all vertices have been matched, i.e. belong to Z.
We will now show that the algorithm can be implemented in polynomial time. For this
we will need to prove the following lemma.
Lemma 2.4. Given a matching M and Z = V (M), we can find, in polynomial time, the
element (a, F ) from A × 2E with the lowest average cost with respect to M such that there
exists a perfect matching over Z ∪ V (F ) using edges from M ∪ F .
Proof. Refer to Appendix A2.
To analyze the approximation factor for the algorithm, we number the vertices in V in
the order in which they are added to Z and resolve ties arbitrarily. Let V = {v1, v2 ··· vn}.
Let OP T be the price of the optimal solution. We have the following lemma to bound the
potential of every vertex.
Lemma 2.5. For each i ∈ {1, 2··· n} , p(vi) ≤ OP T /(n − i + 1).
Proof. Similar to the proof of Claim 3.
Since the price of our solution is at most the sum of the potentials for the vertices the
following theorem follows immediately from lemma 2.5
Theorem 2.6. There exists a log n approximate algorithm for the discounted perfect match-
ing problem.
2.4 Shortest s-t Path
In this section we consider the discounted shortest path problem between two given vertices
s and t. We provide a log n approximation algorithm for this problem.
Unlike any of the previous problems, it seems to difficult to design a greedy algorithm
for the discounted shortest path problem. However using ideas developed in section 2.3
and by a factor preserving reduction we obtain a log n approximate algorithm this problem.
Our reduction also implies that the shortest path problem is easier than the minimum
weight perfect matching problem, and this technique can be applied to very general models
whereby any algorithm for perfect matching problem can be used to solve the shortest path
problem while preserving the approximation factor.
8
First of all, we describe our reduction:
Lemma 2.7. Let A be a β-approximate algorithm for the perfect matching problem, then
we can get a β-approximation for the shortest path problem using A as a subroutine
Proof. Suppose we are given a graph G = (V, E). Construct an auxiliary graph ¯G in the
following way: Replace every vertex v ∈ V by v(cid:48) and v(cid:48)(cid:48) and add an edge connecting them.
The cost of this edge is zero for every agent. For every edge uv ∈ E where u, v /∈ {s, t} we
replace uv with the gadget shown in figure 1.
Figure 1: Gadget for edges not incident on s or t
If d = ca(uv) was the cost of the edge uv for agent a ∈ A then two of the edges in our
gadget have the same cost as shown above. Rest of the edges have zero cost for all agents.
Replace any edge su ∈ E with the gadget shown in figure 2. Do the same for any edge ut.
Figure 2: Gadget for edges incident on s or t
On this graph ¯G, use the algorithm A to get the minimum cost matching. Let M be
the matching returned. We can interpret M as a s − t path in G in the following way. Let
g(uv) be the edges in ¯G corresponding to the edge uv for the gadget shown in figure 1.
Observe that either one or two edges of every such gadget must belong to M. Let S be the
set of edges in G such that two edges in their corresponding gadget belong to M. One can
check that every vertex in V is incident with zero or two edges from S, whereas s and t
are each incident with exactly one edge in S. Therefore S consists of an s − t path PS and
some other circuits. Now the circuits in S must have cost zero. This is because if a circuit
has positive cost then the cost of the matching can be reduced further by pairing up the
vertices in the circuit as shown in figure 3.
Figure 3: Circuits not involving edges in S should have zero costs
9
Note that this defines a cost preserving bijection between s − t paths in G to perfect
matchings in ¯G.
Using the algorithm from section 2.3, by Lemma 2.7, we have:
Theorem 2.8. There exists a log n approximate algorithm for the discounted shortest path
problem.
3 Hardness of Approximation
In this section we present hardness of approximation results for the problems defined earlier.
Unlike some of the previous work on combinatorial optimization [9, 18] over non-linear cost
functions, the bounds presented here are not information theoretic but are contingent on
N P = DT IM E(nO(log log n)). It should be noted that the discount functions for the agents
are the only part of the input that is not presented explicitly. However it is well known
that concave functions that accessible through an oracle can be closely approximated by a
piecewise linear function using few queries to the oracle. So, we are not handicapped by
the lack of information from solving these problems optimally.
To show the hardness of approximation of the problems stated earlier we consider the
following general problem and use a reduction from set cover to establish its hardness of
approximation.
Discounted Reverse Auction: We are given a a set E of n items and a set A of agents
each of whom specifies a function ca : E → R+. Here ca(e) is the cost for procuring item e
from agent a. Each agent also specifies a discounted price function given by da : R+ → R+.
ca(e)) is
The task is to find a partition P = {P1 ··· Pk} of E such that (cid:80)
a∈A da((cid:80)
e∈Pa
minimized.
Lemma 3.1. It is hard to approximate the discounted reverse auction problem within factor
(1 − o(1)) log n unless N P = DT IM E(nO(log log n)).
Proof. Refer to Appendix A4
This reduction can be extended to other combinatorial problems in this setting to give
logarithmic hardness of approximation for many combinatorial problems. This can be
achieved by considering an instance of the problem where we have just one combinatorial
object and our task is the allocate it optimally among the agents. For example for the
discounted spanning tree problem we consider the instance when the input graph is itself
a tree and we have to optimally allocate its edges among the agents to minimize the total
price. Thus, we have the following:
Theorem 3.2. It is hard to approximate any of the problems of the discounted edge cover,spanning
tree,perfect matching and shortest path can be approximated within factor (1 − o(1)) log n
on a graph over n vertices unless N P = DT IM E(nO(log log n)).
10
References
[1] Gruia Calinescu, Ra Chekuri, Martin Pl, and Jan Vondrk. Maximizing a submodular
set function subject to a matroid constraint. In Proc. of 12 th IPCO, page 2007, 2007.
[2] William Cook and Andr Rohe. Computing minimum-weight perfect matchings. IN-
FORMS Journal on Computing, 11:138 -- 148, 1998.
[3] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein. Intro-
duction to Algorithms. The MIT Press, 2001.
[4] Sven de Vries and Rakesh V. Vohra. Combinatorial auctions: A survey. INFORMS
Journal on Computing, (3):284 -- 309, 2003.
[5] Edsger W. Dijkstra. A note on two problems in connexion with graphs. Numerische
Mathematik, 1:269 -- 271, 1959.
[6] Jack Edmoonds. Paths,trees and flowers. Canad J.Math, 17, 1965.
[7] Uriel Feige. A threshold of ln n for approximating set cover. Journal of the ACM,
45:314 -- 318, 1998.
[8] Uriel Feige and Vahab S. Mirrokni. Maximizing non-monotone submodular functions.
In In Proceedings of 48th Annual IEEE Symposium on Foundations of Computer Sci-
ence (FOCS, page 2007, 2007.
[9] Gagan Goel, Chinmay Karande, Pushkar Tripathi, and Lei Wang. Approximability
of combinatorial problems with multi-agent submodular cost functions. In FOCS '09:
Proceedings of the 2009 50th Annual IEEE Symposium on Foundations of Computer
Science, 2009.
[10] M.X. Goemans, N.J.A. Harvey, S. Iwata, and V. Mirrokni. Approximating submodular
functions everywhere. In SODA '09: Proceedings of the 20th ACM-SIAM Symposium
on Discrete Algorithms, 2009.
[11] Satoru Iwata and Kiyohito Nagano. Submodular function minimization under covering
constraints. In FOCS '09: Proceedings of the 2009 50th Annual IEEE Symposium on
Foundations of Computer Science, 2009.
[12] Johnson and David S. Approximation algorithms for combinatorial problems. In STOC
'73: Proceedings of the fifth annual ACM symposium on Theory of computing, pages
38 -- 49, New York, NY, USA, 1973. ACM.
[13] Joseph B. Kruskal. On the shortest spanning subtree of a graph and the traveling
salesman problem. Proceedings of the American Mathematical Society, 7(1):48 -- 50,
February 1956.
11
[14] Jon Lee, Vahab S. Mirrokni, Viswanath Nagarajan, and Maxim Sviridenko. Non-
monotone submodular maximization under matroid and knapsack constraints.
In
STOC '09: Proceedings of the 41st annual ACM symposium on Theory of comput-
ing, pages 323 -- 332, New York, NY, USA, 2009. ACM.
[15] Vahab Mirrokni, Michael Schapira, and Jan Vondrak. Tight information-theoretic lower
bounds for welfare maximization in combinatorial auctions. In EC '08: Proceedings of
the 9th ACM conference on Electronic commerce, pages 70 -- 77, New York, NY, USA,
2008. ACM.
[16] Noam Nisam, Tim Roughgarden, ´Eva Tardos, and Vijay V. Vazirani. Algorithmic
Game Theory. Cambridge University Press, September 2007.
[17] Seth Pettie and Vijaya Ramachandran. An optimal minimum spanning tree algorithm.
J. ACM, 49(1):16 -- 34, 2002.
[18] Zoya Svitkina and Lisa Fleischer. Submodular approximation: Sampling-based algo-
rithms and lower bounds. In FOCS '08: Proceedings of the 2008 49th Annual IEEE
Symposium on Foundations of Computer Science, pages 697 -- 706, Washington, DC,
USA, 2008. IEEE Computer Society.
Appendix
A1 Proof of Claim 1
To find the desired set we construct a graph G(cid:48) = (V (cid:48), E(cid:48)) as follows: Add a set X ∪ Y to
the set of vertices in G, where X = Q and Y = n− d. Match every vertex X to a vertex
Q with an edge of cost 0. Connect each vertex in Y to each vertex in V by an edge of very
large cost. Set the cost of each edge e ∈ E as ca(e). Find the minimum cost edge cover in
G(cid:48). Let S∗ be such a cover. It is easy to verify that S∗ is the desired set.
A2 Proof of Lemma 2.4
We may assume that the agent a, and the size of V (F ) − Z in the formula for αM
a (F ) are
fixed by iterating over all values of these parameters. So the above statement may be viewed
as one concerning a single agent for a fixed value of V (F ) − Z. Let this agent be a ∈ A
and let t = V (F ) − Z.
i.e. We wish to find the minimum discounted weight matching
which saturates t new vertices(not in Z) and keeps the vertices in Z saturated. We create
an auxiliary graph G(cid:48) by adding n − (Z + t) vertices to V . We add edges connecting each
of the newly added vertex with each vertex in V − Z. Let E(cid:48) be the set of newly added
a by setting the cost of edges in E(cid:48) to be 0. Finally we find
edges. We extend cM
a
the minimum weight perfect matching in G(cid:48) under the cost function cM
a . Let M(cid:48) be this
matching. It is easy to check that M(cid:48)− E(cid:48) is the desired matching saturating t new vertices.
to cM
12
A3 Proof of Claim 3
For i ∈ {1··· n}, suppose vi is covered in phase r. Let Gr be the underlying graph at the
beginning of phase r. Since vi, vi+1, ..., vn are not covered before phase r, Gr contains at
least n − i + 1 vertices. Since the optimal spanning tree can cover the vertices in Gr by a
price of OP TST , by our greedy choice, p(vi) ≤ OP TST /(n − i + 1).
Similarly, let 1 ≤ j ≤ n(cid:48) and assume zj is covered in phase r. Since we should be able
to produce zj+1, zj+2, ..., zn(cid:48) from contraction on vertices of Gr, there are at least n(cid:48) − r
vertices in Gr. Therefore we have p(zj) ≤ OP TST /(n(cid:48) − j).
A4 Proof of Lemma 3.1
We reduce set cover to the discounted reverse auction problem to prove this result.
Consider an instance I = (U, C, w) of set cover where we wish to cover all elements
in the universe U using sets from C to minimize the sum of weights under the weight
function w : U → R+. We define an instance, I(cid:48) of our discounted reverse auction problem
corresponding to I in the following way. Let U be the set of items. For every set S ∈ C
define an agent aS, whose cost function ca assigns the value w(S) for every element s ∈ S
and sets the cost of all other elements in U to be infinity. The discount price function for
the agent is shown in figure A1. Here the slope of the second segment is small enough.
Figure A1: Discount function for agent corresponding to set S
Consider a solution for I(cid:48) where we procure at least one item from agent aS; then we
can buy all elements in S from aS without a significant increase in our payment. So the
cost of the optimal solution to I can be as close to the price of the optimal solution for
I(cid:48) as we want. By [7], set cover is hard to approximate beyond a factor of log n unless
N P = DT IM E(nO(log log n)). Therefore the discounted reverse auction problem can not be
approximated within factor (1 − o(1)) log n unless N P = DT IM E(nO(log log n)).
13
|
1909.01741 | 1 | 1909 | 2019-09-04T12:46:08 | B\"uchi automata for distributed temporal logic | [
"cs.MA"
] | The distributed temporal logic DTL is a logic for reasoning about temporal properties of distributed systems from the local point of view of the system's agents, which are assumed to execute sequentially and to interact by means of synchronous event sharing. Different versions of DTL have been provided over the years for a number of different applications, reflecting different perspectives on how non-local information can be accessed by each agent. In this paper, we propose a novel notion of distributed B\"uchi automaton envisaged to encompass DTL with a model-checking mechanism. | cs.MA | cs |
Buchi automata for distributed
temporal logic
Jaime Ramos
Dep. Matem´atica, Instituto Superior T´ecnico, Universidade de Lisboa, Portugal
SQIG, Instituto de Telecomunica¸coes, Portugal
[email protected]
September 5, 2019
Abstract
The distributed temporal logic DTL is a logic for reasoning about
temporal properties of distributed systems from the local point of view
of the system's agents, which are assumed to execute sequentially and
to interact by means of synchronous event sharing. Different versions of
DTL have been provided over the years for a number of different applica-
tions, reflecting different perspectives on how non-local information can
be accessed by each agent. In this paper, we propose a novel notion of
distributed Buchi automaton envisaged to encompass DTL with a model-
checking mechanism.
Keywords: Distributed Temporal Logic (DTL), Buchi automata, dis-
tributed systems, specification and verification, model-checking.
1
Introduction
The distributed temporal logic DTL was introduced in [13] as a logic for speci-
fying and reasoning about distributed information systems. DTL allows one to
reason about temporal properties of distributed systems from the local point of
view of the system's agents, which are assumed to execute sequentially and to
interact by means of synchronous event sharing. In DTL, distribution is implicit
and properties of entire systems are formulated in terms of the local properties
of the agents and their interaction. The logic was shown to be decidable, as well
as trace-consistent, which makes it suitable for model-checking tasks.
Different versions of distributed temporal logic have been given over the
years for a number of different applications, reflecting different perspectives on
how non-local information can be accessed by each agent. In particular, DTL
has proved to be useful in the context of security protocol analysis in order to
reason about the interplay between protocol models and security properties [8,
9, 5]. However, most of the results for security protocol analysis and for other
case studies were obtained directly by semantic arguments.1 To overcome this
1DTL is closely related to the family of temporal logics whose semantics are based on
the models of true concurrency introduced and developed in [17, 18, 21]. In particular, the
1
problem, a labeled tableaux system for DTL was proposed in [3, 4]. The main
goal was to have a usable deductive system in which deductions followed closely
semantic arguments, also thanks to the labeling of the formulas along with a
labeling algebra capturing the different semantic properties.
The labeled tableaux system was proved to be sound and complete, but
decidability was not considered in [3, 4] and the system included an infinite
closure rule to capture eventualities that are always delayed. Hence, the labeled
system proved to be quite hard to use in practice although several properties
can still be proved using only the tableaux system. For instance, the correctness
of the two-phase commit protocol
is one of such examples where a decision
procedure is not needed. The DTL specification for a simplified version of the
protocol as well as a proof of correctness using labelled tableaux can be found
in [4].
Nevertheless, DTL was shown to be decidable via a translation to linear
temporal logic (LTL). However, when translating DTL specifications into LTL
specifications, we lose one of the main advantages of DTL, namely the natural-
ness of the distributed nature of DTL, which allows for more natural and simpler
specifications. Later, in [7], a decidable tableaux system was proposed for DTL.
The tableaux system was built on top of a tableaux system for LTL as presented
in [15]. Similar systems for LTL have also been proposed, e.g., [16]. In the case
of DTL, the tableaux system integrated in a smooth way both the usual rules
for the temporal operators and rules for tackling the specific communication
features of DTL.
In this paper, we take a first step towards empowering DTL with model-
checking tools. Nowadays, systems are becoming more and more complex which
makes the task of verification such systems harder. Model-checking stands out
as a tool well suited for automatic verification, which has been successfully
used in industry with several well documented cases [19, 14, 2]. Depending
on the temporal logic considered [20, 11], the approach to model-checking is
different[10, 12, 22]. In the case of DTL, we adopt an approach closer to the
usual approaches in LTL, based on Buchi automata [22]. Our goal is to use
Buchi automata to capture DTL models.
For the local component of our automata, we follow closely the ideas in [22,
It is worth mentioning that, similar to [1], in which an anchored version
1].
of LTL is considered, in this paper we consider an anchored version of DTL.
This anchored version of DTL is less expressive in terms of global reasoning
since DTL does not include global temporal operators and, thus, we cannot use
the usual correspondence between anchored and floating semantics of temporal
logic. However, it let us focus on the distributed nature and synchronization
primitives of the logic.
We proceed as follows. In Section 2, we briefly introduce DTL, its syntax,
semantics and some auxiliary notions that will be useful later. In Section 3, we
present distributed Buchi automata for DTL and prove the correctness of the
construction with respect to the semantics of DTL. In Section 4, we conclude
and discuss future work.
semantics of these logics are based on a conflict-free version of Winskel's event structures [23],
enriched with information about sequential agents.
2
2 The Distributed Temporal Logic DTL∅
As we mentioned above, a number of variants of DTL have been considered in
the past, especially to adapt it to specific applications and case studies. In this
paper, we consider an anchored variant of DTL that we call DTL∅ and that has
the following syntax and semantics.
2.1 Syntax
The logic is defined over a distributed signature
Σ = hId , {Prop}i∈Id i ,
where Id is a finite non-empty set (of agent identifiers) and, for each agent
i ∈ Id , Prop i is a set of local state propositions, which, intuitively characterize
the current local states of the agents. We assume that Prop i ∩ Prop j = ∅, for
i 6= j.
The local language Li of each agent i ∈ Id is defined by
Li ::= Prop i ¬ Li Li ⇒ Li X Li G Li c(cid:13)j[Lj]
with j ∈ Id . We will denote such local formulas by the letters ϕ and ψ. As the
names suggests, local formulas hold locally for the different agents. For instance,
locally for an agent i, the operators X and G are the usual next (tomorrow) and
always in the future temporal operators, whereas the communication formula
c(cid:13)j[ψ] means that agent i has just communicated (synchronized) with agent j,
for whom ψ held.
Other logical connectives (conjunction ∧, disjunction ∨, true ⊤, etc.) and
temporal operators (sometime in the future F) can be defined as abbreviations
as is standard.
The global language L is defined by
L ::= @i[Li] ¬ L L ⇒ L
with i ∈ Id . We will denote the global formulas by α, β and δ. A global formula
@i[ϕ] means that the local formula ϕ holds for agent i.
In the sequel, we will need some auxiliary notions. The set of i-literals is the
set of all state propositions and their negations:
Liti = Prop i ∪ {¬ p p ∈ Prop i}.
An i-valuation v is a set of i-literals such that for each p ∈ Prop i, p ∈ v iff
¬ p /∈ v. The set of all i-valuations is denoted by Vi. Observe that Vi ⊆ 2Liti .
Given i ∈ Id and ϕ, ψ ∈ Li, we say that ψ is an i-subformula of ϕ if ψ is ϕ
or:
• ϕ is ¬ ϕ1 and ψ is a subformula of ϕ1;
• ϕ is ϕ1 ⇒ ϕ2 and ψ is a subformula of ϕ1 or of ϕ2;
• ϕ is X ϕ1 and ψ is a subformula of ϕ1;
• ϕ is G ϕ1 and ψ is a subformula of ϕ1.
3
We denote by subfi(ϕ) the set of all i-subformulas of ϕ. When no confusion
arises, we drop the reference to i and talk about subformulas. For instance, the
set of subformulas of G(p ⇒ c(cid:13)j[q1 ⇒ q2]) ∈ Li, subfi(G(p ⇒ c(cid:13)j[q1 ⇒ q2])), is
{G(p ⇒ c(cid:13)j[q1 ⇒ q2]), p ⇒ c(cid:13)j[q1 ⇒ q2], p, c(cid:13)j[q1 ⇒ q2]}.
Note that, from the point of view of agent i, formula c(cid:13)j[q1 ⇒ q2] has no further
structure and is treated as atomic.
The i-closure of ϕ is the set of all its subformulas and their negations with
the proviso that ¬ ¬ ψ is identified with ψ, that is,
closurei(ϕ) = subfi(ϕ) ∪ {¬ ψ ψ ∈ subfi(ϕ)}.
Again, when no confusion arises, we will talk about the closure of a formula.
We also define similar concepts for the global language. Given α, β ∈ L, we
say that β is a subformula of α if β is α or:
• α is @i[ψ] and β ∈ subfi(ϕ)
• α is ¬ α1 and β is a subformula of α1;
• α is α1 ⇒ α2 and β is a subformula of α1 or of α2.
For instance, the set of subformulas of formula @i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q],
subf(@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q]), is:
{@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], @i[X(p ⇒ c(cid:13)j[q])], @j[X q]}
∪ subfi(X(p ⇒ c(cid:13)j[q])) ∪ subfj(X q) =
{@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], @i[X(p ⇒ c(cid:13)j[q])], @j[X q]}
∪ {X(p ⇒ c(cid:13)j[q]), p ⇒ c(cid:13)j[q], p, c(cid:13)j[q], X q, q}.
The closure of α is the set of all its subformulas and their negations with
the proviso that ¬ ¬ β is identified with β, that is,
closure(α) = subf(α) ∪ {¬ β β ∈ subf(α)}.
Finally, given a set B ⊆ closure(α) and i ∈ Id we denote B↓i the subset of
B that contains all the global formulas of B and no local formulas other than
those of agent i, that is, B↓i satisfies the following conditions:
• B↓i⊆ B;
• B↓i ∩L = B ∩ L;
• B↓i ∩Li = B ∩ Li;
• B↓i ∩Lj = ∅, for j 6= i.
For instance, if B = {@i[X(p ⇒ c(cid:13)j[q])]⇒@j[X q], X(p⇒ c(cid:13)j[q]), c(cid:13)j[q], X q} then
B↓i= {@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], X(p ⇒ c(cid:13)j[q]), c(cid:13)j[q]}
and
B↓j= {@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], X q}.
4
2.2 Semantics
The interpretation structures of L are labeled distributed life-cycles, built upon
a simplified form of Winskel's event structures [23].
A local life-cycle of an agent i ∈ Id is a countable infinite, discrete, and well-
founded total order λi = hEv i, ≤ii, where Ev i is the set of local events and ≤i
the local order of causality. We define the corresponding local successor relation
→i ⊆ Ev i × Ev i to be the relation such that e →i e′ if e <i e′ and there is no
e′′ such that e <i e′′ <i e′. As a consequence, we have that ≤i = →∗
i , i.e., ≤i is
the reflexive and transitive closure of →i.
A distributed life-cycle is a family λ = {λi}i∈Id of local life-cycles such that
≤= (Si∈Id ≤i)∗ defines a partial order of global causality on the set of all events
Ev =Si∈Id Ev i.
Communication is modeled by event sharing, and thus for some event e we
may have e ∈ Ev i ∩ Ev j, with i 6= j. In that case, requiring ≤ to be a partial
order amounts to requiring that the local orders are globally compatible, thus
excluding the existence of another e′ ∈ Ev i ∩ Ev j such that e <i e′ but e′ <j e.
We denote by Ids(e) the set {i ∈ Id e ∈ Ev i}, for each e ∈ Ev .
We can check the progress of an agent by collecting all the local events that
have occurred up to a given point. This yields the notion of the local state of
agent i, which is a finite set ξi ⊆ Ev i down-closed for local causality, i.e., if
e ≤i e′ and e′ ∈ ξi then also e ∈ ξi. The set Ξi of all local states of an agent i
is totally ordered by inclusion and has ∅ as the minimal element.
Each non-empty local state ξi is reached, by the occurrence of an event that
we call last (ξi), from the local state ξi \ {last(ξi)}. The local states of each agent
are totally ordered, as a consequence of the total order on local events. Since
they are discrete and well-founded, we can enumerate them as follows: ∅ is the
0th state; {e}, where e is the minimum of hEv i, ≤ii, is the 1st state; and if ξi
is the kth state of agent i and last (ξi) →i e, then ξi ∪ {e} is agent i's (k + 1)th
state.
We will denote by ξk
i the kth state of agent i, so ξ0
i = ∅ is the initial state
i is the state reached from the initial state after the occurrence of the first
is the only state of agent i that contains k elements, i.e.,
i = k. Given e ∈ Ev i, e ↓i= {e′ ∈ Ev i e′ ≤i e} is always a local state.
and ξk
k events. In fact, ξk
i
where ξk
Moreover, if ξi is non-empty, then last (ξi) ↓i= ξi.
We can also define the notion of a global state: a finite set ξ ⊆ Ev closed for
global causality, i.e. if e ≤ e′ and e′ ∈ ξ, then also e ∈ ξ. The set Ξ of all global
states constitutes a lattice under inclusion and has ∅ as the minimal element.
Clearly, every global state ξ includes the local state ξi = ξ ∩ Ev i of each agent
i. Given e ∈ Ev , e↓= {e′ ∈ Ev e′ ≤ e} is always a global state.
Figure 1 depicts a distributed life-cycle where each row comprises the local
life-cycle of one agent. In particular, Ev i = {e1, e4, e5, e8, . . . } and →i corre-
sponds to the arrows in i's row. We can think of the occurrence of event e1 as
leading agent i from its initial state ∅ to the state {e1}, and the the occurrence
of event e4 as leading to state {e1, e4}, and so on. Shared events at communica-
tion points are highlighted by the dotted vertical lines. Note that the numbers
annotating the events are there only for convenience since, in general, no global
total order on events is imposed. Figure 2 shows that corresponding lattice of
global states.
An interpretation structure µ = hλ, ϑi consists of a distributed life-cycle λ
5
i
j
k
e1
e2
e3
e4
e4
e4
e5
e6
e8
e8
e9
. . .
. . .
. . .
e7
e7
Figure 1: A distributed life-cycle for agents i, j and k.
{e1}
{e1, e2}
{e1, e2, e3, e4, e5}
∅
{e2}
{e1, e3}
{e1, e2, e3}
{e1, e2, e3, e4}
. . .
{e3}
{e2, e3}
{e1, e2, e3, e4, e6}
Figure 2: The lattice of global states.
and a family ϑ = {ϑi}i∈Id of local labeling functions, where, for each i ∈ Id ,
ϑi : Ξi → ℘(Prop i) associates a set of local state propositions to each local
state. We denote the tuple hλi, ϑii also by µi.
We can the define a global satisfaction relation as follows. Given a global
interpretation structure µ and a global state ξ then
• µ, ξ (cid:13) ¬α if µ, ξ 6(cid:13) α;
• µ, ξ (cid:13) α ⇒ β if µ, ξ 6(cid:13) α or µ, ξ, (cid:13) β;
• µ, ξ (cid:13) @i[ϕ] if µi, ξi (cid:13)i ϕ.
The local satisfaction relations at local states are defined by
• µi, ξi (cid:13)i p if p ∈ ϑi(ξi);
• µi, ξi (cid:13)i ¬ ϕ if µi, ξi 6(cid:13)i ϕ;
• µi, ξi (cid:13)i ϕ ⇒ ψ if µi, ξi 6(cid:13)i ϕ or µi, ξi (cid:13)i ψ;
• µi, ξi (cid:13)i X ϕ if there is e ∈ Ev i\ξi such that ξi∪{e} ∈ Ξi and µi, ξi∪{e} (cid:13)i ϕ;
• µi, ξi (cid:13)i G ϕ if µi, ξ′
i
(cid:13)i ϕ, for every ξ′
i ∈ Ξi such that ξi ⊆ ξ′
i;
• µi, ξi (cid:13)i c(cid:13)j[ϕ] if ξi 6= ∅, last (ξi) ∈ Ev j and µj, last(ξi)↓j (cid:13)j ϕ.
We say that µ (globally) satisfies α, or that µ is a model of α, written µ (cid:13) α,
whenever µ, ∅ (cid:13) α. As expected, α is said to be satisfiable whenever there is
µ such that µ (cid:13) α. We denote by Mod(α) the set of all models of α. We
define similar notion for the local languages. We say that µi (locally) satisfies
ϕ, written µi (cid:13)i ϕ if µi, ∅ (cid:13)i ϕ.
The following result will be useful in the future. It basically captures the
traditional fixed-point characterization of the G temporal operator:
G ϕ ⇔ (ϕ ∧ X G ϕ).
6
Lemma 2.1 Let µi be a local interpretation structure and ξi ∈ Ξi any of its
local states. Then
1. µi, ξi (cid:13)i G ϕ iff µi, ξi (cid:13)i ϕ and µi, ξi ∪ {e} (cid:13)i G ϕ, provided that ξi ∪ {e} ∈
Ξi.
3 Distributed Buchi Automata
In this section we present Buchi automota for DTL. We start by presenting
the traditional notion of nondeterministic Buchi automaton and generalized
nondeterministic Buchi automaton. We use these notions to capture the local
behaviour of the agents, given that each agent is essentially linear. In this case,
we follow very closely the ideas presented in [1]. Then, we propose a novel notion
of distributed Buchi automaton to capture the distributed nature of DTL.
A nondeterministic Buchi automation (NBA) is a tuple A = hQ, Σ, δ, Q0, F i
where:
• Q is a nonempty finite set of states;
• Σ is a finite set alphabet symbols such that Q ∩ Σ = ∅;
• δ : Q × Σ → 2Q is the transition function;
• Q0 ⊆ Q is a set of initial states;
• F ⊆ Q is a set of acceptance states (also called final states).
a−→ qk′ instead. Let Σω denote the set of
When q′ ∈ δ(q, a), we may write q
all infinite words over Σ. A run for w = a0a1a2 · · · ∈ Σω in A is an infinite
ak−→ qk+1, for k ∈ N:
sequence q0q1q2 . . . of states in A such that q0 ∈ Q0 and qk
q0
a0−→ q1
a1−→ q2
a2−→ . . .
A run q0q1q2 . . . is accepting if qk ∈ F for infinitely many indices k ∈ N. The
accepted language of A is
L(A) = {w ∈ Σω there exists an accepting run for w in A}.
A generalized nondeterministic Buchi automaton (GNBA) is a tuple G =
hQ, Σ, δ, Q0, Fi where Q, Σ, δ and Q0 are defined just as for NBA and F is a
(possibly empty) subset of 2Q. The elements of F are called acceptance sets.
A run for w = a0a1a2 · · · ∈ Σω in G is defined as in the case of an NBA. A
run q0q1q2 . . . is accepting if for each acceptance set F ∈ F there are infinitely
many indices k ∈ N such that qk ∈ F . The accepted language for a GNBA is
defined just as for the case of an NBA.
The classes of NBA's and GNBA's are equivalent in the sense that they
accept exactly the same languages. Every NBA is a particular case of a GNBA.
Furthermore, for each GNBA G there exists an NBA AG such that L(AG) =
L(G). Details of this equivalence can be found in [1].
In the sequel, we overload the ↓i notation and use q↓i to denote the projection
of tuple q over component i.
Next, we present the novel notion of distributed Buchi automata for DTL.
From now on, we assume fixed a distributed signature Σ = hId , {Prop}i∈Id i.
7
For each i ∈ Id , let Ai = hQi, Σi, δi, Q0i, Fii be an NBA such that for distinct
i, j ∈ Id :
• Qi ∩ Qj = ∅;
A distributed nondeterministic Buchi automaton (DNBA) based on {Ai}i∈Id is
a tuple
D = hQ, Σ, δ, Q0, Fi
such that:
• Q =Ni∈Id Qi;
• Σ = {a ⊆ ⊎i∈Id Σi a 6= ∅ and a ∩ Σi ≤ 1, for i ∈ Id };
• δ : Q × Σ → 2Q is such that δ(q, a) is the set of all states q′ satisfying:
-- if a ∩ Σi = ∅ then q′↓i= q↓i;
-- if a ∩ Σi 6= ∅ then q′↓i∈ δi(q↓i, a ∩ Σi);
• Q0 =Ni∈Id Q0i ;
• F = {Fi ⊆ Q
for i ∈ Id } such that Fi = {q ∈ Q q↓i∈ Fi};
The states of the DNBA are tuples of states from the local automata, one
for each agent. Each symbol of the distributed alphabet is a nonempty set of
symbols of the local automata with the proviso that in each global symbol there
is at most one symbol from each agent. The transition from one state to the
next at the global level is guided by the local behaviour of each component. If,
for a particular global symbol a, agent i is not involved, that is, if a ∩ Σi = ∅
then for this transition the agent's local state will not change. If, on the other
hand, the agent is involved in a, that is, if a ∩ Σi 6= ∅ then the agent's local
state will change according to its local behaviour, which is dictated by δi. Note
that, in this case, we are abusing notation. If a ∩ Σi 6= ∅ then a ∩ Σi is a set,
a singleton {a′} with a′ ∈ Σi, but nevertheless, a set. Hence, when we write
δi(q↓i, a ∩ Σi) we obviously mean δi(q↓i, a′).
The language accepted by the distributed automaton will be as expected. It
will accepted all the local words of the local automata. However, we need one
additional proviso: we only consider fair words. A global word a0a1a2 is fair
if, for every i ∈ Id , ak ∩ Σi 6= ∅, for infinitely many indices k ∈ N. We need
to ensure that a global accepting run is locally accepting for each agent. So,
we cannot simply promote a state q to accepting because one of its components
is accepting in the local automaton. This would allow for the acceptance of
other words from other local automata. A global run for a fair word w in D is
ak−→ qk+1, just as for the local case.
a sequence of states q0q1q2 . . . such that qk
is accepting if, for each i ∈ Id , qk ∈ Fi, for infinitely
A global run a0a1a2 . . .
many indices k ∈ N.
Let w = a0a1a2 · · · ∈ Σω be a fair global word. Then, we denote by w↓i the
local word obtained from w as follows:
• first, consider the projection w′ = (a0 ∩ Σi)(a1 ∩ Σi)(a2 ∩ Σi) . . . over the
alphabet Σi;
• then, let w↓i be the local word obtained from w′ by removing all the empty
sets and replacing each nonempty set {a} by its element a.
8
Recall that ak ∩ Σi ≤ 1 hence ak ∩ Σi is either a singleton or the empty set.
Similarly, let τ = q0q1q2 . . . be a global run for w = a0a1a2 . . . . Then, we
denote by τ↓w,i the local run for w↓i obtained from τ as follows:
• first, consider the projection τ ′ = q0↓i q1↓i q2↓i . . . over the states of Ai;
• then, let τ ↓w,i be the local run obtained from τ ′ by removing qk+1 if
(ak ∩ Σi) = ∅, for k ∈ N.
In this case, we project each global state on its local component for agent i
and then remove all the states resulting from transitions where agent i was not
involved. The follow lemma proves that w↓i is indeed a word in Σω
i and that
τ↓w,i is a local run for w↓i in Ai.
Lemma 3.1 Let D be a DNBA based on {Ai}i∈Id . If τ = q0q1q2 . . . is a global
run for a fair global word w = a0a1a2 · · · ∈ Σω then, for each i ∈ Id :
1. w↓i∈ Σω
i ;
2. τ↓w,i is a local run for w↓i in Ai.
Furthermore, τ is accepting if and only if τ↓w,i is accepting, for every i ∈ Id .
Proof: 1. Straightforward from the definition of w↓i and the fact that w is fair.
2. We briefly sketch the intuition behind this result. Consider the situation:
. . . h. . . , qk, . . .i ak−→ h. . . , qk+1, . . .i
ak+1−→ h. . . , qk+2, . . .i . . .
where ak ∩ Σi = ∅ and ak+1 ∩ Σi 6= ∅. Then, in τ ↓w,i, we get the following
situation:
. . . qk
ak+1∩Σi−→ qk+2 . . .
where qk+1↓i was deleted. By definition of δ,
qk+1↓i= qk↓i and qk+2↓i∈ δi(qk+1↓i, ak+1 ∩ Σi).
It is not very difficult to conclude that this leads to a run in Ai. Furthermore,
the fact that τ is accepting if and only if τ↓w,i is accepting, for every i ∈ Id , is
an immediate consequence of the definition of global acceptance.
(cid:3)
Example 3.2 Consider the NBA's A1 and A2 depicted in Figure 3, with Σ1 =
{0, 1} and Σ2 = {a, b}. A1 accepts all the infinite words over {0, 1} with in-
finitely many 0's, and A2 accepts all the infinite words over {a, b} with finitely
many a's.
Now let us consider the DNBA D based on {A1, A2}. The alphabet is com-
posed of sets with one symbol from one or from the two agents:
Σ = {{0}, {1}, {a}, {b}, {0, a}, {0, b}, {1, a}, {1, b}}.
The set of states is
Q = {hq0, p0i, hq0, p1i, hq1, p0i, hq1, p1i}.
9
0
q1
1
q0
0
1
a, b
p0
b
b
p1
Figure 3: NBA's A1 (on the left) and A2 (on the right).
{1}, {a}, {b}
{1, a}, {1, b}
{1}, {b}, {1, b}
hq0, p0i
{b}, {1, b}
hq0, p1i
{1}
{1, a}
{1, b}
{0, b}
{0}
{0, a}
{0, b}
{1}
{1, b}
{0}
{0, b}
hq1, p0i
{b}, {0, b}
hq1, p1i
{0}, {a}, {b}
{0, a}, {0, b}
{0}, {b}, {0, b}
Figure 4: DNBA D based on {A1, A2}.
Of these, only hq0, p0i is initial, and
F1 = {hq1, p0i, hq1, p1i} and F2 = {hq0, p1i, hq1, p1i}.
The transition function δ is depicted is Figure 4.
It is not very difficult to observe that words with an infinite number of 0's
and a finite number of a's are accepted. For instance, the word
{1}{a}{1, b}{0, b}{0}{b}{0, b}{0}{b} . . .
is accepted given that state hq1, p1i is visited infinitely often, that is, a final state
form A1 and a final state from A2 are visited infinitely often. However, if this
was the only requirement for acceptance, then the word
{1}{a}{1, b}{0, b}{0}{0}{0}{0} . . .
would also be accepted given that state hq1, p1i, in this case, is also visited in-
finitely often. We don't want this word to be accepted because its projection on
A2 yields the finite word aba which is not part of the language of A2. But, this
global word is not a fair word hence it will not be accepted by D.
10
Our goal now is to define a DNBA Dα for a given formula α ∈ L that
accepts all the models of α and only those. To this end, we start by defining
some auxiliary notions.
From now on, assume fixed a global formula α. A set B ⊆ closure(α) is said
to be consistent with respect to propositional logic if:
• γ1 ⇒ γ2 ∈ B if and only if ¬ γ1 ∈ B or γ2 ∈ B, for γ1 ⇒ γ2 ∈ closure(α);
• if γ1 ∈ B then ¬ γ1 /∈ B;
• if ⊤ ∈ closure(α) then ⊤ ∈ B.
Herein, γ1, γ2 denote either local or global formulas.
A set B ⊆ closure(α) is said to be locally consistent with respect to the
temporal operator G if:
• if G ϕ1 ∈ B then ϕ1 ∈ B, for every G ϕ1 ∈ closure(α).
A set B ⊆ closure(α) is said to be i-consistent with respect to global formulas
if
• @i[ϕ] ∈ B iff ϕ ∈ B, for every @i[ϕ] ∈ subf(α).
A set B ⊆ closure(α) is said to be maximal if for all γ ∈ closure(α):
• if γ /∈ B then ¬ γ ∈ B.
A set B ⊆ closure(α) is i-elementary if it is consistent with respect to propo-
sitional logic, maximal and locally consistent with respect to the temporal op-
erator G and i-consistent with respect to global formulas. Elementary sets try
to capture all the properties that can be asserted locally. When i is clear from
context, we may write elementary instead of i-elementary.
Recall formula @i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q]. The set
B1 = {@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], @i[X(p ⇒ c(cid:13)j[q])], @j[X q],
X(p ⇒ c(cid:13)j[q]), X q, ¬ X p, ¬ c(cid:13)j[q], q}
is an example of an elementary sets. However, for instance, sets
B2 = {@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], @i[X(p ⇒ c(cid:13)j[q])], ¬ @j[X q], . . . }
and
B3 = {@i[X(p ⇒ c(cid:13)j[q])] ⇒ @j[X q], ¬ @i[X(p ⇒ c(cid:13)j[q])], @j[X q], ¬ X q, . . . }
are not elementary. Set B2 is not consistent with propositional logic and set B3
is not j-consistent with global formulas.
We have all we need to define the envisaged DNBA. We start by defining
the local GNBA's Gi for each agent i ∈ Id . The construction is similar the one
presented in [1]. From these, we can then obtain equivalent NBAs Ai that will
be used to define the DNBA. Each GNBA Gi = hQi, Vi, δi, Q0i, Fii is as follows:
• Qi = {B↓i B ⊆ closure(α) and B is i-elementary};
• Q0i = {B ∈ Qi α ∈ B and c(cid:13)j[ϕ] /∈ B, for c(cid:13)j[ϕ] ∈ closure(α)};
11
• Fi = {FG ϕ G ϕ ∈ closure(α)} where
-- FG ϕ = {B ∈ Qi G ϕ ∈ B or ϕ /∈ B};
• δi : Qi × Vi → 2Qi is such that:
-- if v 6= B ∩ Liti then δi(B, v) = ∅;
-- if v = B ∩ Liti then δi(B, v) is the set of all elementary sets B′ such
that:
1. X ϕ ∈ B iff ψ ∈ B′, for every X ϕ ∈ closure(α);
2. G ϕ ∈ B iff ϕ ∈ B and G ϕ ∈ B′, for every G ϕ ∈ closure(α).
Recall that Vi is the set of all i-valuations and that a valuation is a set of literals
such that, for each propositional symbol either the symbol is in the valuation
or its negation is. This is the alphabet of the automaton. The states of the
automaton are all the elementary sets (restricted to the relevant formulas, that
is, all the global formulas and all the local formulas for the agent at hand).
Each state contains all the formulas that are intended to hold at that point. In
particular, initial states characterize the initial set-up conditions. We want α
to hold initially and, as imposed by the semantics of DTL, there can only by
synchronizations after the first event occurs. Hence, there can be no commu-
nication formulas in any initial state. Regarding the transition function, given
an alphabet symbol v and state B, the transition will only be enabled if the
valuation v agrees with the information in B, i.e., if the state propositions in
v are also present in B, meaning that they should be true, and the negation of
state propositions in v are also present in B, meaning that they should be false.
Additionaly, conditions (1) and (2) reflect the semantics of temporal operators.
In particular, condition (2) is based on the fixed-point semantics of the G oper-
ator. The final states are defined in order to capture the temporal semantics of
the G operator. They are basically used to exclude runs where, from a certain
point on, a formula ϕ is always true (that is, it present in all the states) but G ϕ
is not (that is, it is not present in the states of the run), for instance, as follows:
B0
v0−→ . . .
vk−1−→ {. . . , ϕ, . . . } vk−→ {. . . , ϕ, . . . }
vk+1−→ . . .
In this run, ϕ is present in all the states starting from k. Then, this means that
G ϕ is true from that point on. In order for the run to be accepting, FG ϕ must
be visited infinitely often. This means that after k and as ϕ ∈ Bn for n ≥ k
then G ϕ must be in infinitely of these states. Then, by condition (2) of the
transition function, G ϕ must be in all of them, as intended.
For each i ∈ Id , let Ai be an NBA equivalent to the GNBA Gi. Then
Dα = hQ, Σ, δ, Q0, Fi is the DNBA based on {Ai}i∈Id satisfying the following
conditions, for every q, q′ ∈ Q, a ∈ Σ and i, j ∈ Id :
(LC) if c(cid:13)j[ϕ] ∈ q↓i then ϕ ∈ q↓j;
(SC1) if q′ ∈ δ(q, a) and c(cid:13)j[ϕ] ∈ q′↓i and a ∩ Vi 6= ∅ then a ∩ Vj 6= ∅;
(SC2) if a ∩ Vi 6= ∅ and a ∩ Vj 6= ∅ and ϕ ∈ q′↓j then c(cid:13)j[ϕ] ∈ q′↓i.
12
Condition (LC) states that in a global state, if for agent i it holds c(cid:13)j[ϕ] then it
must be the case that ϕ holds for agent j, as intended for the semantics of the
communication primitive. Condition (SC1) states that in every state q′ reached
by a transition where i was an active participant (expressed by a ∩ Vi 6= ∅),
if c(cid:13)j[ϕ] holds for i then i and j must have just synchronized and so j must
also have been an active participant in a (expressed by a ∩ Vj 6= ∅). Finally,
condition (SC1) states that if i and j were both active in the last transition and
if ψ holds for agent j then, in the event that i wants to communicate with j, it
will be able to infer that ψ holds for j, that is, c(cid:13)j[ψ] holds for i.
We now proceed to show the correctness of this construction. We aim at
proving that any word accepted by the automaton is captured by a DTL model
of α and that any DTL model of α is represented by words accepted by the
automaton.
Let w = a0a1a2 . . . · · · ∈ L(Dα). Then, there is an accepting run τ =
q0q1q2 . . . in Dα. We denote by µτ = hλ, ϑi the interpretation structure induced
by τ (and consequently by w), defined as follows:
• Ev = {ek k ≥ 1};
• Ev i = {ek ∈ Ev ak−1 ∩ Vi 6= ∅};
• λi = hEv i ≤ii is the local life-cycle such that ek1 ≤i ek2 if k1 ≤ k2;
• λ = {λi}i∈Id is the corresponding distributed life-cycle;
• ϑi : Ξi → 2Propi is such that, for every p ∈ Prop i and ek ∈ Ev i:
0 if ¬ p ∈ q0↓i
-- ϑi(∅) =(1 if p ∈ q0↓i
-- ϑi(ξ ∪ {ek}) =(1
if p ∈ qk↓i
if ¬ p ∈ qk↓i
0
, for ξ ∈ Ξi;
• ϑ = {ϑi}i∈Id .
Note that each ϑi is well defined because qk is elementary. In the sequel, we can
consider the following enumeration of global events:
• ξ0 = ∅
• ξk = {e1, . . . , ek}, for k ≥ 1.
Theorem 3.3 If w ∈ L(Dα) then µτ ∈ Mod(α), for some accepting run τ for
w in Dα.
Proof: Let τ = q0q1q2 . . . . Our goal is to prove that µτ (cid:13) α. We start by
establishing a preliminary result for the local level. We prove that µτ
i , ξki (cid:13)i ψ
iff ψ ∈ qk↓i, for every ψ ∈ closure(α) ∩ Li. The proof is done by induction on
the structure of ψ, simultaneously for all agents.
Basis: ψ ∈ Prop i. Then, if k = 0 then ξ0 = ∅ and p ∈ q0 ↓i iff ϑi(∅) = 1
iff µτ
If ek ∈ Ev i then
ξki = ξ′ ∪ {ek}, for some local state ξ′ ∈ Ξi, and p ∈ qk↓i iff ϑi(ξ′ ∪ {ek}) = 1 iff
µτ
i , ξki (cid:13)i p. If ek /∈ Ev k, and consequently ak−1 ∩ Vi = ∅, then we distinguish
If k > 0, then ξk 6= ∅ and last (ξk) = ek.
i , ξ0 (cid:13)i p.
13
two cases: either (i) ak1 ∩ Vi = ∅, for every k1 < k; or (ii) there is k1 < k such
that Vk1 ∩ Vi 6= ∅. In the first case, this means that ek1 /∈ Ev i, for k1 < k and so
ξki = ∅ and, by definition of DNBA, it also follows that q0↓i= q1↓i= · · · = qk↓i.
The proof then follows as in the case of k = 0. In case (ii), let k1 be the greatest
k′ < k such that ak′ ∩ Vi 6= ∅. Then, ek1+1 ∈ Ev i, last (ξki) = ek1+1 and
ξki = ξ′ ∪ {ek1+1}, for some local state ξ′ ∈ Ξi. Furthermore, it follows by
definition of DNBA that qk1+1↓i= · · · = qk↓i. Hence, p ∈ qk↓i iff p ∈ qk1+1↓i iff
ϑi(ξ′ ∪ {ek1+1}) = 1 iff µτ
i , ξ′ ∪ {ek1+1} (cid:13)i p iff µτ
i , ξki (cid:13)i p.
Induction step: The case of propositional formulas is an immediate consequence
of the definition elementary set and we omit the details.
Assume that ψ = X ψ1. Additionally, assume also that X ψ1 ∈ qk ↓i. If ak ∩
Vi 6= ∅ then ψ1 ∈ qk+1 ↓i, by definition of Gi, and, by induction hypothesis,
i , ξk+1i (cid:13)i ψ1. But ak ∩ Vi 6= ∅ also implies that ek+1 ∈ Ev i and, so, ξk+1i =
µτ
i , ξki ∪ {ek+1} (cid:13)i ψ1 which implies
(ξk ∪ {ek+1})i = ξki ∪ {ek+1}. Hence, µτ
that µτ
i , ξki (cid:13)i X ψ1. If ak ∩ Vi = ∅ then let k1 be the least index greater than
k such that ak1 ∩ Vi 6= ∅. Then, by definition of DNBA, qk ↓i= · · · = qk1 ↓i
and so X ψ1 ∈ qk1 ↓i. Consequently, ψ1 ∈ qk1+1 ↓i. By induction hypothesis,
i , ξk1+1i (cid:13)i ψ1. By definition of Ev i, it follows that ek1+1 ∈ Ev i, ξk1 i =
µτ
ξki and thus ξk1+1i = (ξk1 ∪ {ek1+1})i = ξk1 i ∪ {ek1+1} = ξki ∪ {ek1+1}.
i , ξki (cid:13)i X ψ1. Assume now that
Hence, µτ
µτ
i , ξki (cid:13)i X ψ1. Then, there is e ∈ Ev i \ ξki such that ξki ∪ {e} ∈ Ξi and
i , ξki ∪ {e} (cid:13)i ψ1. Clearly, there is k1 > k such that e = ek1 and ξki ∪ {e} =
µτ
ξk1 i. Using the induction hypothesis, it follows that ψ1 ∈ qk1↓i. If k1 = k + 1
then ak ∩ Vi 6= ∅ and, by definition of Gi, X ψ1 ∈ qk ↓i.
If k1 > k + 1 then
ak ∩ Vi = · · · = ak1−2 ∩ Vi = ∅ and ak1−1 ∩ Vi 6= ∅. Hence qk↓i= · · · = qk1−1↓i
and if ψ1 ∈ qk1↓i then X ψ1 ∈ qk1−1↓i= qk↓i.
i , ξki ∪ {ek1+1} (cid:13)i ψ1 and thus µτ
i , ξk′
i , ξ′
i
i , ξki (cid:13)i G ψ1. Assume now that µτ
Assume now that ψ is G ψ1. We start by observing that for any run B0B1B2 . . .
in Gi if G ψ1 ∈ Bk then G ψ1 ∈ Bk′ for every k′ ≥ k. This is an immediate
consequence of condition (2) in the definition of δi and can easily be established
by induction. Assume first that G ψ ∈ qk↓i. Then, by the previous claim and
Lemma 3.1, it follows that G ψi ∈ qk′ ↓i for every k′ ≥ k. As each set qk′ ↓i
is elementary then it is locally consistent with respect to temporal operator
G and so ψ1 ∈ qk′ ↓i, for every k′ ≥ k. Using the induction hypothesis, it
i (cid:13)i ψ1, for every k′ ≥ k. And this last condition implies
follows that µτ
that µτ
(cid:13)i ψ1,
for every ξ′
i ⊇ ξki. A simple inductive argument allows us to conclude that
i , ξk′
i = ξ′
µτ
i ∪ {ek′ }
i = ξk′−1i and
which is in Ξi and satisfies ξ′
again ξk′
i ⊇ ξki. Hence, by induction hypothesis, it follows that ψ1 ∈ qk′ ↓i,
for every k′ ≥ k. As the run is accepting then, some of these states must be in
. Let qk1↓i, with k1 ≥ k, be the first of such states. Clearly, it must be the
FG ψ1
case that k1 = k. In fact, if k1 > k, given that ψ ∈ qk1↓i then, by condition (2)
in the definition of δi this would imply that G ψ1 ∈ qk1−1↓i forcing qk1−1↓i to
also be in FG ψ1
and thus contradicting the fact the k1 was the first final state
after k. Hence, G ψ1 ∈ qk↓i.
i (cid:13)i ψ1, for every k′ ≥ k. Note that if ek′ ∈ Ev i then ξk′
i ∪{ek′ } ⊇ ξki. If ek′ /∈ Ev i then ξk′
i , ξki (cid:13)i G ψ1. Then, µτ
Finally, assume that ψ = c(cid:13)j[ψ1]. If k = 0 then, by definition of initial state and
of the local satisfaction relation, ψ /∈ q0↓i and µτ
i , ∅ 6(cid:13)i ψ and the result follows.
14
If k > 0 then last (ξki) = ek. Assume first that c(cid:13)j[ψ1] ∈ qk↓i. Furthermore, let
k1 be the greatest index less that k such that ak1 ∩ Vi 6= ∅. Then ek1+1 ∈ Ev i.
We have to consider two cases: either (i) k1 = k − 1; or (i) k1 < k − 1. Let us
consider case (i). Then, by condition (SC1), it follows that ak1 ∩Vj 6= ∅ and thus
ek ∈ Ev j. Furthermore, by condition (LC), it follows that ψi ∈ qk ↓j. Hence,
j , ξkj (cid:13)j ψ1. But, as last (ξki) = ek = last (ξkj)
by induction hypothesis, µτ
j , last (ξki)↓j (cid:13)j ψ1 which implies
then last (ξki)↓j= last (ξkj)↓j= ξkj and so µτ
that µτ
i , ξki (cid:13)i c(cid:13)j[ψ1]. Consider now the case (ii). In this case, we know that
last (ξki) = ek1+1. We also know that qk1+1↓i= · · · = qk ↓i. Hence c(cid:13)j[ψ1] ∈
qk1+1↓i and, by condition (LC), it also follows that ψ1 ∈ qk1+1↓j. Additionally,
by condition (SC1), ak1 ∩ Vj 6= ∅, which implies that ek1+1 ∈ Ev j. Reasoning
i , ξk1+1i (cid:13)i c(cid:13)j[ψ1]. But ξk1+1i = ξki
as in case (i), we can conclude that µτ
given that last (ξki) = ek1+1 = last (ξk1+1i). Hence, µτ
i , ξki (cid:13)i c(cid:13)j[ψ1]. To
i , ξki (cid:13)i c(cid:13)j[ψ1]. Then last (ξki) ∈ Ev j and
prove the converse, assume that µτ
µτ
j , last(ξki)↓j (cid:13)j ψ1. Let last (ξki) = ek1 ∈ Ev i. Clearly, k1 ≤ k. Again, we
need to consider two cases: either (i) k1 = k; or (ii) k1 < k. In the first case,
ek ∈ Ev i implies that ak−1 ∩ Vi 6= ∅, ek ∈ Ev j implies that ak−1 ∩ Vj 6= ∅ and,
by induction hypothesis, ψ1 ∈ qk↓j, given that last (ξki) = ek = last (ξkj) and
j , ξkj (cid:13)j ψ1. Hence, using condition (SC2),
so µτ
we conclude that c(cid:13)j[ψ1] ∈ qk ↓i. If condition (ii) holds then ak1−1 ∩ Vi 6= ∅
and ak1 ∩ Vi = · · · = ak−1 ∩ Vi = ∅, which implies that qk1 ↓i= · · · = qk ↓i.
Furthermore, given that last ((ξki) ↓j) = ek1 then (ξki) ↓j= ξk1 ↓j. Thus,
µτ
j , ξk1 ↓j (cid:13)j ψ1 and, by induction hypothesis, ψ1 ∈ qk1 ↓j. Furthermore, like
in case (i), we also know that ak1−1 ∩ Vi 6= ∅ and ak1−1 ∩ Vj 6= ∅. Hence,
c(cid:13)j[ψ1] ∈ qk1↓i, which implies that c(cid:13)j[ψ1] ∈ qk↓i.
j , last (ξki)↓j (cid:13)j ψ1 implies µτ
Next, we prove a similar result for the global level. For each α1 ∈ closure(α),
α1 ∈ q0 iff µτ , ξ0 (cid:13) α1. We abuse notation and write α1 ∈ q0 to mean that
α1 ∈ qk ↓i, for some i ∈ Id . Clearly, q0 ↓i ∩subf(α) = q0 ↓j ∩subf(α), that
is, the initial states of all the local automata have exactly the same global
subformulas of α. Again, the proof follows by induction in the structure of
α. The propositional cases are immediate consequences of the properties of
elementary sets. So, let α1 = @i[ϕ]. Then, µτ , ξ0 (cid:13) @i[ϕ] iff µτ
i , ξ0i (cid:13)i ϕ iff
ϕ ∈ q0↓i, by the previous result, iff @i[ϕ] ∈ q0↓i, by the properties of elementary
sets, iff @i[ϕ] ∈ q0↓i.
(cid:3)
We now prove the converse, i.e, we prove that any DTL model of α can
be captured by Dα. Let µ = hλ, ϑi be an interpretation structure and let
hEv , ≤Ev i be the underlying global order on events.
It is always possible to
linearize hEv , ≤Ev i, i.e., it is always possible to define a bijection ℓ : N1 → Ev
such that if k1 <N k2 then ℓ(k1) <Ev ℓ(k2), where <N is the usual ordering on
the natural numbers. See, e.g., [6]. From now on, we assume fixed a linearization
function (or just linearization) ℓ, which induces an enumeration of the global
states as follows
• ξ0 = ∅;
• ξk = ξk−1 ∪ {ℓ(k)} for each k ≥ 1.
15
Consider the word wµ,ℓ = a0a1a2 . . . where
ak =
[i∈Ids (ℓ(k+1))
{p ∈ Liti µi, ξki (cid:13)i p}
This word represent one possible evolution of the system represented by µ. Our
goal is to show that wµ,ℓ is captured by Dα, that is, to show that wµ,ℓ ∈ L(Dα).
Theorem 3.4 If µ ∈ Mod(α) then wµ,ℓ ∈ L(Dα).
Proof: To show that wµ,ℓ ∈ L(Aα) we need to present an accepting run for wµ,ℓ.
For each k ∈ N consider the sets of formulas induced by µ:
• xi
0 = {α ∈ closure(α) µ, ξ0 (cid:13) α};
• xi
k =({α ∈ closure(α) µ, ξk (cid:13) α}
k−1
xi
if last (ξk) ∈ Ev i
otherwise
• yi
k = {ϕ ∈ closure(α) ∩ Li µi, ξki (cid:13)i ϕ}.
• qk = ⊗i∈Id (xi
k ∪ yi
k).
Each xi has information about the global formulas and each yi has information
about the local formulas of agent i. We start by establishing a structural result
If, for i ∈ Id , ℓ(k + 1) /∈ Ev i then
on the local component of the states.
k+1 = yi
yi
k. This is an immediate consequence of the fact that if ℓ(k + 1) /∈ Ev i
then ξk+1i = ξki and so yi
k+1 = yi
k.
k∪yi
k ∪ yi
Having established this result, we prove that each qk is a state in Dα. By
construction, we have that (xi
k ⊆ closure(α). Further-
more, each xi
k is i-elementary. The conditions concerning the connectives
and temporal operators are a consequence of the definition of the satisfaction
relation. The only condition worth checking is the one related to global formu-
k iff µ, ξk (cid:13) @i[ϕ] iff µi, ξki (cid:13)i ϕ
las. In this case, @i[ϕ] ∈ xi
iff ϕ ∈ yi
k iff @i[ϕ] ∈ xi
k)↓i= xi
k and xi
k ∪ yi
k∪yi
k∪yi
k iff ϕ ∈ xi
k ∪ yi
k.
Next, we prove that q0q1q2 . . . is a run for wµ,ℓ in Dα, that is, we need to
ak−→ qk+1, for every k ∈ N. The fact that q0 ∈ Q0
establish that q0 ∈ Q0 and qk
is straightforward. Indeed, q0 ∈ Q0 iff q0↓i∈ Q0i iff α ∈ q0↓i and q0↓i has no
0 because µ, ξ0 (cid:13) α given that
communication formulas. Observe that α ∈ xi
µ ∈ Mod(α). And q0↓i has no communication formulas because ξ0 = ∅ and so
ak−→ qk+1 we
it does not satisfy any communication formula. To prove that qk
consider two cases: (i) ak ∩ Vi = ∅; and (ii) ak ∩ Vi 6= ∅, for i ∈ Id .
(i) If ak ∩ Vi = ∅ then i /∈ Ids(ℓ(k + 1)) which implies that yi
more, by definition, xi
(ii) If ak ∩ Vi 6= ∅ then ℓ(k + 1) ∈ Ev i and ξk+1i = ξki ∪ {ℓ(k + 1)}. We prove
that, in this case, qk+1↓i∈ δi(qk↓i, ak ∩ Vi):
k. Hence, qk+1↓i= qk↓i.
k. Further-
k+1 = xi
k+1 = yi
• ak ∩ Vi = qk↓i ∩Vi: straightforward by construction of ak and qk;
• let X ψ ∈ closure(α) ∩ Li: X ψ ∈ qk ↓i iff X ψ ∈ yi
µi, ξki ∪ {ℓ(k + 1)} (cid:13)i ψ iff µi, ξk+1i (cid:13)i ψ iff ψ ∈ yi
k iff µi, ξki (cid:13)i X ψ iff
k+1↓i iff ψ ∈ qk+1↓i;
16
• let G ψ ∈ closure(α) ∩ Li: G ψ ∈ qk ↓i iff G ψ ∈ yi
k iff µi, ξki (cid:13)i G ψ
iff µi, ξ′ (cid:13)i ψ, for ξ′ ⊇ ξki, iff µi, ξki (cid:13)i ψ and µi, ξ′ (cid:13)i ψ, for ξ′ ⊇
ξki ∪ {ℓ(k + 1)}, iff µi, ξki (cid:13)i ψ and µi, ξki ∪ {ℓ(k + 1)} (cid:13)i G ψ iff
µi, ξki (cid:13)i ψ and µi, ξk+1i (cid:13)i G ψ iff ψ ∈ yi
k+1 iff ψ ∈ qk↓i
and G ψ ∈ qk+1↓i.
k and G ψ ∈ yi
Hence, from (i) and (ii), if follows that qk+1 ∈ δ(ak, qk), as desired.
Finally, we just need to establish that the run is accepting. We start by
showing that each local run is accepting. Let G ϕ ∈ closure(α) ∩ Li. We need
to show that there are infinitely many indices k such that qk↓i∈ FG ϕ. Assume
that this is not the case, i.e., assume that there are only finitely many indices
k such that qk↓i∈ FG ϕ. Then, there is n ∈ N such that qk↓i /∈ FG ϕ, for every
k ≥ n. But, if qk↓i /∈ FG ϕ then G ϕ /∈ qk↓i and ϕ ∈ qk↓i, for k ≥ n. If G ϕ /∈ qk↓i
k and so µi, ξki 6(cid:13) G ϕ. This implies that there is ξ′ ⊇ ξki such
then G ϕ /∈ yi
that µi, ξ′
6(cid:13) ϕ. It is not very difficult to see that there is m ≥ k such that
ξ′ = ξmi. Hence µi, ξmi 6(cid:13) ϕ which implies that ϕ /∈ yi
m and, consequently,
ϕ /∈ qm ↓i. But, as m ≥ k ≥ n then ϕ ∈ qm ↓i and we reach a contradiction.
This means that each set FG ϕ is visited infinitely often and, thus, the local run
is accepting. The fact that the global word wµ,ℓ is fair is a consequence of the
linearization. Then, by Lemma 3.1, we can conclude that the global run is also
accepting.
(cid:3)
We now study the relationship between µ and µτ where τ is an accepting
run for wµ,ℓ, and between w and wµτ ,ℓ. If we start with a model µ, construct
wµ,ℓ and then define µwµ,ℓ
, we may wonder what is the relation between µ and
µwµ,ℓ
? Conversely, if we start with a word w (with accepting run τ ) and we
build the model µτ , can we can find a linearization ℓ such that wµτ ,ℓ is w? The
following lemmas answer both these questions.
Lemma 3.5 Let w ∈ L(Dα) with accepting run τ . Then, there is a linearization
ℓ such that w = wµτ ,ℓ.
Proof: Let w ∈ L(Dα) with accepting run τ and consider µτ defined as above.
Consider the linearization ℓ : N1 → Ev such that ℓ(k) = ek. First, we observe
that ξ0 = ∅ and ξk = {ℓ(1), . . . , ℓ(k)} = {e1, . . . , ek}. Let p ∈ Prop i. Then, for
k ∈ N
p ∈ wµτ ,ℓ
k
i , ξki (cid:13)i p (by definition of wµ,ℓ)
iff µτ
iff p ∈ ϑi(ξki)
iff p ∈ qk↓i
iff p ∈ wk ∩ Vi
iff p ∈ wk
(by definition of satisfaction)
(by definition of µτ )
(by definition of δi)
(because p ∈ Prop i).
Hence, wµτ ,ℓ
k = wk, for every k ∈ N, and so wµτ ,ℓ = w.
(cid:3)
We can state a similar result for µ and µwµ,ℓ
. But in case, we cannot say that
the two interpretation structures are equal but only isomorphic. We say that
two distributed life-cycles λ1 = {hEv 1
i ii}i∈Id are
isomorphic, written λ1 ∼=f λ2, if there is an bijection f : Ev 1 → Ev 2 such that
e ≤1 e′ if and only if f (e) ≤2 f (e′), for every e, e′ ∈ Ev 1. Function f establishes
i , ≤1
i ii}i∈Id and λ2 = {hEv 2
i , ≤2
17
a bijection between states of λ1 and λ2 as would be expected. We say that
two interpretation structures µ1 = hλ1, ϑ1i and µ2 = hλ2, ϑ2i are isomorphic,
written µ1 ∼=f µ2, if λ1 ∼=f λ2 and ϑ1
i (f (ξi)), for every state ξi ∈ Ξ1
i .
We may drop the reference to f and simply write λ1 ∼= λ2. We can now state
the converse result of Lemma 3.5.
i (ξi) = ϑ2
Lemma 3.6 Let µ ∈ Mod(α). Then, µ ∼= µτ , where τ is an accepting run for
wµ,ℓ, for a given linearization ℓ.
Proof: We start by defining a bijection between Ev and Ev τ , where Ev τ denotes
the set of events of µτ . Let f : Ev → Ev τ be such that f (e) = ek where k is
such that ℓ(k) = e that exists because ℓ is a linearization of Ev . Furthermore,
let e, e′ ∈ Ev such that ℓ(k) = e and ℓ(k′) = e′, for some k, k′ ∈ N1. Then e ≤ e′
iff ℓ(k) ≤ ℓ(k′) iff k ≤ k′ iff ek ≤τ ek′ . Hence, λ ∼= λτ . As it was said before,
f (∅) = ∅ and f (ξk) = f ({ℓ(1), . . . , ℓ(k)}) = {e1, . . . , ek}. Next, we prove that
µ ∼= µτ . Let p ∈ Prop i. Then for k ∈ N
p ∈ ϑτ
i (f (ξk)i)
i ({e1, . . . , ek}i)
iff p ∈ ϑτ
iff p ∈ qk↓i
iff p ∈ wµτ ,ℓ
iff p ∈ wµτ ,ℓ
iff µτ
i , ξki (cid:13)i p
iff p ∈ ϑi(ξki)
k
∩ Vi
k
(as observed above)
(by definition of µτ )
(by definition of δi)
(because p ∈ Prop i)
(by definition of wµ,ℓ)
(by definition of satisfaction).
Hence, we conclude that µ ∼= µτ .
(cid:3)
These two lemmas allow us to conclude that Mod(α) and L(Dα) have essen-
tially the same information.
4 Concluding remarks
We have proposed a notion of distributed Buchi automaton. We have then
endowed DTL with an operation semantics based on DNBA's, where for the
local components (that have a linear behaviour) an approach similar to the
followed in [22, 1] was adopted. The construction was proved correct with
respect to DTL semantics.
As future work, we believe that it would be interesting to extend our ap-
proach to other temporal operators, like the until operator and past operators.
No surprises are expected as these have been widely studied for LTL and the
local agents of DTL have a linear time behaviour.
The main goal of this ongoing work is to endow DTL with a model-checking
tool. The work presented in this paper is the first step towards that goal.
Having such a tool will allows us to verify some of the problems to which DTL
has successfully been applied, but in an automated way [5, 8, 9]. It is also our
goal to study the complexity of our intendend approach and compare it with
existing tools.
18
Acknowledgements
This work was supported by the Portuguese Funda¸cao para a Ciencia e a Tec-
nologia (FCT) by way of grant UID/EEA/50008/2013 to Instituto de Teleco-
munica¸coes (IT).
References
[1] F. Baier and J.-P. Katoen. Principles of Model Checking. The MIT Press,
2008.
[2] E. Bartocci and C. R. C. R. Ramakrishnan. Preface of the special issue on
model checking of software. International Journal on Software Tools for
Technology Transfer, 18(4):355 -- 357, 2016.
[3] D. Basin, C. Caleiro, J. Ramos, and L. Vigan`o. A labeled tableaux for
the distributed temporal logic DTL. In Proceedings of the 15th Int. Symp.
on Temporal Representation and Reasoning (TIME 2008), pages 101 -- 109.
IEEE Computer Society Press, 2008.
[4] D. Basin, C. Caleiro, J. Ramos, and L. Vigan`o. Labelled tableaux for
distributed temporal logic. Journal of Logic and Computation, 19:1245 --
1279, 2009.
[5] D. Basin, C. Caleiro, J. Ramos, and L. Vigan`o. Distributed temporal logic
for the analysis of security protocol models. Theoretical Computer Science,
412(31):4007 -- 4043, 2011.
[6] E. Best and C. Fern´andez. Nonsequential processes -- A Petri net view.
Springer-Verlag, 1988.
[7] C. Caleiro, P. Gouveia, J. Ramos, and L. Vigan`o. A tableaux-based deci-
sion procedure for distributed temporal logic. In C. Caleiro, F. Dion´ısio,
P. Gouveia, P. Mateus, and J. Rasga, editors, Essays in Honour of Amil-
car Sernadas, Logic and Computation, pages 73 -- 124. College Publications,
London, 2017.
[8] C. Caleiro, L. Vigan`o, and D. Basin. Metareasoning about security pro-
tocols using distributed temporal logic. Electronic Notes in Theoretical
Computer Science, 125(1):67 -- 89, 2005. Preliminary version presented at
IJCAR'04 ARSPA Workshop.
[9] C. Caleiro, L. Vigan`o, and D. Basin. Relating strand spaces and distributed
temporal logic for security protocol analysis. Logic Journal of the IGPL,
13(6):637 -- 664, 2005.
[10] E. A. Clarke and E. M. Edmund. Characterizing correctness properties of
parallel programs using fixpoints.
In J. de Bakker and J. van Leeuwen,
editors, Automata, Languages and Programming. ICALP 1980, Lecture
Notes in Computer Science. Springer, Berlin, 1980.
[11] E. M. Clarke and E. A. Emerson. Design and synthesis of synchronization
skeletons using branching time temporal logic. In D. Kozen, editor, Logics
of Programs, pages 52 -- 71. Springer Berlin Heidelberg, 1982.
19
[12] E. M. Clarke, E. A. Emerson, and A. P. Sistla. Automatic verification of
finite-state concurrent systems using temporal logic specifications. ACM
Trans. Program. Lang. Syst., 8(2):244 -- 263, 1986.
[13] H.-D. Ehrich and C. Caleiro. Specifying communication in distributed
information systems. Acta Informatica, 36:591 -- 616, 2000.
[14] G. Holzmann. The SPIN Model Checker: Primer and Reference Manual.
Addison-Wesley Professional, 2004.
[15] F. Kroger and S. Merz. Temporal logic and state systems. Springer, 2008.
[16] O. Lichtenstein and A. Pnueli. Propositional Temporal Logic: Decidability
and Completeness. Logic Journal of the IGPL, 8(1):55 -- 85, 2000.
[17] K. Lodaya, R. Ramanujam, and P. Thiagarajan. Temporal logics for com-
municating sequential agents: I. International Journal of Foundations of
Computer Science, 3(1):117 -- 159, 1992.
[18] K. Lodaya and P. Thiagarajan. A modal logic for a subclass of event struc-
tures. In Proceedings of 14th ICALP, LNCS 267, pages 290 -- 303. Springer,
1987.
[19] K. L. McMillan. Symbolic Model Checking. Kluwer Academic Publishers,
Norwell, MA, USA, 1993.
[20] A. Pnueli. The temporal logic of programs. In FOCS, pages 46 -- 57, 1977.
[21] R. Ramanujam. Locally linear time temporal logic. In In Proceeding of
11th LICS, pages 118 -- 127. IEEE Computer Society Press, 1996.
[22] M.Y. Vardi and P. Wolper. Reasoning about infinite computations. Infor-
mation and Computation, 115(1):1 -- 37, 1994.
[23] G. Winskel. Event structures. In W. Brauer, W. Reisig, and G. Rozen-
berg, editors, Petri Nets: Applications and Relationships to Other Models
of Concurrency, LNCS 255, pages 325 -- 392. Springer-Verlag, 1987.
20
|
1805.03691 | 2 | 1805 | 2018-05-13T00:36:13 | Self-Stabilizing Task Allocation In Spite of Noise | [
"cs.MA",
"cs.DC"
] | We study the problem of distributed task allocation inspired by the behavior of social insects, which perform task allocation in a setting of limited capabilities and noisy environment feedback. We assume that each task has a demand that should be satisfied but not exceeded, i.e., there is an optimal number of ants that should be working on this task at a given time. The goal is to assign a near-optimal number of workers to each task in a distributed manner and without explicit access to the values of the demands nor the number of ants working on the task.
We seek to answer the question of how the quality of task allocation depends on the accuracy of assessing whether too many (overload) or not enough (lack) ants are currently working on a given task. Concretely, we address the open question of solving task allocation in the model where each ant receives feedback that depends on the deficit defined as the (possibly negative) difference between the optimal demand and the current number of workers in the task. The feedback is modeled as a random variable that takes value lack or overload with probability given by a sigmoid of the deficit. Each ants receives the feedback independently, but the higher the overload or lack of workers for a task, the more likely it is that all the ants will receive the same, correct feedback from this task; the closer the deficit is to zero, the less reliable the feedback becomes. We measure the performance of task allocation algorithms using the notion of regret, defined as the absolute value of the deficit summed over all tasks and summed over time.
We propose a simple, constant-memory, self-stabilizing, distributed algorithm that quickly converges from any initial distribution to a near-optimal assignment. We also show that our algorithm works not only under stochastic noise but also in an adversarial noise setting. | cs.MA | cs |
Self-Stabilizing Task Allocation In Spite of Noise
Anna Dornhaus1, Nancy Lynch2, Frederik Mallmann-Trenn∗2, Dominik Pajak2, and Tsvetomira Radeva2
1University of Arizona, Department of Ecology and Evolutionary Biology
2MIT, CSAIL
Abstract
We study the problem of distributed task allocation inspired by the behavior of social insects,
which perform task allocation in a setting of limited capabilities and noisy environment feedback.
We assume that each task has a demand that should be satisfied but not exceeded, i.e., there is an
optimal number of ants that should be working on this task at a given time. The goal is to assign
a near-optimal number of workers to each task in a distributed manner and without explicit access
to the values of the demands nor the number of ants working on the task.
We seek to answer the question of how the quality of task allocation depends on the accuracy
of assessing whether too many (overload) or not enough (lack) ants are currently working on a
given task. Concretely, we address the open question of solving task allocation in the model where
each ant receives feedback that depends on the deficit defined as the difference between the optimal
demand and the current number of workers in the task. The feedback is modeled as a random
variable that takes values lack or overload with probability given by a sigmoid of the deficit.
Each ants receives the feedback independently. The higher the overload or lack of workers for a
task, the more likely is that an ant receives the correct feedback from this task; the closer the deficit
is to zero, the less reliable the feedback becomes. We measure the performance of task allocation
algorithms using the notion of regret, defined as the absolute value of the deficit summed over all
tasks and summed over time.
We propose a simple, constant-memory, self-stabilizing, distributed algorithm that converges
from any initial distribution to a near-optimal assignment. We then show that our algorithm works
not only under stochastic noise but also in an adversarial noise setting. Finally we prove a lower
bound on the regret for any constant-memory algorithm, which matches, up to a constant factor,
the regret achieved by our algorithm.
keywords- distributed task allocation; self-stabilization; randomized algorithms; load balancing;
social insects
∗[email protected]
Contents
1 Introduction
1.1 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Model
2.1 Notation and Assumptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Noisy feedback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Regret metric, closeness and farness.
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
3 Results and Discussion
3.1 High-Level Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Results for Both Noise Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Results for Sigmoid Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Results for Adversarial Noise
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Algorithm Ant - Theorem 3.1
5 Algorithm Precise Sigmoid - Theorem 3.2
6 Conclusion and Open Problems
A Missing Proofs of Section 4 and Proof of Theorem 3.1
B Sigmoid Noise - Missing Proofs of Section 3.3
C Adversarial Noise
D Trivial Algorithm
D.1 Sequential Model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
D.2 Synchronous Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
E Auxiliary Tools
2
3
5
5
6
7
7
7
8
8
9
10
13
14
18
21
23
25
26
26
26
1
1
Introduction
Task allocation in social insect colonies is the process of assigning workers to tasks such as foraging,
scouting, nursing, etc. in a way that maximizes the reproductive success of the colony. Each ant can
work on each of the tasks but the demands of the tasks are different and might also vary over time.
The ants probably neither know the number of ants needed nor can count the current number of ants
working on a given task [20]. Therefore the allocation has to be performed based only on the feedback
from the tasks. The feedback, in biology called 'task stimulus', corresponds for example to sensing a
too-high temperature in the nest, seeing light through a hole in the nest-wall, or smelling a pheromone
produced by hungry brood. Despite using limited communication, local observations, and noisy sens-
ing, many ant species are known to excel at task allocation. How do ants perform task allocation and
what can we learn from their behavior?
In [11], the authors proposed a solution to the problem of task allocation in the case where the
feedback received by the ants is always correct. More precisely, in [11], each task j ∈ [k] has a demand
d(j), that represents the number of workers needed at that task. If the load, i.e., the number of ants
working on the task, exceeds the demand, then all ants receive feedback overload. Conversely, if the
load is below or equals the demand of the task, then all ants receive feedback lack. Such a feedback
function is rather unrealistic in ant colonies due to its sharp transition between overload and lack-
it requires each ant to be able to tell the difference between d(j) and d(j) + 1 number of workers at a
task. The authors in [11] therefore pose the open problem of considering a weaker, noisy version of the
binary-feedback-the focal point of this paper.
We study the performance of task allocation in two noise models: a realistic, stochastic model and
a more theoretical, adversarial model. In the stochastic model, the feedback from the tasks for each
ant in each step is a random variable with possible values lack and overload. The probability that
it takes value lack equals to the sigmoid function of the deficit (demand minus load1) of the task. In
the adversarial noise model, the feedback is deterministic and it is always correct if the absolute value
of the deficit is large but can be arbitrary if the absolute value of the deficit is small.
We assume that the ants regularly receive the feedback from all the tasks. However, there is a delay
between the moments when an ant collects the feedback and when it changes its allocation during which
other ants may also make some decisions. To model this delay we assume, similarly to [11], synchronous
rounds: at the beginning of each round each ant receives binary feedback of the load of the task. The
ants then concurrently make a decision of whether to join or to leave their current task.
How can the ants make independent decisions and achieve a 'good' task allocation in spite of out-
dated observations, in spite of noise and in spite of the lack of global information-not knowing the
demands nor the current load of a task?
In order to define what a 'good' task allocation means we propose to use the notion of regret. In-
tuitively, the regret in our setting measures the sum over all rounds over all tasks of the absolute value
of the deficits (demand minus load) of the tasks. Here we penalize overload and underload equally:
An underload corresponds to work that is not being done, and each ant exceeding the demand of a
task corresponds to work being wasted (or, even worse, sometimes the excessive number of workers in
a task may block each other and decrease the efficiency [15, 12]). Note, that we do not charge any cost
for switching tasks.
It turns out that any constant-memory algorithm for task allocation can be lower bounded in terms
of this 'cost' function. We show that the quality of task allocation is determined by the critical value
(see Section 2 for a formal definition). Intuitively, the critical value determines a value of the deficit
(seen as a fraction of the demand) for which the feedback is correct for each ant with high probabil-
ity. This corresponds to a smallest value of the deficit at which the sigmoid is very close to 1 and a
1Note, that the deficit can be negative.
2
largest value for which it is very close to 0. The lower bound shows that the regret of the optimal
constant-memory algorithm is linear in the sum of demands times this critical value.
We then provide a simple, constant-memory Algorithm Ant that utilizes the oscillations in the num-
ber of workers in each task in order to achieve a stable allocation in the synchronous model. The size of
the oscillation at each task in our algorithm is proportional to the demand of the task times the critical
value. In the light of our lower bound the algorithm achievess a constant factor approximation of the
optimal regret. Our algorithm assumes a slightly stronger form of synchronization where the ants take
actions through cycles (phases) of length two (for example, day and night) and all of the ants are at the
same step of the cycle2. Our algorithm is parameterized by a learning parameter γ that upper bounds
the critical value. Smaller values of γ yield better bounds for the regret but slower convergence time.
On a high level, in our Algorithm Ant, the ants repeatedly collect two samples each to asses which
decision (joining, leaving or remaining) would be optimal in terms of regret. To obtain the first sample,
all ants assigned to a task work on it. For the second sample ants collectively leave their tasks with
a small probability. If both samples of a working ant indicate an overload, then it will permanently
leave the task with a small probability. If both samples of an idle ant indicate underload, then such
ant will join this task (or if there are multiple such tasks then the idle ant joins one chosen uniformly
at random). This two-sample approach can be seen as computing the slope of the regret function in
each task. The ants collectively, in a distributed fashion, increase or decrease the loads in the tasks by
a small fraction depending on the computed "direction of the gradient" and learning parameter γ. The
algorithm bears similarities to gradient descent; in our case a noisy and distributed version. We show
that this algorithm achieves a task allocation that is up to a constant factor optimal with respect to
the aforementioned lower bound.
We then build upon our two-sample algorithm to derive a more theoretical algorithm Algorithm
Precise Sigmoid that uses more memory as well as longer synchronous phases (consisting of more than
two rounds). This algorithm, together with a lower bound, establish a theoretical tradeoff between
the memory available at each ant and the optimal regret that can be achieved by an algorithm using
this memory. Moreover, we also consider the adversarial noise setting in which we propose algorithm
Algorithm Precise Adversarial that also achieves a constant approximation of the best possible regret.
We also show that if the deficit is too close to 0 for some number of steps then (due to the noisy
feedback) it drastically increases. It is therefore impossible to keep the absolute value of the deficit
very small for too long. This means that small oscillations in the number of workers at each task are
unavoidable in any algorithm that 'tries' to achieve a good and stable allocation. Our proposed solution
is to avoid getting too close to 0 with the absolute value of the deficit and use the oscillations (jumping
between positive and negative deficit) to achieve stable allocation and asymptotically optimal regret.
1.1 Related work
Distributed task allocation in the context of social insect colonies and similar simple models have been
studied for years in both the theoretical distributed computing and the behavioral ecology communities.
The most related previous work on task allocation is [11], in which the authors also assume syn-
chronous rounds and binary feedback. The authors present a very simple algorithm that converges to
an almost-optimum allocation (the allocation that differs from the demand by at most 1 at each task)
and analyze its convergence time. Considering a noisy version of the model was left as an open question.
Moreover, the author of [29] provide a model similar to that of [11] but they also study different
versions of the feedback that ants receive from the environment, which varies in the amount of infor-
mation the ants receive about the deficits of the tasks. In short, the model consists of two feedback
components: a success component that informs each ant in each round whether it is successful, e.g.,
needed at the task it is currently working on, and a choice component that provides unsuccessful ants
2In a theoretical model, this can be achieved using one bit of additional memory for each worker and using very
limited communication, e.g. [4].
3
with an alternative task to work on. The results in [29] analyze the convergence time of task allocation,
and as such are not directly comparable to our work here. In [29], the noise model is very rudimentary:
in each round the feedback of the binary success component can be noisy for at most a small fixed
number of ants. The results do not generalize to our setting.
The problem of task allocation in social insect colonies has been well studied in the communi-
ties of theoretical and experimental biology. The observations show that social insect colonies are
self-organized, with no individuals directing the task choices of others, with interactions between indi-
viduals potentially affecting task selection [18]. Workers in a colony may switch tasks as needed [17],
although this may come at additional cost [25]. The concept of task switching gives rise to an intrigu-
ing question: what is the algorithm used by the ants to decide whether to switch and which tasks to
choose? Some notable examples of models of task allocation [3] include (1) the threshold-based model
where ants compared the stimulus of a task to their built-in threshold to determine whether to work
on a given task, and (2) the 'foraging for work' model [33] where the ants are believed to actively look
for work when they are idle or redundant in the current task. In some species the ants are believed
to choose the tasks based on physical suitability (physical polyethism) [24], whereas in other species,
the ants are physically similar and suitable to do any task (temporal polyethism) [5]. In this paper, we
assume that the ants are identical (no thresholds) and universal (they can work on any task).
Some biological studies have focused on the efficiency of the task allocation process itself, and how
it is determined by the specific algorithm used by the ants. For example, [27] and [13] model task allo-
cation determined by social interactions and response thresholds, respectively, and both demonstrate
that perfect task allocation of workers to tasks cannot be achieved, potentially due to the speed and
accuracy of task allocation trading off against each other. In [32], an algorithm of task allocation is
analyzed in a setting where there are no thresholds, the tasks are arranged in a line and there is no
noise in sensing of the demand. The goal of [32] is to explain the experimental observations where
certain tasks were preferred by older ants.
Additional factors such as individual experience, interactions with other workers, spatial and hier-
archical position in the colony, and random encounters with tasks are also known to affect the specific
task allocation mechanism employed [6, 14, 18]. Unfortunately, most often it is not precisely known
what is the actual algorithm that the ants use to select tasks or how the factors listed above interact
to produce variation in preferences across tasks or across individuals [28].
A key property that we observe in our results-oscillations in the task allocation behavior of ants-
is also a commonly observed biological phenomenon more generally known as cyclical activity patterns
[9]. Although the role of cyclical activity patterns is not completely understood [10], several studies
make conjectures that may be related to the conclusions in our paper. First, our assumption that
ants perform actions in synchronized rounds and phases as a means of introducing 'delay' between
one another's actions is also observed in biological studies. Ants perform actions in bursts of activity
and inactivity in order to clear stale information from spreading through the colony [30]. Second, our
results suggest that, assuming that ants have constant memory, and noisy environmental feedback, the
oscillations are inevitable as the deficit becomes small. We conjecture that such cyclic activity patterns
(switching between different tasks and being idle) are necessary and a product of the limitations of the
ants and the noisy feedback about number of workers at a task.
Another key assumption we make is that the noise follows a sigmoid function (also known as
a logistic sigmoid activation function). Such functions appear in countless biological contexts (e.g.
[16, 22, 31]), to model the uncertainty with which the ants sense the need for work at different tasks.
We believe that the versatility and applicability to the real-world problems of the sigmoid noise model
makes it a good choice to model the noise of the environment in our setting.
Finally, somewhat related load-balancing processes have been studied under the term user-based
migration in which the tasks move in a network of resources [1, 23, 2] by querying the load of the cur-
rent resource and moving to a neighbor in case of an overload. However, this line of research assumes
4
that each resource knows an upper bound on how many tasks it can accept. Furthermore, the setting
is noise-free and we hope that this paper can be used to derive more realistic models, in which the load
cannot be determined precisely as it is unclear how long each task will need to be processed.
2 Model
The sections consists of three parts. First, we present the model, the assumptions and the bulk of
the notation. Second, we define the noise models (adversarial noise model and sigmoid noise model).
Finally, we define the regret metric and the closeness of a task allocation.
t = d(j) − W (j)
2.1 Notation and Assumptions
We have a collection of n ants and k tasks where each task j ∈ [k] has a fixed demand d(j) meaning
, j ∈ [k]
that the task requires d(j) many ants assigned to it. Let d be the demand vector. Let W (j)
denote the load of resource j at time t, i.e., the number of ants performing the task. For task j ∈ [k]
we define the deficit as ∆(j)
and a negative deficit signifies an overload. Unless specified
otherwise, each ant has memory that is linear in the number of tasks, but independent of n. We assume
there is no communication among the ants.
We assume synchronous rounds each consisting of two sub-rounds. We define round t ≥ 1 to be the
time interval between times t − 1 and t. Fix round t. In sub-round 1 (of this round), each ant receives
noisy feedback of the load situation at time t − 1 (see definition below). Then, in sub-round 2-based
on the obtained feedback and the internal state (memory) of the ant-the ant decides whether to work
and on which task during round t. We will bundle two consecutive rounds into a phase and we assume
that each phase starts for all ants at the same time step.
t
t
We assume that the demands do not change, but our results trivially extend to changing demands
due to the self-stabilizing nature of our algorithms.
We now give our assumptions on the demand vector. First, we require the demands to be at least
of logarithmic size in the number of ants. Second we assume that there is is sufficient slack of the
demands, so that not all ants are required to work.
Indeed, it is commonly observed that a large
fractions of the ants do not work (e.g. [7, 8]).
Assumptions 2.1. Assume that for all j ∈ [k] we have d(j) = Ω(log n) for j ∈ [k]. Moreover, assume
that the sum of the demands satisfies(cid:80)
j∈[k] dj ≤ n/2.
We model ants as finite state automata and assume that the states of all the ants must be always
reachable from each other. We do not allow for example algorithms where an ant working on some task
can never leave this task. It is well-known (e.g. [19]) that ants do not stabilize to a fixed allocation
but switch between the tasks whenever it is needed.
Assumptions 2.2. We assume that for any pair of states s1, s2 (idle or working on one of the tasks
j ∈ [k]), there must exists a finite sequence of feedbacks F1, F2, . . . such that an ant in state s1 transi-
tions with nonzero probability to state s2.
We use the shorthand w.p.
'with high
probability' to mean that the event happens w.p. at least 1 − O(1/n). We say an event happens with
overwhelming probability is it happens w.p. at least 1 − e−Ω(n). We define the configuration at time τ
to be the internal state of all ants at time τ.
for 'with probability'. We say an event happens w.h.p.
5
2.2 Noisy feedback
We seek to model the noise in the sensing such that the following axioms are fulfilled. First, in case of
a very large deficit almost all ants will notice this (w.h.p. all ants receive feedback lack). Similarly,
in case of a high overload (negative deficit) almost all ants will notice this as well. Second, whenever
exactly the correct number of ants are working on a given task, then the 'uncertainty' in this task is
the largest and each ant receives lack and overload with equal probability. In the following we define
the sigmoid feedback and the adversarial feedback that both fulfill these axioms.
Sigmoid feedback The noisy feedback is modeled by a sigmoid function
s(x) =
1
1 + e−λx ,
(cid:40)
for fixed λ ∈ R. At the beginning of round t, each ant i receives for each task j noisy feedback F (j)
of the deficit:
t
(i)
F (j)
t
(i) =
lack
overload
with probability s(∆(j)
otherwise
t−1)
.
We write Ft(i) to denote the vector of feedbacks for all tasks that ant i receives at time t. It is not
crucial for our results to have a sigmoid function; in fact all our results apply for any monotone antisym-
metric function s with exponential decay and limx→−∞ s(x) = 0 and limx→∞ s(x) = 1 and s(0) = 1/2.
Adversarial feedback This feedback is parameterized by threshold γad ∈ R. The adversarial feed-
back F (j)
(i) received at the beginning of round t by each ant i for each task j is defined as:
t
lack
F (j)
t
(i) =
∆(j)
t−1 > γadd(j)
t−1 ∈ [−γadd(j), γadd(j)]
arbitrary value in {lack, overload} ∆(j)
t−1 < −γadd(j)
∆(j)
overload
.
Critical value The critical value is the deficit which ensures that all the ants receive with high
probability the correct feedback provided that the overload/lack is far enough away from 0. The idea
is that an ant will be able to see an overload (lack, respectively) if the number of ants working at a
task exceeds (1 + c)d (or is lower than (1− c)d, respectively) for some small constant c > 0 and demand
d. The value c is the critical value. Note that, by symmetry, a deficit of −cd also ensures that all ants
receive feedback overload.
Definition 2.3 (critical value and grey zone). We define the critical (feedback) value for the respective
models as follows.
• Sigmoid feedback model. Let y(x) = minx(cid:48)∈R{s(−x(cid:48) · d(j)) ≤ x :
for all j ∈ [k]}. We define the
critical (feedback) value to be γ∗ = y(1/n8). Observe that, due to the antisymmetry of the sigmoid,
s(−γ∗·d(j)) = 1−s(γ∗·d(j)). Wwe make no assumption on γ∗ apart from being smaller than 1/2.
• Adversarial feedback model. We define the critical (feedback) value γ∗ to be γad.
We define for each task j ∈ [k] the grey zone to be gj = [−γ∗d(j), γ∗d(j)].
6
Figure 1: Whenever the overload is in the green (red, respectively) region, all ants will receive w.h.p.
the feedback lack (overload, respectively). Whenever the overload is in the grey region (falling tilling
pattern), and the closer the overload is to 0, the more unpredictable is the feedback received by the ants.
2.3 Regret metric, closeness and farness.
In order to compare the quality of task allocation strategies we use the following notion of regret r(t),
defined for time t ∈ R as:
r(t) =
d(j) − W (j)
=
t
∆(j)
.
t
(cid:88)
j∈[k]
(cid:88)
j∈[k]
Furthermore, we define the (total) regret up to time t as:
R(t) =(cid:80)
i≤t r(i).
The idea behind the regret metric is that if over large scales of time the deficit is only rarely large,
then this should not affect the performance (i.e., the survival of the colony) too much. Moreover, as
we will see later, any constant-memory algorithm will have oscillations (or a significant deficit) and
hence classical metrics such as convergence time carry little meaning. It is worth pointing out that we
penalize lack and overload (measuring a waste of energy) equally and we leave it as a future direction
to use different weights.
We say that an algorithm produces an assignment that is c-close if the average regret of the alloca-
tion is at most c times the critical value and the sum of demands, i.e., limt→∞ R(t)
+
O(1). Similarly, we say an the assignment of an algorithm is c-far if the average regret of is at least c
times the critical value and the sum of demands, i.e., limt→∞ R(t)
cγ∗(cid:80)
t ≥(cid:16)
t ≤(cid:16)
cγ∗(cid:80)
j∈[k] d(j)(cid:17)
.
j∈[k] d(j)(cid:17)
3 Results and Discussion
In this section we present our results. We start with an overview of all the results. Second, we give
results that hold for both noise models (sigmoid and adversarial). We then give the results for the sig-
moid noise model followed by the results for the adversarial noise model. The corresponding algorithms
are defined in Section 4, Section 5 and Appendix C.
3.1 High-Level Overview
Our results on sigmoid noise are threefold.
(i) In our main theorem we show that the somewhat realistic Algorithm Ant with parameter γ results
γ∗ -close provided that γ ≥ γ∗. In particular, the assignment is
w.h.p. in an assignment that is 5 γ
5-close if γ = γ∗. (Theorem 3.1)
(ii) Second, we develop Algorithm Precise Sigmoid that is slightly more artificial but achieves arbi-
trary closeness: for any ε ∈ (0, 1), the algorithm produces w.h.p. an assignment that is ε-close
provided that γ = γ∗. The algorithm uses O(log(1/ε)) memory and synchronous phases of length
O(1/ε). (Theorem 3.2)
7
overload(−∆)γ∗d−γ∗d0greyzoneProb.ofreceivingfeedbackoverload121n6w.h.pallantsreceivefeedbacklackw.h.pallantsreceivefeedbacksurplus (iii) Finally, we show that this bound is in fact tight: any algorithm using c log(1/ε) memory, for
small enough constant c will produce, with overwhelming probability, an assignment that is at
least ε-far for almost all time steps. (Theorem 3.3)
In the adversarial noise setting, we prove the following bounds.
(i) Our Algorithm Ant also achieves in the adversarial noise setting w.h.p. an assignment that is
γ∗ -close. (Theorem 3.1)
5 γ
(ii) Second, we develop Algorithm Precise Adversarial that is slightly more artificial but achieves
for any ε ∈ (0, 1), the algorithm produces w.h.p. an assignment that is
arbitrary closeness:
(1 + ε)-close provided that γ = γ∗. The algorithm uses O(log(1/ε)) memory and synchronous
phases of length O(1/ε). (Theorem 3.6)
(iii) Any algorithm, regardless of the memory, produces an assignment that is in expectation at least
(1 − o(1))-far. (Theorem 3.5)
3.2 Results for Both Noise Models
We proceed by giving the precise statements of Theorem 3.1 showing that the assignment of Algorithm
Ant is w.h.p. 5 γ
γ∗ -close.
Theorem 3.1. The following holds for the sigmoid noise model as well as for the adversarial noise
model. Consider an arbitrary initial allocation at time 0. Fix an arbitrary t ∈ N. Algorithm Ant with
learning rate γ ≥ γ∗ has w.h.p. a total regret during the first t rounds that is bounded by
for some constant c.
R(t) ≤ c
nk
γ
+
5γ
(cid:88)
j∈[k]
· t,
d(j) + 3
Moreover for any interval of time of length at most n4, for each task j ∈ [k], the absolute value of
the deficit is w.h.p. in all but O(k log n/γ) rounds bounded by 5γd(j) + 3.
Note that we allow t to take arbitrary values-in particular, values that are super-exponential in n.
3.3 Results for Sigmoid Noise
In the following we show that Algorithm Precise Sigmoid is ε-close but requires synchronous cycles of
length O(1/ε). For applications where this is tolerable, the algorithm becomes superior and in fact
optimal w.r.t. to the closeness as we will see later.
Theorem 3.2. The following holds for the sigmoid noise model. Fix an arbitrary ε = Ω(1/ log n).
Algorithm Precise Sigmoid with parameter γ ≥ γ∗ and ε has a regret of
lim
t→∞
R(t)
t
= γε
d(j) + O(1),
(cid:88)
j∈[k]
where we suppress all terms that are independent of t. The algorithm uses O(log(1/ε)) memory and
synchronous rounds of length O(1/ε).
We finally show a lower bound on the memory size for any algorithm that is ε-close establishing
optimality of Algorithm Precise Sigmoid.
8
Furthermore, if the deficit for all tasks is below 2εγ∗d(j) for a constant number of consecutive steps
(of sufficient length), then, with overwhelming probability, there will be a task j with an oscillation
(i.e., a deficit) of order ω(γ∗d(j)).
In the proof of our lower we show that, unless the deficit is of order γ∗(cid:80)
j∈[k] d(j) oscillations are
unavoidable.
The lower bound focuses on the perpetual cost of the algorithm that, once t is large enough,
marginalizes the initial costs. Nevertheless, we point out that if the demands can change arbitrarily in
any round, then any algorithm that does not know when the demands, will pay an initial regret of Ω(n).
Remark 3.4. The guarantees from Theorem 3.1 and Theorem 3.2 even apply if the feedback is arbi-
trarily correlated as long as the marginal probability for each ant to receive incorrect feedback outside
the grey zone is 1/nc for some small constant c > 1.
Moreover, our algorithm trivially also works-due to its self-stabilizing nature-for changing de-
mands.
Theorem 3.3. The following holds for the sigmoid noise model. Assume γ∗ ≤ 1 and let c ≤ 1 be a
small enough constant. Let ε ∈ (0, 1/4) and let n (number of ants) be large enough integer. There exists
a demand vector (d(1), d(2), . . . , d(k)) such that for any collection of n ants executing (possibly distinct)
algorithms A1, A2, . . . , An, each using at most c(cid:98)log(1/ε)(cid:99) bits of memory, the following holds. For any
√
time t ≥ 1/
ε with overwhelming probability
R(t) ≥
εγ∗(cid:88)
j∈[k]
t.
d(j)
Furthermore, even though we state Algorithm Ant and Algorithm Precise Sigmoid in a way that the
feedback of all the tasks is collected (as in [11]), this is not necessary and only the initial cost would
change if each ant could only receive feedback from one (adaptively) chosen task. In addition for the
ease of presentation, our algorithms store the samples inefficiently (e.g. the median). The bounds we
claim on the space complexity assume slightly smarter, but obvious techniques.
constant need not be 1/2; it must only guarantee that(cid:80)
Finally, the required bound on the sum of demands of Assumptions 2.1 can be relaxed and the
j∈[k](1+5γ∗)d(j) ≤ c∗n for some constant c∗ < 1.
3.4 Results for Adversarial Noise
In the following we present our results for the adversarial noise model.
Theorem 3.5. In the adversarial noise model, any algorithm, possibly randomized, using unlimited
memory and communication has expected regret of at least:
E [ R(t) ] ≥ (1 − o(1))tγ∗(cid:88)
d(j).
j∈[k]
Note that in the sigmoid noise model a ε-closeness is possible showing a separation between the
two models. Using a modification of Algorithm Ant we can achieve (1 + ε)-closeness.
Theorem 3.6. Let γ∗ ≤ 1/4. Fix an arbitrary ε = Ω(1/ log n). Algorithm Precise Adversarial with
parameter γ ≥ γ∗ using synchronized cycles of length O(1/ε) and memory O(log(1/ε)) achieves in the
adversarial noise model the following asymptotic bound on regret:
lim
t→∞
R(t)
t
= γ(1 + ε)
d(j) + O(1).
(cid:88)
j∈[k]
Interestingly, our algorithm also minimizes the total number of switches of ants between tasks in
comparison to Algorithm Ant; this might be of interest, if one changes the regret to incorporate costs
for switching between tasks.
9
4 Algorithm Ant - Theorem 3.1
In this section we introduce Algorithm Ant. We start by giving the intuition and we refer the reader to
page 11 for full definition of the algorithm. We present the intuition behind the algorithm by focusing
on the single task setting, however, the full algorithm is written and analyzed in the general setting
with k > 1. We divide time into phases, where each phase consists of two consecutive rounds. This
allows each ant in each phase to take two samples where a sample is simply the binary feedback (lack
or overload) for the task. We say that the samples are taken at different points which means that the
number of ants working at the considered task is temporarily reduced between the first and the second
sample. The intuition of taking two samples is as follows. If both samples indicate an overload, then
since the second sample is taken while the number of workers is reduced (and it still shows overload),
we should decrease the number of workers in the task. If one sample indicates overload and the other
lack, then load is likely to be close to the optimal demand and there is no need to change the number
of ants working on the task in the next phase. Finally, if both samples indicate a lack, then additional
ants are required to join the task.
How can we, without any central control, take two samples of feedback for different values of load?
The ants achieve it by independently with probability Θ(γ), temporarily, pausing their work on the
current task. With this idea we can obtain two samples that are taken at two different values of load in
the task, and the distance between the samples can be regulated by the probability of pausing. More
precisely, assume that W ants are working on some fixed task with demand d in the first round of some
phase. Then the first sample is simply the feedback of this load (i.e., lack with probability s(W − d)
and overload otherwise). The second sample (in the second round of the phase) is taken after each ant
independently stopped working with probability csγ, where cs is a constant. Therefore, the number of
ants working in the second sample is roughly W(1 − csγ) (see Figure 2 for an illustration).
Why do we need to take two samples that are 'spaced' apart? The reason is that if the deficit of
task j ∈ [k] in the current step is in the range [−γ∗d(j), γ∗d(j)] ("grey zone") then the feedback in this
step might be unreliable. In our algorithm we take two samples-spaced far enough apart-so that at
least one of them must lie w.h.p. outside of the above range. When the deficit is outside this range,
then with high probability, all the ants receive the same, correct feedback. If W ≥ d(1 + γ∗) then the
first sample shows overload hence after this phase we may decrease the number of ants in this task
but we cannot increase (to increase we need lack in both samples). Similarly if W ≤ d(1 + γ∗) then
the second samples shows lack (because the samples are sufficiently spaced apart) hence we cannot
decrease but we may increase in this phase. This shows that, informally speaking, the load of a task
w.h.p. can move only in the correct direction (if we look only at the loads after the decision at the
end of each phase and disregard the loads in the second sample in each phase). Finally we can identify
a "stable zone" in which the first sample lies on the right side of grey zone (Figure 2) and the second
sample lies to left of the grey zone. Note that once the load of a task is in the stable zone, it can
neither increase nor decrease (because the first sample always gives overload and the second lack).
Analysis The idea of our proof is to analyze the regret in three different regions. For this we split the
regret R(t) into R+(t), R≈(t) and R−(t) which intuitively measure the regret in the overload region,
the region that is close to 0 deficit and in a region in which there is a lack of ants.
We divide time into intervals of length n4 and show, by using a series of potential arguments, that
R−(t) is w.h.p. bounded by O(nk) during any fixed interval. We do this by arguing that in any phase
w.h.p. either the 'lack' of the resources decreases by a significant amount or the number of tasks that
are underloaded does. It turns out that both quantities are monotonically decreasing every second
round and will w.h.p. quickly reach 0. We then bound R+(t), by showing, that w.h.p. each task can
only cause at most once in the interval an overload of Θ(n) and once this happens, the overload of
the task will decrease in a geometrical fashion. Summing over all tasks gives w.h.p. a total regret of
10
Figure 2: A typical execution of one phase in which two samples are taken-each ant temporarily
drops out with probability linear in γ ≈ γ∗.
It is convenient to think of the process as gradient
descent-like with learning rate γ. The left figure shows a phase in which both samples for all ants
indicate an overload (red). As a result, at the end of the phase, a few ants drop out-again with w.p.
linear in γ; this time the drop out is 'permanent'.
The next phase is depicted in the right figure. Here the first sample indicate an overload (for all ants
w.h.p.), the second a lack (green) for ants (for all ants w.h.p.). From here on no ant will join or leave
the task for a polynomial number of steps (w.h.p.).
t )j∈[k] for each task, learning parameter γ ∈ [γ∗, 1/16]
Algorithm Ant
(see Theorem 3.1)
3
Input at round t: Noisy feedback Fj = (F j
Output at round t: Assignment of the ant to a task at ∈ {idle, 1, 2, . . . , k}
1: cd ← 19 and cs ← 2 1
2: for At every step t ≥ 1 do
t mod 2 = 1 then
3:
currentT ask ← at−1
4:
s1 ← Ft
5:
if currentT ask (cid:54)= idle, then at ←
t mod 2 = 0 then
s2 ← Ft
[receive feedback]
if currentT ask = idle then
idle
currentT ask
w.p. csγ
otherwise
[receive feedback]
(cid:40)
if
if
underloadedT asks ← {j ∈ [k] : s(j)
at ←
Uniform (underloadT asks)
idle
1 = s(j)
2 = lack}
if underloadT asks (cid:54)= ∅
otherwise
6:
7:
8:
9:
10:
11:
12:
13:
(cid:40)
(cid:40)
else
at ←
output at
idle
currentT ask
w.p. γ/cd if scurrentT ask
otherwise
1
= scurrentT ask
2
= overload
O(nk). Finally, R≈(t) can simply, by definition, be bounded by O(γ(cid:80)
this is, up to constants, tight as our lower bounds show.
j∈[k] d(j)) per time step; in fact
In the remainder we give the formal statements. See Appendix A for the missing proofs.
In the following claims we consider a continuous interval of time steps I of length t ≤ n4 and we
will analyze the dynamics of the system in this interval assuming an arbitrary initial configuration
in step 0. For each ant i we define ai
t to be the task the ant i is assigned to at the end of round
0 denotes the initial assignment of ants. We start by
t, i.e., ai
defining the pair of 'good' events that occur w.h.p. during I. For each task j we define the grey zone
gj = [−γ∗d(j), γ∗d(j)] which intuitively corresponds to deficits, in which the feedback is very 'noisy'.
t ∈ {idle, 1, 2, . . . , k}. In particular, ai
11
overload(−∆)regretγ∗d−γ∗d0greyzoneoverload(−∆)regretγ∗d−γ∗d0greyzone Let Ef eedback be the event that for any step t of the interval I, any task j, and any ant i the
t−1 − d(j) (cid:54)∈ gj then the feedback received by ant i for task j is correct, meaning
t−1 ≥ γ∗d(j) and F j
Let Econcentration be the event that for any step t of the interval I, such that t mod 2 = 0 and for
following holds. If W (j)
t = lack if ∆(j)
that F j
t = overload if ∆(j)
t−1 ≤ −γ∗d(j).
any task j such that W (j)
t ≥ (1 − γ)d(j) we have:
(1 − 0.9csγ)],
t
t
(1 − 1.1csγ), W (j)
(1 − 3γ
t+1 ∈ [W (j)
t+2 ≥ W (j)
t ≥ d(j)(1 + (1 + 1.2cs)γ), then W (j)
• W (j)
• W (j)
• if W (j)
In the following we show, using standard techniques, that w.h.p. each of these events occurs.
t+2 ≤ W (j)
(1 − γ
2cd
2cd
),
).
t
t
Claim 4.1. We have that:
1. P [Ef eedback ] ≥ 1 − n−2 and
2. P [Econcentration ] ≥ 1 − 2n−2,
In the following we condition on both of these events.
Claim 4.2. Conditioned on events Ef eedback and Econcentration. For each task j ∈ [k] there exists at
most one time step τj ∈ I such that W (j)
τj−1 and if τj exists, then for any τ ∈ I
τj ≥ d(j)(1 + γ) > W (j)
such that τ > τj and τ mod 2 = 0 we have:
τ ≥ d(j)(1 + γ).
τj ≥ W (j)
W (j)
In the following, we will split the cost of the regret into three different more tractable costs: the
regret induced by a significant overload R+(t), by being close to the demands R≈(t) and by significant
lacks R−(t). Let c+ = 1.2cs and c− = 1 + 1.2cs Let
(W (j)
t − (1 + c+γ)d(j)) · 1
) · 1
((1 − c−γ)d(j) − W (j)
t
W (j)
t −(1+c+γ)d(j)>0
(1−c−γ)d(j)−W (j)
t >0
(cid:88)
(cid:88)
j∈[k]
j∈[k]
r+(t) =
r−(t) =
(cid:88)
(cid:88)
γ
r≈(t) = r(t) − r+(t) − r−(t).
(cid:88)
Similarly, define
R+(t) =
r+(τ ),
R≈(t) =
r≈(τ ),
and R−(t) =
We have R(t) = R+(t) + R≈(t) + R−(t).
τ≤t
We will simply bound R≈(t) by using its definition resulting in R+(t) ≤ 2kncd
τ≤t
τ≤t
r−(τ ).
. Bounding the other
two quantities is more involved; we will start by bounding R+(t).
Using Claim 4.2 we can upper bound the total regret related to the overload in all the tasks. The
idea of the proof is that for a fixed task, we have that during the interval I, there can only be at most
one significant increase in the load. After such an increase in the load of that task j, the number of
ants working on task j will from there on drop every phase by roughly a factor γ until the regret due
to overload (r+(τ )) becomes zero. Integrating over the entire period results in a geometric series and
summing over all tasks gives a regret of R+(t) ≤ 2nkcd
Claim 4.3. Conditioned on Ef eedback and Econcentration. If γ ≤ 1/16, then we have that: R+(t) ≤ 2nkcd
and r+(τ ) > 0 for at most 4kcd log n/γ different rounds of τ ∈ I.
.
γ
γ
,
12
We say that all tasks are saturated in step t if for all tasks j ∈ [k] we have in step t that
t ≥ d(j)(1 − γ).
W (j)
Claim 4.4. Condition on Ef eedback and Econcentration. If in time step t ∈ I all tasks are saturated, then
1. in any step t(cid:48) > t such that t(cid:48) mod 2 = 0, all tasks are saturated,
2. in any step t(cid:48) ≥ t, r−(t(cid:48)) = 0.
In order to bound R−(t) we use the following potential functions, which measure the number
of ants working on a significantly overloaded task and the number of significantly overloaded tasks,
respectively. For any t ∈ N let
Φ(t) =
and
((1 + γ)d(j) − W (j)
2t ) · 1
(1+γ)d(j)>W (j)
2t
,
(cid:88)
j∈[k]
(cid:88)
j∈[k]
Ψ(t) =
1
(1+γ)d(j)>W (j)
2t
.
Observe that functions Φ and Ψ are defined for phase numbers not for steps. Phase number t consists
of two time steps 2t and 2t + 1. In the following we show that after 2 phases we have that either one
of three events must occur: 1) the number of ants working on underloaded tasks increases significantly
2) the number of saturated tasks increases 3) all tasks are saturated.
Claim 4.5. Condition on Ef eedback and Econcentration. We have that the functions Φ(t) and Ψ(t) are
non-increasing and moreover for an arbitrary t such that in step 2t not all tasks are saturated and
2t ≤ n4 − 4, we have that at least one of the following three events must happen:
1. Φ(t + 2) − Φ(t) ≤ −cγn for some constant c > 0
2. Ψ(t + 2) − Ψ(t) ≤ −1,
3. all tasks are saturated in step 2(t + 2).
From this we are able to derive a bound on R−(t) and by putting everything together we derive
the desired bound on R(t) = R+(t) + R≈(t) + R−(t).
Lemma 4.6. Condition on Ef eedback and Econcentration. We have that
R(t) ≤ ckn
γ
+ c
γtd(j),
(cid:88)
j∈[k]
where c is a constant. Moreover for all tasks j ∈ [k] the absolute value of the deficit is bounded by
5γd(j) in all but O(k log n/γ) rounds.
From Lemma 4.6 and Claim 4.1 we are able to establish our main theorem Theorem 3.1. Since the
claims have been shown for any initial configuration we can simply cut the time horizon into intervals
of length at most n4 and use the bounds on the regret for each interval. See Appendix A for the proof.
5 Algorithm Precise Sigmoid - Theorem 3.2
In this section we present Algorithm Precise Sigmoid that is built on the foundations of Algorithm Ant
and to takes advantage of the sigmoid noise to achieve an even better total regret than Algorithm Ant.
Recall that all Algorithm Ant requires is that whenever the deficit is outside of the grey zone, then the
probability to receive an incorrect feedback is at most 1/n8. Now if we choose a step size of roughly εγ
then, due to the sigmoid noise, the probability of failure increases roughly to 1/nε8. However, by taking
13
multiple samples instead of a single sample and by computing the median, we can amplify the proba-
bility that the median sample is correct to 1− 1/n8. This means that the result of Theorem 3.1 applies
with the same guarantees but at a much smaller step size, ultimately resulting in a much smaller regret.
More precisely, the modified version uses a step size of εγ/cχ and relies on synchronized phases of
length 2m, with m = (cid:100)2cχ/ε + 1(cid:101). Each such phase consists of m rounds during which the algorithm
calculates the median of the feedback. This results in 2 samples per phase just as in Algorithm Ant
and the rest of the algorithm is exactly the same as Algorithm Ant. Up to the different step size and
the computation of the median, the proof is along the same lines as the proof of Algorithm Ant. See
Appendix B for the proof.
Algorithm Precise Sigmoid
(see Theorem 3.2)
t )j∈[k] for each task, learning parameter γ < 1/2,
3
precision parameter ε < 1.
Input at round t: Noisy feedback Ft = (F j
Output at round t: Assignment of the ant to a task at ∈ {idle, 1, 2, . . . , k}.
1: cχ ← 10, cd ← 19 and cs ← 2 1
2: m ← (cid:100)2cχ/ε + 1(cid:101)
3: for At every step t ≥ 1 do
r ← t mod 2m
4:
if r = 1 then
5:
6:
if r ∈ [1, m] then
7:
8:
9:
10:
11:
currentT ask ← at−1
sr ← Ft
at ← at−1
s1 ← median(sr(cid:48) : r(cid:48) ∈ [1, m])
if currentT ask (cid:54)= idle, then at ←
if r ∈ [m + 1, 2m − 1] ∪ {0} then
w.p. εcsγ/cχ
otherwise
idle
currentT ask
[receive feedback]
if r = m then
(cid:40)
[receive feedback]
sr ← Ft
at ← at−1
s2 ← median(sr(cid:48) : r(cid:48) ∈ [m + 1, 2m − 1] ∪ {0})
if currentT ask = idle then
if r = 0 then
underloadedT asks ← {j ∈ [k] : s(j)
at ←
Uniform (underloadT asks)
idle
1 = s(j)
2 = lack}
if underloadT asks (cid:54)= ∅
otherwise
idle
currentT ask
w.p. γ/(cχcd) if scurrentT ask
otherwise
1
= scurrentT ask
2
= overload
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
(cid:40)
(cid:40)
else
at ←
output at
6 Conclusion and Open Problems
We presented a simple proof-of-concept algorithm that achieves, given a suitable learning parameter γ,
a fairly good assignment of ants. The algorithm is very resilient to noise, changes in demands, changes
of the number of ants and even changes of the number of tasks. The algorithm embraces the seeming
14
obstacle of full-synchronization (which we introduced to model the delay of information) to perform
controlled oscillations. It would be interesting to see if variations of this algorithm also work in settings
of less synchronization.
Moreover, it remains an open problem to understand if and by how much simple communication
among the ants can help. In the adversarial setting for example, it is clear that even unlimited com-
munication cannot help. This leads to the question of which other noise models would make sense to
study and how to design experiments with real ants to gather more knowledge about the way noise
affects the sensing.
15
References
[1] H. Ackermann, S. Fischer, M. Hoefer, and M. Schöngens. Distributed algorithms for QoS load
balancing. Distributed Computing, 23(5-6):321–330, 2011.
[2] P. Berenbrink, T. Friedetzky, F. Mallmann-Trenn, S. Meshkinfamfard, and C. Wastell. Threshold
load balancing with weighted tasks. J. Parallel Distrib. Comput., 113:218–226, 2018.
[3] S. N. Beshers and J. H. Fewell. Models of division of labor in social insects. Annual review of
entomology, 46(1):413–440, 2001.
[4] L. Boczkowski, A. Korman, and E. Natale. Minimizing message size in stochastic communication
patterns: Fast self-stabilizing protocols with 3 bits. In Proceedings of the Twenty-Eighth Annual
ACM-SIAM Symposium on Discrete Algorithms, SODA 2017, Barcelona, Spain, Hotel Porta
Fira, January 16-19, pages 2540–2559, 2017.
[5] P. Calabi. Behavioral flexibility in hymenoptera: a re-examination of the concept of caste.
Advances in myrmecology, pages 237–258, 1988.
[6] D. Charbonneau and A. Dornhaus. When doing nothing is something. How task allocation mech-
anisms compromise between flexibility, efficiency, and inactive agents. Journal of Bioeconomics,
17:217–242, 2015.
[7] D. Charbonneau, C. Poff, H. Nguyen, M. C. Shin, K. Kierstead, and A. Dornhaus. Who are the
"lazy" ants? the function of inactivity in social insects and a possible role of constraint: Inactive
ants are corpulent and may be young and/or selfish.
Integrative and Comparative Biology,
57(3):649–667, 2017.
[8] D. Charbonneau, T. Sasaki, and A. Dornhaus. Who needs "lazy" workers? inactive workers act
as a reserve labor force replacing active workers, but inactive workers are not replaced when they
are removed. PLOS ONE, 12(9):1–20, 09 2017.
[9] B. J. Cole. Short-term activity cycles in ants: generation of periodicity by worker interaction.
The American Naturalist, 137(2):244–259, 1991.
[10] B. J. Cole and F. I. Trampus. Activity cycles in ant colonies: worker interactions and decentralized
control. In Information processing in social insects, pages 289–307. Springer, 1999.
[11] A. Cornejo, A. R. Dornhaus, N. A. Lynch, and R. Nagpal. Task allocation in ant colonies. In
Distributed Computing - 28th International Symposium, DISC 2014, Austin, TX, USA, October
12-15, 2014. Proceedings, pages 46–60, 2014.
[12] T. J. Czaczkes, C. Grüter, and F. L. Ratnieks. Negative feedback in ants: crowding results in
less trail pheromone deposition. Journal of the Royal Society Interface, 10(81):20121009, 2013.
[13] A. Duarte, I. Pen, L. Keller, and F. J. Weissing. Evolution of self-organized division of labor in
a response threshold model. Behavioral Ecology and Sociobiology, 66(6):947–957, 2012.
[14] A. Duarte, F. J. Weissing, I. Pen, and L. Keller. An evolutionary perspective on self-organized
insects. Annual Review of Ecology, Evolution, and Systematics,
division of labor in social
42:91–110, 2011.
[15] A. Dussutour, S. C. Nicolis, J.-L. Deneubourg, and V. Fourcassié. Collective decisions in ants
when foraging under crowded conditions. Behavioral Ecology and Sociobiology, 61(1):17–30, 2006.
16
[16] C. A. Edwards and P. J. Bohlen. Biology and ecology of earthworms, volume 3. Springer Science
& Business Media, 1996.
[17] D. M. Gordon. Dynamics of task switching in harvester ants. Animal Behaviour, 38(2):194–204,
1989.
[18] D. M. Gordon. The organization of work in social insect colonies. Nature, 380(14 March):121–124,
1996.
[19] D. M. Gordon. Ants at work: how an insect society is organized. Simon and Schuster, 1999.
[20] D. M. Gordon and N. J. Mehdiabadi. Encounter rate and task allocation in harvester ants.
Behavioral Ecology and Sociobiology, 45(5):370–377, Apr 1999.
[21] T. Hagerup and C. Rüb. A guided tour of chernoff bounds.
33(6):305–308, 1990.
Information Processing Letters,
[22] M. Hassell and H. Comins. Sigmoid functional responses and population stability. Theoretical
Population Biology, 14(1):62–67, 1978.
[23] M. Hoefer and T. Sauerwald. Brief announcement: Threshold load balancing in networks.
In
32nd Symposium on Principles of Distributed Computing PODC, pages 54–56, Montreal, Canada,
2013. ACM. The full version is available at https://arxiv.org/abs/1306.1402.
[24] B. Hölldobler and E. O. Wilson. The ants. Harvard University Press, 1990.
[25] G. M. Leighton, D. Charbonneau, and A. Dornhaus. Task switching is associated with temporal
delays in temnothorax rugatulus ants. Behavioral Ecology, 28(1):319–327, 2016.
[26] M. Mitzenmacher and E. Upfal. Probability and Computing: Randomized Algorithms and
Probabilistic Analysis. Cambridge University Press, 2005.
[27] H. M. Pereira and D. M. Gordon. A trade-off in task allocation between sensitivity to the
environment and response time. Journal of Theoretical Biology, 208(2):165–184, 2001.
[28] N. Pinter-Wollman, J. Hubler, J. A. Holley, N. R. Franks, and A. Dornhaus. How is activity
distributed among and within tasks in Temnothorax ants? Behavioral Ecology and Sociobiology,
66(10):1407–1420, 2012.
[29] T. Radeva, A. Dornhaus, N. Lynch, R. Nagpal, and H.-H. Su. Costs of task allocation with local
feedback: Effects of colony size and extra workers in social insects and other multi-agent systems.
PLoS computational biology, 13(12):e1005904, 2017.
[30] T. O. Richardson, J. I. Liechti, N. Stroeymeyt, S. Bonhoeffer, and L. Keller.
Short-term
activity cycles impede information transmission in ant colonies. PLoS computational biology,
13(5):e1005527, 2017.
[31] H. R. Thieme. Mathematics in population biology. Princeton University Press, 2003.
[32] C. Tofts. Algorithms for task allocation in ants.(a study of temporal polyethism: theory). Bulletin
of mathematical biology, 55(5):891–918, 1993.
[33] C. Tofts and N. R. Franks. Doing the right thing: ants, honeybees and naked mole-rats. Trends
in ecology & evolution, 7(10):346–349, 1992.
17
[34] A. C.-C. Yao. Probabilistic computations: Toward a unified measure of complexity. In Proceedings
of the 18th Annual Symposium on Foundations of Computer Science, SFCS '77, pages 222–227.
IEEE Computer Society, 1977.
A Missing Proofs of Section 4 and Proof of Theorem 3.1
t
Proof of Claim 4.1. The probability that any ant receives an incorrect feedback from task j in step t
assuming that ∆(j)
≥ γ∗dj is by the definition of γ∗ at most n−8. By taking union bound over n4
step, at most n ants and at most n tasks we get that the total probability of all 'bad' events is at most
n−2, hence P [Ef eedback ] ≥ 1 − n−2.
In the following we condition on Ef eedback. Fix any task j and round t such that t mod 2 = 0. We
assume that γ ≤ 1/16 and W (j)
t ≥ (1 − γ)d(j) ≥
(1 − γ)50000 log n/γ2 ≥ 120 · max{c2
t ≥ (1 − γ)d(j). By Assumptions 2.1 we have W (j)
Recall that W (j)
sample for the considered task j. We have E(cid:104)
P(cid:104)
(cid:105)
t+1 ≥ (1 + δ2)E(cid:104)
for any 0 < δ2 < 1:
t+1 denotes the random variable indicating the number of ants working in the second
(1 − csγ). By Chernoff bound we have
(cid:105)(cid:105) ≤ e
(cid:114) 24 log n
E(cid:104)
(cid:105) ≤ 1 each of these probabilities is upper bounded by n−8. Moreover
(cid:115)
t+1 ≤ (1 − δ2)E(cid:104)
(cid:105)(cid:105) ≤ e
and by choosing δ2 =
we have, since γ ≤ 1/16, cs < 1/(2γ), and w1 ≥ 120c2
d} · log n/γ2.
W (j)
t+1
W (j)
t+1
W (j)
t+1
W (j)
E(cid:104)
E(cid:104)
s, c2
W (j)
t+1
W (j)
t+1
W (j)
t+1
−δ2
−δ2
(cid:105)
/2
/3
,
t
2
2
W (j)
= W (j)
(cid:105)
, P(cid:104)
(cid:115)
(cid:105) ≥ 1 − n−6, because W (j)
t+2 ≤ (1 − δstay)E(cid:104)
s log n/γ2 that:
≤ csγ
24 log n
10
w1(1 − csγ)
,
t
Hence P(cid:104)
E(cid:104)
t
24 log n
E [ W3 ]
δ2 =
=
(1 − 1.1csγ), (1 − 0.9csγ)]
t
t
t
t
t
,
/2
stay
(cid:105)
W (j)
t+2
W (j)
t+2
W (j)
W (j)
W (j)
t+2
W (j)
t+2
P(cid:104)
t+1 ∈ [W (j)
W (j)
(cid:105)(cid:105) ≤ P(cid:104)
(1 − 1.1csγ) and W (j)
(1 − csγ)(1 − csγ/10) ≥
(1 − 0.9csγ). Now if in step t + 2 some ants
W (j)
decide to transition to state idle then each of them decides independently with probability γ/cd hence
(1 − csγ)(1 + csγ/10) ≤ W (j)
(1 − γ/cd). We have by Chernoff bound
(cid:105) ≥ W (j)
t+2 ≤ (1 − γ/(2cd))E(cid:104)
(cid:114) 24 log n
E(cid:104)
t ≥ d(j)(1 + (1 + 1.2cs)γ). We have with probability 1 − n−8 that W (j)
t+1 ≥
d(j)(1 + 2.2csγ)(1 − 0.9csγ) = d(j)(1 + 2.1γ − 1.98γ2) ≥ d(j)(1 + γ), where in the last inequality we use
the fact that γ ≤ 1/16. Then, conditioned on Ef eedback, each ant working on task j receives feedback
overload in both samples in step t and t + 1. By the definition of the algorithm, Each such ant leaves
the task in step t + 2 with probability γ/cd. Let random variable Wleave denote the number of ants
that leave task j in step t + 2. We can bound the value of this variable using Chernoff bound:
by choosing δstay =
(1 − γ/cd)(1 − γ/(2cd)) ≥ (1 − 3γ/(2cd))).
(cid:105) ≤ γ/(2cd), we get that P(cid:104)
E(cid:104)
(cid:105) ≤ n−8 (because
−δ2
W (j)
t+2
(1 − 3γ/(2cd))
(cid:105)(cid:105) ≤ e
t+2 ≤ W (j)
W (j)
Assume that W (j)
P [ Wleave ≤ E [ Wleave ] /2Ef eedback ] ≤ P [ Wleave ≤ (1 − δleave)E [ Wleave ] ] ≤ e−δ2
(cid:113) 18 log n
E[ Wleave ] ≤ 1/2, the inequality is true with probability at most n−9. We ob-
t /cd we have that with probability of at most n−9 at most
E[ Wleave ]/2,
t γ/(2cd) ants decide to leave task j in step t + 2. Thus P(cid:104)
and by choosing δleave =
serve that since E [ WleaveEf eedback ] = γW (j)
W (j)
n−9.
We showed that each bad event happens with probability at most n−8. By the union bound
over all ants (at most n), steps (n4) and tasks (at most k) and 4 different bad events we get that
(1 − γ/(2cd))Ef eedback
t+2 ≥ W (j)
W (j)
(cid:105) ≤
leave
t
18
with probability at least 1 − n−2 no bad event happens in the considered interval. This shows
that P [EconcentrationEf eedback ] ≥ 1 − n2 but since P [Ef eedback ] ≥ 1 − n−2 we get P [Econcentration ] ≥
P [EconcentrationEf eedback ] P [Ef eedback ] ≥ (1 − n−2)2 ≥ 1 − 2n−2.
Second, observe that if W (j)
Proof of Claim 4.2. First observe that, since the number of ants working on a task increases only in
τ ≥ dj(1 + γ), then the smallest
steps τ such that τ mod 2 = 0 then if there exists τ such that W (j)
such τ (call it τj) must satisfy τj mod 2 = 0. Fix a task j and assume that τj exists. We observe that
conditioned on Ef eedback all idle ants receive feedback overload in the first sample for task j in step
τj. Hence no ant can join task j in this phase (recall that in order to join, both samples of the phase
need to show lack). This shows that whenever the load is at least dj(1 + γ) in the first step of a phase
then it cannot be larger in the first step of the next phase.
τj ≤ d(j)(1 + (0.9cs − 1)γ) then conditioning on event Econcentration, the
τj (1 − 0.9csγ) ≤ d(j)(1 − γ) thus
second sample in the phase starting in step τj satisfies W (j)
conditioning on Ef eedback, all ants receive feedback lack in the second sample thus no ant can leave this
task in this phase. These two claims show the existence of a stable zone [d(j)(1+γ), d(j)(1+(0.9cs−1)γ)]
(which is nonempty because cs > 2), where the number of ants in the task can neither increase nor
decrease (if we look only at steps τ such that τ mod 2 = 0).
τj ≥ d(j)(1 + (0.9cs − 1)γ), then conditioning on event Econcentration al-
τj (1− 3γ/(2cd)) ≥
ways at least W (j)
d(j)(1 + (0.9cs − 1)cs)(1 − 3γ/(2cd)). The last expression is lower bounded by d(j)(1 + γ) if cs ≥
20/9 + 2/(cd − 1). Observe that with our choice of cs and cd this holds. This shows that it is not
possible to 'jump over' the stable zone.
Now we can inductively consider steps τj + 2, τj + 4, . . . and assuming that τj + 2i, τj + 2(i + 1) ∈ I
(1− 3γ/(2cd)) ants remain in this task. We have then W (j)
Observe moreover that if W (j)
τj +1 ≤ W (j)
τj +2 ≥ W (j)
τ
and conditioned on Econcentration and Ef eedback we have that W (j)
τj +2i ≥ W (j)
τj +2(i+1) ≥ d(j)(1 + γ).
Proof of Claim 4.3. Fix any task j ∈ [k]. Let us define the regret due to overload with respect to task j:
r+
j (τ ) = (W (j)
τ − (1 + c+γ)d(j)) · 1
W (j)
τ −(1+c+γ)d(j)>0
and
Clearly R+(t) =(cid:80)
j∈[k] R+
j (t).
(cid:88)
τ≤t
R+
j (t) =
r+
j (τ ).
tl
tl+1
j (t) = 0.
step τ∗ does not exist in interval I then R+
if for some l ≥ 1, we have W (j)
≥ d(j)(1 + c+γ), then W (j)
So for some l∗ ≤ log1/(1−γ/(2cd)) n we must have that W (j)
step τ ≥ tl∗ we have W (j)
steps τ, τ + 1 such that τ (mod 2) = 0 then in step τ + 1 ants only leave the task hence W (j)
and W (j)
τj ≥ d(j)(1 + c+γ). If such a
Let us define the smallest step τ∗ such that τ∗(mod 2) = 0 and W (j)
In the opposite case we can define a sequence of steps tl = τ∗ + 2 · l. Conditioned on Econcentration
τj (1 − γ/(2cd))l.
tl∗ ≤ d(j)(1 + c+γ). And by Claim 4.2 in any
τl∗ ≤ d(j)(1 + c+γ). Observe that for any phase consisting of two time
τ +1 ≤ W (j)
(cid:88)
(cid:88)
j (τ + 1) ≤ 2r+
l∗−1(cid:88)
j (τ ) ≤ (cid:88)
thus r+
j (τ ) + r+
r+
j (τ ) =
j (τ ). We have:
r+
j (τ ) +
τ +1 ≤ W (j)
j (t) =
(cid:88)
l∗−1(cid:88)
(1 − γ/(2cd)) ≤ W (j)
(cid:88)
∞(cid:88)
τ ≤ W (j)
≤ W (j)
tl
r+
j (τ ) +
τ∗≤τ <tl∗
τ∗≤τ <tl∗
τ
R+
n · (1 − γ/(2cd))l ≤ 2n
(1 − γ/(2cd))l ≤ 2ncd/γ
τ <τ∗
j (tl) ≤ 2
r+
r+
τ≥tl∗
τ≤t
≤ 2
r+
j (τ )
τ
l=0
l=0
l=0
19
Finally R+(t) =(cid:80)
j∈[k] R+
j (t) ≤ 2nkcd/γ. Observe that r+(τ ) > 0 only for at most
2kl∗ = 2k log1/(1−γ/(2cd)) n =
2k ln n
1−γ/(2cd)
1
ln
≤ 4kcd log n/γ
t
steps in the considered interval, where we in the last inequality we used that ln x ≥ (x − 1)/x.
Proof of Claim 4.4. To prove the first claim take any task j ∈ [k] and observe that if W (j)
t ≥ d(j)(1+γ)
then it remains at least d(j)(1 + γ) in all even time steps by Claim 4.2. On the other hand if
∈ [d(j)(1 − γ), d(j)(1 + γ)] then in the second sample in this phase we have by Econcentration
W (j)
at most d(j)(1 + γ)(1− 0.9csγ) ≤ d(j)(1− (0.9cs − 1)γ) and since 0.9cs ≥ 2, the last expression is upper
bounded by d(j)(1 − γ) hence by Ef eedback all the ants receive feedback lack in the second sample in
the considered phase. Hence no ant can leave the task in this phase. Thus we also have that in step
t + 2 all tasks are saturated.
t ≥ d(j)(1− γ), then conditioned on Econcentration, in the second
sample in this phase the number of workers in this task is at least d(j)(1−γ)(1−1.1csγ) > d(j)(1−c−γ)
combined with the fact that no ant can leave the task at the end of the phase this shows the second
claim.
Proof of Claim 4.5. First, we observe that Φ(·) and Ψ(·) are non-increasing, because due to Ef eedback
and Claim 4.2, no ant can join a task j that has load of at least d(j)(1 + γ) at the first step of a phase.
We observe that if for a task j, W (j)
We distinguish between two cases depending on ιt: the number of idle ants at time 2t.
1. Suppose ιt ≥ γn
64 . We know that not all tasks are saturated hence there exists in at the beginning
of step 2t + 1 at least one task j with at most d(j)(1 − γ) workers. By Ef eedback all idle ants
receive feedback lack from task j at the beginning of step 2(t + 1). Hence, by the definition of
the algorithm, all the idle ants join some tasks in step 2(t + 1). By Ef eedback all idle ants will join
tasks {j(cid:48) : W (j(cid:48))
2t < d(j(cid:48))(1 + γ)}. Again, there are two cases, either there is now at least one task j
for which load is now above d(j)(1 + γ) and we have that Ψ(t + 2)− Ψ(t) ≤ Ψ(t + 1)− Ψ(t) ≤ −1,
due to monotonicity of Ψ(·). Otherwise, Φ(t + 2) − Φ(t) ≤ Φ(t + 1) − Φ(t) ≤ − γn
64 .
assume γ ≤ 1/16 and cs = 2 1
we get that the sum of ants working on tasks j with W (j)
64 . Since we assume (Assumptions 2.1) that(cid:80)
j∈[k] d(j) ≤ n/2. Moreover, we
(cid:18)
8 )n and
(cid:88)
j∈[k](1 + (1 + 1.2cs)γ)d(j) ≤(cid:0)1 + 1
(cid:19)
3 and hence(cid:80)
t ≥(1+(1+1.2cs))γd(j) ≥ n − γn
t ≥ (1 + (1 + 1.2cs))γd(j) is at least
(cid:1) n
2 = (1− 3
2. Now suppose ιt < γn
W =
W (j)
t
W (j)
n ≥ n
4
.
1 − 3
8
−
64
1
4
j∈[k]
By definition of the algorithm, after one phase each of these W many ants leaves w.p. γ/cd and
by Econcentration, at least γW/(2cd) resulting in ιt+1 ≥ γn
idle ants at the beginning of the next
phase. From here on the same argument as above holds and we will have that either all tasks
are saturated in step 2(t + 1) or Ψ(t + 2) − Ψ(t) ≤ −1 or Φ(t + 2) − Φ(t) ≤ − γn
4cd
.
4cd
Proof of Lemma 4.6. The proof consists of three parts.
First observe that for any round τ, the regret in that round is trivially bounded by:
r(τ ) ≤ (cid:88)
τ − d(j) ≤ (cid:88)
X j
j∈[k]
X j
τ +
j∈[k]
(cid:88)
j∈[k]
d(j) ≤ 2n
We know by Claim 4.3 that:
R+(t) ≤ 2kncd
γ
.
20
(1)
(2)
Secondly, by the definition of R≈(t):
R≈(t) ≤ max{c+, c−}(cid:88)
j∈[k]
γtd(j).
(3)
Finally we need to show that, for some constant c2:
R−(t) ≤ c2kn
γ
.
(cid:16) 2n
(4)
To prove this we use the potential argument from Claim 4.5. By Claim 4.5 we have that for any step
τ such that τ mod 2 = 0 if Φ(τ ) = 0 and Ψ(τ ) = 0 then all the tasks are saturated which in turn
implies by Claim 4.4 that r−(τ(cid:48)) = 0 for all τ(cid:48) ≥ τ. Observe that Ψ(0) ≤ k and Φ(0) ≤ 2n and by
Claim 4.5 from the monotonicity of Ψ and Φ and the rate at which these potentials decrease there can
be at most 2
γ ) steps in which not all tasks are saturated. The regret per round
is, by (1), bounded by 2n, which yields (4).
γ steps r−(t) > 0. By Claim 4.3 in at most O(k log n/γ) steps
we have r+(t) > 0. Thus in all but O(k log n/γ) steps in all tasks j ∈ [k] the absolute value of the
deficit is bounded by 5γd(j).
We also showed that in at most k + 1
= O(k + 1
cγn + k
(cid:17)
Proof of Theorem 3.1. From Lemma 4.6 it follows that the total regret is w.h.p. during the interval I
is, by Union bound, w.p. p ≥ 1 − 1/n bounded by
+ c−1(cid:88)
j∈[k]
ckn
γ
γtd(j).
regret in any interval of length n4 is bounded by 2n5, by (1).
In the case that the high probability event fails, we simply use the following crude bound: The
Now consider an arbitrary time step t ∈ N (possibly super-exponential in n). Let (cid:96) =(cid:4)t/n4(cid:5) denote
B ∼ Binomial((cid:96), P [Econcentration ∩ Ef eedback ]) (cid:45) Binomial((cid:96), n−1),
the number of complete intervals that are contained in the first t time steps. Let B denote the number
of complete intervals among the first (cid:96) in which the regret is exceeds the upper bound in the statement
of Lemma 4.6. Observe that E [ B ] ≤ (1 − p)(cid:96) ≤ t
where we use the symbol ∼ to denote "distributed as" and the symbol (cid:45) to express stochastic domi-
(cid:3) ≤ P [ B ≥ nE [ B ] ] ≤ 1/n and therefore, by Union
nation. Hence, by Markov inequality, P(cid:2) B ≥ t
(cid:88)
bound, we have that w.h.p.
n6 . We have
(cid:88)
n5
+ c−1γ
(t − (cid:96)n4)d(j)
R(t) ≤ B2n5 + ((cid:96) − X)
ckn
γ
n4d(j)
j∈[k]
+
(cid:18)
ckn
(cid:19) ckn
γ
t
n4
γ
j∈[k]
+ c−1γ
(cid:88)
j∈[k]
td(j)
+ c−1γ
(cid:88)
j∈[k]
td(j) + 3t.
+ c−1γ
td(j) ≤ 2t +
1 +
≤ 2t + (1 + (cid:96))
≤ ckn
γ
+ c−1γ
This concludes the proof.
ckn
γ
(cid:88)
j∈[k]
B Sigmoid Noise - Missing Proofs of Section 3.3
Proof of Theorem 3.3. Consider the setting where k = O(1). We choose the constant c in the statement
of Theorem 3.3 such that each Ai uses at most
bits of memory.
log
√
(cid:106)
(cid:17)(cid:107)
(cid:16) 1
16
ε
21
√
= 1
16
∆(j)
This implies that the corresponding state-machine has at most s = 2
ε states. Fur-
thermore, we assume that all transition probabilities are either 0 or lower bounded by some p ∈ (0, 1),
which can be a function of ε, but not of n.
Consider an interval of length s and remark that s is a constant since ε is a constant. In the remain-
der we will show that there exists for each task j ∈ [k] w.p. at least 1 − e−Ω(n) a time step τ such that
(5)
j∈[k] d(j) for all tasks j ∈ [k] with
overwhelming probability, by taking Union bound. From this we easily get that for any t large enough:
Hence, the regret in the interval of length s is at least 2sεγ∗(cid:80)
t,
τ ≥ 2sεγ∗(cid:88)
εγ∗(cid:88)
of time intervals in which the regret is larger than 2sεγ∗(cid:80)
≤ P
Xt(cid:48) ≥ sεγ∗(cid:88)
Xt(cid:48) ≥ E [ Xt(cid:48) ]
due to the following reasoning. Consider t(cid:48) many time intervals of length s. Let Xt(cid:48) be the number
j∈[k] d(j). We have that E [ Xt(cid:48) ] = t(cid:48)e−Ω(n).
Thus, by Chernoff bounds (the random choices made in each interval are independent),
≤ e−t(cid:48)Ω(n).
R(t) ≥
d(j)t(cid:48)
j∈[k]
j∈[k]
d(j).
(cid:21)
(cid:20)
d(j)
2
Thus, by Union bound,(cid:80)∞
P
√
j∈[k]
P [ Xt(cid:48) ] ≤ e−Ω(n)+1. It remains to show (5).
t(cid:48)=1
We set d(j) =
n, for j ∈ [k]. Let γ(cid:48) = 2sεγ∗. Recall that for any task j ∈ [k]:
(cid:106)
log
(cid:17)(cid:107)
(cid:16) 1
√
ε
16
1
1 + eλγ∗d(j) = 1 − s(γ∗d(j)) = s(−γ∗d(j)) =
1
n8 .
Assuming that the load of a task j is in [(1 − γ(cid:48))d(j), (1 + γ(cid:48))d(j)], we have that the probability of
receiving incorrect feedback for each ant is at least, using the symmetry of the sigmoid function,
q = min{s(x) : x ∈ [−γ(cid:48)d(j), γ(cid:48)d(j)]} = s(−γ(cid:48)d(j)) =
(cid:18) 1
(cid:19)2sε ≥ 1
(cid:18)
eλγ∗d(j)
e
1 + eλγ∗d(j)
1
(cid:19)2sε ≥ 1
en16sε .
=
1
e
1
1 + eλsεγ∗d(j) ≥ 1
e
1
eλ2sεγ∗d(j)
By the Pigeonhole principle and Assumptions 2.2, there must exist a task i such that there must
exist a sequence of feedback signals (possibly different for each ant) such that n/k ants joins task
i. W.l.o.g. we assume that the probability of receiving an incorrect feedback signal for task i in
every round throughout the interval is at least q;
if it is less than this value, then this implies
that the regret is already of order γ(cid:48)d(j). Thus, the probability for every ant to join is at least
p(cid:48) = (p · q)s ≥ (p ·
distributed with parameters n/k and p(cid:48). We have E [ Yi ] ≥ n
δ = 1/2 we get
Let Yi denote the number of ants joining task i in the interval of length s. Note that Yi is binomially
k p(cid:48) ≥ 2n2/3. By Chernoff inequality with
(cid:32)
en16sε )s = p16sεn−16/(16)2es.
(cid:33)
1
(cid:21)
P(cid:104)
Yi ≥ n2/3(cid:105) ≤ P
(cid:20)
Yi ≥ E [ Yi ]
2
≤ exp
− 2n2/3
12
= e−Ω(n).
(6)
Summing over all k tasks gives that
22
s(cid:88)
(cid:88)
τ =1
j∈[k]
d(j) − W (j)
τ
≥ (cid:88)
≥ (cid:88)
j∈[k]
j∈[k]
This yields (5) and completes the proof.
min{Yi − d(j), 0}e−Ω(n) ≥ n2/3e−Ω(n)
2εγ∗d(j),
Proof of Theorem 3.2. Recall that s(−γ∗d(j)) ≤ 1/n8. For any task j ∈ [k]
1
1 + eλγd(j) = s(−γd(j)) ≤ s(−γ∗d(j)) =
1
n8 .
Moreover,
and hence, by putting both together, we get
1/e
eλγd(j) ≤
1
1 + eλγd(j)
1
eλγd(j) ≤ e
n8 .
Thus, the event Y that the median computed for a single round for a single ant is incorrect is, by
From this we get that for a step size of γ(cid:48) = εγ/cχ the probability p that the sample is incorrect, is at
most,
Theorem E.3 with α = 1/2, bounded by
P(cid:104)
Y ≥ m
2
(cid:105) ≤(cid:16)
p = s(−γ(cid:48)d(j)) =
1
≤
eλεγd(j)/cχ
1
(cid:18)(cid:16) e
n8
=
n8
(cid:16) e
(cid:17)ε/cχ
(cid:17)ε/(2cχ)(cid:19)m ≤ 1
· t,
n8 .
1 + eλεγd(j)/cχ
(2p)1/2 (2)1/2(cid:17)m ≤ 2m
(cid:88)
εγ
nk
γ(cid:48) +
j∈[k]
R(t) ≤ c
d(j) + O(1)
We observe that these are exactly the same guarantees that Algorithm Ant requires and hence we
get from Theorem 3.1 using a step size of γ(cid:48) instead of γ a regret of
for some constant c. We note that it will take longer to converge to right value as the step size is
smaller and the rounds are longer.
C Adversarial Noise
We first give the intuition behind Algorithm Precise Adversarial. Again, the algorithm builds on Algo-
rithm Ant. This time, every phase consists of two sub-phases. In the first sub-phase, the algorithm takes
samples spaced roughly εγ/32 apart. Each ant then remembers the task j (including the idle state) it
executed when it received for the first time the feedback lack (during the first-sub phase the feedback).
Throughout the second sub-phase, the ant will execute task j.
The idea behind this is that when an ant received for the first time the feedback lack, the deficit
must have been close to 0. Hence, taking many samples in this region yields a small regret.
We now give the sketch of Theorem 3.6.
Proof of Theorem 3.6. Algorithm Precise Adversarial is variation of Algorithm Ant with the difference
that O(1/ε) samples are taken in each phase instead of just 2. The main ideas are the same; in partic-
ular, if all samples indicate an overload, then after the phase, the number of ants reduces by a factor
of roughly γ.
Each phase consists of two sub-phases, the first one covering O(1/ε) rounds that is used to find the
first round rmin in the sub-phase where the the feedback received by the ant switched from overload
to lack.
23
Algorithm Precise Adversarial
(see Theorem 3.6)
Input at round t: Noisy feedback Ft = (F j
precision parameter ε < 1
Output at round t: Assignment of the ant to a task at ∈ {idle, 1, 2, . . . , k}
r1 ← 32
for At every step t ≥ 1 do
ε , r2 ← 4r1
t )j∈[k] for each task, learning parameter γ ∈ [γ∗, 1/16],
r ← t mod(r1 + r2)
if r = 1 then
[receive feedback]
s1 ← Ft
currentT ask ← at−1
t(cid:48) ← t
sr ← Ft
if currentT ask (cid:54)= idle, then at ←
[receive feedback]
if r ∈ [2, r1) then
if r = r1 then
rmin ← min{r(cid:48) < r1 : s(j)
sr ← Ft
if currentT ask (cid:54)= idle, then at ←
[receive feedback]
(cid:40)
r(cid:48) = lack or r(cid:48) = r1}
if r ∈ [r1 + 1, r1 + r2 − 1] then
at ← armin
if r = 0 then
s0 ← Ft
[receive feedback]
if currentT ask = idle then
(cid:40)
idle
currentT ask
w.p. εγ
32
otherwise
idle
currentT ask
if at(cid:48)+rmin−1 = idle
otherwise
(cid:40)
(cid:40)
else
at ←
output at
underloadedT asks ← {j ∈ [k] : s(j)
at ←
Uniform (underloadT asks)
idle
0 = s(j)
r1+r2−1 = lack}
1 = ··· = s(j)
if underloadT asks (cid:54)= ∅
otherwise
idle
currentT ask
32 if scurrentT ask
0
w.p. εγ
otherwise
= scurrentT ask
1
= ··· = scurrentT ask
r1+r2−1
= overload
Then in the second sub-phase of length r2 = O(1/ε) samples are taken as follows: if an ant was
working in round rmin (which may differ from ant to ant) then it will work throughout r2; and otherwise
it will be idle throughout r2.
We now argue that w.h.p. the average regret per phase is small for the case that load is not in a zone
where it will either decrease or increase w.r.t. to the beginning of the following phase. Let c = 1/16. Fix
an arbitrary task j ∈ [k]. Assume that at the beginning of the phase we had W (j)
t ≤ (1 + 3γ)d(j) since
otherwise w.h.p. all ants working on j will see an overload. Moreover, we assume that at the beginning
of the phase it must hold that W (j)
t ≥ (1 − 2γ)d(j) since otherwise, the resource is underloaded.
Therefore, throughout one can show that w.h.p. the first sub-phase we have ∆(j)
Furthermore, we have w.h.p. that throughout the second sub-phase it holds that ∆(j)
≤ (1 + 4γ)d(j).
≤ (1 + cεγ)d(j).
t
t
By definition,
r2 = 4r1
24
This implies that w.h.p. the total regret during the phase is at most:
(4r1 + cεr2)
d(j)γ ≤ (1 + ε)(r1 + r2)γ
d(j),
(cid:88)
j∈[k]
(cid:88)
j∈[k]
where the last inequality holds since r2 = 4r1 ≥ 4−(1+ε)
1+ε−cε r1. Similar as in Theorem 3.1, we can
bounded the regret in case that the high probability events did not hold. Summing over all phases
yields the desired result.
We now give the proof of Theorem 3.5.
Proof of Theorem 3.5. We argue that an algorithm cannot distinguish between two different demands.
We define two demand vectors. Let d = (n/(2k), n/(2k), . . . , n/(2k)) and d(cid:48) = (n/(2k) − 2τ, n/(2k) −
2τ, . . . , n/(2k) − 2τ ), where τ = (1 − o(1))γadn/(2k) is the threshold parameter defined in Section 2.2.
We now define the responses of the adversary
For the case of d we define the responses
1 ∆(j)
0 ∆(j)
Similarly, for the case of d(cid:48) we define the responses
i (j) =
F t
(cid:40)
(cid:40)
.
t−1 ≥ −γadd(j)
t−1 < −γadd(j)
t−1 ≥ γadd(j)
t−1 < γadd(j)
.
1 ∆(j)
0 ∆(j)
F t
i (j) =
It can easily be verified that the feedback in both cases is identical for any load distribution.
We apply Yao's principle Theorem E.1: With probability 1/2 we choose the demand vector d and
otherwise d(cid:48); we assume that the algorithm knows this input distribution. The cost of the algorithm
is simply the demand.
Consider the best deterministic algorithm A on this input distribution. We have
E [ R(t) ] ≥(cid:88)
(cid:88)
(cid:18) 1
2
(cid:12)(cid:12)(cid:12)W (j)
(cid:12)(cid:12)(cid:12) +
(cid:12)(cid:12)(cid:12)W (j)
By Theorem E.1 any randomized algorithm has at least an regret of tkτ = (1 − o(1))tγ∗(cid:80)
j∈[k]
+ 2τ
t(cid:48)≤t
τ − n
2k
τ − n
2k
≥ tkτ.
1
2
(7)
j∈[k] d(j)
over the course of t time steps. This concludes the proof.
(cid:12)(cid:12)(cid:12)(cid:19)
D Trivial Algorithm
In this section we consider for the sake of completeness the following trivial algorithm. Each ant that
is currently not working and receives the feedback that there is lack of ants a task, joins one of those
tasks (chosen u.a.r.); if the ant is already working at a task, then it continues working until it receives
feedback that there is an overload. We will analyze this algorithm in two models:
1. Sequential model: here exactly one ant is chosen u.a.r. and receives the feedback of the round
before.
2. Synchronous model: all ants receive feedback simultaneously. See Section 2 for details.
25
D.1 Sequential Model
In this section we show that the trivial algorithm performs reasonably well in this model. The reason
is that once some fixed task j is slightly overloaded (by a constant factor), then all other ants will see
this and refrain from joining. To make this more concrete, we can again use the critical value γ∗. Once
the overload (or underload) is at least γ∗d(j), each following ant-for the next n4 time steps, by the
same argument as in Claim 4.1-will w.h.p. receive the correct feedback and will refrain from joining
(or leaving).
It is also not too hard to see that the obtained regret is up to a constant factor intrinsic to this
algorithm (and any other constant memory sequential algorithm): The proof follows from a very similar
algorithm as Theorem 3.3. Consider the setting with one task with demand d(1) =
n. Suppose the
overload is less than γ∗d(1)/20, then the probability for an idle ant to join the task is s(−γ∗d(1)/20)
which is roughly 1/(en8/20). Hence, the expected increase in the number of ants working (assuming
ants are chosen u.a.r.) is at least:
√
t − W (j)
W (j)
t−1 Ft−1
1
− W (j)
t−1
n
≥ 1
en8/20
− 2W (j)
t−1
n
≥
1
,
2en8/20
n
en8/20
where we used our assumption that γ∗ is a constant and we generously assumed that every working
ant leaves the task. This positive drift, shows that as long the overload is less than γ∗d(1)/20, the
system will slowly converge to an overload of γ∗d(1)/20. From here one can use standard techniques
to conclude and we note that the proof generalizes to any constant number of jobs k. Therefore, the
regret is of size limt→∞ R(t)
j∈[k] d(j)) , which asymptotically matches the optimal regret in
the synchronous setting (Theorem 3.1 and Theorem 3.3).
t = Θ(γ∗(cid:80)
E(cid:104)
(cid:105) ≥ (n − W (j)
t−1)
D.2 Synchronous Model
In this section we argue that the trivial algorithm fails to achieve a good task allocation in the syn-
chronous model. Concretely, we show that the trivial algorithm does not converge in eΩ(n) steps.
Consider the example with one task having a demand d(1) = n/4 and assume γ∗ to be some constant
smaller than 1/2. Assume that the initial load is 0, then with overwhelming probability all ants will
receive lack and join. In the following step all ants will receive with overwhelming probability the
feedback overload and all leave. Union bound, these oscillations continue for at least eΩ(n) steps.
E Auxiliary Tools
Given a problem, let X be the set of inputs and A be the set of all deterministic algorithms solving it.
Then, for any x ∈ X and A ∈ A, the cost of executing A on x is denoted by cost(a, x).
Theorem E.1 (Yao's principle [34]). Let p be a probability distribution over A, and let A be an al-
gorithm chosen at random according to p. Let q be a probability distribution over X , and let X be an
input chosen at random according to q. Then
max
x∈X
Ep[cost(A, x)] ≥ min
a∈A
Theorem E.2 (Chernoff bound [26, Theorem 4.4 and 4.5]). Let X =(cid:80)
(cid:19)E[ X ]
pendent random variables. Then,
Eq[cost(a, X)].
1. for any δ > 0,
(cid:18)
P [ X ≥ (1 + δ)E [ X ] ] <
eδ
(1 + δ)1+δ)
.
i Xi be the sum of 0/1 inde-
26
2. for any 0 < δ ≤ 1,
3. for R ≥ 6E [ X ],
4. for 0 < δ < 1,
P [ X ≥ (1 + δ)E [ X ] ] ≤ e−E[ X ]δ2/3.
P [ X ≥ R ] ≤ 2−R.
P [ X ≤ (1 − δ)E [ X ] ] ≤
(cid:18)
e−δ
(1 − δ)1−δ)
(cid:19)E[ X ]
.
5. for 0 < δ < 1,
Theorem E.3 ([21, Equation 10]). Let Y = (cid:80)m
(cid:32)(cid:16) p
P [ Y ≥ α · m ] ≤
P [ Yi = 1 ] = p and P [ Yi = 0 ] = 1 − p. We have for any α ∈ (0, 1) that
P [ X ≤ (1 − δ)E [ X ] ] ≤ e−E[ X ]δ2/2.
(cid:19)1−α(cid:33)m
(cid:17)α(cid:18) 1 − p
1 − α
.
i=1 Yi be the sum of m i.i.d. random variables with
α
27
|
1704.02935 | 1 | 1704 | 2017-04-10T16:20:43 | A Cooperative Enterprise Agent Based Control Architecture | [
"cs.MA"
] | The paper proposes a hierarchical, agent-based, DES supported, distributed architecture for networked organization control. Taking into account enterprise integration engineering frameworks and business process management techniques, the paper intends to apply control engineering approaches for solving some problems of coordinating networked organizations, such as performance evaluation and optimization of workflows. | cs.MA | cs | A COOPERATIVE ENTERPRISE AGENT BASED CONTROL ARCHITECTURE
Simona Caramihai, Ioan Dumitrache, Aurelian Stanescu, Janetta Culita
Politehnica University of Bucharest, Dept. of Automatic Control and Computer Science
E-mail(s): [email protected] , [email protected], [email protected],
[email protected]
Abstract: The paper proposes a hierarchical, agent-based, DES supported, distributed
architecture for networked organization control. Taking into account enterprise
integration engineering frameworks and business process management techniques, the
paper intends to apply control engineering approaches for solving some problems of
coordinating networked organizations, such as performance evaluation and optimization
of workflows.
Keywords: Manufacturing Modelling for Management and Control, Discrete Event
Dynamic Systems.
1. INTRODUCTION
this
the
capabilities
dependent
on
increasingly
competition,
The globalization of business and commerce makes
enterprises
their
cooperation partners, clients and suppliers. At
present, competition occurs between networks of
enterprises instead of between individual enterprises.
In
for
interoperability of the enterprises become critical.
Interoperability has been used initially for the
analysis of software systems, from the process
control point of view. This means that a software
system is actually composed by a number of
subsystems able to communicate with each other and
with the environment through well defined interfaces.
All the communications are based on pre-defined
function implementation.
The most important challenge raised by the new
interpretation of the interoperability idea is how to
represent
semantics
according to the given purpose. The recent research
has highlighted the following conclusions:
Interoperability covers technical, semantic and
specification
the
system
pragmatic interoperability (Berre, 2005).
Technical interoperability means that messages
can be transmitted from one application to
another.
Semantic
the
message content becomes understood in the
same way by the senders and the receivers. This
may require
information
representation or messaging sequences.
interoperability means
transformations of
that
The pragmatic
interoperability captures
the
willingness of partners for the actions necessary
for
the collaboration. The willingness of
participating presumes both capability of
performing a requested action and policies
dictating whether
the potential action
is
preferable for the enterprise to be involved in.
Based on these premises, recent publications have
revealed that the interoperability only at data and
information level is not longer sufficient. The
collaboration must be done at knowledge level. At
this level all the organizational
interoperability
senses can be applied. For an efficient knowledge
sharing process, a new research field appeared:
knowledge management (KM).
and
through
feedback control
experimenting with
Knowledge and particularly the capacity to manage,
create and share knowledge, are becoming now the
centre of the scope for a successful company.
Knowledge becomes another attribute of
the
changing economic reality. KM and the learning
culture of the company, as a "necessary" condition
for Knowledge Economy (KE) represent an attitude
and a way of working with management. This
process is one of redefining the target of the company
from a profit making or share value increasing entity
to a knowledge creating and sharing unit.
In order to build a learning organization or a
corporate learning culture, companies should be
skilled in systematic problem solving, learning from
their own experience and from the others, processing
the
knowledge quickly and efficiently
organization
new
approaches.
There is a certain necessity to create a new type of
company with high autonomy, by using a distributed
intelligent system with most important attributes:
perception, learning, thinking, communication and
planning. The large FMS (Flexible Manufacturing
Systems) could be viewed as an Adaptive Complex
Dynamic System with a high capacity
to be
reconfigurable and a large knowledge base. The feed
forward and
strategies are
integrated on the general philosophy of knowledge
management into a company.
The process of KM means the identification and
analysis of knowledge,
for developing new
knowledge that will be used to realize organizational
goals. Knowledge is gathered from a geographical
and informational distributed system.
The best choice to define and to model a distributed
system
is represented by multi-agent systems
(MAS). A MAS is defined by modularity, abstraction
and
intelligent agent must be
autonomous, but in the same time it has to develop
the capacity of interoperability with the other entities
of the system. The concept of autonomy can be
interpreted in different ways (Wooldrige, 2002) by
reactivity, pro-activity and socialization.
Usually, the agent strategy consists in an ordered set
of activities and constraints – and can be modeled by
workflows. Knowledge management implies, in this
case, the appropriate choice of knowledge that can be
used for partial or full completion of the workflow
tasks. Moreover, a new workflow can be designed
using pieces of other workflows.
Agent's interoperability must be translated into
identification of the knowledge that can be used for
the accomplishment of each other goals. This
knowledge, having a structured modularity and
representation, can be the basics for the collaboration
support between the agents that have different
intelligence. An
systems,
(manufacturing
for diagnostics. Few
modeling technologies and also that are physically
placed into another structure. After the classes of
knowledge have been identified, the simulation of
business objective workflows can be done in order to
determine the failure and success rate. Different
applications
business
processes and also information technology (Brooke,
et al., 2003), make obvious that there is necessary
this type of approach.
Another challenge is related to the existence of a
support
interoperability
architectures support simulation, verification and
validation of process designs. Few systems support
collection and interpretation of real-time data. (van
der Aalst, et al., 2003). Anytime an agent can decide
that it no longer desires to share its knowledge,
therefore it is necessary to define protocols and
standards for the recovering – simulation becomes
critical for finding the scenarios when this approach
is useful.
The paper intends to present an original approach for
solving the above-mentioned problems in terms of
control engineering – architecture, modelling support
and evaluation. This approach is based on intelligent
agent-based
for Flexible
Manufacturing Systems (FMS), derived itself from
agent-based negotiation procedures proposed by Lin
and Solberg (1994) and from holonic architectures
(Wyns, et al., 1999). It was based on a Petri Nets
(PN) Discrete Event System modeling framework,
allowing control policies synthesis and evaluation.
The second part of the paper will shortly introduce
control
above mentioned
architecture. The third one will present its extension
for networked organisation as well as an illustrative
case study. Conclusions will include some comments
and further research directions.
architecture
concepts
control
and
the
2.THE SUPERVISED CONTROL
ARCHITECTURE
The control architecture (Caramihai, et al., 2001) was
designed to have two main levels: the agent level and
supervisory one.
The agent
level has extended autonomy and
communication capabilities. If the system is not
perturbed, the agents will act following given rules
and according to some pre-established patterns of
association, with a global behavior similar to a
hierarchical architecture. When a perturbation occurs,
agents have to take their own decisions, acting as
autonomous units. The agent-level itself acts as a
real-time control level for the manufacturing system.
On the other hand, the supervisory level has a more
complex behavior: on-line functioning for monitoring
the agents and taking some decisions on their
performances and an off-line supervisory1 policy
synthesis for ensuring robustness and efficiency in
the FMS global control.
The generic structure of an agent includes the
following basic blocks:
The World Model block - is a knowledge base
containing the internal model of the agent
(including relevant knowledge about the real
world, i.e. the state of temporary links with other
agents, the state of the controlled processes, the
long and short-term goals of the agent and the
value of the criteria of their achievement) and a
rule base on "how to take decisions". The world
model must have intrinsic updating mechanisms
for its two main parts, verifying and ensuring the
global consistency of data;
The decision making block – it decides either
how to process the received data (with direct and
explicit consequences on the immediate actions
of
to modify
information and knowledge contained in the
world model block (with implicit consequences
on the whole behavior of the agent);
the control agent) or how
The perception block - receives information
from the environment, via the Sensors block and
modifies, if necessary, the internal model of the
agent;
The actuators block - is basically an interface
with the environment – the real world: in this
situation it is rather a communication interface,
either with the control module of the controlled
process or with the other agents
The sensors block – is the interface between the
real world and the perception block, usually
consisting of sensors and transducers with the
purpose to record modifications induced by
actuators on the real world. In this case, it is also
a communication interface with the control
module of the controlled process (for updates in
the world model block) and with other agents.
The generic agent structure could be conceived as
being "split" in two components:
The computational agent part is working with
symbolic information and has as modules the
perception, world model and decision making
blocks;
the actuators blocks and works with
The real agent part is composed of the sensors
and
the
environment (the real world), modifying it and
recording the effect of both its own actions and of
those of other agents.
1 Supervisor – the term is used with its both connotations:
1. supervisory
if necessary,
correcting agent activities, and 2. supervisor in the sense
given by the classical DES theory; when using the
"synthesis of the supervisor" expression, the second
connotation is taken into account.
level, monitoring and,
the
the FMS
agents
are
considered
is achieved by
The two agent components act as a closed loop
structure, where the real agent is the controlled
process and the computational one is an adaptive
control module. It is important to note that the
actions of other
as
perturbations in the real world, for which, because of
agent-based philosophy, the world model is only a
partial model.
It was assumed that a FMS can be modeled
exclusively by two basic types of agents, each one
having a specific internal organization: the product
resource agent. The desired
agent and
functioning of
the
negotiations between the different types of agents.
A product agent is in charge with the completion of
a specific product or activity. In order to accomplish
its objective it negotiates with resource agents the
necessary operations and tasks, following a given
workflow model. A product agent owns a
"manufacturing account" consisting in virtual cost
units, which it can use for "paying" every operation it
requires from other agents. Its goal is to finalize the
products or activities it is responsible with following
the associated workflow, in a given time interval,
with minimum manufacturing cost and an adequate
quality.
A resource agent controls a resource in order to
ensure the required processing for product agents. It
is supposed that it acts as a high level control
structure and communicates with
the real-time
control module of
the resource for obtaining
information about its state and for starting specific
tasks. Its goal is to maximize the amount of cost units
received in a given time interval.
The internal model of a resource agent should
contain at least: the list of processing tasks that the
resource could perform - with their respective
duration and cost - and the current state of the
resource.
an
The
the main
interconnected
implementation
agent-based
architecture is derived from the large amount of
inter-agents communication, which could become
difficult to manage and overload the software system.
The architecture presented in this paper restricts this
communication flow by the fact that an agent has to
contact only agents of different type.
The PN part of the world model of a resource agent
takes
two aspects
mentioned above. The detailed internal model of the
agents is not needed because the control architecture
behavior is based on the global model. The PN model
of a product agent describes the workflow it has to
implement. The global model of a processing FMS is
obtained by the synchronous composition of the
respective resource and products PN models.
structure. One of
an
problems
into consideration only
the
communicate,
agents
can
creating
of
is
functioning
represented
the state-space analysis,
The supervisor presents two functioning regimes: one
for the off-line synthesis and one for the on-line
behavior.
When the system is working off-line, the supervisor
have
to be determined by firstly qualitatively
analyzing the PN model of the FMS for ensuring that
the system functioning will respect the desired
specifications, without deadlocks or forbidden states.
After performing
the
synthesis process determines a desired functioning
(based on optimality criteria, as the time spent in the
system by a product agent) and, based on this, it will
prescribe an "optimum functioning" for agents. This
prescribed
as
recommended partners for negotiations.
During its on-line regime, the supervisor monitors
the system. The monitoring is aimed to verify that the
agents respect their recommended evolutions. If a
perturbation occurs, then the global system will start
a degraded functioning regime with respect to the
pre-established optimal one.
In this paper the breakdown of a resource has been
considered as perturbation. Obviously, the respective
resource agent will no longer participate to the
negotiations. This will degrade the functioning of the
product agents whose recommended partner it was.
Furthermore, if the break-down has occurred during a
processing task, then a product agent could be more
specifically affected and the supervisor will have to
take some decisions concerning it.
Other kind of decisions the supervisory module
should take during the on-line functioning is linked
especially with agent-based architectures and
manufacturing processes specificities. Other kind of
decisions is concerned with the proper resource
sharing and with eventual bottlenecks and undesired
states that can appear in the degraded functioning
regime. The supervisor is still in possession of the
overall model of the system and can simulate on it
the results of achieved negotiations in order to
estimate in advance the problems that can occur.
A procedure for
the supervisor synthesis was
presented in detail in (Bratosin, et al., 2005) and
there was designed
application
implementing
it. The synthesis methodology is
highly general and it can be applied for any size of
the system.
a
software
3. ENTREPRISE CONTROL ARCHITECTURE
At enterprise level, the main problem is how to
implement a general control architecture allowing to
different and heterogeneous structures
to share
knowledge without having a predetermined idea on
the actual goal and without having to modify the
communication interfaces.
Fig. I Distributed control architecture
The main idea of the design is to realize a distributed
intelligent
architecture, where highly general
activities initiated at the strategic level should be
stepwise decomposed in more particular activities,
using partial workflow models. Figure I illustrates
the composition of partial workflow models into
complex models. Basically, every workflow unit
should encompass a "piece of knowledge" used for
solving a given type of problem. So, the global
knowledge of the system is distributed, and every
agent acting with respect to a given workflow will
include a part of the global intelligence of the system.
This characteristic will result only by the interaction
of agents, without requirements of real intelligence at
agent level. Moreover, the actual management and
evaluation of knowledge are performed at workflow
level, and they do not depend on
the actual
implementation. At the strategic level, there are
processes to be defined, each of them implying the
planning and execution of a sequence of activities.
Fig. II Knowledge bases - agents interfaces
types of
Agents could have
communications: with
base
(dynamical part) – via a dedicated interface, and with
some of the other agents, for eventual negotiation and
cooperative tasks. Static and dynamic knowledge-
two different
knowledge
the
Static knowledge base (workflows) Dynamic knowledge base (alternatives in-course) Strategic Operational Implementation – Real-time
interfaces with
bases are in a general format representation (as Petri
Nets, for instance).
The actual implementation of agents will necessitate
communication
the dynamical
knowledge base, but the knowledge transfer from one
agent to another could be made at the workflow
level, as well as the combination of different
workflows in order to design new activities (Figure
II). From the management level point of view, the
static part of the database can be the support for
strategic decisions and conception and the dynamic
part for the operational ones.
The structure provides great flexibility in the choice
of workflow alternatives and possibilities of dynamic
reconfiguration in failure situations, in the sense
described in Section 2. Moreover, it can be extended
from shop-floor level until to networks of enterprises
by the basic extension of the knowledge base, and
without regard to the actual implementation of
control agents; as well as the workflow definition is
performed and the interfaces between the workflow
representation and real-time control part exist.
4. CASE STUDY
result
three different activities performed by
The following case-study aims to illustrate how
workflows can be modeled by Petri Nets and,
consequently, different pieces of workflows can be
combined to obtain a new one. It will be underlined
how this composition can affect the execution of
initial workflows and how limited resources sharing
will
in several possibilities of activity
scheduling, therefore necessitating a decision support
system for choosing the most appropriate one.
Let be
different actors that have to share common resources.
At operational level they can be executed with the
following constraints:
- A1 consists of three jobs: J11, J12, J13, that are
executed strictly sequential;
- A2 consists of jobs J21, J22, J23, where J21 should
be performed first; the other two can be performed in
any order, but sequentially, as long as they are
starting after J21.
- A3 consists of jobs J31, J32, J33, where J33 is the
last to be performed, staring after the end of the other
two, which can be performed in any order, including
concurrency.
Each job needs some resources, as presented in Table
1. There are two resources R1 available and one of
each other type. The workflows representing the
activities are contained at the static knowledge base
level while the global model including them together
with their relationships at the dynamical knowledge
base level. The global model could be stored for an
eventually employment, after that being removed
from its location.
Table 1 Resource allocation and time duration for
each component job of workflow activities
Job
Res.
(time)
Job
Res.
(time)
J21
J22
J13
J12
R3(1) R4(2) R1(2) R2(1)
J11
R1(3)
R3(3)
J23
R3(2) R1(3) R4(2) R2(2)
J31
J32
J33
Figure III illustrates the global workflow model for
this case study.
Fig. III Petri net global model
From the modeling point of view, there are some
representation rules for workflows:
- workflows in static database are only represented at
a structural
job
sequencing with execution constraints and necessary
resources;
thus representing
level,
jobs,
P1P2P3P4P5P6P7P8P9R1J11J12J13P10P11J21P12P13P15P14P16P17P18P19R2J22J23J23J22P20P21P22P23P24P25P26J31J32J33 interoperability
Retrieved June 13, 2006, from http://interop-
noe.org/deliv/DAP1B.
for non-functional aspects"
of
the
15th
Bratosin, C, S. Caramihai and H. Alla (2005).
Supervisory Synthesis for Safe Petri Nets. In:
Proceedings
International
Conference on Control Sytems and Computer
Science (I.Dumitrache & C.Buiu (Eds)), Vol. 1,
pp. 149-154. Politehnica Press, Bucharest, RO.
Brooke, J., K. Garwood and C. Goble (2003).
Interoperability of Grid Resource Descriptions:
A Semantic Approach. Retrieved May 15, 2004,
from http://www.semanticgrid.org/GGF/ggf9/
Caramihai, S., C. Bratosin and C. Munteanu (2005).
A DES – supported agent based control
architecture for FMS. In: 17th IMACS World
Congress, CD-ROM, Paris, France.
Caramihai, S., C. Munteanu, J. Culita and A.
Stănescu (2001). Control Architecture for FMSs
based on Supervised Control of Autonomous
Agent. In: Proceedings of the 13th International
Conference on Control Systems and Computer
Science (I.Dumitrache & C.Buiu (Eds)), Vol. 1,
pp. 164-169. Politehnica Press, Bucharest, RO.
Lin, Y and Solberg, J (1994). Autonomous control
for open manufacturing systems. In: Computer
Control of Flexible Manufacturing Systems. (S.
Joshi and J. Smith (Eds)), pp. 169-206.
Chapman & Hall, London, UK.
Wyns, J., H. Van Brussel and P. Valckenaers (1999).
Design pattern for deadlock handling in holonic
manufacturing systems. In: Production Planning
and Control, Vol. 10, No. 7, pp. 616-626.
(2003). Business
Van der Aalst, W. M. P, A. H. M. ter Hofstede and
M. Weske
process
management: A survey. In: Proceedings of the
International Conference on Business Process
Management (W.M.P. van der Aalst, A.H.M.
ter Hofstede and M. Weske (Eds)), Volume
Lecture Notes in Computer Science, pp. 1-12.
Springer-Verlag, Berlin.
Wooldridge, M. (2002), An Introduction to Multi
Agent Systems, John Wiley & Sons Ltd, New
York.
- real-time process data as number of entities to be
processed (markings of P1, P10, P20) according to
workflow and actual resource capacities (marking of
P7, P8, P9, P19) – i.e. the complete initial marking of
the Petri Net are specified (by an initialization
procedure) only in the dynamic part of the data base;
- every job is represented by a block transition-place-
transition, as its execution is considered indivisible
(places meaning "job in execution" and transitions
"start" and respectively "end" job);
- between every two successive jobs, there is a place
with the meaning of "waiting", in order to avoid
deadlocks due to resources sharing.
It could be observed that even for a unary initial
marking of P1, P10, P20, there are several effective
conflicts that will affect the execution of activities,
thus delaying some of them.
Using the software application mentioned in Section
2 (Caramihai, et al., 2005), the above mentioned
initial marking will result in 8 execution possibilities
that
the application will underline. Therefore,
allocating an adequate cost function for every
activity, priorities can be established in order to
obtain the "better" activity scheduling for the context.
5. CONCLUSION
interoperability
The paper presents an architectural approach, based
on agent-based knowledge management, Petri Net
modeling and
for cooperative
enterprise networks.
The approach continues
in
supervisory synthesis for agent-based architectures
and it offers the foundation for a Decision Support
System in activity planning.
the previous work
REFERENCES
Berre, A-J
(2005).
"State-of-the
for
Interoperability architecture approaches" with a
focus on "Model driven and dynamic, federated
enterprise
interoperability architectures and
art
|
1908.09184 | 1 | 1908 | 2019-08-24T18:36:17 | Universal Policies to Learn Them All | [
"cs.MA",
"cs.LG"
] | We explore a collaborative and cooperative multi-agent reinforcement learning setting where a team of reinforcement learning agents attempt to solve a single cooperative task in a multi-scenario setting. We propose a novel multi-agent reinforcement learning algorithm inspired by universal value function approximators that not only generalizes over state space but also over a set of different scenarios. Additionally, to prove our claim, we are introducing a challenging 2D multi-agent urban security environment where the learning agents are trying to protect a person from nearby bystanders in a variety of scenarios. Our study shows that state-of-the-art multi-agent reinforcement learning algorithms fail to generalize a single task over multiple scenarios while our proposed solution works equally well as scenario-dependent policies. | cs.MA | cs | Universal Policies to Learn Them All
Hassam Ullah Sheikh
Department of Computer Science
University of Central Florida
Orlando, Florida, United States
[email protected]
9
1
0
2
g
u
A
4
2
]
A
M
.
s
c
[
1
v
4
8
1
9
0
.
8
0
9
1
:
v
i
X
r
a
Abstract
We explore a collaborative and cooperative multi-
agent reinforcement learning setting where a team
of reinforcement learning agents attempt to solve
a single cooperative task in a multi-scenario set-
ting. We propose a novel multi-agent reinforce-
ment learning algorithm inspired by universal value
function approximators that not only generalizes
over state space but also over a set of different sce-
narios. Additionally, to prove our claim, we are in-
troducing a challenging 2D multi-agent urban se-
curity environment where the learning agents are
trying to protect a person from nearby bystanders
in a variety of scenarios. Our study shows that
state-of-the-art multi-agent reinforcement learning
algorithms fail to generalize a single task over mul-
tiple scenarios while our proposed solution works
equally well as scenario-dependent policies.
1 Introduction
Recent research in deep reinforcement learning (RL) has led
to wide range of accomplishments in learning optimal poli-
cies for sequential decision making problems. These ac-
complishments include training agents in simulated environ-
ments such as playing Atari games [Mnih et al., 2015], beat-
ing the best players in board games like Go and Chess [Sil-
ver et al., 2016] as well as learning to solve real world
problems. Similar to single agent reinforcement learning,
multi-agent reinforcement (MARL) is also producing break
through results in challenging collaborative-competitive en-
vironments such as [OpenAI, 2018; Jaderberg et al., 2018;
Liu et al., 2019].
The success in reinforcement learning has prompted inter-
est in more complex challenges as well as a shift towards
cases in which an agent tries to learn multiple tasks in a single
environment. Formally this paradigm of learning is known
as multitask reinforcement learning [Teh et al., 2017]. The
essence of multitask reinforcement learning is to simultane-
ously learn multiple tasks jointly to speed up learning and
induce better generalization by exploiting the common struc-
tures among multiple tasks.
Despite having success in single agent multitask reinforce-
ment learning [Borsa et al., 2019; Teh et al., 2017], multitask
Ladislau Boloni
Department of Computer Science
University of Central Florida
Orlando, Florida, United States
[email protected]
multi-agent reinforcement learning has been explored in only
one recent study [Omidshafiei et al., 2017]. In this paper, we
are exploring an opposite problem where multiple reinforce-
ment learning agents are trying to master a single task across
multiple scenarios. Consider multiple RL agents trying to
master a single task in multiple scenarios. In order for the
agents to generalise, they need to able to identify and exploit
common structure of the single task under multiple scenarios.
One possible structure is the similarity between the solutions
of the single task over multiple scenarios either in the policy
space or associated value-function space. For this, we build
our solution upon two frameworks. The first framework is
universal value function approximators (UVFAs) by [Schaul
et al., 2015] and the second framework is multi-agent deep
deterministic policy gradient (MADDPG) by [Lowe et al.,
2017]. UVFAs are extension of value functions that also in-
clude the notion of a task or a scenario thus exploiting com-
mon structure in associated optimal value functions. MAD-
DPG is a multi-agent reinforcement learning algorithm that
uses the centralized training and distributed testing paradigm
to stabilize learning. The outcome of combining these two
frameworks is a solution for multi-agents that learn to gener-
alize over both state space and a set of multiple scenarios.
To investigate the emergence of collaboration in multi-
scenario learning for multi-agent learning agents, we de-
signed a challenging environment with simulated physics in
Multi-Agent Particle Environment [Mordatch and Abbeel,
2017]. We have developed 4 different simulated scenarios
representing a challenging urban security problem of pro-
viding physical protection to VIP from nearby bystanders of
more than one different class. The complexity in our environ-
ment arises primarily from the different moving patterns of
these bystanders that a standard state-of-the-art MARL algo-
rithm such as MADDPG [Lowe et al., 2017] fail to capture.
The goal here is to learn a stable and a consistent multi-agent
cooperative behavior across all the known scenarios. Here,
we are not dealing with unknown scenarios.
2 Background
Partially observable Markov Game [Littman, 1994] is a multi-
agent extension of MDP characterized by S, N agents with
partial observations O = {O1, . . . ,ON} of the environment
with a collective action space of A = {A1, . . . ,AN}, a re-
ward function R and a state transition function T . At every
time step, each agent chooses an action ai from it's policy
π parameterized by θi conditioned on its private observation
oi i.e ai = πθi (oi) and receives a reward ri = R (s, ai) The
goal of each agent is to maximise its own total expected return
i is the collected reward by
E [Gi] = E(cid:104)(cid:80)T
where rt
t=0 γtrt
i
(cid:105)
agent i at time t.
2.1 Policy Gradients
Policy gradient methods have been shown to learn the opti-
mal policy in a variety of reinforcement learning tasks. The
main idea behind policy gradient methods is to maximize the
objective function by parameterizing the policy π with θ and
updating the policy parameters in the direction of the gradient
∇J(θ) of the objective function J (θ) = E [Gt]. The gradient
is defined as
∇J(θ) = E [∇θ log πθ (as) Qπ (s, a)]
[Silver et al., 2014] has shown that it is possible to extend
the policy gradient framework to deterministic policies i.e.
πθ : S → A. In particular we can write ∇J (θ) as
∇J(θ) = E(cid:2)∇θπ (as)∇aQπ (s, a)a=π(s)
(cid:3)
(cid:3)
A variation of this model, Deep Deterministic Policy Gra-
dients (DDPG) [Lillicrap et al., 2015] is an off-policy algo-
rithm that approximates the policy π and the critic Qπ with
deep neural networks. DDPG also uses an experience replay
buffer alongside a target network to stabilize the training.
∇J(θi) = E(cid:2)∇θiπi (aioi)∇ai Qπ
Multi-agent deep deterministic policy gradients (MAD-
DPG) [Lowe et al., 2017] extends DDPG for the multi-agent
setting where each agent has it's own policy. The gradient of
each policy is written as
i (s, a1, . . . , aN )ai=πi(oi)
where s = (o1, . . . , oN ) and Qπ
i (s, a1, . . . , aN ) is a cen-
tralized action-value function that takes the actions of all the
agents in addition to the state of the environment to estimate
the Q-value for agent i. Since every agent has it's own Q-
function, the model allows the agents to have different ac-
tion spaces and reward functions. The primary motivation be-
hind MADDPG is that knowing all the actions of other agents
makes the environment stationary that helps in the stabiliza-
tion of the training, even though the policies of the agents
change.
The Universal Value Function Approximator [Schaul et al.,
2015] is an extension of DQN [Mnih et al., 2015] where it
generalizes not only over a set of states but also on a set
of goals. At the beginning of the episode, a state-goal pair
is sampled from a probability distribution, the goal remains
constant throughout the episode. At each timestep, the agent
receives a pair of current state and goal and gets the reward
rt = rg (st, at). As a result, the Q-function is not only depen-
dent on state-action pair but also on the goal. The extension
of this approach is straight forward for DDPG [Marcin et al.,
2017] and MADDPG.
3 Multi-Agent Universal Policy Gradient
We propose multi-agent universal policy gradient: a multi-
agent deep reinforcement learning algorithm that learns dis-
tributed policies not only over state space but also over a set
of scenarios.
ig
While generalization across multi-task in multi-agent rein-
forcement learning has been studied in [Omidshafiei et al.,
2017], to our best knowledge we are the first one to con-
sider the multi-scenario multi-agent deep RL system. Our ap-
proach uses Universal Value Function Approximators [Schaul
et al., 2015] to train policies and value functions that take a
state-scenario pair as an input. The outcome are universal
multi-agent policies that are able to perform on multiple sce-
narios as well as policies that are trained separately.
The main idea is to represent the different value function
approximators for each agent i by a single unified value func-
tion approximator that generalizes of over both state space
and a set of scenarios. For agent i we consider Vi (s, g; φ) ≈
V ∗
(s, a) that approximate the
ig
optimal unified value functions over multiple scenarios and a
large state space. These value functions can be used to extract
policies implicitly or as critics for policy gradient methods.
(s) or Qi (s, a, g; φ) ≈ Q∗
The learning paradigm we used is similar to the central-
ized training with decentralized execution during testing used
by [Lowe et al., 2017]. In this setting, additional information
is provided for the agents during training that is not available
during test time. Thus, extracting policies from value func-
tions is not feasible in this model. However, the value func-
tions can be used as critics in a multi-agent deep deterministic
policy gradient setting.
Concretely, consider an environment with N agents
with policies πππ = {π1π1π1, . . . , πNπNπN} parameterized by θθθ =
{θ1θ1θ1, . . . , θNθNθN} then the multi-agent deep deterministic policy
gradient for agent i can written as
∇J(θi) = E(cid:2)∇θiπi (aioi)∇aiQπ
i (s, a1, . . . , aN )ai=πi(oi)
where s = (o1, . . . , oN ) and Qπ
i (s, a1, . . . , aN ) is a central-
ized action-value function parameterized by φi that takes the
actions of all the agents in addition to the state of the environ-
ment to estimate the Q-value for agent i. We extend the idea
of MADDPG with universal functional approximator, specif-
ically we augment the centralized critic with an embedding
of the scenario. Now the modified policy gradient for each
agent i can be written as
(cid:3)
(cid:34)
∇J(θi) = Es,a,g∼D
∇θiπi (aioi, g)
(cid:35)
(1)
∇aiQπ
i (s, a1, . . . , aN , g)
where ai = πi (oi, g) is action from agent i following pol-
icy πi and D is the experience replay buffer. The centralized
critic is Qπ
i is updated as:
L (φi) = Es,a,r,s(cid:48),g [(Qπ
i (s, a1, . . . , aN , g)) − y]
y = rg
1, . . . , a(cid:48)
where y is defined as:
(s(cid:48), a(cid:48)
i + γQπ(cid:48)
φ(cid:48)
where π(cid:48)π(cid:48)π(cid:48) = {π(cid:48)
N} are target policies parameterized
1π(cid:48)
Nπ(cid:48)
1, . . . , π(cid:48)
1π(cid:48)
Nπ(cid:48)
by θ(cid:48)θ(cid:48)θ(cid:48) = {θ(cid:48)
N}.
1θ(cid:48)
Nθ(cid:48)
Nθ(cid:48)
1θ(cid:48)
1, . . . , θ(cid:48)
The overall algorithm to which we refer as multi-agent uni-
versal policy gradient (MAUPG) is described in algorithm 1.
N , g)a(cid:48)
i=π(cid:48)
i(o(cid:48)
i,g)
i
Additionally, we refer the learnt policies as universal policies.
The overview of the architecture can be seen in fig. 1.
i, g)
i
for t = 1 to episode -- length do
For each agent i, select action at
Execute actions at = [at
1, . . . , at
For each agent i, get next observation ot+1
i = πθi (ot
N ]
end for
Sample an additional scenario k
for t = 1 to episode -- length do
Algorithm 1 Multi-agent Universal Policy Gradient
1: Sample a random scenario g
2: for episode = 1 to M do
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
for agent i = 1 to N do
i := rg
i, ot+1
i, rt
/* Hindsight Replay */
Get reward rt
i (ot
i, at
end for
Set g = k
for agent i = 1 to N do
Store(cid:0)ot
Store(cid:0)ot
Get reward rt
i, at
i := rk
i, ot+1
i, rt
i
end for
i (ot
i
i, at
i)
i, at
i)
, g(cid:1) in replay buffer
, k(cid:1) in replay buffer
(cid:16)
sj, aj, rj, s(cid:48)j, gj(cid:17)
(cid:17)
N , gj(cid:17)
N , gj(cid:17)(cid:17)
1, . . . , a(cid:48)
(cid:48)j
N , gj)
1, . . . , aj
(o
(cid:16)
θi
Sample minibatch of size S
(cid:48)j
π(cid:48)
i , gj), . . . , π(cid:48)
w :=
(o
(a(cid:48)
1, . . . , a(cid:48)
N ) := w
s(cid:48)j, a(cid:48)
i + γQπ(cid:48)
Set yj = rj
φ(cid:48)
Update critic by minimizing
(cid:16)
θN
i
yj − Qπ
φi
sj, aj
(cid:16)
1
S
(cid:16)
(cid:88)
θi+ =(cid:80)
j
j
21:
22:
23:
24:
25:
26:
27:
∇θi πi(aioj
i ,gj)∇ai Qπ
φi(sj ,aj
1,..,aj
N ,gj)
end for
Update target network parameters for each agent i
S
i ← τ θi + (1 − τ ) θ(cid:48)
θ(cid:48)
i ← τ φi + (1 − τ ) φ(cid:48)
φ(cid:48)
i
i
(a) An overview of the mutli-agent de-
centralized actor with centralized critic
represented by a universal value function
approximator.
(b) Representation of
the single centralized
critic by a universal
function approximator.
Figure 1: An overview of the multi-agent universal policy gradient
architecture where both actors and critics are augmented with the
goal that the agents are trying to achieve.
a given moment from the nearby bystanders -- the aim of the
bodyguards is to reduce this value.
Two agents x and y have a line of sight LoS(x, y) ∈ [0, 1]
if x can directly observe y and with no obstacle between
them. A bystander b can only pose a threat to the VIP if it
is closer than the safe distance SafeDist. The threat level
TL (VIP, b) [Bhatia et al., 2016] is defined as the probability
that a bystander b can successfully assault the VIP, defined as
a value exponentially decaying with distance:
TL (VIP, b) = exp−A(Dist(VIP,b))/B
(2)
where the VIP should be in line of sight of b and
Dist(VIP, b) < SafeDist. A and B are positive constants
that control the decay rate of the threat level.
The residual threat RT (V IP,B, R) is defined as the threat
to the VIP at time t from bystanders B. Bodyguards can block
the line of sight from the bystanders, thus the residual threat
is always smaller than the threat level and depends on the po-
sition of the bodyguards with respect to the bystanders and
the VIP. The cumulative residual threat to the VIP from by-
standers B in the presence of bodyguards R over the time
period [0, T ] is defined as:
T(cid:90)
1 − k(cid:89)
28: end for
CRT =
(1 − RT (V IP, bi, R)) dt
(3)
0
i=1
4 Experimental Setup
4.1 The VIP Protection Problem
To demonstrate the effectiveness of our proposed algorithm,
we simulated an urban security problem of VIP protection
where a team of learning agents (bodyguards) are providing
physical protection to a VIP from bystanders in a crowded
space. This problem is briefly explored in [Sheikh and
Boloni, 2018a; Sheikh and Boloni, 2018b].
We are considering a VIP moving in a crowd of bystanders
B = {b1, b2, . . . , bM} protected from assault by a team of
bodyguards R = {r1, r2, . . . , rN}. To be able to reason about
this problem, we need to quantify the threat to the VIP at
Our end goal is to minimize CRT through multi-agent rein-
forcement learning.
4.2 Simulation and Scenarios
We designed four scenarios inspired from possible real world
situations of VIP protection and implemented them as behav-
iors in the Multi-Agent Particle Environment( fig. 2) [Mor-
datch and Abbeel, 2017].
In each scenario, the participants are the VIP, 4 bodyguards
and 10 bystanders of one or more classes. The scenario de-
scription contains a number of landmarks, points on a 2D
space that serve as a starting point and destinations for the
goal-directed movement by the agents. For each scenario,
𝑔𝜋1𝑎1𝑜1𝑄1𝑑𝜋𝑛𝑎𝑛𝑜𝑛𝑔𝑄𝑛𝑑Neural Neta𝑄sg (a) Random Landmarks
(b) Shopping Mall
(c) Street
(d) Pie-in-the-face
Figure 2: Visual representation of the four different scenarios. Emergence of complex behavior can be clearly seen where the bodyguards(in
blue) have positioned themselves between the VIP(in brown) and the bystanders(in red) shielding from potential threat.
(a) Random Landmarks
(b) Shopping Mall
(c) Street
(d) Red Carpet
Figure 3: Learning curve of the scenarios in terms of average cumulative reward for the bodyguards. Notice that start-of-the-art multi-agent
reinforcement learning algorithms such as COMA and Q-MIX fail to even take off in most of the challenging scenarios. MADDPG was the
only consistent algorithm that was able to learn in the environment.
the VIP (brown disk) starts from the starting point and moves
towards the destination landmark (green disk). The VIP ex-
hibits a simple path following behavior, augmented with a
simple social skill metric:
it is about to enter the personal
space of a bystander, it will slow down or come to a halt.
The scenarios differ in the arrangement of the landmarks
and the behavior of the different classes of bystanders.
A Random Landmark: In this scenario, 12 landmarks are
placed randomly in the area. The starting point and des-
tination for the VIP are randomly selected landmarks.
The bystanders are performing random waypoint nav-
igation:
they pick a random landmark, move towards
it, and when they reached it, they choose a new desti-
nation. A set of fixed seeds were used for placement
of landmarks and a different set of seeds were used for
spawning bystanders in the environment.
B Shopping Mall:
In this scenario, 12 landmarks are
placed in fixed position on the periphery of the area, rep-
resenting shops in a market. The bystanders visit ran-
domly selected shops and were spawned using a fixed
set of random seeds.
C Street: This scenario aims to model the movement on a
crowded sidewalk. The bystanders are moving towards
waypoints that are outside the current area. However,
due to their proximity to each other, the position of the
other bystanders influence their movement described by
laws of particles motion [Vicsek et al., 1995].
D Pie-in-the-Face: While the in other scenarios the by-
standers treat the VIP as just another person, in this "red
carpet" scenario the bystanders take an active interest in
the VIP. We consider two distinct classes of bystanders
with different behaviors. Rule-abiding bystanders stay
behind a designated line observing as the VIP passes in
front of them. Unruly bystanders break the limit im-
posed by the line and try to approach the VIP (presum-
ably, to throw a pie in his/her face).
Observation and Action Space
Following the model of Multi-Agent Particle Environ-
ment [Mordatch and Abbeel, 2017], the action space Ai of
each bodyguard i consists of 2D vector of forces applied
on the bodyguard and to promote collaboration and coopera-
tion [Mordatch and Abbeel, 2017], a c dimensional commu-
nication channel.
The observation of each bodyguard is the physical state
of the nearest m ⊂ M bystanders, all the N bodyguards
in the scenario and their verbal utterances such that oi =
[xj,...N +m, ck,...N ] ∈ Oi where xj is the observation of the
entity j from the perspective of agent i and ck is the verbal
utterance of the agent k.
In this problem, we are assuming that all bodyguards have
010002000300040005000600070008000Number of Episodes−30−25−20−15−10−5Average Cumu ative RewardMADDPGQ-MixCOMAIQLVDN010002000300040005000Number of Episodes−18−16−14−12−10−8−6−4−2Average Cu ulative RewardMADDPGQ-MixCOMAIQLVDN010002000300040005000Number of E isodes−35−30−25−20−15−10−5Average Cumulative RewardMADDPGQ-MixCOMAIQLVDN010002000300040005000Number of Episodes−14−12−10−8−6−4−2Average Cu ulative RewardMADDPGQ-MixCOMAIQLVDN identical observation space and action space. Moreover, each
scenario embedding g is represented as a one hot vector.
4.3 Reward Function
Using the definitions of threat level and cumulative residual
threat defined in eq. (2) and eq. (3) respectively, the reward
function rb for bodyguard i can be written as
rb = −1 +
(1 − RT (V IP, bi, R))
(4)
To encourage the bodyguards to stay at a limited distance
from the VIP and discourage them to attack the bystanders,
a distance regularizer D is added to eq. (4) to form the final
reward function.
D (VIP , xi) =
−1
m ≤ (cid:107)xi − VIP(cid:107)2 ≤ d
otherwise
(5)
k(cid:89)
i=1
0
where m is the minimum distance the bodyguard has to main-
tain from VIP and d is the SafeDist mentioned in section 4.1.
The final reward function is represented as
(cid:32)
k(cid:89)
i=1
+ β (D (VIP , xi))
rb =α
−1 +
(1 − RT (V IP, bi, R))
(cid:33)
(6)
Depending upon on the scenario g, different values of α,
and β were chosen for the optimal performance. A different
value of α and β also fulfills the requirement of a different
reward function to train a UVFA.
5 Experiments and Results
In this section, we first evaluate the usefulness of multi-agent
reinforcement learning on the given problem by comparing
it's results with an hand-engineered solution for the VIP prob-
lem. Then we demonstrate the inability of the state-of-the-
art MARL algorithms to generalize over different scenarios.
Finally we compare the results of scenario dependant poli-
cies with the results of universal policies. Our primary eval-
uation metric is Cumulative Residual Threat (CRT) defined
in eq. (3).
5.1 Multi-Agent Reinforcement Learning vs
Quadrant Load Balancing
In order to verify that multi-agent reinforcement learning so-
lutions are better than explicitly programmed behavior of the
bodyguards, we evaluate policies trained on individual sce-
narios with quadrant load balancing technique (QLB) intro-
duced in [Bhatia et al., 2016].
To identify the best MARL algorithm to compete with
the hand-engineered solution, we trained five state-of-the-
art MARL algorithms such as Q-Mix, VDN, IQL, COMA1
and MADDPG on our environment. It can be seen in fig. 3
1https://github.com/oxwhirl/pymarl/
Figure 4: Comparing the average cumulative residual threat values
of MADDPG and QLB on four different scenarios.
that MADDPG was the only algorithm that was successful in
learning in our environment while other algorithms fail even
to take off. Therefore, we dropped the CRT graphs.
We then compared the results of MADDPG with quadrant
load-balancing (QLB). From the results in fig. 4 we can see
that the outcome are different depending on the characteris-
tics of the scenario. For the Pie-in-the-face scenario, where
most of the bystanders stay away behind the lines, both the
RL-learned agent and the QLB model succeeded to essen-
tially eliminate the threat. This was feasible in this specific
setting, as there were four bodyguard agents for one "unruly"
bystander. For the other scenarios, the average cumulative
residual threat values are higher for both algorithms. How-
ever, for the Random Landmark and Shopping Mall scenarios
the RL algorithm is able to reduce the threat to less than half,
while in the case of the Street scenario, to less than one ninth
of the QLB value.
Overall,
these experiments demonstrate that the multi-
agent reinforcement learning can learn behaviors that im-
prove upon algorithms that were hand-crafted for this specific
task.
5.2 Universal Policies Vs Scenario-Dependant
Policies
In order to verify the claim that MARL algorithms trained on
specific scenario fail to generalize over different scenarios,
we evaluate policies trained via MADDPG on specific sce-
nario and test them on different scenarios. Policies on specific
scenarios were trained using the same settings and configura-
tions from experiments performed in section 5.1.
From the results shown in fig. 5 we can see that MADDPG
policies trained on specific scenarios performed poorly when
tested on different scenarios as compared to when tested on
same scenario with different seeds.
In order to tackle the
generalization problem, we train the agents using multi-agent
universal policy gradient and compare its results with the re-
sults of scenario-dependant MADDPG policies.
From the results in fig. 6 we can see that our proposed
method performs better than policies trained on specific sce-
narios as well as quadrant-load balancing. Overall, these ex-
periments demonstrate that start-of-the-art MARL algorithms
such as MADDPG fail to generalize a single task over mul-
tiple scenarios while our proposed solution MAUPG learn
3.862.942.750.0011.391.570.310.00100.511.522.533.544.5Random LandmarkShopping MallStreetPie-in-the-faceResidual ThreatQuadrant Load BalancingMADDPG Figure 5: A confusion matrix representing the average residual
threat values of MADDPG policies trained on specific scenario
when tested on different scenarios over 100 episodes.
Figure 7: Learning curves of the ablated version of MAUPG. Notice
the increase in performance in terms of average cumulative reward
with an addition of UVFAs and hindsight replay step.
Figure 6: Comparing the average cumulative residual threat values
for universal policy agents with MADDPG and QLB agents
policies that allows a single task to be learnt across multiple
scenarios and improve upon the start-of-the-art multi-agent
reinforcement learning algorithm.
6 Ablation Study
The first natural question that can be asked here is that why
can't we just sample scenarios during training of a standard
MADDPG? To answer this question and to see the effect of
UVFA and hindsight replay used in MAUPG, we perform an
ablation study in which we gradually add important building
blocks to MADDPG to transform the solution into MAUPG.
First to answer the question, we trained a MADDPG and
sampled different scenarios. Second we replaced the stan-
dard centralized critic with an UVFA and finally we added
the hindsight replay step. All the training settings and hyper-
parameters were kept same across all the experiments.
From figs. 7 and 8, we can see that MADDPG does not
learn half as good as MAUPG with or without the hind-
sight replay step. MAUPG learns better and faster than
MAUPG without hindsight replay. This happens because
MAUPG with hindsight replay benefits from replaying trajec-
tories from one scenario in other scenarios(see lines 14 and 15
of algorithm 1) thus providing more experience to learn effi-
ciently.
Figure 8: Learning curves of the ablated version of MAUPG. Notice
the decline in average cumulative residual threat with an addition of
UVFAs and hindsight replay step.
7 Conclusion
In this paper, we highlighted the issue with MARL algorithms
of failing to generalize a single task over multiple known sce-
narios. To solve the generalization problem, we proposed
multi-agent universal policy gradient, a universal value func-
tion approximator inspired policy gradient method that not
only generalizes over state space but also over set of differ-
ent scenarios. We also built a 2D challenging environment
simulating an urban security problem that can be used as a
benchmark for similar problems. Experimental studies have
demonstrated that our proposed method generalizes well on
different scenario and performs better than MADDPG when
trained on different scenarios individually.
References
[Bhatia et al., 2016] T.S. Bhatia, G. Solmaz, D. Turgut, and
L. Boloni. Controlling the movement of robotic body-
In Proc. of the
guards for maximal physical protection.
29th International FLAIRS Conference, pages 380 -- 385,
May 2016.
[Borsa et al., 2019] Diana Borsa, Andre Barreto, John Quan,
Daniel J. Mankowitz, Hado van Hasselt, Remi Munos,
3.862.942.750.0011.391.570.310.0011.011.250.210.00100.511.522.533.544.5Random LandmarksShopping MallStreetPie-in-the-faceResidual ThreatQuadrant Load BalancingMADDPGMAUPG02000400060008000Number f Epis des−22.5−20.0−17.5−15.0−12.5−10.0−7.5−5.0−2.5Average Cumulative RewardMADDPGMAUPG-w/o-HERMAUPG-w/-HER02000400060008000Number of Episodes1.01.52.02.53.0Average Cumulative Residual ThreatMADDPGMAUPG-w/o-HERMAUPG-w/-HER David Silver, and Tom Schaul. Universal successor fea-
In International Conference on
tures approximators.
Learning Representations, 2019.
[Jaderberg et al., 2018] Max Jaderberg, Wojciech M. Czar-
necki, Iain Dunning, Luke Marris, Guy Lever, Anto-
nio Garcia Castaneda, Charles Beattie, Neil C. Rabi-
nowitz, Ari S. Morcos, Avraham Ruderman, Nicolas Son-
nerat, Tim Green, Louise Deason, Joel Z. Leibo, David
Silver, Demis Hassabis, Koray Kavukcuoglu, and Thore
Graepel. Human-level performance in first-person mul-
tiplayer games with population-based deep reinforcement
learning. arXiv preprint arXiv: 1807.01281, 2018.
[Lillicrap et al., 2015] Timothy P. Lillicrap,
Jonathan J.
Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
Tassa, David Silver, and Daan Wierstra. Continuous con-
trol with deep reinforcement learning. In Proc. of the 3rd
Int'l Conf. on Learning Representations (ICLR), 2015.
[Littman, 1994] Michael L. Littman. Markov games as a
framework for multi-agent reinforcement learning.
In
Proc. of the 11th Int'l Conf. on Machine Learning(ICML),
pages 157 -- 163. Morgan Kaufmann, 1994.
[Liu et al., 2019] Siqi Liu, Guy Lever, Nicholas Heess, Josh
Merel, Saran Tunyasuvunakool, and Thore Graepel. Emer-
In International
gent coordination through competition.
Conference on Learning Representations, 2019.
[Lowe et al., 2017] Ryan Lowe, Yi Wu, Aviv Tamar, Jean
Harb, Pieter Abbeel, and Igor Mordatch. Multi-agent
actor-critic for mixed cooperative-competitive environ-
ments. In Advances in Neural Information Processing Sys-
tems 30, pages 6379 -- 6390, 2017.
[Marcin et al., 2017] Andrychowicz Marcin, Wolsk Filip,
Ray Alex, Schneider Jonas, Fong Rachel, Welinde Peter,
McGrew Bob, Tobin Josh, Pieter Abbeel, and Zaremba
In Advances in
Wojciech. Hindsight experience replay.
Neural Information Processing Systems 30, pages 5048 --
5058, 2017.
[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, 02 2015.
[Schaul et al., 2015] Tom Schaul, Daniel Horgan, Karol
Gregor, and David Silver. Universal value function ap-
proximators. In Proc. of the 32st Int'l Conf. on Machine
Learning(ICML), pages 1312 -- 1320, 2015.
[Sheikh and Boloni, 2018a] H. U. Sheikh and L. Boloni.
Designing a multi-objective reward function for creat-
ing teams of robotic bodyguards using deep reinforce-
In Prof. of 1st Workshop on Goal Spec-
ment learning.
ifications for Reinforcement Learning (GoalsRL-2018) at
ICML 2018, July 2018.
[Sheikh and Boloni, 2018b] H. U. Sheikh and L. Boloni. The
emergence of complex bodyguard behavior through multi-
In Proc. of Autonomy in
agent reinforcement learning.
Teams (AIT-2018) workshop at ICML-2018, July 2018.
[Silver et al., 2014] David Silver, Guy Lever, Nicolas Heess,
Thomas Degris, Daan Wierstra, and Martin Riedmiller.
Deterministic policy gradient algorithms. In Proc. of the
31st Int'l Conf. on Machine Learning(ICML), pages 387 --
395, 2014.
[Silver et al., 2016] David Silver, Aja Huang, Christopher J.
Maddison, Arthur Guez, Laurent Sifre, George van den
Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda
Panneershelvam, Marc Lanctot, Sander Dieleman, Do-
minik Grewe,
Ilya
Sutskever, Timothy Lillicrap, Madeleine Leach, Koray
Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mas-
tering the game of Go with deep neural networks and tree
search. Nature, 529:484 -- 503, 2016.
John Nham, Nal Kalchbrenner,
[Teh et al., 2017] Yee Teh, Victor Bapst, Wojciech M. Czar-
necki, John Quan, James Kirkpatrick, Raia Hadsell, Nico-
las Heess, and Razvan Pascanu. Distral: Robust multitask
reinforcement learning. In Advances in Neural Informa-
tion Processing Systems 30, pages 4496 -- 4506. 2017.
[Vicsek et al., 1995] Tam´as Vicsek, Andr´as Czir´ok, Eshel
Ben-Jacob, Inon Cohen, and Ofer Shochet. Novel type of
phase transition in a system of self-driven particles. Phys-
ical review letters, 75(6):1226, 1995.
[Mordatch and Abbeel, 2017] Igor Mordatch and Pieter
lan-
arXiv preprint
Abbeel.
guage in multi-agent populations.
arXiv:1703.04908, 2017.
Emergence of grounded compositional
[Omidshafiei et al., 2017] Shayegan Omidshafiei,
Jason
Pazis, Christopher Amato, Jonathan P. How, and John
Vian. Deep decentralized multi-task multi-agent rein-
forcement learning under partial observability. In Proc. of
the 34th Int'l Conf. on Machine Learning(ICML), pages
2681 -- 2690, 2017.
[OpenAI, 2018] OpenAI. OpenAI Five, 2018. https://blog.
openai.com/openai-five/.
|
1503.05317 | 1 | 1503 | 2015-03-18T09:35:16 | Model Checking AORTA: Verification of Organization-Aware Agents | [
"cs.MA"
] | As agent systems grow larger and more complex, there is an increasing need to formally verify them. Furthermore, it is often suggested that complex systems can be regulated using organizational models, imposing constraints on the agents in the systems. Agents that can understand the organizational model and constraints in a system is said to be organization-aware. This paper is concerned with verification of organization-aware agents. We show how agents using AORTA, a framework for making agents organization-aware, can be formally verified using an extended version of the Agent Java PathFinder (AJPF), a model checking system designed specifically for agent programming languages. We integrate AORTA with the Agent Infrastructure Layer (AIL), which is an intermediate layer on top of which APLs can be implemented, and use our extension of AJPF to verify a system of agents aiming to write a paper together by using an organization for coordination. | cs.MA | cs |
Model Checking AORTA:
Verification of Organization-Aware Agents
Andreas Schmidt Jensen
Technical University of Denmark, Kongens Lyngby, Denmark
[email protected]
Abstract. As agent systems grow larger and more complex, there is an
increasing need to formally verify them. Furthermore, it is often sug-
gested that complex systems can be regulated using organizational mod-
els, imposing constraints on the agents in the systems. Agents that can
understand the organizational model and constraints in a system is said
to be organization-aware. This paper is concerned with verification of
organization-aware agents. We show how agents using AORTA, a frame-
work for making agents organization-aware, can be formally verified us-
ing an extended version of the Agent Java PathFinder (AJPF), a model
checking system designed specifically for agent programming languages.
We integrate AORTA with the Agent Infrastructure Layer (AIL), which
is an intermediate layer on top of which APLs can be implemented, and
use our extension of AJPF to verify a system of agents aiming to write
a paper together by using an organization for coordination.
Keywords: Model Checking, Agent Programming Languages, Organi-
zational Reasoning, Organization-Aware Agents
1
Introduction
In many of the areas where multi-agent systems (MASs) are used, there is a need
for dependability and security. Therefore, it is increasingly necessary to consider
formal verification of such systems [4]. Furthermore, we have seen an increase of
interest in the area of organization-oriented MAS, i.e. systems in which agents
have to consider organizational constraints. The motivation for organizational
MASs is the increasing complexity of heterogeneous agents in open systems. The
owner of an open system cannot in general assume much about agents entering
the system, and it is therefore important to be able to regulate their behavior
to ensure that it is within the acceptable boundaries of the system.
Organizational models (e.g. Moise+ [16]) are designed to describe what is
expected of agents in the system without taking the individual agents and their
implementation into account. This is done using the notion of roles: an agent
can a enact roles, giving it certain responsibilities (the objectives of the role)
while providing certain capabilities (access to objects in the system, access to
groups of other agents, etc.). Furthermore, organizational models often have a
normative aspect: behavior, or states of affair, that is expected of the agents in
1
the system, but is not directly enforced. That is, the agents are free to violate the
norms of a system, but they should then expect to be punished. Organizational
models thus provide a way for the designer of a system to explain to the agents
entering the system, what is expected of them.
However, if agents are expected to fulfill the system's expectations of them,
they need a way to understand the organizational model of that system. Agents
that are able to do this are organization-aware [3]. Organization-aware agents
will naturally tend to be more complex than their "unaware" counterparts; even
though programming them may be easier, since certain aspects may be auto-
mated (task allocation, coordination, etc.), the reasoning cycle of the agents will
include more steps. This makes it even harder to convince ourselves that our
implementation is correct.
Since agent-oriented programming (AOP) differs from the well-known object-
oriented programming (OOP), the verification techniques from OOP must be
extended to capture the agent metaphor. That is, since agents are autonomous
and their behavior is based on beliefs and intentions, we need to be able to not
only check what the agent does (similar to verification in OOP), but also why it
did so.
The principles of model checking as defined in [2] "is an automated technique
that, given a finite-state model of a system and a formal property, systematically
checks whether this property holds for (a given state in) that model ". Model
checking agent programming languages (APLs) can thus be reduced to translat-
ing the system into a finite-state model in which we can prove certain properties.
Since agents are usually enriched with mental attitudes such as beliefs, goals and
intentions, model checking APLs is only interesting, if we can verify properties
about these mental attitudes. For example, it is possible to check whether agents
in a system only intend to achieve goal states by checking the temporal formula
✷(I(ag, φ) → G(ag, φ)). Here, ag refers to an agent, φ is a state and I and G
are modal operators referring to intentions and goals, respectively. Quite some
work has been done to make it possible to perform model checking on existing
APLs, and for example, the Agent Java PathFinder project [11] is an example
of a practical system in which model checking is feasible.
In this paper, we present an extension to AJPF, which makes it possible to
perform verification of organization-aware agents. We use AORTA [18] to make
agents organization-aware, and extend the specification language to incorporate
modalities about organizational information. Our contribution is two-fold: first,
we integrate AORTA with the Agent Infrastructure Layer (AIL), which is an
intermediate layer on top of which APLs can be implemented. By integrating
AORTA with AIL, we enable verification of organization-aware agents from po-
tentially any APL with a clear semantics. Second, we use the integration to verify
properties in a system of agents working together to write a scientific paper.
The rest of the paper is organized as follows. First, we provide a brief de-
scription of the AORTA framework in section 2. We then describe the AJPF
system, which is used to translate MASs into finite-state models on which model
checking can be done (section 3). In section 4 we present our extension to AJPF,
2
making it possible to perform model checking on organization-aware agents. We
evaluate the system in section 5 and we conclude the paper in section 6.
2 The AORTA framework
AORTA [18] is an organizational reasoning component that can be integrated into
an agent's reasoning mechanism, allowing it to reason about (and act upon) reg-
ulations specified by an organizational model using reasoning rules. That is, the
organization is preexisting and independent from the agent and the component
is agent-centered, focusing on letting the agent reason about the organization.
By separating the organization from the agent, the architecture of the agent is
independent from the organizational model, and the agent is free to decide on
how to use AORTA in its reasoning. The separation is achieved by basing the
component on reasoning rules using an organizational metamodel, designed to
support different organizational models. A prototype of AORTA has been imple-
mented in Java1 [19], designed such that it can provide organizational reasoning
capabilities to agents implemented in existing APLs.
In this paper, we use an extended version of the AORTA architecture de-
scribed in [19]: options are generated automatically, action deliberation and
coordination are merged into a single phase, and we incorporate obligations.
Organizational reasoning in AORTA is then divided into three phases: obligation
check (OC), option generation (OG) and action execution (AE). The OC-phase
uses the agent's mental state and organizational state to determine if obligations
are activated, satisfied or violated, and updates the organizational state accord-
ingly. The OG-phase uses the organizational specification to generate possible
organizational options. The agent considers these options in the AE-phase using
reasoning rules, which can alter the organizational state and the agent's inten-
tions, or send messages to other agents. The component is shown in figure 1.
We assume is connected to a cognitive agent, i.e., agents with mental attitudes
(such as beliefs and goals) and practical reasoning rules.
Checking obligations: In the OC-phase, AORTA uses the agent's state to deter-
mine for each obligation if it should change to a new state. We use conditional
obligations with a deadline, so an obligation can change state in several different
situations. If the condition for activating an obligation has happened, the com-
ponent activates the obligation by updating the organizational state. Similarly,
it checks whether the obligation has been satisfied (the objective is completed)
or violated (the deadline was reached before the objective was completed).
Option generation: In the OG-phase, AORTA uses the mental state of the agent
and the organizational state to consider what the agent can do regarding the
organization. The following organizational aspects are considered in the OG-
phase:
1 Java was chosen since many existing agent platforms are built in Java.
3
AORTA
Cognitive Agent
Organizational beliefs
Beliefs
Goals
OC
OG
AE
Reasoning rules
Mailbox
Fig. 1. The AORTA component. The arrows indicate flow of information. Obligations
and options are generated from the organizational beliefs, and actions are based on the
generated options.
Role enactment: Roles that are possible to enact given the agent's goals.
Role deactment: (Currently enacting) roles that have been fulfilled or are no
longer useful.
Obligations: States the agent is currently obliged to achieve.
Delegation: Objectives that can be delegated based on a dependency relation.
Information: Obtained information that other agents will benefit from know-
ing.
The options that are generated in this phase are then available to act upon in
the AE-phase.
Action execution: The AE-phase uses reasoning rules to decide how to react on
a given option in a given context. The AE-phase selects at most one option to
act upon. The reasoning is based on rules of the form2
option : context → action
where option is a previously generated option, context is a state description
that should hold for an action to be applicable, and action is the action to be
executed.
The agent has actions available to enact or deact a role, commit to complete
or drop an objective, and send messages. This corresponds to the options that
can be generated in the previous phase.
2.1 The AORTA organizational metamodel
Reasoning in AORTA uses an organizational metamodel, which is based on roles,
objectives and obligations, as these concepts are commonly used in existing
organizational models (e.g. Moise+).
2 Inspired by the plan syntax of AgentSpeak(L) [20].
4
Definition 1 (Organizational metamodel). The organizational metamodel
of AORTA is defined by the following predicates:
role(Role , Objs)
obj(Obj , SubObjs)
dep(Role 1 , Role 2 , Obj )
Role is the role name, and Objs is a set of
objectives.
Obj is the name of an objective, and SubObjs is a
set of sub-objectives.
Role Role 1 depends on role Role 2 for completion
of objective Obj .
rea(Ag , Role)
Agent Ag enacts role Role.
cond(Role, Obj, Deadline, Cond) A conditional obligation for role Role to complete
obl(Ag, Role, Obj, Deadline)
viol(Ag, Role, Obj)
Obj before Deadline when Cond holds.
An obligation for agent Ag playing role Role to
complete Obj before Deadline.
Agent Ag playing role Role has violated the obli-
gation to complete Obj.
A role is defined only by its name and its main objectives. Sub-objectives of
an objective are specified using obj-predicates. We distinguish between the differ-
ent states of obligations by using different predicates. For example, a conditional
obligation is represented by the predicate cond(borrower, return(Book ), Dead-
line, borrowed (Book )). If an agent Bob enacts the borrower role and borrows
the book "1984", the obligation is activated, which is represented by the predi-
cate obl(bob, borrower, return(1984 ), Deadline). A violation of the obligation is
represented by the predicate viol(bob, borrower, return(1984 )).
2.2 Operational semantics of AORTA
The AORTA framework has a well-defined operational semantics, which has been
implemented in Java and integrated with Jason. We will not go into details
with the semantics in this paper, but give the relevant definitions required to
understand our integration of AORTA with AIL.
One of the key ideas of AORTA is the notion of the organizational knowledge
base used by the component for reasoning about options and actions. Further-
more, the component contains an options base containing the options generated
in the OG-phase.
Definition 2 (Mental state). The AORTA mental state is based on knowledge
bases. Each knowledge base is based on a predicate language, L, with typical
formula φ. The agent's belief base and intention base are denoted Σa and Γa,
respectively. The language of the organization is denoted Lorg, and Lorg ⊆ L,
and the option language is denoted Lopt, and Lopt ⊆ L. The organizational
specification and options are denoted Σo and Γo, respectively. The mental state,
MS, is then a tuple of knowledge bases:
where Σa, Γa ⊆ L, Σo ⊆ Lorg and Γo ⊆ Lopt
MS = hΣa, Γa, Σo, Γoi,
5
Definition 3 (Options). The option language, Lopt with typical element γ is
defined as follows:
γ ::= role(R) obj(O) send(R, ilf, φ),
where R is a role identifer, O is an objective, ilf is tell or achieve, and φ ∈ L is
a message.
Each of the knowledge bases in the mental state can be queried using rea-
soning formulas.
Definition 4 (Formulas). AORTA uses reasoning formulas, LR, with typical
element ρ, which are based on organizational formulas, option formulas, belief
formulas and goal formulas:
ρ ::= ⊤ org(φ) opt(φ) bel(φ) goal(φ) ¬ρ ρ1 ∧ ρ2,
where φ ∈ L.
Organizational formulas, org(φ), queries the organizational beliefs, option
formulas, opt(φ), queries the options base, belief formulas, bel(φ), queries the
belief base and goal formulas, goal(φ), queries the goal base.
Definition 5 (Semantics of reasoning formulas). The semantics are based
on the agent's mental state, MS = hΣa, Γa, Σo, Γoi.
MS = ⊤
MS = bel(φ)
iff φ ∈ Σa
MS = goal(φ) iff φ ∈ Γa
MS = org(φ) iff φ ∈ Σo
MS = opt(φ) iff φ ∈ Γo
MS = ¬ρ
MS = ρ1 ∧ ρ2 iff MS = ρ1 and MS = ρ2
iff MS 6= ρ
We define the configuration of an agent with an AORTA component as follows:
Definition 6 (AORTA-agent). An AORTA-agent configuration is defined by
the following tuple:
A = hα, MS, AR, F, µi,
where α is the name of the agent, MS is the mental state, AR is the agents rea-
soning rules, AR ⊆ RA, F is the set of transition functions and µ = hµin, µouti
is the mailbox, contains incoming and outgoing messages.
The initial configuration consists of a set of initial beliefs and goals, and the
organizational specification. The agent has a number of state transition rules
available, which can be used to change its state. The execution of an entire orga-
nizational cycle will check for messages and external changes, apply obligation
rules, generate options and execute an action. Then, the reasoning cycle of the
agent enriched with the component is executed (e.g. the Jason reasoning cycle).
6
3 Model Checking Agent Programming Languages
Much of the work done in the area of model checking MASs and agent pro-
gramming languages has been in the setting of AgentSpeak(L) [5,6,7]. While
interesting, such approaches are generally hard to extend to other languages
without a lot of hard work. Furthermore, verification of heterogeneous MASs3
is not possible. Recently, others have proposed a way to verify heterogeneous
MASs by translating programs into a common metalanguage, meta-APL [13].
However, since this approach requires a translation of the program, we need to
convince ourselves that the translation is faithful to the original program.
In this paper, we are focusing on another approach for verifying agent sys-
tems, which is based on an extended version of Java PathFinder (JPF) [22]
called Agent JPF (AJPF), which takes advantage of the advanced model check-
ing features of JPF, while making it possible to verify properties relevant to
intelligent agents. AJPF can be used as-is for potentially any APL implemented
in Java, but its real power shows, when combined with the agent infrastructure
layer (AIL). AIL is designed so that interpreters of semantically well-defined
agent programming languages can be implemented using it [11], and has been
optimized for model checking in AJPF, by using techniques such as state-space
reduction. AIL comes with a simple APL, Gwendolen [10], which provides the
default semantics for AIL. An AIL agent has a belief base, possibly a rule base,
goals, plans and intentions. Furthermore, the agent has a reasoning cycle, which
executes the implementation of the operational semantics. We will not go into
details with all the different components of AIL, but refer to [11] for a detailed
description.
3.1 Specifying properties
Model checking agent systems is only useful, if we can specify desirable properties
in a language that can incorporate the mental attitudes of agents. In AJPF, these
properties are specified in the property specification language (PSL) [11]. PSL is
a linear-time temporal logic (LTL) with additional modal operators for beliefs,
goals, intentions, actions and percepts. We can thus specify formulas that should
hold in the system.
The full PSL syntax is given below. ag is the agent's name, f is a ground
first-order atomic formula.
φ ::= B(ag , f ) G(ag , f ) A(ag , f ) I(ag , f ) P(f ) φ ∨ φ ¬φ φUφ φRφ
B(ag , f ) is true if agent ag believes f to be true, G(ag , f ) is true if the
agent has f as a goal. A represents actions, I intentions and P properties of the
environment. The LTL formulas U and R represents "until" and "release", re-
spectively. The temporal operators ✸ (eventually) and ✷ (always) can be derived
from U and R.
3 That is, MASs comprised of agents implemented in different APLs.
7
The underlying semantics of the modal operators depends on the MAS being
verified. When using AORTA we can specify the semantics of beliefs as follows:
MAS = B(ag , f )
iff MS ag = bel(f ),
where MAS is the multi-agent system, and MS ag is agent ag's mental state as
defined in section 2. The semantics of the other modal operators can be given in
a similar way.
3.2 AJPF
Agent JPF is a module for the Java PathFinder [22]. JPF consists of an im-
plementation of the Java Virtual Machine (JVM), which can execute all paths
through a program in order to verify some predefined properties about the pro-
gram. Since the state space often explodes, JPF employs state matching in order
to reduce the number of states explored.
AJPF implements a controller, which takes care of execution of each of the
agents in the system. At each time step, it checks whether the system is in an end
state4 and should terminate, and otherwise it decides which agent to execute.
This decision is made by a scheduler, which keeps a list of active agents, i.e.
agents that do not want to sleep. The model checker can then branch out at
each of these states and execute each of the active agents (by choosing one path,
executing it until reaching an end state and then backtracking).
Each agent in an APL (either an AIL-enabled or an existing APL) needs
to implement the MCAPLLanguageAgent interface, which is used by AJPF to
perform all the steps necessary for the verification of a system:
-- Perform a reasoning step (MCAPLreason()).
-- Decide to put the agent to sleep (MCAPLwantstosleep()) or wake it up
(MCAPLwakeup()).
-- Check if a property holds (MCAPLbelieves(fml), MCAPLhasGoal(fml), etc.).
AJPF provides a listener, which is used to verify the properties specified in
PSL. This can be done by first building a Buchi automaton that represents the
property, and then compute the global behavior of the system by executing it.
The product of these, the product automaton, can then be used to check if the
property is violated. A property is violated if there exists a path to an accepting
state [8]. The implementation of the model checker in AJPF employs techniques
that allows progressively building the product automaton [14], making the veri-
fication process more efficient.
4 Verification of Organizations in Multi-Agent Systems
In the previous section, we described the AJPF framework, which can be used
for verification of MASs. We now turn to verification of organizational MASs.
4 An end state is defined as a state where every agent is sleeping and the environment
is not changing, thus nothing new can happen after such state has been reached.
8
In [12], several desirable properties of organizational MASs were define: what
makes an organization well-defined, good, effective, etc. These properties re-
quire not only a way to express the beliefs of the agents in the system, but also
the state of the organization. For example, a good organization "is an organi-
zation such that if the organization has the capability to achieve φ and there is
a group of roles in the organization responsible for realizing it, then the roles
being in charge have a chain of delegation to roles that are played by agents in
Ai that are actually capable of achieving it" [12]. Being able to verify that a sys-
tem satisfies these properties would be a large step towards convincing oneself
that the system actually works. Verification of organizational aspects has been
investigated before [1,9,17,21], but usually only by considering the internals of
each agent as a black box. Model checking of electronic institutions specified in
the ISLANDER framework was explored in [17]. By translating an ISLANDER
specification into MABLE, a language for automatic verification of MASs, the
system can be verified using the SPIN model checker.
The work most similar to ours is described in [9], where a programming
language for normative MASs is implemented in AIL. Agents in the system can
interact with an organization, and the system can then verify various properties
of both the agents and the organization. Our integration with AIL differs in
that we only verify properties of the agents, but these properties may include
organizational properties, as defined in the AORTA component. Since AORTA
is not tightly coupled to a specific APL, the integration with AIL allows us to
perform verification of existing agents with additional properties concerning an
organization.
In the remainder of this section, we show how AORTA can be integrated into
AIL, such that 1) interpreters implemented in AIL can make use of AORTA, and
2) verification of these systems is possible using AJPF5.
4.1 Specifying organizational properties
In order to verify properties about organizational beliefs and options, we need
to be able to express the properties in PSL. We therefore extend the PSL syntax
to incorporate such properties:
ψ ::= φ Org(ag , f ) Opt(ag , f )
The interpretation of Org(ag , f ) is given as:
MAS = Org(ag , f )
iff MS ag = org(f ),
where MAS is the multi-agent system (AIL+AORTA) and MSag is agent ag's
mental state as defined in section 2. Similarly, the interpretation of Opt(ag , f )
is:
MAS = Opt(ag , f )
iff MS ag = opt(f ).
5 AORTA and the integration with AIL is open source and is available at
http://www2.compute.dtu.dk/~ascje/AORTA/
9
In AJPF, we have implemented the extended PSL by adding functions check-
ing whether an agent has organizational beliefs or options to the MCAPLLanguage-
Agent interface, which defines the methods needed by AJPF to perform model
checking.
4.2 Verifying AORTA
Verification of the agents with an AORTA component requires (1) an integration
of the AORTA architecture within existing AIL agents, and (2) the ability to
verify properties about organizational beliefs and organizational options.
We have integrated the AORTA architecture in AIL, allowing existing in-
terpreters implemented in AIL to take advantage of the AORTA organizational
reasoning component. The AortaAILAgent extends the AILAgent class as fol-
lows:
MCAPLreason(int flag) Executes the AORTA reasoning cycle before calling
the AIL agent's own reasoning cycle.
MCAPLhasOrganizationalBelief(MCAPLFormula phi) Returns true if Σo con-
tains phi.
MCAPLhasOrganizationalOption(MCAPLFormula phi) Returns true if Γo con-
tains phi.
wantstosleep() Returns true if both AORTA and the AIL agent wants to sleep.
AORTA wants to sleep if the last execution did not change anything.
addBel(...)/addGoal(...)/delBel(...)/removeGoal(...) Responsible for
synchronization of knowledge bases.
newMessages(Set<Message> msgs) Checks if any of the incoming messages are
organizational messages and if so, lets AORTA handle them. Otherwise, they
are forwarded to the AIL agent.
We have furthermore implemented an AILBridge, which is responsible for
updating the AIL agent, when AORTA performs actions that change the belief
base or goal base.
5 Evaluation of AIL+AORTA
In this section, we evaluate our integration of AORTA in AIL. As we shall see,
the example is small enough to generate the entire state space within reasonble
time (takes approximately 10 minutes), so we evaluate the system in two ways:
1. We generate the product automata on the fly by executing the agent system,
while verifying each of the properties.
2. We first generate the entire state space for the system, and use it to verify
each of the properties.
The first method is practical for large (possible infinite) systems, or properties
that can be verified quickly (e.g. that the agents eventually enacts a role, since
this is the first thing happening in our system). The second method is practical
10
Listing 1. Gwendolen-program for writing a paper.
GWENDOLEN
:name: alice
: I n i t i a l B e l i e f s :
: B e l i e f R u l e s :
: I n i t i a l G o a l s :
editor [ achieve ]
: P l a n s :
+! editor [ achieve ] : { True } <- + editor ;
+! wtitle [ achieve ] : { True } <- + wtitle ;
+! wabs [ achieve ] : { True } <- + wabs ;
+! w s e c t i t l e [ achieve ] : { True } <- + w s e c t i t l e;
+! fdv [ achieve ] : { True } <- + fdv ;
+! wcon [ achieve ] : { True } <- + wcon ;
+! sv [ achieve ] : { True } <- + sv ;
:name: bob
: I n i t i a l B e l i e f s :
: B e l i e f R u l e s :
: I n i t i a l G o a l s :
writer [ achieve ]
: P l a n s :
+! writer [ achieve ] : { True } <- + writer ;
+! wsec [ achieve ] : { True } <- + wsec ;
+! wref [ achieve ] : { True } <- + wref ;
for verifying many properties in a single, finite system, since the time used for
verification of each property is significantly lower than the time spend generating
the state space.
5.1 Example: Writing a paper
To illustrate the capabilities of the model checker for AIL+AORTA, we use a
simple example of a group of agents aiming to write a scientific paper using
an organizational specification to help them collaborate (inspired by [15]). In
the example, an editor should create a first draft version (fdv ), consisting of
a title (wtitle) and an abstract (wabs) and the section titles (wsectitle). The
submission version (sv ) is then created by letting a number of writers write the
sections (wsec) and the references (wref ), while the editor writes the conclusion
(wconc). The writers depend on the editor for the completion of fdv, while the
editor depends on the writers for the completion of wsec and wref.
We consider two agents: Alice, capable of editing the paper and Bob, capable
of being a writer. We have implemented the agents in Gwendolen. The imple-
mentation of the agents is shown in listing 1. Since the focus is not on verification
of agents writing a paper, but rather on the organizational coordination mech-
anisms of AORTA, the implementation of each objective is very simple. Note
that the initial goals of each agent are artificial goals, which are used to generate
a role enactment option, since AORTA currently only supports generating role
enactment options based on the agent's goals.
The agents are enriched with an AORTA component that enables them to
enact roles, commit to objectives and coordinate. The AORTA-program for the
agents is shown in listing 2. The agents are able to delegate goals (the send(R,
11
Listing 2. AORTA-program for writing a paper.
role ( R ) : true = > enact ( R ).
obj ( bel ( O )) : bel ( me ( Me )) , org ( obl ( Me ,_ , bel ( O ) , _ )) = > commit ( O ).
send (_ , tell , org ( rea ( Me , R )))
bel ( me ( Me ) , agent ( Ag ) , Ag \= Me ) , ∼( bel ( sent ( Ag , org ( rea ( Me , R )))))
:
= > send ( Ag , org ( rea ( Me , R ))).
send (R , achieve , O )
org ( rea ( Ag , R )) , bel ( me ( Me ) , Ag \= Me ) , ∼( bel ( sent ( Ag , goal ( O ))))
:
= > send ( Ag , goal ( O )).
send (R , tell , O )
org ( rea ( Ag , R )) , bel ( me ( Me ) , Ag \= Me ) , ∼( bel ( sent ( Ag , bel ( O ))))
:
= > send ( Ag , bel ( O )).
achieve, O) option) and inform about completion of goals (the send(R, tell,
O) option). They commit to objectives that they are obliged to complete, and
for simplicity simply enact a role, if it is considered an option.
5.2 Main results
Our system was evaluated6 using the properties listed in table 1. Our results
are divided into two sets: 1) the on the fly verification, for which we specify the
number of states explored and the time used, and 2) the complete state space
verification, in which we only state the time used, since in that case, the number
of states is constant (the example contains 251 states).
Properties 1 and 2 check that the agents eventually enacts their roles. Prop-
erty 3 tries to verify that Alice eventually enacts the writer role, but fails, since
that is not the case. Properties 4 and 5 check that the agents furthermore eventu-
ally knows about the other agent's enactment. Property 7 verifies that whenever
Alice is obliged to achieve wabs, she will eventually believe that she has achieved
it. Property 9 verifies that when Alice believes she has completed fdv, she will
eventually inform Bob, because of the dependency relation between their roles.
Finally, property 11 verifies that the paper is eventually written. Note that the
time difference between properties 1 and 2 are due to the way the AJPF sched-
uler executes the agents: Bob is only executed once AJPF has detected that
Alice has nothing more to do.
Since the time complexity of LTL model checking is exponential in the length
of the formula [2], we furthermore verified larger formulas to get an assessment
of the model checker's capabilities. Property 6 verifies that the agents know
their own role and the role of the other agent. Property 8 verifies that for every
obligation it is the case that it is eventually satisfied. Property 10 verifies that
all dependency relations are used for generating options, and finally, 12 verifies
all of the properties above.
As expected, once the state space has been fully generated, verification of
each property is quite fast. We also see that verification of larger formulas takes
much more time (e.g. property 8 compared to property 7), which is expected,
6 We evaluated the system using Java 7 on a laptop with a dual core 2.80 GHz Intel
i7 CPU and 8 GB RAM running Windows 8.1.
12
Table 1. The properties that were verified by AJPF. The results are specified for 1)
the on the fly verification and 2) the complete state space verification.
Property
✸Org(alice, rea(alice, editor ))
✸Org(bob, rea(bob, writer ))
✸Org(alice, rea(alice, writer ))
✸Org(alice, rea(bob, writer ))
✸Org(bob, rea(alice, editor ))
(1) ∧ (2) ∧ (4) ∧ (5)
✷(Org(alice, obl(alice, editor , wabs, fdv ))
→ ✸B(alice, wabs))
✷(V Org(ag, obl(ag , role, obj , deadline))
→ ✸B(ag , obj ))
✷(Org(alice, dep(writer , editor , fdv )) ∧ B(alice, fdv )
→ ✸B(alice, sent (bob, bel(fdv ))))
✷ V Org(ag, dep(role1 , role2 , obj )) ∧ B(ag, obj )
→ ✸B(ag , sent (ag rea(role1 ), bel(obj )))
✸B(alice, sv )
(6) ∧ (8) ∧ (10) ∧ (11)
1
2
3
4
5
6
7
8
9
10
11
12
States1
6
6
11
21
27
31
Time1
0:14
0:26
0:40
1:38
1:48
2:24
Time2
<10ms
<10ms
<10ms
15ms
15ms
85ms
251
251
251
251
167
251
8:46
110ms
10:33
2740ms
8:51
9:15
8:34
13:10
25ms
95ms
<10ms
15812ms
since the formulas are LTL formulas. Obviously, generating the entire model
beforehand is only necessary, when we want to validate several properties, since
in several of the cases, the entire state space does not need to be explored.
Our experiments have further shown that one execution of an AORTA cycle
is on average 10 times slower when executed using JPF compared to executing
the system on the host JVM (i.e. without running the verification process).
Even though some decrease in performance is expected, it should be possible
to improve this by implementing the operational semantics of AORTA using
AIL. In that case, we have to convince ourselves that the semantics are correctly
implemented in AIL (i.e., the functionality should correspond to the functionality
of the existing implementation). Still, the advantage of verifying the existing
implementation is that the results are directly applicable (i.e., we know that the
agents will inform each other about their roles, that they will conform with their
obligations, etc.).
In section 4, we mentioned that an organization could be considered good if
the roles are related in such a way that the objectives of the organization will
be delegated to the agents that can actually achieve them. We cannot directly
specify general properties like this in PSL, so instead, they must be specified
using the specific properties relevant to the given system. In our case, we can
specify this using the dependency relations (property 10) and the fact that the
paper is eventually written (property 11). Using the organizational model and
the agent programs it should be possible to generate such specifications, but
that is out of scope for this paper.
6 Conclusion
As agent systems gain popularity and become increasingly complex, the possi-
bility to understand every part of a system becomes difficult. By model checking
13
agent systems, it will be possible to verify that the agents of the system be-
have as expected and that the outcome is satisfactory. We have discussed some
of the work done concerning model checking of agent programming languages,
focusing especially on the generic Agent Java PathFinder, which enables model
checking of potentially any kind of agent programming language. Furthermore,
by allowing the implementation of an APL interpreter within AJPF (using the
Agent Infrastructure Layer), it is possible to optimize the model checking pro-
cess, making it feasible for larger systems.
In increasingly complex systems, there is often a need for regulation, since
agents may come from different sources and cannot as such be forced to perform
actions required to achieve the system objectives. This further creates a need
for model checking, since the complexity again increases. In this paper, we have
shown that our framework for organizational reasoning, AORTA, can be model
checked using an extended version of AJPF. We have integrated AORTA in
AIL and have verified properties about a system implemented in an APL using
AORTA for organizational reasoning. We have verified that the agents of the
system enacts roles, coordinate role enactment and that they can successfully
delegate tasks using an organizational model describing roles and their relations.
The AORTA framework is not tightly integrated with AIL, but is rather used
as a library, which means that many of the optimization techniques of AJPF
cannot be used. Furthermore, a single step of the AORTA reasoning cycle is
considered atomic in the current implementation, making it impossible to ver-
ify properties about the internals of AORTA. Even though the system, in its
current state, makes it possible to verify interesting properties about a system,
it would be interesting to address these shortcomings in the future. However,
as we have mentioned, even though the advantages of implementing the opera-
tional semantics of AORTA in AIL rather than using it as a library are desirable,
they should be weighed against the drawback of not verifying the actual system,
but a (hopefully) equivalent one (in terms of operational semantics). We believe
that both approaches have their merits, and an implementation of the oper-
ational semantics in AIL will be also a useful contribution to model checking
of organization-aware agents. An obvious direction for future work is thus to
implement the operational semantics in AIL.
Finally, even though the example used in this paper shows that it is possible
to verify properties about organization-aware agents, it is rather small and it
would be interesting to verify larger, more complex systems containing more
than a handful of agents.
References
1. A¸stefanoaei, L., Dastani, M., Meyer, J.J., de Boer, F.S.: A verification framework
for normative multi-agent systems. In: Intelligent and Multi-Agent Systems. pp.
54 -- 65 (2008)
2. Baier, C., Katoen, J.P.: Principles of Model Checking (Representation and Mind
Series). The MIT Press (2008)
14
3. Boissier, O., van Riemsdijk, M.B.: Organisational Reasoning Agents. Agreement
Technologies pp. 309 -- 320 (2013)
4. Bordini, R.H., Dennis, L.A., Farwer, B., Fisher, M.: Automated Verification of
Multi-Agent Programs. In: 2008 23rd IEEE/ACM International Conference on
Automated Software Engineering. pp. 69 -- 78. IEEE (2008)
5. Bordini, R.H., Fisher, M., Visser, W., Wooldridge, M.: Verifiable multi-agent pro-
grams. Programming Multi-Agent Systems LNCS(LNAI 3067), 72 -- 89 (2004)
6. Bordini, R.H., Fisher, M., Visser, W., Wooldridge, M.: Verifying Multi-agent Pro-
grams by Model Checking. Autonomous Agents and Multi-Agent Systems 12(2),
239 -- 256 (2006)
7. Bordini, R.H., Fisher, M., Wooldridge, M., Visser, W.: Model checking rational
agents. IEEE Intelligent Systems (2004)
8. Courcoubetis, C., Vardi, M., Wolper, P., Yannakakis, M.: Memory-efficient algo-
rithms for the verification of temporal properties. In: Formal Methods in System
Design. pp. 275 -- 288 (1992)
9. Dennis, L., Tinnemeier, N., Meyer, J.J.: Model checking normative agent organi-
sations. In: Computational Logic in Multi-Agent Systems (2009)
10. Dennis, L.A., Farwer, B.: Gwendolen: A BDI Language for Verifiable Agents. In:
Lowe, B. (ed.) Logic and the Simulation of Interaction and Reasoning. AISB,
Aberdeen (2008), AISB'08 Workshop
11. Dennis, L.A., Fisher, M., Webster, M.P., Bordini, R.H.: Model Checking Agent
Programming Languages. Automated Software Engineering 19(1), 5 -- 63 (2011)
12. Dignum, V., Dignum, F.: A logic of agent organizations. Logic Journal of IGPL
pp. 283 -- 316 (2011)
13. Doan, T.T., Yao, Y., Alechina, N., Logan, B.: Verifying heterogeneous multi-agent
programs. In: Autonomous Agents and Multi-Agent Systems. pp. 149 -- 156 (2014)
14. Gerth, R., Peled, D., Vardi, M.Y., Wolper, P.: Simple on-the-fly automatic verifi-
cation of linear temporal logic. In: Protocol Specification, Testing and Verification.
pp. 3 -- 18. Chapman & Hall, Ltd. (1996)
15. Hubner, J.F., Boissier, O., Kitio, R., Ricci, A.: Instrumenting multi-agent organ-
isations with organisational artifacts and agents. Autonomous Agents and Multi-
Agent Systems 20(3), 369 -- 400 (2010)
16. Hubner, J.F., Sichman, J.S., Boissier, O.: A model for the structural, functional,
and deontic specification of organizations in multiagent systems. In: SBIA '02
Proceedings. pp. 118 -- 128 (2002)
17. Huget, M.P., Esteva, M., Phelps, S., Sierra, C., Wooldridge, M.: Model checking
electronic institutions. In: MoChArt 2002. pp. 51 -- 58 (2002)
18. Jensen, A.S., Dignum, V.: AORTA: Adding Organizational Reasoning to Agents.
In: Proc. AAMAS '14. pp. 1493 -- 1494 (2014)
19. Jensen, A.S., Dignum, V., Villadsen, J.: The AORTA architecture: Integrating
organizational reasoning in Jason. In: 2nd International Workshop on Engineering
Multi-Agent Systems (EMAS 2014). pp. 112 -- 128 (2014)
20. Rao, A.S.: AgentSpeak (L): BDI agents speak out in a logical computable language.
Agents Breaking Away (L) (1996)
21. Vigan`o, F.: A framework for model checking institutions. In: Model Checking and
Artificial Intelligence. pp. 129 -- 145 (2007)
22. Visser, W., Havelund, K., Brat, G., Park, S., Lerda, F.: Model checking programs.
Automated Software Engineering 10, 203 -- 232 (2003)
15
|
1712.08507 | 1 | 1712 | 2017-12-20T19:02:33 | Limits for Rumor Spreading in stochastic populations | [
"cs.MA",
"cs.OH"
] | Biological systems can share and collectively process information to yield emergent effects, despite inherent noise in communication. While man-made systems often employ intricate structural solutions to overcome noise, the structure of many biological systems is more amorphous. It is not well understood how communication noise may affect the computational repertoire of such groups. To approach this question we consider the basic collective task of rumor spreading, in which information from few knowledgeable sources must reliably flow into the rest of the population.
In order to study the effect of communication noise on the ability of groups that lack stable structures to efficiently solve this task, we consider a noisy version of the uniform PULL model. We prove a lower bound which implies that, in the presence of even moderate levels of noise that affect all facets of the communication, no scheme can significantly outperform the trivial one in which agents have to wait until directly interacting with the sources. Our results thus show an exponential separation between the uniform PUSH and PULL communication models in the presence of noise. Such separation may be interpreted as suggesting that, in order to achieve efficient rumor spreading, a system must exhibit either some degree of structural stability or, alternatively, some facet of the communication which is immune to noise.
We corroborate our theoretical findings with a new analysis of experimental data regarding recruitment in Cataglyphis niger desert ants. | cs.MA | cs | Limits for Rumor Spreading in stochastic populations∗
Lucas Boczkowskia, Ofer Feinermanc, Amos Kormana, and Emanuele Nataleb
aCNRS, IRIF, Universit Paris Diderot, 75013 Paris, France, [email protected]
bMax-Planck-Institut fr Informatik, 66123 Saarbrcken, Germany, [email protected]
cWeizmann Institute of Science, 76100 Rehovot, Israel, [email protected]
Abstract
Biological systems can share and collectively process information to yield emer-
gent effects, despite inherent noise in communication. While man-made systems often
employ intricate structural solutions to overcome noise, the structure of many biolog-
ical systems is more amorphous. It is not well understood how communication noise
may affect the computational repertoire of such groups. To approach this question we
consider the basic collective task of rumor spreading, in which information from few
knowledgeable sources must reliably flow into the rest of the population.
In order to study the effect of communication noise on the ability of groups that
lack stable structures to efficiently solve this task, we consider a noisy version of the
uniform PULL model. We prove a lower bound which implies that, in the presence of
even moderate levels of noise that affect all facets of the communication, no scheme
can significantly outperform the trivial one in which agents have to wait until directly
interacting with the sources. Our results thus show an exponential separation be-
tween the uniform PUSH and PULL communication models in the presence of noise.
Such separation may be interpreted as suggesting that, in order to achieve efficient
rumor spreading, a system must exhibit either some degree of structural stability or,
alternatively, some facet of the communication which is immune to noise.
We corroborate our theoretical findings with a new analysis of experimental data
regarding recruitment in Cataglyphis niger desert ants.
7
1
0
2
c
e
D
0
2
]
A
M
.
s
c
[
1
v
7
0
5
8
0
.
2
1
7
1
:
v
i
X
r
a
∗This work has received funding from the European Research Council (ERC) under the European Union's
Horizon 2020 research and innovation programme (grant agreement No 648032).
1
1
Introduction
1.1 Background and motivation
Systems composed of tiny mobile components must function under conditions of unrelia-
bility. In particular, any sharing of information is inevitability subject to communication
noise. The effects of communication noise in distributed living systems appears to be
highly variable. While some systems disseminate information efficiently and reliably
despite communication noise [2, 22, 11, 31, 37], others generally refrain from acquiring
social information, consequently losing all its potential benefits [25, 35, 38]. It is not
well understood which characteristics of a distributed system are crucial in facilitating
noise reduction strategies and, conversely, in which systems such strategies are bound
to fail. Progress in this direction may be valuable towards better understanding the
constraints that govern the evolution of cooperative biological systems.
Computation under noise has been extensively studied in the computer science com-
munity. These studies suggest that different forms of error correction (e.g., redundancy)
are highly useful in maintaining reliability despite noise [3, 1, 40, 39]. All these, however,
require the ability to transfer significant amount of information over stable communica-
tion channels. Similar redundancy methods may seem biologically plausible in systems
that enjoy stable structures, such as brain tissues.
The impact of noise in stochastic systems with ephemeral connectivity patterns is
far less understood. To study these, we focus on rumor spreading - a fundamental
information dissemination task that is a prerequisite to almost any distributed system
[10, 12, 16, 28]. A successful and efficient rumor spreading process is one in which a
large group manages to quickly learn information initially held by one or a few informed
individuals. Fast information flow to the whole group dictates that messages be relayed
between individuals. Similar to the game of Chinese Whispers, this may potentially
result in runaway buildup of noise and loss of any initial information [9]. It currently
remains unclear what are the precise conditions that enable fast rumor spreading. On
the one hand, recent works indicate that in some models of random noisy interactions,
a collective coordinated process can in fact achieve fast information spreading [20, 23].
These models, however, are based on push operations that inherently include a certain
reliable component (see more details in Section 1.3.2). On the other hand, other works
consider computation through noisy operations, and show that several distributed tasks
require significant running time [26]. The tasks considered in these works (including the
problem of learning the input bits of all processors, or computing the parity of all the
inputs) were motivated by computer applications, and may be less relevant for biological
contexts. Moreover, they appear to be more demanding than basic tasks, such as rumor
spreading, and hence it is unclear how to relate bounds on the former problems to the
latter ones.
In this paper we take a general stance to identify limitations under which reliable
and fast rumor spreading cannot be achieved. Modeling a well-mixed population, we
consider a passive communication scheme in which information flow occurs as one agent
1
observes the cues displayed by another. If these interactions are perfectly reliable, the
population could achieve extremely fast rumor spreading [28]. In contrast, here we focus
on the situation in which messages are noisy. Informally, our main theoretical result
states that when all components of communication are noisy then fast rumor spreading
through large populations is not feasible. In other words, our results imply that fast
rumor spreading can only be achieved if either 1) the system exhibits some degree of
structural stability or 2) some facet of the pairwise communication is immune to noise.
In fact, our lower bounds hold even when individuals are granted unlimited computa-
tional power and even when the system can take advantage of complete synchronization.
Finally, we corroborate our theoretical findings with new analyses regarding the
efficiency of information dissemination during recruitment by desert ants. More specif-
ically, we analyze data from an experiment conducted at the Weizmann Institute of
Science, concerning recruitment in Cataglyphis niger desert ants [34]. These analyses
suggest that this biological system lacks reliability in all its communication components,
and its deficient performances qualitatively validate our predictions. We stress that this
part of the paper is highly uncommon. Indeed, using empirical biological data to vali-
date predictions from theoretical distributed computing is extremely rare. We believe,
however, that this interdisciplinary methodology may carry significant potential, and
hope that this paper could be useful for future works that will follow this framework.
1.2 The problem
In this section, we give an intuitive description of the problem. Formal definitions are
provided in Section 2.
Consider a population of n agents. Thought of as computing entities, assume that
each agent has a discrete internal state, and can execute randomized algorithms - by
internally flipping coins. In addition, each agent has an opinion, which we assume for
simplicity to be binary, i.e., either 0 or 1. A small number, s, of agents play the role of
sources. Source agents are aware of their role and share the same opinion, referred to
as the correct opinion. The goal of all agents is to have their opinion coincide with the
correct opinion.
To achieve this goal, each agent continuously displays one of several messages taken
from some finite alphabet Σ. Agents interact according to a random pattern, termed
as the parallel-PULL model: In each round t ∈ N+, each agent u observes the message
currently displayed by another agent v, chosen uniformly at random (u.a.r) from all
agents.
Importantly, communication is noisy, hence the message observed by u may
differ from that displayed by v. The noise is characterized by a noise parameter δ > 0.
Our model encapsulates a large family of noise distributions, making our bounds highly
general. Specifically, the noise distribution can take any form, as long as it satisfies the
following criterion.
2
Definition 1.1 (The δ-uniform noise criterion). Any time some agent u observes an
agent v holding some message m ∈ Σ, the probability that u actually receives a message
m(cid:48) is at least δ, for any m(cid:48) ∈ Σ. All noisy samples are independent.
When messages are noiseless, it is easy to see that the number of rounds that are
required to guarantee that all agents hold the correct opinion with high probability is
O(log n) [28]. In what follows, we aim to show that when the δ-uniform noise criterion
is satisfied, the number of rounds required until even one non-source agent can be
moderately certain about the value of the correct opinion is very large. Specifically,
thinking of δ and s as constants independent of the population size n, this time is at
least Ω(n).
To prove the lower bound, we will bestow the agents with capabilities that far surpass
those that are reasonable for biological entities. These include:
• Unique identities: Agents have unique identities in the range {1, 2, . . . n}. When
observing agent v, its identity is received without noise.
• Complete knowledge of the system: Agents have access to all parameters of the
system (including n, s, and δ) as well as to the full knowledge of the initial
configuration except, of course, the correct opinion and the identity of the sources.
In addition, agents have access to the results of random coin flips used internally
by all other agents.
• Full synchronization: Agents know when the execution starts, and can count
rounds.
We show that even given this extra computational power, fast convergence cannot be
achieved.
1.3 Our contributions
1.3.1 Theoretical results
In all the statements that follow we consider the parallel-PULL model satisfying the
δ-uniform noise criterion, where cs/n < δ ≤ 1/2 for some sufficiently large constant
c. Note that our criterion given in Definition 1.1 implies that δ ≤ 1/Σ. Hence,
the previous lower bound on δ implies a restriction on the alphabet size, specifically,
Σ ≤ n/(cs).
Theorem 1.2. Any rumor spreading protocol cannot converge in less than Ω(
rounds.
nδ
s2(1−2δ)2 )
Recall that a source is aware that it is a source, but if it wishes to identify itself as
such to agents that observe it, it must encode this information in a message, which is, in
3
turn, subject to noise. We also consider the case in which an agent can reliably identify
a source when it observes one (i.e.,, this information is not noisy). For this case, the
following bound, which is weaker than the previous one but still polynomial, applies (a
formal proof appears in Section B):
Corollary 1.3. Assume that sources are reliably detectable. There is no rumor spread-
ing protocol that converges in less than Ω((
s2(1−2δ)2 )1/3) rounds.
nδ
Our results suggest that, in contrast to systems that enjoy stable connectivity, struc-
tureless systems are highly sensitive to communication noise. More concretely, the two
crucial assumptions that make our lower bounds work are: 1) stochastic interactions,
and 2) δ-uniform noise (see the right column of Figure 1). When agents can stabilize
their interactions the first assumption is violated. In such cases, agents can overcome
noise by employing simple error-correction techniques, e.g., using redundant messag-
ing or waiting for acknowledgment before proceeding. As demonstrated in Figure 1
(left column), when the noise is not uniform, it might be possible to overcome it with
simple techniques based on using default neutral messages, and employing exceptional
distinguishable signals only when necessary.
Figure 1: Non-uniform noise vs. uniform noise. On the left, we consider an example
with non-uniform noise. Assume that the message vocabulary consists of 5 symbols, that is,
Σ = {m1, m2, m3, m4, m5}, where m1 = 0 and m5 = 1, represent the opinions. Assume that
noise can occur only between consecutive messages. For example, m2 can be observed as either
m2, m3 or m1, all with positive constant probability, but can never be viewed as m4 or m5.
In this scenario, the population can quickly converge on the correct opinion by executing the
following. The sources always display the correct opinion, i.e., either m1 or m5, and each other
agent displays m3 unless it has seen either m1 or m5 in which case it adopts the opinion it saw
and displays it. In other words, m3 serves as a default message for non-source agents, and m1
and m5 serve as attracting sinks. It is easy to see that the correct opinion will propagate quickly
through the system without disturbance, and within O(log n) number of rounds, where n is the
size of the population, all agents will hold it with high probability. In contrast, as depicted on
the right picture, if every message can be observed as any other message with some constant
positive probability (for clarity, some of the arrows have been omitted from the sketch), then
convergence cannot be achieved in less than Ω(n) rounds, as Theorem 1.2 dictates.
4
Non-uniformnoiseδ-uniformnoisem1m2m3m4m5m1m2m3m4m5 1.3.2 Exponential separation between PUSH and PULL
Our lower bounds on the parallel-PULL model (where agents observe other agents)
should be contrasted with known results in the parallel-PUSH model, which is the push
equivalent to parallel-PULL model, where in each round each agent may or may not
actively push a message to another agent chosen u.a.r. (see also Section 2.3). Although
never proved, and although their combination is known to achieve more power than each
of them separately [28], researchers often view the parallel-PULL and parallel-PUSH
models as very similar on complete communication topologies. Our lower bound result,
however, undermines this belief, proving that in the context of noisy communication,
there is an exponential separation between the two models. Indeed, when the noise level
is constant for instance, convergence (and in fact, a much stronger convergence than we
consider here) can be achieved in the parallel-PUSH using only logarithmic number of
rounds [20, 23], by a simple strategy composed of two stages. The first stage consists of
providing all agents with a guess about the source's opinion, in such a way that ensures
a non-negligible bias toward the correct guess. The second stage then boosts this bias
by progressively amplifying it. A crucial aspect in the first stage is that agents remain
silent until a certain point in time that they start sending messages continuously, which
happens after being contacted for the first time. This prevents agents from starting to
spread information before they have sufficiently reliable knowledge. It further allows to
control the dynamics of the information spread in a balanced manner. More specifically,
marking an edge corresponding to a message received for the first time by a node, the
set of marked edges forms a spanning tree of low depth, rooted at the source. The depth
of such tree can be interpreted as the deterioration of the message's reliability.
On the other hand, as shown here, in the parallel-PULL model, even with the
synchonization assumption, rumor spreading cannot be achieved in less than a linear
number of rounds. Perhaps the main reason why these two models are often considered
similar is that with an extra bit in the message, a PUSH protocol can be approximated
in the PULL model, by letting this bit indicate whether the agent in the PUSH model
was aiming to push its message. However, for such a strategy to work, this extra bit
has to be reliable. Yet, in the noisy PULL model, no bit is safe from noise, and hence,
as we show, such an approximation cannot work. In this sense, the extra power that
the noisy PUSH model gains over the noisy PULL model, is that the very fact that
one node attempts to communicate with another is reliable. This, seemingly minor,
difference carries significant consequences.
1.3.3 Generalizations
Several of the assumptions discussed earlier for the parallel-PULL model were made for
the sake of simplicity of presentation. In fact, our results can be shown to hold under
more general conditions, that include: 1) different rate for sampling a source, and 2) a
more relaxed noise criterion. In addition, our theorems were stated with respect to the
5
parallel-PULL model. In this model, at every round, each agent samples a single agent
u.a.r. In fact, for any integer k, our analysis can be applied to the model in which, at
every round, each agent observes k agents chosen u.a.r. In this case, the lower bound
would simply reduce by a factor of k. Our analysis can also apply to a sequential variant,
in which in each time step, two agents u and v are chosen u.a.r from the population and
u observes v. In this case, our lower bounds would multiply by a factor of n, yielding,
for example, a lower bound of Ω(n2) in the case where δ and s are constants1.
1.3.4 Recruitment in desert ants
Our theoretical results assert that efficient rumor spreading in large groups could not be
achieved without some degree of communication reliability. An example of a biological
system whose communication reliability appears to be deficient in all of its components
is recruitment in Cataglyphis niger desert ants. In this species, when a forager locates an
oversized food item, she returns to the nest to recruit other ants to help in its retrieval
[4, 34].
We complement our theoretical findings by providing new analyses from an exper-
In such
iment on this system conducted at the Weizmann Institute of Science [34].
experimental setting, we interpret our theoretical findings as an abstraction of the in-
teraction modes between ants. While such high-level approximation may be considered
very crude, we retain that it constitutes a good trade-off between analytical tractability
and experimental data.
In our experimental setup recruitment happens in the small area of the nest's en-
trance chamber (Figure 2a). We find that within this confined area, the interactions
between ants are nearly uniform [32], such that an ant cannot control which of her nest
mates she meets next (see Figure 2b). This random meeting pattern coincides with the
first main assumption of our model. Additionally, it has been shown that recruitment
in Cataglyphis niger ants relies on rudimentary alerting interactions [18, 27] which are
subject to high levels of noise [34]. Furthermore, the responses to a recruiting ant and
to an ant that is randomly moving in the nest are extremely similar [34]. Although this
may resemble a noisy push interaction scheme, ants cannot reliably distinguish an ant
that attempts to transmit information from any other non-communicating individual.
In our theoretical framework, the latter fact means that the structure of communication
is captured by a noisy-pull scheme (see more details about PUSH vs. PULL in Section
1.3.2).
It has previously been shown that the information an ant passes in an interaction
can be attributed solely to her speed before the interaction [34]. Binning ant speeds
into four arbitrary discrete messages and measuring the responses of stationary ants
to these messages, we can estimate the probabilities of one message to be mistakenly
1This increase in not surprising as each round in the parallel-PULL model consists of n observations,
while the sequential model consists of only one observation in each time step.
6
Figure 2: Unreliable communication and slow recruitment by desert ant
(Cataglyphis niger ). a. The experimental setup. The recruiter ant (circled) returns to
the nest's entrance chamber (dark, 9cm diameter, disc) after finding the immobilized food item
(arrow). Group size is ten. b. A pdf of the number of interactions that an ant experiences
before meeting the same ant twice. The pdf
is compared to uniform randomized interaction
pattern. Data summarizes N = 671 interactions from seven experiments with a group size of
6 ants. c. Interactions with moving ants where classified into four different messages ('a' to
'd') depending on the ants' speed. The noise at which messages were confused with each other
was estimated according to the response recipient, initially stationary, ants (see Materials and
Methods). Gray scale indicates the estimated overlap between every two messages δ(i, j). Note
that δ = min(δ(i, j)) ≈ 0.2. Data collected over N = 64 interactions. d. The mean time it takes
an ant that is informed about the food to recruit two nest-mates to exit the nest is presented
for two group size ranges.
perceived as another one (see Materials and Methods). Indeed, we find that this com-
munication is extremely noisy and complies with the uniform-noise assumption with a
δ of approximately 0.2 (Figure 2c).
Given the coincidence between the communication patterns in this ant system and
the requirements of our lower bound we expect long delays before any uninformed
ant can be relatively certain that a recruitment process is occurring. We therefore
measured the time it takes an ant, that has been at the food source, to recruit the
help of two nest-mates. We find that this time increases with group size (p < 0.05
Kolmogorov-Smirnov test over N = 24 experiments, Figure 2d). Thus, in this system,
inherently noisy interactions on the microscopic level have direct implications on group
7
level performance. While group sizes in these experiments are small, we nevertheless
find these recruitment times in accordance with our asymptotic theoretical results. More
details on the experimental methodology can be found in Appendix A.
1.4 Related work
Lower bound approaches in biological contexts are still extremely rare [8, 21]. Our ap-
proach can be framed within the general endeavour of addressing problems in theoretical
biology through the algorithmic perspective of theoretical computer science [14, 13].
The computational study of abstract systems composed of simple individuals that
interact using highly restricted and stochastic interactions has recently been gaining con-
siderable attention in the community of theoretical computer science. Popular models
include population protocols [7], which typically consider constant size individuals that
interact in pairs (using constant size messages) in random communication patterns, and
the beeping model [41], which assumes a fixed network with extremely restricted com-
munication. Our model also falls in this framework as we consider the PULL model
[16, 28, 29] with constant size messages. So far, despite interesting works that consider
different fault-tolerant contexts [5, 6], most of the progress in this framework considered
noiseless scenarios.
In Rumor Spreading problems (also referred to as Broadcast) a piece of information
typically held by a single designated agent is to be disseminated to the rest of the
population. It is the subject of a vast literature in theoretical computer science, and
more specifically in the distributed computing community, see, e.g., [10, 12, 16, 17,
20, 26, 28, 33]. While some works assume a fixed topology, the canonical setting does
not assume a network. Instead agents communicate through uniform PUSH/PULL
based interactions (including the phone call model), in which agents interact in pairs
with other agents independently chosen at each time step uniformly at random from all
agents in the population. The success of such protocols is largely due to their inherent
simplicity and fault-tolerant resilience [19, 28]. In particular, it has been shown that
under the PUSH model, there exist efficient rumor spreading protocol that uses a single
bit per message and can overcome flips in messages (noise) [20].
The line of research initiated by El-Gamal [15], also studies a broadcast problem
with noisy interactions. The regime however is rather different from ours: all n agents
hold a bit they wish to transmit to a single receiver. This line of research culminated
in the Ω(n log log n) lower bound on the number of messages shown in [26], matching
the upper bound shown many years sooner in [24].
2 Formal description of the models
We consider a population of n agents that interact stochastically and aim to converge
on a particular opinion held by few knowledgable individuals. For simplicity, we assume
8
that the set of opinions contain two opinions only, namely, 0 and 1.
As detailed in this section, we shall assume that agents have access to significant
amount of resources, often exceeding reasonable more realistic assumptions. Since we
are concerned with lower bounds, we do not loose generality from such permissive
assumptions. These liberal assumptions will actually simplify our proofs. One of these
assumptions is the assumption that each agent is equipped with a unique identity id(v)
in the range {1, 2, . . . , n} (see more details in Section 2.4).
2.1 Initial configuration
The initial configuration is described in several layers. First, the neutral initial con-
figuration corresponds to the initial states of the agents, before the sources and the
desired opinion to converge to are set. Then, a random initialization is applied to the
given neutral initial configuration, which determines the set of sources and the opinion
that agents need to converge to. This will result in what we call the charged initial
configuration. It can represent, for example, an external event that was identified by
few agents which now need to deliver their knowledge to the rest of the population.
Neutral Initial Configuration x(0). Each agent v starts the execution with an
input that contains, in addition to its identity, an initial state taken from some discrete
set of states, and2 a binary opinion variable λv ∈ {0, 1}. The neutral initial configuration
for i ∈ [n], is the input of the agent with identity
x(0) is the vector whose i'th index, x(0)
i.
i
Charged Initial Configuration and Correct Opinion. The charged initial con-
figuration is determined in three stages. The first corresponds to the random selection
of sources, the second to the selection of the correct opinion, and the third to a possible
update of states of sources, as a result of being selected as sources with a particular
opinion.
• 1st stage - Random selection of sources. Given an integer s ≤ n, a set S of
size s is chosen uniformly at random (u.a.r) among the agents. The agents in S
are called sources. Note that any agent has equal probability of being a source.
We assume that each source knows it is a source, and conversely, each non-source
knows it is not a source.
• 2nd stage - Random selection of correct opinion. In the main model we
consider, after sources have been determined in the first stage, the sources are
randomly initialized with an opinion, called the correct opinion. That is, a fair
coin is flipped to determine an opinion in {0, 1} and all sources are assigned with
this opinion.
• 3rd stage - Update of initial states of sources. To capture a change in
behavior as a result of being selected as a source with a particular opinion, we
2The opinion of an agent could have been considered as part of the state of the agent. We separate
these two notions merely for the presentation purposes.
9
assume that once the opinion of a source u has been determined, the initial state
of u may change according to some distribution fsource−state that depends on (1)
its identity, (2) its opinion, and (3) the neutral configuration. Each source samples
its new state independently.
2.2 Alphabet and noisy messages
Agents communicate by observing each other according to some random pattern (for
details see Section 2.3). To improve communication agents may choose which content,
called message, they wish to reveal to other agents that observe them. Importantly,
however, such messages are subject to noise. More specifically, at any given time, each
agent v (including sources) displays a message m ∈ Σ, where Σ is some finite alphabet.
The alphabet Σ agents use to communicate may be richer than the actual information
content they seek to disseminate, namely, their opinions. This, for instance, gives them
the possibility to express several levels of certainty [30]. We can safely assume that the
size of Σ is at least two, and that Σ includes both symbols 0 and 1. We are mostly
concerned with the case where Σ is of constant size (i.e., independent of the number of
agents), but note that our results hold for any size of the alphabet Σ, as long as the
noise criterion is satisfied (see below).
δ-uniform noise. When an agent u observes some agent v, it receives a sample
of the message currently held by v. More precisely, for any m, m(cid:48) ∈ Σ, let Pm,m(cid:48) be
the probability that, any time some agent u observes an agent v holding some message
m ∈ Σ, u actually receives message m(cid:48). The probabilities Pm,m(cid:48) define the entries of the
noise-matrix P [23], which does not depend on time. We hereby also emphasize that
the agents' samples are independent.
The noise in the sample is characterized by a noise parameter 0 < δ ≤ 1/2. One of
the important aspects in our theorems is that they are general enough to hold assuming
any distribution governing the noise, as long as it satisfies the following noise criterion.
Definition 2.1 (The noise ellipticity parameter δ). We say that the noise has ellipticity
δ if Pm,m(cid:48) ≥ δ for any m, m(cid:48) ∈ Σ.
Observe that the aforementioned criterion implies that δ ≤ 1/Σ, and that the case
δ = 1/Σ corresponds to messages being completely random, and the rumor spreading
problem is thus unsolvable. We next define a weaker criterion, that is particularly
meaningful in cases in which sources are more restricted in their message repertoire than
general agents. This may be the case, for example, if sources always choose to display
their opinion as their message (possibly together with some extra symbol indicating
that they are sources). Formally, we define Σ(cid:48) ⊆ Σ as the set of possible messages that
a source can hold together with the set of messages that can be observed when viewing
a source (i.e., after noise is applied). Our theorems actually apply to the following
10
criterion, that requires that only messages in Σ(cid:48) are attained due to noise with some
sufficient probability.
Definition 2.2 (The relaxed noise ellipticity parameter δ). We say that the noise has
Σ(cid:48)-relaxed ellipticity δ if Pm,m(cid:48) ≥ δ for any m ∈ Σ and m(cid:48) ∈ Σ(cid:48).
2.3 Random interaction patterns
We consider several basic interaction patterns. Our main model is the parallel-PULL
model. In this model, time is divided into rounds, where at each round i ∈ N+, each
agent u independently selects an agent v (possibly u = v) u.a.r from the population and
then u observes the message held by v. The parallel-PULL model should be contrasted
with the parallel-PUSH model, in which u can choose between sending a message to
the selected node v or doing nothing. We shall also consider the following variants of
PULL model.
• parallel-PULL(k). Generalizing parallel-PULL for an integer 1 ≤ k ≤ n, the
parallel-PULL(k) model allows agents to observe k other agents in each round.
That is, at each round i ∈ N+, each agent independently selects a set of k agents
(possibly including itself) u.a.r from the population and observes each of them.
In each time step t ∈ N+, two agents u and v are selected
uniformly at random (u.a.r) among the population, and agent u observes v.
It each time step t ∈ N+ one agent is chosen u.a.r.
• sequential-PULL.
• broadcast-PULL.
from
the population and all agents observe it, receiving the same noisy sample of its
message3.
Regarding the difference in time units between the models, since interactions occur in
parallel in the parallel-PULL model, one round in that model should informally be
thought of as roughly n time steps in the sequential-PULL or broadcast-PULL model.
2.4 Liberal assumptions
As mentioned, we shall assume that agents have abilities that surpass their realistic
ones. These assumption not only increases the generality of our lower bounds, but also
simplifies their proofs. Specifically, the following liberal assumptions are considered.
• Unique identities. Each agent is equipped with a unique identity id(v) ∈
{1, 2, . . . , n}, that is, for every two agents u and v, we have id(u) (cid:54)= id(v). More-
over, whenever an agent u observes some agent v, we assume that u can infer
3The broadcast-PULL model is mainly used for technical considerations. We use it in our proofs as
it simplifies our arguments while not harming their generality. Nevertheless, this broadcast model can
also capture some situations in which agents can be seen simultaneously by many other agents, where
the fact that all agents observe the same sample can be viewed as noise being originated by the observed
agent.
11
the identity of v. In other words, we provide agents with the ability to reliably
distinguish between different agents at no cost.
• Unlimited internal computational power. We allow agents to have unlimited
computational abilities including infinite memory capacity. Therefore, agents can
potentially perform arbitrarily complex computations based on their knowledge
(and their id).
• Complete knowledge of the system. Informally, we assume that agents have
access to the complete description of the system except for who are the sources
and what is their opinion. More formally, we assume that each agent has access
to:
– the neutral initial configuration x(0),
– all the systems parameters, including the number of agents n, the noise pa-
rameter δ, the number of sources s, and the distribution fsource−state govern-
ing the update the states of sources in the third stage of the charged initial
configuration.
• Full synchronization. We assume that all agents are equipped with clocks
that can count time steps (in sequential-PULL or broadcast-PULL) or rounds (in
parallel-PULL(k)). The clocks are synchronized, ticking at the same pace, and
initialized to 0 at the beginning of the execution. This means, in particular, that
if they wish, the agents can actually share a notion of time that is incremented at
each time step.
• Shared randomness. We assume that algorithms can be randomized. That
is, to determine the next action, agents can internally toss coins and base their
decision on the outcome of these coin tosses. Being liberal, we shall assume that
randomness is shared in the following sense. At the outset, an arbitrarily long
sequence r of random bits is generated and the very same sequence r is written in
each agent's memory before the protocol execution starts. Each agent can then
deterministically choose (depending on its state) which random bits in r to use as
the outcome of its own random bits. This implies that, for example, two agents
can possibly make use of the very same random bits or merely observe the outcome
of the random bits used by the other agents. Note that the above implies that,
conditioning on an agent u being a non-source agent, all the random bits used by
u during the execution are accessible to all other agents.
• Coordinated sources. Even though non-source agents do not know who the
sources are, we assume that sources do know who are the other sources. This
means, in particular, that the sources can coordinate their actions.
2.5 Considered algorithms and solution concept
Upon observation, each agent can alter its internal state (and in particular, its message
to be seen by others) as well as its opinion. The strategy in which agents update these
12
variables is called "algorithm". As mentioned, algorithms can be randomized, that is,
to determine the next action, agents can use the outcome of coin tosses in the sequence
r (see Shared randomness in Section 2.4). Overall, the action of an agent u at time t
depends on:
1. the initial state of u in the charged initial configuration (including the identity of
u and whether or not it is a source),
2. the initial knowledge of u (including the system's parameters and neutral config-
3. the time step t, and the list of its observations (history) up to time t− 1, denoted
uration),
x(<t)
u
,
4. the sequence of random bits r.
2.6 Convergence and time complexity
At any time, the opinion of an agent can be viewed as a binary guess function that is
used to express its most knowledgeable guess of the correct opinion. The agents aim to
minimize the probability that they fail to guess this opinion. In this context, it can be
shown that the optimal guessing function is deterministic.
Definition 2.3. We say that convergence has been achieved if one can specify a par-
ticular non-source agent v, for which is it guaranteed that its opinion is the correct
opinion with probability at least 2/3. The time complexity is the number of time steps
(respectively, rounds) required to achieve convergence.
We remark that the latter definition encompasses all three models considered.
Remark 2.4 (Different sampling rates of sources). We consider sources as agents in the
population but remark that they can also be thought of as representing the environment.
In this case, one may consider a different rate for sampling a source (environment)
vs. sampling a typical agent. For example, the probability to observe any given source
(or environment) may be x times more than the probability to observe any given non-
source agent. This scenario can also be captured by a slight adaptation of our analysis.
When x is an integer, we can alternatively obtain such a generalization by considering
additional artificial sources in the system. Specifically, we replace each source ui with
a set of sources Ui consisting of x sources that coordinate their actions and behave
identically, simulating the original behavior of ui. (Recall that we assume that sources
know who are the other sources and can coordinate their actions.) Since the number
of sources increases by a multiplicative factor of x, our lower bounds (see Theorem 3.1
and Corollary 1.3) decrease by a multiplicative factor of x2.
13
3 The lower bounds
Throughout this section we consider δ < 1/2, such that (1−2δ)
section is to prove the following result.
δsn ≤ 1
10 . Our goal in this
Theorem 3.1. Assume that the relaxed δ-uniform noise criterion is satisfied.
• Let k be an integer. Any rumor spreading protocol on the parallel-PULL(k) model
(cid:16)
(cid:17)
cannot converge in fewer rounds than Ω
nδ
ks2(1−2δ)2
.
• Consider either the sequential-PULL or the broadcast-PULL model. Any rumor
spreading protocol cannot converges in fewer rounds than Ω
(cid:16)
(cid:17)
n2δ
s2(1−2δ)2
.
To prove the theorem, we first prove (in Section 3.1) that an efficient rumor spreading
algorithm in either the noisy sequential-PULL model or the parallel-PULL(k) model
can be used to construct an efficient algorithm in the broadcast-PULL model. The
resulted algorithm has the same time complexity as the original one in the context of
sequential-PULL and adds a multiplicative factor of kn in the context of parallel-PULL(k).
We then show how to relate the rumor spreading problem in broadcast-PULL to
a statistical inference test (Section 3.2). A lower bound on the latter setting is then
achieved by adapting techniques from mathematical statistics (Section 3.3).
3.1 Reducing to the broadcast-PULL Model
The following lemma establishes a formal relation between the convergence times of the
models we consider. We assume all models are subject to the same noise distribution.
Lemma 3.2. Any protocol operating in sequential-PULL can be simulated by a protocol
operating in broadcast-PULL with the same time complexity. Moreover, for any integer
1 ≤ k ≤ n, any protocol P operating in parallel-PULL(k) can be simulated by a protocol
operating in broadcast-PULL with a time complexity that is kn times that of P in
parallel-PULL(k).
Proof. Let us first show how to simulate a time step of sequential-PULL in the
broadcast-PULL model. Recall that in broadcast-PULL, in each time step, all agents
receive the same observation sampled u.a.r from the population. Upon drawing such
an observation, all agents use their shared randomness to generate a (shared) uniform
random integer X between 1 and n. Then, the agent whose unique identity corresponds
to X is the one processing the observation, while all other agents ignore it. This reduces
the situation to a scenario in sequential-PULL, and the agents can safely execute the
original algorithm designed for that model.
As for simulating a time step of parallel-PULL(k) in broadcast-PULL, agents di-
vide time steps in the latter model into rounds, each composing of precisely kn time
14
steps. Recall that the model assumes that agents share clocks that start when the
execution starts and tick at each time step. This implies that the agents can agree
on the division of time into rounds, and can further agree on the round number. For
1 ≤ i ≤ kn, during the i-th step of each round, only the agent whose identity is (i
mod n)+1 receives4 the observation, while all other agents ignore it. This ensures
that when a round is completed in the broadcast-PULL model, each agent receives
precisely k independent uniform samples as it would in a round of parallel-PULL(k).
Therefore, at the end of each round j ∈ N+ in the broadcast-PULL model, all agents
can safely execute their actions in the j'th round of the original protocol designed for
parallel-PULL(k). This draws a precise bijection from rounds in parallel-PULL(k) and
rounds in broadcast-PULL. The multiplicative overhead of kn simply follows from the
fact that each round in broadcast-PULL consists of kn time steps.
Thanks to Lemma 3.2, Theorem 3.1 directly follows from the next theorem.
Theorem 3.3. Consider the broadcast-PULL model and assume that the relaxed δ-
uniform noise criterion is satisfied. Any rumor spreading protocol cannot converges in
fewer time steps than Ω
(cid:17)
(cid:16)
n2δ
s2(1−2δ)2
.
The remaining of the section is dedicated to proving Theorem 3.3. Towards achieving
this, we view the task of guessing the correct opinion in the broadcast-PULL model,
given access to noisy samples, within the more general framework of distinguishing
between two types of stochastic processes which obey some specific assumptions.
3.2 Rumor Spreading and hypothesis testing
To establish the desired lower bound, we next show how the rumor spreading problem
in the broadcast-PULL model relates to a statistical inference test. That is, from the
perspective of a given agent, the rumor spreading problem can be understood as the
following: Based on a sequence of noisy observations, the agent should be able to tell
whether the correct opinion is 0 or 1. We formulate this problem as a specific task of
distinguishing between two random processes, one originated by running the protocol
assuming the correct opinion is 0 and the other assuming it is 1.
One of the main difficulties lies in the stochastic dependencies affecting these pro-
cesses. In general, at different time steps, they do not consist of independent draws of a
given random variable. In other words, the law of an observation not only depends on
the correct opinion, on the initial configuration and on the underlying randomness used
by agents, but also on the previous noisy observation samples and (consequently) on
the messages agents themselves choose to display on that round. An intuitive version
of this problem is the task of distinguishing between two (multi-valued) biased coins,
4Receiving the observation doesn't imply that the agent processes this observation. In fact, it will
store it in its memory until the round is completed, and process it only then.
15
whose bias changes according to the previous outcomes of tossing them (e.g., due to
wear). Following such intuition, we define the following general class of Adaptive Coin
Distinguishing Tasks, for short ACDT.
Definition 3.4 (ACDT). A distinguisher is presented with a sequence of observations
taken from a coin of type η where η ∈ {0, 1}. The type η is initially set to 0 or 1 with
probability 1/2 (independently of everything else). The goal of the distinguisher is to
determine the type η, based on the observations. More specifically, for a given time
step t, denote the sequence of previous observations (up to, and including, time t − 1)
by x(<t) = (x(1), . . . , x(t−1)). At each time t, given the type η ∈ {0, 1} and the history of
η ∈ Σ, which
previous observations x(<t), the distinguisher receives an observation X (t)
has law5 P (X (t)
η = m x(<t)).
We next introduce, for each m ∈ Σ, the parameter
ε(m, x(<t)) = P (X (t)
1 = m x(<t)) − P (X (t)
0 = m x(<t)).
Since, at all times t, it holds that
0 = m x(<t)) =
P (X (t)
m∈Σ
then (cid:80)
(cid:80)
m∈Σ ε(m, x(<t)) = 0. We shall be interested in the quantity dε(x(<t)) :=
m∈Σ ε(m, x(<t)), which corresponds to the (cid:96)1 distance between the distributions
1 = m x(<t)) given the sequence of previous obser-
0 = m x(<t)) and P (X (t)
m∈Σ
P (X (t)
1 = m x(<t)) = 1,
P (X (t)
vations.
(cid:88)
(cid:88)
Definition 3.5 (The bounded family ACDT(ε, δ)). We consider a family of instances
of ACDT, called ACDT(ε, δ), governed by parameters ε and δ. Specifically, this family
contains all instances of ACDT such that for every t, and every history x(<t), we have:
• dε(x(<t)) ≤ ε, and
• ∀m ∈ Σ such that ε(m, x(<t)) (cid:54)= 0, we have δ ≤ P (X (t)
η = m x(<t)) for η ∈ {0, 1}.
In the rest of the current section, we show how Theorem 3.3, that deals with the
broadcast-PULL model, follows directly from the next theorem that concerns the adap-
tive coin distinguishing task, by setting ε = 2s(1−2δ)
. The actual proof of Theorem 3.6
appears in Section 3.3.
Theorem 3.6. Consider any protocol for any instance of ACDT(ε, δ), The number of
samples required to distinguish between a process of type 0 and a process of type 1
with probability of error less than 1
ε2 . In particular, if
ε
(cid:16)
δ < 10, then the number of necessary samples is Ω(cid:0) δ
3 is at least ln 2
δ3−δ2ε+3δε2−ε3
(cid:17) δ
6(δ−ε)3
(cid:1).
n
9
ε2
5We follow the common practice to use uppercase letters to denote random variables and lowercase
letter to denote a particular realisation, e.g., X(≤t) for the sequence of observations up to time t, and
x(≤t) for a corresponding realization.
16
3.2.1 Proof of Theorem 3.3 assuming Theorem 3.6
Consider a rumor spreading protocol P in the broadcast-PULL model. Fix a node u.
We first show that running P by all agents, the perspective of node u corresponds to a
specific instance of ACDT( 2s(1−2δ)
, δ) called Π(P, u). We break down the proof of such
correspondence into two claims.
n
The ACDT instance Π(P, u). Recall that we assume that each agent knows the com-
plete neutral initial configuration, the number of sources s, and the shared of random
bits sequence r. We avoid writing such parameters as explicit arguments to Π(P, u)
in order to simplify notation, however, we stress that what follows assumes that these
parameters are fixed. The bounds we show hold for any fixed value of r and hence also
when r is randomized.
Each agent is interested in discriminating between two families of charged initial
configurations: Those in which the correct opinion is 0 and those in which it is 1 (each
of these possibilities occurs with probability 1
2 ). Recall that the correct opinion is
determined in the 2nd stage of the charged initial configuration, and is independent
form the choice of sources (1st stage).
We next consider the perspective of a generic non-source agent u, and define the
instance Π(P, u) as follows. Given the history x(<t), we set P (X (t)
η = m x(<t)), for
η ∈ {0, 1}, to be equal to the probability that u observes message m ∈ Σ at time
step t of the execution P. For clarity's sake, we remark that the latter probability is
conditional on: the history of observations being x(<t), ,the sequence of random bits r,
,the correct opinion being η ∈ {0, 1}, ,the neutral initial configuration, ,the identity of u,
,the algorithm P, and the system's parameters (including the distribution fsource−state
and the number of sources s).
Claim 3.7. Let P be a correct protocol for the rumor spreading problem in broadcast-PULL
and let u be an agent for which the protocol is guaranteed to produce the correct opin-
ion with probability at least p by some time T (if one exists), for any fixed constant
p ∈ (0, 1). Then Π(P, u) can be solved in time T with correctness being guaranteed with
probability at least p.
Proof. Conditioning on η ∈ {0, 1} and on the random seed r, the distribution of ob-
servations in the Π(P, u) instance follows precisely the distribution of observations as
perceived from the perspective of u in broadcast-PULL. Hence, if the protocol P at u
terminates with output j ∈ {0, 1} at round T , after the T -th observation in Π(P, u) we
can set Π(P, u)'s output to j as well. Given that the two stochastic processes have the
same law, the correctness guarantees are the same.
Lemma 3.8. Π(P, u) ∈ ACDT
(cid:16) 2(1−2δ)s
(cid:17)
, δ
.
n
17
n
Proof. Since the noise in broadcast-PULL flips each message m ∈ Σ into any m(cid:48) ∈ Σ(cid:48)
with probability at least δ, regardless of the previous history and of η ∈ {0, 1}, at all
times t, if m ∈ Σ(cid:48) then P (X (t)
η = m x(<t)) ≥ δ. Consider a message m ∈ Σ\ Σ(cid:48) (if such
a message exists). By definition, such a message could only be received by observing a
non-source agent. But given the same history x(<t), the same sequence of random bits
r, and the same initial knowledge, the behavior of a non-source agent is the same, no
matter what is the correct opinion η. Hence, for m ∈ Σ \ Σ(cid:48) we have P (X (t)
0 = m
1 = m x(<t)), or in other words, m ∈ Σ \ Σ(cid:48) =⇒ ε(m, x(<t)) = 0.
x(<t)) = P (X (t)
It remains to show that dε(x(<t)) ≤ 2(1−2δ)s
. Let us consider two executions of
the rumor spreading protocol, with the same neutral initial configuration, same shared
sequence of random bits r, same set of sources, except that in the first the correct
opinion is 0 while in the other it is 1. Let us condition on the history of observations
x(<t) being the same in both processes. As mentioned, given the same history x(<t),
the behavior of a non-source agent is the same, regardless of the correct opinion η. It
follows that the difference in the probability of observing any given message is only due
to the event that a source is observed. Recall that the number of sources is s. Therefore,
the probability of observing a source is s/n, and we may write as a first approximation
ε(m, x(<t)) ≤ s/n . However, we can be more precise. In fact, ε(m, x(<t)) is slightly
smaller than s/n, because the noise can still affect the message of a source. We may
interpret ε(m, x(<t)) as the following difference. For a source v ∈ S, let mv
η be the
message of u assuming the given history x(<t) and that v is of type η ∈ {0, 1} (the
message mv
η is deterministically determined given the sequence r of random bits, the
neutral initial configuration, the parameters of the system, and the identity of v). Let
αm(cid:48),m be the probability that the noise transforms a message m(cid:48) into a message m.
Then ε(m, x(<t)) = 1
n
0,m), and
v∈S(αmv
(cid:80)
(cid:88)
m∈Σ
1,m − αmv
ε(m, x(<t)) ≤ 1
n
(cid:88)
(cid:88)
m∈Σ
v∈S
dε(x(<t)) =
αmv
1,m − αmv
0,m.
(1)
By the definition of ACDT(ε, δ), it follows that either αmv
or δ ≤ αmv
(1), we can use the following claim.
0,m (if ε(m, x(<t)) = 0)
0,m ≤ 1 − δ (if ε(m, x(<t)) (cid:54)= 0). Thus, to bound the right hand side in
1,m = αmv
1,m, αmv
element m ∈ Σ, δ ≤ P (m), Q(m) ≤ 1 − δ. Then(cid:80)
Claim 3.9. Let P and Q be two distributions over a universe Σ such that for any
Proof of Claim 3.9. Let Σ+ := {m : P (m) > Q(m)}. We may write
P (m) − Q(m) =
(P (m) − Q(m)) +
(Q(m) − P (m))
(cid:88)
m∈ΣP (m) − Q(m) ≤ 2(1 − 2δ).
(cid:88)
m∈Σ+
(cid:88)
m∈Σ
m∈setminusΣ+
= P (Σ+) − Q(Σ+) + Q(Σ \ Σ+) − P (Σ \ Σ+)
= 2(P (Σ+) − Q(Σ+)),
18
obtain
1
n
(cid:88)
(cid:88)
m∈Σ
v∈S
αmv
1,m − αmv
(cid:16) 2(1−2δ)s
0,m ≤ 1
(cid:17)
n
2
(cid:88)
v∈S
(1 − 2δ) ≤ 2(1 − 2δ)s
n
.
Case 2. Otherwise, Σ+ ≥ 2 and 2(cid:80)
where in the last line we used the fact that Q(Σ \ Σ+) − P (Σ \ Σ+) = 1 − Q(Σ+) −
1 + P (Σ+) = P (Σ+) − Q(Σ+). We now distinguish two cases. Case 1.If Σ+ is a
singleton, Σ+ = {m∗}, then P (Σ+)−Q(Σ+) = P (m∗)−Q(m∗) ≤ 1−2δ, by assumption.
Q(m) ≤
2(1 − δΣ+) ≤ 2(1 − 2δ), using the fact that for any m, Q(m) ≥ δ, and the fact that P
is a probability measure. This completes the proof of Claim 3.9.
(P (m) − Q(m)) ≤ 2 − 2(cid:80)
m∈Σ+
m∈Σ+
Applying Claim 3.9 for a fixed v ∈ S to distributions (αmv
0,m)m and (αmv
1,m)m, we
Hence, we have Π(P) ∈ ACDT
, δ
, establishing Lemma 3.8.
n
Thanks to Claims 3.7 and Lemma 3.8, Theorem 3.3 regarding the broadcast-PULL
(cid:16) ε2
(cid:17)
model becomes a direct consequence of Theorem 3.6 on the adaptive coin distinguishing
task, taking ε = 2(1−2δ)s
δsn ≤ c for some small
constant c, ensures that ε
. More precisely, the assumption (1−2δ)
(cid:17)
δ ≤ c as required by Theorem 3.6. The lower bound Ω
(cid:16)
n
δ
corresponds to Ω
. This concludes the proof of Theorem 3.3.
n2δ
(1−2δ)2s2
To establish our results it remains to prove Theorem 3.6.
3.3 Proof of Theorem 3.6
We start by recalling some facts from Hypothesis Testing. First let us recall two standard
(cid:80)
notions of (pseudo) distances between probability distributions. Given two discrete dis-
tributions P0, P1 over a probability space Ω with the same support6, the total variation
x∈ΩP0(x) − P1(x), and the Kullback-Leibler
distance is defined as T V (P0, P1) := 1
2
divergence KL(P0, P1) is defined7 as KL(P0, P1) :=(cid:80)
x∈Ω P0(x) log P1(x)
P0(x) .
The following lemma shows that, when trying to discriminate between distributions
P0, P1, the total variation relates to the smallest error probability we can hope for.
Lemma 3.10 (Neyman-Pearson [36, Lemma 5.3 and Proposition 5.4]). Let P0, P1 be
two distributions. Let X be a random variable of law either P0 or P1. Consider a
(possibly probabilistic) mapping f : Ω → {0, 1} that attempts to "guess" whether the
and is thus made for simplicity's sake.
6The assumption that the support is the same is not necessary but it is sufficient for our purposes,
7We use the notation log(·) to denote the base 2 logarithms, i.e., log2(·) and for a probability
distribution P , use the notation P (x) as a short for P (X = x).
19
observation X was drawn from P0 (in which case it outputs 0) or from P1 (in which
case it outputs 1). Then, we have the following lower bound,
P0 (f (X) = 1) + P1 (f (X) = 0) ≥ 1 − T V (P0, P1).
The total variation is related to the KL divergence by the following inequality.
Lemma 3.11 (Pinsker [36, Lemma 5.8]). For any two distributions P0, P1,
T V (P0, P1) ≤(cid:112)KL (P0, P1).
We are now ready to prove the theorem.
0
η
Proof of Theorem 3.6. Let us define Pη ( · ) = P ( · "correct distribution is η") for η ∈
{0, 1}. We denote P (≤t)
, η ∈ {0, 1}, the two possible distributions of X(≤t). We refer to
P (≤t)
as the distribution of type 1. Furthermore,
we define the correct type of a sequence of observations X(≤t) to be 0 if the observations
are sampled from P (≤t)
, and to be 1 if they are sampled from P (≤t)
as the distribution of type 0 and to P (≤t)
After t observations x(≤t) = (x(1), . . . , x(t)) we have to decide whether the distribu-
tion is of type 0 or 1. Our goal is to maximize the probability of guessing the type of
the distribution, observing X(≤t), which means that we want to minimize
0
1
.
1
f (X(≤t)) = 1 − η
Pη
P ("correct type is η") .
(2)
(cid:88)
(cid:16)
f =
η∈{0,1}
Recall that the correct type is either 0 or 1 with probability 1
bility described in (2) becomes
2 . Thus, the error proba-
1
2
P0
f (X(≤t)) = 1
+
1
2
P1
f (X(≤t)) = 0
.
(3)
By combining Lemmas 3.10 and 3.11 with X = X(≤t) and Pη = P (≤t)
for η = 0, 1,
we get the following Theorem. Although for convenience we think of f as a deterministic
function, it could in principle be randomized.
η
(cid:17)
Theorem 3.12. Let f be any guess function. Then
f t(X(≤t)) = 1
P0
+ P1
f (X(≤t)) = 0
(cid:16)
(cid:114)
(cid:17) ≥ 1 −
KL
(cid:16)
(cid:17)
.
P (≤t)
0
, P (≤t)
1
Theorem 3.12 implies that for the probability of error to be small, it must be the
is large. Our next goal is therefore to show that
, P (≤t)
P (≤t)
case that the term KL
in order to make this term large, t must be large.
0
1
(cid:17)
(cid:16)
(cid:17)
(cid:16)
(cid:17)
20
(cid:16)
(cid:17)
(cid:16)
η
Note that P (≤T )
for η ∈ {0, 1} cannot be written as the mere product of the marginal
distributions of the X (t)s, since the observations at different times may not necessarily
be independent. Nevertheless, we can still express the term KL(P (≤T )
) as a
sum, using the Chain Rule for KL divergence8. It yields
KL(P (≤T )
KL(P0(x(t) x(<t)), P1(x(t) x(<t)))
, P (≤T )
, P (≤T )
(cid:88)
) =
(4)
0
1
0
1
t≤T
(cid:88)
(cid:88)
x(<t)∈Σt−1
x(<t)∈Σt−1
:=
=
(cid:88)
(cid:88)
x(t)∈Σ
m∈Σ
P0(x(<t))
P0(x(t) x(<t)) log
P0(x(t) x(<t))
P1(x(t) x(<t))
.
P0(x(<t))
P0(X (t)
0 = m x(<t)) log
P (X (t)
P (X (t)
0 = m x(<t))
1 = m x(<t))
(5)
.
Since we are considering an instance of ACDT (ε, δ), we have
• dε(x(<t)) =(cid:80)
m∈Σ ε(m, x(<t)) ≤ ε, and
• for every m ∈ Σ such that ε(m, x(<t)) (cid:54)= 0, it holds that δ ≤ Pη(X (t)
0 = m x(<t))
for η ∈ {0, 1}.
We make use of the previous facts to upper bound the KL divergence terms in the right
hand side of (5), as follows.
(cid:33)
KL(P0(x(t) x(<t)), P1(x(t) x(<t)))
P0(x(<t))
P (X (t)
x(<t)∈Σt−1
=
(cid:88)
= −(cid:88)
x(<t)
(cid:32)
(cid:88)
(cid:32)
(cid:88)
m∈Σ
m∈Σ
0 = m x(<t)) log
(cid:32)
P (X (t)
P (X (t)
0 = m x(<t)) + ε(m, x(<t))
0 = m x(<t))
(cid:33)(cid:33)
ε(m, x(<t))
0 = m x(<t))
P (X (t)
.
(6)
P0(x(<t))
P (X (t)
0 = m x(<t)) log
1 +
Recall that we assume
ε(m,x(<t))
0 =mx(<t))
P (X (t)
≤ ε(m,x(<t))
δ
≤ ε
δ . We make use of the following
claim, which follows from the Taylor expansion of log(1 + u) around 0.
Claim 3.13. Let x ∈ [−a, a] for some a ∈ (0, 1). Then log(1 + x)− x + x2/2 ≤ x3
3(1−a)3 .
8See Lemma 3 in http://homes.cs.washington.edu/anuprao/pubs/CSE533Autumn2010/lecture3.
pdf.
21
(8)
0 = m x(<t))2(cid:17)
(9)
m∈Σ
m ε(m, x(<t)) = 0, thus the first term in (7) disappears. Hence, sub-
δ , we can bound the inner sum appearing in (6) from
(cid:33)(cid:33)
.
(ε(m, x(<t)))3
0 = m x(<t))2
P (X (t)
(7)
m (ε(m, x(<t)))2 ≤ ε2. The latter
0 = m x(<t)) ≥ δ for any m ∈ Σ such that
(cid:32)
(ε(m, x(<t)))2
0 = m x(<t))
P (X (t)
mε(m, x(<t)) ≤ ε, we also have that (cid:80)
±
δ3
3(δ − ε)3
1
ln 2
m∈Σ
above and below with
Using Claim 3.13 with a = ε
ε(m, x(<t)) − 1
2
bound, together with the fact that P (X (t)
(cid:32)
(cid:88)
Since (cid:80)
ε( m, x(<t)) (cid:54)= 0, implies (cid:88)
Finally, we can similarly bound the term(cid:80)
Recall that(cid:80)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)log
(cid:32)
ε(m, x(<t))
0 = m x(<t))
(cid:88)
P (X (t)
P (X (t)
(cid:16)
with
1 +
m
(ε(m, x(<t)))3/P (X (t)
stituting the bounds (8) and (9) in (7), we have
(ε(m, x(<t)))3/P (X (t)
.
m∈Σ
≤ ε2
δ
(ε(m, x(<t)))2
(cid:16)
0 = m x(<t))
0 = m x(<t))2(cid:17) ≤ ε3
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 1
ε2
δ
+
δ2 .
(cid:18) 1
(cid:18) 1
2
ln 2
≤ 1
ln 2
(cid:19)
(cid:19) ε2
δ
δε3
3(δ − ε)3
δ2ε
+
3(δ − ε)3
2
.
(10)
If we define the right hand side (10) to be W (ε, δ) and we substitute the previous bound
in (6), we get
KL(P0(x(t) x(<t)), P1(x(t) x(<t))) ≤ W (ε, δ),
and combining the previous bound with (4), we can finally conclude that for any integer
T , we have KL(P (≤T )
) ≤ T · W (ε, δ). Thus, from Theorem 3.12 and the latter
bound, it follows that the error under a uniform prior of the source type, as defined in
(3), is at least
, P (≤T )
1
0
(cid:16)
(cid:17)
(cid:16)
1
2
P0
f t(X(≤t)) = 1
+
1
2
P1
f (X(≤t)) = 0
(cid:17) ≥ 1
2
≥ 1
2
− 1
2
− 1
2
KL(P (≤T )
(cid:113)
(cid:112)T · W (ε, δ).
(cid:16)
0
, P (≤T )
)
1
1
W (ε,δ) = ln 2
9
6(δ−ε)3
δ3−δ2ε+3δε2−ε3
(cid:17) δ
ε2
Hence, the number of samples T needs to be greater than 1
9
to allow the possibility that the error be less than 1/3.
22
100
In particular, if we assume that 10ε < δ, then we can bound
2187 . It follows that (10) can be bounded with W (ε, δ) ≤ 1
so 1
9
Theorem 3.3.
W (ε,δ) ≥ 0.14 · δ
ε2 = Ω(cid:0) δ
ε2
1
(cid:1) ≤ 0.79 , and
(cid:1) . This completes the proof of Theorem 3.6 and hence of
(cid:0) 1
3(δ−ε)3 ≤ δ3
10·
2 + 100
3(9/10)3δ3 ≤
2187
δ2ε
ln 2
1
References
[1] A. El Gamal a. Young-Han Kim. Network Information Theory. 2011.
[2] M. Abeles. Corticonics: Neural circuits of the cerebral cortex. 1991.
[3] N. Alon, M. Braverman, K. Efremenko, R. Gelles, and B. Haeupler. Reliable
communication over highly connected noisy networks. In PODC, pages 165–173,
2016.
[4] F. Amor, P. Ortega, X. Cerd´a, and R. Boulay. Cooperative prey-retrieving in the
ant cataglyphis floricola: An unusual short-distance recruitment. Insectes Sociaux,
57(1), 2010.
[5] D. Angluin, J. Aspnes, and D. Eisenstat. A simple population protocol for fast
robust approximate majority. Distributed Computing, 21(2):87–102, 2008.
[6] D. Angluin, J. Aspnes, M. J. Fischer, and H. Jiang. Self-stabilizing population
protocols. TAAS, 3(4), 2008.
[7] J. Aspnes and E. Ruppert. An introduction to population protocols. Bulletin of
the EATCS, 93:98–117, 2007.
[8] W. Bialek. Physical limits to sensation and perception. Annual review of biophysics
and biophysical chemistry, 16(1):455–478, 1987.
[9] S. Bikhchandani, D. Hirshleifer, and I. Welch. Learning from the behavior of
others: Conformity, fads, and informational cascades. J. Economic Perspectives,
12(3):151–170, 1998.
[10] L. Boczkowski, A. Korman, and E. Natale. Minimizing message size in stochastic
In SODA,
communication patterns: Fast self-stabilizing protocols with 3 bits.
pages 2540–2559, 2017.
[11] A. Cavagna, A. Cimarelli, I. Giardina, G. Parisi, R. Santagati, F. Stefanini, and
M. Viale. Scale-free correlations in starling flocks. PNAS, 107(26):11865–11870,
2010.
23
[12] K. Censor-Hillel, B. Haeupler, J. A. Kelner, and P. Maymounkov. Global compu-
tation in a poorly connected world: fast rumor spreading with no dependence on
conductance. In STOC, pages 961–970, 2012.
[13] E. Chastain, A. Livnat, C. Papadimitriou, and U. Vazirani. Algorithms, games,
doi:10.1073/pnas.
and evolution. PNAS, 111(29):10620–10623, July 2014.
1406556111.
[14] Bernard Chazelle. Natural algorithms. In SODA, pages 422–431, 2009.
[15] Thomas M Cover and B Gopinath. Open problems in communication and compu-
tation. Springer Science & Business Media, 2012.
[16] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker, H. Sturgis,
D. Swinehart, and D. Terry. Epidemic algorithms for replicated database mainte-
nance. In PODC, 1987.
[17] B. Doerr, L. A. Goldberg, L. Minder, T. Sauerwald, and C. Scheideler. Stabilizing
consensus with the power of two choices. In SPAA, pages 149–158, 2011.
[18] A. Dornhaus and L. Chittka. Food alert in bumblebees (bombus terrestris): Pos-
sible mechanisms and evolutionary implications. Behavioral Ecology and Sociobi-
ology, 50(6):570–576, 2001.
[19] R. Elsasser and T. Sauerwald. On the runtime and robustness of randomized
broadcasting. Theor. Comput. Sci., 410(36):3414–3427, 2009.
[20] O. Feinerman, B. Haeupler, and A. Korman. Breathe before speaking: efficient
information dissemination despite noisy, limited and anonymous communication.
In PODC, 2014.
[21] O. Feinerman and A. Korman. Theoretical distributed computing meets biology:
A review. In ICDCIT, pages 1–18. Springer, 2013.
[22] O. Feinerman, A. Rotem, and E. Moses. Reliable neuronal logic devices from
patterned hippocampal cultures. Nature physics, 4(12):967–973, 2008.
[23] P. Fraigniaud and E. Natale. Noisy rumor spreading and plurality consensus. In
PODC, pages 127–136, 2016.
[24] R. G. Gallager. Finding parity in a simple broadcast network. IEEE Trans. Inf.
Theor., 34(2):176–180, 2006.
[25] L. A. Giraldeau, T. J. Valone, and J.J. Templeton. Potential disadvantages of using
socially acquired information. Philosophical Transactions of the Royal Society of
London B: Biological Sciences, 357(1427):1559–1566, 2002.
24
[26] N. Goyal, G. Kindler, and M. E. Saks. Lower bounds for the noisy broadcast
problem. SIAM J. Comput., 37(6):1806–1841, 2008.
[27] B. Holldobler. Recruitment behavior in camponotus socius (hym. formicidae). J.
of Comparative Physiology A, 75(2):123–142, 6 1971.
[28] R. M. Karp, C. Schindelhauer, S. Shenker, and B. Vocking. Randomized rumor
spreading. In FOCS, pages 565–574, 2000.
[29] D. Kempe, A. Dobra, and J. Gehrke. Gossip-based computation of aggregate
information. In FOCS, pages 482–491. IEEE, 2003.
[30] A. Korman, E. Greenwald, and O. Feinerman. Confidence sharing: An economic
strategy for efficient information flows in animal groups. PLoS Comp. Biology,
10(10), 2014.
[31] S. Marras, R. Batty, and P. Domenici.
Information transfer and antipredator
maneuvers in schooling herring. Adaptive Behavior, 20(1):44–56, 2012.
[32] C. Musco, H. Su, and N. A. Lynch. Ant-inspired density estimation via random
walks: Extended abstract. In PODC, pages 469–478, 2016.
[33] B. Pittel. On spreading a rumor. SIAM J. Appl. Math., 47(1):213–223, 1987.
[34] N. Razin, J.P. Eckmann, and O. Feinerman. Desert ants achieve reliable re-
Journal of the Royal Society Interface;
cruitment across noisy interactions.
10(20170079)., 2013.
[35] G. Rieucau and L. A. Giraldeau. Persuasive companions can be wrong: the use
of misleading social information in nutmeg mannikins. Behavioral Ecology, pages
1217–1222, 2009.
[36] P. Rigollet. High dimensional statistics. Lecture notes for course 18S997., 2015.
[37] S. B. Rosenthal, C. R. Twomey, A. T. Hartnett, H. S. Wu, and I. D. Couzin. Reveal-
ing the hidden networks of interaction in mobile animal groups allows prediction
of complex behavioral contagion. PNAS, 112(15):4690–4695, 2015.
[38] J. J. Templeton and Giraldeau L. A. Patch assessment in foraging flocks of european
starlings: evidence for the use of public information. Behavioral Ecology, 6(1):65–
72, 1995.
[39] J. von Neumann. Probabilistic logics and the synthesis of reliable organisms from
unreliable components. Automata Studies, pages 43–98, 1956.
25
[40] A. Xu and M. Raginsky. Information-theoretic lower bounds for distributed func-
tion computation. IEEE Trans. Information Theory, 63(4):2314–2337, 2017.
[41] Y.Afek, N. Alon, O. Barad, E. Hornstein, N. Barkai, and Z. Bar-joseph. A biological
solution to a fundamental distributed computing problem. Science, 2011.
A Methods
All experimental results presented in this manuscript are re-analysis of data obtained in
Cataglyphis niger recruitment experiments [34]. In short, ants in the entrance chamber
of an artificial nest were given access to a tethered food item just outside the nest's
entrance (Fig 2a). The inability of the ants to retrieve the food induced a recruitment
process [34].
The reaction of the ants to this manipulation was filmed and the locations, speeds and
interactions of all participating ants were extracted from the resulting videos.
Calculation of δ. To estimate the parameter δ we used interactions between ants
moving at four different speed ranges (measured in cm/sec), namely, 'a': 0-1, 'b': 1-
5, 'c': 5-8, and 'd': over 8, and stationary "receiver" ants where used. The message
alphabet is then assumed to be Σ = {a, b, c, d}. The response of a stationary ant v to
the interaction was quantified in terms of her speed after the interaction. Assuming
equal priors to all messages in Σ, and given specific speed of the receiver ant, v, the
probability that it was the result of a specific message i ∈ Σ was calculated as pi(v) =
k∈Σ p(v k), where p(v j) is the probability of responding in speed v after
"observing" j. The probability δ(i, j) that message i was perceived as message j was
then estimated as the weighted sum over the entire probability distribution measured
v p(v j) · pi(v). The parameter δ can then be calculated
using δ = min{δ(i, j) i, j ∈ Σ}.
p(v i)/(cid:80)
as a response to j: δ(i, j) =(cid:80)
B Proof of Corollary 1.3
Let us start with the first item of the corollary, namely the lower bound in the sequential-PULL
model. For any step t, let S(t) denote the set of sources together with the agents that
have directly observed at least one of the sources at some point up to time t. We have
S = S(0) ⊆ S(1) ⊆ S(2) ⊆ . . .. The size of the set S(t) is a random variable which is
expected to grow at a moderate speed. Specifically, letting s(cid:48) = 11
Claim B.1. With probability at least 1 − n−10, we have S(T ) ≤ s(cid:48).
10 · s · T , we obtain:
26
Proof of Claim B.1. The variable S(T ) may be written as a sum of indicator variables
n(cid:88)
≤ n(cid:88)
i=1
i=1
(cid:88)
r≤T
S(T ) =
1(Agent i observed at least one source before step t)
1(Agent i observes a source on step r).
This last expression is a sum of n · T independent Bernoulli variables with parameter
s/n. In other terms, it is a binomial variable with probability s/n and T · n trials. By a
standard Chernoff bound the probability that it deviates by a multiplicative factor 11
from its mean s · T is less than exp(−Ω(sT )) ≤ n−10. The last bound holds because we
10
assume sT ≥ C log n for some large enough constant C.
Denote by E the event that S(t) ≤ s(cid:48) for every t ≤ T . Using Claim B.1, we know
that P (E) ≥ 1 − n−10. Our goal next is to prove that the probability ρ that a given
agent correctly guesses the correct opinion is low for any given time t ≤ cT , where c is a
small constant. For this purpose, we condition on the highly likely event E. Removing
this conditioning will amount to adding a negligible term (of order at most n−10) to ρ.
In order to bound ρ, we would like to invoke Theorem 3.3 with the number of
sources upper bounded by s(cid:48). Let us explain why it applies in this context. To begin
with, we may adversarially assume (from the perspective of the lower bound) that all
agents in S(t) learn the value of the correct bit to spread. Thus, they essentially become
"sources" themselves. In this case the number of sources varies with time, but the proof
of Theorem 3.3 can easily be shown to cover this case as long as s (i.e., s(cid:48) here) is
an upper bound on the number of sources at all times. We can therefore safely apply
Theorem 3.3 with s(cid:48). By the choice of T ,
(cid:18)
(cid:19)
T = Θ
n2δ
(s(cid:48))2(1 − 2δ)2
.
Hence, we can set c to be a sufficiently small constant such that for all times t ≤ cT ,
the probability of guessing correctly, even in this adversarial scenario, is less than 1/3.
In other words, we have ρ ≤ 1/3. All together, this yields a lower bound of Ω(T ) on
the convergence time.
As for the parallel-PULL(k) model, the argument is similar. After T (cid:48) = T /k parallel
rounds, using a similar claim as Claim B.1, we have that with high probability, at most
O(ksT (cid:48)) agents have directly observed one of the s sources by time T (cid:48). Applying
Theorem 3.3 with s(cid:48)(cid:48) = O(ksT (cid:48)) = O(sT ) yields a lower bound (in terms of samples in
the broadcast model) of
(cid:18)
Θ
n2δ
(s(cid:48)(cid:48))2(1 − 2δ)2
n2δ
s2T 2(1 − 2δ)2
= Θ(T ).
(cid:19)
(cid:19)
(cid:18)
= Θ
27
The last line follows by choice of T . Hence T is a lower bound on the number of samples,
which is attained in T (cid:48) rounds of parallel-PULL(k) model.
28
|
1204.5399 | 3 | 1204 | 2013-04-09T21:49:06 | A Consensual Linear Opinion Pool | [
"cs.MA",
"math.ST",
"math.ST"
] | An important question when eliciting opinions from experts is how to aggregate the reported opinions. In this paper, we propose a pooling method to aggregate expert opinions. Intuitively, it works as if the experts were continuously updating their opinions in order to accommodate the expertise of others. Each updated opinion takes the form of a linear opinion pool, where the weight that an expert assigns to a peer's opinion is inversely related to the distance between their opinions. In other words, experts are assumed to prefer opinions that are close to their own opinions. We prove that such an updating process leads to consensus, \textit{i.e.}, the experts all converge towards the same opinion. Further, we show that if rational experts are rewarded using the quadratic scoring rule, then the assumption that they prefer opinions that are close to their own opinions follows naturally. We empirically demonstrate the efficacy of the proposed method using real-world data. | cs.MA | cs | A Consensual Linear Opinion Pool
Arthur Carvalho
Kate Larson
Cheriton School of Computer Science
Cheriton School of Computer Science
University of Waterloo
[email protected]
University of Waterloo
[email protected]
3
1
0
2
r
p
A
9
]
A
M
.
s
c
[
3
v
9
9
3
5
.
4
0
2
1
:
v
i
X
r
a
Abstract
An important question when eliciting opinions
from experts is how to aggregate the reported opin-
ions. In this paper, we propose a pooling method
to aggregate expert opinions. Intuitively, it works
as if the experts were continuously updating their
opinions in order to accommodate the expertise of
others. Each updated opinion takes the form of a
linear opinion pool, where the weight that an ex-
pert assigns to a peer’s opinion is inversely related
to the distance between their opinions.
In other
words, experts are assumed to prefer opinions that
are close to their own opinions. We prove that such
an updating process leads to consensus, i.e., the ex-
perts all converge towards the same opinion. Fur-
ther, we show that if rational experts are rewarded
using the quadratic scoring rule, then the assump-
tion that they prefer opinions that are close to their
own opinions follows naturally. We empirically
demonstrate the efficacy of the proposed method
using real-world data.
1 Introduction
Predicting outcomes of relevant uncertain events plays an es-
sential role in decision-making processes. For example, com-
panies rely on predictions about consumer demand and ma-
terial supply to make their production plans, while weather
forecasts provide guidelines for long range or seasonal agri-
cultural planning, e.g., farmers can select crops that are best
suited to the anticipated climatic conditions.
Forecasting techniques can be roughly divided into statis-
tical and non-statistical methods. Statistical methods require
historical data that contain valuable information about the fu-
ture event. When such data are not available, a widely used
non-statistical method is to request opinions from experts re-
garding the future event [Cooke, 1991]. Opinions usually
take the form of either numerical point estimates or prob-
ability distributions over plausible outcomes. We focus on
opinions as probability mass functions.
The literature related to expert opinions
typi-
cally concerned about how expert opinions are used
[Mosleh et al., 1988], how uncertainty is or should be repre-
sented [Ng and Abramson, 1990], how experts do or should
is
reason with uncertainty [Cooke, 1991], how to score the
quality and usefulness of expert opinions [Savage, 1971;
Boutilier, 2012], and how to produce a single consensual
opinion when different experts report differing opinions
[DeGroot, 1974]. It is this last question that we address in
this paper.
We propose a pooling method to aggregate expert opinions
that works as if the experts were continuously updating their
opinions in order to accommodate the expertise and knowl-
edge of others. Each updated opinion takes the form of a
linear opinion pool, or a convex combination of opinions,
where the weight that an expert assigns to a peer’s opinion
is inversely related to the distance between their opinions. In
other words, experts are assumed to prefer opinions that are
close to their own opinions. We prove that such an updat-
ing process leads to consensus, i.e., the experts all converge
towards the same opinion. We also show that if the opinions
of rational experts are scored using the quadratic scoring rule,
then the assumption that experts prefer opinions that are close
to their own follows naturally.
2 Related Work
The aggregation of expert opinions have been extensively
studied in computer science and, in particular, artificial intel-
ligence, e.g., the aggregation of opinions represented as pref-
erences over a set of alternatives as in social choice theory
[Chevaleyre et al., 2007], the aggregation of point estimates
using non-standard opinion pools [Jurca and Faltings, 2008],
and the aggregation of probabilistic opinions using prediction
markets [Chen and Pennock, 2010].
A traditional way of aggregating probabilistic opinions
is through opinion pooling methods. These methods are
often divided into behavioral and mathematical methods
[Clemen and Winkler, 1999]. Behavioral aggregation meth-
ods attempt to generate agreement among the experts through
interactions in order for them to share and exchange knowl-
edge. Ideally, such sharing of information leads to a consen-
sus. However, these methods typically provide no conditions
under which the experts can be expected to reach agreement
or even for terminating the iterative process.
On the other hand, mathematical aggregation methods con-
sist of processes or analytical models that operate on the in-
dividual probability distributions in order to produce a single,
aggregate probability distribution. An important mathemati-
cal method is the linear opinion pool, which involves taking
a weighted linear average of the opinions [Cooke, 1991].
Several interpretations have been offered for the weights
in the linear opinion pool. The performance-based approach
recommends setting the weights based on previous perfor-
mance of the experts [Genest and McConway, 1990]. A
caveat with this approach is that performance measurements
typically depend on the true outcome of the underlying event,
which might not be available at the time when the opinions
have to be aggregated. Also, previous successful (respective-
ly, unsuccessful) predictions are not necessarily good indica-
tors of future successful (respectively, unsuccessful) ones.
More closely related to this work is the interpretation of
weights as a measure of distance. For example, Barlow et
al. [1986] proposed that the weight assigned to each expert’s
opinion should be inversely proportional to its distance to the
most distant opinion, where distance is measured according
to the Kullback-Leibler divergence. A clear drawback with
this approach is that it only considers the distance to the most
distant opinion when assigning a weight to an expert’s opin-
ion. Thus, even if the majority of experts have similar and
accurate opinions, the weights of these experts’ opinions in
the aggregate prediction can be greatly reduced due to a sin-
gle distant opinion.
For a comprehensive review of different perspectives on
the weights in the linear opinion pool, we refer the interested
reader to the work by Genest and McConway [1990].
3 Model
We consider the forecasting setting where a decision maker
is interested in a probability vector over a set of mutually
exclusive outcomes θ1, . . . , θz, for z ≥ 2. The decision
maker deems it inappropriate to interject his own judgment
about these outcomes. Hence, he elicits probabilistic opin-
ions from n experts. Experts’ opinions are represented by
z-dimensional probability vectors f1, . . . , fn. The probabil-
ity vector fi = (fi,1, . . . , fi,z) represents expert i’s opinion,
where fi,k is his subjective probability regarding the occur-
rence of outcome θk.
Since experts are not always in agreement, belief aggrega-
tion methods are used to combine their opinions into a single
probability vector. Formally, f = T (f1, . . . , fn), where f is
called an opinion pool, and the function T is the pooling op-
erator. The linear opinion pool is a standard approach that
involves taking a weighted linear average of the opinions:
n
T (f1, . . . , fn) =
wifi
(1)
where wi denotes the weight associated with expert i’s opin-
ion. We make the standard assumption that 0 ≤ wi ≤ 1, for
every i ∈ {1, . . . , n}, andPn
i=1 wi = 1.
3.1 Consensus and Weights
DeGroot [1974] proposed a model which describes how a
group can reach agreement on a common probability distri-
bution by pooling their individual opinions.
Initially, each
expert i is informed of the opinion of every other expert. In
Xi=1
order to accommodate the information and expertise of the
rest of the group, expert i updates his own opinion as follows:
n
f (1)
i =
pi,j fj
Xj=1
where pi,j is the weight that expert i assigns to the opinion
of expert j when he carries out this update. Weights must be
chosen on the basis of the relative importance that experts as-
sign to their peers’ opinions. It is assumed that pi,j > 0, for
j=1 pi,j = 1. In this way, each
updated opinion takes the form of a linear opinion pool. The
whole updating process can be written in a slightly more gen-
eral form using matrix notation, i.e., F(1) = PF(0), where:
every expert i and j, and Pn
P =
F(0) =
p1,1
p2,1
...
pn,1
f1
f2
...
fn
p1,2
p2,2
...
pn,2
=
· · ·
· · ·
...
· · ·
f1,1
f2,1
...
fn,1
p1,n
p2,n
...
pn,n
f1,2
f2,2
...
fn,2
· · ·
· · ·
...
· · ·
,
and
f1,z
f2,z
...
fn,z
Since all the opinions have changed, the experts might wish
to revise their new opinions in the same way as they did
before.
If there is no basis for the experts to change their
weights, we can then represent the whole updating process
after t revisions, for t ≥ 1, as follows:
(2)
F(t) = PF(t−1) = PtF(0)
i,1 , . . . , f (t)
Let f
(t)
i = (cid:16)f (t)
i,z(cid:17) be expert i’s opinion after t
j
i = f (t)
updates, i.e., it denotes the ith row of the matrix F(t). We
say that a consensus is reached if f (t)
, for every expert
i and j, as t → ∞. Since P, the matrix with weights, is a
n×n stochastic matrix, it can then be regarded as the one-step
transition probability matrix of a Markov chain with n states
and stationary probabilities. Consequently, one can apply a
limit theorem that says that a consensus is reached when there
exists a positive integer t such that every element in at least
one column of the matrix Pt is positive [DeGroot, 1974].
3.2 Weights as a Measure of Distance
The original method proposed by DeGroot [1974] has some
drawbacks. First,
the experts might want to change the
weights that they assign to their peers’ opinions after learn-
ing their initial opinions or after observing how much the
opinions have changed from stage to stage. Further, opin-
ions and/or identities have to be disclosed to the whole group
when the experts are assigning the weights. Hence, privacy
is not preserved, a fact which might be troublesome when the
underlying event is of a sensitive nature.
In order to tackle these problems, we derive the weights
that experts assign to the reported opinions by interpreting
each weight as a measure of distance between two opinions.
We start by making the assumption that experts prefer opin-
ions that are close to their own opinions, where closeness is
measured by the following distance function:
D(fi, fj) =rPz
k=1(fi,k − fj,k)2
z
i.e., it is the root-mean-square deviation between two opin-
ions fi and fj. Given the above assumption, one can estimate
the weight that expert i assigns to expert j’s opinion at a given
time t, for t ≥ 1, as follows:
(3)
(4)
p(t)
i,j =
α(t)
i
ǫ + D(cid:16)f (t−1)
i
, f (t−1)
j
(cid:17)
i
where α(t)
normalizes the weights so that they sum to one,
and ǫ is a small, positive constant used to avoid division by
zero. We set f (0)
i = fi, i.e., it is the original opinion reported
by expert i. There are some important points regarding equa-
tion (4). First, the distance between two opinions is always
non-negative. Hence, the constant ǫ ensures that every single
weight is strictly greater than 0 and strictly less than 1. Fur-
(t−1)
are, the higher
ther, the closer the opinions f
i
the resulting weight p(t)
0, the weight that each expert assigns to his own opinion is
always greater than or equal to the weights that he assigns to
his peers’ opinions.
i,j will be. Since D(cid:16)f (t−1)
(t−1)
and f
j
(cid:17) =
, f (t−1)
i
i
Now, we can redefine equation (2) so as to allow the experts
to update their weights based on the most recent opinions. Af-
ter t revisions, for t ≥ 1, we have that F(t) = P(t)F(t−1) =
P(t)P(t−1) . . . P(1)F(0), where each element of each matrix
P(k) is computed according to equation (4):
P(k) =
p(k)
1,1
p(k)
2,1
...
p(k)
n,1
p(k)
1,2
p(k)
2,2
...
p(k)
n,2
· · ·
· · ·
...
· · ·
p(k)
1,n
p(k)
2,n
...
p(k)
n,n
Pn
The opinion of each expert i at time t then becomes f
(t)
i =
. Algorithm 1 provides an algorithmic de-
j=1 p(t)
i,j
f (t−1)
j
scription of the proposed method.
In order to prove that all opinions converge towards a con-
sensual opinion when using the proposed method, consider
the following functions:
δ (U) =
1
2
max
i,j
ui,k − uj,k
γ(U) = min
i,j
min(ui,k, uj,k)
z
z
Xk=1
Xk=1
where 0 ≤ δ (U) , γ(U) ≤ 1, and U is a stochastic matrix.
δ (U) computes the maximum absolute difference between
two rows of a stochastic matrix U. Thus, when δ(cid:0)F(t)(cid:1) = 0,
all rows of F(t) are the same, i.e., a consensus is reached. We
use the following results in our proof [Paz, 1971]:
Algorithm 1 Algorithmic description of the proposed method
to find a consensual opinion.
Require: n probability vectors f (0)
Require: recalibration factor ǫ.
1: for t = 1 to ∞ do
for i = 1 to n do
2:
3:
1 , . . . , f (0)
n .
for j = 1 to n do
(t)
α
i
4:
p(t)
i,j =
end for
(t)
f
i =Pn
5:
6:
7:
8: end for
end for
ǫ+D(cid:16)f (t−1)
i
,f (t−1)
j
(cid:17)
j=1 p(t)
i,j
(t−1)
f
j
Proposition 1. Given two stochastic matrices U and V,
δ(UV) ≤ δ(U)δ(V).
Proposition 2. Given a stochastic matrix U, then δ(U) =
1 − γ(U).
Our main result is stated below.
Theorem 1. When t → ∞, f (t)
j.
i = f (t)
j
, for every expert i and
that F(t)
is the stochastic matrix repre-
Proof. Recall
senting the experts’ opinions after t revisions, and that
F(t) = P(t)F(t−1). Now, consider the following sequence:
(cid:0)δ(cid:0)F(0)(cid:1) , δ(cid:0)F(1)(cid:1) , ..., δ(cid:0)F(t)(cid:1)(cid:1). We are interested in the
behavior of this sequence when t → ∞. First, we show that
such a sequence is monotonically decreasing:
δ(cid:16)F(t)(cid:17) = δ(cid:16)P(t)F(t−1)(cid:17)
≤ δ(cid:16)P(t)(cid:17) δ(cid:16)F(t−1)(cid:17)
=(cid:16)1 − γ(cid:16)P(t)(cid:17)(cid:17) δ(cid:16)F(t−1)(cid:17)
≤ δ(cid:16)F(t−1)(cid:17)
The second and third lines follow, respectively, from
Propositions 1 and 2. Since δ (U) ≥ 0 for every stochastic
matrix U, then the above mentioned sequence is a bounded
decreasing sequence. Hence, we can apply the standard
monotone convergence theorem [Bartle and Sherbert, 2000]
and δ(cid:0)F(∞)(cid:1) = 0. Consequently, all rows of the stochastic
matrix F(∞) are the same.
In other words, a consensus is always reached under the
proposed method, and this does not depend on the initial re-
ported opinions. A straightforward corollary of Theorem 1 is
that all revised weights converge to the same value.
Corollary 1. When t → ∞, p(t)
j.
n , for every expert i and
i,j = 1
Hence, the proposed method works as if experts were
continuously exchanging information so that their individual
knowledge becomes group knowledge and all opinions are
equally weighted. Since we derive weights from the reported
opinions, we are then able to avoid some problems that might
arise when eliciting these weights directly, e.g., opinions do
not need to be disclosed to others in order for them to assign
weights, thus preserving privacy.
i,j
f (t−1)
j
j=1 p(t)
j=1 βj f (0)
an instance of the linear opinion pool. Recall that f (t)
The resulting consensual opinion can be represented as
i =
= · · · =
, where β = (β1, β2, . . . , βn) is a probability
vector that incorporates all the previous weights. Hence, an-
other interpretation of the proposed method is that experts
reach a consensus regarding the weights in equation (1).
= Pn
i,jPn
Pn
Pn
k=1 p(t−1)
j=1 p(t)
f (t−2)
k
j,k
j
3.3 Numerical Example
A numerical example may clarify the mechanics of the pro-
posed method. Consider three experts (n = 3) with the fol-
lowing opinions: f1 = (0.9, 0.1), f2 = (0.05, 0.95), and
f3 = (0.2, 0.8). According to (3), the initial distance be-
tween, say, f1 and f2 is:
D(f1, f2) =r (0.9 − 0.05)2 + (0.1 − 0.95)2
2
= 0.85
Similarly, we have that D(f1, f1) = 0 and D(f1, f3) = 0.7.
Using equation (4), we can then derive the weights that each
expert assigns to the reported opinions. Focusing on expert
1 at time t = 1 and setting ǫ = 0.01, we obtain p(1)
1,1 =
α(1)
1 /0.01, p(1)
1 /0.71. Since
these weights must sum to one, we have α(1)
1 ≈ 0.00975 and,
consequently, p(1)
1,3 ≈ 0.014.
Repeating the same procedure for all experts, we obtain the
matrix:
1,2 ≈ 0.011, and p(1)
1 /0.86, and p(1)
1,1 ≈ 0.975, p(1)
1,3 = α(1)
1,2 = α(1)
P(1) =" 0.975 0.011 0.014
0.013 0.058 0.929 #
0.011 0.931 0.058
1
1,j
P3
The updated opinion of expert 1 is then f (1)
j=1 p(1)
=
fj ≈ (0.8809, 0.1191). By repeating the above
procedure, when t → ∞, P(t) converges to a matrix where
all the elements are equal to 1/3. Moreover, all experts’ opin-
ions converge to the prediction (0.3175, 0.6825). An inter-
esting point to note is that the resulting prediction would be
(0.3833, 0.6167) if we had taken the average of the reported
opinions, i.e., expert 1, who has a very different opinion,
would have more influence on the aggregate prediction.
4 Consensus and Proper Scoring Rules
The major assumption of the proposed method is that experts
prefer opinions that are close to their own opinions. In this
section, we formally investigate the validity of this assump-
tion. We start by noting that in the absence of a well-chosen
incentive structure, the experts might indulge in game playing
which distorts their reported opinions. For example, experts
who have a reputation to protect might tend to produce fore-
casts near the most likely group consensus, whereas experts
who have a reputation to build might tend to overstate the
probabilities of outcomes they feel will be understated in a
possible consensus [Friedman, 1983].
Scoring rules are traditional devices used to promote hon-
esty in forecasting settings [Savage, 1971]. Formally, a scor-
ing rule is a real-valued function, R(fi, e), that provides a
score for the opinion fi upon observing the outcome θe.
Given that experts’ utility functions are linear with respect
to the range of the score used in conjunction with the scor-
ing rule,
the condition that R is strictly proper implies
that the opinion reported by each expert strictly maximizes
his expected utility if and only if he is honest. Formally,
i )] = fi, where Efi [R(·)] is the fi-expected
argmaxf ′
i , e). A well-known
value of R, Efi [R(f ′
strictly proper scoring rule is the quadratic scoring rule:
e=1 fi,e R(f ′
Efi [R(f ′
i
i )] = Pz
R(fi, e) = 2fi,e −
f 2
i,k
(5)
z
Xk=1
The proof that the quadratic scoring rule is indeed strictly
proper as well as some of its interesting properties can be
seen in the work by Selten [1998].
Proper scoring rules have been used as a tool to promote
truthfulness in a variety of domains, e.g., when sharing
rewards among a set of agents based on peer evaluations
[Carvalho and Larson, 2010;
Carvalho and Larson, 2011;
Carvalho and Larson, 2012],
to incentivize agents to ac-
curately estimate their own efforts to accomplish a task
[Bacon et al., 2012], in prediction markets [Hanson, 2003],
in weather forecasting [Gneiting and Raftery, 2007], etc.
4.1 Effective Scoring Rules
Scoring rules can also be classified based on monotonicity
properties. Consider a metric G that assigns to any pair of
opinions fi and fj a real number, which in turn can be seen as
the shortest distance between fi and fj. We say that a scoring
rule R is effective with respect to G if the following relation
holds for any opinions fi, fj, and fk [Friedman, 1983]:
G(fi, fj) < G(fi, fk) ⇐⇒ Efi [R(fj)] > Efi [R(fk)]
In words, each expert’s expected score can be seen as a mono-
tone decreasing function of the distance between his true
opinion and the reported one, i.e., experts still strictly max-
imize their expected scores by telling the truth, and the closer
a reported opinion is to the true opinion, the higher the ex-
pected score will be. The property of effectiveness is stronger
than strict properness, and it has been proposed as a desider-
atum for scoring rules for reasons of monotonicity in keeping
an expert close to his true opinion [Friedman, 1983].
By definition, a metric G must satisfy the following condi-
tions for any opinions fi, fj, and fk:
1. Positivity: G(fi, fj) ≥ 0, for all experts i, j, and
G(fi, fj) = 0 if and only if fi = fj;
2. Symmetry: G(fi, fj) = G(fj , fi);
3. Triangle Inequality: G(fi, fk) ≤ G(fi, fj) + G(fj , fk).
The root-mean-square deviation shown in (3) satisfies the
above conditions. However, equation (4), taken as a func-
tion of opinions, is not a true metric, e.g., symmetry does
not always hold. We adjust the original definition of effec-
tive scoring rules so as to consider weights instead of metrics.
We say that a scoring rule R is effective with respect to a set
of weights W = {p(t)
n,n} assigned at
any time t ≥ 1 if the following relation holds for any opinions
f (t−1)
i
2,1, . . . , p(t)
1,1, . . . , p(t)
, and f (t−1)
1,n, p(t)
, f (t−1)
:
j
k
j
k
)]
f (t−1)
i
f (t−1)
i
)] > E
[R(f (t−1)
[R(f (t−1)
i,k ⇐⇒ E
i,j < p(t)
p(t)
In words, each expert’s expected score can be seen as a
monotone increasing function of his assigned weights, i.e.,
the higher the weight one expert assigns to a peer’s opinion,
the greater the expected score of that expert would be if he
reported his peer’s opinion, and vice versa. We prove below
that the quadratic scoring rule shown in (5) is effective with
respect to a set of weights assigned according to (4).
Proposition 3. The quadratic scoring rule shown in (5)
to a set of weights W =
is effective with respect
{p(t)
n,n} assigned at any time t ≥ 1
according to equation (4).
2,1, . . . , p(t)
1,1, . . . , p(t)
1,n, p(t)
Proof. Given an opinion fj, we note that the fi-expected
value of the quadratic scoring rule in (5) can be written as:
z
Efi [R(fj)] =
=
=
=
z
z
fi,e R(fj, e)
Xe=1
Xe=1 2fj,efi,e − fi,e
Xe=1
Xe=1
Xx=1
Xe=1
2fj,efi,e −
2fj,efi,e −
z
z
z
fi,e
f 2
j,x
z
f 2
j,x!
Xx=1
Xx=1
f 2
j,x
z
Now, consider the weights assigned by expert i to the opin-
ions of experts j and k at time t ≥ 1 according to equation
(4). We have that p(t)
i,k if and only if:
i,j < p(t)
i,x − f (t−1)
i,x − f (t−1)
, f (t−1)
α(t)
i
i
i
z
k
ǫ + D(cid:16)f (t−1)
D(cid:16)f (t−1)
Xx=1(cid:16)f (t−1)
Xx=1
Xx=1
2f (t−1)
2f (t−1)
i,x
i,x
z
z
E
f
(t−1)
i
hR(cid:16)f (t−1)
k
α(t)
i
<
, f (t−1)
j
i
i
j
z
z
<
, f (t−1)
(cid:17)
(cid:17) < D(cid:16)f (t−1)
k,x (cid:17)2
ǫ + D(cid:16)f (t−1)
(cid:17)
Xx=1(cid:16)f (t−1)
k,y (cid:17)2
Xy=1(cid:16)f (t−1)
j,y (cid:17)2
Xy=1(cid:16)f (t−1)
hR(cid:16)f (t−1)
(cid:17)i > E
(t−1)
i
>
z
j
f
f (t−1)
k,x −
f (t−1)
j,x −
≡
≡
≡
≡
, f (t−1)
k
(cid:17)
j,x (cid:17)2
(cid:17)i
Proposition 3 implies that there is a correspondence be-
tween weights, assigned according to (4), and expected scores
from the quadratic scoring rule: the higher the weight one ex-
pert assigns to a peer’s opinion, the greater that expert’s ex-
pected score would be if he reported his peer’s opinion, and
vice versa. Hence, whenever experts are rational, i.e., when
they behave so as to maximize their expected scores, and their
opinions are rewarded using the quadratic scoring rule, then
the major assumption of the proposed method for finding a
consensual opinion, namely that experts prefer opinions that
are close to their own opinions, is formally valid. A straight-
forward corollary of Proposition 3 is that a positive affine
transformation of the quadratic scoring rule is still effective
with respect to a set of weights assigned according to (4).
Corollary 2. A positive affine transformation of
the
quadratic scoring rule R in (5), i.e., xR (fi, e) + y, for x > 0
and y ∈ ℜ, is effective with respect to a set of weights
W = {p(t)
n,n} assigned at any time
t ≥ 1 according to equation (4).
2,1, . . . , p(t)
1,1, . . . , p(t)
1,n, p(t)
5 Empirical Evaluation
In this section, we describe an experiment designed to test
the efficacy of the proposed method for finding a consensual
opinion. In the following subsections, we describe the dataset
used in our experiments, the metrics used to compare differ-
ent methods to aggregate opinions, and the obtained results.
5.1 Dataset
Our dataset was composed by 267 games (256 regular-season
games and 11 playoff games) from the National Football
League (NFL) held between September 8th, 2005 and Febru-
ary 5th, 2006. We obtained the opinions of 519 experts for
the NFL games from the ProbabilityFootball1 contest. The
contest was free to enter. Each expert was asked to report his
subjective probability that a team would win a game. Pre-
dictions had to be reported by noon on the day of the game.
Since the probability of a tie in NFL games is very low (less
than 1%), experts did not report the probability of such an
outcome. In particular, no ties occurred in our dataset.
Not all 519 registered experts reported their predictions for
every game. An expert who did not enter a prediction for a
game was removed from the opinion pool for that game. On
average, each game attracted approximately 432 experts, the
standard deviation being equal to 26.37. The minimum and
maximum number of experts were, respectively, 243 and 462.
Importantly, the contest rewarded the performance of experts
via a positive affine transformation of the quadratic scoring
rule, i.e., 100 − 400 × p2
l , where pl was the probability that
an expert assigned to the eventual losing team.
A positive affine transformation of a strictly proper scor-
ing rule is still strictly proper [Gneiting and Raftery, 2007].
The above scoring rule can be obtained by multiplying (5) by
200 and subtracting the result by 100. The resulting proper
scoring rule rewards bold predictions more when they are
right. Likewise, it penalizes bold predictions more when they
are wrong. For example, a prediction of 99% earns 99.96
1Available at http://probabilityfootball.com/2005/
points if the chosen team wins, and it loses 292.04 points if
the chosen team loses. On the other hand, a prediction of
51% earns 3.96 points if it is correct, and it loses 4.04 points
if it is wrong. A prediction of 50% neither gains nor loses
any points. The experts with highest accumulated scores won
prizes in the contest. The suggested strategy at the contest
website was “to make picks for each game that match, as
closely as possible, the probabilities that each team will win”.
We argue that this dataset is very suitable for our purposes
due to many reasons. First, the popularity of NFL games pro-
vides natural incentives for people to participate in the Proba-
bilityFootball contest. Furthermore, the intense media cover-
age and scrutiny of the strengths and weaknesses of the teams
and individual players provide useful information for the gen-
eral public. Hence, participants of the contest can be viewed
as knowledgeable regarding to the forecasting goal. Finally,
the fact that experts were rewarded via a positive affine trans-
formation of the quadratic scoring rule fits perfectly into the
theory developed in this work (see Corollary 2).
5.2 Metrics
We used two different metrics to assess the prediction power
of different aggregation methods.
Overall Accuracy
We say that a team is the predicted favorite for winning a
game when an aggregate prediction that this team will win
the game is greater than 0.5. Overall accuracy is then the per-
centage of games that predicted favorites have indeed won. A
polling method with higher overall accuracy is more accurate.
Absolute Error
Absolute error is the difference between a perfect prediction
(1 for the winning team) and the actual prediction. Thus, it is
just the probability assigned to the losing team (pl). An ag-
gregate prediction with lower absolute error is more accurate.
5.3 Experimental Results
For each game in our dataset, we aggregated the reported
opinions using three different linear opinion pools:
the
method proposed in Section 3, henceforth referred to as the
consensual method, with ǫ = 10−4; the traditional average
approach, where all the weights in (1) are equal to 1/n; and
the method proposed by Barlow et al. [1986], henceforth re-
ferred to as the BMS method. These authors proposed that the
weight assigned to expert i’s opinion should be wi =
i∗ ) ,
where c is a normalizing constant, I(fi, fi∗ ) is the Kullback-
Leibler divergence, and fi∗ achieves max{I(fi, fj) : 1 ≤ j ≤
n}, i.e., fi∗ is the most distant opinion from expert i’s opin-
ion. The BMS method produces indeterminate outputs when-
ever there are probability assessments equal to 0 or 1. Hence,
we recalibrated the reported opinions when using the BMS
method by replacing 0 and 1 by, respectively, 0.01 and 0.99.
Given the aggregated opinions, we calculated the perfor-
mance of each method according to the accuracy metrics pre-
viously described. Regarding the overall accuracy of each
method, the consensual method achieves the best perfor-
mance in this experiment with an overall accuracy of 69.29%.
The BMS and average methods achieve an overall accuracy
of, respectively, 68.54% and 67.42%.
c
I(fi,f
Table 1: Average absolute error of each method over the 267
games. Standard deviations are in parentheses.
Consensual
Average
BMS
0.4115 (0.1813)
0.4176 (0.1684)
0.4295 (0.1438)
Table 1 shows the average absolute error of each method
over the 267 games. The consensual method achieves the
best performance with an average absolute error of 0.4115.
We performed left-tailed Wilcoxon signed-rank tests in order
to investigate the statistical relevance of these results. The
resulting p-values are all extremely small(cid:0)< 10−4(cid:1), showing
that the results are indeed statistically significant.
Despite displaying a decent overall accuracy, the BMS
method has the worst performance according to the absolute
error metric. A clear drawback with this method is that it
only considers the distance to the most distant opinion when
assigning a weight to an opinion. Since our experimental de-
sign involves hundreds of experts, it is reasonable to expect at
least one of them to have a very different and wrong opinion.
The high number of experts should give an advantage to the
average method since biases of individual judgment can off-
set with each other when opinions are diverse, thus making
the aggregate prediction more accurate. However, the aver-
age method achieves the worst overall accuracy, and it per-
forms statistically worse than the consensual method when
measured under the absolute error metric. We believe this re-
sult happens because the average method ends up overweight-
ing extreme opinions when equally weighting all opinions.
On the other hand, under the consensual method, experts
put less weight on opinions far from their own opinions,
which implies that this method is generally less influenced
by extreme predictions as illustrated in Section 3.3.
6 Conclusion
We proposed a pooling method to aggregate expert opin-
ions. Intuitively, the proposed method works as if the experts
were continuously updating their opinions, where each up-
dated opinion takes the form of a linear opinion pool, and the
weight that each expert assigns to a peer’s opinion is inversely
related to the distance between their opinions. We proved that
this updating process leads to a consensus.
A different interpretation of the proposed method is that
experts reach a consensus regarding the weights of a linear
opinion pool. We showed that if rational experts are rewarded
using the quadratic scoring rule, then our major assumption,
namely that experts prefer opinions that are close to their own
opinions, follows naturally. To the best of our knowledge, this
is the first work linking the theory of proper scoring rules to
the seminal consensus theory proposed by DeGroot [1974].
Using real-world data, we compared the performance of
the proposed method with two other methods:
the tradi-
tional average approach and another distance-based aggrega-
tion method proposed by Barlow et al. [1986]. The results of
our experiment show that the proposed method outperforms
all the other methods when measured in terms of both overall
accuracy and absolute error.
[Gneiting and Raftery, 2007] Tilmann
and
Adrian E. Raftery. Strictly proper scoring rules, pre-
diction, and estimation.
the American
Statistical Association, 102(477):359–378, 2007.
Journal of
Gneiting
[Hanson, 2003] R. Hanson. Combinatorial information mar-
ket design. Information Systems Frontiers, 5(1):107–119,
2003.
[Jurca and Faltings, 2008] R. Jurca and B. Faltings. Incen-
tives for expressing opinions in online polls. In Proceed-
dings of the 2008 ACM Conference on Electronic Com-
merce, pages 119–128, 2008.
[Mosleh et al., 1988] A. Mosleh, V.M. Bier, and G. Aposto-
lakis. A critique of current practice for the use of expert
opinions in probabilistic risk assessment. Reliability Engi-
neering & System Safety, 20(1):63 – 85, 1988.
[Ng and Abramson, 1990] K.C. Ng and B. Abramson. Un-
IEEE Expert,
certainty management in expert systems.
5(2):29 – 48, april 1990.
[Paz, 1971] A. Paz. Introduction to Probabilistic Automata.
Academic Press, 1971.
[Savage, 1971] L.J. Savage. Elicitation of Personal Probabil-
ities and Expectations. Journal of the American Statistical
Association, 66(336):783–801, 1971.
[Selten, 1998] R. Selten. Axiomatic characterization of the
quadratic scoring rule. Experimental Economics, 1(1):43–
62, 1998.
References
[Bacon et al., 2012] David F. Bacon, Yiling Chen, Ian Kash,
David C. Parkes, Malvika Rao, and Manu Sridharan. Pre-
dicting Your Own Effort. In Proceedings of the 11th In-
ternational Conference on Autonomous Agents and Multi-
agent Systems, pages 695–702, 2012.
[Barlow et al., 1986] R. E. Barlow, R. W. Mensing, and
N. G. Smiriga. Combination of experts’ opinions based
on decision theory. In A. P. Basu, editor, Reliability and
quality control, pages 9–19. North-Holland, 1986.
[Bartle and Sherbert, 2000] R. G. Bartle and D. R. Sherbert.
Introduction to Real Analysis. Wiley, 3rd edition, 2000.
[Boutilier, 2012] C. Boutilier. Eliciting forecasts from self-
interested experts: Scoring rules for decision makers. In
Proceedings of the 11th International Conference on Au-
tonomous Agents and Multiagent Systems, pages 737–744,
2012.
[Carvalho and Larson, 2010] Arthur Carvalho and Kate Lar-
son. Sharing a reward based on peer evaluations.
In
Proceedings of the 9th International Conference on Au-
tonomous Agents and Multiagent Systems, pages 1455–
1456, 2010.
[Carvalho and Larson, 2011] Arthur Carvalho and Kate Lar-
son. A truth serum for sharing rewards. In Proceedings of
the 10th International Conference on Autonomous Agents
and Multiagent Systems, pages 635–642, 2011.
[Carvalho and Larson, 2012] Arthur Carvalho and Kate Lar-
son. Sharing rewards among strangers based on peer eval-
uations. Decision Analysis, 9(3):253–273, 2012.
[Chen and Pennock, 2010] Y. Chen and D.M. Pennock. De-
signing markets for prediction. AI Magazine, 31(4):42–52,
2010.
[Chevaleyre et al., 2007] Y. Chevaleyre, U. Endriss, J. Lang,
and N. Maudet. A short introduction to computational so-
cial choice. In Proceedings of the 33rd conference on Cur-
rent Trends in Theory and Practice of Computer Science,
pages 51–69, 2007.
[Clemen and Winkler, 1999] R.T. Clemen and R.L. Winkler.
Combining probability distributions from experts in risk
analysis. Risk Analysis, 19:187–203, 1999.
[Cooke, 1991] R.M. Cooke. Experts in uncertainty: opinion
and subjective probability in science. Oxford University
Press, 1991.
[DeGroot, 1974] M.H. DeGroot.
Reaching a consen-
Journal of the American Statistical Association,
sus.
69(345):118–121, 1974.
[Friedman, 1983] D. Friedman. Effective scoring rules for
probabilistic forecasts. Management Science, 29(4):447–
454, 1983.
[Genest and McConway, 1990] C. Genest and K.J. Mc-
Conway. Allocating the weights in the linear opinion pool.
Journal of Forecasting, 9(1):53–73, 1990.
|
1812.04985 | 1 | 1812 | 2018-12-12T15:17:03 | Implementing Argumentation-enabled Empathic Agents | [
"cs.MA"
] | In a previous publication, we introduced the core concepts of empathic agents as agents that use a combination of utility-based and rule-based approaches to resolve conflicts when interacting with other agents in their environment. In this work, we implement proof-of-concept prototypes of empathic agents with the multi-agent systems development framework Jason and apply argumentation theory to extend the previously introduced concepts to account for inconsistencies between the beliefs of different agents. We then analyze the feasibility of different admissible set-based argumentation semantics to resolve these inconsistencies. As a result of the analysis we identify the maximal ideal extension as the most feasible argumentation semantics for the problem in focus. | cs.MA | cs |
Implementing Argumentation-enabled Empathic Agents
Timotheus Kampik (cid:66)[0000−0002−6458−2252], Juan Carlos
Nieves[0000−0003−4072−8795], and Helena Lindgren[0000−0002−8430−4241]
Umea University, 901 87, Umea, Sweden
{tkampik,jcnieves,helena}@cs.umu.se
Abstract. In a previous publication, we introduced the core concepts of em-
pathic agents as agents that use a combination of utility-based and rule-based
approaches to resolve conflicts when interacting with other agents in their envi-
ronment. In this work, we implement proof-of-concept prototypes of empathic
agents with the multi-agent systems development framework Jason and apply
argumentation theory to extend the previously introduced concepts to account
for inconsistencies between the beliefs of different agents. We then analyze the
feasibility of different admissible set-based argumentation semantics to resolve
these inconsistencies. As a result of the analysis, we identify the maximal ideal
extension as the most feasible argumentation semantics for the problem in focus.
Keywords: Agent architectures · Agent-oriented software engineering · Argu-
mentation
1
Introduction
Complex information systems that act with high degrees of autonomy and automation
are ubiquitous in human society and shape day-to-day life. This leads to societal chal-
lenges of increasing frequency and impact, in particular when systems are primarily
optimized towards simple metrics like views and clicks and are then used by malevolent
actors for deceptive purposes, for example, to manipulate political opinions. Emerging
research highlights these challenges and suggests the development of new multi-agent
system concepts to address the problem ([17], [12]). In a recent publication, we outlined
the basic concepts of empathic agents [13], based on an established definition of empa-
thy as the ability to "simulate another's situated psychological states, while maintaining
clear self -- other differentiation" [8]1. Our empathic agent is utility-based but addition-
ally uses acceptability rules to avoid conflicts with agents that act in its environment.
Thereby, we attempt to emulate empathic human behavior, which considers rules and
societal norms, as well as the goals and intentions of others. We consider our agent a
research contribution to the "synergistic combination of modelling methods" for "au-
tonomous agents modelling other agents" as outlined as an open research problem in
the discussion of a survey by Albrecht and Stone [1].
This paper addresses the following research questions:
1 We based our empathic agent on a rationality-oriented definition of empathy, to avoid the
technical ambiguity definitions that focus on emotional empathy imply. A comprehensive dis-
cussion of definitions of empathy is beyond the scope of this work.
2
T. Kampik et al.
-- How can proof-of-concept prototypes of empathic agents be implemented with a
multi-agent systems development framework?
-- How can abstract argumentation (see: Dung [9]) be used to extend the agents to
be able to resolve scenarios in which the beliefs of different agents are initially
inconsistent?
-- Which admissible set-based argumentation semantics are most feasible for resolv-
ing these inconsistencies?
To illustrate our work in a context that can be considered of societal relevance, we
introduce the following example scenario: A persuader agent can select exactly one
item from a list of persuasive messages (ads) it can display to a mitigator agent. Based
on the impact the message will have on a particular end-user the mitigator represents,
the mitigator will either accept the message and allow the end-user to consume the per-
suader's service offering (and the message), or terminate its session with the persuader.
The persuader's messages are considered empathic agent actions; the mitigator does
not act in the sense of the empathic agent core framework (see: Section 2). If the mit-
igator accepts the action proposal of the persuader, the environment will pay different
utility rewards (or punishments) to persuader and mitigator. The goal of the persuader
is to select a message that is utility-optimal for itself, considering that the mitigator
has to accept it. We suggest that the scenario, albeit simple, reflects a potential type
of real-world use case for future empathic agents of greater maturity. Improving the
end-user experience of systems that are traditionally primarily optimized towards one
simple metric can be considered a societally beneficial use case for empathic agents:
(self-)regulation could motivate advertisement-financed application providers to either
implement both persuader and mitigator to create a more purposeful user experience,
or to open up their APIs to third-party clients that try to defend the interests of the
end-users.
The rest of this paper is organized as follows: First, we elaborate on the core con-
cepts of the empathic agent (Section 2), taking into consideration the belief-desire-
intention (BDI) architecture in which the concepts are to be implemented. Then, we
describe the implementation of the concepts with the multi-agent systems development
framework Jason [6] (Section 3). Next, we apply argumentation theory and extend the
core concepts and their implementation with Jason to handle inconsistencies between
the beliefs of different agents (Section 4). To clarify our choice of argumentation se-
mantics, we then analyze the applicability of different admissible set-based semantics
(Section 5). Subsequently, we examine other work related to argumentation with Jason,
briefly discuss the work in the context of existing negotiation approaches, discuss lim-
itations of the implemented agents, and outline possible steps towards more powerful
empathic agents (Section 6), before we conclude the paper (Section 7).
2 Empathic Agent Concepts
In this section, we elaborate on the empathic agent core concepts as initially sketched
out in [13].
A set of interacting empathic agents can be described as follows:
Implementing Argumentation-enabled Empathic Agents
3
, ..., ActAim
-- In an environment, n empathic agents {A0, ..., An} are acting.
-- Each agent Ai(0 ≤ i ≤ n) can execute a finite set of actions
}. All agents execute all actions simultaneously
ActsAi = {ActAi0
at one specific point in time; (inter)actions in continuous time or over a series of
discrete time steps are beyond the scope of the basic concept.
-- In its belief base, each agents has a utility function that assigns a numeric value to
all possible action combinations: uAi : ActsA0 × ... × ActsAn → {−∞, R, +∞}.
-- Each agent strives to determine a set of actions it should execute that is considered
acceptable by all agents. For this, an agent Ai determines the sets of actions that
maximize its own utility function (arg max uAi)2: If there is no own best set of ac-
tions actsami,k ∈ arg max uAi that matches a set of best actions of all of the other
agents, a conflict of interests exists. Hence, a conflict of interest can be determined
by the following boolean function:
c(arg max uA0, ..., arg max uAi, ..., arg max uAn ) =
true, if :
(cid:54) ∃ actsam0,l , ..., actsami,k , ..., actsamn,m
in arg max uA0 × ... × arg max uAi × ... × arg max uAn :
{actsam0,l ∩ ... ∩ actsami,k ∩ ... ∩ actsamn,m} (cid:54)= {}
else : f alse
-- In addition to its utility mapping, each agent Ai has a set of acceptability rules
∈ AccsAi is a boolean func-
AccsAi in its belief base. An acceptability rule AccAij
tion that takes as it input the to-be-executed actions and returns a boolean value3:
ActsA0 × ... × ActsAn → {true, f alse}.
-- If the conflict determination function does not determine a conflict, the agents can
go ahead and execute their optimal actions. If the conflict determination function
returns true (determines a conflict), each agent applies the acceptability rules to
check whether the actions it would ideally execute are in fact not acceptable or
whether they can potentially be executed nevertheless. Here, the agents can employ
different game-theoretical approaches (for example: minimizing the maximal loss)
for ensuring that the execution of conflicting, but acceptable action sets by multiple
agents does not lead to utility outcomes that are "bad" (in particular: worse than
executing the actions that maximize combined utility) for all parties4. If the actions
are not acceptable or are acceptable but chosen to not be executed, each agent can
either choose to maximize the shared utility of all agents (arg max(uA0 × ... ×
uAn)) or to select the next best action set and execute the conflict and acceptability
check for them. We refer to agent variants that employ the former approach as
lazy empathic agents, and agents that use the latter as full empathic agents. Lazy
2 Note that arg max uAi returns a set of sets.
3 Note that a single acceptability rule does not necessarily consider all to-be-executed actions,
i.e. it might ignore some of its input arguments.
4 As the simple examples we implement in this paper feature only one acting agent (a second
agent is merely approving or disapproving of the actions), such game-theoretical considera-
tions are beyond scope. Hence, we will not elaborate further on them.
4
T. Kampik et al.
empathic agents save computation time by not iterating through action sets that
do not optimize the ideally possible individual utility but provide better individual
utility than the action sets that maximize combined utility. However, they might not
find "good" solutions that are represented by action sets they do not cover5.
Listing 1 describes the empathic agent base algorithm (lazy variant) in a two-agent
scenario.
Listing 1: Empathic agent base algorithm
1. Determine actions that maximize own utility (arg max uA self ).
2. Check for conflict with actions that maximize utility of other agent (arg max uA other).
3. If conflicts appear (apply conflict determination function c):
Check if actions that maximize own utility are acceptable despite conflicts.
if conflicts are acceptable
(apply acceptability rules as function Actsself × Actsother → {true, f alse}):
else: Execute (arg max(uA self × uA other)) ∩ Actsself .
Execute (arg max uA self ) ∩ Actsself .
else: Execute (arg max uA self )∩Actsself (or (arg max(uA self ×uA other))∩Actsself ).
3
Implementation of an Empathic Agent with Jason
To provide proof-of-concept prototypes of empathic agents, we implemented a simple
empathic agent example scenario with Jason, a development framework that is well-
established in the multi-agent systems community6.
The agents implement the scenario type we explained in the introduction as follows:
1. The persuader agent has a set of utility mappings -- (utility value, unique action
name)-tuples -- in its belief base7:
revenue(3, "Show vodka ad").
revenue(1, "Show university ad").
...
The mappings above specify that the persuader can potentially receive three utility
units for showing a vodka advertisement, or one utility unit for showing a university
advertisement. The persuader communicates these utility mappings to the mitigator.
5 For now, we assume all agents in a given scenario have the same implementation variant.
Empathic agents that are capable to effectively interact with empathic agents of other imple-
mentation variants or with non-empathic agents are -- although interesting -- beyond scope.
6 The implementation of our empathic agents with Jason (including the Jason extension we
introduce below, as well as a technical report that documents the implementation) is available
at https://github.com/TimKam/empathic-jason.
7 The mappings are end-user specific. In a scenario with multiple end-users, the persuader would
have one set of mappings per user.
Implementing Argumentation-enabled Empathic Agents
5
2. The mitigator has its own utility mappings:
benefit(-100, "Show vodka ad").
benefit(10, "Show university ad").
...
The mitigator's utility mappings are labelled benefit, in contrast to the persuader's
revenue label. While this should reflect the difference in impact the actions have on
persuader and mitigator (or rather: on the end-user the mitigator is proxying), it is
important to note that we consider the utility of the two different mappings com-
parable. The mitigator responds to the persuader's announcement by sending back
its own mapping. In addition to the utility mappings, both persuader and mitigator
have a set of acceptability rules in their belief base8 :
acceptable(
"Show university ad",
"Show community college ad").
acceptable(
"Show community college ad",
"Show university ad").
...
The intended meaning the empathic agents infer from these rules is that both agents
agree that Show university ad is always acceptable if the preferred action of the mit-
igator is Show community college ad and vice versa. After sending the response, the
mitigator determines the action it thinks should be executed (the expected action),
using the algorithm as described in Section 2. In the current implementation, the
agents are lazy empathic agents, i.e. they will only consider the actions with maxi-
mal individual and maximum shared utility (as explained above) 9.
3. The persuader determines its action proposal in the same way and announces it to
the mitigator.
4. When the response is received and the own expected action is determined, the mit-
igator compares the received action with the determined action. If the actions are
identical, the mitigator sends its approval of the action proposal to the persuader. If
the actions are inconsistent, the mitigator sends a disapproval message and cancels
the session.
5. If the persuader receives the approval from the mitigator, it executes the agreed-
upon action.
Figure 1 shows a sequence diagram of the basic empathic agent example we imple-
mented with Jason. The message labels show the internal actions the agents use for
communicating with each other.
8 Note that in Jason terminology, acceptability rules are beliefs and not rules.
9 If at any step of the decision process, several actions could be picked because they provide
the same utility, the agents will always pick the first one in the corresponding list to reach a
deterministic result.
6
T. Kampik et al.
Persuader
Mitigator
.broadcast(tell, announce(utility, Utility))
.send(Source, tell, respond(OwnUtility))
Note: compute
expected action
while waiting for
reply
.broadcast(tell, propose(action, DeterminedAction))
alt
.send(Source, tell, approve(ReceivedAction))
If: action acceptable
.broadcast(tell, execute(action, AgreedAction))
.send(Source, tell, disapprove(ReceivedAction))
Else
Note:
cancel session
Fig. 1: Empathic agent interaction: sequence diagrams, basic example
4 Argumentation for Empathic Agents: Reaching Consensus in
Case of Inconsistent Beliefs
In real-world scenarios, agents will often have inconsistent beliefs because of an only
partially observable environment or subjective interpretations of environment informa-
tion. Then, further communication is necessary for the agents to reach consensus on
what actions to execute and to prevent mismatches between the different agents' deci-
sion making processes that lead to poor results for all involved agents. In this section,
we show how abstract argumentation can be used to synchronize inconsistent agents ac-
ceptability rules in the agents' belief bases. I.e., we improve the autonomy for the agent
from the perspective of moral philosophy in that we enhance ability of an empathic
agent to "impose the [...] moral law on [itself]" [7].
Abstract argumentation, as initially introduced by Dung formalizes a theory on the
exchange of arguments and the acceptability of arguments within a given context (ar-
gumentation framework) as a tuple (cid:104)A, R(cid:105) where A is a finite set of arguments (for
example: {a, b, c} and R denotes the arguments' binary attack relations (for example:
{(a, b), (b, c)}, i.e. a attacks b and b attacks c). Dung introduces different properties
of argument sets in argumentation frameworks, such as acceptability, admissibility,
and completeness [9]. Dung's initial theory has since been extended, for example by
Bench-Capon, who introduced value-based argumentation to account for values and
value preferences in argumentation [3]. In multi-agent systems research, argumentation
is a frequently applied concept, "both for automating individual agent reasoning, as well
Implementing Argumentation-enabled Empathic Agents
7
as multiagent interaction" [16]. It can be assumed that applying (abstract) argumenta-
tion to enhance the consensus-finding capabilities of the empathic agents is an approach
worth investigating.
In the previous example implementation, which primarily implements the empathic
agent core concepts, inconsistencies between acceptability rules are irreconcilable (and
would lead to the mitigator's rejection of the persuader's action proposal). To address
this limitation, we use an abstract argumentation approach based on Dung's argumen-
tation framework [9] and its maximal ideal extension definition [10]. We enable our
agents to launch attacks at acceptability rules. An attack is a tuple (a, b), with a being
the identifier of an action or of another attack and b being the identifier of the attack
itself. Each attack is added to a list of arguments, with the rule it attacks as its target.
Attacked rules are added as arguments without attack targets. Additional arguments
can be used to launch attacks on attacks. The argumentation framework is resolved by
determining its maximal ideal extension10. The instructions in Listing 2 describe the
algorithm the argumentation-enabled empathic agents use for belief synchronization.
The algorithm is executed if the agents initially cannot agree on a set of actions, which
would lead to a cancellation of the interaction using only the basic approach as intro-
duced above. We extended Jason to implement an argumentation function as an internal
action. The function calls the argumentation framework solver of the Tweety Libraries
for Logical Aspects of Artificial Intelligence and Knowledge Representation [18], which
we wrapped into a web service with a RESTful HTTP interface. In our example imple-
mentation, we extend our agents as follows:
1. The persuader starts with some acceptability rules in its belief base that are not
existent in the belief set of the mitigator:
acceptable("Show vodka ad", "Show university ad").
...
In the provided example, the persuader believes (for unknown reasons) that the
action Show vodka ad is acceptable even if the action Show university ad provides
greater utility to the mitigator.
2. When the mitigator assesses the action proposal, it detects that its own expected
action (Show university ad) is inconsistent with the persuader's expected action
(Show vodka ad). It sends a disapproval message to the persuader that includes an
attack on the acceptability rules of the Show vodka ad action:
attack("Show vodka ad", "Alcoholic").
3. The persuader then constructs an argumentation framework from acceptability rules
and attacks. With the help of the Jason argumentation extension, it determines all
preferred extensions of the argumentation framework and removes the arguments
(acceptability rules and attacks) that are not part of any preferred extension. Subse-
quently, the persuader updates its belief base accordingly, re-determines the action
proposal, and sends the proposal to the mitigator, who then re-evaluates it (and
ideally accepts it). Potentially, the persuader could also launch attacks on the miti-
gator's attacks and so on, until consensus is reached or the session is aborted11.
10 Note that we compare different argumentation semantics in Section 5.
11 However, the provided example code implements only one argumentation cycle.
8
T. Kampik et al.
Listing 2: Argumentation-enabled belief synchronization
1. Mitigator agent (After having received not acceptable action proposal): Request acceptabil-
ity rules from mitigator.
2. Persuader agent: Send acceptability rules.
3. Mitigator agent (Upon receiving acceptability rules):
(a) Consider acceptability rules as initial arguments {acc0, ..., accn} in argumentation
framework AF = (cid:104){acc0, ..., accn},{}(cid:105). Resolve AF by determining maximal ideal
extension.
(b) Decide whether:
new attacks should and can be launched,
or AF can be accepted as-is,
or session should be cancelled.
(c) If attacks should be launched:
Add additional set of arguments to AF that attack acceptability rules that are
inconsistent with own beliefs. Send updated AF to persuader (go to 4a).
else if AF can be accepted as-is:
Accept and resolve AF .
else:
Cancel session.
4. Persuader agent (Upon receiving AF ):
(a) Generate/update own argumentation framework AF . Resolve AF by determining max-
imal ideal extension.
(b) Decide whether:
new attacks should and can be launched,
or AF can be accepted as-is,
or session should be cancelled.
(c) If attacks should be launched:
Add additional set of arguments to AF that attack acceptability rules that are
inconsistent with own beliefs. Send updated AF to mitigator (go to 3c).
else if AF can be accepted as-is:
Accept and resolve AF .
else:
Cancel session.
Figure 2 shows a sequence diagram of the argumentation-capable empathic agent ex-
ample we implemented with Jason.
5 Argumentation Semantics Analysis
In this section, we explain why our argumentation-enabled empathic agents use ar-
gumentation semantics that select the maximal ideal extension. The purpose of the
required argumentation reasoner is to remove all arguments from an argumentation
framework that have been attacked by arguments that in turn have not been success-
fully attacked. I.e., the naive requirements are as follows:
Implementing Argumentation-enabled Empathic Agents
9
Persuader
Mitigator
.broadcast(tell, announce(utility, Utility))
.send(Source, tell, respond(OwnUtility))
Note: compute
expected action
while waiting for
reply
.broadcast(tell, propose(action, DeterminedAction))
alt
.send(Source, tell, approve(ReceivedAction))
If: action acceptable
.broadcast(tell, execute(action, AgreedAction))
Else
.send(Source, tell, disapprove(Rules, Attacks, ApprovedAction))
Note: cancel
session if
conversation does
not terminate
Fig. 2: Empathic agent interaction: sequence diagrams, argumentation example
1. Given a set of arguments S, the semantics must exclude all successfully attacked
arguments from S, and only these.
2. An argument in S is considered successfully attacked if it is attacked by any argu-
ment in S that is not successfully attacked itself.
According to Dung, a "preferred extension of an argumentation framework AF is a max-
imal (with respect to set inclusion) admissible set of AF". In this context, "a conflict-free
set of arguments S is admissible iff each argument in S is acceptable with respect to S",
given an "argument A ∈ AR is acceptable with respect to a set S of arguments iff for
each argument B ∈ AR: if B attacks A then B is attacked by S" [9]. As Dung's defini-
tion of preferred extensions is congruent with the stipulated requirements, it seems rea-
sonable for the empathic agent implementations to resolve argumentation frameworks
by using the corresponding argumentation semantics.
However, if there are arguments that attack each other directly or that form any
other circular structure with an even number of nodes, multiple preferred extensions
exist. For example, given the arguments {a, b} and the attacks {(a, b), (b, a)} both {a}
and {b} are preferred extensions.
This behavior implies ambiguity, which prevents the empathic agents from reaching
a deterministic resolution of their belief inconsistency. As this is to be avoided, we
introduce an additional requirement:
3. The semantics must determine exactly one set of arguments by excluding ambigu-
ous arguments.
10
T. Kampik et al.
Dung defines the grounded extension of an argumentation framework as "the least (with
respect to set inclusion) complete extension" and further stipulates that an "admissible
set S of arguments is called a complete extension iff each argument, which is acceptable
with respect to S, belongs to S" [9]. In the above example, the grounded extension is
{}. As determining the grounded extension avoids ambiguity, using grounded semantics
seems to be a reasonable approach at first glance. However, as Dung et al. show in a
later work, grounded semantics are sceptical beyond the stipulated requirements, in
that they exclude some arguments that are not successfully attacked [10]12. In the same
work, the authors define a set of arguments X as ideal "iff X is admissible and it is
contained in every preferred set of arguments" and show that the maximal ideal set is
"a proper subset of all preferred sets" [10]. Hence, we adopt maximal ideal semantics
for determining the set of acceptable arguments of our empathic agents.
The following two running examples highlight the relevance of the distinction be-
tween preferred, maximal ideal, and grounded semantics in the context of our empathic
agents.
Example 1 Example 1 highlights the advantage of grounded and maximal ideal seman-
tics over preferred and complete semantics for the use case in focus. The to-be-solved
empathic agent scenario is as follows:
-- Acceptability rules persuader: {(Show steak ad,∗)} (a1)13
-- Acceptability rules mitigator: {}.
To resolve the acceptability rule inconsistency, the agents exchange the following argu-
ments:
1. Mitigator: {(a1, Steak ad too unhealthy)} (b1)
2. Persuader: {(b1, Steak ad healthy enough)} (c1)
3. Mitigator: {(c1, b1)}
As a result, we construct the following argumentation framework:
AF1 = (cid:104){a1, b1, c1},{(b1, a1)(c1, b1), (b1, c1)}(cid:105)
Applying preferred, complete, grounded, and maximal ideal semantics to solve the
framework yields the following results:
-- Preferred: {b1},{a1, c1}
-- Complete: {b1},{a1, c1},{}
-- Grounded: {}
-- Maximal ideal: {}
semantics are overly strict.
12 In Example 2, we illustrate an empathic agent argumentation scenario, in which grounded
13 For the sake of simplicity, we use a wild card (∗) to denote that the acceptability rule applies
no matter which preference the mitigator agent has. Note that this syntax is not supported by
our implementation.
Implementing Argumentation-enabled Empathic Agents
11
As can be seen, applying preferred or complete semantics does not allow for a clear
decision on whether the acceptability rule should be discarded or not, whereas both
grounded and maximal ideal semantics are addressing this problem by discarding all
arguments that are ambiguous in complete or ideal semantics, respectively. Figure 3
contains a graphical overview of how the different argumentation semantics solve the
argumentation framework of the first example.
c1
b1
a1
Fig. 3: Example 1: argumentation semantics visualization: preferred (bold and grey, respectively).
Grounded and maximal ideal extensions are {}. The preferred argument sets are also complete.
{} is an additional complete set of arguments.
Example 2 Example 2 highlights the advantage of maximal ideal over grounded se-
mantics for the use case in focus. We start with the following empathic agent scenario:
-- Acceptability rules persuader: (Show steak ad,∗) (a2)
-- Acceptability rules mitigator: {}
To resolve the acceptability rule inconsistency, the agents exchange the following argu-
ments:
1. Mitigator: {(a2, Steak ad too unhealthy)} (b2)
2. Persuader: {(b2, Steak ad shows quality meat)} (c2)
3. Mitigator: {(c2, U ser pref ers sweets)} (d2)
4. Persuader: {(d2, T his pref erence does not matter)} (e2)
5. Mitigator: {(e2, d2)}
6. Persuader: {(c2, b2)}
From the scenario specification, the following argumentation framework can be con-
structed:
AF2 = (cid:104){a2, b2, c2, d2, e2},{(b2, a2)(b2, d2), (c2, b2), (d2, c2), (e2, d2)}(cid:105)
Applying preferred, complete, grounded, and maximal ideal semantics to solve the
framework yields the following results:
T. Kampik et al.
12
-- Preferred: {a2, c2, e2}
-- Complete: {a2, c2, e2}, {}
-- Grounded: {}
-- Maximal ideal: {a2, c2, e2}
As can be seen, applying grounded semantics removes acceptability rule a2, although
this rule is not successfully attacked according to the definition stipulated in our re-
quirements. In contrast, applying maximal ideal semantics does not remove a2. Figure 4
contains a graphical overview of how the different argumentation semantics solve the
argumentation framework of the second example.
e2
b2
a2
d2
c2
Fig. 4: Example 2: argumentation semantics visualization: preferred (grey) and maximal ideal
(dashed border). The grounded extension is {}. The preferred/maximal ideal argument set is also
complete. {} is an additional complete set of arguments.
6 Discussion
6.1 Argumentation and Jason
Argumentation in agent-oriented programming languages has been covered by previous
research. In particular, both Berariu [4] and Panisson et al. [14] describe the implemen-
tation of argumentation capabilities with Jason. In contrast to these works, we present
a service-oriented approach for argumentation support that provides one simple func-
tion to solve arguments and thus offers a higher level of abstraction for application
developers. For future work, we suggest considering advanced abstract argumentation
approaches like value-based argumentation [3] to account for agent preferences and
possibilistic argumentation [2] to account for uncertainty. Generally, we consider the
approach of writing service interface extensions for Jason as promising to integrate
with existing relevant technologies, for example to provide technology bridges to tools
and frameworks developed by the machine learning community.
Implementing Argumentation-enabled Empathic Agents
13
6.2 Alternative Negotiation Approaches
We consider the empathic agent approach as fundamentally different from other ne-
gotiation approaches (see for an overview: Fatima and Rahwan [11]) in that the em-
pathic agents only optimize their own utility if acceptability rules explicitly permit
doing so and strongly consider the utility of other agents in their environment. I.e.,
empathic agent application scenarios can be considered at the intersection of mixed-
motive and fully cooperative games. Comprehensive comparisons, in particular with
other argumentation-based negotiation approaches, as for example developed by Black
and Atkinson [5], are certainly relevant future work. For example, simulations that elab-
orate on the benefits and limitations of specific empathic agent implementations could
be developed and executed in scenarios similar to the one presented in the examples of
this paper.
6.3 Limitations
This paper explains how to implement basic empathic agents with Jason and how to
extend Jason to handle scenarios with initially inconsistent beliefs by applying argu-
mentation theory. However, the implemented agents are merely running examples to
illustrate the empathic agent concept and do not solve any practical problem. The fol-
lowing limitations need to be addressed to facilitate real-world applicability:
-- Although we extended Jason to better support the use case in focus, the current
implementation does not provide sufficient abstractions to conveniently implement
complex empathic agents. Moreover, our empathic agent implementation does not
make use of the full range of BDI-related concepts Jason offers.
-- While the agents can find consensus in case of simple belief inconsistencies, they do
not use scalable methods for handling complex scenarios with large state and action
spaces, and are required to explicitly exchange their preferences (utility mappings)
with each other, which is not possible in many application scenarios, in which the
utility mappings need to be learned.
-- So far our empathic agent only interacts with other empathic agents. This simplifi-
cation helps explain the key concepts with basic running examples and is permis-
sible in some possible real-world scenarios, in particular when designing empathic
agents that interact with agents of the same type in closed environments (physically
or virtually) that are not accessible to others. To interact with agents of other de-
signs, the empathic agent concept needs to be further extended, for example, to be
capable of dealing with malevolent agents.
-- The analysis of the argumentation approach is so far limited to the admissible set-
based semantics as introduced in the initial abstract argumentation paper by Dung
[9] that considers arguments as propositional atoms. The structure of the arguments
is not addressed.
6.4 Towards a Generic Empathic Agent
For future research, we suggest further extending Jason -- or to work with alternative pro-
gramming languages and frameworks -- to support more powerful abstractions for imple-
menting empathic agents, for example:
14
T. Kampik et al.
-- Develop engineering-oriented abstractions for empathic agents, possibly following
a previously introduced architecture proposal [13]. While further extending Jason
can facilitate scientific applications and re-use, providing a minimally invasive li-
brary or framework for a general-purpose programming language could help in-
troduce the concept to the software engineering mainstream. User interfaces could
be developed that facilitate the explainability of the empathic agents' decision pro-
cesses.
-- Leverage existing research to better address partial observability and subjectivity.
The aforementioned survey by Albrecht and Stone shows that a body of reinforce-
ment learning research from which can be drawn upon exists [1]. Yet, the general
problem of scaling intelligent agents in complex and uncertain environments re-
mains an open challenge. It would be interesting to investigate how this problem
affects the applicability of our empathic agent.
-- Design and implement empathic agents that can meaningfully interact with agents
of other architectures and ultimately with humans. As a starting point, one could
develop an empathic Jason agent that interacts with a deceptive Jason agent as
introduced by Panisson et al. [15].
-- Extend the analysis of a suitable argumentation semantics for the empathic agent
by considering advanced approaches, for example value-based argumentation as in-
troduced by Bench-Capon [3], and by identifying a feasible structure the arguments
can use internally.
7 Conclusion
In this paper, we show how to implement proof-of-concept prototypes of empathic
agents with an existing multi-agent development framework. Also, we devise an exten-
sion of the previously established empathic agent concept that employs argumentation
theory to account for inconsistent beliefs between agents. In our analysis of different
admissible set-based argumentation semantics, we have determined that maximal ideal
semantics are the most feasible for addressing the problem in focus, as they provide
both an unambiguous solution of a given argumentation framework and are not overly
sceptical (in contrast to grounded semantics) as to which arguments are acceptable.
However, the provided empathic agent implementations do not yet scale to solve
real-world problems. It is important to identify or devise more powerful methods to
handle subjectivity and partial observability, and to ultimately enable empathic agents
to meaningfully interact with agents of other architecture types. On the technology side,
an abstraction specifically for empathic agents is needed that forms a powerful frame-
work for implementing empathic agents. We suggest creating a Jason extension that
can facilitate the implementation of empathic agents with a framework that is famil-
iar to members of the academic multi-agent systems community, but to also consider
providing abstractions for programming frameworks that are popular among industry
software engineers and are not necessarily BDI-oriented or even agent-oriented per se.
Acknowledgements We thank the anonymous reviewers for their constructive feed-
back. This work was partially supported by the Wallenberg AI, Autonomous Systems
and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation.
Implementing Argumentation-enabled Empathic Agents
15
References
1. Albrecht, S.V., Stone, P.: Autonomous agents modelling other agents: A comprehensive sur-
vey and open problems. Artificial Intelligence 258, 66 -- 95 (May 2018)
2. Alsinet, T., Chesnevar, C.I., Godo, L., Simari, G.R.: A logic programming framework for
possibilistic argumentation: Formalization and logical properties. Fuzzy Sets and Systems
159(10), 1208 -- 1228 (2008)
3. Bench-Capon, T.J.: Persuasion in practical argument using value-based argumentation
frameworks. Journal of Logic and Computation 13(3), 429 -- 448 (2003)
4. Berariu, T.: An argumentation framework for BDI agents. In: Intelligent Distributed Com-
puting VII, pp. 343 -- 354. Springer (2014)
5. Black, E., Atkinson, K.: Choosing persuasive arguments for action. In: The 10th Interna-
tional Conference on Autonomous Agents and Multiagent Systems-Volume 3. pp. 905 -- 912.
International Foundation for Autonomous Agents and Multiagent Systems (2011)
6. Bordini, R.H., Hubner, J.F.: BDI agent programming in AgentSpeak using Jason. In: Inter-
national Workshop on Computational Logic in Multi-Agent Systems. pp. 143 -- 164. Springer
(2005)
7. Christman, J.: Autonomy in moral and political philosophy. In: Zalta, E.N. (ed.) The Stanford
Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University, spring 2018
edn. (2018)
8. Coplan, A.: Will the real empathy please stand up? a case for a narrow conceptualization.
The Southern Journal of Philosophy 49(s1), 40 -- 65 (2011)
9. Dung, P.M.: On the acceptability of arguments and its fundamental role in nonmonotonic
reasoning, logic programming and n-person games. Artificial Intelligence 77(2), 321 -- 357
(1995)
10. Dung, P.M., Mancarella, P., Toni, F.: Computing ideal sceptical argumentation. Artificial
Intelligence 171(10-15), 642 -- 674 (2007)
11. Fatima, S., Rahwan, I.: Negotiation and bargaining. In: Weiss, G. (ed.) Multiagent Systems,
Second Edition, chap. 4, pp. 143 -- 176. MIT Press, Cambridge (2013)
12. Kampik, T., Nieves, J.C., Lindgren, H.: Coercion and deception in persuasive technologies.
In: 20th International TRUST Workshop (Jul 2018)
13. Kampik, T., Nieves, J.C., Lindgren, H.: Towards empathic autonomous agents. In: 6th In-
ternational Workshop on Engineering Multi-Agent Systems (EMAS 2018). Stockholm (Jul
2018)
14. Panisson, A.R., Meneguzzi, F., Vieira, R., Bordini, R.H.: An approach for argumentation-
based reasoning using defeasible logic in multi-agent programming languages. In: 11th In-
ternational Workshop on Argumentation in Multiagent Systems (2014)
15. Panisson, A.R., Sarkadi, S., McBurney, P., Parson, S., Bordini, R.H.: Lies, Bullshit, and De-
ception in Agent-Oriented Programming Languages. In: 20th International TRUST Work-
shop. Stockholm (Jul 2018)
16. Rahwan, I.: Argumentation among agents. In: Weiss, G. (ed.) Multiagent Systems, Second
Edition, chap. 5, pp. 177 -- 210. MIT Press, Cambridge (2013)
17. Sen, S., Crawford, C., Rahaman, Z., Osman, Y.: Agents for Social (Media) Change. In: Pro-
ceedings of the 17th International Conference on Autonomous Agents and Multiagent Sys-
tems (AAMAS 2018). Stockholm (Jul 2018)
18. Thimm, M.: Tweety - A Comprehensive Collection of Java Libraries for Logical Aspects
of Artificial Intelligence and Knowledge Representation. In: Proceedings of the 14th Inter-
national Conference on Principles of Knowledge Representation and Reasoning (KR'14).
Vienna, Austria (Jul 2014)
|
0902.2751 | 4 | 0902 | 2009-03-01T10:35:34 | Object Classification by means of Multi-Feature Concept Learning in a Multi Expert-Agent System | [
"cs.MA",
"cs.LG"
] | Classification of some objects in classes of concepts is an essential and even breathtaking task in many applications. A solution is discussed here based on Multi-Agent systems. A kernel of some expert agents in several classes is to consult a central agent decide among the classification problem of a certain object. This kernel is moderated with the center agent, trying to manage the querying agents for any decision problem by means of a data-header like feature set. Agents have cooperation among concepts related to the classes of this classification decision-making; and may affect on each others' results on a certain query object in a multi-agent learning approach. This leads to an online feature learning via the consulting trend. The performance is discussed to be much better in comparison to some other prior trends while system's message passing overload is decreased to less agents and the expertism helps the performance and operability of system win the comparison. | cs.MA | cs | Object Classification by means of
Multi-Feature Concept Learning in a Multi Expert-
Nima Mirbakhsh
CS Department, IASBS
[email protected]
Agent System
Arman Didandeh
CS Department, IASBS
[email protected]
Mohsen Afsharchi
CS Department, IASBS
[email protected]
Abstract -- Classification of some objects in classes of related
concepts is an essential and even breathtaking task in many
applications. A solution is discussed here based on Multi-Agent
systems. A kernel of some expert agents in several classes is to
consult a central agent decide among the classification problem
of a certain object. This kernel is moderated with the center
agent, trying to manage the querying agents for any decision
problem by means of a data-header like feature set. Agents have
cooperation among concepts related to the classes of this
classification decision-making; and may affect on each others'
results on a certain query object in a multi-agent learning
approach. This leads to an online feature learning via the
consulting trend. The performance is discussed to be much better
in comparison to some other prior trends while system's message
passing overload id decreased to less agents and the expertism
helps the performance and operability of system win the
comparison.
Keywords-component; Agent, Feature, Concept Learning,
Expertism
I.
INTRODUCTION
In this part, we represent some definitions to clarify the
expressions we are going to use in the following parts of this
paper:
AGENTS, MAS AND EXPERTISM
Fortunately or unfortunately,
there are very several
definitions for an Agent like [8] or [9]. To meet the term in a
more general look, an agent is a tool that carries out some task
or tasks on behalf of a human. Intelligent agents have
additional domain knowledge that enables them to carry out
their tasks even when the parameters of the task change or
when unexpected situations arise. Many intelligent agents are
able to learn, from their own performance, from other agents,
from the user, or from the environment in which they are
situated. An intelligent Agent has properties as: Autonomy,
Ability to Learn, Cooperation &many other ones [4]. Giving a
more formal definition, an agent is defined in [5].
A Multi Agent System is a system consisting of several,
not-the-same-type agents which are in collaboration with
eachother to do a complicated task, usually with some
constraints like time and recourses. The MAS is studied for
many reasons and by so many pioneers of AI, esp. for Agent-
Studying, Decision Making and also Expert Systems. The most
important obstacle to be met in perusal of MAS is the trade-off
between performance-based and goal-based nature of agents
gathering together to model the MAS, where usually a trade-off
seems to be undeniable.
"The goal of Multi Agent Systems' research is to find
methods that allow us to build complex systems composed of
autonomous agents who, while operating on local knowledge
and possessing only limited abilities, are nonetheless capable of
enacting the desired global behaviors." [6]
Regarding an Expert-Agent, we usually mean an agent
which is able to decide about a specific topic so much better
than some others agents. This ability may be gained with agent
collaboration in a MAS, with learning among some dataset or
using any other way of decision optimization. If we have a
MAS in which agent are to be experts in their own area of
deciding problems, then we will have a Multi Expert-Agent
System.
CONCEPTS, OBJECTS AND FEATURES
As we look around into the world we live, it is available to
understand several things and differentiate them from others.
This task of differentiation is because of that these being
compared things are under some sort of experiment to show
their specific characteristics as a react in response to some
query-like act. These responses are generated due to the
characteristics which are known under the set of "Features". It
is clear that the sets we are mentioning here are not disjoint, but
have several in common among eachother.
The THINGS which are carrying these feature sets as a
collection of properties are called "Objects". Objects are
exactly what we see in the world around. But what is making
the difference between these objects to generate the responses?
Several objects may carry several feature sets which have
nothing in common or may share some or even all features in
their feature sets. Besides the set of all features that can
introduce the objects to us, one more important matter to be
remi
takin
is de
and
amo
inded is the v
ng in compari
efined with the
also with the
ng the set of a
values that th
son to other o
e set of all fea
specific valu
all values they
hese features
objects in com
atures carrying
es that these f
y can be instan
can take and
mmon. So an o
g in its feature
features are ta
ntiated with.
d are
object
e set;
aking
B
But what is ha
at le
ast very likely
course that the
Of c
also
they have f
hother. When
each
ng looked, the
bein
like,
we are menti
esentation ca
repre
colle
ection of all
h most proxim
with
share
e in their feat
take-, the mor
can t
It is
s clear for su
mbers of a C
mem
stand
dpoint of Obje
appening when
y to be the sam
ese objects are
feature values
we have suc
ey remind us
ioning a more
alled
the "C
objects that a
mity. The mo
ture sets -- an
re we call them
ure that the le
Concept-famil
ect Orientation
n some objects
me when they
e near in the s
s that make t
ch groups of
of some othe
e abstract elem
Concept". C
are usually ca
ore these coll
nd also in the
m an integrate
evel of proxi
y can be m
n.
s seem the sam
are being stud
set of features
them come a
objects that w
er objects of
ment of knowl
Concepts are
arrying featur
ections of ob
set of values
ed unique con
imity between
modeled under
me or
died?
s and
about
while
their
ledge
the
re set
bjects
they
ncept.
n the
r the
CONCEPT
LEARNING
A
A very import
Agen
nt systems is
deal with each
to d
w nothing or
know
ut. This mak
abou
ning. Several
learn
ning where so
learn
set o
of peer agent
ther. A good
toge
be m
met in [1].
munication in M
tant hindrance
always put on
h other about
even are not
kes us deal
attempts was
ome agent lear
ts into the sy
example of "
e about comm
n where some
t some concep
t sure to have
with the pro
made on the p
rns some conc
ystem they ar
"New Concep
agents are wi
pts of which
e unanimous i
oblem of con
process of con
cept from one
re collaboratin
pt Learning" c
Multi
illing
they
ideas
ncept
ncept
e or a
ng in
could
We here addre
W
of which w
roach. In [3],
-unanimous co
oncept fluctu
phery objects.
analogy of it w
ess the attemp
we believe is
Afsharchi et
oncepts for wh
uate between
. We will later
with [3] in bas
idea
appr
non-
a co
perip
the a
ts made in [3]
having a lot
al. brings to
hich they allow
two bounda
r talk more on
e ideas.
] while it has s
t common to
o front the ide
w the definitio
aries of core
n our approach
some
o our
ea of
on of
and
h and
II. APPROA
ACH OVERVIEW
W
A
Assume that
dinality of M
sisting of an ob
ome of these m
s is so similar
ch we mention
cts are assum
ed system for
acted from kn
port of initial d
Naïve-Bayes
we have som
M, let's say {
bject to decide
main classes t
to many appl
n here [2] and
med to be text
the task of
nown Machine
datasets. [10]
classifier in a
me set of ma
C1, C2, …,
e about, could
to be classifie
lication-based
d [10]. In both
in a web pag
classification
e Learning me
has this task d
hierarchical s
ain classes w
CM}. Any q
d go mapped to
ed into our sys
d previous task
mentioned w
ge. [2] uses a
and the rule
ethods with a
done with the
structure.
with a
uery,
o one
stem.
ks, of
works,
rule-
s are
good
help
card
cons
or so
This
whic
obje
base
extra
supp
of a
H
Here we intro
M} which each
AgM
ses. We try t
class
decid
ding over on
em, annotatin
syste
ording to the p
acco
conc
cepts for incom
oduce M agen
is willing to b
to make any
ne and only o
ng the known
probability of
ming queries.
nts, let to nam
be an expert o
Agent, Agi to
one of the ma
features of t
f those feature
These agents
me {Ag1, Ag2,
on one of the
o be an expe
ain classes o
their main cla
es representing
have collabor
. . .,
main
ert in
f the
asses,
g the
ration
unde
with
thes
fact
any
over
er the name o
h eachother. It
e main classe
is not an obs
agent over its
r such hindran
of MAS and
t's somehow
es may have s
stacle to our a
s own class is
nces.
have the rela
clear that at m
something in
approach whil
important her
ation called "P
most attention
common. Bu
le the expertis
re, and is a triu
Peer"
n, yet
ut this
sm of
umph
T
To moderate t
ong agents/cla
amo
nt, called the C
agen
h any agent r
with
em maintains
syste
the s
system at any
ward the quer
forw
patching is hel
disp
gree of confi
"deg
Cen
nterAgent thin
n. It then inc
upon
ing job on t
mixi
fidence and g
conf
uld the incom
wou
outp
putted as a se
ure 1 shows h
Figu
from
m outside.
the action of
asses, we utili
CenterAgent a
epresenting a
an abstract in
y time, and du
ry to a subset
ld under some
idence" whic
nks the object
epts the resul
these results
gets a percepti
ming object a
et of concepts
how this syst
dispatching th
ize a more ge
and has the rel
a main class.
nformation of
ue to this infor
t of all its ag
e value of pro
ch shows to
is for the pa
lts of all its p
with regard
ion on in whi
a consonant b
s with the mo
tem works wh
he incoming q
eneral and ab
lation called "
This entity o
f all other agen
rmation, decid
gents. The ta
obability, calle
what extents
air agent to d
pair agents, do
to the degre
ch class or cl
be. This resu
ost probable p
hen you look
query
stract
"Pair"
of our
nts of
des to
sk of
ed the
s our
decide
oes a
ee of
lasses
ult is
posse.
k at it
Figure 1.
i
is the big picture
of the system bei
ing discussed
A
Another impo
nt, having a se
agen
ble to revise
is ab
conc
cept of that m
ries. These qu
quer
data
aset to learn th
ortant matter
et of features f
the probabili
main class in t
ueries are also
he agent on fea
in our appr
for its corresp
ities of featur
time with the
o utilized as a
atures related t
roach is that
ponding main c
res expressing
e help of inco
a shape of sui
to its main cla
each
class,
g the
oming
itable
ass.
A
A talk on mo
syste
ems operabilit
ore details com
ty and perform
mes next, and
mance after.
d the discussio
on of
III. MOR
N
RE DISCUSSION
A. B
Before Runnin
ng the System
H
Here we prov
ume that our
se main class
erts before th
vide the proced
system is equ
ses are gathere
e system initi
dure of our sy
uipped with s
ed and approv
ialization, and
ystem. At firs
some base cla
ved by a grou
d then are fed
st, we
asses.
up of
d into
assu
The
expe
the k
abstr
Supe
conc
child
abstr
more
hiera
kernel of syst
ract concept w
erConcept is
cepts in the di
dren of the ma
ract agent lik
e concrete ag
archy is shown
tem. Each ma
which is name
the parent
scussed criter
ain classes. Th
ke CenterAgen
ents like its p
n in Figure 2.
ain class repre
ed here as a S
of a group
ria, which thes
his is as we ar
nt holding the
pair decision-e
esents a some
uperConcept.
of more det
se concepts ar
re talking abo
e control of s
expert agents.
ehow
This
tailed
re the
ut an
some
. The
Figu
ure 2. shows the
e hierarchy of con
to
o the system mod
ncepts and the rela
del being discusse
ation of this hiera
ed
archy
is a
E
Each of these
r own, are m
their
cept
conc
racteristics of t
char
here
named as f
archical repre
hiera
node in the lo
any
he SuperConc
of th
repre
esenting the p
abstract main
modeled with
group of o
their own in g
features and a
esentation we
ower levels of
cept, holding
arent SuperCo
n classes holdi
a group of
objects havin
general. These
are held in a
are using he
f the hierarchy
a subset of
ing the Conce
features, wh
ng
the com
characteristic
a feature set.
ere lets us pro
y as a SubCon
the base fea
ept of
hile a
mmon
cs are
The
ovide
ncept
atures
oncept.
T
These base fea
system befor
our
runn
are r
of ou
poin
ning the system
reserved into t
ur CenterAge
nting the conce
atures we are
re its initializ
m. A set of al
the CenterAge
ent's base feat
epts of that cla
discussing her
zation with he
l base feature
ent and some
tures into any
ass.
re are also fed
elp of the ex
s for all M cla
M disjoint su
main class a
d into
xperts
asses
ubsets
agent,
B
Before discus
sing about th
essen
ntial to expla
ain more on s
ne two thresh
defin
holds, after W
babilities of ea
prob
ach class's fea
y feature inclu
Any
uding the bas
m experts at th
from
he beginning o
a reg
gion above the
e first threshol
on as [1]. Any
regio
y feature betw
yet another re
in y
egion, named
w the second
belo
d threshold is
regio
on features ha
ave the most p
er a main cl
unde
lass, while b
formance of c
perfo
classification
ures have a lo
featu
ower probabili
authorized to
are a
either enter t
featu
ure set for the
e D-region. T
may not refl
who
ect any prope
reserved in the
are r
e agent featur
track
k -- maybe the
dispatching w
e- and to be a
time
allowed to ye
impa
acting on gen
nerating the re
.
nt's feature set
agen
he starting of
some idea we
Williams [1]
atures, being h
se features o
of system runn
ld. We name th
ween the two th
as the M re
placed into a
probability to
eing revised
to the system
ity showing th
the K-region
The D-region
erty of the cla
re collection i
was false ove
et come back
esult. Figure
f the system,
e are utilizing
for governing
held in each a
f subjective i
ning, is placed
his region as t
hresholds is pl
gion. Any fe
D-region. Th
show the con
to give the
m. The M-re
he main class
or even leave
features are t
ass they are in
in order to be
er a long perio
to the region
3 shows one
it is
. We
g the
agent.
ideas
d into
the K
laced
eature
he K-
cepts
best
egion
s, but
e the
those
n, but
kept
od of
ns of
e pair
Fig
gure 3. Figure (3
including the K
3) is the represen
K, M and D-regio
ntation of an agen
n features with th
nt's feature collect
heir probabilities
tion,
T
This all mean
K-re
egion features
a set of M-reg
and
M threshold
and
erating and se
gene
patched query.
disp
of al
ll classes are g
ide about the
deci
This
s is shown in F
ns that each ag
s with upper
gion features w
ds. These feat
ending results
Yet another s
gathered toget
dispatching ta
Figure 4 below
gent of a main
threshold K-
with probabili
ture sets are
back to the C
set of all these
ther in the Cen
ask of any late
w.
n class has a s
-region proba
ities between t
the main hel
CenterAgent o
e K-region fea
nterAgent to h
er incoming q
set of
ability
the K
lp on
on its
atures
help it
query.
Figu
ure 4. show the
relation between
feature c
n CenterAgent and
collection
d its pair agents o
on the
B. R
Running the sy
ystem
W
When a newl
cess is done
ect, which are
hen forwarded
. The CenterA
ection of tags
agents -- whic
ly come obje
to extract th
named as tag
d to CenterAge
Agent, now re
s and based o
ch is the base f
ect arises to o
he collection
g collection. T
ent as the coll
eads the queri
on its abstract
feature set dis
our system, a
of features in
This queried o
lection of extr
ied object usin
t knowledge o
cussed above
a pre-
n the
object
racted
ng its
of its
to be
proc
obje
is th
tags
colle
pair
held
to di
tags
disp
degr
any
our
pair
prob
sear
prob
This
Cen
and
of c
vect
inco
shou
d in the kernel
ispatch this qu
to one or s
patching task i
ree our Center
pair agent. T
CenterAgent o
agent Agi n
babilities of e
rches for ma
bability and pa
s package is
nterAgent sorts
generates a v
onfidence for
tor consists o
oming queried
uld be what is
l of this syste
uery in a stand
some pair ag
is under a deg
rAgent thinks
These pair age
on deciding ab
now viewing
expertism, say
apping featur
ackages this a
sent back t
s these packag
vector of outgo
r any result co
of main clas
d object repres
seen as the re
em- and their
dard view incl
gents. As onc
gree of confid
the dispatchin
ents would be
bout the classi
g over its se
ay {<fj, pj>
res to some
s a result of th
towards the
ges of results f
oing output lo
oming from an
ses and the
senting any of
esult of the sys
expertism, de
luding its extr
ce mentioned,
dence which i
ng is done trul
as a consulta
ification task.
t of features
fj Ki !
concept wi
he incoming q
CenterAgent.
from its pair a
ooking at its d
ny pair agent.
likelihood o
f them. This v
stem but its us
ecides
racted
, this
is the
ly for
ant to
Each
s and
Mi},
ith a
query.
Our
agents
egree
This
f the
vector
ser.
C. L
Learning more
e on each age
ent's feature pr
robabilities
O
Our object-cla
h the help of t
with
ch it stores so
whic
subconcepts o
the
ures on the h
featu
expe
ert agent m
terAgent. The
Cent
ures, here tags
featu
ection or not.
colle
agent to indiv
any
their corresp
and
set. Wheneve
datas
ed to an agen
pass
query and the
the q
expe
ert agent -- in
esponding pr
corre
mory, named t
mem
t recently inco
most
wn in Figure
show
featu
ure for describ
he queries bein
of th
and
over for Agi
bability that th
prob
cepts is releva
conc
featu
ure on decisio
ut it repeating
abou
assification m
the collection
me K, M and
of its own m
hierarchy of p
meets some
multi-agent sys
of features an
d D region fea
main class. W
probabilities,
dispatched q
ese objects, f
for sure inclu
s, which may
be a member
This may lea
ad to have a
vidually learn
and revise its
ponding prob
babilities due
er a new obje
ect consisting
nt, it is going
to keep track
eir density ov
ver time. For
addition to
its collection
obabilities- is
s equipped w
the "Time-Inte
erval Memory
oming feature
es. This time-i
5. Here the a
agent can dis
bing its own c
concepts over
ng passed to
it. If a feature
and it is in t
the base featu
his feature is
more helpful
ant. Unlike it i
is cool to low
on making ta
asks which w
s.
in Ci concepts
stem does its
nd probabilitie
atures, represen
While having t
any willing-to
queries from
ude some se
of any main
feature datase
s feature colle
to the incom
of several ta
k of the tags in
this purpose,
n of features
with a time-b
y" to keep trac
interval memo
scover any ne
the whole fea
e fj is coming
ures of it, then
l for describin
wer the impact
we are not so
task
es, in
nting
these
o-be-
the
veral
class
et for
ction
ming
ags is
nside
, any
and
based
ck of
ory is
eeded
atures
over
n the
ng Ci
t of a
sure
Fig
gure 5. Figure (5
feature c
5) shows the struc
collection besides
cture of any pair
s its time-interval
agent on holding
l memory
g the
D. L
Learning best f
features by an
ny agent
T
The above to
nts to collabor
agen
featu
ures with the
cess of object
proc
h its peer age
with
MAS
S by consultin
w deciding feat
now
ld tag collect
rate within an
e help of the
t classification
ents, trying to
ng on whethe
tures.
tion is also a
nd do some e
e peer agents
n, any agent
o improve the
er they know
a good datase
extra learning
s. Along with
has a correl
e expertism o
anything abou
et for
over
h the
lation
f the
ut its
are parts
F
For this purpo
keep
ping track of m
tags
Cent
help
whic
begi
its p
cons
its m
terAgent, ther
ful to Agi in i
ch could be a
nning by the
peer agents.
sults on the n
memory time s
se, again we u
most recently i
of a query
re may come
its future deci
base feature o
expert team.
Simply talki
on-familiar fe
slice over and
utilize the tim
incoming feat
y which
is
some feature
ision making p
of Ci but is no
Here Agi nee
ng about thi
eatures which
over.
me-interval mem
tures. Almost t
dispatched
es which coul
processes, i.e.
ot mentioned a
eds to consult
is procedure,
are repeated
mory
these
from
ld be
. tags
at the
with
Agi
over
If this process
I
wly-come featu
new
there
e has to be so
ure probabiliti
featu
of consulting
ure should be
ome raise/redu
ies.
g leads to this p
added to Agi
uce modificati
perception tha
i's feature set,
ion on peer ag
at this
, then
gents'
L
Let us suppose
ts time-interva
in it
featu
ure set. Now A
ropriate proba
appr
on. As it fin
regio
bability of this
prob
one over and o
is do
the M
M and K regi
cedure is call
proc
nd out that th
foun
on features an
regio
is re
epresenting, i
in and put it
agai
blem while w
prob
ses to be disjo
clas
he CenterAgen
in th
som
me other agents
tacle, we prov
obst
e that Agi has
al memory in
Agi adds this
ability, let to b
nds the featur
s feature to som
over until this
ons, which is
led the raise
his feature ha
nd could be a b
it would like
up to its K
e have govern
oint subsets of
nt), and it is p
s' K-region fe
ide an approac
s found some n
nteresting for
feature to its
be the floor pr
re more intere
me extent, let
s feature reach
the first defin
procedure. N
s a lot in com
base feature fo
to add the f
K-region. But
ned the K-reg
f all-features s
possible for fj
eature sets. In
ch as below:
new feature, s
itself to add
feature set wi
robability of it
esting, it add
us suppose .
hes to the bord
ned threshold.
Now that Ag
mmon with it
for the main cl
feature proba
this may cau
gion features o
set (being rese
to be a memb
order to solve
say fj,
to its
ith an
ts M-
ds the
. This
der of
This
i has
ts K-
lass it
ability
use a
of all
erved
ber of
e this
"
"Whenever an
ature fj suitab
a fea
with
h its peer agen
egion. If not s
K-re
ropriate proba
appr
on. But in the
regio
peer
r agents, let A
nts having it i
agen
pose . This p
supp
w this fact tha
know
procedure co
fall
abov
ve, now."
n agent Agi re
ble to enter its
nts on whether
so, add fj to t
ability, let to
e case of hav
Agi reduce th
in their K-reg
procedure is ca
at no peer agen
omprises the c
eaches to the p
s K-region, le
r they have suc
the feature set
be the floor
ing the contra
he probability
gion to some e
alled the fall p
nt has fj in its
contradiction.
point where it
t the agent co
ch a feature in
ts of Ci with
probability o
adiction withi
of fj for the
extent, let us
procedure. Le
K-region whe
Let Agi add
finds
onsult
n their
some
of K-
in the
peer
again
et Agi
en the
fj as
T
The more inte
all peer agen
ses the -value
can have yet
s by adding fj
eresting this fe
nts for their a
e probability r
again disjoin
to Agi's class
eature is for Ag
advisory sight
reduce happen
nt sets of base
s feature set.
for
caus
we
clas
gi, the more it
t, and the mo
n. This is done
e features for
t calls
ost it
e until
each
E. D
Differences w
with prior attem
mpts on the top
pic
S
Several attemp
ect classificati
obje
point of view
our p
he list of its
in th
diffe
erences in a co
ed on [1] and W
base
pts have been
ion, of which
is Williams' [
citations! So
omparison bet
Williams hims
n made on con
the one that
[1]. This impo
o we provide
tween our app
self.
ncept learning
is so importa
ortance shows
some outstan
proach and the
g and
ant in
itself
nding
e ones
W
Williams' wor
K, M and D
cepts. This is
approaches.
ures in an age
iding about w
oming query o
ures. The next
s and then us
performance
rk is based on
D regions for
the first poin
We utilize t
ent's feature co
which features
or easier to say
t is Williams u
ses these rules
of such syste
n features, as
deciding abo
nt of disagree
these regions
ollection. This
s to let in d
y, to decide ab
uses features
s to decide. B
ems is directly
ours. But he
out the additio
ement betwee
s to moderate
s moderation
eciding about
bout the additi
to generate se
But we believe
y dependant t
e uses
on of
n the
e the
is for
t any
ion of
everal
e that
to the
the
conc
two
featu
deci
inco
featu
rule
the
number of rules we use. Better to say, each time we ask
something in our system, we are using some resources. So it is
better to lower the number of these queries. Another matter is
that the probabilities which Williams use are put over the rules
that he uses. In our approach, the rule set is somehow constant
and the decision making task is due to the feature collection
and their corresponding probabilities. It is once told above that
this feature-probability collection is allowed to shrink or
expand during system working time.
As a big picture to get the differences, Williams uses some
agent to collaborate about learning some concepts with
questioning eachother whether anyone knows anything about
the mentioned concept. We are using a MAS with several
expert agents, each in some realm of knowledge, to decide on
classifying one object into one or some of the realms, known as
concepts. This is done by utilizing the implicit knowledge of
our expert agents. This knowledge is represented as a
collection of coupled feature-probability metadata.
F. Fault Tolerance, Performance and Operability
It is clear from the above section "Learning more on each
agent's feature probabilities" that if some feature fj is wrongly
supposed to be added to some class's feature sets and it is
inserted to the M-region for that reason-and even increased
during some interval of false tag mentioning-, after a period of
time that it is not mentioned in its forwarded objects by the
CenterAgent, we would have the system automatically lower
its probability to lower M-region and may also to D-region not
to have bad impact on our system. This fact exists in the steady
state for the wrong features that some expert would falsely put
into some agent's body.
Another matter is the performance of this system, trying to
make its agents expert in some fields. As time goes by, the
features of a same kind and similar -- not in essence but in
concept recognition- gather together in collections, showing the
concepts they are the representative to identify with. Besides,
the reserved collection of all the disjoint class features sets in
the CenterAgent makes the system lower the message passing
rate among the pair and peer agents after a period of system
execution. This means that the CenterAgent has to consult with
fewer consultants and even get a better result and it is indeed
an economical thrift in the realm of decision making. Of course
that it will not appear when the class numbers are in the order
of 10 or 100. But suppose that we are willing to decide upon
100000 or 1000000 classes to correlate some objects into them.
Using the methods based on [1] and of course itself, any agent
should consult to any other agent on knowing anything about
the decision matter. Unavoidably a great extent of this message
passing is for nothing while there are so many agents that are
not aware of what the concept is or what objects point to it.
Using our approach in a MAS, the CenterAgent consults only
to some agents -of an order of for example 50 or 100- whether
they could help the decision making problem. It is discussed
before that these agent are the most probable agents on most
probable concepts that the pointed object could be a match for.
We believe that it is clear that such a system is better in
performance in comparison to its parents. We also believe that
the operability of such approach in any realm of application is
fully applicable and also implementable.
CONCLUSION AND FUTURE WORK
Here we addressed a method of concept learning in a Multi
Agent system, in which several agents are labeled to be an
expert in a specific domain of itself. This is supposed to be a
good way for object classification -into some known main
classes- in almost any application domain. Along with this
going on, we would like to let our system add needed (willing
to be expert) agents for any newly accepted main concept
classes. Another work is to revise the likelihood of concept
features, based on their agent-rules along the time. We would
like to mention these tasks in near future.
REFERENCES
[1] A. B. Williams: Learning to Share Meaning in a Multi-Agent System;
Autonomous Agents and Multi-Agent Systems, 8, 165-193; 2004
Kluwer Academic Publishers, Manufactured in Netherlands.
[2] L. Gravano, P. G. Ipeirotis, M. Sahami: QProber: A System for
Automatic Classification of Hidden-Web Databases
[3] M. Afsharchi, J. Denzinger, B.H. Far: Enhancing Communication with
Groups of Agents using Learned Non-Unanimous Ontology Concepts;
Web Intelligence and Agent Systems 3 (2007) 1-3; IOS Press;
[4] B. Coppin: Artificial Intelligence Illuminated; Jones and Bartlett
Publishers 2004; Sudbury, Massachusetts; pages 543-546.
[5] M. Afsharchi: Ontology-Guided Collaborative Concept Learning in
Multiagent Systems; PhD Thesis; DEPARTMENT OF ELECTRICAL
AND COMPUTER ENGINEERING, UNIVERSITY OF CALGARY;
pages 13-15.
J. M. Vidal: Fundamentals of Multiagent Systems; 24th August 2007;
page 9.
[6]
[7] N. J. Nilsson: Artificial Intelligence, a New Synthesis; Morgan
Kaufmann Publishers Inc., 1998.
[8] S. J. Russell, P. Norvig: Artificial Intelligence, a Modern Approach, 2nd
Edition; Prentice Hall Series in AI, 2002.
[9] M. Afsharchi, B.H. Far, J. Denzinger: Ontology-Guided Learning to
Improve Communication between Groups of Agents; Proceedings of the
5th International Joint Conference on Autonomous and Multi-Agent
Systems, 923-930; Hakodate, Japan 2006.
[10] D. Koller, M. Sahami: Hierarchically Classifying Documents using few
words.
|
1509.04826 | 1 | 1509 | 2015-09-16T06:48:04 | Inter-Robot Interactions in Multi-Robot Systems Using Braids | [
"cs.MA",
"cs.RO",
"eess.SY"
] | This paper describes a framework for multi-robot coordination and motion planning with emphasis on inter-agent interactions. We focus on the characterization of inter-agent interactions with sufficient level of abstraction so as to allow for the enforcement of desired interaction patterns in a provably safe (i.e., collision-free) manner, e.g., for achieving rich movement patterns in a shared space, or to exchange sensor information. We propose to specify interaction patterns through elements of the so-called braid group. This allows us to not focus on a particular pattern per se, but rather on the problem of being able to execute a whole class of patterns. The result from such a construction is a hybrid system driven by symbolic inputs that must be mapped onto actual paths that both realize the desired interaction levels and remain safe in the sense that collisions are avoided. | cs.MA | cs | Inter-Robot Interactions in
Multi-Robot Systems Using Braids
Yancy Diaz-Mercado and Magnus Egerstedt
1
5
1
0
2
p
e
S
6
1
]
A
M
.
s
c
[
1
v
6
2
8
4
0
.
9
0
5
1
:
v
i
X
r
a
Abstract -- This paper describes a framework for multi-robot
coordination and motion planning with emphasis on inter-agent
interactions. We focus on the characterization of inter-agent
interactions with sufficient level of abstraction so as to allow
for the enforcement of desired interaction patterns in a provably
safe (i.e., collision-free) manner, e.g., for achieving rich movement
patterns in a shared space, or to exchange sensor information. We
propose to specify interaction patterns through elements of the
so-called braid group. This allows us to not focus on a particular
pattern per se, but rather on the problem of being able to execute
a whole class of patterns. The result from such a construction is
a hybrid system driven by symbolic inputs that must be mapped
onto actual paths that both realize the desired interaction levels
and remain safe in the sense that collisions are avoided.
I. INTRODUCTION
M ANY applications have been proposed for multi-
robot systems. For example, the multi-robot foraging
paradigm, in which agents wander around an environment
searching for items of interest [1], [2], can relate to many
real-world problems (e.g., waste or specimen collection in
hazardous environments, explosive detection and disposal, and
search and rescue operations after a crisis such as floods and
earthquakes). Another application of interest in the research
community and in industry is cooperative assembly [3], [4] and
self-assembly [5], [6]. In the cooperative assembly scenario,
multiple robots need to coordinate their motion in order to
cooperatively assemble a structure using a possibly heteroge-
neous set of tools, end effectors, skills, or parts. Similarly, in
the self-assembly scenario, multiple robots need to coordinate
their motion in order to collectively form a structure or achieve
a geometric configuration. There has been a recent push for
robotic farming and precision agriculture [7], [8], where a fleet
of robots is sent to gather data on the status of crops, tend
to and harvest them. In some communication architectures,
mobile agents called data MULEs or message ferries [9] --
[11] are used to transport data between sensors, access points,
or base stations,
to
establish physical communication. Multi-robot simultaneous
localization and mapping (SLAM) takes advantage of the robot
team's size to attain a more complete (or more detailed) map
faster than with a single robot by combining information from
multiple sources and coordinating motion [12], [13]. Other
applications include transportation systems (e.g., intelligent
highways, air traffic control) [14], and the convoy protection
scenario (e.g., surveillance, coverage) [15].
in situations where it
is impractical
The authors are is with the Department of Electrical and Computer
Engineering, Georgia Institute of Technology, Atlanta, GA, 30332 USA.
E-mail: {[email protected], [email protected]}.
In many of these applications, the overall objectives can
be stated in terms of making a team of robots follow a
physical path, such as a road or the movements of a ground
convoy, while ensuring that particular search patterns are
executed [16] -- [18]. These patterns should be selected in such
a way that certain secondary geometric objectives are met,
including ensuring that an area along the path is covered, that
multiple views of the same objects are achieved, that an aerial
vehicle is always on top of the convoy, or that sufficiently
many vehicle-to-vehicle interactions take place for the purpose
of information sharing [19], [20]. In this work, we collect
all of these different secondary objectives under one unified
banner, namely multi-robot mixing. In particular, we specify
interaction patterns with certain desired levels of mixing, and
then proceed to generate the actual cooperative movements
that realize these mixing levels.
In this paper we study the problem of characterizing inter-
robot interactions for the sake of coordination and collision
avoidance. We specify the mixing patterns through elements
of the so-called braid group [21], [22], where each element
corresponds to a particular mixing pattern, i.e., we do not focus
on a particular pattern per se, but rather on the problem of
being able to execute a whole class of patterns. The result
from such a construction is a hybrid system driven by symbolic
inputs [23], i.e., the braids, that must be mapped onto actual
paths that both obtain the mixing level specified through the
braid, and remain safe in the sense that collisions are avoided.
The use of symbolic inputs allows us to abstract away
the geometry and physical constraints involved in the multi-
robot motion planning. As described in [23], this provides
the advantage of hierarchical abstractions that are typically
broken into three stages. At the top layer is the specification
level, which describes the motion tasks (such as robots A and
B should interact, and arrive at the goal simultaneously). The
second layer is the execution level, which describes how to
obtain the motion plans, e.g., by generating trajectories based
on optimality conditions. The bottom layer is the implemen-
tation level, which concerns itself with constructing the robot
controller, e.g., to track a reference trajectory. This work also
extends the notions presented in [24] for trajectory tracking of
virtual vehicles. The idea there is to have the physical vehicles
track a virtual vehicle, as opposed to a path (or trajectory)
itself. The virtual vehicle, which is being controlled directly
with simplified dynamics, can be controlled in order to track a
reference path while satisfying the constraints, compensating
for the physical vehicle's dynamics and other disturbances.
The outline of the paper is as follows: in Section II we
start with a brief summary of the existing literature on braid-
ing for the sake of multi-robot motion planning; in Section
II-B, the braid group is introduced as a way of specifying
mixing levels and the corresponding symbolic braid objects
are given a geometric interpretation in terms of planar robot
paths; controllers are then proposed so that mixing strategies
satisfying a given specification can be executed by a class of
robots as described in Section III, together with a bound on the
highest achievable mixing level. In Section IV, implementation
of these controllers on actual robotic platforms is addressed,
and these ideas are deployed on a team of actual mobile robots.
II. BRAIDS AND ROBOTICS
The use of braids for robot motion planning has been
considered before. Using the notion of configuration space
[25] for robot motion planning, [26] studies the problem
of construction and classification of configuration spaces for
graphs, e.g., robots on a manufacturing floor constrained on
rails or paths. By studying the topological data associated with
these graphs, such as the braid groups, he is able to provide
a measure of the complexity of the control problem (e.g., the
construction of potential field controllers on homeomorphic
spaces).
In [27], the graph braid group is used as the fundamental
group of the configuration space of graphs that describe robot
motion. There, each graph in the configuration space repre-
sents discretized collision-free robotic motion plan (e.g. road
maps), where at each discrete time the graph vertices represent
the positions of robots and (possibly moving) obstacles, and
the edges represent fixed tracks connecting these vertices.
They present an algorithm to construct the presentation of the
graph braid group of N agents, where the group generators
(i.e., the braids) represent actual paths between configurations
of robots on the graph, i.e., the motion plan to transition from
one configuration to another. However, they only consider
zero-size robots where they rely on a "one edge separation"
between points at all times to avoid collisions, and as such,
their approach is purely academic and mainly focuses on the
combinatorics of ideal robots moving on a graph.
We will use the braid group's generators as symbolic
inputs to multi-robot hybrid controllers which characterize
and enforce collision-free interactions, take into account kine-
matic and geometric constraints, and are executable on actual
robotic platforms. This is the first approach that uses braids
in multi-robot motion planning to symbolically characterize
and enforce rich interaction patterns, implementable on ac-
tual robotic platforms. In particular, we extend the notions
presented in [28], [29]. The definition on planar braids and
their geometric interpretations are generalized from their orig-
inal form presented in [28]. A controller was designed for
nonholonomic vehicles to track a geometric path in [29].
In this paper, we modify this controller to instead produce
optimal trajectories that are provably safe in the collision-free
sense, satisfy a set of spatio-temporal constraints, and follow
desired geometric paths (Section III-B). This also allows us
to obtain tighter bounds on the amount of interaction patterns
that are achievable in a space than the one found in [29]. Two
new contributions to this work are found in Section III-C,
2
1
2
1
2
1
2
2
1
1
2
2
1
σ1
σ0
σ1
(a) Two robots, orig-
inally at one side of
a square, map to the
other.
Fig. 1. Two-robot interactions and corresponding symbolic braids.
(b) Three braids for two planar agents. Note that
"strand" 1 goes over "strand" 2 in σ1 and vice versa
in σ1.
where these trajectory generating controllers are extended to
non-rectangular regions, and Section IV-A, which includes
development of an optimal trajectory tracking controller with
formal guarantees on optimality and spatio-temporal constraint
satisfaction.
A. Planar Braids
Consider two agents on a square, initially located at the
two left vertices of the square as in Fig. 1a. The agents' task
is to move to the two right vertices of the square. There are
two ways in which these target vertices can be assigned. The
first is to simply let the robots move along a straight line
while the second is to have them cross paths and move to
vertices diagonally across from their initial placement. If the
robots are not to collide with each other, one agent can cross
the intersection of the two paths first, and then the other (or
vice versa). In the braid group, these two options correspond
to different "braids," and we have thus identified three planar
braids for two agents, as shown in Fig. 1b. Let us momentarily
denote these three braids, σ0, σ1, σ1.
Now, given these three braids, we can concatenate them
together to form other braids, as seen in Fig. 2. The left braid is
given by σ1·σ1 and the right braid is σ1·σ1. In the braid group,
what really matters is not the geometric layout of the paths, but
how the paths wrap around each other. As can be seen, if we
were to "pull" the right corners in σ1· σ1, the "strands" would
get "tangled" in the middle, while a "stretched-out" σ1 · σ1 is
simply σ0. Thus we let σ0 be the identity braid, such that σ1
and σ1 are each others' inverses in the sense that
σ1 · σ1 = σ1 · σ1 = σ0.
1 = σ1 (and σ−1
In fact, every braid has an inverse and, as such, the set of braids
(together with the concatenation operation) is indeed a group.
And, as σ−1
1 = σ1), σ0 and σ1 (or σ0 and
σ1 for that matter) are the so-called generator braids for this
group in that all planar braids can be written as concatenations
of these two braids and their inverses [21], [22].
This notion of generator braids can be extended to the case
when there are N ≥ 2, with the only difference being that
we will have N generators rather than just two, i.e., let σ0
be the trivial generator with no interactions and σk denote the
interactions between agents k and k + 1, k = 1, . . . , N − 1.
If we let ΣN be the set of all planar generator braids over
N agents, this set will serve as the alphabet over which braid
strings (themselves braids) are produced from, and we let ΣM
N
denote the set of all braids of length M (i.e., braids composed
of M generators) over N agents.
σ(1)(cid:121) σ(2)(cid:121)
ξ (0, 1)
ξ (1, 1)
ξ (2, 3)
ξ (0, 2)
ξ (1, 3)
ξ (2, 1)
3
} h
ξ (0, 3)
(cid:124)
ξ (1, 2)
(cid:123)(cid:122)
ξ (2, 2)
(cid:125)
(cid:96)
P 0
3 = {(0, 0) , (0, 0.5h) , (0, h)}
P 2
3 = {((cid:96), 0) , ((cid:96), 0.5h) , ((cid:96), h)}
P 1
3 = {(0.5(cid:96), 0) , (0.5(cid:96), 0.5h) , (0.5(cid:96), h)}
Fig. 3. The geometric interpretation of braid string σ = σ2 · σ1 for the
three-agent case. In this example, σ(1) = σ2 and maps P 0
3 , while
σ(2) = σ1 and maps P 1
3 to P 1
3 to P 2
3 .
where σ (k) is the kth braid2 in the string σ and
(cid:110)(cid:0) q
M (cid:96), 0(cid:1) ,
(cid:16) q
P q
N =
(cid:17)
M (cid:96),
1
(N−1) h
,
(cid:16) q
(cid:17)
, . . . ,(cid:0) q
M (cid:96), h(cid:1)(cid:111)
,
M (cid:96),
2
(N−1) h
as shown for the three-agent case in Fig. 3.
We moreover use the notation ξ(i, j) ∈ R2 to denote the
point agent j should go to at step i, i = 1, . . . , M. We use
the convention that ξ(0, j) = (0, (j − 1)h/(N − 1)), j =
1, 2, . . . , N, to denote agent j's initial position. In other words,
ξ(1, j) = σ(1)(cid:104)ξ(0, j)(cid:105),
ξ(2, j) = σ(2)(cid:104)ξ(1, j)(cid:105) = σ(2) ◦ σ(1)(cid:104)ξ(0, j)(cid:105),
or more generally,
ξ(i, j) = σ(i)(cid:104)ξ(i − 1, j)(cid:105)
= σ(i) ◦ σ(i − 1) ◦ ··· ◦ σ(1)(cid:104)ξ(0, j)(cid:105),
where we use the (cid:104)·(cid:105) notation to denote the argument to
σ(i) and ◦ to denote composition. This construction is also
illustrated in Fig. 3 for the three-agent case.
The geometric interpretation we will make of the planar
braids is that the mobile agents that are to execute them must
traverse through these braid points. They must moreover do
so in an orderly and safe manner, which will be the topic of
the next section.
III. EXECUTING BRAIDS
Given a collection of N agents with dynamics
σ1 · σ1
σ1 · σ1
Fig. 2. Two concatenated braids σ1 · σ1 and σ1 · σ1. The latter of these two
braids is the same as the identity braid σ0.
B. A Geometric Interpretation
Although heavily inspired by geometry in [21], the braid
group is not concerned with the actual geometry of the braid
strands. For the sake of describing the robot motion plans,
we will associate geometric paths with the different braids.
First of all, we assume that the braid is geometrically located
in a rectangular area of height h and length (cid:96) no matter
how long the braid string is. Using the particular two-agent
braids discussed in the previous paragraphs, we assume that
the two agents are initially located at the points (0, 0) and
(0, h), while the final locations are at ((cid:96), 0) and ((cid:96), h). If
the total braid results from the use of one single generator
braid σ0, or σ1 then no additional points are needed. However,
if the braid has length 2, then we also need to introduce
intermediary "half-way" points ((cid:96)/2, 0) and ((cid:96)/2, h). As such,
we let P q
spaced1 positions for these intermediary points, where the
subscript 2 denotes the two-agent case, M is the length of
the braid to be executed, and q = 0, 1, . . . , M.
M (cid:96), h(cid:1)(cid:9) be a set of uniformly
M (cid:96), 0(cid:1) ,(cid:0) q
2 = (cid:8)(cid:0) q
Using this notation, we can refer back to Fig. 1a and say that
each of the two generator braids correspond to an assignment,
i.e., a bijective map, between P 0
2 , and we use the
following notation to denote this fact
2 and P 1
σi : P 0
2 →b P 1
2 ,
i = 0, 1,
where →b denotes "bijection". Note that this is a slight abuse
of notation in that σi now denotes both an element in the braid
group as well as a map -- this distinction, however, should be
entirely clear from the context. Further, we will refer to these
points which agents are bijectively mapped to and from as
braid points.
If we generalize this to N ≥ 2 agents and let σ denote a
N , we will
string of generators of length M ≥ 2, i.e., σ ∈ ΣM
use the notation
σ(k) : P (k−1)
N →b P k
N ,
k = 1, . . . , M,
and planar output
xj = f (xj, uj),
yj = h(xj) ∈ R2,
j = 1, . . . , N,
1Unless otherwise stated, the rest of this document we will assume these
sets of intermediary points are uniformly spaced in h and (cid:96), but the notions
presented here extend to points which are not, as the application demands.
More on this in Section III-C.
2Note that a "braid" here refers to a member of the braid group, e.g., a
single generator (viz., a single bijective map) or a concatenation of several
generators (viz., a composition of bijective maps).
then it is of interest to have these agents execute a braid
σ ∈ ΣM
N . We now define what it means for this braid to be
executed.
Given an input braid string σ, what each individual agent
should do is "hit" the intermediary braid points ξ(i, j) at
specified time instances. We let T denote the time it should
take for the entire string to be executed. As such, the first
condition for a multi-agent motion to be feasible with respect
to the braid is the following:
Definition 3.1 (Braid-Point Feasibility):
A multi-robot trajectory is braid-point feasible if
yj(ti) = ξ(i, j), i = 0, . . . , M, j = 1, . . . , N.
where the ti's form a partition of a given time window [0, T ],
i.e.,
(cid:5)
t0 = 0 < t1 < ··· < ti < ··· < tM = T.
On top of braid-point feasibility, we also insist on the robots
not colliding during the maneuvers. To a certain degree, this
condition is what restricts the level of mixing that is possible,
i.e., since the braid is constrained in a rectangle of fixed height
and width, what length strings the multi-robot system can
execute while maintaining a desired level of safety separation.
Definition 3.2 (Collision-Free):
A multi-robot trajectory is collision-free if
(cid:107)yi(t) − yj(t)(cid:107) ≥ δij,
∀i (cid:54)= j,
t ∈ [0, T ],
where δij > 0 is the desired level of safety separation between
(cid:5)
agents i and j.
For convenience, we will refer to ¯δ = max δij as the
maximum safety separation such that no agent collides. This
will come up in Theorem 2. We are missing a notion to
describe what the multi-robot mixing problem is, that is, what
constitutes a braid controller.
Definition 3.3 (Braid Controller):
A multi-robot controller is a braid controller if the resulting
trajectories are both braid-point feasible and collision free, for
all collision-free initial conditions such that
4
A. Braid Controllers: Stop-Go-Stop Hybrid Strategy
Our first attempt at executing braids will be a hybrid control
strategy called the Stop-Go-Stop. We will assume that agent
dynamics are given by single integrator dynamics, i.e., xj =
uj ∈ R2 with yj = xj, j = 1, . . . , N. Further, for practical
considerations, assume that there is a cap on the maximum
velocity achievable by the agents, i.e., (cid:107)uj(cid:107) ∈ [0, vmax]. The
idea is then at each braid step to "send" agents off straight
to their next braid point in order of the distance they'll need
to travel, waiting just long enough to avoid collisions before
sending an agent off. To that end, we define si : {1, . . . , N} →
{1, . . . , N} to be a bijective mapping that denotes the farthest
distance ordering at step i, that is
si(p) < si(q)
=⇒
(cid:107)ξ(i, p) − ξ(i − 1, p)(cid:107) ≥ (cid:107)ξ(i, q) − ξ(i − 1, q)(cid:107)
(cid:17)
tan−1(cid:16) (ξ(i,j)−ξ(i−1,j))2
where ties are arbitrarily broken such that si
remains a
bijection. We let the agents heading angle be given by θi,j =
where the subscript indicates the
first or second component, and the unit heading vector be
ρi,j = [cos(θi,j), sin(θi,j)]T . Lastly, the time the agents will
wait before entering GO mode will be given by their ordering
as (si(j) − 1)τ where
(ξ(i,j)−ξ(i−1,j))1
τ =
δ
vmax cos(θ∗)
is the time required to be δ apart horizontally3, with cos(θ(cid:63)) =
(cid:96)/M√
(cid:96)2/M 2+h2 being the maximum horizontal distance an agent
could travel given σ ∈ ΣM
N . To ensure that the agents do
not overtake the first agent (horizontally), the speed of the
agents should be scaled by the velocity of the first agent, i.e.,
(1))/ cos(θi,si(j)). The hybrid automaton
uj = vmax cos(θi,s
describing the stop-go-stop controller is given in Fig. 4.
−1
i
Theorem 1: [28] The STOP-GO-STOP controller in Fig. 4 is
a braid controller if the braid points themselves are sufficiently
separated and
(cid:16)
(cid:17)
≥(cid:112)(cid:96)2/M 2 + h2.
yj(0) = ξ(0, j), j = 1, . . . , N.
(cid:5)
cos(θ(cid:63))vmax
min
i
(ti − ti−1) − (N − 1)τ
As a final notion, we are interested in how much mixing a
particular system can support.
Definition 3.4 (Mixing Limit):
The mixing limit M (cid:63) is the largest integer M such that there
(cid:5)
exists a braid controller for every string in ΣM
N .
Whenever two strands of the braid associated with a given
braid string cross, the two associated agents will have to inter-
act. The mixing limit therefore serves as an input-independent
bound on how much mixing is achievable for a given team of
agents operating in a given environment. The mixing limit is
in general quite hard to compute, it needs to consider every
permutation of strings of varying lengths up to some number,
the geometry assigned to each string and is dependent on
the kinematical response of the multi-robot system. However,
under certain assumptions it is possible to find bounds on M (cid:63)
for a given braid controller.
Proof:
The STOP-GO-STOP controller ensures that the agents are
never within δ of each other by virtue of the fact that they
have to wait until they are indeed at least that far apart (hor-
izontally) before entering GO mode. As such, the trajectories
are collision-free.
What remains to show is that they are also braid-point
feasible. Consider the agent that has to wait the longest before
it can move, i.e., the agent that has to wait a total of (N −1)τ,
and at its worst has mini (ti − ti−1) − (N − 1)τ time left
to reach the next braid point. In other words, we need that
the distance traveled in that amount of time at the speed
3Since we are interested in the mixing limit, the analysis is done with the
horizontal direction being the limiting factor for safe execution of the braids.
t = (si (j) − 1) τ + ti−1
Mixing Bounds
i := 1
STOP
xj = 0
i := i + 1
t = ti
xj = vmax
STOP
xj = 0
xi = ξ (i, j)
GO
(cid:16)
cos(cid:0)θi,si(j)
θi,s
−1
i
cos
(cid:17)
(cid:1) ρi,j
(1)
Fig. 4. Hybrid STOP-GO-STOP braid controller.
vmax cos(θi,s
required. But, we note that
−1
i
(1))/ cos(θi,si(j)) is greater than the distance
vmax cos(θi,s
−1
i
(1))/ cos(θi,si(j)) ≥ vmax cos(θ(cid:63))
and, as we are only looking for a bound, we assume that we
use this lower speed and that the distance required to travel is
the largest distance possible (which it really is not). In other
words, we need
(cid:16)
cos(θ(cid:63))vmax
min
i
(ti − ti−1) − (N − 1)τ
(cid:17)
≥(cid:112)(cid:96)2/M 2 + h2,
and the proof follows.
Note that Theorem 1 implicitly provides a lower bound on
the mixing limit as long as the agents' paths are straight lines.
In Fig. 5 we can see the mixing limit as a function of the
number of agents in the team for parameters vmax = 5, T =
20, (cid:96) = 5, h = 10, and ∆ = 0.2. The problem with this
strategy is that it does not allow for more general geometric
paths (which could be interpreted as feasible trajectories under
a given robot dynamical model), nor does it ensure that agents
get within a specified distance from each other (which could be
necessary when collaboration is required). The next section we
present a new strategy called braid reparameterization which
will explicitly consider more general strand geometries, and
allow us to obtain analytical bounds on the mixing limit.
B. Braid Controller: Braid Reparameterization
We are now seeking a strategy that will allow us to follow
a given geometry while achieving a mixing strategy encoded
as a braid string σ ∈ ΣM
N . Further, we wish to enforce inter-
agent interaction as dictated in σ by having agents get as close
as the safety separation δjk. Before moving forward, consider
the following lemma.
Lemma 1: If at any braid step, the generators in the braid
substring σ(k) = σ ⊆ Σm
N , m ≤ M, have indices that are two
or more apart, then any agent interacts with at most one other
agent at this step.
Proof:
Let Bk
N ∈ RN×2 be a matrix that contains the set of
braid points at time k such that B0
N = [ξ(0, 1)··· ξ(0, N )]T .
Consider the two-generator concatenation σi·σj. As a bijective
map, if σi = σ0, then the agents do not interact, and the agent
in braid point position [Bk−1
N ]n gets mapped to braid point
5
60
70
70
M
s
r
o
t
a
r
e
n
e
G
f
o
r
e
b
m
u
N
10
9
8
7
6
5
4
3
2
1
0
10
10
30
30
20
Number of Agents N
50
50
40
Fig. 5. Lower bound on the mixing limit using the Stop-Go-Stop braid
controller.
N ]n, where [B]n corresponds to the nth row of B,
position [Bk
n = 1, . . . , N. If σi (cid:54)= σ0, then σi will swap the position of the
two agents occupying the braid point positions i and i + 1 at
step k−1, i.e., it maps the agent occupying [Bk−1
N ]i to [Bk
N ]i+1
and the agent occupying [Bk−1
N ]i. Similarly, σj
swaps the position of the agents occupying the braid point
positions j and j + 1.
The two agents in positions i and i + 1 at k − 1 would only
interact with the two agents in positions j and j + 1 at k − 1
if i (or i + 1 for that matter) is equal to either j or j + 1. But
if we let i − j ≥ 2, then we get that
N ]i+1 to [Bk
i − j ≥ 2 ⇔ 2 ≤ i − j or 2 ≤ j − i
⇒ j < j + 1 < j + 2 ≤ i < i + 1
i < i + 1 < i + 2 ≤ j < j + 1
or
i (cid:54)= j
i (cid:54)= j + 1
i + 1 (cid:54)= j
i + 1 (cid:54)= j + 1,
⇒
and as such these two-generator concatenation maps the agents
from one set of braid points to the next with at most two
interaction between at most two agents per interaction.
One of the two braid group relations [21], [22] tells us that
if i − j ≥ 2 then σi · σj = σj · σi. More generally, if we let
h : {1, . . . , m} → {1, . . . , N − 1}
h(i) − h(j) ≥ 2 for all
be a surjective map such that
i, j ∈ {1, . . . , m} with i (cid:54)= j, then for any bijective map
g : {1, . . . , m} → {1, . . . , m} we have that
σ(k) = σh(1) · σh(2) · ··· · σh(m)
= σh(g(1)) · σh(g(2)) · ··· · σh(g(m)).
As such, the braid generators can be rearranged to obtain any
permutation of two-generator concatenations from generators
in σ(k). Since for all permutations of two-generator concate-
nations we will have indices that are two or more apart, these
{σ1 · σ3 · σ2}
= {σ3 · σ1 · σ2}
(a)
{σ1 · σ3} · σ2
= {σ3 · σ1} · σ2
(b)
σ1 · σ3 · σ2
(c)
σ3 · σ1 · σ2
(d)
Fig. 6. A braid string σ = σ1 · σ3 · σ2 taking place in a varying number of
braid steps.
will map the agents at braid step k from the set of braid
points Bk−1
N with at most
m interactions total, and at most one interaction per agent.
to the next set of braid points Bk
N
subject to
6
in order to safely execute braids, one need only consider the
case when two agents interact. We will devise a strategy for
reparameterizing the geometry such that should two agents
interact, the resulting parameters are at least δ from each other.
By tracking the parameterized paths, the agents' trajectories
will be collision-free and braid point feasible. The resulting
controllers can be combined to satisfy a given braid string.
Consider the geometric path agent j must follow at step
i (0) = ξ(i − 1, j) and
i given by γj
i
γj
i (1) = ξ(i, j). Let ∆ be the arclength of this path, given by
: [0, 1] → R2 with γj
(cid:90) 1
(cid:114)(cid:16)
(cid:17)T
∆ =
0
γj
i (p)
γj
i (p) dp.
We wish to find a parameterization of the strand geometries
such that the parameters of two intersecting strands, thought of
as a virtual vehicles that live on their geometries, are collision-
free. The strategy to do so will be to impose constraints on
the agents separation from the path intersection at a specified
time. To ensure braid-point feasibility, the strategy will be to
impose constraints on the time in which the agents must get
from the beginning of one braid step to the end of that step. To
that end, we propose the following constrained optimization
problem
(v(cid:63)
j , v(cid:63)
k) = arg min
(vj ,vk)
J(vj, vk) = arg min
(vj ,vk)
1
2
(cid:90) ti
ti−1
v2
j + v2
k dτ
(1)
If the geometric interpretation of the braid string is restricted
to the case of only pairwise interactions at every braid step,
then it is possible to devise a hybrid strategy with the desired
properties, which we can then compose together to achieve
the desired interaction patterns. As such, we will restrict the
geometric interpretation of braid strings to those satisfying
pairwise interactions as in Lemma 1.
Restriction 1: As a bijective map from one set of braid
points to another, the braid σ(k) will only contain generators
whose indices are two or more apart, i.e., for some h : N →
{0, . . . , N − 1}
σ(k) = σh(1) · σh(2) · ··· , where h(i) − h(j) ≥ 2 ∀i (cid:54)= j
As an example, suppose that a braid string contains the
substring of three concatenated generators σ1 · σ3 · σ2. The
first two generators, σ1 · σ3, may take place simultaneously at
braid step 1 without incurring in more than one interaction per
agent with another agent, but we would require an additional
braid step for σ2 in order to avoid multiple interactions in the
same step. This is illustrated geometrically in Fig. 6. In (a),
an agent interacts with more than one other agent since all
the generators are not at least two indices apart and take place
at the same braid step. Note that in the remaining cases (b)-
(d), we introduce intermediate braid points while retaining the
desired level of interaction, final configuration of the agents,
and reducing to pairwise interactions.
Braid strings that satisfy Restriction 1 will have at most
interactions involving two agents at any given braid step. So
pj
and
pj = vj,
(cid:18) ti−1 + ti
2
2
pk = vk,
(cid:18) ti−1 + ti
pk
(cid:19)
(cid:19)
=
∆ + δ
2∆
,
∆ − δ
2∆
,
=
pj(ti−1) = 0,
pj(ti) = 1,
pk(ti−1) = 0,
pk(ti) = 1.
2∆
The constraints at ti−1 and ti will ensure that the reparame-
terization is braid-point feasible. For collision avoidance we
will define a safety separation region, as depicted in Fig.
7, to be the region from the intersection point to the point
where it is possible for the two agents to be within δjk of
(cid:3) for δ ∈ [0, ∆] such
i (pk)(cid:13)(cid:13) ≥ δjk for all pj, pk ∈ [0, 1], where
each other, i.e., the set (cid:2) ∆−δ
that (cid:13)(cid:13)γj
i (pj) − γk
2∆ , ∆+δ
we let δ be the distance from the intersection point to the
boundary of the safety separation region along the path. Note
that if the geometry of the braid strands are straight lines, the
distance δ can be easily computed by δ = δjk csc(θ), where
θ is the angle between the two intersecting lines, and this
would also imply the parameters get as close as δjk csc (θ/2).
Similarly, when the geometry is "city-block"-like, we can find
2(N−1) and the agent get as close as 2δjk. For
δ = δjk +
more general geometries, δ may be conservatively selected.
At ¯ti
, i.e., half-time along the braid step, we
will require one parameter to exit the safety separation region
as the other one is about to enter it. All these cases are
:= ti−1+ti
h
2
δjk
θ
δ
δ(cid:48)
δjk
δ
Fig. 7. Safety separation region for three different-geometry, two-agent braids.
For straight lines (left) or "city-block"-like (right), the distance δ can be
computed exactly. For arbitrary curves (center), the distance δ(cid:48) can be selected
conservatively.
illustrated in Fig. 7. This way we guarantee that the two
parameters are never inside the region simultaneously and thus
their separation will always be of at least δjk.
The optimality conditions for equation (1) are given by
λj = λk = 0,
pj(¯ti) = ∆+δ
2∆ ,
pj = −λj,
pk(¯ti) = ∆−δ
pk = −λk,
2∆ , pj(ti) = pk(ti) = 1.
Since λj = λk = 0, we will get that λj and λk are piecewise
constant. By using the midway condition, we have that for
t ∈ (ti−1, ¯ti]
pj(¯ti) =
vj dτ =
∆ + δ
2∆
vj(¯ti − ti−1) =
⇒
∆ + δ
2∆
vj =
1
∆
∆ + δ
(ti − ti−1)
and similarly for t ∈ (¯ti, ti], the terminal condition tells us
that
(cid:90) ¯ti
ti−1
⇒
(cid:90) ti
¯ti
⇒
7
performed in a space of height h and length (cid:96) in time T is
bounded above by
(cid:113)
(cid:96)
M (cid:63) ≤ min
4h2 − ¯δ2 (N − 1)2
,
¯δh
(N − 1)(cid:0)vmaxT −(cid:0)(cid:96) + ¯δ(cid:1)(cid:1)
h
(cid:41)
.
− 1
2
Proof:
Consider σ ∈ ΣM
N . Since at any step an agent can either inter-
act with another agent or move straight ahead, the maximum
mixing will be achieved if an agent interacts with another
agent at every step. Thus, at every step it must be enforced
that there are no collisions. Assuming a uniform partition of
the time window, the arclength-normalized parameter velocity
will be given by
(cid:26) (cid:0) M ∆±M δ
(cid:0) M ∆∓M δ
T ∆
(cid:1)
(cid:1)
T ∆
if t ∈(cid:0) i−1
if t ∈(cid:0) 2i−1
2M T(cid:3)
M T(cid:3)
M T, 2i−1
2M T, i
vj =
for i = 1, 2, . . . , M, where the sign on the numerator depends
on the interpretation of whether the "strand" goes over or
under, ∆ is the arclength of the strand geometry connecting
two braid point, assumed equal at each braid step and for both
agents due to symmetry, and δ is the safety separation distance
along the braid as described above. The total braid path length
∆ heavily depends on the geometry of path. However, for the
sake of obtaining bounds on the mixing limit, we may assume
that the braid points are uniformly distributed in height and
length, such that the length of sufficiently regular paths will
be bounded by(cid:115)(cid:18) h
(cid:19)2
N − 1
+
(cid:19)2 ≤ ∆ ≤ h
(cid:18) (cid:96)
M
N − 1
+
(cid:96)
M
(3)
pj(ti) =
vj dτ +
2∆
∆ + δ
∆
=
∆
vj(ti − ¯ti) =
1
2∆
⇒
(2∆ − ∆ + δ)
vj =
1
∆
∆ − δ
(ti − ti−1)
where the lower bound assumes straight lines connecting the
braid points and the upper bound assumes "city block"-like
paths stepping midway between the two points. But after
normalizing the bounds on the parameter velocity, we see that
Note that for agent k, the signs are reversed on the numer-
ator. As such, the resulting braid parameterization will have
velocities given by
pj(t) =
∆
1
∆
∆±δ
∆∓δ
(ti−ti−1)
(ti−ti−1)
t ∈ (ti−1, ¯ti]
t ∈ (¯ti, ti]
(2)
(cid:40) 1
where the sign in the numerator is determined by the interpre-
tation given to the braid strand going "under" (e.g., the agent
crosses the intersection point first) or "over" (e.g., the agent
crosses the intersection point second), and in cases where there
are no intersections we set δ = 0 in the numerator.
It is possible to come up with an upper bound on the
length of the braid attainable under this mixing scheme. Under
Restriction 1, the following theorem provides an upper bound
on the mixing limit.
Theorem 2: Given the maximum safety separation ¯δ and
bounds on the agents' velocities such that vj(t) ∈ [0, vmax]
∀t, j, the mixing limit M (cid:63) for N-agent braids that can be
≤ M (∆ + δ)
T ∆
≤ vmax
∆
.
(4)
The lower bound of (4) tells us that for the parameter to not
go backwards we need δ ≤ ∆. To ensure this, we set δ =
¯δ csc θ ≤
¯δ ≤ h
relationships to solve for M yields
(cid:1)2 ≤ ∆, and since we require
N−1 for collision-free braid points, using the geometric
M
T ∆
0 ≤ M (∆ − δ)
(cid:114)(cid:16) h
(cid:17)2
+(cid:0) (cid:96)
(cid:113)
M ≤ (cid:96)
N−1
4h2 − ¯δ2 (N − 1)2
.
¯δh
M
=
N−1 )+¯δ)
Similarly, the right-hand side inequality of (4) tells us that
∆ ≤ vmaxT−δ
, and to ensure this we
set ∆ ≤ h
. Solving for M
yields
vmaxT−( 1
2 ( h
M
M ≤ vmaxT−( 1
M ≤ (N − 1)(cid:0)vmaxT −(cid:0)(cid:96) + ¯δ(cid:1)(cid:1)
N−1 + (cid:96)
N−1 )+¯δ)
2 ( h
M
− 1
2
.
h
M
d
n
u
o
B
t
i
m
L
i
g
n
i
x
i
M
20
0
10
20
Number of Agents N
40
20
60
)
s
(
T
d
e
t
o
l
l
A
e
m
T
i
Fig. 8. Upper bound on the Mixing Limit presented in Theorem 2 for
parameters (cid:96) = 2 m, h = 4 m, δ = 0.13 m, vmax = 2 m/s
(cid:113)
Thus
M (cid:63) ≤ min
(cid:96)
4h2 − ¯δ2 (N − 1)2
,
¯δh
(N − 1)(cid:0)vmaxT −(cid:0)(cid:96) + ¯δ(cid:1)(cid:1)
h
(cid:41)
− 1
2
. (cid:4)
Theorem 2 provides a compact expression to obtain an
upper bound on the mixing limit that abstracts away strand
geometry. It provides a notion of the whether or not desirable
mixing levels are achievable in the space, regardless of what
the actual movement patterns to achieve these mixing levels
are (encoded in the braid string of length M ≤ M (cid:63)). Fig. 8
includes a plot of the bound on the mixing limit for varying
number of agents and time window size.
C. Non-Rectangular Regions
Up to this point, only the problem of braiding on a rect-
angular region of height h and length l has been considered.
On this region, the braid points were uniformly distributed
along both dimensions and bounds on the mixing limit were
provided through the use of the proposed braid controller. In
this section, the scheme is extended to more generally shaped
regions, e.g., the road on Fig. 9. As has been done previously,
discussion begins by first considering the two agent case.
Consider the two agents attempting to perform a braid
of length one on the arbitrarily curved region on Fig. 9b.
Connecting the braid points together results in the quadrilateral
depicted in the red dotted line. Let
this quadrilateral be
considered as the space where the agent needs to perform a
braid of length one, rather than the curved region itself. Note
that as longer length braids are included in this road segment,
other quadrilaterals appended together will be obtained which
approximate the road slightly better. Since it is of interest to
obtain mixing strategies near the mixing limit, as longer length
8
(a) Three agents braiding on a curved region.
(b) Two agents need to move
from the left-most circles to
the right-most circles along
the curved region.
(c) The longer the braid length
in this curved road segment,
the closer the quadrilaterals
resemble the curved region.
Fig. 9. Agents braiding on a region that curves.
braids are included in this road segment, better approximations
of the curved region will be obtained by these composition of
quadrilaterals. This is depicted in Fig. 9c.
The strategy for performing a mixing strategy in curved re-
gion will be to transform the curved region into a straightened
rectangular region of known height and length, as illustrated
in Fig. 10a. In this way the braid controller can be fashioned
as in previous sections and the resulting braid controller can
be transformed back into the actual curved region. After
distributing the braid points on both the curved region and
the rectangular region, as in Fig. 10a, the next step is to find
a transformation to map between these two regions.
Let agents j and k interact in at braid step i. Denote S q
i,j
to be the quadrilateral formed by connecting together the
braid points ξ(i − 1, j), ξ(i − 1, k), ξ(i, j), and ξ(i, k), like
the one depicted in Fig. 9b. Let S r
i,j be a rectangular plane
of specified height and length whose corners are given by
ξr(i − 1, j), ξr(i − 1, k), ξr(i, j), and ξr(i, k).
With knowledge of these braid points and through the use of
a projective transform as in [30], it is possible to obtain a local
diffeomorphism that maps from a rectangle of specified height
and length to the convex arbitrarily shaped quadrilateral, i.e.,
Ti,j : S r
i,j → S q
i,j.
i,j and S r
Note that by selecting transforms based on the corners of
these quadrilaterals, a continuous curve that spans across the
boundary between S r
i+1,j might be mapped to a
discontinuous curve that spans across the boundary between
S q
i,j and S q
i+1,j when transformed using Ti,j and Ti+1,j in their
respective spaces. However, there will certainly be continuity
in the mapping of curves passing through the braid points,
since these points are shared by the quadrilaterals and are
used to compute the transforms, i.e., Ti,j (ξr(i, j)) = ξ(i, j) =
Ti+1,j (ξr(i, j)).
Recall
the braid controller proposed for the rectangular
9
point along the curve where the two curves intersect (resp.
where the agent exits the safety separation region).
In the special case where the geometry is given by straight
lines, then by setting
i (p) = (1 − p)ξ(i − 1, n) + pξ(i, n), p ∈ [0, 1], n = j, k.
γn
as the path agent j and k must follow, the intersection point
s may be found by setting s = γj
i (πk) where
i (πj) = γk
= A−1 (ξ(i − 1, k) − ξ(i − 1, j))
(cid:20)πj
(cid:21)
πk
with A = [(ξ(i, j) − ξ(i − 1, j)) ,− (ξ(i, k) − ξ(i − 1, k))].
Recall that the distance in S q
i,j from the intersection point s,
for the special case of the geometry being straight lines, was
given by δ = δjk csc(θ) with θ being the angle between these
two lines, i.e., θ = cos−1(cid:0)x(cid:62)
(cid:1) where xn is the unit vector
j xk
pointing towards the next point, i.e.,
(ξ(i, n) − ξ(i − 1, n))
(cid:107)ξ(i, n) − ξ(i − 1, n)(cid:107) ,
xn =
n = j, k.
(cid:90) 1
(cid:113)
By setting γ(p) = (1 − p)s ± pδ xj, where sign depends on
whether the braid goes "over" or "under," it is possible to
determine δ in the non-rectangular plane directly as
δr =
(±δ xj − s)
(cid:62) M (γ(p)) (±δ xj − s) dp.
0
i,d(t) = γj
(cid:13)(cid:13)(cid:13)DTi,j(γj
With this information, vj can be found in S r
i,j for interactions
between agents j and k. Note that a reparameterization of the
path can now be set equivalent to the desired trajectory of
agent j by setting γj
i,r(pj(t)) (see (5) below). Thus,
the braid controller parameter velocity vq
j (t) for agent j in
S q
i,j will be given by vq
This strategy was implemented in simulation over the
curved region illustrated in Fig. 11. In the figure, agents are
performing the mixing strategy given to them by a braid of
length 80 on the curved region (top) and simultaneously on the
straightened rectangular region (bottom). The parameters used
for this simulation were δjk = 7.7 cm and vmax = 1.5 m/s
∀j, k, and T = 30 s.
i,d(t)) γj
(cid:13)(cid:13)(cid:13).
j (t) =
i,d(t)
IV. IMPLEMENTING BRAIDS
Section II approached the problem of multi-robot mixing
from an execution level, where given a specification it is
possible to synthesize a braid controller that satisfies the
specification. In this section we consider the implementation
level, that is, we address how to find controllers that are
implementable on actual robotic systems that follow the braid
controllers found in the previous section. We then validate the
framework by implementing a mixing strategy on a team of
six robots.
A. Optimal Tracking Controller
We will now utilize the braid parameter velocity to find the
braid controller for the agents. By integrating (2) we obtain
the braid parameterization of the path γj
i
(cid:40) t−ti−1
(ti−ti−1)
(ti−ti−1)
t−¯ti
∆±δ
∆
∆ + ∆±δ
∆∓δ
2∆
t ∈ (ti−1, ¯ti]
t ∈ (¯ti, ti] .
(5)
pj(t) =
(a) The curved region will be mapped to the rectangular region of
known width and height where control design will take place.
(b) The bijective transformation T maps points in the rectangle to
points in the quadrilateral. Both shapes are defined by the braid points
which determine the corners.
Fig. 10. Rectangular and non-rectangular regions.
∆
1
∆
vj(t) =
1
(cid:16) ∆−δ
(cid:16) ∆+δ
if t ∈ (ti−1, ¯ti] ,
if t ∈ (¯ti, ti] .
region. For agent j, at braid step i, this was given by
(cid:17)
(cid:17)
ti−ti−1
ti−ti−1
where it is expected that T −1
i,j (x (ti−1)) = ξr(i − 1, j) and
T −1
i,j (x (ti)) = ξr(i, j). In the expression, ∆ corresponds to
the length of the geometric path agent j must follow to move
between ξr(i − 1, j) and ξr(i, j), while δ corresponds to the
distance along the path agent j switches velocities in order to
avoid collisions. Note that for a given parameterization of the
i,j with parameter p ∈ [0, 1], the
geometric path γj
arclength ∆ may be computed as follows
γj(cid:62)
i,r (p) γj
0
∆ =
i,r(p) dp.
If the geometric curve is directly given in S q
that the curve in S r
T −1
i,j (γj
i,j as γj
i,j may be parameterized as γj
i (p)), then the arclength may be computed by
i (p) such
i,r(p) =
i,r(p) in S r
(cid:90) 1
(cid:113)
(cid:90) 1
(cid:114)(cid:16)
∆ =
0
(cid:17)(cid:62) M(γj
(cid:16)
γj
i (p)
(cid:16)
i (p)) γj
i (p) dp.
(cid:17)(cid:17)(cid:62)
(cid:16)
(cid:17)
i,j
i,j
i,j
DT −1
where M(γj
γj
γj
i (p)
i (p)
i (p)) =
is the Jacobian of T −1
and DT −1
i,j . In the special case
where the geometry is given by straight lines, then ∆ =
(cid:107)ξr(i, j) − ξr(i − 1, j)(cid:107).
DT −1
Finally, it is of interest to find δ in order to avoid collisions.
i,j is known,
i,j as before
i (p) of the curve in S q
If the parameterization γj
then the safety separation ball may be set in S q
and the safety separation distance may be computed as
(cid:90) b
(cid:114)(cid:16)
δ =
a
(cid:17)(cid:62)M(γj
γj
i (p)
i (p)) γj
i (p) dp.
where [a, b] ⊂ [0, 1], and if the agent is meant to braid "over"
(resp. "under") then γj
i (a) corresponds to the point along the
curve where the agent enters the safety separation region (resp.
where the two curves intersect) and γj
i (b) corresponds to the
10
(a) Simulation Time 5.04 seconds.
(b) Simulation Time 9.57 seconds.
Fig. 11. Five agents performing a mixing strategy given by a braid of length 80. The top left plot represents the actual curved region the agents are mixing
in and the top right is a close-up of the agents. The lower plot is the virtual "straightened" rectangular region where the design of the braid took place.
(cid:17)
(cid:17)T
(cid:16)
(cid:90) ti
(cid:16)
1
2
Assume that agents have single integrator dynamics, i.e., xj =
uj with yj = xj ∈ R2. The agent's controller will be found
by optimally tracking the reparameterized path γj
i (pj(t)) to
minimize the cost
xj − γj
xj − γj
i
i
Q
ti−1
+ uT
j Ruj dτ (6)
J(uj) =
for Q = QT (cid:31) 0 and R = RT (cid:31) 0, with constraints
xj = uj, xj(ti−1) = γj
i (1). Using
the standard variational argument together with Pontryagin's
minimum principle, the first order necessary conditions for
optimality tell us that the optimal tracking controller u∗
j is
given by
i (0), and xj(ti) = γj
where λ is the so-called costate and satisfies
j = −R−1λj
u∗
(cid:16)
λj = −Q
xj − γj
i
(cid:17)
with unknown terminal condition λj(ti). Suppose that simi-
larly to [31, Chapter 5.3] we can construct λj as an affine
combination of the unknown λj(ti) and the state, i.e.,
λj(t) = H(t)xj(t) + K(t)λj(ti) + E(t)
and similarly, the terminal state as
xj(ti) = ξ(i, j) = F (t)xj(t) + G(t)λj(ti) + D(t)
for some yet unknown functions H, K, E, F, G, and D. One
can differentiate these equations and manipulate the equations
to obtain
(cid:16)
HR−1H − Q − H
(cid:17)
(cid:16)
(cid:16)
xj +
+
HR−1K − K
HR−1E + Qγj
(cid:17)
(cid:16)
F R−1K − G
λj(ti)
F R−1E − D
+
(cid:17)
(cid:17)
λj(ti)
i − E
(cid:17)
= 0
= 0.
and(cid:16)
(cid:17)
xj +
(cid:16)
F R−1H − F
In order to satisfy these equations for any value of xj(t) and
λj(ti), the terminal conditions, and after noticing that F =
K T , we obtain that
H = HR−1H − Q,
K = HR−1K,
G = K T R−1K,
E = HR−1E + Qγj
i ,
D = K T R−1E,
H(ti) = 02×2
K(ti) = I2
G(ti) = 02×2
E(ti) = 02×1
D(ti) = 02×1.
(7)
Note that G(ti) = 0 and G(t) (cid:23) 0 for all t, which suggests
that G(t) (cid:22) 0 for t < ti. If the problem is not abnormal, i.e.,
there exists a neighboring minimum solution, then G will be
invertible at some t < ti. In particular, by solving backwards
in time in the sequence H → K → E → G → D up to
t = ti−1, we can find that
G−1 (ti−1)(cid:0)ξ (i, j) − K T (ti−1) ξ (i − 1, j) − D (ti−1)(cid:1)
j (xj, t) = −R−1(cid:104)
resulting in the feedback optimal trajectory tracking control
law
u∗
H(t)xj(t) + K(t)G−1(ti−1)
λj (ti) =
ξ(i, j)
(cid:16)
− K T (ti−1)ξ(i − 1, j) − D(ti−1)
+ E(t)
(8)
(cid:17)
(cid:105)
where H, K, E, G and D are the solutions to terminal value
problems in (7). which can be solved numerically backwards
from ti. As it turns out, these conditions are also sufficient for
optimality as presented in the following theorem.
(cid:104)
(cid:18) 1
J(u∗) =
Theorem 3: The tracking controller in (8) is a minimizer to
the cost functional (6) whose optimal value is given by
ξ(i − 1, j)
(cid:19)
(cid:105)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)t=ti−1
+ ξ(i − 1, j)T(cid:0)KG−1 (ξ(i, j) − D) + E(cid:1) + ϕ
H − KG−1K T
ξ(i − 1, j)T
2
01234560123Track (Real World)East (m)North (m)3.5400.20.40.60.8Zoomed in Track (Real World)East (m)North (m)246810121400.20.40.6Straightened Track (Virtual World)Counter Clockwise along Path (m, along Center of Path)From Outer to InnerTrack Boundary (m)01234560123Track (Real World)East (m)North (m)55.25.45.65.80.40.60.811.2Zoomed in Track (Real World)East (m)North (m)246810121400.20.40.6Straightened Track (Virtual World)Counter Clockwise along Path (m, along Center of Path)From Outer to InnerTrack Boundary (m) 11
(a) The robots start at the beginning of the braid.
(b) Collision-free -- robots get as close as δ.
Fig. 12. Actual robots executing the mixing strategy in (10). The geometric paths and spatio-temporal constraints are being projected on the workspace with
an overhead projector for the sake of visualization.
(c) The controller is braid point feasible -- braid points are reached simultaneously.
where ϕ(t) is the solution to the terminal boundary problem
(cid:16)
(cid:17)T
γj
i (t)
Qγj
i (t)
1
2
(Λj(t))T R−1Λj(t) − 1
ϕ =
2
ϕ(ti) = −ξ(i, j)T Λj(ti)ξ(i, j)
with
Λj(t) = E(t)+
K(t)G−1(ti−1)(ξ(i, j) − K T (ti−1)ξ(i − 1.j) − D(ti−1)).
Proof:
As the control law was derived from the necessary conditions
for optimality, we only need to show that it is sufficient for
optimality. We will do so by leveraging the Hamilton-Jacobi-
Bellman theorem [32, Chapter 2].
Note that the optimization problem is regular as there exists
a uj that allows the Hamiltonian to achieve a minimum with
respect to it, i.e.,
H(xj, uj, λj)
(cid:20)(cid:16)
(cid:17)T
(cid:1)T
(cid:0)uj + R−1λj
xj − γj
i
Q
(cid:17)
(cid:16)
R(cid:0)uj + R−1λj
xj − γj
i
+ uT
=
=
1
2
1
2
j Ruj
(cid:21)
(cid:1) − 1
+ λT
j uj
(cid:16)
(cid:17)T
j R−1λj
λT
xj − γj
Q
2
i
(cid:17)
(cid:16)
+
1
2
xj − γj
i
which attains a minimum with respect to uj when
j = −R−1λj.
u∗
1
0
)
m
(
h
t
r
o
N
−1
−2
−1
0
East (m)
1
2
12
)
m
(
e
c
n
a
t
s
i
D
0.5
0.4
0.3
0.2
0.1
0
min ≈ 0.132867m
0
10
20
30
Time (s)
40
50
60
(a) Robot trajectories in the plane.
(b) Instantaneous minimum inter-robot distance.
Fig. 13. Data associated with robotic implementation in Fig. 12.
Define V (z, t) as
−zT(cid:16)
V (z, t) =
1
2
zT H(t)z + zT K(t)G−1(ti−1)ξ(i, j)
K(t)G−1(ti−1)(cid:0)K T (ti−1)ξ(i−1, j)−D(ti−1)(cid:1)+E(t)
(cid:17)
in order to obtain the fully closed-loop optimal
controller
j (xj, t) = −R−1(cid:2)(cid:0)H(t) − K(t)G−1(t)K T (t)(cid:1) xj(t)
+K(t)G−1(t) (ξ(i, j) − D(t)) + E(t)(cid:3) .
u∗
tracking
(9)
+ ϕ(t)
B. Robotic Implementation
where ϕ(t) is as defined above. It can be verified that
V (z, t) satisfies the terminal condition V (ξ(i, j), ti) = 0, that
∂V (z,t)
= λj, and that is satisfies the Hamilton-Jacobi-
(cid:12)(cid:12)(cid:12)T
In order to validate the above results in a practical setting,
the braid controllers were implemented on a team of Khepera
III differential-drive robots, which may be modelled using
unicycle dynamics, i.e.,
νj sin θj
,
θj = ωj.
where xj ∈ R2 is the robot's position and θj its heading.
The single integrator control uj from (9) can be mapped to
unicycle dynamics as
(cid:3)T
xj =(cid:2)νj cos θj
νj =(cid:2)cos θj
ωj =
(cid:3) · uj,
(cid:26) κ [ − sin θj cos θj ] · uj(cid:107)uj(cid:107) ,
κ [ − sin θj cos θj ] · uj,
sin θj
if (cid:107)uj(cid:107) > 1
otherwise
for tuning gain κ > 0. Fig. 12 illustrates the team of 6 real
robots executing the braid string
σ = {σ1·σ3·σ5}·σ2·σ3·σ4·{σ3·σ5}·{σ2·σ4}·σ1
(10)
where the braids grouped in braces are executed simultane-
ously. A visual representation of the braid string geometry
is being projected onto the robot workspace via an overhead
projector mounted on the ceiling of the lab space. Fig.
13b illustrates the minimum pairwise distance throughout the
execution of the braid. It can be seen that
the minimum
distance achieved is approximately 0.132867m -- greater than
the size of the robots which is approximately 0.13m. Fig. 13a
illustrates the trajectories taken by the robots and how they
follow the strand geometry, which were selected as straight
lines, while still accommodating for the agents' unicycle
dynamics.
∂z
z=xj
Bellman equation, i.e.,
0 =
∂V (z, t)
∂t
xj, u∗
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)z=xj
+ H
j (xj, t) ,
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)T
z=xj
∂V (z, t)
∂z
.
As a consequence, u∗
and
j is a minimizer to the cost functional
J(u∗) = V (ξ(i − 1, j), ti−1).
(cid:4)
As a final note, the terminal costate value λj(ti) was com-
puted using the initial conditions for the problem. However,
as the gains involved are solved from terminal conditions,
the choice of initial conditions is arbitrary, and evaluating at
t = ti−1 results in control law (8) being open-loop in the
terminal costate value. This could yield undesired results under
the influence of disturbances and errors. To alleviate this, we
can rewrite the terminal costate as a function of the current
state value instead, i.e.,
λj(ti) =
G−1 (ti−1)(cid:0)ξ (i, j) − K T (ti−1) ξ (i − 1, j) − D (ti−1)(cid:1)
= G−1 (t)(cid:0)ξ (i, j) − K T (t) xj (t) − D (t)(cid:1)
V. CONCLUSION
The notion of multi-robot mixing is presented as a frame-
work with which to characterize rich movement patterns in
a multi-robot system symbolically in terms of inter-robot
interactions. Two controllers are presented which allow to the
multi-robot system to achieve provably safe interactions, in the
sense that agents get close without colliding, while meeting
certain spatio-temporal constraints. Under these controllers, it
is possible to obtain theoretical bounds on the mixing limit, the
greatest level of mixing or inter-robot interactions achievable
in a given region. Further, tracking controllers that optimally
follow desired geometries are provided which can be mapped
to classes of nonholonomic dynamical systems. The concepts
presented here are validated on an actual multi-robot system
executing a interaction patterns with a desired mixing level.
Not found in this paper is the question of how to select the
interaction pattern that should be executed by these controllers.
This question addresses the specification level of our symbolic
approach. This discussion will be presented in a different paper
where motion tasks are encoded in linear temporal logic,
allowing for syntactically rich specifications at a high level,
such as "agents 1 and 3 need to collected data before arriving
at a goal location," which are then translated into a mixing
strategy encoded as a braid.
ACKNOWLEDGMENT
This work is partially supported under a grant from the Air
Force Office of Scientific Research.
The authors would like to thank the members of the Georgia
Robotics and Intelligent Systems Laboratory (GRITS Lab) for
the many fruitful discussions.
REFERENCES
[1] T. Balch, "The impact of diversity on performance in multi-robot
foraging," in Proceedings of the third annual conference on Autonomous
Agents. ACM, 1999, pp. 92 -- 99.
[2] K. Sugawara, T. Kazama, and T. Watanabe, "Foraging behavior of
interacting robots with virtual pheromone," in Intelligent Robots and
Systems, 2004. (IROS 2004). Proceedings. 2004 IEEE/RSJ International
Conference on, vol. 3, Sept 2004, pp. 3074 -- 3079 vol.3.
[3] M. Bonert, L. Shu, and B. Benhabib, "Motion planning for multi-
robot assembly systems," International Journal of Computer Integrated
Manufacturing, vol. 13, no. 4, pp. 301 -- 310, 2000.
[4] W. Nguyen and J. K. Mills, "Multi-robot control for flexible fixtureless
assembly of flexible sheet metal auto body parts," in Robotics and
Automation, 1996. Proceedings., 1996 IEEE International Conference
on, vol. 3.
IEEE, 1996, pp. 2340 -- 2345.
[5] R. Gross, M. Bonani, F. Mondada, and M. Dorigo, "Autonomous self-
assembly in swarm-bots," Robotics, IEEE Transactions on, vol. 22, no. 6,
pp. 1115 -- 1130, 2006.
[6] E. Klavins, "Programmable self-assembly," Control Systems, IEEE,
vol. 27, no. 4, pp. 43 -- 56, 2007.
[7] S. M. Pedersen, S. Fountas, H. Have, and B. Blackmore, "Agricultural
robotssystem analysis and economic feasibility," Precision agriculture,
vol. 7, no. 4, pp. 295 -- 308, 2006.
[8] N. Noguchi, J. Will, J. Reid, and Q. Zhang, "Development of a master --
slave robot system for farm operations," Computers and Electronics in
agriculture, vol. 44, no. 1, pp. 1 -- 19, 2004.
13
[9] R. C. Shah, S. Roy, S. Jain, and W. Brunette, "Data mules: Modeling
and analysis of a three-tier architecture for sparse sensor networks," Ad
Hoc Networks, vol. 1, no. 2, pp. 215 -- 233, 2003.
[10] I. Vasilescu, K. Kotay, D. Rus, M. Dunbabin, and P. Corke, "Data
collection, storage, and retrieval with an underwater sensor network," in
Proceedings of the 3rd international conference on Embedded networked
sensor systems. ACM, 2005, pp. 154 -- 165.
[11] W. Zhao, M. Ammar, and E. Zegura, "Controlling the mobility of
multiple data transport ferries in a delay-tolerant network," in INFO-
COM 2005. 24th annual joint conference of the IEEE computer and
communications societies. Proceedings IEEE, vol. 2.
IEEE, 2005, pp.
1407 -- 1418.
[12] A. Howard, "Multi-robot simultaneous localization and mapping using
particle filters," The International Journal of Robotics Research, vol. 25,
no. 12, pp. 1243 -- 1256, 2006.
[13] N. E. Ozkucur and H. L. Akın, "Cooperative multi-robot map merging
using fast-slam," in RoboCup 2009: Robot Soccer World Cup XIII.
Springer, 2010, pp. 449 -- 460.
[14] R. M. Murray, "Recent research in cooperative control of multivehicle
systems," Journal of Dynamic Systems, Measurement, and Control, vol.
129, no. 5, pp. 571 -- 583, 2007.
[15] X. C. Ding, A. R. Rahmani, and M. Egerstedt, "Multi-uav convoy
protection: an optimal approach to path planning and coordination,"
Robotics, IEEE Transactions on, vol. 26, no. 2, pp. 256 -- 268, 2010.
[16] R. Hall, "Efficient spiral search in bounded spaces," IEEE Transactions
on Pattern Analysis and Machine Intelligence, vol. 4, no. 2, pp. 208 -- 215,
March 1982.
[17] W. Huang, "Optimal
line-sweep-based decompositions for coverage
algorithms," in IEEE International Conference on Robotics and Automa-
tion, 2001.
[18] R. Hashemi, L. Jin, G. Anderson, E. Wilson, and M. Clark, "A
comparison of searchpatterns for cooperative robots operating in remote
environment," in International Conference on Information Technology:
Coding and Computing, 2001.
[19] J. Kim, F. Zhang, and M. Egerstedt, "Simultaneous cooperative explo-
ration and networking based on Voronoi diagrams," in IFAC Workshop
on Networked Systems, October 2009.
[20] B. Tovar, L. Freda, and S. M. LaValle, "Using a robot to learn geometric
information from permutations of landmarks," in Contemporary Math-
ematics. American Mathematical Society, 2007, vol. 438, pp. 33 -- 45.
[21] E. Artin, "Theory of braids," Annals of Mathematics, vol. 48, no. 1, pp.
101 -- 126, 1947.
[22] J. S. Birman, Braids, links, and mapping class groups.
Princeton
University Press, 1975, no. 82.
[23] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins, and G. J. Pap-
pas, "Symbolic planning and control of robot motion," IEEE Robotics
and Automation Magazine, vol. 14, no. 1, pp. 61 -- 70, March 2007.
[24] M. Egerstedt, X. Hu, and A. Stotsky, "Control of mobile platforms using
a virtual vehicle approach," Automatic Control, IEEE Transactions on,
vol. 46, no. 11, pp. 1777 -- 1782, 2001.
[25] T. Lozano-Perez, "Spatial planning: A configuration space approach,"
Computers, IEEE Transactions on, vol. C-32, no. 2, pp. 108 -- 120, Feb
1983.
[26] R. Ghrist, "Configuration spaces and braid groups on graphs in robotics,"
AMS IP STUDIES IN ADVANCED MATHEMATICS, vol. 24, pp. 29 -- 40,
2001.
[27] V. Kurlin et al., "Computing braid groups of graphs with applications to
robot motion planning," Homology, Homotopy and Applications, vol. 14,
no. 1, pp. 159 -- 180, 2012.
[28] Y. Diaz-Mercado and M. Egerstedt, "Multi-robot mixing using braids,"
in Decision and Control (CDC), 2013 IEEE 52nd Annual Conference
on, Dec 2013, pp. 2001 -- 2005.
[29] -- -- , "Multi-robot mixing of nonholonomic mobile robots," in Control
Applications (CCA), 2014 IEEE Conference on, Oct 2014, pp. 524 -- 529.
[30] A. Criminisi, I. Reid, and A. Zisserman, "A plane measuring device,"
Image and Vision Computing, vol. 17, no. 8, pp. 625 -- 634, 1999.
[31] A. E. Bryson, Applied optimal control: optimization, estimation and
[32] A. Locatelli, Optimal control: An introduction.
Springer Science &
control. CRC Press, 1975.
Business Media, 2001.
|
1611.01006 | 1 | 1611 | 2016-11-02T16:57:42 | Bayesian Heuristics for Group Decisions | [
"cs.MA",
"cs.SI",
"eess.SY",
"math.ST",
"stat.AP",
"math.ST"
] | We propose a model of inference and heuristic decision-making in groups that is rooted in the Bayes rule but avoids the complexities of rational inference in partially observed environments with incomplete information, which are characteristic of group interactions. Our model is also consistent with a dual-process psychological theory of thinking: the group members behave rationally at the initiation of their interactions with each other (the slow and deliberative mode); however, in the ensuing decision epochs, they rely on a heuristic that replicates their experiences from the first stage (the fast automatic mode). We specialize this model to a group decision scenario where private observations are received at the beginning, and agents aim to take the best action given the aggregate observations of all group members. We study the implications of the information structure together with the properties of the probability distributions which determine the structure of the so-called "Bayesian heuristics" that the agents follow in our model. We also analyze the group decision outcomes in two classes of linear action updates and log-linear belief updates and show that many inefficiencies arise in group decisions as a result of repeated interactions between individuals, leading to overconfident beliefs as well as choice-shifts toward extremes. Nevertheless, balanced regular structures demonstrate a measure of efficiency in terms of aggregating the initial information of individuals. These results not only verify some well-known insights about group decision-making but also complement these insights by revealing additional mechanistic interpretations for the group declension-process, as well as psychological and cognitive intuitions about the group interaction model. | cs.MA | cs |
Bayesian Heuristics for Group Decisions
?
Institute for Data, Systems, and Society (IDSS), Massachusetts Institute of Technology (MIT), Cambridge, MA 02139, USA.
Emails: rahimian , jadbabai @mit.edu
M. Amin Rahimian & Ali Jadbabaie ?
We propose a model of inference and heuristic decision making in groups that is rooted in the Bayes
rule but avoids the complexities of rational inference in partially observed environments with incomplete
information. According to our model, the group members behave rationally at the initiation of their interac-
tions with each other; however, in the ensuing decision epochs, they rely on a heuristic that replicates their
experiences from the first stage. Subsequently, the agents use their time-one Bayesian update and repeat
it for all future time-steps; hence, updating their actions using a so-called Bayesian heuristic. This model
avoids the complexities of fully rational inference and also provides a behavioral and normative foundation
for non-Bayesian updating. It is also consistent with a dual-process psychological theory of decision mak-
ing, where a controlled (conscious/slow) system develops the Bayesian heuristic at the beginning, and an
automatic (unconscious/fast) system takes over the task of heuristic decision making in the sequel.
We specialize this model to a group decision scenario where private observations are received at the
beginning, and agents aim to take the best action given the aggregate observations of all group members.
We present the implications of the choices of signal structure and action space for such agents. We show
that for a wide class of distributions from the exponential family the Bayesian heuristics take the form of
an affine update in the self and neighboring actions. Furthermore, if the priors are non-informative (and
possibly improper), then these action updates become a linear combination. We investigate the requirements
on the modeling parameters for the action updates to constitute a convex combination as in the DeGroot
model. The results reveal the nature of assumptions that are implicit in the DeGroot updating and highlights
the fragility and restrictions of such assumptions; in particular, we show that for a linear action update to
constitute a convex combination the precision or accuracy of private observations should be balanced among
all neighboring agents, requiring a notion of social harmony or homogeneity in their observational abilities.
Following the DeGroot model, agents reach a consensus asymptotically. We derive the requirements on the
signal structure and network topology such that the consensus action aggregates information efficiently. This
involves additional restrictions on the signal likelihoods and network structure. In the particular case that
all agents observe the same number of i.i.d. samples from the same distribution, then efficiency arise in
degree-regular balanced structures, where all nodes listen to and hear from the same number of neighbors.
We next shift attention to a finite state model, in which agents take actions over the probability simplex;
thus revealing their beliefs to each other. We show that the Bayesian heuristics, in this case, prescribe a
log-linear update rule, where each agent's belief is set proportionally to the product of her own and neigh-
boring beliefs. We analyze the evolution of beliefs under this rule and show that agents reach a consensus.
The consensus belief is supported over the maximizers of a weighted sum of the log-likelihoods of the initial
observations. Since the weights of the signal likelihoods coincide with the network centralities of their respec-
tive agents, these weights can be equalized in degree-regular and balanced topologies, where all nodes have
the same in and out degrees. Therefore, in such highly symmetric structures the support of the consensus
belief coincides with the maximum likelihood estimators (MLE) of the truth state; and here again, balanced
regular structures demonstrate a measure of efficiency. Nevertheless, the asymptotic beliefs systematically
reject the less probable alternatives in spite of the limited initial data, and in contrast with the optimal
(Bayesian) belief of an observer with complete information of the environment and private signals. The
latter would assign probabilities proportionally to the likelihood of every state, without rejecting any of the
possible alternatives. The asymptotic rejection of less probable alternatives indicates a case of group polar-
ization, i.e. overconfidence in the group aggregate that emerges as a result of the group interactions. Unlike
the linear action updates and the DeGroot model which entail a host of knife-edge conditions on the signal
structure and model parameters, we observe that the belief updates are unweighted; not only they effectively
internalize the heterogeneity of the private observations, but also they compensate for the individual priors.
Thence, we are led to the conclusion that multiplicative belief updates, when applicable provide a relatively
robust description of the decision making behavior.
Key words : opinion dynamics; social learning; Bayesian learning; non-Bayesian learning; rational learning;
observational learning; statistical learning; distributed learning; distributed hypothesis testing;
distributed detection; DeGroot model; linear regression; conjugate priors; exponential families
MSC2000 subject classification : Primary: 91B06; secondary: 91A35, 62C10
OR/MS subject classification : Primary: Games/group decisions: Voting/committees; secondary:
Organizational studies: Decision making Effectiveness/performance Information; Networks/graphs:
Heuristics
History : This version is dated September 12th, 2016. A preliminary version was presented in NBER-NSF
Seminar on Bayesian Inference in Econometrics and Statistics on April 30, 2016.
1
2
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
1. Introduction. Daniel Kahneman in his highly acclaimed work, "Thinking, Fast and Slow ",
points out that the proper way to elicit information from a group is not through a public discus-
sion but rather confidentially collecting each person's judgment [52, Chapter 23]. Indeed, decision
making among groups of individuals exhibit many singularities and important inefficiencies that
lead to Kahneman's noted advice. As a team converges on a decision expressing doubts about the
wisdom of the consensus choice is suppressed; subsequently teams of decision makers are afflicted
with groupthink as they appear to reach a consensus.1 The mechanisms of uncritical optimism,
overconfidence, and the illusions of validity in group interactions also lead to group polarization,
making the individuals more amenable toward extreme opinions [92].
In more abstract terms, agents in a social network exchange opinions to benefit from each other's
experience and information when making decisions about adoption of technologies, purchasing
products, voting in elections and so on. The problem of social learning is to characterize and
understand such interactions and it is a classical focus of research in microeconomics [45, Chapter
8], [39, Chapter 5], [18]. Research on formation and evolution of beliefs in social networks and
subsequent shaping of the individual and group behaviors have attracted much attention amongst
diverse communities in engineering [58, 97], statistics [66], economics [47, 65], and sociology [82].
An enhanced understanding of decision making and learning in social networks sheds light on
the role of individuals in shaping public opinion and how they influence efficiency of information
transmissions. These in turn help us improve our predictions about group behavior and provide
guidelines for designing effective social and organizational policies.
1.1. Rational social learning. The rational approach advocates application of Bayes rule to
the entire sequence of observations successively at every step. However, such repeated applications
of Bayes rule in networks become very complex, especially if the agents are unaware of the global
network structure; and as they use their local data to make inferences about all possible contin-
gencies that can lead to their observations. While some analytical properties of rational learning
is deduced and studies in the literature [32, 1, 68, 66, 42], their tractable modeling and analysis
remains an important problem in network economics and continues to attract attention.
Some of the earliest results addressing the problem of social learning are due to Banerjee (1992)
[9], and Bikhchandani et al. (1998) [13] who consider a complete graph structure where the agent's
observations are public information and also ordered in time, such that each agent has access to
the observations of all the past agents. These assumptions help analyze and explain the interplay
between public and private information leading to fashion, fads, herds etc. Later results by Gale
and Kariv [32] relax some of these assumptions by considering the agents that make simultane-
ous observations of only their neighbors rather than the whole network, but the computational
complexities limit the analysis to networks with only two or three agents. In more recent results,
Mueller-Frank [68] provides a framework of rational learning that is analytically amenable. Mossel,
Sly and Tamuz [66] analyze the problem of estimating a binary state of the world from a single
initial private signal that is independent and identically distributed amongst the agents conditioned
on the true state; and they show that by repeatedly observing each other's best estimates of the
unknown as the size of the network increases with high probability Bayesian agents asymptotically
learn the true state. Hence, the agents are able to combine their initial private observations and
learn the truth. Further results by Kanoria and Tamuz [53] provide efficient algorithms to calculate
each agent's estimate in the case of tree networks.
On the one hand, the properties of rational learning models are difficult to analyze beyond some
simple asymptotic facts such as convergence. On the other hand, these models make unrealistic
1Gar Klein proposes a famous method of project premortem to overcome the groupthink through an exercise: imagining
that the the planned decision was failed in implementation and writing a brief report of the failure [57].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
3
assumptions about the complexity and amount of computations that agents perform before com-
mitting to a decision. To avoid these shortcomings, an alternative "non-Bayesian" approach relies
on simple and intuitive heuristics that are descriptive of how agents aggregate the reports of their
neighbors before coming up with a decision.
1.2. Heuristic decision making. Heuristics are used widely in the literature to model social
interactions and decision making [34, 35, 36]. They provide tractable tools to analyze boundedly
rational behavior and offer insights about decision making under uncertainty. Hegselmann and
Krause [43] investigate various ways of averaging to model opinion dynamics and compare their
performance for computations and analysis. Using such heuristics one can avoid the complexities of
fully rational inference, and their suitability are also verified in experimental studies by Grimm and
Mengel [40] and Chandrasekhar, Larreguy and Xandri [19]. The study of such heuristics started in
1974 with the seminal work of DeGroot [20] in linear opinion pooling, where agents update their
opinions to a convex combination of their neighbors' beliefs and the coefficients correspond to the
level of confidence that each agent puts in each of her neighbors. More recently, Jadbabaie, Molavi
and Tahbaz-Salehi [47, 48, 65] consider a variation of this model for streaming observations, where in
addition to the neighboring beliefs the agents also receive private signals. Despite their widespread
applications, theoretical and axiomatic foundations of social inferences using heuristics and non-
Bayesian updates have received limited attention and only recently [65, 69], and a comprehensive
theory of non-Bayesian learning that reconciles the rational and boundedly rational approaches
with the widely used heuristics remains in demand. The main goal of our paper is to address this
gap and to do so from a behavioral perspective.
A dual process theory for the psychology of mind and its operation identifies two systems of
thinking [24]: one that is fast, intuitive, non-deliberative, habitual and automatic (system one);
and a second one that is slow, attentive, effortful, deliberative, and conscious (system two).1 Major
advances in behavioral economics are due to incorporation of this dual process theory and the
subsequent models of bounded rationality [51]. Reliance on heuristics for decision making is a
distinctive feature of system one that avoids the computational burdens of a rational evaluation;
system two on the other hand, is bound to deliberate on the options based on the available infor-
mation before making recommendations. The interplay between these two systems and how they
shape the individual decisions is of paramount importance [17].
Tversky and Kahneman argue that humans have limited time and brainpower, therefore they
rely on simple rules of thumb, i.e. heuristics, to help them make judgments. However, the use of
these heuristics causes people to make predictable errors and subjects them to various biases [96].
Hence, it is important to understand the nature and properties of heuristic decision making and
its consequences to individual and organizational choice behavior. This premise underlies many of
the recent advances in behavioral economics [93], and it motivates our work as well.
1.3. Our contribution.
In this work we are concerned with the operations of system one: we
aim to study heuristics for information aggregation in group decision scenarios when the relevant
information is dispersed among many individuals. In such situations, individuals in the group are
subjected to informational (but not strategic) externalities. By the same token, the heuristics that
are developed for decision making in such situations are also aimed at information aggregation.
In our model as the agent experiences with her environment her initial response would engage
1While many decision science applications focus on developing dual process theories of cognition and decision mak-
ing (cf. [25, 63] and the references therein); other researchers identify multiple neural systems that derive decision
making and action selection: ranging from reflexive and fast (Pavlovian) responses to deliberative and procedural
(learned) ones; and these systems are in turn supported by several motoric, perceptual, situation-categorization and
motivational routines which together comprise the decision making systems [80, Chapyter 6].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
4 h
er system two: she rationally evaluates the reports of her neighbors and use them to make a
decision. However, after her initial experience and by engaging in repeated interactions with other
group members her system one takes over her decision processes, implementing a heuristic that
imitates her (rational/Bayesian) inferences from her initial experience; hence avoiding the burden
of additional cognitive processing in the ensuing interactions with her neighbors.
On the one hand, our model of inference based on Bayesian (initial) heuristics is motivated by
the real-world behavior of people induced by their system one and reflected in their spur-of-the-
moment decisions and impromptu behavior: basing decisions only on the immediately observed
actions and disregarding the history of the observed actions or the possibility of correlations among
different observations; i.e. "what you see is all there is" [52]. On the other hand, Bayesian heuristics
offer a boundedly rational approach to model decision making over social networks. The latter is
in the sense of the word as coined by Herbert A. Simon: to incorporate modifications that lead to
substantial simplifications in the original choice problem [87].1 This in contrast with the Bayesian
approach which is not only unrealistic in the amount of cognitive burden that it imposes on the
agents, but also is often computationally intractable and complex to analyze.
Our main contribution is to offer a normative framework for heuristic decision making, by relying
on the time-one Bayesian update and using it for all future decision epochs. This model offers a
behavioral foundation for non-Bayesian updating that is compatible with the dual-process psycho-
logical theory of decision making. In Section 2, we describe the mathematical details of our model;
in particular, we explain the mathematical steps for deriving the so-called Bayesian heuristics in
a given decision scenario. Specific cases of Bayesian heuristics that we explore in the following
sections are the log-linear (multiplicative) updating of beliefs over the probability simplex, and the
linear (weighted arithmetic average) updating of actions over the Euclidean space. In Section 3, we
specialize our group decision model to a setting involving exponential family of distributions for
both signal likelihoods and agents' beliefs. The agents aim to estimate the expected values of the
sufficient statistics for their signal structures. We show that the Bayesian heuristics in this case
are affine rules in the self and neighboring actions, and we give explicit expressions for their coeffi-
cients. Subsequently, we provide conditions under which these action updates constitute a convex
combination as in the DeGroot model, with actions converging to a consensus in the latter case. We
also investigate the efficiency of the consensus action in aggregating the initial observations of all
agents across the network. Next in Section 4, we discuss a situation where agents exchange beliefs
about a truth state that can takes one of the finitely many possibilities. The Bayesian heuristics
in this case set the updated beliefs proportional to the product of self and neighboring beliefs,
as reported in every decision epoch. We investigate the evolution of beliefs under the prescribed
update rules and compare the asymptotic beliefs with that of a Bayesian agent with direct access
to all the initial observations; thus characterizing the inefficiencies of the asymptotic beliefs. We
summarize our findings from the analysis of linear action and log-linear belief updates in Section
5, where we reiterate the assumptions that are implicit in the adoption of popular aggregation
heuristics such as the DeGroot model; moreover, we discuss the inefficiencies that arise as a result
of their application. Such heuristics allow us to aggregate the information in our environment, and
provide for desirable asymptotic properties such as consensus; however, this consensus often fails as
an efficient group aggregate for the individuals' private data. We provide the mathematical proofs
and the relevant details for many of the results in the appendices at end of the paper.
1Simon advocates "bounded rationality" as compatible with the information access and the computational capacities
that are actually possessed by the agents in their environments. Most importantly he proposes the use of so-called
"satisficing" heuristrics; i.e. to search for alternatives that exceed some "aspiration levels" by satisfying a set of
minimal acceptability criteria [88, 89].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
5
2. The model. Consider a group of n agents that are labeled by [n] and interact according
to a digraph G = ([n],E).1 The neighborhood of agent i is the set of all agents whom she observes
including herself, and it is denoted by Ni = {j ∈ [n]; (j, i) ∈ E}∪{i}; every node having a self-loop:
(i, i) ∈ E for all i. We refer to the cardinality of Ni as the degree of node i and denote it by
deg(i). There is a state θ ∈ Θ that is unknown to the agents and it is chosen arbitrarily by nature
from an underlying state space Θ, which is measurable by a σ-finite measure Gθ(·). For example
if a space (Θ or S) is a countable set, then we can take its σ-finite measure (Gθ or Gs) to be
the counting measure, denoted by K(·); and if the space is a subset of Rk with positive Lebesgue
measure, then we can take its σ-finite measure to be the Lebesgue measure on Rk, denoted by
Λk(·). Associated with each agent i, Si is a measurable space called the signal space of i, and given
θ, Li(· θ) is a probability measure on Si, which is referred to as the signal structure of agent i.
Furthermore, (Ω, F ,Pθ) is a probability triplet, where Ω = S1 × . . . × Sn is a product space, and
F is a properly defined sigma field over Ω. The probability measure on Ω is Pθ(·) which assigns
probabilities consistently with the signal structures Li(· θ), i ∈ [n]; and in such a way that with θ
fixed, the random variables si, i ∈ [n] taking values in Si, are independent. These random variables
represent the private signals that agents i ∈ [n] observe at time 0. Note that the private signals are
independent across the agents. The expectation operator Eθ{·} represents integration with respect
to Pθ(dω), ω ∈ Ω.
2.1. Beliefs, actions and rewards. An agents' belief about the unknown allows her to
make decisions even as the outcome is dependent on the unknown value θ. These beliefs about the
unknown state are probability distributions over Θ. Even before any observations are made, every
agent i ∈ [n] holds a prior belief Vi(·) ∈ ∆Θ; this represents her subjective biases about the possible
values of θ. For each time instant t, let Mi,t(·) be the (random) probability distribution over Θ,
representing the opinion or belief at time t of agent i about the realized value of θ. Moreover, let
the associated expectation operator be Ei,t{·}, representing integration with respect to Mi,t(dθ).
We assume that all agents share the common knowledge of signal structures Li(·θ),∀θ ∈ Θ, their
priors Vi(·), and their corresponding sample spaces Si and Θ for all i ∈ [n]. 2
Let t ∈ N0 denote the time index; at t = 0 the values θ ∈ Θ followed by si ∈ Si of si are realized
and the latter is observed privately by each agent i for all i ∈ [n]. Associated with every agent i is
an action space Ai that represents all the choices available to her at every point of time t ∈ N0, and
a utility ui(·,·) : Ai × Θ → R which in expectation represents her von Neumann-Morgenstern pref-
erences regarding lotteries with independent draws from Ai and/or Θ. These utilities are additive
over time corresponding to successive independent draws. The utility functions and action spaces
are common knowledge amongst the agents. Subsequently, at every time t ∈ N0 each agent i ∈ [n]
chooses an action ai,t ∈ Ai and is rewarded ui(ai,t, θ).
1 Some notations: Throughout the paper, R is the set of real numbers, N denotes the set of all natural numbers,
and N0 := N ∪ {0}. For n ∈ N a fixed integer the set of integers {1, 2, . . . , n} is denoted by [n], while any other set
is represented by a capital Greek or calligraphic letter. For a measurable set X we use ∆X to denote the set of all
probability distributions over the set X . Furthermore, any random variable is denoted in boldface letter, vectors are
represented in lowercase letters and with a bar over them, measures are denoted by upper case Greek or calligraphic
Latin letters, and matrices are denoted in upper case Latin letters. For a matrix A, its spectral radius ρ(A) is the
largest magnitude of all its eigenvalues.
2The signal structures Li(·θ),∀θ ∈ Θ and the priors Vi(·), as well as the corresponding sample spaces Si and Θ are
common knowledge amongst the agents for all i ∈ [n]. The assumption of common knowledge in the case of fully
rational (Bayesian) agents implies that given the same observations of one another's beliefs or private signals distinct
agents would make identical inferences; in the sense that starting form the same belief about the unknown θ, their
updated beliefs given the same observations would be the same; in Aumann's words, rational agents cannot agree to
disagree [5].
6
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
about the value of θ and chooses her action ai,0 ←- arg maxai∈Ai RΘ ui(ai, θ)Mi,0(dθ), maximizing
2.2. Aggregation heuristics. Given si, agent i forms an initial Bayesian opinion Mi,0(·)
her expected reward. Here for a set A, we use the notation a ←- A to denote an arbitrary choice
from the elements of A that is assigned to a. Not being notified of the actual realized value for
ui(ai,0, θ), she then observes the actions that her neighbors have taken. Given her extended set of
observations {aj,0, j ∈ Ni} at time t = 1, she refines her opinion into Mi,1(·) and makes a second,
and possibly different, move ai,1 according to:
ai,1 ←- arg max
ai∈Ai ZΘ
ui(ai, θ)Mi,1(dθ),
(1)
maximizing her expected pay off conditional on everything that she has observed thus far; i.e.
maximizing Ei,1{ui(ai, θ)} = Eθ{ui(ai,1, θ)si, aj,0 : j ∈ Ni} = RΘ ui(ai, θ)Mi,1(dθ). Subsequently,
she is granted her net reward of ui(ai,0, θ) + ui(ai,1, θ) from her past two plays. Following realization
of rewards for their first two plays, in any subsequent time instance t > 1 each agent i ∈ [n] observes
the preceding actions of her neighbors aj,t−1 : j ∈ Ni and takes an option ai,t out of the set Ai. Of
particular significance in our description of the behavior of agents in the succeeding time periods
t > 1, is the relation:
fi(aj,0 : j ∈ Ni) := ai,1 ←- arg max
ai∈Ai
Ei,1{ui(ai, θ)}
(2)
derived in (1), which given the observations of agent i at time t = 0, specifies her (Bayesian) pay-off
maximizing action for time t = 1. Once the format of the mapping fi(·) is obtained, it is then
used as a heuristic for decision making in every future epoch. The agents update their action
by choosing: ai,t = fi (aj,t−1 : j ∈ Ni) ,∀t > 1. We refer to the mappings fi :Qj∈Ni Aj → Ai thus
obtained, as Bayesian heuristics.1,2,3
1The heuristics thus obtained suffer from same fallacies of snap judgments that are associated with the recommen-
dations of system one in "Thinking, Fast and Slow"; flawed judgments that rely on simplistic interpretations: "what
you see is all there is", in Kahneman's elegant words [52]. Indeed, the use of the initial Bayesian update for future
decision epochs entails a certain level of naivety on the part of the decision maker: she has to either assume that the
structure of her neighbors' reports have not departed from their initial format, or that they are not being influenced
back by her own or other group members and can thus be regarded as independent sources of information. Such
naivety in disregarding the history of interactions has been highlighted in our earlier works on Bayesian learning
without recall [75], where we interpret the use of time-one Bayesian update for future decision epochs, as a rational
but memoryless behavior: by regarding their observations as being direct consequences of inferences that are made
based on the initial priors, the agents reject any possibility of a past history beyond their immediate observations.
2Similar and related forms of naivety have been suggested in the literature. Eyster and Rabin [26, 27] propose the
autarkic model of naive inference, where players at each generation observe their predecessors but naively think that
any predecessor's action relies solely on that player's private information, thus ignoring the possibility that successive
generations are learning from each other. Bala and Goyal [7]study another form of naivety and bounded-rational
behavior by considering a variation of observational learning in which agents observe the action and pay-offs of their
neighbors and make rational inferences about the action/pay-off correspondences, based on their observations of
the neighboring actions; however, they ignore the fact that their neighbors are themselves learning and trying to
maximize their own pay-offs. Levy and Razin look at a particularly relevant cognitive bias called correlation neglect,
which makes individuals regard the sources of their information as independent [60]; they analyze its implications to
diffusion of information, and focus in particular, on the voting behavior.
3Cognitive and psychological roots of the Bayesian heuristics as aggregation rules can be traced to Anderson's seminal
theory of information integration, developed throughout 1970s and 1980s [3]. Accordingly, a so-called "value function"
assigns psychological values to each of the stimuli and these psychological values are then combined into a single
psychological (and later an observable) response through what is called the "integration function". A fundamental
assumption is that valuation can be represented at a higher (molar) level as a value on the response dimension for each
stimulus, as well as a weight representing the salience of this stimulus in the overall response. These valuations and
weights are themselves the result of integration processes in the lower (molecular) level. At the heart of information
integration theory is the "cognitive algebra" which describes the rules by which the values and weights of stimuli are
integrated into an overall response [4].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
7
3. Affine action updates, linear updating and DeGroot learning.
In this section we
explore the essential modeling features that lead to a linear structure in the Bayesian Heuristics.
We present a general scenario that involves the exponential family of distributions and leads to
linear action updates.
To describe the signal structures, we consider a measurable sample space S with a σ-finite
measure Gs(·), and a parametrized class of sampling functions {L(·θ; σi) ∈ ∆S : σi > 0 and δi > 0}
belonging to the k-dimensional exponential family as follows:
'(sθ; σi, δi) :=
dL(·θ; σi, δi)
dGs
Λk(ξ(ds))
Gs(ds)
(cid:12)
(cid:12)
(cid:12) τ (σiξ(s), δi) eσiη(θ)T ξ(s)−δiγ(η(θ)),
(cid:12)
(3)
where ξ(s) : S → Rk is a measurable function acting as a sufficient statistic for the random samples,
η : Θ → Rk is a mapping from the parameter space Θ to Rk, τ : Rk ×(0, +∞) → (0, +∞) is a positive
weighting function, and
Λk(ξ(ds))
Gs(ds)
(cid:12)
(cid:12)
(cid:12) τ (σiξ(s), δi)eσiη(θ)T ξ(s)Gs(ds),
(cid:12)
lnZs∈S
1 δ
i
γ(η(θ)) :=
is a normalization factor that is constant when θ is fixed, even though δi > 0 and σi > 0
vary. This normalization constant for each θ is uniquely determined by the functions η(·),
ξ(·) and τ (·). The parameter space Θ and the mapping η(·) are such that the range space
Ωθ := {η(θ) : θ ∈ Θ} is an open subset of the natural parameter space Ωη := {η ∈ Rk :
Rs∈S Λk(ξ(ds))/Gs(ds)τ (ξ(s), 1) eηT ξ(s)Gs(ds) < ∞}. In (3), σi > 0 and δi > 0 for each i are scal-
ing factors that determine the quality or informativeness of the random sample si with regard to
the unknown θ: fixing either one of the two factors σi or δi, the value of the other one increases
with the increasing informativeness of the observed value ξ(si). The following conjugate family of
priors1 are associated with the likelihood structure (3). This family is determined uniquely by the
transformation and normalization functions: η(·) and γ(·), and it is parametrized through a pair
of parameters (α, β), α ∈ Rk and β > 0:
ν(θ; α, β) :=
dV(·; α, β)
Fγ,η :=(cid:26)V(θ; α, β) ∈ ∆Θ, α ∈ Rk, βi > 0 :
(cid:12)
(cid:12)
(cid:12)
(cid:12)
κ(α, β) := Zθ∈Θ(cid:12)(cid:12)(cid:12)(cid:12)
=(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)
(cid:12) eη(θ)T α−βγ(η(θ))Gθ(dθ) < ∞(cid:27).
(cid:12)
(cid:12)
dGθ
Λk(η(dθ))
Gθ(dθ)
Λk(η(dθ))
Gθ(dθ)
eη(θ)T α−βγ(η(θ))
,
κ(α, β)
mi,θ), ∀a ∈ Ai = Rk, where mi,θ := Ei,θ{ξ(si)} :=Rs∈S
Furthermore, we assume that agents take actions in Rk, and that they aim for a minimum variance
estimation of the regression function or conditional expectation (given θ) of the sufficient statistic
ξ(si). Hence, we endow every agent i ∈ [n] with the quadratic utility ui(a, θ) = −(a − mi,θ)T (a −
Our main result in this section prescribes a scenario in which each agent starts from a prior
belief V(·; αi, βi) belonging to Fγ,η and she observes a fixed number ni of i.i.d. samples from the
distribution L(· θ; σi, δi). The agents then repeatedly communicate their actions aimed at minimum
variance estimation of mi,θ. These settings are formalized under the following assumption that we
term the Exponential Family Signal-Utility Structure.
ξ(s)L(dsθ; σi, δi) ∈ Rk.
1Consider a parameter space Θ, a sample space S, and a sampling distribution L(·θ) ∈ ∆S, θ ∈ Θ. Suppose that s
is a random variable which is distributed according to L(·θ) for any θ. A family F ⊂ ∆Θ is a conjugate family for
L(·θ), if starting from any prior distribution V(·) ∈ F and for any signal s ∈ S, the posterior distribution given the
observation s = s belongs to F.
= σi(cid:12)(cid:12)(cid:12)(cid:12)
σi(cid:12)(cid:12)(cid:12)(cid:12)
8
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
Assumption 1 (Exponential family signal-utility structure).
(i) Every agent i ∈ [n] observes ni i.i.d. private samples si,p, p ∈ [ni] from the common sample
space S and that the random samples are distributed according to the law L(·θ; σi, δi) given
by (3) as a member of the k-dimensional exponential family.
(ii) Every agent starts from a conjugate prior Vi(·) = V(·; αi, βi) ∈ Fγ,η, for all i ∈ [n].
(iii) Every agent chooses actions a ∈ Ai = Rk and bears the quadratic utility ui(a, θ) = −(a −
mi,θ)T (a− mi,θ), where mi,θ := Ei,θ{ξ(si)} :=Rs∈S
ξ(s)L(dsθ; σi, δi) ∈ Rk.
The Bayesian heuristics fi(·), i ∈ [n] under the settings prescribed by the exponential family
signal-utility structure (Assumption 1) are linear functions of the neighboring actions with specified
coefficients that depend only on the likelihood structure parameters: ni, σi and δi as well as the
prior parameters: αi and βi, for all i ∈ [n].1
Theorem 1 (Affine action updates). Under the exponential family signal-utility structure
specified in Assumption 1, the Bayesian heuristics describing the action update of every agent i ∈ [n]
Tijaj,t−1 + i, where for all i, j ∈ [n] the constants Tij
and δi are as follows:
are given by: ai,t = fi(aj,t−1 : j ∈ Ni) =Pj∈Ni
Tij =
δiσj(nj + δ−1
j βj)
σi(βi +Pp∈Ni
npδp)
, i = −
δi
σi(βi +Pp∈Ni
npδp) Xj∈Ni\{i}
αj.
1,t, . . . , aT
The action profile at time t is the concatenation of all actions in a column vector: at =
(aT
n,t)T . The matrix T with entries Tij, i, j ∈ [n] given in Theorem 1 is called the social
influence matrix. The constant terms i in this theorem appear as the rational agents attempt to
compensate for the prior biases of their neighbors when making inferences about the observations
in their neighborhood; we denote = (T
n )T and refer to it as the vector of neighborhood
biases. The evolution of action profiles under conditions of Theorem 1 can be specified as follows:
at+1 = (T ⊗ Ik)at + , where Ik is the k × k identity matrix and (T ⊗ Ik) is a Kronecker product.
Subsequently, the evolution of action profiles over time follows a non-homogeneous positive linear
discrete-time dynamics, cf. [28]. If the spectral radius of T is strictly less than unity: ρ(T ) < 1, then
I − T is non-singular; there is a unique equilibrium action profile given by ae = ((I − T )−1 ⊗ Ik)
and limt→∞ at = ae. If unity is an eigenvalue of T , then there may be no equilibrium action profiles
or an infinity of them. If ρ(T ) > 1, then the linear discrete-time dynamics is unstable and the action
profiles may grow unbounded in their magnitude, cf. [49].
1 , . . . , T
Example 1 (Gaussian Signals with Gaussian Beliefs). Mossel and Tamuz [67] consider
the case where the initial private signals as well as the unknown states are normally distributed
and the agents all have full knowledge of the network structure. They show that by iteratively
observing their neighbors' mean estimates and updating their beliefs using Bayes rule all agents
converge to the same belief. The limiting belief is the same as what a Bayesian agent with direct
access to everybody's private signals would have hold; and furthermore, the belief updates at each
step can be computed efficiently and convergence occurs in a number of steps that is bounded
in the network size and its diameter. These results however assume complete knowledge of the
network structure by all the agents.
Here, we consider the linear action updates in the Gaussian setting. Let Θ = R be the param-
eter space associated with the unknown parameter θ ∈ Θ. Suppose that each agent i ∈ [n] holds
1Some of the non-Bayesian update rules have the property that they resemble the replication of a first step of a
Bayesian update from a common prior. For instance, DeMarzo, Vayanos and Zwiebel [21] interpret the weights in
the DeGroot model as those assigned initially by rational agents to the noisy opinions of their neighbors based on
their perceived precision. However, by repeatedly applying the same weights over and over again, the agents ignore
the need to update these weights and to account for repetitions in their information sources (the so-called persuasion
bias); as one of our main objectives, we formalize this setup as a Bayesian heuristic.
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
9
i
i
i = δ−1
i
and variance β−1
onto a Gaussian prior belief with mean αiβ−1
; here, γ(θ) = θ2/2 and η(θ) =
θ. Further suppose that each agent observes an independent private Gaussian signal si with
mean θ and variance σ−1
, for all i ∈ [n]; hence, ξ(si) = si and τ (σiξ(si), δi) = τ (σisi, σi) =
(2π/σi)−1/2 exp(σis2
i /2). After observing their private signals all engage in repeated communica-
tions with their neighbors. Finally, we assume that each agents is trying to estimate the mean
mi,θ = Ei,θ{si} of her private signal with as little variance as possible. Under the prescribed set-
ting, Theorem 1 applies and the Bayesian heuristic update rules are affine with the coefficients as
specified in the theorem with ni = 1 and σi = δi for all i. In particular, if αi = (0, . . . , 0) ∈ Rk and
βi → 0 for all i, then i = 0 for all i and the coefficients Tij = σj/Pp∈Ni
σp > 0 specify a convex
combination:Pj∈Ni
Example 2 (Poisson signals with gamma beliefs). As the second example, suppose that
each agent observe ni i.i.d. Poisson signals si,p : p ∈ [ni] with mean δiθ, so that Θ = Ai = (0, +∞)
for all i ∈ [n]. Moreover, we take each agent's prior to be a Gamma distribution with parameters
αi > 0 and βi > 0, denoted Gamma(αi,βi):
Tij = 1 for all i.
νi(θ) :=
dVi
dΛ1
=
αi
βi
Γ(αi)
θαi−1e−βiθ,
for all θ ∈ (0,∞) and each i ∈ [n]. Note that here η(θ) = log θ, γ(η(θ)) = exp(η(θ)) = θ, κ(αi, βi) =
si,p
Γ(αi)βi−αi, mi,θ = δiθ, ξ(si,p) = si,p, σi = 1 and τ (σiξ(si,p), δi) = δ
i /(si,p!), for all i, p. This setting
corresponds also to a case of Poisson observers with common rate θ and individual exposures
δi, i ∈ [n], cf. [33, p. 54]. The posterior distribution over Θ after observing of the sum of ni Poisson
p=1 si,p + αi
and niδi + βi, [33, pp. 52–53]. Using a quadratic utility −(a− δiθ)2, the expected pay-off at time zero
is maximized by the δi-scaled mean of the posterior Gamma belief distribution [33, p. 587]: ai,0 =
j )aj,0 − αj,
[(nj +
βjδ−1
njδj. The subsequent optimal action at time 1 and the resultant
Bayesian heuristics are as claimed in Theorem 1 with σi = 1 for all i ∈ [n]. Here if we let αi, βi → 0
np > 0 again
Tij = 1 for all i as in the DeGroot model. In the following
two subsections, we shall further explore this correspondence with the DeGroot updates and the
implied asymptotic consensus amongst the agents.
mean δiθ samples is again a Gamma distribution with updated (random) parametersPni
δi(Pni
p=1 si,p + αi)/(niδi + βi). Given the neighboring informationPnj
∀j ∈ Ni, agent i can refine her belief into a Gamma distribution with parameters αi +Pj∈Ni
j )aj,0 − αj] and βi +Pj∈Ni
and δi = δ > 0 for all i, then i = 0 for all i and the coefficients Tij = nj/Pp∈Ni
specify a convex combination: Pj∈Ni
p=1 sj,p = (nj + βjδ−1
3.1. Linear updating and convergence.
In general, the constant terms i in Theorem 1
depend on the neighboring prior parameters αj, j ∈ Ni \ {i} and can be non-zero. Accumulation
of constant terms over time when ρ(T ) ≥ 1 prevents the action profiles from converging to any
finite values or may cause them to oscillate indefinitely (depending upon the model parameters).
However, if the prior parameters are vanishingly small, then the affine action updates in Theorem
1 reduce to linear update and i = 0. This requirement on the prior parameters is captured by our
next assumption.
Assumption 2 (Non-informative priors). For a member V(·; α, β) of the conjugate family
Fγ,η we denote the limit limαi,βi→0 V(·; α, β) by V∅(·) and refer to it as the non-informative (and
improper, if V∅(·) 6∈ Fγ,η) prior. 1 All agents start from a common non-informative prior: Vi(·) =
V∅(·), ∀i.
1Conjugate priors offer a technique for deriving the prior distributions based on the sample distribution (likelihood
structures). However, in lack of any prior information it is impossible to justify their application on any subjective
basis or to determine their associated parameters for any agent. Subsequently, the use of non-informative priors is
suggested by Bayesian analysts and various techniques for selecting non-informative priors is explored in the literature
[55]. Amongst the many proposed techniques for selecting non-informative priors, Jeffery's method sets its choice
10
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
As the name suggest non-informative priors do not inform the agent's action at time 0 and
the optimal action is completely determined by the observed signal si,p : p ∈ [ni] and its likelihood
structure, parameterized by σi and δi. If we let αi, βi → 0 in the expressions of Tij and i from
Theorem1, then the affine action updates reduce to linear combinations and the preceding corollary
is immediate.
Corollary 1 (Linear updating). Under the exponential
family signal-utility structure
(Assumption 1) with non-informative priors (Assumption 2); the Bayesian heuristics describe
each updated action ai,t as a linear combination of the neighboring actions aj,t−1, j ∈ Ni: ai,t =
npδp).
Pj∈Ni
Tijaj,t−1, where Tij = δiσjnj/(σiPp∈Ni
The action profiles under Corollary 1 evolve as a homogeneous positive linear discrete-time system:
at+1 = (T ⊗ Ik)at and if the spectral radius of T is strictly less than unity, then limt→∞ at = 0. For a
strongly connected social network with Tii > 0 for all i the Perron-Frobenius theory [85, Theorems
1.5 and 1.7] implies that T has a simple positive real eigenvalue equal to ρ(T ). Moreover, the
left and right eigenspaces associated with ρ(T ) are both one-dimensional with the corresponding
eigenvectors l = (l1, . . . , ln)T and r = (r1, . . . , rn)T , uniquely satisfying klk2 = krk2 = 1, li > 0, ri > 0,
i=1 liri = 1. The magnitude of any other eigenvalue of T is strictly less than ρ(T ). If
ρ(T ) = 1, then limt→∞ at = limt→∞(T t ⊗ Ik)a0 = (rl
⊗ Ik) a0; in particular, the asymptotic action
profile may not represent a consensus although every action converges to some point within the
convex hull of the initial actions {ai,0, i ∈ [n]}. If ρ(T ) > 1, then the linear discrete-time dynamics
is unstable and the action profiles may increase or decrease without bound; pushing the decision
outcome to extremes, we can associate ρ(T ) > 1 to cases of polarizing group interactions.
∀i and Pn
T
3.2. DeGroot updates, consensus and efficiency.
In order for the linear action updates
in Corollary 1 to constitute a convex combination as in the DeGroot model,1 we need to introduce
some additional restrictions on the likelihood structure of the private signals.
Assumption 3 (Locally balanced likelihoods). The likelihood structures given in (3) are
δjnj) /Pj∈Ni
called locally balanced if for all i ∈ [n], (δi/σi) = (Pj∈Ni
structures of every agent i and her neighborhood to satisfy: δiPj∈Ni
Assumption 3 signifies a local balance property for the two exponential family parameters σi
and δi and across every neighborhood in the network. In particular, we need for the likelihood
δjnj. Since
parameters σi and δi are both measures of accuracy or precision for private signals of agent i,
the balance condition in Assumption 3 imply that the signal precisions are spread evenly over the
agents; i.e. the quality of observations obey a rule of social balance such that no agent is in a
σjnj = σiPj∈Ni
σjnj.
proportional to the square root of Fisher's information measure of the likelihood structure [81, Section 3.5.3], while
Laplace's classical principle of insufficient reason favors equiprobability leading to priors which are uniform over the
parameter space.
1The use of linear averaging rules for modeling opinion dynamics has a long history in mathematical sociology and
social psychology [30]; their origins can be traced to French's seminal work on"A Formal Theory of Social Power"
[29]. This was followed up by Harary's investigation of the mathematical properties of the averaging model, including
the consensus criteria, and its relations to Markov chain theory [41]. This model was later generalized to belief
exchange dynamics and popularized by DeGroot's seminal work [20] on linear opinion pools. In engineering literature,
the possibility to achieve consensus in a distributed fashion (through local interactions and information exchanges
between neighbors) is very desirable in a variety of applications such as load balancing [6], distributed detection and
estimation [15, 95, 54], tracking [56], sensor networks and data fusion [98, 99], as well as distributed control and
robotics networks [46, 64]. Early works on development of consensus algorithms originated in 1980s with the works
of Tsitsiklis et.al [94] who propose a weighted average protocol based on a linear iterative approach for achieving
consensus: each node repeatedly updates its value as a weighted linear combination of its own value and those received
by its neighbors.
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
11
position of superiority to everyone else. Indeed, fixing δi = δ for all i, the latter condition reduces
to a harmonic property for the parameters σi, when viewed as a function of their respective nodes
(cf. [61, Section 2.1] for the definition and properties of harmonic functions):
σi =Xj∈Ni
njσjPk∈Ni
, δi = δ,∀i.
nk
(4)
However, in a strongly conceded social network (4) cannot hold true unless σi is a constant: σi = σ
for all i. Similarly, when σi = σ is a constant, then under Assumption 3 δi is spread as a harmonic
function over the network nodes, and therefore can only take a constant value δi = δ for all i, cf.
[61, Section 2.1, Maximum Principle]. In particular, fixing either of the parameters σi or δi for all
agents and under the local balance condition in Assumption 3, it follows that the other parameter
should be also fixed across the network; hence, the ratio σi/δi will be a constant for all i. Later
when we consider the efficiency of consensus action we introduce a strengthening of Assumption 3,
called globally balanced likelihood (cf. Assumption 4), where the ratio δi/σi should be a constant
for all agents across the network. Examples 1 and 2 above provide two scenarios in which the
preceding balancedness conditions may be satisfied: (i) having σi = δi for all i, as was the case
with the Gaussian signals in Example 1, erasures that the likelihoods are globally balanced; (ii)
all agents receiving i.i.d. signals from a common distribution in Examples 2 (Poisson signals with
the common rate θ and common exposure δ) makes a case for likelihoods being locally balanced.
Theorem 2 (DeGroot updating and consensus). Under the exponential family signal-
utility structure (Assumption 1), with non-informative priors (Assumption 2) and locally balanced
likelihoods (Assumption 3); the updated action ai,t is a convex combination of the neighboring
Tij = 1 for all i. Hence, in a strongly connected
social network the action profiles converge to a consensus, and the consensus value is a convex
combination of the initial actions ai,0 : i ∈ [n].
actions aj,t−1, j ∈ Ni: ai,t =Pj∈Ni
Tijaj,t−1,Pj∈Ni
In light of Theorem 2, it is of interest to know if the consensus action agrees with the min-
imum variance unbiased estimator of mi,θ given all the observations of every agent across the
network, i.e. whether the Bayesian heuristics efficiently aggregate all the information amongst the
networked agents. Our next result addresses this question. For that to hold we need to introduce
a strengthening of Assumption 3:
Assumption 4 (Globally balanced likelihoods). The likelihood structures given in (3) are
called globally balanced if for all i ∈ [n] and some common constant C > 0, δi/σi = C.
In particular, under Assumption 4, σiδj = σjδi for all i, j, and it follows that the local balance
of likelihoods is automatically satisfied. We call the consensus action efficient if it coincides with
the minimum variance unbiased estimator of mi,θ for all i and given all the observations of every
agent across the network. Our next result indicates that global balance is a necessary condition
for the agents to reach consensus on a globally optimal (efficient) action. To proceed, let the
network graph structure be encoded by its adjacency matrix A defined as [A]ij = 1 ⇐⇒ (j, i) ∈ E,
and [A]ij = 0 otherwise. To express the conditions for efficiency of consensus, we need to consider
the set of all agents who listen to the beliefs of a given agent j; we denote this set of agents by
N out
:= {i ∈ [n] : [I + A]ij = 1} and refer to them as the out-neighborhood of agent j. This is in
contrast to her neighborhood Nj, which is the set of all agents whom she listens to. Both sets Nj
and N out
include agent j as a member.
Theorem 3 ((In-)Efficiency of consensus). Under the exponential
family signal-utility
structure (Assumption 1) and with non-informative priors (Assumption 2); in a strongly connected
social network the agents achieve consensus at an efficient action if, and only if, the likelihoods are
npδp, for all i and j. The efficient consensus action is
j
j
globally balanced andPp∈N out
then given by a? =Pn
j=1(cid:16)δjnjaj,0/Pn
npδp =Pp∈Ni
p=1 npδp(cid:17).
j
12
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
Following our discussing of Assumption 3 and equation (4), we pointed out that if either of
the two parameters σi and δi that characterize the exponential family distribution of (3) are held
fixed amongst the agents, then the harmonicity condition required for the local balancedness of
the likelihoods implies that the other parameter is also fixed for all the agents. Therefore the local
balancedness in many familiar cases (see Example 2) restricts the agents to observing i.i.d. signals:
allowing heterogeneity only in the sample sizes, but not in the distribution of each sample. This
special case is treated in our next corollary, where we also provide the simpler forms of the Bayesian
heuristics and their linearity coefficients in the i.i.d. case:
j
Corollary 2 (DeGroot learning with i.i.d. samples). Suppose that each agent i ∈ [n]
observes ni i.i.d. samples belonging to the same exponential family signal-utility structure (Assump-
tion 1 with σi = σ and δi = δ for all i). If the agents have non-informative priors (Assump-
tion 2) and the social network is strongly connected social network, then following the Bayesian
Tijaj,t−1,
np, and reach a consensus. The consensus action is efficient if, and only
np for all i and j, and the efficient consensus action is given by a? =
heuristics agents update their action according to the linear combination: ai,t =Pj∈Ni
where Tij = nj/Pp∈Ni
if, Pp∈N out
np =Pp∈Ni
j=1(cid:16)aj,0nj/Pn
p=1 np(cid:17) .
Pn
It is notable that the consensus value pinpointed by Theorem 2 does not necessarily agree with
the MVUE of mi,θ given all the private signals of all agents across the network; in other words,
by following Bayesian heuristics agents may not aggregate all the initial data efficiently. As a
simple example, consider the exponential family signal-utility structure with non-informative priors
(Assumptions 1 and 2) and suppose that every agent observes an i.i.d. sample from a common
distribution L(·θ; 1, 1). In this case, the action updates proceed by simple iterative averaging:
aj,t−1 for all i ∈ [n] and any t ∈ N. For an undirected graph G it is well-known
that the asymptotic consensus action following simple iterative averaging is the degree-weighted
i=1(deg(i)/E)ai,0, cf. [38, Section II.C]; and the consensus action is different form the
i=1 ai,0 unless the social network is a regular graph in which case,
ai,t = (1/Ni)Pj∈Ni
averagePn
global MVUE a? = (1/n)Pn
tion for efficiency of consensus,Pp∈N out
j
Remark 1 (Efficiency of Balanced Regular Structures).
deg(i) = d is fixed for all i, and E = n.d.
In general, if we assume
that all agents receive the same number of i.i.d. samples from the same distribution, then the condi-
np, is satisfied for balanced regular structures.
In such highly symmetric structures, the number of outgoing and incoming links are the same for
every node and equal to fixed number d.
np =Pp∈Ni
Our results shed light on the deviations from the globally efficient actions, when consensus is
being achieved through the Bayesian heuristics. This inefficiency of Bayesian heuristics in globally
aggregating the observations can be attributed to the agents' naivety in inferring the sources of
their information, and their inability to interpret the actions of their neighbors rationally, [38];
in particular, the more central agents tend to influence the asymptotic outcomes unfairly. This
sensitivity to social structure is also due to the failure of agents to correct for the repetitions in
the sources of the their information: agent i may receive multiple copies that are all influenced by
the same observations from a far way agent; however, she fails to correct for these repetition in the
sources of her observations, leading to the co-called persuasion bias, [21].1
1Sobel [90] provides a theoretical framework to study the interplay between rationality and group decisions and points
out the subsequent inefficiencies in information aggregation. The seminal work of Janis [50] provides various examples
involving the American foreign policy in the mid-twentieth where the desire for harmony or conformity in the group
have resulted in bad group decisions, a phenomenon that he coins groupthink. Various other works have looked at the
choice shift toward more extreme options [23, 91] and group polarization [44, 83].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
13
4. Log-Linear belief updates. When the state space Θ is finite, the action space is the
probability simplex and the agents bear a quadratic utility that measures the distance between
their action and the point mass on the true state, the communication structure between the agents
is rich enough for them to reveal their beliefs at every time period. The Bayesian heuristics in
this case lead to a log-linear updating of beliefs similar to what we analyze in [73, 75] under the
Bayesian without recall model; the chief question of interest to the latter works is whether the
agents, after being exposed to sequence of private observations and while communicating with each
other, can learn the truth using the Bayesian without recall update rules. 1 The learning framework
of [73, 74, 75, 78] in which agents have access to an stream of new observations is in contrast with
the group decision model of this paper; the difference being in the fact that here the agents have a
single initial observation and engage in group decision making to come up with the best decision
that aggregates their individual private data with those of the other group members.
Consider an environment where the state space is a finite set of cardinality m and agents take
actions over the (m− 1)-simplex of probability measures while trying to minimize their distance to
a point mass on the true state. Specifically, let Θ = {θ1, . . . , θm}, Ai = ∆Θ and for any probability
measure M(·) ∈ Ai with probability mass function µ(·) := dM/dK, suppose that
ui(M, θj) = −(1− µ(θj))2 −
µ(θk)2.
m
X
k=1,
k6=j
Ei,t{ui(M, θ)}, and the agents proceed by truthfully announc-
Subsequently, Mi,t = arg maxM∈∆Θ
ing their beliefs to each other at every time step. In particular, if we denote the belief probability
mass functions νi(·) := dVi/dK and µi,t(·) := dMi,t/dK for all t, then we can follow the steps of [77]
to derive the Bayesian heuristic fi in (2) by replicating the time-one Bayesian belief update for all
future time-steps:
µi,t(θ) =
for all θ ∈ Θ and at any t > 1.2,3
µj,t−1(θ)
µi,t−1(θ)(cid:16)Qj∈Ni\{i}
Pθ∈Θ µi,t−1(θ)(cid:16)Qj∈Ni\{i}
νj (θ) (cid:17)
νj (θ) (cid:17) ,
µj,t−1(θ)
(5)
1Naivety of agents in these cases impedes their ability to learn; except in simple social structures such as cycles or
rooted trees (cf. [74]). Rahimian, Shahrampour and Jadbabaie [78] show that learning in social network with complex
neighborhood structures can be achieved if agents choose a neighbor randomly at every round and restrict their belief
update to the selected neighbor each time. In the literature on theory of learning in games [31], log-linear learning
refers to a class of randomized strategies, where the probability of each action is proportional to an exponential of the
difference between the utility of taking that action and the utility of the optimal choice. Such randomized strategies
combine in a log-linear manner [14], and they have desirable convergence properties: under proper conditions, it can
be shown that the limiting (stationary) distribution of action profiles is supported over the Nash equilibria of the
game [62].
2In writing (5), every time agent i regards each of her neighbors j ∈ Ni as having started from some prior belief
νj(·) and arrived at their currently reported belief µj,t−1(·) upon observing their private signals, hence rejecting any
possibility of a past history, or learning and correlation between their neighbors. Such a rule is of course not the
optimum Bayesian update of agent i at any step t > 1, because the agent is not taking into account the complete
observed history of beliefs and is instead, basing her inference entirely on the initial signals and the immediately
observed beliefs.
3It is notable that the Bayesian heuristic in (5) has a log-linear structure. Geometric averaging and logarithmic opinion
pools have a long history in Bayesian analysis and behavioral decision models [37, 84] and they can be also justified
under specific behavioral assumptions [65]. The are also quite popular as a non-Bayesian update rule in engineering
literature for addressing problems such as distributed detection and estimation [86, 79, 71, 59, 8]. In [8] the authors use
14
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
4.1. An algebra of beliefs. Both the linear action updates studied in the previous chapter
as well as the weighted majority update rules that arise in the binary case and are studied in
[76, 75] have a familiar algebraic structure. It is instructive to develop similar structural properties
for belief updates in (5) and over the space ∆Θ, i.e. the points of the standard (m − 1)-simplex.
Given two beliefs µ1(·) and µ2(·) over Θ we denote their "addition" as
µ1 ⊕ µ2(θ) =
µ1(θ)µ2(θ)
Pθ∈Θ µ1(θ)µ2(θ)
.
Indeed, let ∆Θo denote the (m − 1)-simplex of probability measure over Θ after all the edges are
excluded; ∆Θo endowed with the ⊕ operation, constitutes a group (in the algebraic sense of the
word). It is easy to verify that the uniform distribution ¯µ(θ) = 1/Θ acts as the identity element
for the group; in the sense that ¯µ ⊕ µ = µ for all µ ∈ ∆Θo, and given any such µ we can uniquely
identify its inverse as follows:
µinv(θ) =
1/µ(θ)
.
Pθ∈Θ 1/µ(θ)
Moreover, the group operation ⊕ is commutative and we can thus endow the abelian group (∆Θo,⊕)
with a subtraction operation:
µ1 (cid:9) µ2(θ) = µ1 ⊕ µinv
2 (θ) =
µ1(θ)/µ2(θ)
Pθ∈Θ µ1(θ)/µ2(θ)
.
We are now in a position to rewrite the Bayesian heuristic for belief updates in terms of the group
operations ⊕ and (cid:9) over the simplex interior:
µi,t = ⊕j∈Ni
µj,t−1 (cid:9)j∈Ni\{i}
νj.
The above belief update has a structure similar to the linear action updates studied in (1): the
agents incorporate the beliefs of their neighbors while compensating for the neighboring priors to
isolate the observational parts of the neighbors' reports. A key difference between the action and
belief updates is in the fact that action updates studied in Section 3 are weighted in accordance
with the observational ability of each neighbor, whereas the belief updates are not. Indeed, the
quality of signals are already internalized in the reported beliefs of each neighbor; therefore there
is no need to re-weight the reported beliefs when aggregating them.
a logarithmic opinion pool to combine the estimated posterior probability distributions in a Bayesian consensus filter;
and show that as a result: the sum of KullbackLeibler divergences between the consensual probability distribution
and the local posterior probability distributions is minimized. Minimizing the sum of KullbackLeibler divergences as
a way to globally aggregate locally measured probability distributions is proposed in [11, 10] where the corresponding
minimizer is dubbed the KullbackLeibler average. Similar interpretations of the log-linear update are offered in [70]
as a gradient step for minimizing either the KullbackLeibler distance to the true distribution, or in [72] as a posterior
incorporation of the most recent observations, such that the sum of KullbackLeibler distance to the local priors is
minimized; indeed, the Bayes' rule itself has a product form and the Bayesian posterior can be characterized as the
solution of an optimization problem involving the KullbackLeibler divergence to the prior distribution and subjected
to the observed data [100]. In the past, we have investigated the implications of such log-linear behavior and properties
of convergence and learning when agents are exposed to a stream of private observation [73, 74, 77, 78].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
15
Given the abelian group structure we can further consider the "powers" of each element µ2 =
µ ⊕ µ and so on; in general for each inetger n and any belief µ ∈ ∆Θo, let the n-th power of µ be
denoted by n(cid:12) µ := µn, defined as follows:1
µn(θ) =
µn(θ)
.
Pθ∈Θ µn(θ)
([I + A]ij (cid:12) µj,t) (cid:9)j∈[n]
Using the ⊕ and (cid:12) notations, as well as the adjacency matrix A we get:
µi,t+1 = ⊕j∈Ni
µj,t (cid:9)j∈Ni\{i}
νj = ⊕j∈[n]
([A]ij (cid:12) νj).
(9)
With some abuse of notation, we can concatenate the network beliefs at every time t into a column
vector µt = (µ1,t, . . . , µn,t)T and similarly for the priors ν = (ν1, . . . , νn)T ; thus (9) can be written
in the vectorized format by using the matrix notation as follows:
µt = {(I + A)(cid:12) µt−1}(cid:9){A(cid:12) ν}
Iterating over t and in the common matrix notation we obtain:
µt =n(I + A)t (cid:12) µ0o(cid:9)n(Pt
τ =0(I + A)τ A)(cid:12) νo.
The above is key to understanding the evolution of beliefs under the Bayesian heuristics in (5), as
we will explore next. In particular, when all agents have uniform priors νj = ¯µ for all j, then (10)
and (11) simplify as follows: µt = (I + A) (cid:12) µt−1 = (I + A)t (cid:12) µ0. This assumption of a common
uniform prior is the counterpart of Assumption 1 (non-informative priors) in Subsection 3.1, which
paved the way for transition from affine action updates into linear ones. In the case of beliefs over
a finite state space Θ, the uniform prior ¯µ is non-informative. If all agents start form common
uniform priors, the belief update in (5) simplifies as follows:
(10)
(11)
µi,t(θ) = Qj∈Ni
Pθ∈ΘQj∈Ni
µj,t−1(θ)
µj,t−1(θ)
.
(12)
Our main focus in the next section is to understand how the individual beliefs evolve under (5),
or (12) which is a spacial case of (5). The gist of our analysis is encapsulated in the group theoretic
iterations: µt = (I + A)t (cid:12) µ0, derived above for the common uniform priors case. In particular,
our understanding of the increasing matrix powers (I + A)t plays a key role. When the network
graph G is strongly connected, the matrix I + A is primitive. The Perron-Frobenius theory [85,
Theorems 1.5 and 1.7] implies that I + A has a simple positive real eigenvalue equal to its spectral
radius ρ(I + A) = 1 + ρ, where we adopt the shorthand notation ρ := ρ(A). Moreover, the left and
right eigenspaces associated with this eigenvalue are both one-dimensional and the corresponding
eigenvectors can be taken such that they both have strictly positive entries. The magnitude of
any other eigenvalue of I + A is strictly less than 1 + ρ. Hence, the eigenvalues of I + A denoted
by λi(I + A), i ∈ [n], can be ordered in their magnitudes as follows: λn(I + A) ≤ λn−1(I + A) ≤
1This notation extends to all real numbers n ∈ R, and it is easy to verify that the following distributive properties are
satisfied:
n(cid:12) (µ1 ⊕ µ2) = (n(cid:12) µ1)⊕ (n(cid:12) µ2),
(m + n)(cid:12) µ1 = (m(cid:12) µ1)⊕ (n(cid:12) µ1),
(m.n)(cid:12) µ1 = m(cid:12) (n(cid:12) µ1),
for all m, n ∈ R and µ1, µ2 ∈ ∆Θo.
16
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
(13)
! .
T i
(λi(I + A)/(1 + ρ))tril
n
X
i=2
(I + A)t = (1 + ρ)t r1α
T
+
. . . < λ1(I + A) = 1 + ρ. Subsequently, we can employ the eigendecomposition of (I + A) to analyze
the behavior of (I + A)t+1. Specifically, we can take a set of bi-orthonormal vectors li, ri as the
left and right eigenvectors corresponding to the ith eigenvalue of I + A, satisfying: klik2 = krik2
rj = 0, i 6= j; in particular, the left eigenspace associated with ρ is
= 1, l
one-dimensional with the corresponding eigenvector l1 = α = (α1, . . . , αn)T , uniquely satisfying
i=1 αi = 1, αi > 0, ∀i ∈ [n], and αT A = (ρ + 1)αT . The entry αi is called the centrality of agent i
and as the name suggests, it measures how central is the location of agent in the network. We can
now use the spectral representation of A to write [49, Section 6]:
ri = 1 for all i and l
Pn
T i
T i
Asymptotically, we get that all eigenvalues other than the Perron-Frobenius eigenvalue 1 + ρ are
subdominant; hence, (I + A)t → (1 + ρ)tr1α
(cid:12) µ0 as t → ∞; the latter holds
true for the common uniform priors case and also in general, as we shall see next.
and µt = (1 + ρ)tr1α
T
T
4.2. Becoming certain about the group aggregate. We begin our investigation of the
evolution of beliefs under (12) by considering the optimal response (belief) of an agents who has
been given access to the set of all private observations across the network; indeed, such a response
can be achieved in practice if one follows Kahneman's advice and collect each individual's infor-
mation privately before combining them or allowing the individuals to engage in public discussions
[52, Chapter 23]. Starting from the uniform prior and after observing everybody's private data our
aggregate belief about the truth state is given by the following implementation of the Bayes rule:
µ?(θ) = Qj∈Ni
Pθ∈ΘQj∈Ni
'j(sjθ)
'j(sjθ)
.
(14)
as opposed to the global (network-wide) likelihood maximizer Θ?
Our next theorem describes the asymptotic outcome of the group decision process when the agents
report their beliefs and follow the Bayesian heuristic (12) to aggregate them. The outcome indi-
cated in Theorem 4 departs from the global optimum µ? in two major respects. Firstly, the
agents reach consensus on a belief that is supported over Θ♦ := arg maxθ∈ΘPn
i=1 αi log('i(siθ)),
:= arg maxθ∈Θ µ?(θ) =
arg maxθ∈ΘPn
i=1 log('i(siθ)); note that the signal log-likelihoods in the case of Θ♦ are weighted
by the centralities, αi, of their respective nodes. Secondly, the consensus belief is concentrated
uniformly over Θ♦, its support does not include the entire state space Θ and those states which
score lower on the centrality-weighted likelihood scale are asymptotically rejected as a candidate
for the truth state; in particular, if {θ♦} = Θ♦ is a singlton, then the agents effectively become
certain about the truth state of θ♦, in spite of their essentially bounded aggregate information and
in contrast with the rational (optimal) belief µ? that is given by the Bayes rule in (14) and do not
discredit or reject any of the less probable states. This unwarranted certainty in the face of limited
aggregate data is a manifestation of the group polarization effect that derive the agent to more
extreme beliefs, rejecting the possibility of any alternatives outside of Θ♦.
Theorem 4 (Certainty about the group aggregate). Following the Bayesian heuristic
belief updates in (5), limt→∞ µi,t(θ) = 1/Θ♦ for all i ∈ [n] and any θ ∈ Θ?, where Θ♦ :=
arg maxθ∈ΘPn
i=1 αi log('i(siθ)). In particular, if the sum of signal log-likelihoods weighted by node
centralities is uniquely maximized by θ♦, i.e. {θ♦} = Θ♦, then limt→∞ µi,t(θ♦) = 1 almost surely for
all i ∈ [n].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
17
Remark 2 (Efficiency of balanced regular networks). The fact that log-likelihoods
in Θ♦ are weighted by the node centralities is a source of inefficiency for the asymptotic outcome of
the group decision process. This inefficiency is warded off in especially symmetric typologies, where
in and out degrees of all nodes in the network are the same. In these so-called balanced regular
digraphs, there is a fixed integer d such that all agents receive reports from exactly d agents, and
also send their reports to some other d agents; d-regular graphs are a special case, since all links
are bidirectional and each agent sends her reports to and receive reports from the same d agents.
In such structures α = (1/n)1 so that Θ? = Θ♦ and the support of the consensus belief identifies
the global maximum likelihood estimator (MLE); i.e. the maximum likelihood estimator of the
unknown θ, given the entire set of observations from all agents in the network.
5. Conclusions. We propose the Bayesian heuristics framework to address the problem of
information aggregation and decision making in groups. Our model is consistent with the dual
process theory of mind with one system developing the heuristics through deliberation and slow
processing, and another system adopting the heuristics for fast and automatic decision making:
once the time-one Bayesian update is developed, it is used as a heuristic for all future decision
epochs. On the one hand, this model offers a behavioral foundation for non-Bayesian updating; in
particular, linear action updates and log-linear belief updates. On the other hand, its deviation from
the rational choice theory captures common fallacies of snap-judgments and history neglect that
are observed in real life. Our behavioral method also complements the axiomatic approaches which
investigate the structure of belief aggregation rules and require them to satisfy specific axioms such
as label neutrality and imperfect recall, as well as independence or separability for log-linear and
linear rules, respectively [65].
We showed that under a natural quadratic utility and for a wide class of distributions from the
exponential family the Bayesian heuristics correspond to a minimum variance Bayes estimation
with a known linear structure. If the agents have non-informative priors, and their signal structures
satisfy certain homogeneity conditions, then these action updates constitute a convex combination
as in the DeGroot model, where agents reach consensus on a point in the convex hull of their
initial actions. In case of belief updates (when agents communicate their beliefs), we showed that
the agents update their beliefs proportionally to the product of the self and neighboring beliefs.
Subsequently, their beliefs converge to a consensus supported over a maximum likelihood set, where
the signal likelihoods are weighted by the centralities of their respective agents.
Our results indicate certain deviations from the globally efficient outcomes, when consensus is
being achieved through the Bayesian heuristics. This inefficiency of Bayesian heuristics in globally
aggregating the observations is attributed to the agents' naivety in inferring the sources of their
information, which makes them vulnerable to structural network influences: the share of centrally
located agents in shaping the asymptotic outcome is more than what is warranted by the quality
of their data. Another source of inefficiency is in the group polarization that arise as a result of
repeated group interactions; in case of belief updates, this is manifested in the structure of the
(asymptotic) consensus beliefs. The latter assigns zero probability to any alternative that scores
lower than the maximum in the weighted likelihoods scale: the agents reject the possibility of less
probable alternatives with certainty, in spite of their limited initial data. This overconfidence in
the group aggregate and shift toward more extreme beliefs is a key indicator of group polarization
and is demonstrated very well by the asymptotic outcome of the group decision process.
We pinpoint some key differences between the action and belief updates (linear and log-linear,
respectively): the former are weighted updates, whereas the latter are unweighted symmetric
updates. Accordingly, an agent weighs each neighbor's action differently and in accordance with
the quality of their private signals (which are inferred from the actions). On the other hand, when
communicating their beliefs the quality of each neighbor's signal is already internalized in their
18
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
reported beliefs; hence, when incorporating her neighboring beliefs, an agent regards the reported
beliefs of all her neighbors equally and symmetrically. Moreover, in the case of linear action updates
the initial biases are amplified and accumulated in every iteration. Hence, the interactions of biased
agents are very much dominated by their prior beliefs rather than their observations. This issue can
push their choices to extremes, depending on the aggregate value of their initial biases. Therefore, if
the Bayesian heuristics are to aggregate information from the observed actions satisfactorily, then
it is necessary for the agents to be unbiased, i.e. they should hold non-informative priors about the
state of the world and base their actions entirely on their observations. In contrast, when agents
exchange beliefs with each other the multiplicative belief update can aggregate the observations,
irrespective of the prior beliefs. The latter are asymptotically canceled; hence, multiplicative belief
updates are robust to the influence of priors.
The Bayesian heuristics approach is strongly motivated by the behavioral processes that underlie
human decision making. These processes often deviate from the predictions of the rational choice
theory, and our investigation of the Bayesian heuristics highlights both the mechanisms for such
deviations and their ramifications. In our ongoing research, we expand this behavioral approach by
incorporating additional cognitive biases such as inattentiveness, and investigate how the decision
processes are affected. On the one hand, the obtained insights highlight the value of educating
the public about benefits of rational decision making and unbiased judgment, and how to avoid
common cognitive errors when making decisions. On the other hand, by investigating the effect of
cognitive biases, we can improve the practice of social and organizational policies, such that new
designs can accommodate commonly observed biases, and work well in spite of them.
Appendix A: Proof of Theorem 1.
If agent i starts from a prior belief Vi(·) = V(·; αi, βi) ∈
Fγ,η, then we can use the Bayes rule to verify that, cf. [81, Proposition 3.3.13], the Radon-Nikodym
derivative of the Bayesian posterior of agent i after observing ni samples si,p ∈ S, p ∈ [ni], with
p=1 ξ(si,p), βi + niδi), and in particular the Bayesian posterior at time
likelihood (3) is ν(·; αi + σiPni
zero belongs to the conjugate family Fγ,η: Mi,0(·) = V(·; αi + σiPni
Subject to the quadratic utility ui(a, θ) = −(a − mi,θ)T (a − mi,θ), the expected pay-off at any
time time t is maximized is by choosing [12, Lemma 1.4.1]:
p=1 ξ(si,p), βi + niδi).
ai,t = Ei,t{mi,θ} :=Zθ∈Θ
mi,θMi,t(dθ),
which coincides with her minimum variance unbiased estimator (Bayes estimate) for mi,θ. The
members of the conjugate family Fγ,η satisfy the following linearity property of the Bayes estimates
that is key to our derivations.1
Lemma 1 (Proposition 3.3.14 of [81]). Let ζ ∈ Rk be a parameter and suppose that the
parameter space Ωζ is an open set in Rk. Suppose further that ζ ∈ Ωζ has the prior distribution
W(·; α, β) with density κ0(α, β)eζT α−βγ(ζ) w.r.t. Λk where κ0(α, β) is the normalization constant. If
s0 ∈ S0 ⊂ Rk is a random signal with distribution D(·; ζ) and density τ 0(s)eζts−γ0(ζ) w.r.t. Λk, then
.
α β
sD(ds; ζ)W(dζ; α, β) =
Zζ∈Ωζ Zs0∈S0
1In fact, such an affine mapping from the observations to the Bayes estimate characterizes the conjugate family Fγ,η
and every member of this family can be uniquely identified from the constants of the affine transform [22].
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
19
Hence for any V(·; α, β) ∈ Fγ,η we can write
(cid:12)
(cid:12)
(cid:12)
(cid:12)
Zθ∈Θ
mi,θV(dθ; α, β) = Zθ∈Θ Zs∈S
= Zθ∈Θ(cid:12)(cid:12)(cid:12)(cid:12)
ξ(s)σ(cid:12)(cid:12)(cid:12)(cid:12)
Zs∈S
= Zζ∈Ωθ (cid:12)(cid:12)(cid:12)(cid:12)
ξ(s)σ(cid:12)(cid:12)(cid:12)(cid:12)
Zs∈S
= Zζ∈Ωη
Λk(η(dθ))
Gθ(dθ)
Λk(ξ(ds))
Gs(ds)
Λk(η(dθ))
Gθ(dθ)
Λk(ξ(ds))
Gs(ds)
δ γ0(ζ)
eζT α− β
κ(α, β)
αδ
σβ
ξ(s)L(dsθ; σ, δ)V(dθ; α, β)
eη(θ)T α−βγ(η(θ))
κ(α, β)
Gθ(dθ)×
κ(α, β)
Gθ(dθ)×
eη(θ)T α−βγ(η(θ))
(cid:12)
(cid:12)
(cid:12) τ (σξ(s), δ) eση(θ)T ξ(s)−δγ(η(θ))Gs(ds)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12)
(cid:12) τ (σξ(s), δ) eση(θ)T ξ(s)−δγ(η(θ))Gs(ds)
(cid:12)
Λk(dζ) Zs0∈S0
eζT s0−γ0(ζ)Λk(ds0)
s0τ 0(s0)
σ
=
,
(15)
where in the penultimate equality we have employed the following change of variables: ζ = η(θ),
s0 = σξ(s), γ0(ζ) = δγ(ζ), τ 0(s0) = τ (s0, δ); and the last equality is a direct application of Lemma 1.
p=1 ξ(si,p), βi + niδi), the expectation maximizing action
at time zero coincides with:
In particular, given Mi,0(·) = V(·; αi + σiPni
ai,0 =Pni
p=1 ξ(si,p) + σ−1
i αi
ni + δ−1
i βi
.
(16)
Subsequently, following her observations of aj,0, j ∈ Ni and from her knowledge of her neighbor's
p=1 ξ(sj,p) for all her
neighbors. Hence, we get
priors and signal likelihood structure, agent i infers the observed values ofPnj
ξ(sj,p) = (nj + δ−1
j βj)aj,0 − σ−1
j αj,∀j ∈ Ni.
(17)
njXp=1
The observations of agent i are therefore augmented by the set of independent samples from her
p=1 ξ(sj,p) : j ∈ Ni}, and her refined belief at time 1 is again a member of the conjugate
neighbors: {Pnj
family Fγ,η and is give by:
Mi,1(·) = V(·; αi +Xj∈Ni
σj
njXp=1
ξ(sj,p), βi +Xj∈Ni
njδj).
We can again invoke the linearity of the Bayes estimate for the conjugate family Fγ,η and the
subsequent result in (15), to get that the expected pay-off maximizing action at time 1 is given by:
ai,1 =
δi(cid:16)αi +Pj∈Ni
p=1 ξ(sj,p)(cid:17)
σjPnj
njδj(cid:17)
σi(cid:16)βi +Pj∈Ni
.
(18)
Finally, we can use (17) to replace for the neighboring signals and derive the expression of the
action update of agent i at time 1 in terms of her own and the neighboring actions aj,0, j ∈ Ni;
(cid:3)
leading to the expression of linear Bayesian heuristics as claimed in Theorem 1.
20
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
that the coefficients of the linear conbination from Corollary 1 sum to one:Pj∈Ni
by setting ai,0 =Pni
Appendix B: Proof of Theorem 2. The balancedness of likelihoods (Assumption 3) ensures
Tij = 1, for all
i; thus forming a convex combination as in the DeGroot model. Subsequently, the agents begin
p=1 ξ(si,p)/ni according to (16), and at every t > 1 they update their actions
according to at = T at−1, where at = (a1,t, . . . , an,t)T and T is the n× n matrix whose i, j-th entry
is Tij. Next note from the analysis of convergence for DeGroot model, cf. [38, Proporition 1], that
for a strongly connected network G if it is aperiodic (meaning that one is the greatest common
divisor of the lengths of all its circles; and it is the case for us, since the diagonal entries of T
are all non-zero), then limτ→∞ T τ = 1s
, where s := (s1, . . . , sn)T is the unique left eigenvector
i=1 si = 1, si > 0, ∀i. Hence, starting from
non-informative priors agents follow the DeGroot update and if G is also strongly connected, then
(cid:3)
associated with the unit eigenvalue of T and satisfyingPn
they reach a consensus at sT a0 =Pn
Appendix C: Proof of Theorem 3. We begin by a lemma that determines the so-called
global MVUE for each i, i.e. the MVUE of mi,θ given all the observations of all agents across the
network.
i=1 si(Pni
p=1 ξ(si,p)/ni).
T
Lemma 2 (Global MVUE). Under the exponential family signal-utility structure (Assump-
tion 1), the (global) MVUE of mi,θ given the entire set of observations of all the agents across the
network is given by:
If we further impose non-informative priors (Assumption 2), then the global MVUE for each i can
be rewritten as
a?
i =
δi(cid:16)αi +Pn
p=1 ξ(sj,p)(cid:17)
j=1 σjPnj
j=1 njδj(cid:17)
σi(cid:16)βi +Pn
p=1 ξ(sj,p)(cid:17)
j=1 σjPnj
σi(cid:16)Pn
j=1 njδj(cid:17)
X
δi
σi
j=1
=
n
p=1 npδp
σjnjPn
δi(cid:16)Pn
a?
i =
.
(19)
aj,0.
(20)
This lemma can be proved easily by following the same steps that lead to (18) to get (19); making
the necessary substitutions under Assumption 2 yields (20). From (20), it is immediately clear that
if some consensus action is to be the efficient estimator (global MVUE) for all agents i ∈ [n], then
we need δiσj = σiδj for all i, j; hence, the global balance is indeed a necessary condition. Under this
condition, the local balance of likelihoods (Assumption 3) is automatically satisfied and given non-
informative priors Theorem 2 guarantees convergence to a consensus in strongly connected social
network. Moreover, we can rewrite (20) as a?
p=1 npδp, for all i. Hence, if
the consensus action (sT a0 in the proof of Theorem 2, Appendix B) is to be efficient then we need
j=1 njδj for all i; s = (s1, . . . , sn) being the unique normalized left eigenvector associated
with the unit eigenvalue of T : sT T = sT , as defined in Appendix B. Using δiσj = σiδj, we can also
si = δini/Pn
rewrite the coefficients Tij of the DeGroot update in Theorem 2 as Tij = δjnj/(Pp∈Ni
Therefore, by expanding the eigenvector condition sT T = sT we obtain that in order for the
consensus action sT a0 to agree with the efficient consensus a?, it is necessary and sufficient to have
that for all j
j=1 δjnjaj,0)/Pn
i = a? = (Pn
npδp).
siTij =
n
X
i=1
n
i=1 δiniPn
X
j=1 δjnj! δjnj[I + A]ij
npδp
Pp∈Ni
= sj =
,
j=1 δjnj
δjnjPn
(21)
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
21
(22)
= 1,
or equivalently,
Xi:j∈Ni
δiniPp∈Ni
npδp
= Xi∈N out
j
δiniPp∈Ni
npδp
δjnj/(Pp∈Ni
for all j. Under the global balance condition (Assumption 4), δiσj = δjσi, the weights Tij =
npδp) as given above, correspond to transition probabilities of a node-weighted ran-
dom walk on the social network graph, cf. [16, Section 5]; where each node i ∈ [n] is weighted
by wi = niδi. Such a random walk is a special case of the more common type of random walks
on weighted graphs where the edge weights determine the jump probabilities; indeed, if for any
edge (i, j) ∈ E we set its weight equal to wi,j = wiwj then the random walk on the edge-weighted
graph reduces to a random walk on the node-weighted graph with node weights wi, i ∈ [n]. If the
social network graph is undirected and connected (so that wi,j = wj,i for all i, j), then the edge-
weighted (whence also the node-weighted) random walks are time-reversible and their stationary
distributions (s1, . . . , sn)T can be calculated in closed form as follows [2, Section 3.2]:
(23)
In a node-weighted random walk we can replace wi,j = wiwj for all j ∈ Ni and (23) simplifies into
Similarly to (21), the consensus action will be efficient if and only if
wj
wi,j
si =
.
wi,j
.
wj
si = Pj∈Ni
i=1Pj∈Ni
Pn
wiPj∈Ni
Pn
i=1 wiPj∈Ni
wiPj∈Ni
wiPn
Pn
i=1 wiPj∈Ni
X
wk)Xj∈Ni
X
(wiXj∈Ni
which holds true only if Pj∈Ni
Pj∈Ni
Pi∈N out
wj =Pj∈Ni
or equivalently:
wj =
=
wj
si =
n
(
k=1
n
i=1
wj
k=1 wk
,∀i,
wj),∀i,
wj is a common constant that is the same for all agents, i.e.
δjnj = C0 > 0 for all i ∈ [n]. Next replacing in (22) yields that, in fact, C0 =
(cid:3)
δini for all j, completing the proof for the conditions of efficiency.
j
Appendix D: Proof of Theorem 4. We address the more general case of (5), which includes
(12) as a special case (with common uniform priors). For any pair of states θ and θ, define the log
ratio of beliefs, likelihoods, and priors as follows:
By concatenating the log-ratio statistics of the n networked agents, we obtain the following three
vectorizations for the log-ratio statistics:
φi,t(θ, θ) := log(cid:16)µi,t(θ)/µi,t(θ)(cid:17) ,
λi(θ, θ) := log(cid:16)'i(siθ)/'i(siθ)(cid:17) ,
γi(θ, θ) := log(cid:16)νi(θ)/νi(θ)(cid:17) ,
φt(θ, θ) := (φ1,t(θ, θ), . . . , φn,t(θ, θ)),
λ(θ, θ) := (λ1(θ, θ), . . . , λn(θ, θ)),
γ(θ, θ) := (γ1(θ, θ), . . . , γn(θ, θ)).
22
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
Using the above notation, we can rewrite the log-linear belief updates of (5) in a linearized vector
format as shown below:
(I + A)τ Aγ(θ, θ)
φt+1(θ, θ) =(I + A)φt(θ, θ)− Aγ(θ, θ)
X
τ =0
t
=(I + A)t+1φ0(θ, θ)−
=(I + A)t+1(cid:16)λ(θ, θ) + γ(θ, θ)(cid:17)−
X
=(I + A)t+1λ(θ, θ) + (I + A)t+1 −
X
τ =0
τ =0
t
t
(I + A)τ Aγ(θ, θ)
(I + A)τ A! γ(θ, θ).
Next we use the spectral decomposition in (13) to obtain: 1
φt+1(θ, θ) = (1 + ρ)t+1r1Λ(θ, θ) + ((1 + ρ)t+1 −
X
= (1 + ρ)t+1 r1Λ(θ, θ) + (1−
→ (1 + ρ)t+1r1Λ(θ, θ),
τ =0
t
(1 + ρ)τ ρ)r1β(θ, θ) + o((1 + ρ)t+1)
t
τ =0
X
(1 + ρ)τ−t−1ρ)r1β(θ, θ) + o(1)!
(24)
τ =0 ρ(1 + ρ)τ−1 = 1.
where we adopt the following notations for the global log likelihood and prior ratio statistics:
β(θ, θ) := αT γ(θ, θ) and Λ(θ, θ) := αT λ(θ, θ); furthermore, in calculation of the limit in the last
step of (24) we use the geometric summation identityP∞
To proceed denote Λ(θ) :=Pn
i=1 αi'i(siθ) so that Λ(θ, θ) = Λ(θ) − Λ(θ). Since Θ♦ consists of
the set of all maximizers of Λ(θ), we have that Λ(θ, θ) < 0 whenever θ ∈ Θ♦ and θ 6∈ Θ♦. Next recall
from (24) that φt+1(θ, θ) → (1 + ρ)t+1r1Λ(θ, θ) and r1 is the right Perron-Frobenius eigenvector
with all positive entries; hence, for all θ ∈ Θ♦ and any θ, φi,t(θ, θ) → −∞ if θ 6∈ Θ♦ and φi,t(θ, θ) = 0
whenever θ ∈ Θ♦; or equivalently, µi,t(θ)/µi,t(θ) → 0 for all θ 6∈ Θ♦, while µi,t(θ) = µi,t(θ) for any
θ ∈ Θ♦. The latter together with the fact thatP θ∈Θ µi,t(θ) = 1 for all t implies that with probability
one: limt→∞ µi,t(θ) = 1/Θ♦,∀θ ∈ Θ♦ and limt→∞ µi,t(θ) = 0,∀θ 6∈ Θ♦ as claimed in the Theorem.
In the special case that Θ♦ is a singleton, {θ♦} = Θ♦, we get that limt→∞ µi,t(θ♦) = 1 almost surely
for all i ∈ [n].
Acknowledgments. This work was supported by ARO MURI W911NF-12-1-0509.
References
[1] Acemoglu D, Dahleh MA, Lobel I, Ozdaglar A (2011) Bayesian learning in social networks. The Review
of Economic Studies 78(4):1201–1236.
[2] Aldous D, Fill J (2002) Reversible markov chains and random walks on graphs. Book in progress.
[3] Anderson NH (1971) Integration theory and attitude change. Psychological review 78(3):171.
[4] Anderson NH (1981) Foundations of information integration theory .
[5] Aumann RJ (1976) Agreeing to disagree. The annals of statistics 1236–1239.
[6] Baikerikar J, Surve S, Prabhu S (2011) Consensus based dynamic load balancing for a network of
heterogeneous workstations. Advances in Computing, Communication and Control, 116–124 (Springer).
[7] Bala V, Goyal S (1998) Learning from neighbours. The Review of Economic Studies 65(3):595–621.
1Given two functions f (·) and g(·) we use the asymptotic notation f (t) = o(g(t)) to signify the relations
limt→∞ f (t)/g(t) = 0.
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
23
[8] Bandyopadhyay S, Chung SJ (2014) Distributed estimation using bayesian consensus filtering. Amer-
ican Control Conference 634–641.
[9] Banerjee AV (1992) A simple model of herd behavior. The Quarterly Journal of Economics 107(3):797–
817.
[10] Battistelli G, Chisci L (2014) Kullback–leibler average, consensus on probability densities, and dis-
tributed state estimation with guaranteed stability. Automatica 50(3):707–718.
[11] Battistelli G, Chisci L, Morrocchi S, Papi F (2011) An information-theoretic approach to distributed
state estimation. IFAC Proceedings Volumes 44(1):12477–12482.
[12] Bickel PJ, Doksum KA (2015) Mathematical Statistics: Basic Ideas and Selected Topics, volume I
(CRC Press).
[13] Bikhchandani S, Hirshleifer D, Welch I (1998) Learning from the behavior of others: Conformity, fads,
and informational cascades. The Journal of Economic Perspectives 12(3):pp. 151–170, ISSN 08953309.
[14] Blume LE (1993) The statistical mechanics of strategic interaction. Games and Economic Behavior
5(3):387–424.
[15] Borkar V, Varaiya P (1982) Asymptotic agreement in distributed estimation. IEEE Transactions on
Automatic Control 27(3):650–655.
[16] Brightwell GR, Winkler P (1999) Graph homomorphisms and phase transitions. Journal of Combina-
torial Theory, Series B 77(2):221–262.
[17] Brocas I, Carrillo JD (2014) Dual-process theories of decision-making: A selective survey. Journal of
economic psychology 41:45–54.
[18] Chamley CP (2004) Rational Herds: Economic Models of Social Learning (Cambridge University
Press).
[19] Chandrasekhar AG, Larreguy H, Xandri JP (2015) Testing models of social learning on networks:
Evidence from a lab experiment in the field. Technical report, National Bureau of Economic Research.
[20] DeGroot MH (1974) Reaching a consensus. Journal of American Statistical Association 69:118 – 121.
[21] DeMarzo PM, Vayanos D, Zwiebel J (2003) Persuasion bias, social influence, and unidimensional
opinions. The Quarterly Journal of Economics 118:909–968.
[22] Diaconis P, Ylvisaker D (1979) Conjugate priors for exponential families. The Annals of statistics
7(2):269–281.
[23] Eliaz K, Ray D, Razin R (2006) Choice shifts in groups: A decision-theoretic basis. The American
economic review 1321–1332.
[24] Evans JSB (2003) In two minds: dual-process accounts of reasoning. Trends in cognitive sciences
7(10):454–459.
[25] Evans JSB, Stanovich KE (2013) Dual-process theories of higher cognition advancing the debate.
Perspectives on psychological science 8(3):223–241.
[26] Eyster E, Rabin M (2010) Naive herding in rich-information settings. American economic journal:
microeconomics 2(4):221–243.
[27] Eyster E, Rabin M (2014) Extensive imitation is irrational and harmful. The Quarterly Journal of
Economics qju021.
[28] Farina L, Rinaldi S (2011) Positive linear systems: theory and applications, volume 50 (John Wiley &
Sons).
[29] French Jr JR (1956) A formal theory of social power. Psychological review 63(3):181–194.
[30] Friedkin NE (1986) A formal theory of social power. Journal of Mathematical Sociology 12(2):103–126.
[31] Fudenberg D (1998) The Theory of Learning in Games (MIT Press).
[32] Gale D, Kariv S (2003) Bayesian learning in social networks. Games and Economic Behavior 45:329–
346.
[33] Gelman A, Carlin JB, Stern HS, Dunson DB, Vehtari A, Rubin DB (2014) Bayesian data analysis.
Texts in Statistical Science Series (CRC Press, Boca Raton, FL), third edition.
24
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
[34] Gigerenzer G, Gaissmaier W (2011) Heuristic decision making. Annual review of psychology 62:451–
482.
[35] Gigerenzer G, Goldstein DG (1996) Reasoning the fast and frugal way: models of bounded rationality.
Psychological review 103(4):650.
[36] Gigerenzer G, Todd PM (1999) Simple heuristics that make us smart (Oxford University Press, USA).
[37] Gilardoni GL, Clayton MK (1993) On reaching a consensus using DeGroot's iterative pooling. The
Annals of Statistics 391–401.
[38] Golub B, Jackson MO (2010) Naıve Learning in Social Networks and the Wisdom of Crowds. American
Economic Journal: Microeconomics 2(1):112–149.
[39] Goyal S (2012) Connections: an introduction to the economics of networks (Princeton University Press).
[40] Grimm V, Mengel F (2014) An experiment on belief formation in networks. Available at SSRN 2361007
.
[41] Harary F (1959) A criterion for unanimity in french's theory of social power. Studies in Social Power,
168–182 (University of Michigan), d. Cartwright, Ed.
[42] Harel M, Mossel E, Strack P, Tamuz O (2014) When more information reduces the speed of learning
ArXiv preprint arXiv:1412.7172.
[43] Hegselmann R, Krause U (2005) Opinion dynamics driven by various ways of averaging. Computational
Economics 25(4):381–405.
[44] Isenberg DJ (1986) Group polarization: a critical review and meta-analysis. Journal of personality and
social psychology 50(6):1141.
[45] Jackson MO (2008) Social and Economic Networks (Princeton, NJ: Princeton University Press).
[46] Jadbabaie A, Lin J, Morse AS (2003) Coordination of groups of mobile autonomous agents using
nearest neighbor rules. IEEE Transactions on Automatic Control 48(6):988–1001.
[47] Jadbabaie A, Molavi P, Sandroni A, Tahbaz-Salehi A (2012) Non-bayesian social learning. Games and
Economic Behavior 76(1):210 – 225.
[48] Jadbabaie A, Molavi P, Tahbaz-Salehi A (2013) Information heterogeneity and the speed of learning
in social networks. Revise and Resubmit, Review of Economic Studies .
[49] James G, Rumchev V (2005) Stability of positive linear discrete-time systems. Bulletin of the Polish
Academy of Sciences. Technical Sciences 53(1):1–8.
[50] Janis IL (1982) Groupthink: Psychological studies of policy decisions and fiascoes (Houghton Mifflin
Boston).
[51] Kahneman D (2003) Maps of bounded rationality: Psychology for behavioral economics. The American
economic review 93(5):1449–1475.
[52] Kahneman D (2011) Thinking, fast and slow (Farrar Straus Giroux).
[53] Kanoria Y, Tamuz O (2013) Tractable bayesian social learning on trees. Selected Areas in Communi-
cations, IEEE Journal on 31(4):756–765.
[54] Kar S, Moura JM (2009) Distributed consensus algorithms in sensor networks with imperfect commu-
nication: Link failures and channel noise. IEEE Transactions on Signal Processing 57(1):355–369.
[55] Kass RE, Wasserman L (1996) The selection of prior distributions by formal rules. Journal of the
American Statistical Association 91(435):1343–1370.
[56] Katragadda S, SanMiguel JC, Cavallaro A (2014) Consensus protocols for distributed tracking in wire-
less camera networks. 17th International Conference on Information Fusion (FUSION), 1–8 (IEEE).
[57] Klein G (2007) Performing a project premortem. Harvard Business Review 85(9):18–19.
[58] Krishnamurthy V, Poor HV (2013) Social learning and bayesian games in multiagent signal processing:
How do local and global decision makers interact? IEEE Signal Processing Magazine, 30(3):43–57.
[59] Lalitha A, Sarwate A, Javidi T (2014) Social learning and distributed hypothesis testing. IEEE Inter-
national Symposium on Information Theory 551–555.
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
25
[60] Levy G, Razin R (2015) Correlation neglect, voting behavior, and information aggregation. The Amer-
ican Economic Review 105(4):1634–1645.
[61] Lyons R, Peres Y (2016) Probability on Trees and Networks (Cambridge Series in Statistical and
Probabilistic Mathematics).
[62] Marden JR, Shamma JS (2012) Revisiting log-linear learning: Asynchrony, completeness and payoff-
based implementation. Games and Economic Behavior 75(2):788–808.
[63] Masicampo EJ, Baumeister RF (2008) Toward a physiology of dual-process reasoning and judgment:
Lemonade, willpower, and expensive rule-based analysis. Psychological Science 19(3):255–260.
[64] Mesbahi M, Egerstedt M (2010) Graph Theoretic Methods in Multiagent Networks (Princeton Univer-
sity Press).
[65] Molavi P, Tahbaz-Salehi A, Jadbabaie A (2016) Foundations of non-bayesian social learning. Columbia
Business School Research Paper .
[66] Mossel E, Sly A, Tamuz O (2014) Asymptotic learning on bayesian social networks. Probability Theory
and Related Fields 158(1-2):127–157, ISSN 0178-8051.
[67] Mossel E, Tamuz O (2010) Efficient bayesian learning in social networks with gaussian estimators.
arXiv preprint arXiv:1002.0747 .
[68] Mueller-Frank M (2013) A general framework for rational learning in social networks. Theoretical
Economics 8(1):1–40.
[69] Mueller-Frank M, Neri C (2015) A general model of boundedly rational observational learning: Theory
and experiment. Available at SSRN 2566210 .
[70] Nedi´c A, Olshevsky A, Uribe C (2016) Distributed learning with infinitely many hypotheses. arXiv
preprint arXiv:1605.02105 .
[71] Nedi´c A, Olshevsky A, Uribe CA (2014) Nonasymptotic convergence rates for cooperative learning
over time-varying directed graphs. arXiv preprint arXiv:1410.1977 .
[72] Nedi´c A, Olshevsky A, Uribe CA (2015) Fast convergence rates for distributed non-bayesian learning.
arXiv preprint arXiv:1508.05161 .
[73] Rahimian MA, Jadbabaie A (2015) Learning without recall: A case for log-linear learning. 5th IFAC
Workshop on Distributed Estimation and Control in Networked Systems .
[74] Rahimian MA, Jadbabaie A (2015) Learning without recall in directed circles and rooted trees. Amer-
ican Control Conference 4222–4227.
[75] Rahimian MA, Jadbabaie A (2016) Bayesian learning without recall Preprint.
[76] Rahimian MA, Jadbabaie A (2016) Naive social learning in ising networks. American Control Confer-
ence .
[77] Rahimian MA, Molavi P, Jadbabaie A (2014) (Non-) bayesian learning without recall. IEEE Conference
on Decision and Control (CDC) 5730–5735.
[78] Rahimian MA, Shahrampour S, Jadbabaie A (2015) Learning without recall by random walks on
directed graphs. IEEE Conference on Decision and Control (CDC) .
[79] Rahnama Rad K, Tahbaz-Salehi A (2010) Distributed parameter estimation in networks. 49th IEEE
Conference on Decision and Control (CDC) 5050–5055.
[80] Redish AD (2013) The mind within the brain: How we make decisions and how those decisions go
wrong (Oxford University Press).
[81] Robert C (2001) The Bayesian Choice: From Decision-Theoretic Foundations to Computational Imple-
mentation. Springer Texts in Statistics (Springer).
[82] Rogers EM (2003) Diffusion of Innovations (Simon and Schuster), 5th edition.
[83] Roux N, Sobel J (2015) Group polarization in a model of information aggregation. American Economic
Journal: Microeconomics 7(4):202–32.
[84] Rufo MJ, Martin J, P´erez CJ, et al. (2012) Log-linear pool to combine prior distributions: A suggestion
for a calibration-based approach. Bayesian Analysis 7(2):411–438.
26
Rahimian and Jadbabaie: Bayesian Heuristics for Group Decisions
[85] Seneta E (2006) Non-negative matrices and Markov chains (Springer).
[86] Shahrampour S, Jadbabaie A (2013) Exponentially fast parameter estimation in networks using dis-
tributed dual averaging. 52nd IEEE Conference on Decision and Control (CDC) 6196–6201.
[87] Simon HA (1955) A behavioral model of rational choice. The Quarterly Journal of Economics 99–118.
[88] Simon HA (1956) Rational choice and the structure of the environment. Psychological review 63(2):129.
[89] Simon HA (1979) Rational decision making in business organizations. The American economic review
69(4):493–513.
[90] Sobel J (2014) On the relationship between individual and group decisions. Theoretical Economics
9(1):163–185.
[91] Stoner JA (1968) Risky and cautious shifts in group decisions: The influence of widely held values.
Journal of Experimental Social Psychology 4(4):442–459.
[92] Sunstein CR (2000) Deliberative trouble? why groups go to extremes. The Yale Law Journal 110(1):71–
119.
[93] Thaler RH (2015) Misbehaving: The making of behavioral economics (WW Norton & Company).
[94] Tsitsiklis J, Bertsekas D, Athans M (1986) Distributed asynchronous deterministic and stochastic
gradient optimization algorithms. IEEE Transactions on Automatic Control 31(9):803–812.
[95] Tsitsiklis JN, Athans M (1984) Convergence and asymptotic agreement in distributed decision prob-
lems. Automatic Control, IEEE Transactions on 29(1):42–50.
[96] Tversky A, Kahneman D (1975) Judgment under uncertainty: Heuristics and biases. Utility, probability,
and human decision making, 141–162 (Springer).
[97] Wang Y, Djuric PM (2015) Social learning with bayesian agents and random decision making. IEEE
Transactions on Signal Processing 63(12):3241–3250.
[98] Xiao L, Boyd S, Lall S (2005) A scheme for robust distributed sensor fusion based on average consensus.
Fourth International Symposium on Information Processing in Sensor Networks (IPSN), 63–70.
[99] Xiao L, Boyd S, Lall S (2006) A space-time diffusion scheme for peer-to-peer least-squares estimation.
Proceedings of the 5th international conference on Information Processing in Sensor Networks, 168–176.
[100] Zellner A (1988) Optimal information processing and bayes's theorem. The American Statistician
42(4):278–280.
|
1508.02674 | 1 | 1508 | 2015-08-11T18:14:27 | Space-Time Diagram Generation for Profiling Multi Agent Systems | [
"cs.MA",
"cs.SE"
] | Advances in Agent Oriented Software Engineering have focused on the provision of frameworks and toolkits to aid in the creation of Multi Agent Systems (MASs). However, despite the need to address the inherent complexity of such systems, little progress has been made in the development of tools to allow for the debugging and understanding of their inner workings.
This paper introduces a novel performance analysis system, named AgentSpotter, which facilitates such analysis. AgentSpotter was developed by mapping conventional profiling concepts to the domain of MASs. We outline its integration into the Agent Factory multi agent framework. | cs.MA | cs | Space-Time Diagram Generation for Profiling
Multi Agent Systems
Dinh Doan Van Bien, David Lillis, and Rem W. Collier
School of Computer Science and Informatics
[email protected], {david.lillis,rem.collier}@ucd.ie
University College Dublin
Abstract. Advances in Agent Oriented Software Engineering have fo-
cused on the provision of frameworks and toolkits to aid in the creation
of Multi Agent Systems (MASs). However, despite the need to address
the inherent complexity of such systems, little progress has been made in
the development of tools to allow for the debugging and understanding
of their inner workings.
This paper introduces a novel performance analysis system, named
AgentSpotter, which facilitates such analysis. AgentSpotter was developed
by mapping conventional profiling concepts to the domain of MASs. We
outline its integration into the Agent Factory multi agent framework.
1
Introduction
Recent developments in the area of Multi Agent Systems (MASs) have been
concerned with bridging the gap between theory and practice, by allowing concrete
implementations of theoretical foundations to be built and deployed. However, the
dearth of agent-specific development and debugging tools remains a significant
obstacle to MASs being adopted in industry on a large scale.
While some simple debugging and logging tools exist for MAS analysis,
these tend not to aid in reasoning about large-scale system when viewed at the
high agent-oriented abstraction level. Such tools typically allow for traditional
debugging actions such as state stepping and breakpoint insertion.
One popular performance analysis technique is known as profiling. Profiling is
based on the observation that the majority of the execution time of a program can
be attributed to a small number of bottlenecks (or hot spots). By improving the
efficiency of these portions of a program, overall performance can be dramatically
improved. Profiling was initially introduced by Donald E. Knuth in an empirical
study conducted on FORTRAN programs [1]. Since then, the technique has been
successfully applied to a variety of languages, platforms and architectures.
The aim of this paper is to apply the principles of traditional profiling systems
in a multi agent environment, so as to facilitate the developers of MASs in
debugging their applications by gaining a better understanding of where the
bottlenecks exist and performance penalties are incurred.
This paper is organised as follows: Section 2 provides a brief overview of
existing tools aimed at aiding in the analysis of MASs. In Section 3, we introduce
5
1
0
2
g
u
A
1
1
]
A
M
.
s
c
[
1
v
4
7
6
2
0
.
8
0
5
1
:
v
i
X
r
a
the AgentSpotter profiling system, with particular focus on outlining a conceptual
model for generic MAS profiling. A concrete implementation of this work, aimed at
the Agent Factory MAS framework, is outlined in Section 4. Section 5 presents the
space-time diagram produced by AgentSpotter in more detail, with an evaluation
of its usefulness given in Section 6. Finally, Section 7 presents our conclusions
and ideas for future work.
2 Related Work
In designing a profiler for MASs, the features that tend to be present in traditional
profilers for non-MAS applications must be identified. It is also necessary to
examine those debugging and analysis tools that already exist for MASs.
The motivation behind the use of profiling on computer applications is clearly
outlined in Knuth's observation that "less than 4% of a program accounts for
more than half of its running time" [1]. This statement implies that a developer
can achieve substantial increases in performance by identifying and improving
those parts of the program that account for the majority of the execution time.
The key aim of profilers is to identify these bottlenecks.
Another observation leading to the widespread adoption of profilers as debug-
ging tools is that there frequently exists a mismatch between the actual run-time
behaviour of a system and the programmers' mental map of what they expect
this behaviour to be. Profilers are useful in enlightening developers to particular
aspects of their programs that they may not otherwise have considered.
A traditional profiler typically consists of two logical parts. Firstly, an in-
strumentation apparatus is directly weaved into the program under study or run
side-by-side to gather and record execution data. Secondly, a post-processing
system uses this data to generate meaningful performance analysis listings or
visualisations.
In the traditional software engineering community, historical profilers such
as gprof [2] or performance analysis APIs like ATOM [3] and the Java Virtual
Machine Tool Interface (JVMTI) [4] have made performance analysis more
accessible for researchers and software engineers. However, the MAS community
does not yet have general access to these types of tools.
Unique amongst all of the mainstream MAS development platforms, Cougaar
is alone in integrating a performance measurement infrastructure directly into
the system architecture [5]. Although this is not applicable to other platforms,
it does provide a good insight into the features that MAS developers could
reasonably expect from any performance measurement application. The principal
characteristics of this structure are as follows:
-- Primary data channels consist of raw polling sensors at the heart of the
system execution engine that gather simple low-impact data elements such
as counters and event sensors.
-- Secondary channels provide more elaborate information, such as summaries
of the state of individual components and history analysis that stores perfor-
mance data over lengthy running times.
-- Computer-level metrics provide data on such items as CPU load, network
load and memory usage.
-- The message transport service gathers data on messages flowing through it.
-- An extension mechanism based on servlets allows the addition of visualisation
plugins that bind to the performance metrics data source.
-- The service that is charged with gathering these metrics is designed so as to
have no impact on system performance when not in use.
Other analysis tools exist for aiding the development of MASs. However,
these tend to be narrower in their focus, concentrating only on specific aspects of
debugging MASs and typically being only applicable to a specific agent platform.
The Agent Factory Debugger [6] is an example of a tool that is typical of most
multi agent frameworks. Its principal function is inspecting the status and mental
state of an individual agent: its goals, beliefs, commitments and the messages
it has exchanged with other agents. Tools such as this give limited information
about the interaction between agents and the consequences of these interactions.
The Brahms toolkit features an AgentViewer that allows developers to view
along a time line the actions that particular agents have taken, so as to enable
them to verify that the conceptual model of the MAS is reflected in reality [7].
An administrator tool for the LS/TS agent platform provides some high-level
system monitoring information, such as overall memory consumption and data on
the size of the agent population [8]. Another type of agent debugging tool is the
ACLAnalyzer that has been developed for the JADE platform [9]. Rather than
concentrating on individual agents, it is intended to analyse agent interaction in
order to see how the community of agents interacts and is organised. In addition
to visualising the number and size of messages sent between specific agents, it
also employs clustering in order to identify cliques in the agent community.
These latter tools are focused mostly on identifying what actions an agent is
carrying out, together with identifying the reasons why such actions are taken (in
response to the agents' own belief set or as a result of receiving communication
from other agents).
3 AgentSpotter Overview
The overriding objective of AgentSpotter is to map the traditional concepts
of profiling to agent-oriented concepts so as to build a profiler tool for MAS
developers. It could be argued that most mainstream agent toolkits are written
in Java, hence the existing profiling tools for the Java programming language
are appropriate for the analysis of such platforms and their agents. However, to
do so would necessitate the mapping of low-level method profiles to high-level
agent-specific behaviour. Thus, tools aimed specifically at Java operate at an
inappropriate conceptual level to be of use in agent analysis. Although it may be
useful to incorporate some Object-Oriented metrics into an MAS profiling tool,
the focus of this paper is on the agent-specific metrics that are appropriate.
Ideally, MASs should be capable of managing their own performance and
identifying their own bottlenecks which hamper system efficiency, and indeed
Fig. 1. AgentSpotter abstract architecture
much work is being undertaken towards this goal [10]. However, until this aim is
realised, the provision of analysis tools aimed at aiding human developers identify
issues with their systems remains of paramount importance.
This section outlines the abstract infrastructure of the AgentSpotter system,
which is capable of being integrated into any agent platform. An analysis of
the integration of AgentSpotter into a specific agent platform (namely Agent
Factory) is contained in Section 4.
The AgentSpotter abstract architecture is displayed in Figure 1, using the
following graphical conventions:
-- Top-level architectural units are enclosed in dashed lines and are titled in
slanted capital letters, e.g.
AGENT PLATFORM
-- Self-contained software packages are enclosed in solid lines e.g Profiler
-- Logical software modules (groups of packages) are titled using slanted capi-
talised names e.g. AgentSpotter Service
-- Arrows denote data or processing interactions e.g.
queries -
At the highest level, the AgentSpotter Service should communicate with the
Run-Time Environment to capture the profiling data from a Profiled Application
running inside an Agent Platform. The captured data should be stored into a
Snapshot File which would then be processed by a Query Engine to generate the
input data for AgentSpotter Station, the visualisation application.
The AgentSpotter profiler monitors performance events generated by the
Agent Platform's Run-Time Environment. These include such events as agent
management events, messaging and other platform service activity. Additionally,
the AgentSpotter service may employ system monitors to record performance
information such as CPU load, memory usage or network throughput. This
provides a general context for the event-based information.
Run-Time EnvironmentPro(cid:127)led ApplicationAgentSpotter ServicePro(cid:127)lerSystem MonitorsAGENT PLATFORMAGENTSPOTTER STATION (visualisation application)Snapshot File (.aspot)Session Summary TablesSpace-Time DiagramCall Graph Tree ViewSessionsAgentsEventsMessagesmessagesactivityeventseventsdatadataqueriesQUERY ENGINESession summary queryAgent activity queryMessages sent/received queryCall graph querySystem activity query The event data and other information collected by the AgentSpotter profile
is stored in a snapshot file, which contains the results of a single uninterrupted
data capture session. This snapshot contains a series of raw instrumentation data.
Because large MASs may generate potentially hundreds of events per second, it
is necessary to introduce a Query Engine that is capable of extracting summaries
and other information and make it available to visualisation tools in a transparent
manner. Ideally, this should be through a data manipulation language such as
SQL so as to facilitate the specification of rich and complex queries.
The final component of the abstract architecture is the AgentSpotter Station,
which is the visualisation tool that summarises the information gathered from
the Query Engine in a visual form. The principal focus of this paper is the
Space-Time Diagram, which is presented in Section 5.
When profiling any application, it is important to identify the appropriate
execution unit for profiling (e.g. in the context of object-oriented programming,
this would typically be an object or a method). For profiling a MAS, we believe
that the appropriate execution units are individual agents.
At the agent level, when considering only the autonomous computational entity
abstracted away from the interaction with its peers, the focus is on responsiveness.
It is obvious that the main influence on the responsiveness of an agent is the
amount of computation it requires to carry out the set of tasks required to meet
its design objectives. Agreeing with the BT researchers in [11], we make a further
distinction based on the rationality level of the agent architecture. For reactive
agents, this set of tasks includes only purely reactive behaviour, which is similar
to the behaviour of a traditional process. For deliberative agents that implement
a reasoning system, the computational cost of the reasoning activity must be
considered separately from the actual task execution. Based on this analysis, the
minimum information provided by the system should include:
-- Agent description: name, role, type (deliberative or reactive) of the agent.
-- Cumulative activity: cumulative computation time used by the agent.
-- Perception time: perception time used by a deliberative agent.
-- Action time: task execution time used by a deliberative agent.
-- Reasoning time: reasoning time used by a deliberative agent.
-- % session activity: percentage of the global session computation time used
by the agent.
-- Number of iterations: number of non-zero duration iterations used by the
agent.
-- Total number of messages sent and received: total number of ACL
messages exchanged by the agent.
Additionally, a number of global statistics should also be maintained:
-- Total duration: session run-time recorded.
-- Total activity: amount of computation time recorded over the session.
-- Total number of messages: number of messages sent or received by agents
on the platform being profiled.
Table 1. Benchmark application flat profile
Total Session Time
Total Activity
Messages Sent
Messages Received
Time Slice Duration
18:50.691
10:29.164
1206
1206
1000 ms
Agent
agent001
agent009
agent004
agent014
agent003
agent006
agent005
agent013
agent007
agent008
agent010
agent015
agent002
agent011
agent012
master2
master1
agent024
agent019
agent026
agent030
agent017
agent025
agent027
agent018
agent020
agent022
agent021
agent016
agent029
agent028
agent032
agent031
agent023
agent033
iterations
338
365
349
284
401
361
367
301
378
357
330
285
348
357
225
901
873
46
31
42
26
46
40
31
38
39
47
32
219
38
45
34
36
40
30
22
21
22
14
13
12
12
9
11
7
8
9
8
5
3
0
0
2
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
0
T > 0 T > 100% Activity % Session M ax(T ) Average(T ) Msg. Msg.
rec.
57
77
69
36
76
73
76
71
71
72
81
42
70
39
41
86
82
7
5
4
8
3
3
2
7
3
13
7
6
8
4
4
2
1
0
overload mm:ss.ms
1:08.564
1:04.257
1:01.529
46.413
43.881
40.141
34.903
34.716
31.864
30.850
30.280
29.382
23.196
19.363
13.172
6.681
6.485
6.281
4.449
4.400
4.002
3.811
3.767
3.694
3.384
1.762
1.523
1.300
1.194
1.039
0.749
0.749
0.742
0.598
0.043
ss.ms sent
6
0.202
13
0.176
10
0.176
0.163
2
12
0.109
12
0.111
17
0.095
0.115
14
21
0.084
14
0.086
21
0.091
4
0.103
0.066
9
4
0.054
9
0.058
0.007
504
0.007 514
0.136
3
0
0.143
0
0.104
2
0.153
0.082
0
0
0.094
0
0.119
2
0.089
0
0.045
0.032
5
2
0.040
2
0.005
2
0.027
1
0.016
0.022
1
0
0.020
0
0.014
0.001
0
activity
10.90
10.21
9.78
7.38
6.97
6.38
5.55
5.52
5.06
4.90
4.81
4.67
3.69
3.08
2.09
1.06
1.03
1.00
0.71
0.70
0.64
0.61
0.60
0.59
0.54
0.28
0.24
0.21
0.19
0.17
0.12
0.12
0.12
0.10
0.01
ss.ms
3.740
3.425
3.235
3.148
3.323
3.279
3.325
3.190
3.356
3.201
3.147
3.257
3.147
3.095
3.049
0.183
0.227
3.045
3.014
3.084
3.132
3.031
3.006
3.103
3.044
0.547
0.559
0.555
0.555
0.550
0.546
0.561
0.545
0.543
0.003
-- Average number of active agents per second: This gives an idea of the
level of concurrency in the application.
Following the convention of traditional profiling tools, we describe this in-
formation as a flat profile. AgentSpotter displays this by means of a JTable
(provided by Java's Swing interface tools). An example of how the information
is presented is given in Table 1. This is not necessarily an exhaustive list of
every piece of information a developer may desire for identifying problems with
a system, however we believe that other metrics (such as memory consumption)
are not as crucial for the purposes of profiling the application.
4 Agent Factory Integration
Following the definition of the abstract architecture outlined above, a concrete (i.e.
platform-specific) implementation was created for Agent Factory. Agent Factory
is a cohesive framework that supports a structured approach to the development
of agent-oriented applications [12]. This implementation is illustrated in Figure 2,
which uses the same graphical conventions as Figure 1.
Fig. 2. AgentSpotter concrete architecture
To create a concrete implementation, only the platform-specific details must
change, as the mechanisms required to monitor events vary from one agent
platform to another. In contrast, the AgentSpotter file processing and visualisation
components (shown in the lower part of Figure 2) are identical to those in the
abstract architecture (Figure 1). Thus, when implementing AgentSpotter for
a new type of agent platform, only the AgentSpotter Service that is coupled
directly with the platform needs to be reprogrammed. Provided this service
creates snapshot files in a consistent way, the Query Engine need not differentiate
Metrics ServiceAgent Management ServiceSchedulerMessage Transport ServiceAgent Factory Run-Time EnvironmentPro(cid:127)led ApplicationAgentSpotter ServicePro!lerCustom MetricsAgentsCPU MonitorHeap MonitorAGENT FACTORY PLATFORMAGENTSPOTTER STATION (visualisation application)Snapshot File (.aspot)Session Summary TablesSpace-Time DiagramCall Graph Tree ViewSessionsAgentsEventsMessagesdatadataqueriesQUERY ENGINESession summary queryAgent activity queryMessages sent/received queryCall graph querySystem activity query Fig. 3. AgentSpotter Space-Time Diagram specification annotated with AgentSpotter
for Agent Factory infrastructure links (see Figure 2).
between agent platforms. AgentSpotter snapshot files are actually transportable
single-file databases managed by the public domain SQLite Database Engine [13].
As a result, profiling data is stored as queryable relational database tables.
Within the Agent Factory Run-Time Environment, there are three specific
subsystems that generate events of interest in agent profiling, and as such are
recorded by the AgentSpotter service. First, the Agent Management Service is
responsible for creating, destroying, starting, stopping and suspending agents.
It generates events corresponding to each of these actions, which are recorded
in the snapshot file. The Scheduler is charged with scheduling which agents are
permitted to execute at particular times and generates events based on this.
Finally, the Message Transport Service records the sending and receipt of FIPA
messages by agents.
Metrics ServiceAgent Management ServiceSchedulerMessage Transport ServiceAgent Factory Run-Time EnvironmentPro!led ApplicationAgentSpotter ServicePro!lerCustom MetricsAgentsCPU MonitorHeap MonitorAGENT FACTORY PLATFORMSession Time LineCPU LineAgent Life Lineintra-platform message linesAgent Life Lineinformation window on mouse hoverExternal Platform Life LineHeap MonitorMessage Transport ServiceMessage Transport ServiceMessage Transport Serviceinter-platform messages linestimed events (with duration)"Bird's Eye View"♣♦♠simple events (no duration)Session Time LineCPU LineAgent Life LineAgent Life LineExternal Platform Life Line♣♦♠current viewport 5 Space-Time Diagram
In Section 3, we outlined the minimum amount of information that should be
made available by an agent profiler. However, this information can be presented
merely by the creation of a simple table. We believe that proper visualisation
tools will be far more useful to a developer in understanding a MAS. This section
introduces the Space-Time Diagram, which is at the core of the AgentSpotter
Station visualisation application. The aim of this diagram is to give as much
detail and context as possible about the performance of the MAS to the developer.
The user may pan the view around and zoom in and out so as to reveal hidden
details or focus on minute details.
The Session Time Line represents the running time of the application being
profiled. Regardless of the position and scale of the current viewport, this time
line remains visible to provide temporal context to the section being viewed and
also to allow a developer to move to various points in the session.
The CPU Line is a graphical plot of the CPU load of the host system during
the session. A vertical gradient going from green (low CPU usage) to red (high
CPU usage) provides a quick graphical sense of system load. A popup information
window reveals the exact usage statistics once the mouse is hovered over the line.
Perhaps the most important feature of the space-time diagram is the Agent
Time Lines. Each of these display all the performance and communication events
that occur for a single agent during a profiling session. A number of visual features
are available to the developer so as to gain greater understanding of the status
and performance of the system. For instance, an agent time line begins only at
the point in time when the agent is created. This facilitates the developer in
viewing the fluctuations in the agent population. Another simple visual aid is
that a time line's caption (i.e. the name of the associated agent) is always visible,
regardless of what position along the line a developer has scrolled to. Visual
clutter may also be reduced by temporarily hiding certain time lines that are not
of interest at a particular point in time.
The time line also changes colour to distinguish busier agents from the rest
of the community. Darker lines indicate agents that have consumed a greater
proportion of the system's CPU time. In a situation where system performance
has been poor, this will allow a developer to quickly identify candidate agents for
debugging, if they are consuming more resources than is appropriate or expected.
The default ordering of the time lines shares this aim. The time lines are in
descending order of total computation time, again visually notifying the developer
of those agents consuming more processing resources. However, a developer may
alter this default order by dragging time lines into different positions, perhaps to
group lines with particularly interesting interactions.
In addition to this simple information, the main purpose of the time line
is to show events performed by an agent that are likely to be of interest from
a performance point of view. These performance events are divided into two
categories. Simple performance events are those that have a time stamp only.
These are shown by means of standard icons (such as an envelope icon to denote
that a message was received by the agent).
Fig. 4. Agent Factory agent activity representation in the Space-Time Diagram
The other category of performance events are timed performance events. These
events are typically actions being performed by an agent. A basic timed perfor-
mance event displays as a rectangle with a fixed height and a width proportional
to its duration. More elaborate event representations can be implemented. For
example, in the context of Agent Factory's time slice based scheduler, we have
represented the concept of time slice overshoots i.e. when an agent has overused
its time slice allocation.
An example of how these timed events are represented is given in Figure 4.
Each agent has a particular time slice within which it is expected to perform
all of its actions in a particular iteration. Agents exceeding their time slice may
prevent or delay other agents from accessing CPU. As a visual aid to identifying
when this situation occurs, timing events are represented by coloured rectangles.
The size of these rectangles is proportional to the duration of the event and
the percentage of the allocated timeslice used. The colour code also indicates
how the agent has used its available time. A green rectangle indicates that the
agent has used anything up to 75% of the time available. From 75% to 100%, an
orange rectangle indicates that the agent may require further analysis from the
developer to avoid the danger of exceeding the time slice. Finally, whenever an
agent exceeds its time, a red rectangle is used. It must be acknowledged that as
a result of different approaches to scheduling agent platforms, timed performance
events may not be available to a specific AgentSpotter implementation. When
this arises, all events will be recorded as simple performance events.
Communication between agents is shown by lines linking the appropriate agent
timelines. These include arrows to indicate the direction of the communication.
Hovering the mouse pointer over such a line causes a popup window to display the
Time% time slice50%75%100%0%{{{1 time slice150% of time slice (exceeded allocation: RED)80% of time slice (over 75%: ORANGE)50% of time slice (under 75%: GREEN)proportional dimensions FIPA headers and content of the message. There is also a distinction made between
messages passed between agents housed on the same agent platform (intra-
platform) and those passed between agents on different platforms (inter-platform).
Since agents on other platforms will not have an agent time line, an external
platform life line is drawn for each other platform with which agents communicate.
Rather than linking with individual agent time lines, communications with these
platforms are drawn directly to the external platform life line.
The combination of these communication lines and the performance event
indicators are very useful in identifying the causes of agent activity. Given
the inherently social nature of MASs, it is very common for agent activity to
be motivated by communication. For example, an agent may be requested to
perform a task by some other agent. Alternatively, an agent may receive a piece
of information from another agent that it requires in order to perform a task to
which it has previously committed as a result of its own goals and plans.
Providing such detailed visualisation of a MAS requires a substantial amount
of screen space. The basic features of zooming and panning are complemented
by the provision of a "bird's eye view", which displays a zoomed-out overview of
the entire session. This allows the user to quickly move the current viewport to
focus on a particular point in time during the session, as illustrated in Figure 3.
6 Evaluation
Having outlined the required features of AgentSpotter, along with details of
its implementation, it is necessary to demonstrate how it can be utilised on a
running MAS. To this end, a specialist benchmark application was developed
that will allow the features of the AgentSpotter application to be shown.
6.1 Specification
The aim of the benchmark application is to perform all the activities necessary
for AgentSpotter to display its features. The requirements for the application
can be summarised as follows:
-- Load history: a normally distributed random load history should be gener-
ated so that we can get an idea of a "normal" profile which can be contrasted
with "abnormal" profiles where, for example, a single agent is monopolising
all the load, or the load is spread equally among all agents.
-- Agent population: the number of active agents should be changeable
dynamically to simulate process escalation.
-- Interactions: in addition to direct interactions, the application should
exercise some task delegation scenarios. The idea is to generate multiple hops
messaging scenarios and see their impact on performance.
-- Messages: agents should generate a steady flow of messages with occasional
bursts of intense communication.
-- Performance events: all three performance behaviours described in Sec-
tion 3 should be represented, i.e. green (t ≤ 50% time slice), orange (50% ≤
t ≤ 75% time slice), and red (t > 100%).
These requirements were satisfied by creating a MAS with overseer agents that
request worker agents to execute small, medium or large tasks. Worker agents
that have been recently overloaded will simply refuse to carry out the tasks (in a
real application they would inform requester about their refusal). From time to
time, overseer agents would request agents to delegate some tasks. In this case,
worker agents will behave as overseers just for one iteration. A simple interface
allows the user to start and pause the process, along with the ability to set the
number of active worker agents.
6.2 Evaluation Scenario and Objective
The following simple scenario was played out in order to generate a flat profile
and space-time diagram.
1. Start the session with 12 worker agents and 2 overseer agents.
2. After 10 minutes add 15 worker agents to spread the load.
3. After 4 further minutes, suspend the process for 20 seconds.
4. At this point, reduce the number of worker agents to 12.
5. Run for 5 minutes more and then stop the session.
6.3 Flat profile
The resulting flat profile of this test is reproduced in Table 1. For the reader's
convenience, the maximum value for each column is identified by an enclosing
box. Overseer agents are called "master1" and "master2". The worker agents are
called "agent" followed by a number e.g. "agent007".
Firstly, the benchmark appears to make a good job of producing a load history
following a normal distribution.
Secondly, we can draw the following conclusions from a quick study of Table 1:
-- The most active agents in terms of number of iterations are the overseer agents,
"master1" and "master2", however in terms of CPU load and overload, three
worker agents are topping the list with 30% of the total activity: "agent001",
"agent009", and "agent003".
-- The agents with the highest CPU load also display a high number of time
slice overshoots, and a high average time slice duration.
-- As expected, the overseer agents were very busy exchanging messages with
the workers. However, it seems that messaging is not CPU intensive. This
possibly results from the way in which message sending is implemented, with
the CPU load indicated here corresponding to the scheduling of a message for
sending, rather than the actual sending of the message. It may be necessary
to attach a specialist monitor to the Message Transport Service to gain full
information about the impact of sending messages. This causes the activity
percentage of the overseer agents to be very low, at only 1%.
Fig. 5. Benchmark application sample space-time diagram (18 minute long session)
In this instance, the flat profile lends evidence to the notion that the actual
behaviour of the system matches the design principles on which it was built.
6.4 Space-Time Diagram
The space-time diagram for this session is shown in Figure 5. The individual
agent time lines can clearly be seen as horizontal bars in the main window of
the application. Within these, rectangular boxes represent processing tasks being
carried out by each agent. The vertical lines between the time lines represent
messages being passed between agents. For this simple scenario, only a single
agent platform was used, meaning that there are no external platform time lines
to indicate messages travelling to and from other agent platforms. A number of
points of interest are labelled on the diagram. These can be described as follows:
-- This portion of the diagram shows what happens when the initial 12 workers
are active. The large red rectangles illustrate the time-consuming tasks ordered
by the overseers. As mentioned in Section 5, these are also identifiable by
their size, which increases proportionally to the processing time taken. These
blocks never overlap because of the way Agent Factory schedules agents (i.e.
agents are given access to the CPU sequentially). It is also noteworthy that
the Agent Factory scheduler does not preempt agents that have exceeded
their time allocation. The red rectangles also come in bursts, because both
overseers send the same order to the same worker at the same time. This was
revealed by zooming into what initially appeared to be a single message line.
At a high magnification level, there were in fact two messages lines within a
few microseconds interval to the same worker.
-- At this point, 15 more workers are added to the system, following a slight
pause that is indicated by the temporary absence of message lines. The agent
time lines for the additional agents only begin at this point, clearly indicating
an increase in the agent population.
-- This third portion shows the impact of the new workers. The red blocks are
still present, but they are better spread among the agents, with the new
agents taking some of the load from their predecessors.
-- The Bird's Eye View reveals the bigger picture, and reminds us that we are
looking only at one third of the overall session.
7 Conclusions and Future Work
Currently, the only concrete implementation of AgentSpotter is for the Agent
Factory platform. As noted in Section 3, only the data capture apparatus should
require a separate implementation for another platform. It is intended to develop
such an implementation for other platforms, such as JADE [14].
The most obvious source of improvement for the AgentSpotter application
is the addition of extra information above that which is already available. For
instance, the performance of additional system services should be recorded, and
more details should be collected about agents' performance events, such as the
distribution of an agent's execution time among its sensors, actuators, reasoning
engine and other components. Finally, the AgentSpotter application currently
supports only one agent platform at any given time. The capability to visualise
multiple platforms concurrently would be desirable.
References
1. Knuth, D.E.: An empirical study of FORTRAN programs. j-SPE 1(2) (April/June
1971) 105 -- 133
2. Graham, S.L., Kessler, P.B., Mckusick, M.K.: Gprof: A call graph execution profiler.
SIGPLAN Not. 17(6) (1982) 120 -- 126
3. Srivastava, A., Eustace, A.: Atom: a system for building customized program
analysis tools. In: PLDI '94: Proceedings of the ACM SIGPLAN 1994 conference
on Programming language design and implementation, New York, NY, USA, ACM
(1994) 196 -- 205
4. Sun Microsystems, Inc.: JVM Tool Interface (JVMTI), Version 1.0. Web pages
at http://java.sun.com/j2se/1.5.0/docs/guide/jvmti/ (accessed August 4th,
2008) (2004)
5. Helsinger, A., Thome, M., Wright, T., Technol, B., Cambridge, M.: Cougaar: a
scalable, distributed multi-agent architecture. In: Systems, Man and Cybernetics,
2004 IEEE International Conference on. Volume 2. (2004)
6. Collier, R.: Debugging Agents in Agent Factory. Lecture Notes in Computer
Science 4411 (2007) 229
7. Seah, C., Sierhuis, M., Clancey, W., Cognition, M.: Multi-agent modeling and
simulation approach for design and analysis of MER mission operations.
In:
Proceedings of 2005 International conference on human-computer interface advances
for modeling and simulation (SIMCHI'05). (2005) 73 -- 78
8. Rimassa, G., Calisti, M., Kernland, M.E.: Living Systems R(cid:13)Technology Suite.
Whitestein Series in Software Agent Technologies and Autonomic Computing. In:
Software Agent-Based Applications, Platforms and Development Kits. Birkhauser
Basel (2005) 73 -- 93
9. Botia, J., Hernansaez, J., Skarmeta, F.: Towards an Approach for Debugging
MAS Through the Analysis of ACL Messages. In: Multiagent System Technologies:
Second German Conference, MATES 2004, Erfurt, Germany, September 29-30,
2004: Proceedings, Springer (2004)
10. Horn, P.: Autonomic Computing: IBM's Perspective on the State of Information
Technology. IBM TJ Watson Labs, NY, 15th October (2001)
11. Lee, L.C., Nwana, H.S., Ndumu, D.T., Wilde, P.D.: The stability, scalability and
performance of multi-agent systems. BT Technology Journal 16(3) (1998) 94 -- 103
12. Collier, R., O'Hare, G., Lowen, T., Rooney, C.: Beyond Prototyping in the Factory
of Agents. Multi-Agent Systems and Application III: 3rd International Central
and Eastern European Conference on Multi-Agent Systems, Ceemas 2003, Prague,
Czech Republic, June 16-18, 2003: Proceedings (2003)
13. Hwaci: Web site for the SQLite Database Engine (2008) http://www.sqlite.org/
(accessed October, 2008).
14. Bellifemine, F., Poggi, A., Rimassa, G.: JADE -- A FIPA-compliant agent framework.
In: Proceedings of PAAM. Volume 99. (1999) 97 -- 108
|
1801.09584 | 1 | 1801 | 2018-01-29T15:45:05 | A Unifying Framework for Manipulation Problems | [
"cs.MA",
"cs.DS"
] | Manipulation models for electoral systems are a core research theme in social choice theory; they include bribery (unweighted, weighted, swap, shift, ...), control (by adding or deleting voters or candidates), lobbying in referenda and others.
We develop a unifying framework for manipulation models with few types of people, one of the most commonly studied scenarios. A critical insight of our framework is to separate the descriptive complexity of the voting rule R from the number of types of people. This allows us to finally settle the computational complexity of R-Swap Bribery, one of the most fundamental manipulation problems. In particular, we prove that R-Swap Bribery is fixed-parameter tractable when R is Dodgson's rule and Young's rule, when parameterized by the number of candidates. This way, we resolve a long-standing open question from 2007 which was explicitly asked by Faliszewski et al. [JAIR 40, 2011].
Our algorithms reveal that the true hardness of bribery problems often stems from the complexity of the voting rules. On one hand, we give a fixed-parameter algorithm parameterized by number of types of people for complex voting rules. Thus, we reveal that R-Swap Bribery with Dodgson's rule is much harder than with Condorcet's rule, which can be expressed by a conjunction of linear inequalities, while Dodson's rule requires quantifier alternation and a bounded number of disjunctions of linear systems. On the other hand, we give an algorithm for quantifier-free voting rules which is parameterized only by the number of conjunctions of the voting rule and runs in time polynomial in the number of types of people. This way, our framework explains why Shift Bribery is polynomial-time solvable for the plurality voting rule, making explicit that the rule is simple in that it can be expressed with a single linear inequality, and that the number of voter types is polynomial. | cs.MA | cs |
A Unifying Framework for Manipulation Problems ∗
Dusan Knop†
Martin Kouteck´y‡
Matthias Mnich§
Abstract
Manipulation models for electoral systems are a core research theme in social choice theory; they
include bribery (unweighted, weighted, swap, shift, . . . ), control (by adding or deleting voters or candi-
dates), lobbying in referenda and others.
We develop a unifying framework for manipulation models with few types of people, one of the most
commonly studied scenarios. A critical insight of our framework is to separate the descriptive complexity
of the voting rule R from the number of types of people. This allows us to finally settle the computational
complexity of R-Swap Bribery, one of the most fundamental manipulation problems. In particular, we
prove that R-Swap Bribery is fixed-parameter tractable when R is Dodgson's rule and Young's rule,
when parameterized by the number of candidates. This way, we resolve a long-standing open question
from 2007 which was explicitly asked by Faliszewski et al. [JAIR 40, 2011].
Our algorithms reveal that the true hardness of bribery problems often stems from the complexity of
the voting rules. On one hand, we give a fixed-parameter algorithm parameterized by number of types
of people for complex voting rules. Thus, we reveal that R-Swap Bribery with Dodgson's rule is much
harder than with Condorcet's rule, which can be expressed by a conjunction of linear inequalities, while
Dodson's rule requires quantifier alternation and a bounded number of disjunctions of linear systems.
On the other hand, we give an algorithm for quantifier-free voting rules which is parameterized only
by the number of conjunctions of the voting rule and runs in time polynomial in the number of types
of people. This way, our framework explains why Shift Bribery is polynomial-time solvable for the
plurality voting rule, making explicit that the rule is simple in that it can be expressed with a single
linear inequality, and that the number of voter types is polynomial.
1
Introduction
Problems of manipulation, bribery and control constitute a fundamental part of computational social choice.
Many such problems are known to be NP-hard (or worse). However, their input can naturally be partitioned
into several parts, like the number of voters, the number of candidates, and others. This motivates the study
of such problems by the powerful tools of parameterized complexity. One of the most fundamental parameters
is the number of candidates C, which in many real-life scenarios can be expected to be reasonably small.
A by-now classical example in this direction is the R-Swap Bribery problem, which takes as input an
election consisting of a set C of candidates and a set V of voters with their individual preference lists ≻v
(for v ∈ V ), which are total orders over C. Additionally, for each voter v ∈ V and each pair of consecutive
candidates c ≻v c′, there is some cost σv(c, c′) ∈ Z of swapping the order of c and c′ in ≻v. The objective
is to find a minimum-cost set of swaps of consecutive candidates in the preference lists in order to make
a designated candidate c⋆ ∈ C the winner of the thus-perturbed election under a fixed voting rule R.
This problem was introduced by Elkind et al. [EFS09] and has since been studied for many classical
In particular, its computational complexity has been
voting rules R [DS12, FRRS14, KKM17c, SFE17].
∗Research supported by CE-ITI grant project P202/12/G061 of GA CR, project NFR MULTIVAL, and ERC Starting Grant
306465 (BeyondWorstCase).
†[email protected] Department of Applied Mathematics, Charles University, Prague, Czech Republic and Department of
Informatics, University of Bergen, Bergen, Norway
‡[email protected] Technion - Israel Institute of Technology, Haifa, Israel
§[email protected] Institut fur Informatik, Universitat Bonn, Bonn, Germany and Department of Quantitative Economics,
Maastricht University, Maastricht, The Netherlands
1
thoroughly analyzed with respect to the number of candidates C. The observation that C is often small
motivated the search for fixed-parameter algorithms for R-Swap Bribery parameterized by C, which are
algorithms that run in time f (C) · nO(1) for some computable function f , here n denotes the size of the
input election; if such an algorithm exists, we then say that the problem is fixed-parameter tractable with
respect to the parameter C.
Despite the problem's importance, for a long time, only the "uniform cost" case of R-Swap Bribery was
known to be fixed-parameter tractable for various voting rules, parameterized by the number of candidates;
here, uniform cost refers to the special case that all voters have the same cost function, that is, σv ≡ σ for all
v ∈ V . This is a fundamental result due to Dorn and Schlotter [DS12], who showed that R-Swap Bribery
with uniform cost can be solved in time 22C
· nO(1) for all voting rules R that are "linearly describable".
Many classical voting rules are indeed linearly describable, like any scoring protocol, Copelandα, Maximin,
or Bucklin.
O(1)
Recently, Knop et al. [KKM17c] gave the first fixed-parameter algorithms for R-Swap Bribery for
general cost functions for most voting rules R studied in the literature (scoring protocol, Copelandα, Maximin,
Bucklin etc.), thereby removing the uniform cost assumption. This way, they resolved a long-standing open
problem. Moreover, their algorithm runs in time 2CO(1)
· nO(1) for many rules R, and thus improves the
double-exponential run time by Dorn and Schlotter. Their key idea was to reduce the problem to so-called
n-fold integer programming, which allowed them to solve the problem efficiently for bounded number of
candidates despite their integer program having an unbounded number of variables. Their approach also
solved R-Swap Bribery for R being the Kemeny rule, even for general cost functions, though the Kemeny
rule is not known to be linearly describable (cf. [FHH11, p. 338]). However, this does not apply for Dodgson's
and Young's rules.
1.1 The challenge
Even so, there are some notable voting rules R for which the complexity of R-Swap Bribery remained open
even in the uniform cost case. This includes the Dodgson rule and the Young rule. Those rules are based
on the notion of Condorcet winner, which is a candidate who beats any other candidate in a head-to-head
contest. The Condorcet voting rule is very natural and dates back to the 18th century; however, clearly there
exist elections without a Condorcet winner. In such a situation one proclaims those candidates as winners
who are "closest" to being a Condorcet winner; different notions of closeness then yield different voting rules:
• Closeness measured as the of number of swaps in voter's preference orders defines the Dodgson rule.
• Closeness measured as the number of voter deletions defines the Young rule.
Thus, a candidate c is a Dodgson winner if s/he can be made a Condorcet winner by a minimum number of
swaps in the voter's preference orders over all candidates; analogously for the Young rule and voter deletions.
Kemeny rule1,
When considering R-Swap Bribery, the Dodgson rule and the Young rule are much more complicated
to handle than other rules; the reasons are several. First, for many voting rules R, the winner of an election
can be found in polynomial time, and solving this winner determination problem is certainly a necessary
subtask when solving R-Swap Bribery. However, for R ∈ {Dodgson, Young, Kemeny}, already winner
determination is NP-hard, and so even verifying a solution (that c⋆ is indeed the winner of the perturbed
election) is intractable. However, for R ∈ {Dodgson, Young}, already winner determination is NP-hard,
and so even verifying that c⋆ is indeed the winner of the perturbed election is intractable. In fact, winner
determination for these voting rules is complete for parallel access to NP [HSV05, HHR97, RSV03], denoted
-complete2. Second, for more than 25 years the winner determination problem for the Dodgson rule and
PNP
1Faliszewski [FHH09] give a fixed-parameter algorithm for the R-Bribery problem when R =Kemeny; R-Bribery is, in a
sense, simpler than R-Swap Bribery because in Bribery the cost of bribing a voter does not depend on how we bribe, while
in Swap Bribery the cost is the sum of costs for each performed swap.
2The class PNP
contains all problems solvable in polynomial time by a deterministic Turing machine which has access to an
NP oracle, but must ask all of its oracle queries at once (i.e., the queries can not depend on each other).
2
Young rule was only known to be solvable by an ILP-based algorithm [BITT89] with doubly-exponential
dependence in C; a single-exponential algorithm is only known since recently [KKM17c]. Even though
winner determination for these rules turns out to be fixed-parameter tractable with parameter C, there
is a sharp difference:
for the Kemeny rule, a simple procedure enumerating all C! possible preference
orders suffices to determine the winner, while for Dodgson and Young, only a double-exponential ILP-
based algorithm was known for a long time [BITT89] and a single-exponential algorithm is only known
recently [KKM17c]. This provides a sharp contrast to the Kemeny rule, for which simply enumerating all
C! possible preference orders suffices to determine the winner. Faliszewski et al. [FHH09] describe these
difficulties:
It is interesting to consider which features of Kemeny elections allow us to employ the above [ILP-based]
attack, given that the same approach does not seem to work for either Dodgson or Young elections. One
of the reasons is that the universal quantification implicit in Dodgson and Young elections is over an
exponentially large search space, but the quantification in Kemeny is, in the case of a fixed candidate set,
over a fixed number of options.
Thus, it is not clear how to solve R-Swap Bribery even for uniform cost with any fixed-parameter algorithm
for R being the Dodgson rule or the Young rule. These complications led Faliszewski et al. [FHH11] to
explicitly ask for the complexity of R-Swap Bribery parameterized by the number of candidates under
these rules.
1.2 Our contributions
We start by making a key observation about the majority of fixed-parameter algorithms for R-Swap Bribery
when C is small. A typical such result is an algorithm for R-Swap Bribery for R being Condorcet's voting
rule. That algorithm uses two key ingredients:
1. There are at most C! preference orders of C, and hence each voter falls into one of C! types; thus, an
input election is expressible as a society s = (s1, . . . , sC!), where si is the number of voters of type i.
2. Expressing that a candidate c⋆ is a Condorcet winner is possible using a conjunction of C − 1 linear
inequalities in terms of s.
As those key properties hold almost universally for voting rules R, one might be tempted to think that if
there are many types of voters, the R-Swap Bribery problem must be hard, and if there are few types of
voters, the problem must be easy. However, two points arise as counter-evidence. First, very recently, Knop
et al. [KKM17c] showed that even if there are many types of voters who differ by their cost functions, the
R-Swap Bribery problem remains fixed-parameter tractable for a wide variety of voting rules R. Second,
as already mentioned, it was open since 2007 whether R-Swap Bribery with Dodgson's and Young's voting
rule are fixed-parameter tractable for few candidates, even for uniform cost functions.
From voters and candidates to societies. Here, we take a novel perspective. We observe that the two key
ingredients (1) and (2) apply much more widely than for R-Swap Bribery; namely, they are also present
in many other manipulation, bribery and control problems. We therefore abstract away the specifics of such
problems and introduce general notions of "society", "moves in societies", and "winning conditions". Let
τ ∈ N be the number of types of people (e.g., voters in an election or a referendum). A society s is simply
a non-negative τ -dimensional integer vector encoding the numbers of people of each type. A move m is
a τ 2-dimensional integer vector whose elements sum up to zero; it encodes how many people move from
one type to another. A change ∆ is a τ -dimensional vector (typically associated with a move) encoding
the effect of a move on a society, such that s + ∆ is again a society. Finally, a winning condition Ψ(s) is
a predicate encoding some desirable property of a society, such as that a preferred candidate has won or
that a preferred agenda was selected in a referendum. Specifically, we study winning conditions which are
describable by formulas in Presburger Arithmetic (PA). PA is a logical language whose atomic formulas are
linear inequalities over the integers, which are then joined with logical connectives and quantifiers. Thus,
3
winning conditions describable by PA formulas widely generalize the class of linearly describable voting rules
by Dorn and Schlotter [DS12].
Our main technical contribution informally reads as follows:
Theorem 1 (informal). Deciding satisfiability of PA formulas with two quantifiers is fixed-parameter
tractable with respect to the dimension and length of formula, provided its coefficients and constants are
given in unary.
The importance of Theorem 1 arises from its applicability to the following general manipulation problem
that we introduce here. This general manipulation problem, which we call Minimum Move, captures that
many manipulation problems can be cast as finding a minimum move with respect to some objective function;
in particular, it encompasses the well-studied R-Swap Bribery problem. We study Minimum Move for
linear objective functions and winning conditions Ψ expressible with PA formulas of the form "∃∀". For
all such Ψ, with the help of Theorem 1, we show that Minimum Move is fixed-parameter tractable for
combined parameter the descriptive complexity (length) of the winning condition Ψ and the number τ of
"types of people", that is, it is fixed-parameter tractable for parameter lengthe of Ψ plus τ . As an important
special case, we obtain the first fixed-parameter algorithm for R-Swap Bribery for R the Dodgson rule
and the Young Rule with uniform costs. To this end, we model the winning condition of the Dodgson
rule and Young rules as a PA formula. For intuition, consider the Young rule: a candidate c⋆ is a Young
winner (with score d) if there exists a set V ⋆ ⊆ V of at most d voters such that c⋆ is a Condorcet winner
of the election (C, V \ V ⋆), and for all sets V ′ ⊆ V of at most d − 1 voters any other candidate c 6= c⋆
is not a Condorcet winner of the election (C, V \ V ′). This formula has one quantifier alternation, and its
length (for a fixed score d) is bounded by some function of C; finally we have to take a disjunction of such
formulas over all possible scores d. For a candidate set C, the number τ of types of people is bounded by C!.
Consequently, we finally settle the long-standing open question about the complexity of R-Swap Bribery
for R ∈ {Dodgson, Young}, that was explicitly raised by Faliszewski [FHH11]:
Theorem 2. R-Swap Bribery with uniform cost is fixed-parameter tractable parameterized by the number
of candidates for R being the Dodgson rule or the Young rule; it can be solved in time f (C) · V O(1) for
some computable function f .
Beyond this fundamental problem, we show that a host of other well-studied manipulation problems are
captured by our fixed-parameter algorithm for Minimum Move:
Corollary 3. For R ∈ {Dodgson, Young}, the following problems are fixed-parameter tractable for uniform
costs when parameterized by the number C of candidates: R-$Bribery, R-CCDV/CCAV, R-Possible
Winner, and R-Extension Bribery.
Let us turn our attention to the parameter "number of types of people" τ . Our main contribution here
is the following:
Theorem 4 (informal). For any quantifier-free winning condition Ψ, Minimum Move can be solved in time
polynomial in the number of types and exponential only in the number of linear inequalities of Ψ.
Note that in many models of bribery and control, the number of potential types of people (i.e., types
that can occur in any feasible solution) is polynomial in the number of people on input. For example, in
Shift Bribery, every voter can be bribed to change their preferences order to one of C − 1 orders; thus
the number of potential types is (C − 1)V . Similarly, in CCAV / CCDV (constructive control by adding
or deleting voters), every voter has an active/latent bit; thus the number of potential types is 2V . Similar
arguments also work for Support Bribery where we change voters' approval counts, and with a more
intricate argumentation also for some voting rules and Bribery and $Bribery. In this sense, the fact that
we need to consider C! potential voter types in R-Swap Bribery almost seems like an anomaly, rather than
a rule. In summary, the complexity of Minimum Move depends primarily on the descriptive complexity of
the winning condition Ψ, because in many cases the number of types of people is polynomially bounded.
4
Another consequence of Theorem 1 are the first fixed-parameter algorithms for two important manip-
ulation problems beyond R-Swap Bribery. The Resilient Budget problem asks, for a given society
whether allocating budget B is sufficient in order to repel any adversary move of cost at most Ba with a
counter-move of cost at most B (so that the winning condition is still satisfied). Similarly, Robust Move
asks for a move of cost at most B which causes the winning condition to be satisfied even after any adversary
move of cost at most Ba. For formal definitions and results, cf. Sect. 4.2.
1.3
Interpretation of results
Intuitively, the results obtained with Theorem 1 can be interpreted as follows. Dodgson-Swap Bribery
is fixed-parameter tractable parameterized by C; however, this comes with at least two limitations as
compared to prior work for simpler voting rules R. First, our methods do not extend beyond the uniform
cost scenario, and this remains a major open problem. Second, our result requires the input election to
be given in unary, while prior work allows it to be given in binary (this is sometimes called the succinct
case [FHH09]). This is easily explained by the different descriptive complexities of the respective voting rules:
for example, while Condorcet's voting rule can be formulated as a quantifier-free PA formula, formulating
Dodgson's rule requires a long disjunction of formulas which use two quantifiers and a bounded number of
disjunctions.
Theorem 4 lets us discuss more specifically the complexity of various voting rules. For example, the
Plurality voting rule can be expressed with a single linear inequality encoding that a preferred candidate
obtained more points than the remaining candidates altogether. Thus, all problems which can be modeled
as Minimum Move are polynomial-time solvable with the Plurality voting rule. This interprets the re-
sult of Elkind et al. [EFS09, Theorem 4.1] that Plurality-Shift Bribery is polynomial-time solvable: the
number of potential voter types is polynomial, and Plurality has a simple description. Continuing, we may
compare R=Borda with R=Copeland. The winning condition for R=Borda can be described with C − 1
inequalities, while R=Copeland requires O(C2) inequalities. Thus Borda-Swap Bribery is solvable in
time CO(C2) log V , while Copeland-Swap Bribery requires time CO(C4) log V . Finally, all descrip-
tions of Kemeny's voting rule we are aware of require C! inequalities, and thus result in Kemeny-Swap
Bribery being solvable in time C!(C!)O(1)
log V . We do not claim these complexities to be best possible,
but conjecture the existence of lower bounds separating the various voting rules; in particular, we believe
that Kemeny-Swap Bribery requires double-exponential time.
Finally, our work provides a natural next step in unifying the many different models that have been
proposed for voting, bribing and manipulation problems. In this direction, Faliszewski et al. [FHH11] study
what happens when multiple bribery and manipulation actions can occur in an election; e.g., CCAV asks
for constructive control by adding voters while CCDV by deleting voters; similarly for CCAC and CCDC
for adding/deleting candidates. Faliszewski et al. unify those various (up to that point separately studied)
attacks. Similarly, Knop et al. [KKM17c] formulate the R-Multi Bribery problem, which also incorporates
swaps and perturbing approval counts. The problem we put forward in this paper, Minimum Move, in some
sense generalizes and simplifies all those "meta"-problems.
1.4 Related work
We have reviewed most of the relevant computational social choice work already. However, there seems
to be some confusion in the literature that deserves clarification. The paper of Faliszewski et al. [FHH09]
pioneering the concept of bribery in elections indeed considers the voting rules Kemeny, Dodgson and Young,
and provides a fixed-parameter algorithm for Kemeny-Bribery. There are three features of their paper that
we wish to discuss.
First, turning their attention to Dodgson-Bribery, they write:
Applying the integer programming attack for the case of bribery within Dodgson-like election systems [...]
is more complicated. These systems involve a more intricate interaction between bribing the voters and
then changing their preferences. For Dodgson elections, after the bribery, we still need to worry about
5
the adjacent switches within voters' preference lists that make a particular candidate a Condorcet winner.
[...] This interaction seems to be too complicated to be captured by an integer linear program, but building
on the flavor of the Bartholdi et al. [BITT89] ILP attack we can achieve the following: Instead of making
p a winner, we can attempt to make p have at most a given Dodgson or Young score.
They call this problem DodgsonScore-Bribery and provide positive results for it. Notice, however, that
finding a bribery which makes c⋆ have a certain Dodgson score does not prevent another candidate to have
a lower score and winning the bribed election. Thus, solving DodgsonScore-Bribery can be very far from
the desired result.
Second, the authors then observe that a brute force approach enumerating all V C! briberies solves
the Dodgson-Bribery problem in polynomial time for constantly many candidates; however, theirs is not a
fixed-parameter algorithm for parameter C.
Third, they then introduce another voting system called Dodgson′, which is similar to Dodgson, and
provide a fixed-parameter algorithm for winner determination. However, as in the case of Dodgson-Bribery,
they do not provide a fixed-parameter algorithm for Dodgson′-Bribery.
The issue is then that a subsequent paper of Falisezwski et al. [FHH11] claims that the Dodgson rule is
"integer-linear-program implementable" and that this implies a certain election control problem generalizing
Bribery to be fixed-parameter tractable [FHH11, Theorem 6.2]. We believe the authors do not sufficiently
differentiate between determining the winner with one ILP, as is the case for most simple voting rules, and
with multiple ILPs, as is the case for Dodgson. Thus, we believe there is no evidence that the Dodgson
rule is "integer-linear-program implementable". Yet, this may be possible and this question still deserves
attention. Whatever the reason, we are convinced that their [FHH11, Theorem 6.1] does not hold for
R=Dodgson. Hence, we believe that ours are the first fixed-parameter algorithms for any Bribery-like
problem for R ∈ {Dodgson, Young}.
2 Preliminaries
Let m, n be integers. We define [m, n] := {m, m + 1, . . . , n} and [n] := [1, n]. Throughout, we reserve bold
face letters (e.g. x, y) for vectors. For a vector x its i-th coordinate is xi.
Next, we provide notions and notations for R-Swap Bribery.
Elections. An election (C, V ) consists of a set C of candidates and a set V of voters, who indicate their
preferences over the candidates in C, represented via a preference order ≻v which is a total order over C.
We often identify a voter v with their preference order ≻v. Denote by rank(c, v) the rank of candidate c
in ≻v; v's most preferred candidate has rank 1 and their least preferred candidate has rank C. For distinct
candidates c, c′ ∈ C, write c ≻v c′ if voter v prefers c over c′.
Swaps. Let (C, V ) be an election and let ≻v∈ V be a voter. For candidates c, c′ ∈ C, a swap s = (c, c′)v
v. A swap (c, c′)v is
means to exchange the positions of c and c′ in ≻v; denote the perturbed order by ≻s
admissible in ≻v if rank(c, v) = rank(c′, v) − 1. A set S of swaps is admissible in ≻v if they can be applied
sequentially in ≻v, one after the other, in some order, such that each one of them is admissible. Note that
the perturbed vote, denoted by ≻S
v , is independent from the order in which the swaps of S are applied. We
extend this notation for applying swaps in several votes and denote it V S. We specify v's cost of swaps by
a function σv : C × C → Z.
Voting rules. A voting rule R is a function that maps an election (C, V ) to a subset W ⊆ C, called the
winners. A candidate c ∈ C is a Condorcet winner if any other c′ ∈ C \ {c} satisfies (cid:12)(cid:12){≻v∈ V c ≻v c′}(cid:12)(cid:12) >
(cid:12)(cid:12){≻v∈ V c′ ≻v c}(cid:12)(cid:12); then we say that c beats c′ in a head-to-head contest. The Young score of c ∈ C is
the size of the smallest subset V ′ ⊆ V such that c is a Condorcet winner in (C, V \ V ′). Analogously, the
Dodgson score of c ∈ C is the size of the smallest admissible set of swaps S such that c is a Condorcet winner
in (C, V S). Then, c is a Young (Dodgson) winner if it has minimum Young (Dodgson) score.
We aim to solve the following problem:
6
R-Swap Bribery
Input: An election (C, V ), a designated candidate c⋆ ∈ C and swap costs σv : C × C → Z for
Find:
v ∈ V .
A set S of admissible swaps of minimum cost so that c⋆ wins the election (C, V S) under
the rule R.
3 Moves in Societies and Presburger Arithmetic
Let τ ∈ N be the number of types of people.
Definition 5. A society is a non-negative τ -dimensional integer vector s = (s1, . . . , sτ ).
In most problems, we are interested in modifying a society by moving people between types.
Definition 6. A move is a vector m = (m1,1, . . . , mτ,τ ) ∈ Zτ 2
Intuitively, mi,j is the number of people of type i turning type j.
Definition 7. A change is a vector ∆ = (∆1, . . . , ∆τ ) ∈ Zτ whose elements sum up to 0. We say that ∆
is the change associated with a move m if ∆i = Pτ
j=1 mj,i − mi,j, and we write ∆ = ∆(m). A change ∆ is
feasible with respect to society s if s + ∆ ≥ 0, i.e., if applying the change ∆ to s results in a society.
.
One more useful notion is that of a move costs vector:
Definition 8. A move costs vector is a vector c = (c1,1, . . . , cτ,τ ) in (N∪{+∞})τ 2
inequality, i.e., ci,k ≤ ci,j + cj,k for all distinct i, j, k.
which satisfies the triangle
Definition 9 ((c, k)-move). Let k ∈ N and c be a move costs vector. A move m is a (c, k)-move if c⊺m ≤ k.
Finally, we want to check that the society (e.g., resulting from applying some moves) satisfies a certain
desired condition. This condition depends on the problem we are modeling: in variants of bribery, it says
that a preferred candidate is elected as a winner or to be a part of a committee under a given voting rule; in
the context of lobbying, it says that a preferred agenda was selected. To allow large expressibility, we make
a very broad definition:
Definition 10. A winning condition of width τ is a predicate Ψ(s) with τ free variables.
3.1 Presburger Arithmetic
For two formulas Φ and Ψ, we write Φ ∼= Ψ to denote their equivalence.
Definition 11 (Presburger Arithmetic). Let
P0,(n0),δ,γ,α,β = n Ψ(x0)o
be the set of quantifier-free Presburger Arithmetic (PA) formulas with n0 free variables x0 which are a
disjunction of at most δ conjunctions of linear inequalities a⊺x0 ≤ b, each of length at most γ, where
kak∞ ≤ α and b ≤ β for each inequality. Then, let
P0,(n0),δ,γ,α,β = nΨ(x0) ∃ Ψ(x0) ∈ P0,(n0),δ,γ,α,β : Ψ ∼= Ψo
be the set of PA formulas equivalent to some DNF formula from P0,(n0),δ,γ,α,β. Finally, let
Pk,n,δ,γ,α,β = {Ψ(x0) ≡ ∃/∀x1∃/∀x2 · · · ∃/∀xk : Φ(x0, x1, . . . , xk)}
be the set of PA formulas with quantifier depth k, n0 free variables x0, and dimension n = (n0, n1, . . . , nk);
here, xi ∈ Zni for each i and Φ(x0, . . . , xk) ∈ P0,(n),δ,γ,α,β with n = Pk
i=1 ni. The length of Ψ(x0) is the
number of symbols it contains, which is polynomially bounded in (n, γ, δ). By ∃P and ∀P we denote the
sets of PA formulas whose leading quantifier is ∃ or ∀, respectively.
7
∃/∀x1
...
∃/∀xk
(∨), δ
(∧), γ
(∧), γ
. . .
(∧), γ
Φ
≤ γ
a1 · (x0, . . . , xk) ≤ b1
...
ap · (x0, . . . , xk) ≤ bp
≤ α
≤ β
Figure 1: Visualization of PA formula Ψ(x0) in DNF.
Example. A simple example of PA is the following formula.
Ψ(y) ≡ ∀x1x2∃z1z2z3
:
(x1 + y = z3 ∧ y ≥ 0) ∨
(cid:0)3x1 + 10y − 3z1 ≤ 13 ∧ 2x2 + 5y − z2 ≤ 11
∧ x1 + 1y − z3 ≥ 9 ∧ z1 − z2 + 2z3 ≤ 6(cid:1)
Here k = 2, n = (1, 2, 3) , δ = 2, γ = 4, α = 10, and β = 13.
We study winning conditions Ψ(s) expressible in PA, and state our complexity results with respect to
the descriptive complexity of Ψ, which is its number of variables, quantifiers, logical connectives, and unary
encoding length of coefficients and constants.
Vocabulary. We express relevant definitions by simple PA formulas over integral variables and with integer
coefficients and constants:
• society(s) ≡ s ≥ 0 ∈ P0,(τ ),0,τ,1,0,
• move(m) ∈ P0,(τ 2),0,0,0,0,
• ∆ = ∆(m) is a linear map ∆i = Pτ
Ψ(m, ∆) ∈ P0,(τ +τ 2),0,τ,1,0
j=1 mj,i − mi,j; thus if we let Ψ(m, ∆) ≡ ∆ = ∆(m), then
• feasible(s, ∆) ≡ (s + ∆(m) ≥ 0 ∧ 1⊺∆ = 0) ∈ P0,(2τ ),0,τ,2,0, and,
• (c, k)-move(m) ≡ c⊺m ≤ k ∈ P0,(τ ),0,1,kck∞,k.
We note that, for elections, our definition of winning condition generalizes the notion of "linearly-definable
voting rules" by Dorn and Schlotter [DS12]. Precisely, those rules belong to ∃P with k = 1; we will show
that Dodgson and Young are in ∃P with k = 2. Thus, our winning conditions capture an extensive set of
voting rules.
3.2 Modeling Problems as Minimum Move
We model moves in societies by the following general problem:
Minimum Move
Input: A society s, an objective function f : Zτ 2
Find:
A move m minimizing f (m) s.t. Ψ(s + ∆(m)) ∧ feasible(s, ∆(m)).
→ Z, a winning condition Ψ.
8
It models many well-studied problems:
Multi bribery. Knop et al. [KKM17c] introduce a generalization of various bribery problems called R-
Multi Bribery. Informally, we are given an election (C, V ) where each voter further has an approval count,
and is either active or latent the status of which can be changed at certain cost; likewise, there are costs for
perturbing their preference order or approval count. This problem generalizes Bribery, $Bribery, Swap
Bribery, Shift Bribery, Support Bribery, Extension Bribery, Possible Winner, Constructive
Control by Adding/Deleting Voters and other problems.
Notice that there are at most C! possible preference orders, at most C possible approval counts, and
2 states "active" or "latent". Thus, there are at most τ ≤ 2C · C! potential types of voters, and we
can express the input election as a society s. A move costs vector c describing the costs of moving a voter
from one type to another is obtained by calculating (possibly using a shortest path algorithm) the least costs
based on the given cost functions. Let Ψ(s) be a PA formula which is satisfied if the preferred candidate wins
under the voting rule R in a society s. Then, a bribery of minimum cost in a R-Multi Bribery instance
can be modeled as solving Minimum Move with f (m) = c⊺m. This modeling, combined with Theorem 2
and Corollary 17, yields Corollary 3.
Multiwinner elections. Bredereck et al. [BFN+16] study the complexity of Shift Bribery in committee
elections, that is, in elections with multiple winners. The modeling is exactly the same as above, except for
the winning condition Ψ which will be a long disjunction over all committees which include the preferred
candidate.
Lobbying in referenda. Bredereck et al. [BCH+14] study the complexity of Lobbying in referenda. There,
voters cast ballots with their "yes"/"no" answers to issues. The task is to push an agenda, i.e., a certain
outcome. Again, voters fall into groups according to their ballots, the costs of changing their opinions forms
a move costs vector, and a winning condition Ψ expresses that the selected agenda succeeded.
4 Sentences With Two Quantifiers
We shall now introduce the building blocks of our proof of Theorem 1. Woods [Woo15] gives an algorithm
that efficiently converts any quantifier-free PA formula Φ into an equivalent DNF formula Φ of bounded
length:
Lemma 12 (Woods [Woo15, Proposition 5.1]). Let Φ(x) be a quantifier-free PA formula with x ∈ Zd
containing N inequalities, whose coefficients and right-hand sides are bounded in absolute value by α and
β, respectively. Then Φ(x) can be converted into an equivalent DNF formula Φ(x) with at most δ = N O(d)
disjunctions, each containing at most N conjunctions with the same bound on α and β.
It is often useful for the quantifiers of a PA formula to range over integer points of polyhedra, e.g. ∀x ∈ Q
(we do not write Q ∩ Zn for brevity, as we assume everything to be integer); again, our definition is not
restrictive by the fact that we can always rearrange:
Ψ(x0) ≡ ∃x1 ∈ Q1 · · · ∀/∃xk ∈ Qk : Φ(x0, x1, . . . , xk) ≡
∃x1 · · · ∀/∃xk :(cid:16)Φ(x0, x1, . . . , xk)
^(x1 ∈ Q1 ∧ x3 ∈ Q3 · · · )∧_(x2 6∈ Q2 ∨ x4 6∈ Q4 · · · )(cid:17)
Parametric ILP. A special case of PA are parametric ILPs, which can be viewed3 as deciding the sentence
∀x ∈ Zp : Ax ≤ b ∃y ∈ Zn : B(x, y) ≤ e,
where A ∈ Zℓ×p and B ∈ Zm×n are integer matrices. A consequence of an algorithm of Eisenbrand and
Shmonin [ES08] is the following:
3Parametric ILPs are typically viewed as ILPs with a varying right hand side, that is, deciding the sentence ∀b∃x : Ax ≤ b;
it is known that our formulation is equivalent, as shown by Crampton et al. [CGKW17], who call it ILP Resiliency.
9
Corollary 13 ([ES08, Theorem 4.2], [CGKW17, Corollary 1]). Any parametric ILP whose entries of A, B,
b and e are given in unary, is fixed-parameter tractable when parameterized by n, m and p.
ILP and disjunctions. We shall use a folklore result about implementing disjunctions in ILP when the
domains of variables can be bounded. For that, we need another definition.
Definition 14 (B-bounded, B-small PA formula). Let Ψ(x0) ≡ ∃/∀x1 · · · ∃/∀xk : Φ(x0, . . . , xk) be a PA
formula, and let ΨB(x0) ≡ ∃/∀x1 ∈ [−B, B]n1 · · · ∃/∀xk ∈ [−B, B]nk : Φ(x0, . . . , xk). Then we say that
Ψ(x0) is B-bounded if
{x ∈ Zn0 ΨB(x)} ∩ [−B, B]n0 = {x ∈ Zn0 Ψ(x)} ,
i.e., the set of feasible solutions does not change by restricting all quantifiers and free variables to the
corresponding box of size B.
Moreover, we say that any Ψ ∈ Pk,n,δ,γ,B,B is B-small if it is B-bounded, that is, its coefficients and
constants are bounded by B.
A special case are ILPs which are B-small; they correspond to PA formulas with k = 0 and δ = 0; for
such formulas we show:
Lemma 15 (ILP disjunctions [folklore]). Let Aix ≤ bi for i ∈ [d] be B-small ILPs with Ai ∈ Zm×n for
each i ∈ [d]. Then, a (B2n)-small system Ax ≤ b with A ∈ Z(md+d+1)×(n+d) can be constructed in time
O (cid:16)dm + n + Pd
i=1hAi, bii(cid:17) such that
∃(x, y) ∈ Zn+d : A(x, y) ≤ b ⇐⇒ ∃x ∈ Zn : _
i∈[d]
Aix ≤ bi
.
Proof. Let M = B2n, let yi for i ∈ [d] be binary variables, and consider the following system:
d
X
i=1
yi = 1^ yi ≥ 0, Aix ≤ bi + M (1 − yi) for all i ∈ [d] .
Assume it has an integer solution y. Then there is an index i ∈ [d] such that yi = 1 and thus Aix ≤ bi + 0
holds; thus, the system Aix ≤ bi has an integer solution. In the other direction, assume that the system
Aix ≤ bi has a solution x; then let yi = 1. We shall prove that Ai′ x ≤ bi′ + M holds for all i′ 6= i. Since
Ai′ x ≤ bi′ is B-small, each of its row sums has n terms which are a multiple of two numbers, each bounded
by B, and thus is at most B2n. Moreover, since yi = 1, we have yi′ = 0 and thus the right hand side is
bi′ + M and every assignment of x feasible for Aix ≤ bi satisfies it. Clearly, the new system has n + d
variables, md + d + 1 inequalities, is B-bounded and kAk∞ = B2n and thus it is (B2n)-small, and can be
constructed in the claimed time.
We now prove Theorem 1; we restate it in formal terms here:
Theorem 16 (formal version of Theorem 1). Let Ψ be a β-small P2,n,δ,γ,α,β sentence (i.e., without free
variables and thus n0 = 0). Then Ψ can be decided in time g(n, δ, γ) poly(α, β) for some computable function
g.
Proof. Let Ψ ≡ ∃x1∀x2 : Φ(x1, x2). Clearly, to decide Ψ we can instead decide ¬Ψ ≡ ∀x1∃x2 : ¬Φ(x1, x2).
Consider the formula ¬Φ(x1, x2): by Lemma 12, there exists an equivalent β-small DNF formula ¯Φ(x1, x2)
such that the number of its disjunctions and conjunctions is a function of just the original δ and γ.
Thus, from now on focus on the case Ψ ≡ ∀x1∃x2 : Φ(x1, x2). Our next task is to construct an
instance of Parametric ILP equivalent to deciding Ψ. To do that, replace Φ(x1, x2) with Ξ(x1, x2) ≡
∃x3 : A(x1, x2, x3) ≤ b, where x3 is of dimension δ and the system of linear inequalities A(x1, x2, x3) ≤ b
has bounded length, coefficients and right sides. Now we use Lemma 15. Assume that Φ(x1, x2) is a
disjunction of δ linear systems, each of at most γ conjunctions, and by the assumptions of the theorem
we know that Ψ is β-small. Plugging into Lemma 15, we have B = β, d = δ, m = γ and n = n1 + n2,
and we obtain a formula ∃x3 : A(x1, x2, x3) ≤ b equivalent to Φ(x1, x2). Thus, we are left with deciding
∀x1∃(x2, x3) : A(x1, x2, x3) ≤ b with the following parameters:
10
• The coefficients and right-hand sides k(A, b)k∞ are bounded by some computable function f (n, δ, γ)·(αβ)2.
• The dimensions of A are bounded by f (n, δ, γ).
Thus, we are in the setting of Corollary 13 and we can decide the above sentence by a fixed-parameter
algorithm.
Corollary 17. Let Ψ(x0) be a disjunction of D many β-small ∃P2,(τ,n1,n2),δ,γ,α,β formulas. Then
Minimum Move with objective f (x0) = c⊺x0 and winning condition Ψ can be solved in time
g(τ, n1, n2, δ, γ) poly(D, α, β, c) for some computable function g.
Proof. Let Ψ ≡ Ψ1 ∨ · · · ∨ ΨD, and let Ξi(m) ≡ Ψi(s + ∆(m)) ∧ (c, B)-move(m) ∧ feasible(∆(m), s), for
each i ∈ [D]. Now we need to perform binary search with parameter B on the sentence
∃mΨ(s + ∆(m) ∧ (c, B)-move(m) ∧ feasible(∆(m), s) .
This sentence is obviously equivalent to
∃mΞ1(m) ∨ · · · ∨ ∃mΞD(m),
which can be decided by D application of Theorem 16. This is possible in the claimed time, as, for each
i ∈ [D], ∃mΞi(m) has quantifier depth 2, is β′-small for β′ = poly(β, τ + n1 + n2, kck∞) and has dimension
n = (0, τ + n1, n2). A minimum move can then be constructed by coordinate-wise binary search; cf. [KQR10,
Thm. 4.1].
4.1 Application: Swap Bribery for the Dodgson Rule and Young Rule
We now prove Theorem 2 by giving our fixed-parameter algorithm for R-Swap Bribery with R the Dodgson
Rule or the Young Rule.
Proof of Theorem 2. Fix an instance (C, V, c⋆, σ) of R-Swap Bribery. As there are C! possible total
orders on C, each voter has one of these orders. Thus we view the election as a society s = (s1, . . . , sC!).
Let cswap be a move costs vector defined as ci,j ="swap distance between types i and j" (this is simply the
number of inversions between the permutations i and j [EFS09, Proposition 3.2]); observe that a society s
is in swap distance at most d from s′ if s′ = s + ∆(m) and m is a feasible (cswap, d)-move for s. Moreover,
let cdel be defined as ci,0 = 1 for every type i and ci,j = +∞ for every two types i, j 6= 0, where 0 is a type
for latent voters; observe that s is in voter deletion distance at most d from s′ if s′ = s + ∆(m) and m is a
(cdel, d)-move.
Our plan is to express the winning condition for R using a PA formula ΨDodgson(s) which is a disjunction
of polynomially many (in V and C) formulas from ∃P2,n,δ,γ,α,β, and then solve Minimum Move with
f (m) = σ⊺m and Ψ = ΨDodgson(s) using Corollary 17; recall that since the instance is uniform, we have
σv ≡ σ for all v ∈ V , and we let σ be the move costs vector obtained from σ. For ΨYoung, we proceed
analogously.
In the end, we will verify that n, δ, γ are bounded by a function of C, and that α, β are
polynomial in the input size.
Expressing ΨDodgson and ΨYoung. The winning condition for candidate c⋆ in both the Dodgson and
Young rule can be viewed as c⋆ being closest to being a Condorcet winner with respect to some distance
measure. Specifically, for the Dodgson rule, this distance is the number of swaps in the preference orders,
and in the Young rule, the distance is the number of voter deletions. For this reason, finding a bribery which
makes c⋆ a winner corresponds to finding a bribery which, for some d ∈ N, makes c⋆ be in distance d or less
from being a Condorcet winner, and, simultaneously, making every other candidate c 6= c⋆ be in distance at
least d from being a Condorcet winner.
Let us fix d ∈ N; later we will argue that we can go through all D relevant choices of d. First, we will
express the condition that in a society s, candidate c beats candidate c′ in a head-to-head contest:
beats(c, c′, s) ≡ X
i:c≻ic′
si > X
i:c′≻ic
si,
11
where ≻i is the preference order shared by all voters of type i. Then, it is easy to express that c⋆ is a winner
in s under Condorcet's rule:
ΦCondorcet(c, s) ≡ ^
c′6=c
beats(c, c′, s) .
Finally, we express that c⋆ is a Dodgson-winner with score d by ΨDodgson,d as follows:
ΨDodgson,d(s) ≡(cid:0)∃m : (cswap, d)-move(m) ∧ feasible(∆(m), s)∧
ΦCondorcet(c⋆, s + ∆(m))(cid:1)∧
(cid:0)∀m : (cswap, d − 1)-move(m) ∧ feasible(∆(m), s)∧
^
c6=c⋆
¬ΦCondorcet(c, s + ∆(m))(cid:1)
Then, ΨDodgson ≡ ΨDodgson,1 ∨ · · · ∨ ΨDodgson,D. Clearly, ΨYoung,d is obtained simply by replacing cswap with
cdel.
Complexity. The number D of relevant choices of d is bounded by C2V : at most C2 swaps suffice for
any bribery of a single voter, and there are at most V voters, thus D ≤ C2V is a bound on the Dodgson
score of a candidate. For the Young score, D ≤ V .
Since for each d ∈ [D], ΨDodgson,d ≡ ∃m : Ξ1(m) ∧ ∀m : Ξ2(m) can be equivalently rewritten as
∃m∀m′ : Ξ1(m) ∧ Ξ2(m′), it has quantifier depth 2, and thus belongs to ∃P2,n,δ,γ,α,β. Let us determine the
parameters:
• n is the vector of dimensions; thus n0 = τ = C! are the dimensions of a vector encoding a society, and
n1 = τ 2 = C!2 are the dimensions of a vector encoding a move.
• δ is the number of disjunctions, and it is polynomial in C,
• γ is the number of conjunctions, and it is polynomial in τ = C!,
• α is the largest coefficient, which is kcswapk∞ ≤ C2,
• β is the largest right-hand side, which is d ≤ C2V .
Thus, n, δ, γ are functions of the parameter C and α, β are polynomial in the size of the input election
(C, V, {≻v v ∈ V }), as required by Corollary 17. Analogous analysis applies to ΨYoung,d.
Replacing cswap with 1⊺ produces the winning condition for Dodgson′ as introduced by Faliszewski et
al. [FHH09]. Furthermore, it is interesting to consider voting rules obtained by replacing Condorcet's rule
in the definition of Dodgson's and Young's rule. For example, the Majority rule also might not produce a
winner, and most rules (Copelandα, Scoring protocol etc.) allow ties. Let the score of a candidate be their
distance (swap, deletion, etc.) from being a (unique) winner under rule R. We remark that if we replace
Condorcet's rule with R with a "simple" PA desciription, it corresponds to replacing ΦCondorcet in the proof
above, thus yielding fixed-parameter tractable algorithms for all such rules as well.
4.2 Application: Resilient and Robust Moves
Theorem 1 allows us to develop fixed-parameter algorithms for problems related to moves in society. Problem
Robust Move asks for a move that is robust to any adversary move of cost at most Ba:
Robust Move
Input: A society s, move costs vectors c and ca, a winning condition Ψ, budgets B, Ba ∈ N
Find:
A (c, B)-move m such that for every adversary (ca, Ba)-move ma, Ψ(s + ∆(ma) + ∆(m))
holds.
12
The second problem, Resilient Budget, asks if a budget B suffices to counter any adversary move of
cost at most Ba. Crampton et al. [CGKW17] consider a specialization of this problem for R-Swap Bribery.
Resilient Budget
Input: A society s, move costs vectors c and ca, a winning condition Ψ, budgets B, Ba ∈ N
Find:
Does for every adversary (ca, Ba)-move ma exist a (c, B)-move m such that Ψ(s + ∆(ma) +
∆(m))?
Theorem 18. Robust Move and Resilient Budget with Ψ ∈ P0,(n0),δ,γ,α,β can be solved in time
g(τ, n0, δ, γ)(α + β + B + Ba + kck∞ + kcak∞)O(1), that is, FPT parameterized by τ + n0 + δ + γ.
Proof. We apply Theorem 1 to decide the following formulas, which is clearly equivalent to deciding the
problems at hand:
Ψ(s)RB ≡∃m : feasible(s, ∆(m)) ∧ (c, B)-move(m) ∧
∀ma : feasible(s + ∆(m), ∆(ma)) ∧ (ca, Ba)-move(ma) ∧
Ψ(s + ∆(m) + ∆(ma))
Ψ(s)RM ≡∀ma : feasible(s, ∆(ma)) ∧ (ca, Ba)-move(ma) ∧
∃m : feasible(s + ∆(ma), ∆(m)) ∧ (c, B)-move(m) ∧
Ψ(s + ∆(m) + ∆(ma))
5 Polynomially Many Types
We now prove Theorem 4, which formally reads:
Theorem 19. Let Ψ(s) ∈ P0,(τ,n1),δ,γ,α,β be a winning condition. Minimum Move can be solved in time
g(γ, α)(τ + δ)O(1) log(s) for any linear function f (m) = c⊺m, where g is some computable function.
Proof. Since k = 0, Φ(y) is a disjunction of δ linear systems Aiy ≤ bi, with kAik∞ ≤ α and with at most
γ rows, for every i ∈ [δ]. Thus we instead solve δ instances of Minimum Move with Φi(y) ≡ Aiy ≤ bi
and pick the best solution among them. So from now on assume that Φ(y) ≡ Ay ≤ b with kAk∞ ≤ α and
A ∈ Zγ×τ .
Observe that A can have at most αO(γ) different columns. For two types i, j ∈ [τ ], we say they are
equivalent and write i ∼ j if the columns Ai and Aj are identical. Thus, the τ types of people fall into
C ≤ αO(γ) equivalence classes. For every type i ∈ [τ ], let C[i] = {j ∈ [τ ] j ∼ i} be the equivalence class
containing i, and let ¯A ∈ Zγ×C be a matrix with, for every i ∈ [C], ¯Ai = Aj where j ∈ C[i]. Now, for every
type i ∈ [τ ], we shall create a reduced custom move costs vector ci ∈ NC 2
. For every j ∈ [C], i 6= j, let
ci
i,j = min
j′∈C[j]
ci,j′
be the cost of moving from i to the cheapest equivalent of j, and let ci
¯c = (c1, . . . , cτ )
j,k = +∞ for any i 6= j, k ∈ [C]. Let
Then, consider the following ILP with variables x = (x1, . . . , xτ ) ∈ Zτ C; we obtain the minimum move
m from its optimal solution by taking mi,j = xi
j′ where j′ ∼ j:
min ¯c⊺x s.t. X
i=1,...,τ
¯Axi ≤ b, X
j=1,...,C
xi
j = si
∀i ∈ [τ ] .
This ILP is a combinatorial pre-n-fold IP ; by Knop et al. [KKM17a] (detailed in [KKM17b, Corollary 23]),
it can be solved in the claimed time.
13
6 Discussion
We raise three important questions which naturally arise from this work. First, we ask whether our fixed-
parameter algorithm for {Dodgson,Young}-Swap Bribery with uniform cost extends to general cost func-
tions. This parallels [BCF+14, Challenge #2]. For much simpler voting rules an analogous result was
shown only recently [KKM17c]. We believe that if the answer is positive, proving it would require provid-
ing new powerful integer programming tools, in particular, some analogue of Theorem 1 for n-fold integer
programming, which is the engine behind the recent progress [KKM17c].
Second, can the run time of our algorithm be improved? This analogously recalls [BCF+14, Challenge
#1]. The run time of Corollary 17 is double-exponential in the dimension O(τ ); thus Theorem 2 shows
that Dodgson-Swap Bribery is solvable in triple-exponential time in parameter C. We believe it can be
improved, but we are sceptical that it could be made single-exponential, and thus ask if a double-exponential
lower bound holds. A related question is to show lower bounds separating the complexity of R-Swap
Bribery for different voting rules such as Borda, Copeland and Kemeny.
Third, we note that, unlike most previous results, the run time of our algorithm depends polynomially
on the number V of voters. Many previous results solve the succinct variant of the problem (cf. Falizewski
et al. [FHH09]) and depend polynomially only on log V . Thus we ask whether {Dodgson,Young}-Swap
Bribery is fixed-parameter tractable also in the succinct variant.
Finally, we discuss our usage of Presburger arithmetic as a generalization of ILPs. Solving ILPs amounts
to deciding ∃x : Ax ≤ b; by Lenstra's algorithm [Len83] and its improvements by Kannan [Kan87] and
Frank and Tardos [FT87], this task is fixed-parameter tractable parameterized by the dimension of x even
for unbounded kA, bk∞ and if A has polynomially (in the length of the input) many rows. In 1990, Kannan
claimed to show that Parametric ILP, which amounts to deciding ∀b ∈ Q ∃x : Ax ≤ b for some
polyhedron Q, is fixed-parameter tractable parameterized by the dimension of x; here, kAk∞ must be
bounded by a polynomial and the number of rows of A also has to be a parameter. However, Kannan's result
relies on Kannan's Partitioning Theorem (KPT), which was recently disproved by Nguyen and Pak [NP17].
Nguyen and Pak [NP17, Theorem 1.9] state that Woods [Woo15] gave a polynomial-time algorithm for
deciding ∀y∃x : Φ(x, y) when the dimensions of x and y are constant ; however, it is unclear if this is a fixed-
parameter algorithm. For this reason, as we have the dimensions of x and y as (non-constant) parameter,
we chose to prove Theorem 1 as a slightly weaker result (still sufficient for our purposes) but using only
elementary techniques.
References
[BCF+14] Robert Bredereck, Jiehua Chen, Piotr Faliszewski, Jiong Guo, Rolf Niedermeier, and Gerhard J.
Woeginger. Parameterized algorithmics for computational social choice: Nine research challenges.
Tsinghua Sci. Tech., 19(4):358–373, 2014.
[BCH+14] Robert Bredereck, Jiehua Chen, Sepp Hartung, Stefan Kratsch, Rolf Niedermeier, Ondrej Such´y,
and Gerhard J. Woeginger. A multivariate complexity analysis of lobbying in multiple referenda.
J. Artificial Intelligence Res., 50:409–446, 2014.
[BFN+16] Robert Bredereck, Piotr Faliszewski, Rolf Niedermeier, Piotr Skowron, and Nimrod Talmon.
In Proc. AAAI 2016, pages 2452–2458,
Complexity of shift bribery in committee elections.
2016.
[BITT89]
John J. Bartholdi III, Craig A. Tovey, and Michael A. Trick. Voting schemes for which it can
be difficult to tell who won the election. Soc. Choice Welfare, 6(2):157–165, 1989.
[CGKW17] Jason Crampton, Gregory Gutin, Martin Kouteck´y, and R´emi Watrigant. Parameterized re-
siliency problems via integer linear programming. In Proc. CIAC 2017, volume 10236 of Lecture
Notes Comput. Sci., pages 164–176, 2017.
14
[DS12]
[EFS09]
[ES08]
[FHH09]
[FHH11]
Britta Dorn and Ildik´o Schlotter. Multivariate complexity analysis of swap bribery. Algorithmica,
64(1):126–151, 2012.
Edith Elkind, Piotr Faliszewski, and Arkadii Slinko. Swap bribery. In Proc. SAGT 2009, volume
5814 of Lecture Notes Comput. Sci., pages 299–310, 2009.
Friedrich Eisenbrand and Gennady Shmonin. Parametric integer programming in fixed dimen-
sion. Math. Oper. Res., 33(4), 2008.
Piotr Faliszewski, Edith Hemaspaandra, and Lane A. Hemaspaandra. How hard is bribery in
elections? J. Artificial Intelligence Res., 40:485–532, 2009.
Piotr Faliszewski, Edith Hemaspaandra, and Lane A. Hemaspaandra. Multimode control attacks
on elections. J. Artificial Intelligence Res., 40:305–351, 2011.
[FRRS14] Piotr Faliszewski, Yannick Reisch, Jorg Rothe, and Lena Schend. Complexity of manipulation,
In Proc. AAMAS 2014,
bribery, and campaign management in Bucklin and fallback voting.
pages 1357–1358, 2014.
Andr´as Frank and ´Eva Tardos. An application of simultaneous Diophantine approximation in
combinatorial optimization. Combinatorica, 7(1):49–65, 1987.
[FT87]
[HHR97]
[HSV05]
[Kan87]
Edith Hemaspaandra, Lane A. Hemaspaandra, and Jorg Rothe. Exact analysis of Dodgson
elections: Lewis Carroll's 1876 voting system is complete for parallel access to NP. J. ACM,
44(6):806–825, 1997.
Edith Hemaspaandra, Holger Spakowski, and Jorg Vogel. The complexity of Kemeny elections.
Theoret. Comput. Sci., 349(3):382–391, 2005.
Ravi Kannan. Minkowski's convex body theorem and integer programming. Math. Oper. Res.,
12(3):415–440, August 1987.
[KKM17a] Dusan Knop, Martin Kouteck´y, and Matthias Mnich. Combinatorial n-fold integer programming
and applications. In Proc. ESA 2017, volume 87 of Leibniz Int. Proc. Informatics, pages 54:1–
54:14, 2017.
[KKM17b] Dusan Knop, Martin Kouteck´y, and Matthias Mnich. Combinatorial n-fold integer programming
and applications. Technical report, 2017. https://arxiv.org/abs/1705.08657.
[KKM17c] Dusan Knop, Martin Kouteck´y, and Matthias Mnich. Voting and bribing in single-exponential
time. In Proc. STACS 2017, volume 66 of Leibniz Int. Proc. Informatics, pages 46:1–46:14, 2017.
[KQR10] Matthias Koppe, Maurice Queyranne, and Chris T. Ryan. Parametric integer programming
algorithm for bilevel mixed integer programs. J. Optim. Theory Appl., 146(1):137–150, 2010.
[Len83]
[NP17]
[RSV03]
[SFE17]
[Woo15]
Hendrik W. Lenstra, Jr. Integer programming with a fixed number of variables. Math. Oper.
Res., 8(4):538–548, 1983.
Danny Nguyen and Igor Pak. Complexity of short presburger arithmetic. In Proc. STOC 2017,
pages 812–820, 2017.
Jorg Rothe, Holger Spakowski, and Jorg Vogel. Exact complexity of the winner problem for
Young elections. Theory Comput. Syst., 36(4):375–386, 2003.
Ildik´o Schlotter, Piotr Faliszewski, and Edith Elkind. Campaign management under approval-
driven voting rules. Algorithmica, 77(1):84–115, 2017.
Kevin Woods. Presburger arithmetic, rational generating functions, and quasi-polynomials. J.
Symbolic Logic, 80(2):433–449, 2015.
15
|
1308.0813 | 3 | 1308 | 2015-08-05T04:43:12 | Multi-agent Systems with Compasses | [
"cs.MA"
] | This paper investigates agreement protocols over cooperative and cooperative--antagonistic multi-agent networks with coupled continuous-time nonlinear dynamics. To guarantee convergence for such systems, it is common in the literature to assume that the vector field of each agent is pointing inside the convex hull formed by the states of the agent and its neighbors, given that the relative states between each agent and its neighbors are available. This convexity condition is relaxed in this paper, as we show that it is enough that the vector field belongs to a strict tangent cone based on a local supporting hyperrectangle. The new condition has the natural physical interpretation of requiring shared reference directions in addition to the available local relative states. Such shared reference directions can be further interpreted as if each agent holds a magnetic compass indicating the orientations of a global frame. It is proven that the cooperative multi-agent system achieves exponential state agreement if and only if the time-varying interaction graph is uniformly jointly quasi-strongly connected. Cooperative--antagonistic multi-agent systems are also considered. For these systems, the relation has a negative sign for arcs corresponding to antagonistic interactions. State agreement may not be achieved, but instead it is shown that all the agents' states asymptotically converge, and their limits agree componentwise in absolute values if and in general only if the time-varying interaction graph is uniformly jointly strongly connected. | cs.MA | cs | Multi-agent Systems with Compasses
Ziyang Meng, Guodong Shi, and Karl Henrik Johansson∗
Abstract
This paper investigates agreement protocols over cooperative and cooperative -- antagonistic multi-
agent networks with coupled continuous-time nonlinear dynamics. To guarantee convergence for such
systems, it is common in the literature to assume that the vector field of each agent is pointing inside the
convex hull formed by the states of the agent and its neighbors, given that the relative states between
each agent and its neighbors are available. This convexity condition is relaxed in this paper, as we show
that it is enough that the vector field belongs to a strict tangent cone based on a local supporting
hyperrectangle. The new condition has the natural physical interpretation of requiring shared reference
directions in addition to the available local relative states. Such shared reference directions can be
further interpreted as if each agent holds a magnetic compass indicating the orientations of a global
frame. It is proven that the cooperative multi-agent system achieves exponential state agreement if and
only if the time-varying interaction graph is uniformly jointly quasi-strongly connected. Cooperative --
antagonistic multi-agent systems are also considered. For these systems, the relation has a negative sign
for arcs corresponding to antagonistic interactions. State agreement may not be achieved, but instead
it is shown that all the agents' states asymptotically converge, and their limits agree componentwise in
absolute values if and in general only if the time-varying interaction graph is uniformly jointly strongly
connected.
Keywords: shared reference direction, nonlinear systems, cooperative-antagonistic network
5
1
0
2
g
u
A
5
]
A
M
.
s
c
[
3
v
3
1
8
0
.
8
0
3
1
:
v
i
X
r
a
1
Introduction
In the last decade, coordinated control of multi-agent systems has attracted extensive attention due to
its broad applications in engineering, physics, biology and social sciences, e.g., [6, 15, 22, 26, 36]. A
∗A brief version of this work was presented in the 33rd Chinese Control Conference, Nanjing, China, July 2014 [23].
Z. Meng is with Department of Precision Instrument, Tsinghua University, Beijing 100084, China. G. Shi is with College
of Engineering and Computer Science, The Australian National University, Canberra, Australia. K. H. Johansson is with
ACCESS Linnaeus Centre, School of Electrical Engineering, Royal Institute of Technology, Stockholm, Sweden. This work
has been supported in part by the Knut and Alice Wallenberg Foundation, the Swedish Research Council, and the Alexander
von Humboldt Foundation of Germany. Email: [email protected], [email protected], [email protected].
1
Meng et al. Multi-agent Systems with Compasses
fundamental idea is that by carefully implementing distributed control protocols for each agent, collective
tasks can be reached for the overall system using only neighboring information exchange. Several important
results have been established, e.g., in the area of mobile systems including spacecraft formation flying,
rendezvous of multiple robots, and animal flocking [18, 8, 34].
Agreement protocols, where the goal is to drive the states of the agents to reach a common value using
local interactions, play a basic role in coordination of multi-agent systems. The state agreement protocol
and its fundamental convergence properties were established for linear systems in the classical work [35].
The convergence of the linear agreement protocol has been widely studied since then for both continuous-
time and discrete-time models, e.g., [5, 15, 29]. Much understandings have been established, such as the
explicit convergence rate in many cases [7, 24, 27, 28]. A major challenge is how to quantitatively char-
acterize the influence of a time-varying communication graph on the agreement convergence. Agreement
protocols with nonlinear dynamics have also drawn attention in the literature, e.g., [4, 14, 19, 25, 30, 31].
Due to the complexity of nonlinear dynamics, it is in general difficult to obtain explicit convergence rates
for these systems. All the above studies on linear or nonlinear multi-agent dynamics are based on the stand-
ing assumption that agents in the network are cooperative. Recently, motivated from opinion dynamics
evolving over social networks [10, 37], state agreement problems over cooperative -- antagonistic networks
were introduced [1, 2]. In such networks, antagonistic neighbors exchange their states with opposite signs
compared to cooperative neighbors.
In most of the work discussed above, a convexity assumption plays an essential role in the local inter-
action rule for reaching state agreement. For discrete-time models, it is usually assumed that each agent
updates its state as a convex combination of its neighbors' states [5, 15]. A precise characterization of this
convexity condition guaranteeing asymptotic agreement was established in [25]. For continuous-time mod-
els, an interpretation of this assumption is that the vector field for each agent must fall into the relative
interior of a tangent cone formed by the convex hull of the relative state vectors in its neighborhood [19].
The recent work [21] generalized agreement protocols to convex metric spaces, but a convexity assumption
for the local dynamics continued to play an important role in ensuring agreement convergence.
In this paper, we show that the convexity condition for agreement seeking of multi-agent systems
can be relaxed at the cost of shared reference directions. Such shared reference directions can be easily
obtained by a magnetic compass, with the help of which the direction of each axis can be observed
from a prescribed global coordinate system. Using the relative state information and the shared reference
direction information, each agent can derive a strict tangent cone from a local supporting hyperrectangle.
This cone defines the feasible set of local control actions for each agent to guarantee convergence to state
agreement. In fact, the agents just need to determine, through sensing or communication, the relative
2
Meng et al. Multi-agent Systems with Compasses
orthant of each of their neighbors' states. The vector field of an agent can be outside of the convex hull
formed by the states of the agent and its neighbors, so this new condition provides a relaxed condition
for agreement seeking. We remark that a compass is naturally present in many systems. For instance, the
classical Vicsek's model [36] inherently uses "compass"-like directional information and the calculation of
each agent's heading relies on the information where the common east is. In addition, scientists observed
that the European Robin bird can detect and navigate through the Earth's magnetic field, providing them
with biological compasses in addition to their normal vision [33]. Engineering systems, such as multi-robot
networks, can be equipped with magnetic compasses at a low cost [13, 32].
Under a general definition of nonlinear multi-agent systems with shared reference directions, we establish
two main results:
• For cooperative networks, we show that the underlying graph associated with the nonlinear inter-
actions being uniformly jointly quasi-strongly connected is necessary and sufficient for exponential
agreement. The convergence rate is explicitly given. This improves the existing results based on
convex hull conditions [25, 19].
• For cooperative-antagonistic networks, we propose a general model following the sign-flipping in-
terpretation along an antagonistic arc introduced in [2]. We show that when the underlying graph
is uniformly jointly strongly connected, irrespective with the sign of the arcs, all the agents' states
asymptotically converge, and their limits agree componentwise in absolute values.
The remainder of the paper is organized as follows. In Section 2, we give some mathematical preliminar-
ies on convex sets, graph theory, and Dini derivatives. The nonlinear multi-agent dynamics, the interaction
graph, the shared reference direction, and the agreement metrics are given in Section 3. The main results
and discussions are presented in Section 4. The proofs of the results are presented in Sections 5 and 6,
respectively, for cooperative and cooperative -- antagonistic networks. A brief concluding remark is given in
Section 7.
2 Preliminaries
In this section, we introduce some mathematical preliminaries on convex analysis [3], graph theory [12],
and Dini derivatives [11].
2.1 Convex analysis
For any nonempty set S ⊆ Rd, kxkS = inf y∈S kx − yk is called the distance between x ∈ Rd and S, where
k · k denotes the Euclidean norm. A set S ⊂ Rd is called convex if (1 − ζ)x + ζy ∈ S when x ∈ S, y ∈ S,
3
Meng et al. Multi-agent Systems with Compasses
and 0 ≤ ζ ≤ 1. A convex set S ⊂ Rd is called a convex cone if ζx ∈ S when x ∈ S and ζ > 0. The convex
hull of S ⊂ Rd is denoted co(S) and the convex hull of a finite set of points x1, x2, . . . , xn ∈ Rd denoted
co{x1, x2, . . . , xn}.
Let S be a convex set. Then there is a unique element PS (x) ∈ S, called the convex projection of x onto
S is continuously
S, satisfying kx − PS (x)k = kxkS associated to any x ∈ Rd. It is also known that kxk2
differentiable for all x ∈ Rd, and its gradient can be explicitly computed [3]:
∇kxk2
S = 2(x − PS (x)).
(1)
Let S ⊂ Rd be convex and closed. The interior and boundary of S is denoted by int(S) and ∂S, respectively.
If S contains the origin, the smallest subspace containing S is the carrier subspace denoted by cs(S). The
relative interior of S, denoted by ri(S), is the interior of S with respect to the subspace cs(S) and the
relative topology used. If S does not contain the origin, cs(S) denotes the smallest subspace containing
S − z, where z is any point in S. Then, ri(S) is the interior of S with respect to the subspace z + cs(S).
Similarly, we can define the relative boundary rb(S).
Let S ⊂ Rd be a closed convex set and x ∈ S. The tangent cone to S at x is defined as the set
= 0o. Note that if x ∈ int(S), then T (x,S) = Rd. Therefore, the
T (x,S) = nz ∈ Rd : lim inf ζ→0 kx+ζzkS
definition of T (x,S) is essential only when x ∈ ∂S. The following lemma can be found in [3] and will be
used.
ζ
Lemma 1 Let S1,S2 ⊂ Rd be convex sets. If x ∈ S1 ⊂ S2, then T (x,S1) ⊂ T (x,S2).
2.2 Graph theory
A directed graph G consists of a pair (V,E), where V = {1, 2, . . . , n} is a finite, nonempty set of nodes
and E ⊆ V × V is a set of ordered pairs of nodes, denoted arcs. The set of neighbors of node i is denoted
Ni := {j : (j, i) ∈ E}. A directed path in a directed graph is a sequence of arcs of the form (i, j), (j, k), . . . .
If there exists a path from node i to j, then node j is said to be reachable from node i. If for node i, there
exists a path from i to any other node, then i is called a root of G. G is said to be strongly connected if
each node is reachable from any other node. G is said to be quasi-strongly connected if G has a root.
2.3 Dini derivatives
Let D+V (t, x(t)) be the upper Dini derivative of V (t, x(t)) with respect to t, i.e., D+V (t, x) = lim supτ→0+
The following lemma [9] will be used for our analysis.
V (t+τ,x(t+τ ))−V (t,x(t))
τ
4
Meng et al. Multi-agent Systems with Compasses
Lemma 2 Suppose for each i ∈ V, Vi
: R × M → R is continuously differentiable. Let V (t, x) =
maxi∈V Vi(t, x), and let bV(t) = {i ∈ V : Vi(t, x(t)) = V (t, x(t))} be the set of indices where the maxi-
mum is reached at time t. Then D+V (t, x(t)) = maxi∈ bV(t)
Vi(t, x(t)).
3 Multi-agent Network Model
In this section, we present the model of the considered multi-agent systems, introduce the corresponding
interaction graph, and define some useful geometric concepts used in the control laws.
Consider a multi-agent system with agent set V = {1, . . . , n}. Let xi ∈ Rd denote the state of agent i.
Let x = (xT
1 , xT
2 , . . . , xT
n )T and denote D = {1, 2, . . . , d}.
3.1 Nonlinear multi-agent dynamics
f 1
p (x1, x2, . . . , xn)
Let P be a given (finite or infinite) set of indices. An element in P is denoted by p. For any p ∈
P, we define a function fp(x1, x2, . . . , xn) : Rdn → Rdn associated with p, where fp(x1, x2, . . . , xn) =
f n
p (x1, x2, . . . , xn)
Let σ(t) : [t0,∞) → P be a piecewise constant function, so, there exists a sequence of increasing time
p : Rdn → Rd, i = 1, 2, . . . , n.
instances {tl}∞0 such that σ(t) remains constant for t ∈ [tl, tl+1) and switches at t = tl.
...
with f i
The dynamics of the multi-agent systems is described by the switched nonlinear system
x(t) = fσ(t)(x(t)).
(2)
We place some mild assumptions on this system.
Assumption 1 There exists a lower bound τd > 0, such that inf l(tl+1 − tl) ≥ τd.
Assumption 2 fp(x) is uniformly locally Lipschitz on Rdn, i.e., for every x ∈ Rdn, we can find a neigh-
borhood Uα(x) = {y ∈ Rdn : ky − xk ≤ α} for some α > 0 such that there exists a real number L(x) > 0
with kfp(a) − fp(b)k ≤ L(x)ka − bk for any a, b ∈ Uα(x) and p ∈ P.
Under Assumptions 1 and 2, the Caratheodory solutions of (2) exist for arbitrary initial conditions,
and they are absolutely continuous functions for almost all t on the maximum interval of existence [11].
All our further discussions will be on the Caratheodory solutions of (2) without specific mention.
5
Meng et al. Multi-agent Systems with Compasses
3.2
Interaction graph
Having the dynamics defined for the considered multi-agent system, similar to [19], we introduce next its
interaction graph.
Definition 1 The graph Gp = (V,Ep) associated with fp is the directed graph on node set V = {1, 2, . . . , n}
p depends on xj, i.e., there exist xj, xj ∈ Rd such that
and arc set Ep such that (j, i) ∈ Ep if and only if f i
f i
p(x1, . . . , xj, . . . , xn) 6= f i
p(x1, . . . , xj, . . . , xn).
The set of neighbors of node i in Gp is denoted by Ni(p). The dynamic interaction graph associated
with system (2) is denoted by Gσ(t) = (V,Eσ(t)). The joint graph of Gσ(t) during time interval [t1, t2) is
defined by Gσ(t)([t1, t2)) = St∈[t1,t2) G(t) = (V,St∈[t1,t2) Eσ(t)). We impose the following definition on the
connectivity of Gσ(t), cf., [31].
Definition 2 Gσ(t) is uniformly jointly quasi-strongly (respectively, strongly) connected if there exists a
constant T > 0 such that G([t, t + T )) is quasi-strongly (respectively, strongly) connected for any t ≥ t0.
For each p ∈ P, the node relation along an interaction arc (j, i) ∈ Ep may be cooperative, or antagonistic.
p . To be
We assume that there is a sign, "+1" or "-1", associated with each (j, i) ∈ Ep, denoted by sgnij
precise, if j is cooperative to i, sgnij
p = +1, and if j is antagonistic to i, sgnij
p = −1.
Definition 3 If sgnij
p = +1, for all (j, i) ∈ Ep and all p ∈ P, the considered multi-agent network is called
a cooperative network. Otherwise, it is called a cooperative-antagonistic network.
3.3 Shared reference direction, hyperrectangle, and tangent cone
We assume that each agent has access to shared reference directions with respect to a common Cartesian
coordinate system. We use (−→r1 ,−→r2 , . . . ,−→rd) to represent the basis of that Rd Cartesian coordinate system.
Here −→rk = (0, . . . , 0, 1, 0, . . . , 0) denotes the unit vector in the direction of axis k ∈ D.
A hyperrectangle is the generalization of a rectangle to higher dimensions. An axis-aligned hyperrect-
angle is a hyperrectangle subject to the constraint that the edges of the hyperrectangle are parallel to the
Cartesian coordinate axes.
Definition 4 Let C ⊂ Rd be a bounded set. The supporting hyperrectangle H(C) is the axis-aligned hyper-
rectangle H(C) = [min(C)1, max(C)1] × [min(C)2, max(C)2] × ··· × [min(C)d, max(C)d], where by definition
min(C)k := miny∈C yk, max(C)k := maxy∈C yk, and yk denotes the kth entry of y.
In other words, a supporting hyperrectangle of a bounded set C is an axis-aligned minimum bounding
hyperrectangle.
6
Meng et al. Multi-agent Systems with Compasses
Figure 1: An example of the γ-strict tangent cone.
Definition 5 Let A ⊂ Rd be an axis-aligned hyperrectangle and γ > 0 a constant. The γ-strict tangent
cone to A at x ∈ Rd is the set
Tγ(x,A) =
if x ∈ ri(A)
cs(A);
T (x,A)Tk∈I{z ∈ Rd : hz,−→rki ≥ γDk(A)};
otherwise,
(3)
where I = {k ∈ D : x ∈ rbk(A)}, rbk(A) denotes the two facets of A perpendicular to the axis −→rk , and
Dk(A) = max(A)k − min(A)k denotes the side length parallel to the axis −→rk .
Figure 1 gives an example of the γ-strict tangent cone to A at x.
3.4 Agreement metrics
We next define uniformly asymptotic agreement and exponential agreement in this section.
Definition 6 The multi-agent system (2) is said to achieve uniformly asymptotic agreement on S0 ⊆ Rd
if
(i). point-wise uniform agreement can be achieved, i.e., for all η ∈ J , and ε > 0, there exists δ(ε) > 0
0 , and the
0 denotes S0×S0×. . . S0;
such that for all t0 ≥ 0, kx(t0) − ηk < δ ⇒ kx(t) − ηk < ε, ∀t ≥ t0, where x(t0) ∈ S n
agreement manifold is defined as J = {x ∈ S n
and
0 : x1 = x2 = ··· = xn} and S n
(ii). uniform agreement attraction can be achieved, i.e., for all ε > 0, there exist η(x(t0)) ∈ J and
T (ε) > 0 such that for all t0 ≥ 0, kx(t) − ηk < ε, ∀t ≥ t0 + T.
7
Meng et al. Multi-agent Systems with Compasses
!
"
Figure 2: Convex hull, supporting hyperrectangle, strict tangent cone and feasible vectors f i
p satisfying
Assumption 3.
Definition 7 The multi-agent system (2) is said to achieve exponential state agreement on S0 ⊆ Rd if
(i). point-wise uniform agreement can be achieved; and
(ii). exponential agreement attraction can be achieved, i.e., there exist η(x(t0)) ∈ J and k(S0) > 0,
λ(S0) > 0, such that for all t0 ≥ 0, kx(t) − ηk ≤ ke−λ(t−t0)kx(t0) − ηk.
4 Main Results
In this section, we state the main results of the paper.
4.1 Cooperative networks
We first study the convergence property of the nonlinear switched system (2) over a cooperative network
defined by an interaction graph. Introduce the local convex hull Ci
p(x) = co{xi, xj : j ∈ Ni(p)}. In order
to achieve exponential agreement, we propose the following strict tangent cone condition for the feasible
vector field.
Assumption 3 For all i ∈ V, p ∈ P, and x ∈ Rdn, it holds that f i
p(x) ∈ Tγ(xi,H(Ci
p(x))).
In Assumption 3, the vector f i
p can be chosen freely from the set Tγ(xi,H(Ci
p(x))). Hence, the assumption
specifies constraints on the feasible controls for the multi-agent system. Here Ci
hull formed by agent i and its neighbors, H(Ci
set Ci
example of the convex hull and the supporting hyperrectangle formed by agent 1 and its' neighbors. Two
p(x))) denotes the γ−strict tangent cone to H(Ci
p(x)) denotes the local supporting hyperrectangle of the
p(x), and Tγ(xi,H(Ci
p(x) denotes the convex
p(x)) at xi. Figure 2 gives an
feasible vectors f 1
p are also presented.
8
Meng et al. Multi-agent Systems with Compasses
In order to implement a controller compatible with Assumption 3, the agents need to determine, through
local sensing or communication, the relative orthant of each of their neighbors' states. This can be realized,
for instance, if each agent is capable of measuring the relative states with respect to its neighbors and is
aware of the direction of each axis of a prescribed global coordinate system. More specifically, when the
agent is in the interior of the hyperrectangle, the vector field for the agent can be chosen arbitrarily. When
the agent is on the boundary of its supporting hyperrectangle, the feasible control is any direction pointing
inside the tangent cone of its supporting hyperrectangle. Note that the absolute state of the agents is not
needed, but each agent needs to identify d − 1 absolute directions such that it can identify the direction
of its neighbors with respect to itself. For example, for the planar case d = 2, in addition to the relative
state measurements with respect to its neighbors, each agent just needs to be equipped with a compass.
The compass together with relative state measurements provide the quadrant location information of the
neighbors.
We state an exponential agreement result for the cooperative multi-agent systems.
Theorem 1 Suppose S0 is compact and that Assumptions 1, 2, and 3 hold. Then, the cooperative multi-
agent system (2) achieves exponential agreement on S0 if and only if its interaction graph Gσ(t) is uniformly
jointly quasi-strongly connected.
In order to compare the proposed "supporting hyperrectangle condition" with respect to the usual
convex hull condition [25, 19], we introduce the following assumption, which is a weaker condition than
Assumption 3.
Assumption 4 For all i ∈ V, p ∈ P, and x ∈ Rdn, it holds that f i
p(x) ∈ ri(cid:0)T (xi,H(Ci
p(x)))(cid:1).
We next present a uniformly asymptotic agreement result based on the relative interior condition of a
tangent cone formed by the supporting hyperrectangle.
Proposition 1 Suppose S0 is compact and that Assumptions 1, 2, and 4 hold. Then, the cooperative
multi-agent system (2) achieves uniformly asymptotic agreement on S0 if and only if its interaction graph
Gσ(t) is uniformly jointly quasi-strongly connected.
The proofs of Theorem 1 and Proposition 1 is deferred to Section 5.
Figure 3 illustrates the relative interior of a tangent cone of the convex hull (Assumption A2 of [19]),
relative interior of a tangent cone of the supporting hyperrectangle (Assumption 4), and strict tangent
cone of the supporting hyperrectangle (Assumption 3). It is obvious that the vector fields can be chosen
more freely under Assumption 4 than under Assumption A2 of [19]. On the other hand, strict tangent cone
9
Meng et al. Multi-agent Systems with Compasses
Figure 3: The relative interiors of a tangent cone of the convex hull and the supporting hyperrectangle,
and γ−strict tangent cone of the supporting hyperrectangle.
condition is a more strict condition than the relative interior condition of a tangent cone. However, expo-
nential agreement can be achieved under strict tangent cone condition while only uniformly asymptotic
agreement is achieved under the relative interior condition of a tangent cone.
Remark 1 Theorem 1 and Proposition 1 are consistent with the main results in [19, 21, 25]. Our analysis
relies on some critical techniques developed in [17, 19]. Proposition 1 allows that the vector field belongs
to a larger set compared with the convex hull condition proposed in [19, 21, 25]. In addition, we allow
the agent dynamics to switch over a possibly infinite set and we show exponential agreement and derive
in the proof of Theorem 1 the explicit exponential convergence rate. It follows that by sharing reference
directions in addition to the available local information, agreement of multi-agent systems has an enlarged
set of interactions and faster convergence speed compared with the case of using only local information.
To further illustrate Assumptions 3 and 4, we discuss two examples.
Example 1. Let us first consider Vicsek's model [36]. In particular, consider agent i ∈ V, moving in
the plane with position (xi(t), yi(t)), the same absolute velocity v, and the heading θi(t) at discrete time
t = 0, 1, . . . . The position and angle updates are described by
(4)
(5)
(6)
,
xi(t + 1) = xi(t) + v cos θi(t),
yi(t + 1) = yi(t) + v sin θi(t),
θi(t + 1) = arctan Pj∈Ni(t) sin θj(t)
Pj∈Ni(t) cos θj(t)
for all i ∈ V, where by convention it is assumed that i ∈ Ni(t). From (6), we see that Vicsek's model
inherently uses a "compass"-like directional information. Then, similar to the analysis of Theorem 1, we
can easily show that the first quadrant is an invariant set for (4) and (5). This can be verified by the fact
10
Meng et al. Multi-agent Systems with Compasses
Figure 4: An invariant set of Vicsek's model.
that θi(t + 1) ∈ [0, π
2 ] for all j ∈ Ni(t). Figure 4 illustrates this point for three agents.
At time t, the vector filed of all the agents are pointing inside the first quadrant, so agents construct
2 ] when θj(t) ∈ [0, π
an "unbounded" hyperrectangle (both the upper and right bounds are at infinity). This "unbounded"
hyperrectangle is the invariant set for the positions of all the agents. The existence of left and lower
bounds of the hyperrectangle guarantees that agents 1 and 2 satisfy Assumption 3. However, it is easy to
verify that agent 3 does not satisfy Assumption 3 since the upper and right bounds of the hyperrectangle
do not exist. Therefore, position agreement cannot be achieved in general for Vicsek's model.
Example 2. Consider the following dynamics for each agent i ∈ V:
xi = f i
σ(t)(x) = Ri
σ(t)(x) Xj∈Ni(σ(t))
aij(x)(cid:0)xj − xi(cid:1),
(7)
σ(t)(x) ∈ Rd×d is a
where aij(x) > 0 is a continuous function representing the weight of arc (j, i), and Ri
state-dependent rotation matrix which is continuous in x for any fixed σ ∈ P. Certainly the dynamics
described in (7) is beyond the convex hull agreement protocols [19, 21, 25]. With the results in Theorem 1
and Proposition 1, it becomes evident that the existence of Ri
σ(t)(x) may still guarantee agreement as long
as Ri
σ(t)(x) rotates the convex hull vector filed, Pj∈Ni(σ(t)) aij(x)(cid:0)xj − xi(cid:1), within the proposed tangent
cones given by the local supporting hyperrectangle. Certainly this does not mean that Ri
σ(t)(x) should be
sufficiently small since from Figure 3 this rotation angle can be large for proper x under certain interaction
rules. This can also be viewed as a structural robustness of the proposed "compass"-based framework.
4.2 Cooperative-antagonistic networks
Next, we study the convergence property of the cooperative -- antagonistic networks. Define C
co{xi, xjsgnij
p : j ∈ Ni(p)}. We impose the following assumption.
i
p(x) :=
11
Meng et al. Multi-agent Systems with Compasses
Assumption 5 For all i ∈ V, p ∈ P and x ∈ Rdn, it holds that f i
p(x) ∈ Tγ(xi,H(C
i
p(x))).
Assumption 5 follows the model for antagonistic interactions introduced in [2], where simple examples
can be found on that state agreement cannot always be achieved for cooperative -- antagonistic networks.
Instead, it is possible that agents converge to values with opposite signs, which is known as bipartite
consensus [2]. We present the following result for cooperative -- antagonistic networks.
Theorem 2 Let Assumptions 1, 2 and 5 hold. Then, if (and in general only if ) the interaction graph Gσ(t)
is uniformly jointly strongly connected, all the agents' trajectories asymptotically converge for cooperative-
antagonistic multi-agent system (2), and their limits agree componentwise in absolute values for every
initial time and initial state.
Here by "in general only if," we mean that we can always construct simple examples with fixed inter-
action rule, for which strong connectivity is necessary for the result in Theorem 2 to stand. The proof
of Theorem 2 will be presented in Section 6. Compared with the results given in [2], Theorem 2 requires
no conditions on the structural balance of the network. Theorem 2 shows that every positive or negative
arc contributes to the convergence of the absolute values of the nodes' states, even for general nonlinear
multi-agent dynamics.
The exponential agreement and uniformly asymptotical agreement results given in Theorem 1 and
Proposition 1 rely on uniformly jointly quasi-strong connectivity, while the result in Theorem 2 needs
uniformly jointly strong connectivity. For cooperative networks, we establish the exponential convergence
rate in the proof of Theorem 1. In contrast, for cooperative -- antagonistic networks in Theorem 2, the
convergence speed is unclear. We conjecture that exponential convergence might not hold in general under
the conditions of Theorem 2. The reason is that Lemmas 5 and 7 given in Section 5 cannot be recovered
for cooperative -- antagonistic networks.
We believe that differences between Theorems 1 and 2 discussed in the previous remarks reveal some
important distinctions of cooperative and cooperative -- antagonistic networks.
5 Cooperative Multi-agent Systems
In this section, we focus on the case of cooperative multi-agent systems. We will prove Theorem 1 and
Proposition 1 by analyzing a contraction property of (2), with the help of a series of preliminary lemmas.
5.1
Invariant set
We introduce the following definition.
12
Meng et al. Multi-agent Systems with Compasses
Definition 8 A set M ⊂ Rdn is an invariant set for the system (2) if for all t0 ≥ 0, x(t0) ∈ M =⇒
x(t) ∈ M, ∀t ≥ t0.
For all k ∈ D, define Mk(x(t)) = maxi∈V{xik(t)}, mk(x(t)) = mini∈V{xik(t)}, where xik denotes
kth entry of xi. In addition, define the supporting hyperrectangle by the initial states of all agents as
H0 := H(C(x(t0))), where C(x) = co{x1, x2, . . . , xn}.
In the following lemma, we show that the supporting hyperrectangle formed by the initial states of all
agents is an invariant set for system (2).
0 is an invariant set,
Lemma 3 Let Assumptions 1, 2, and 3 or Assumptions 1, 2, and 4 hold. Then, Hn
i.e., xi(t) ∈ H0, ∀i ∈ V, ∀t ≥ t0.
Proof. We first show that D+Mk(t) ≤ 0, ∀k ∈ D. Let bV(t) = {i ∈ V : xik(t) = Mk(t)} be the
set of indices where the maximum is reached at t. It then follows from Lemma 2 that for all k ∈ D,
D+Mk(t) = maxi∈ bV(t) xik = maxi∈ bV(t) f ik
σ(t). Consider
any initial state x(t0) ∈ Hn
p(x) ∈
p(x)))(cid:1) ⊆
Tγ(xi,H(Ci
p(x))) ⊆ T (xi,H(Ci
T (xi,H(Ci
p(x))), ∀i ∈ V, ∀p ∈ P, for Assumption 4. It follows from the definition of the tangent cone that
0 , D+Mk(t) ≤ 0.
p (x) ≤ 0 for all i ∈ V satisfying xik = Mk. It follows that for all k ∈ D and any x ∈ Hn
f ik
We can similarly show that for all k ∈ D, D+mk(t) ≥ 0.
0 and any initial time t0. It follows from Definition 5 and Lemma 1 that f i
p(x))), ∀i ∈ V, ∀p ∈ P, for Assumption 3 and f i
σ(t)(x(t)), where f ik
σ(t) denotes kth entry of the vector f i
p(x) ∈ ri(cid:0)T (xi,H(Ci
Therefore, it follows that mk(x(t0)) ≤ xik(t) ≤ Mk(x(t0)), ∀k ∈ D, ∀i ∈ V, ∀t ≥ t0. Then, based on the
definition of H0, we have shown that H0 is an invariant set.
(cid:3)
5.2
Interior agents
In this subsection, we study the state evolution of the agents whose states are interior points of H(C(x)).
In the following lemma, we show that the projection of the state on any coordinate axis is strictly less than
an explicit upper bound as long as it is initially strictly less than this upper bound. Figure 5 illustrates
the following Lemma 4.
The proof follows from a similar argument used in the proof Lemma 4.9 in [17] and the following lemma
holds separately for any k ∈ D.
Lemma 4 Let Assumptions 1, 2, and 3 or Assumptions 1, 2, and 4 hold. Also assume that Gσ(t) is
uniformly jointly quasi-strongly connected. Fix any k ∈ D. For any (t1, x(t1)) ∈ R × Hn
0 , any ε > 0, and
1 T ∗
any T ∗ > 0, if xik(t2) ≤ Mk(x(t1)) − ε at some t2 ≥ t1, then xik(t) ≤ Mk(x(t1)) − δ, where δ = e−L∗
for all t ∈ [t2, t2 + T ∗], and L∗1 is a positive constant related to H0.
ε
13
Meng et al. Multi-agent Systems with Compasses
Figure 5: Illustration of Lemma 4
×H(ψ) × ··· × H(ψ), where H(ψ) = [m1(ψ), M1(ψ)] × ··· × [md(ψ), Md(ψ)], and Ho
Proof. Fix (t1, x(t1)) ∈ R × Hn
Ho
k(ψ)
{z }
[m1(ψ), M1(ψ)] × ··· × [mk−1(ψ), Mk−1(ψ)] × [mk+1(ψ), Mk+1(ψ)] × ··· × [md(ψ), Md(ψ)]. The rest of the
proof will be divided in three steps.
0 and any k ∈ D. Denote ψ = x(t1) and Mik = H(ψ) × . . . H(ψ) ×
k(ψ) =
the ith entry
(Step I). Define the following nonlinear function
gψ,k(χ) : [mk(ψ), Mk(ψ)] → R, χ 7→ sup
p∈P{max
where f ik
p (xik, y) denotes the kth entry of the vector f i
y∈Mik{f ik
p (xik, y) : xik = χ}}},
i∈V { max
p(x), xik denotes the kth entry of the vector xi and
(8)
y denotes all the other components of x except xik. The nonlinear function gψ,k(χ) is used as an upper
bound of f ik
σ(t)(x) and the argument χ is used to describe the state xik. In this step, we establish some
useful properties of gψ,k(·) based on Lemmas 11 and 12 in the Appendices. We make the following claim.
Claim A: (i) gψ,k(χ) = 0 if χ = Mk(ψ); (ii) gψ,k(χ) > 0 if χ ∈ [mk(ψ), Mk(ψ)); (iii) gψ,k(χ) is Lipschitz
continuous with respect to χ on [mk(ψ), Mk(ψ)].
p(x) ∈ Tγ(xi,H(Ci
p(x)))(or ri(cid:0)T (xi,H(Ci
It follows from Definition 5, Lemma 1 and the similar analysis of Lemma 3 (by replacing t0 with t1)
p(x)))(cid:1)) ⊆ T (xi,H(Ci
that ∀t ≥ t1, f i
p(x))) ⊆ T (xi,H(C(x))) ⊆
T (xi,H(C(ψ))), ∀i ∈ V, ∀p ∈ P. Then, it follows from Definition 5 that f ik
p (x) ≤ 0 when xik = Mk(ψ).
This implies that gψ,k(χ) ≤ 0 when χ = Mk(ψ) based on the definition of gψ,k(χ). We next show that
actually gψ,k(χ) = 0 when χ = Mk(ψ). Since Gσ(t) is uniformly jointly quasi-strongly connected, there
must exist a ¯p ∈ P such that G¯p has a nonempty arc set E¯p. We can then choose ¯i ∈ V and ¯p such that
agent ¯i has at least one neighbor agent, i.e., N¯i(¯p) is not empty since E¯p is nonempty. We next choose
xj = x¯i ∈ H(C(ψ)), for all j ∈ N¯i(¯p), where x¯ik = Mk(ψ). In such a case, H(C¯i
¯p(x)) is the singleton {x¯i}
14
Meng et al. Multi-agent Systems with Compasses
and it follows from Assumption 3 (or 4) that f¯i
¯p(x) = 0. Therefore, based on the definition of gψ,k(χ), we
know that gψ,k(χ) = 0 if χ = Mk(ψ). This proves (i).
Next, for any χ ∈ [mk(ψ), Mk(ψ)), we still use the same ¯p and ¯i as those in the proof of Claim A(i). We
choose x¯iko = Mko(ψ), ∀ko ∈ {1, . . . , k − 1, k + 1, . . . , d} and xjk = Mk(ψ), ∀k ∈ D, for all j ∈ N¯i(¯p). Note
that x¯ik = χ < Mk(ψ). In such a case, H(C¯i
to (M1(ψ), M2(ψ), . . . , Md(ψ)). It then follows from Assumption 3 that f¯i
Assumption 4 that f¯i
¯p(x) > 0. This verifies that gψ,k(χ) > 0, ∀χ ∈ [mk(ψ), Mk(ψ)). This proves (ii).
¯p(x) ≥ γ(Mk(ψ)− χ) > 0 or from
¯p(x)) is a line from point (M1(ψ), . . . , Mk−1(ψ), χ, Mk+1(ψ), . . . , Md(ψ))
Finally, it follows from Lemma 12 that gik
p (xik, y) is
locally Lipschitz with respect to xik, ∀k ∈ D, ∀i ∈ V and ∀p ∈ P. Then, it follows from Theorem 1.14
of [20] that gik
p (xik) is (globally) Lipschitz continuous with respect to xik on [mk(ψ), Mk(ψ)]. From the
p (xik) : [mk(ψ), Mk(ψ)] → R, xik 7→ maxy∈Mik f ik
first property of gψ,k(χ), it follows that gψ,k(Mk(ψ)) = 0. Therefore, based on Lemma 11, it follows that
gψ,k(χ) is Lipschitz continuous with respect to χ on [mk(ψ), Mk(ψ)]. This proves (iii) and the claim holds.
(Step II). In this step, we construct and investigate the nonlinear function hH0,k(·), which is derived by
gψ,k(·) with the argument ϕ = χ − Mk(ψ) measuring the difference between xik and the upper boundary
Mk(ψ). Define
hH0,k(ϕ) : [ak − ak, 0] → R, ϕ 7→
gψ,k(ϕ + Mk(ψ));
if ϕ ∈ [mk(ψ) − Mk(ψ), 0]
gψ,k(mk(ψ));
if ϕ ∈ [ak − ak, mk(ψ) − Mk(ψ)),
(9)
where ak = mk(x(t0)) and ak = Mk(x(t0)) are constants determined by H0. Obviously, hH0,k(ϕ) is
continuous. We make the following claim.
Claim B: (i) hH0,k(ϕ) is Lipschitz continuous with respect to ϕ on [ak − ak, 0], where the Lipschitz
constant is denoted by L∗1 and L∗1 is related to the initial bounded set H0; (ii) hH0,k(ϕ) > 0 if ϕ ∈ [ak−ak, 0);
(iii) hH0,k(ϕ) = 0 if ϕ = 0.
Note that hH0,k(ϕ) = gψ,k(ϕ + Mk(ψ)) is compact on the compact set [mk(ψ) − Mk(ψ), 0]. It follows
that hH0,k(ϕ) is Lipschitz continuous with respect to ϕ on the compact set [ak − ak, 0]. This shows that
(i) holds and properties (ii) and (iii) follow directly from the definition of hH0,k(ϕ).
(Step III). In this step, we take advantage of gψ,k(χ) and hH0,k(ϕ) to show that xik will be always strictly
less than the upper bound Mk(ψ) as long as it is initially strictly less than Mk(ψ).
Suppose xik(t2) ≤ Mk(ψ) − ε at some t2 ≥ t1 and let T ∗ > 0. Based on the definition of gψ,k(χ), it
χ = gψ,k(χ) with
σ(t)(x(t)) ≤ gψ,k(xik(t)), ∀t ≥ t2. Let χ(t) be the solution of
follows that xik(t) = f ik
initial condition χ(t2) = xik(t2). Based on the Comparison Lemma (Lemma 3.4 of [16]), it follows that
xik(t) ≤ χ(t), ∀t ≥ t2.
Note that ϕ = χ−Mk(ψ) and ϕ = gψ,k(χ) = hH0,k(ϕ). It follows from the first property of hH0,k(ϕ) that
hH0,k(ϕ) − hH0,k(0) ≤ L∗1ϕ, ∀ϕ ∈ [ak − ak, 0]. This shows that hH0,k(ϕ) ≤ −L∗1ϕ based on the second
15
Meng et al. Multi-agent Systems with Compasses
Figure 6: Illustration of Lemma 6
and the third properties of hH0,k(ϕ). Thus, the solution of
∀t ≥ t2 based on the Comparison Lemma.
ϕ = hH0,k(ϕ) satisfies ϕ(t) ≤ e−L∗
1(t−t2)ϕ(t2),
Therefore, xik(t) ≤ χ(t) = ϕ(t) + Mk(ψ) ≤ e−L∗
1(t−t2)(χ(t2) − Mk(ψ)) + Mk(ψ) ≤ e−L∗
1 T ∗
(xik(t2) −
(cid:3)
Mk(ψ)) + Mk(ψ) ≤ Mk(ψ) − e−L∗
1T ∗
ε for all t ∈ [t2, t2 + T ∗].
The following lemma is symmetric to Lemma 4. The proof can be obtained using the proof of Lemma
4 under the transformation zi = −xi, i = 1, . . . , n and it is therefore omitted.
Lemma 5 Let Assumptions 1, 2, and 3 or Assumptions 1, 2, and 4 hold. Also assume that Gσ(t) is
uniformly jointly quasi-strongly connected. Fix any k ∈ D. For any (t1, x(t1)) ∈ R × Hn
0 , any ε > 0, and
2 T ∗
any T ∗ > 0, if xik(t2) ≥ mk(x(t1)) + ε at some t2 ≥ t1, then xik(t2) ≥ mk(x(t1)) + δ, where δ = e−L∗
for all t ∈ [t2, t2 + T ∗], where L∗2 is a positive constant related to H0.
ε
5.3 "Boundary" agents
In the following lemma, we show that any agent that is attracted by an "interior" agent will become an
"interior" agent after a finite time period. Figure 6 illustrates Lemma 6.
Lemma 6 Let Assumptions 1 and 2 hold and assume that Gσ(t) is uniformly jointly quasi-strongly con-
0 , any δ1 > 0 and any T ∗ > 0, assume that there is an
nected. Fix any k ∈ D. For any (t1, x(t1)) ∈ R × Hn
arc (j, i) and a time t2 ≥ t1 such that j ∈ Ni(σ(t)), and xjk(t) ≤ Mk(x(t1)) − δ1 for all t ∈ [t2, t2 + τd].
Then, there exists a t3 ∈ [t1, t2 + τd] such that xik(t) ≤ Mk(x(t1)) − δ2, for all t ∈ [t3, t3 + T ∗]. Here,
if Assumption 3 is satisfied, δ2 = e−L∗
1 related
, δ1} for some positive constants L∗1 and L+
1 T ∗
min{ γτdδ1
1 τd+1
L+
16
Meng et al. Multi-agent Systems with Compasses
to H0. If Assumption 4 is satisfied, δ2 = e−L∗
positive-definite function υ(·) both related to H0.
1T ∗
υ(δ1) for some positive constant L∗1 and a continuous
L+
1 τd+1
Proof. We first show that there exists t3 ∈ [t1, t2 + τd] such that xik(t3) ≤ Mk(x(t1)) − ε, where
, δ1} given Assumption 3 satisfied or ε = υH0(δ1) given Assumption 4 satisfied. This is
ε (C(x(t1))) and an axis-aligned hyperrectangle Hk
ε (C(x)) = {y ∈ H(C(x)) : yk ≤ Mk(x) − ε}. Obviously, B is compact convex set. Suppose
ε = min{ γτdδ1
equivalent to show that kxi(t3)kB = 0, where B := Hk
defined as Hk
kxi(t3)kB 6= 0. It then follows that 0 < kxi(t)kB ≤ ε for all t ∈ [t1, t2 + τd].
ε
Considering the time interval t ∈ [t2, t2 + τd], we define x(t) = [x11, . . . , x1d, x21, . . . , x2d, . . . ,
xn1, . . . , xnd], xik(t) = Mk(x(t1)) for given i and k, and xioko(t) = xioko(t) for io ∈ V \{i} and ko ∈ D\{k}.
The rest of the proof will be divided into three steps.
p (x) − f ik
p (x(t)) is uniformly locally Lipschitz with respect to x and compact
0 , ∀i ∈ V, ∀p ∈ P based on Assumption 2 and Lemma 3. Therefore, there exists a positive constant
1 ε, ∀p ∈ P, and
(Step I). It has been shown that f ik
on Hn
L+
1 related to H0 such that f ik
∀x, x ∈ Hn
0 .
(Step II - Assumption 3). In this step, we show that the derivative of kxi(t)kB along the solution of (2) has
a lower bound. For any p∗ ∈ P such that there is an arc (j, i) where j ∈ Ni(p∗), and xjk ≤ Mk(x(t1)) − δ1
during t ∈ [t2, t2 + τd], it follows from Assumption 3 of f i
p∗(x))) and xik(t) = Mk(x(t1))
that
1 kx(t) − x(t)k ≤ L+
p∗(x) ∈ Tγ(xi,H(Ci
p (x) ≤ f ik
p (x) ≤ L+
p (x) − f ik
p∗(x) ≥ γDk(H(co{xi, xj : j ∈ Ni(p∗)})) ≥ γDk(H(co{xi, xj})) ≥ γδ1,
f ik
(10)
where the first inequality is based on Assumption 3 by noting that xi ∈ rbkH(co{xi, xj : j ∈ Ni(p∗)}),
and rbkH(co{xi, xj : j ∈ Ni(p∗)}) is the facet of H(co{xi, xj : j ∈ Ni(p∗)}) perpendicular to −→rk . This
together with the preceding deduction f ik
0 , implies that
1 ε for any p∗ ∈ P such that there is an arc (j, i) where j ∈ Ni(p∗),
p∗(x) ≥ f ik
f ik
and xjk ≤ Mk(x(t1))− δ1 during t ∈ [t2, t2 + τd]. Note that ε = min{ γτdδ1
, δ1} is chosen sufficiently small
at the beginning of the proof such that γδ1 − L+
1 ε, ∀p ∈ P, and ∀x, x ∈ Hn
1 ε ≥ γδ1 − L+
p (x) ≤ L+
p (x) − f ik
p∗ (x) − L+
1 ε is positive.
1 τd+1
L+
Therefore, based on the assumptions of Lemma 6, it follows that for all t ∈ [t2, t2 + τd],
σ(t)(x(t)) ≥ γδ1 − L+
1 ε,
D+kxi(t)kB = hsgn(xi(t) − PB(xi(t))), f i
σ(t)(x(t))i = f ik
(11)
where the componentwise sign function sgn(·) is defined as sgn(z) = [sgn(z1), sgn(z2), . . . , sgn(zd)] for a
vector z = [z1, z2, . . . , zd] and sgn(z1) is the sign function: sgn(z1) = 1 if z1 > 0, sgn(z1) = 0 if z1 = 0, and
sgn(z1) = −1 if z1 < 0. Note that sgn(xi − PB(xi)) = −→rk whenever kxikB > 0.
(Step II - Assumption 4). Fix (t1, x(t1)) ∈ R×Hn
H(ψ)×···×H(ψ), where H(ψ) = [m1(ψ), M1(ψ)]×··· × [md(ψ), Md(ψ)], and Ho
0 . Denote ψ = x(t1) and Mik = H(ψ)×. . . H(ψ)×Ho
k(ψ)×
k(ψ) = [m1(ψ), M1(ψ)]×
17
Meng et al. Multi-agent Systems with Compasses
··· × [mk−1(ψ), Mk−1(ψ)] × [mk+1(ψ), Mk+1(ψ)] × ··· × [md(ψ), Md(ψ)]. Define
dψ,k(δ1) = inf
p∈P{min
i∈V { min
y∈Uψ{f ik
p (Mk(ψ), y)}}},
(12)
where Uψ(i, k, p, δ1) = {y ∈ Mik : ∃j ∈ Ni(p) such that xjk ≤ Mk(ψ) − δ1}. Based on the relative interior
condition of Assumption 4, we know that dψ,k(δ1) > 0 for δ1 > 0.
For any p∗ ∈ P such that there is an arc (j, i), where j ∈ Ni(p∗), and xjk ≤ Mk(x(t1)) − δ1, we know
p∗(x(t)) ≥ dψ,k(δ1). This together with the
0 , implies that for all t ∈ [t2, t2 + τd],
from the definition of dψ,k(·) that for all t ∈ [t2, t2 + τd], f ik
preceding deduction f ik
1 ε, ∀p ∈ P, and ∀x, x ∈ Hn
p (x) ≤ L+
p (x)−f ik
D+kxi(t)kB = f ik
σ(t)(x(t)) ≥ dψ,k(δ1) − L+
1 ε.
(13)
(14)
Before moving on, we define υH0,k(δ1) : [0, ak − ak] → [0,∞),
δ1 7→
1 +1 o ;
minnδ1, τddψ,k(δ1)
minnMk(ψ) − mk(ψ), τddψ,k(Mk(ψ)−mk (ψ))
τdL+
τdL+
1 +1
if δ1 ∈ [0, Mk(ψ) − mk(ψ)],
o ;
if δ1 ∈ (Mk(ψ) − mk(ψ), ak − ak],
where ak = mk(x(t0)) and ak = Mk(x(t0)) are constants determined by H0. Obviously, υH0,k(δ1) is a
continuous positive-definite function since υH0,k(δ1) = 0 for δ1 = 0, and υH0,k(δ1) > 0 for δ1 > 0. Also
note that υH0,k(δ1) ≤ δ1, for all δ1 ∈ [0, ak − ak] based on the definition of υH0,k and this fact will be used
in the proof of Proposition 1.
(Step III). In this step, we show that there exists a t3 ∈ [t1, t2 + τd] such that xik(t3) ≤ Mk(x(t1)) − ε and
conclude the proof by using Lemma 4.
Define ε = min{ γτdδ1
1 τd+1
L+
, δ1} for Assumption 3 and ε = υH0,k(δ1) ≤ τddψ,k(δ1)
τdL+
1 +1
that (γδ1 − L+
know that f ik
1 ε)τd ≥ ε for Assumption 3 and (dψ,k(δ1) − L+
σ(t)(x(t)) ≥ ε
σ(t)(x(t)) does not change sign and f ik
τd
for Assumption 4. It follows
1 ε)τd ≥ ε for Assumption 4. Since ε > 0, we
for t ∈ [t2, t2 + τd]. Moreover,
kxi(t2 + τd)kB − kxi(t2)kB =Z t2+τd
t2
D+kxi(τ )kBdτ ≥ τd
ε
τd
= ε.
(15)
This contradicts the assumption that 0 < kxi(t)kB ≤ ε for all t ∈ [t1, t2 + τd]. Thus, there exists a
t3 ∈ [t1, t2 + τd] such that xik(t3) ≤ Mk(x(t1)) − ε.
Finally, based on Lemma 4, we obtain xik(t) ≤ Mk(x(t1)) − δ2 for all t ∈ [t3, T ∗], where δ2 = e−L∗
1T ∗
ε.
(cid:3)
This completes the proof of the lemma.
The following lemma is symmetric to Lemma 6.
Lemma 7 Let Assumptions 1 and 2 hold and assume that Gσ(t) is uniformly jointly quasi-strongly con-
0 , any δ1 > 0 and any T ∗ > 0, assume that there is
nected. Fix any k ∈ D. For any (t1, x(t1)) ∈ R × Hn
an arc (j, i) and a time t2 ≥ t1 such that j ∈ Ni(σ(t)), and xjk(t) ≥ mk(x(t1)) + δ1. Then, there exists
18
Meng et al. Multi-agent Systems with Compasses
a t3 ∈ [t1, t2 + τd] such that xik(t) ≥ mk(x(t1)) + δ2, for all t ∈ [t3, t3 + T ∗]. Here, if Assumption 3 is
satisfied, δ2 = e−L∗
2 related to H0. If Assumption
L+
4 is satisfied, δ2 = e−L∗
υ(δ1) for some positive constant L∗2 and a continuous positive-definite function
υ(·) both related to H0.
, δ1} for some positive constants L∗2 and L+
2T ∗
2T ∗
min{ γτdδ1
2 τd+1
5.4 Proof of Theorem 1
The necessity proof follows a similar argument as the proof of Theorem 3.8 of [19]. It is therefore omitted.
We focus on the sufficiency and first give an outline of how the lemmas on invariant set, "interior" agents,
and "boundary" agents are used to prove Theorem 1.
The sufficiency proof is outlined as follows. We first use Lemma 3 to show that point-wise uniform
agreement is achieved on S0. We then focus on agreement attraction. A common Lyapunov function
is constructed and Lemma 3 is used to show that this Lyapunov function is nonincreasing. When the
Lyapunov function is not equal to zero initially, we know that there exists at least one agent not on
the upper boundary or not on the lower boundary at the initial time. Then, we apply Lemma 4 or 5 to
show that this "interior" agent will not become a "boundary" agent afterwards. Based on the fact that
the interaction graph is uniformly jointly quasi-strongly connected, we show that another agent will be
attracted by this "interior" agent at a certain time instant. Using Lemma 6 or 7, we know that this agent
will become an "interior" agent and will not go back to the boundary. Repeating this process, no agents
will stay on the boundary after certain time. This shows that the Lyapunov function is strictly shrinking,
which verifies the desired theorem.
Choose any η ∈ J and any ε > 0, where J = {x ∈ S n
0 : x1 = x2 = ··· = xn}. We define Aa(η) =
{x ∈ S n
0 : kx − ηk∞ ≤ a}. It is obvious from Lemma 3 that Aa(η) is an invariant set since a hypercube
is a special case of a hyperrectangle. Therefore, by setting δ = ε√n , we know that kx(t0) − ηk ≤ δ ⇒
kx(t) − ηk ≤ ε, ∀t ≥ t0. This shows that point-wise uniform agreement is achieved on S0.
Now define V (x) = ρ(H(C(x))), where ρ(H(C(x))) denotes the maximum side length of the hyperrect-
angle H(C(x)). Clearly, it follows from Lemma 3 that V (x) is nonincreasing along (2) and xi(t) ∈ H0,
∀i ∈ V, ∀t ≥ t0. We next prove the sufficiency of Theorem 1 by showing that V (x) is strictly shrinking
over suitable time intervals.
Since Gσ(t) is uniformly jointly quasi-strongly connected, there is a T > 0 such that the union graph
G([t0, t0 + T ]) is quasi-strongly connected. Define T1 = T + 2τd, where τd is the dwell time. Denote
κ1 = t0 + τd, κ2 = t0 + T1 + τd, . . . , κn2 = t0 + (n2 − 1)T1 + τd. Thus, there exists a node i0 ∈ V such
that i0 has a path to every other nodes jointly on time interval [κli, κli + T ], where i = 1, 2, . . . , n and
1 ≤ l1 ≤ l2 ≤ ··· ≤ ln ≤ n2. Denote T = n2T1. We divide the rest of the proof into three steps.
19
Meng et al. Multi-agent Systems with Compasses
(Step I). Consider the time interval [t0, t0 + T ] and k = 1. In this step, we show that an agent that does
More specifically, define ε1 = M1(x(t0))−m1(x(t0))
not belong to the interior set will become an "interior" agent due to the attraction of "interior" agent i0.
. It is trivial to show that M1(x(t)) = m1(x(t)), ∀t ≥ t0
when M1(x(t0)) = m1(x(t0)) based on Definition 5. Therefore, we assume that M1(x(t0)) 6= m1(x(t0))
without loss of generality. Split the node set into two disjoint subsets V1 = {j xj1(t0) ≤ M1(x(t0)) − ε1}
and V 1 = {jj /∈ V1}.
2
Assume that i0 ∈ V1. This implies that xi01(t0) ≤ M1(x(t0)) − ε1. It follows from Lemma 4 that
xi01(t) ≤ M1(x(t0))− δ1, ∀t ∈ [t0, t0 + T ], where δ1 = e−L∗
1T ε1. Considering the time interval [κl1, κl1 + T ],
we can show that there is an arc (i1, j1) ∈ V1 × V 1 such that i1 is a neighbor of j1 (i1 might be equal or
not to i0) because otherwise there is no arc (i1, j1) for any i1 ∈ V1 and j1 ∈ V 1 (which contradicts the fact
that i0 ∈ V1 has a path to every other nodes jointly on time interval [κl1, κl1 + T ]). Therefore, there exists
a time τ ∈ [κl1, κl1 + T ] = [t0 + (l1 − 1)T + τd, t0 + l1T − τd] such that j1 ∈ Ni(σ(τ )). Based on Assumption
1, it follows that there is time interval [τ 1, τ 1 + τd] ⊂ [t0 + (l1 − 1)T, t0 + l1T ] such that j1 ∈ Ni(σ(τ )), for
all t ∈ [τ 1, τ 1 + τd].
Also note that i1 ∈ V1 implies that xi11(t0) ≤ M1(x(t0)) − ε1. This shows that xi11(t) ≤ M1(x(t0)) −
δ1, ∀t ∈ [t0, t0 + T ] based on Lemma 4. Therefore, it follows from Lemma 6 that there exists a t2 ∈
[t0, τ 1 + τd] such that xj11(t2) ≤ M1(x(t0)) − ε2 and xj11(t) ≤ M1(x(t0)) − δ2, ∀t ∈ [t2, t2 + T ], where
, δ1o. To this end, we have shown that at least two
ε2 = minn γτdδ1
, δ1o and δ2 = e−L∗
1T minn γτdδ1
agents are not on the upper boundary at t0 + l1T .
(Step II). In this step, we show that the side length of the hyperrectangle H(C(x)) parallel to the kth axis
−→rk at t0 + T is strictly less than that at t0.
1 τd+1
1 τd+1
L+
L+
We can now redefine two disjoint subsets V2 = {j xj1(t0) ≤ M1(x(t0)) − ε2} and V 2 = {jj /∈ V2}. It
then follows that V2 has at least two nodes by noting that ε2 ≤ ε1. By repeating the above analysis, we
can show that xi1(t) ≤ M1(x(t0)) − δn, ∀i ∈ V, ∀t ∈ [tn, tn + T ] by noting that δn ≤ δn−1 ≤ ··· ≤ δ1,
where tn ∈ [t0, τ n + τd] ⊆ [t0 + (ln − 1)T1, t0 + lnT1], δn = e−nL∗
1T minn (γτd)n−1
1 τd+1)n−1 , 1o ε1.
Instead, if i0 ∈ V 1, or what is equivalent, xi01(t0) ≥ m1(x(t0)) + ε1, we can similarly show that
xi1(t) ≥ m1(x(t0)) + δn, ∀i ∈ V, ∀t ∈ [tn, tn + T ] using Lemmas 5 and 7, where tn ∈ [t0, τ n + τd] ⊆
[t0 + (ln − 1)T1, t0 + lnT1], δn = e−nL∗
Therefore, it follows that D1(H(x(t0 + T ))) ≤ D1(H(x(t0))) − βD1(H(x(t0))), and β is specified as
β = e−nL∗T minn (γτd)n−1
(Step III) In this step, we show that ρ(H(C(x))) at t0 + dT is strictly less than at t0 and thus prove the
theorem by showing that V is strictly shrinking.
2o, L∗ = max{L∗1, L∗2}, and L+ = max{L+
2T minn (γτd)n−1
2 τd+1)n−1 , 1o ε1.
2(L+τd+1)n−1 , 1
1 , L+
2 }.
(L+
(L+
We consider the time interval [t0 + T , t0 + 2T ] and k = 2. Following similar analysis as of Step I and
20
Meng et al. Multi-agent Systems with Compasses
Step II, we can show that D2(H(x(t0 + 2T ))) ≤ D2(H(x(t0))) − βD2(H(x(t0))).
By repeating the above analysis, it follows that V (x(t0 + dT )) − V (x(t0)) ≤ −β(V (x(t0))).
Then, letting N be the smallest positive integer such that t ≤ t0 + N dT , we know that
V (x(t)) ≤ (1 − β)N−1V (x(t0)) ≤
1
1 − β
t−t0
dT V (x(t0)) =
(1 − β)
1
1 − β
e−β∗(t−t0)V (x(t0)),
(16)
ln 1
1−β . Denote H(S0) as the supporting hyperrectangle of S0. Since x(t0) ∈ Hn
where β∗ = 1
0 ⊆ Hn(S0),
dT
0 . By choosing k = 1
it follows that the above inequality holds for any x(t0) ∈ Hn(S0) or any x(t0) ∈ S n
1−β
and λ = β∗, we have that exponential agreement attraction is achieved on S0. This proves the desired
theorem.
(cid:3)
5.5 Proof of Proposition 1
The necessity proof follows a similar argument as the proof of Theorem 3.8 of [19] and the proof of point-
wise uniform agreement is similar to the one of Theorem 1. We focus on the proof of agreement attraction
and use a similar analysis as in the proof of Theorem 1.
Using the same Lyapunov function V (x) = ρ(H(C(x))) as in the proof of Theorem 1, we first show that
xi01(t0) ≤ M1(x(t0))− ε1 and xi01(t) ≤ M1(x(t0))− δ1, ∀t ∈ [t0, t0 + T ], where ε1 = M1(x(t0))−m1(x(t0))
and
δ1 = e−L∗
1T ε1. Then, we have another agent i1 ∈ V1 satisfying xi11(t0) ≤ M1(x(t0)) − ε1. This shows that
xi11(t) ≤ M1(x(t0))−δ1, ∀t ∈ [t0, t0 +T ] based on Lemma 4. Therefore, it follows from Lemma 6 that there
exists t2 ∈ [t0, τ 1 + τd] such that xj11(t2) ≤ M1(x(t0)) − ε2 and xj11(t) ≤ M1(x(t0)) − δ2, ∀t ∈ [t2, t2 + T ],
where ε2 = υH0,k(δ1) and δ2 = e−L∗
1T υH0,k(δ1).
2
Then, we define two disjoint subsets V2 = {j xj1(t0) ≤ M1(x(t0)) − ε2} and V 2 = {jj /∈ V2}. It
follows that V2 has at least two nodes. Note that υH0,k(δ1) ≤ δ1 for all its definition domain. By repeating
the above analysis, we can show that xi1(t) ≤ M1(x(t0)) − δn, ∀i ∈ V, ∀t ∈ [tn, tn + T ] by noting
that δn ≤ δn−1 ≤ ··· ≤ δ1, where tn ∈ [t0, τ n + τd] ⊆ [t0 + (ln − 1)T1, t0 + lnT1], δn(D1(H(x(t0)))) =
ς ◦ υH0,k ◦ . . . υH0,k ◦ ς( ·2 ), a continuous positive-definite function ς(·) is defined as ς(x) = e−L∗
1T x, and
D1(H(x(t0))) = M1(x(t0))−m1(x(t0)). It is obvious that δn(D1(H(x(t0)))) is a continuous positive-definite
function.
Instead, if i0 ∈ V 1, or what is equivalent, xi01(t0) ≥ m1(x(t0)) + ε1, we can similarly show that
xi1(t) ≥ m1(x(t0))+δn, ∀i ∈ V, ∀t ∈ [tn, tn +T ] using Lemmas 5 and 7, where tn ∈ [t0, τ n +τd] ⊆ [t0 +(ln−
1)T1, t0 +lnT1], a continuous positive-definite function δn(D1(H(x(t0)))) = ς◦υH0◦. . . υH0◦ς( ·2 ), and ς(·) is
defined as ς(x) = e−L∗
2T x. Therefore, it follows that D1(H(x(t0 + T ))) ≤ D1(H(x(t0)))− δ∗(D1(H(x(t0)))),
where δ∗(x) = min{δn(x), δn(x)} is a continuous positive-definite function.
Then, following Lemma 4.3 of [16], there exists a class K function Υ(D1(H(x(t0)))) defined on [0, ak−ak]
satisfying Υ(D1(H(x(t0)))) ≤ δ∗(D1(H(x(t0)))), ∀D1(H(x(t0))) ∈ [0, ak− ak], where a continuous function
21
Meng et al. Multi-agent Systems with Compasses
Υ : [0, b) → [0,∞) is said to belong to class K if it is strictly increasing and Υ(0) = 0. Therefore, it follows
that D1(H(x(t0 + T ))) ≤ D1(H(x(t0))) − Υ(D1(H(x(t0)))), ∀i ∈ V, ∀t ∈ [tn, tn + T ].
We next consider the time interval [t0 + T , t0 + 2T ]. Following the previous analysis, we can show that
D2(H(x(t0 + 2T ))) ≤ D2(H(x(t0))) − Υ(D2(H(x(t0)))).
By repeating the above analysis, it follows that V (x(t0 + dT )) − V (x(t0)) ≤ −Υ(V (x(t0))).
Then, let N be the smallest positive integer such that t ≤ t0 + N dT . It then follows that
V (x(t)) − V (x(t0)) ≤ −Υ(V (x(t0))) − ··· − Υ(V (x(t0 + (N − 1)dT ))) ≤ −N Υ(V (x(t0))).
(17)
Therefore, for any ε > 0, there exists a sufficiently large N such that V (x(t)) ≤ 2ε√n , ∀t ≥ t0 + N dT . This
shows that there exists η ∈ J such that kx(t) − ηk∞ ≤ ε√n , ∀t ≥ t0 + N dT . This implies kx(t) − ηk ≤ ε,
∀t ≥ t0 + N dT , which shows that uniformly agreement attraction is achieved on S0 and proves the
proposition.
(cid:3)
Remark 2 (Extension to global convergence) The convergence is semi-global since the selections of
K class function Υ, and parameters λ and k depend on that the initial common space is given in advance
and compact, i.e., the assumption that S0 is compact is necessary to guarantee uniformly asymptotic or
exponential agreement. On the other hand, if Assumption 2 is changed to "uniformly globally Lipschitz",
we obtain a global convergence result.
6 Cooperative -- antagonistic Multi-agent Systems
In this section, we focus on
cooperative -- antagonistic multi-agent systems and prove Theorem 2 using a contradiction argument, with
the help of a series of preliminary lemmas. Note that since every agent admits a continuous trajectory, we
only need to prove that all the agents' componentwise absolute values reach an agreement.
6.1
Invariant set
In this section, we construct an invariant set for the dynamics under the cooperative -- antagonistic network.
For all k ∈ D, define M†k(x(t)) = maxi∈V xik(t). In addition, define an origin-symmetric supporting
hyperrectangle H(bC(x)) ⊂ Rd as H(bC(x)) := [−M†1 (x), M†1 (x)] × ··· × [−M†d(x), M†d (x)]. The origin-
symmetric supporting hyperrectangle formed by the initial states of all agents is given by bHn
i∈V xid(t0)(cid:3).
i∈V xi1(t0)(cid:3) × ··· ×(cid:2) − max
i∈V xid(t0), max
bH0 =(cid:2) − max
i∈V xi1(t0), max
0 , where
(18)
22
Meng et al. Multi-agent Systems with Compasses
Introduce the state transformation yik = x2
ik, for all i ∈ V, and for all k ∈ D. The analysis will be
carried out on yik, instead of xik to avoid non-smoothness. The following lemma establishes an invariant
set for system (2).
0 is an invariant set, i.e., xi(t) ∈ bH0,
Lemma 8 Let Assumptions 1, 2 and 5 hold. Then, for system (2), bHn
∀i ∈ V, ∀t ≥ t0.
Proof. Let yk = maxi∈V yik, for all k ∈ D. We first show that D+yk ≤ 0, for all k ∈ D. It follows
σ(t)(x), ∀i ∈ V, ∀k ∈ D. Let bV(t) = {i ∈ V : yik(xi(t)) = yk(x(t))} be the
from (1) that yik = 2xikf ik
set of indices where the maximum is reached at t. It then follows from Lemma 2 that for all k ∈ D,
D+yk = 2 maxi∈ bV(t) xikf ik
σ(t)(x). Consider any x(t0) ∈ Hn
0 and any initial time t0. It follows from Definition
5 and Lemma 1 that f i
∀i ∈ V, ∀p ∈ P. Based on Definition
p(x) ∈ Tγ(xi,H(C
p (x) ≥ 0 for xik = −√yk ≤ 0. This shows that
0 , D+yk ≤ 0.
ik(t0), ∀i ∈ V, ∀k ∈ D, which shows that − maxi∈V xik(t0) ≤ xik(t) ≤
p (x) ≤ 0 for i ∈ bV = {i ∈ V : yik = yk}. It follows that for all k ∈ D, and x ∈ bHn
p (x) ≤ 0 for xik = √yk ≥ 0 and f ik
p(x))) ⊆ T (xi,H(bC(x))),
ik(t) ≤ maxi∈V x2
5, it follows that f ik
i
xikf ik
Therefore, x2
maxi∈V xik(t0), ∀i ∈ V, ∀k ∈ D, ∀t ≥ t0. This implies that bHn
0 is an invariant set.
(cid:3)
6.2 "Interior" agents
In the following lemma, we show that the projection of the state on any axis is strictly less than a certain
upper bound as long as it is initially strictly less than this upper bound. The lemma relies on the technical
Lemmas 11 and 13, which can be found in the appendices.
Lemma 9 Let Assumptions 1, 2, and 5 hold and assume that Gσ(t) is uniformly jointly strongly connected.
Fix any k ∈ D. For any (t1, x(t1)) ∈ R × bHn
0 , any ε > 0 and any T ∗ > 0, if yik(t2) ≤ y∗ − ε at some
t2 ≥ t1, where y∗ ≥ yk(x(t1)) is a constant, then yik(t) ≤ y∗− δ for all t ∈ [t2, t2 + T ∗], where δ = e−L∗T ∗
ε,
and L∗ is a positive constant related to bH0.
Proof. Similar to the proof of Lemma 4, we consider any (t1, x(t1)) ∈ R × bHn
k(ψ) × bH(ψ) × ··· × bH(ψ), where bH(ψ) = [−M†1 (ψ), M†1 (ψ)] × ··· ×
and Mik = bH(ψ) × ··· × bH(ψ) × bHo
[−M†d (ψ), M†d (ψ)], and bHo
× [−M†k−1(ψ), M†k−1(ψ)] × [−M†k+1(ψ), M†k+1(ψ)] × ··· × [−M†d(ψ), M†d (ψ)]. Again, for clarity we divide
k(ψ) = [−M†1 (ψ), M†1 (ψ)] × . . .
0 , and let ψ = x(t1)
the rest of the proof into three steps.
(Step I). Define the following function
gψ,k(χ) : [−M†k (ψ), M†k (ψ)] → R, χ → sup
p∈P{max
i∈V { max
y∈Mik{xikf ik
p (xik, y) : xik = χ}}}.
(19)
23
Meng et al. Multi-agent Systems with Compasses
Obviously, gψ,k is continuous. In this step, we establish some useful properties of gψ,k based on Lemmas
11 and 13. We make the following claim.
Claim A: (i) gψ,k(χ) = 0 if χ = ±M†k (ψ); (ii) gψ,k(χ) > 0 if χ ∈ (−M†k(ψ), M†k (ψ)); (iii) gψ,k(χ) is
Lipschitz continuous with respect to χ on [−M†k (ψ), M†k (ψ)].
The first and second properties of Claim A can be obtained following a similar analysis to the proof of
p : [−M†k (ψ), M†k (ψ)] → R,
Lemma 4. For the third property of Claim A, it follows from Lemma 13 that gik
xik 7→ maxy∈Mik xikf ik
p (xik, y) is Lipschitz continuous with respect to xik, ∀k ∈ D, ∀i ∈ V, and ∀p ∈ P.
Also note that gψ,k(M†k (ψ)) = 0. Then, it follows from Lemma 11 that gψ,k(χ) is Lipschitz continuous
with respect to χ on [−M†k (ψ), M†k (ψ)].
(Step II). In this step, we construct another nonlinear function h bH0,k(·) based on the definition of gψ,k(·).
From the definitions of gψ,k(χ), it follows that yik(t) = 2xikf ik
σ(t)(x(t)) ≤ 2gψ,k(xik(t)), ∀t ≥ t2. It also
follows from the properties of gψ,k(χ) that there exists a Lipschitz constant L1 such that gψ,k(χ) =
gψ,k(χ) − gψ,k(M†k (ψ)) ≤ L1χ − M†k(ψ) = L1(M†k (ψ) − χ), ∀χ ∈ [−M†k(ψ), M†k (ψ)] and gψ,k(χ) =
gψ,k(χ)− gψ,k(−M†k (ψ)) ≤ L1χ + M†k(ψ) = L1(M†k (ψ) + χ), ∀χ ∈ [−M†k (ψ), M†k (ψ)], where L1 is related
Therefore, for the case of xik ≥ 0, we have that yik(t) ≤ 2L1(M†k (ψ)−xik) = 2L1(M†k (ψ)−pyik(t)), For
the case of xik < 0, we have that yik(t) ≤ 2L1(M†k (ψ) + xik) = 2L1(M†k (ψ) −pyik(t)). Overall, it follows
that yik(t) ≤ 2L1(M†k (ψ)−pyik(t)), ∀t ≥ t2. Let φ(t) be the solution of φ = hψ,k(φ) with initial condition
φ(t2) = yik(t2), where hψ,k : [0, yk(ψ)] → R, φ 7→ 2L1(M†k (ψ) − √φ). It follows from the Comparison
Lemma that yik(t) ≤ φ(t), ∀t ≥ t2.
to ψ.
Next, by letting ϕ = yk(ψ) − φ and ak = M†k (x(t0)), we define the following function
h bH0,k(ϕ) : [0, (ak)2] → R, ϕ 7→
hψ,k(yk(ψ) − ϕ);
hψ,k(0);
if ϕ ∈ (yk(ψ), (ak)2].
if ϕ ∈ [0, yk(ψ)],
(20)
We have the following claim by easily checking the definition of h bH0,k(ϕ).
Claim B: (i) h bH0,k(ϕ) is Lipschitz continuous with respect to ϕ on [0, (ak)2]; (ii) h bH0,k(ϕ) = 0 if ϕ = 0;
(iii) h bH0,k(ϕ) > 0 if ϕ ∈ (0, (ak)2].
(Step III). In this step, we take advantage of the function h bH0,k(·) to show that yik will be always strictly
less than the upper bound y∗ as long as it is initially strictly less than y∗.
Consider any T ∗ > 0 and t ∈ [t2, t2 + T ∗]. It follows from the first property of h bH0,k(ϕ) that there exists
a constant L∗ related to bH0 such that h bH0,k(ϕ) − h bH0,k(0) ≤ L∗ϕ, ∀ϕ ∈ [0, (ak)2]. From the second and
third properties of h bH0,k(ϕ), it follows that h bH0,k(ϕ) ≤ L∗ϕ, ∀ϕ ∈ [0, (ak)2] and ϕ = − φ = −hψ,k(φ) =
−h bH0,k(ϕ), ∀ϕ ∈ [0, yk(ψ)]. It follows from the Comparison Lemma that the solution of
ϕ = −h bH0,k(ϕ)
24
Meng et al. Multi-agent Systems with Compasses
satisfies ϕ(t) ≥ e−L∗(t−t2)ϕ(t2), ∀t ≥ t2 since −h bH0,k(ϕ) ≥ −L∗ϕ.
Therefore, yik(t) ≤ φ(t) = yk(ψ) − ϕ(t) ≤ yk(ψ) − e−L∗(t−t2)(yk(ψ) − φ(t2)) ≤ yk(ψ) − e−L∗T ∗
(yk(ψ) −
(y∗ − yik(t2)) + (yk(ψ) −
ε for all t ∈ [t2, t2 + T ∗] since y∗ ≥ yk(ψ). This proves the lemma by letting
(y∗ + yk(ψ) − y∗ − yik(t2)) = y∗ − e−L∗T ∗
yik(t2)) = y∗ + yk(ψ) − y∗ − e−L∗T ∗
y∗)(1 − e−L∗T ∗
δ = e−L∗T ∗
ε.
) ≤ y∗ − e−L∗T ∗
(cid:3)
6.3 "Boundary" agents
In the following lemma, we show that any agent that is attracted by an agent strictly inside the upper
bound is drawn strictly inside that bound. This lemma relies on Lemma 14, which can be found in the
Appendices.
γτdδ
Lemma 10 Let Assumptions 1, 2, and 5 hold and assume that Gσ(t) is uniformly jointly strongly con-
nected. Fix any k ∈ D. For any (t1, x(t1)) ∈ R × bHn
0 and any δ > 0, assume that there is an arc (j, i) and
a time t2 ≥ t1 such that j ∈ Ni(σ(t)), and yjk(t) ≤ y∗ − δ for all t ∈ [t2, t2 + τd], where y∗ ≥ yk(x(t1)) is
a constant. Then, there exists a t3 ∈ [t1, t2 + τd] such that yik(t3) ≤ y∗ − ε, where ε =
2(L+τd+γτd+1) and
L+ is a constant related to bH0.
Proof. We prove this lemma by contradiction. Suppose that there does not exist a t3 ∈ [t1, t2 + τd]
such that xik(t3) ≤ √y∗ − ε1, where ε1 =
2(L+τd+γτd+1)√y∗ is a positive constant. Then it follows that
√y∗ − ε1 < xik(t) ≤ M†k (x(t1)) for all t ∈ [t1, t2 + τd].
We focus on the time interval t ∈ [t2, t2 + τd]. Define x(t) by replacing xik(t) in x(t) with xik(t) =
p (x(t)) is uniformly locally Lipschitz with respect to x and
0 , for all i ∈ V, and all p ∈ P. By noting that M†k(x(t1))− ε1 ≤ √y∗− ε1 < xik(t), it follows
p (x) ≤
p (x) ≤ f ik
p (x) − f ik
p (x) − f ik
maxi∈V{xik(t)} = M†k(x(t)). We know that f ik
compact on bHn
that there exists a positive constant L+ related to bH0 such that f ik
L+kx(t) − x(t)k ≤ L+ε1, ∀p ∈ P, and ∀x, x ∈ bHn
It follows from yjk(t) ≤ y∗ − δ that √y∗ − xjk(t) ≥ δ
an arc (j, i) with j ∈ Ni(p∗) and √y∗ − xjk(t) ≥ δ
p∗(x)(cid:12)(cid:12)(cid:12) ≥ γDk(H(co{xi, xjsgnij
(cid:12)(cid:12)(cid:12)f ik
L+ε1 > γ(cid:16) δ
2√y∗ − ε1(cid:17) − L+ε1. Choose ε1 sufficiently small, especially, ε1 =
γ(cid:16) δ
every y∗ > 0. It follows that γ(cid:16) δ
p∗ : j ∈ Ni(p∗)})) ≥ γDk(H(co{xi, xjsgnij
2√y∗ − ε1(cid:19) .
p∗(x(t)) ≥ f ik
2√y∗ > ε1 based on the definition of ε1. Therefore, we know that f ik
2√y∗ − ε1(cid:17) − L+ε1. It then follows that for all t ∈ [t2, t2 + τd], D+xik(t) = f ik
p∗})) > γ(cid:18) δ
2√y∗ , it follows from Assumption 5 that
p∗(x(t)) −
σ(t)(x(t)) >
2(L+τd+γτd+1)√y∗ . Such ε1 exists for
2√y∗ . Therefore, for any p∗ ∈ P such that there is
. Therefore, we know that
Note that
(21)
γτdδ
0 .
δ
2√y∗ − ε1(cid:17) − L+ε1 > ε1
xik(t2 + τd) − xik(t2) =Z t2+τd
τd
t2
γτdδ
(cid:12)(cid:12)D+xik(τ )(cid:12)(cid:12) dτ > τd
ε1
τd
= ε1.
25
(22)
Meng et al. Multi-agent Systems with Compasses
This contradicts the assumption that √y∗− ε1 < xik(t) ≤ M†k(x(t1)) for all t ∈ [t2, t2 + τd]. It then follows
that yik(t3) ≤ y∗ − ε, where ε = √y∗ε1 =
2(L+τd+γτd+1) .
γτdδ
(cid:3)
6.4 Proof of Theorem 2
Unlike the contraction analysis of a common Lyapunov function given in the proof of Theorem 1, we use
a contradiction argument for the proof of Theorem 2.
According to the proof of Lemma 8, we know that D+yk ≤ 0 and yk ≥ 0, for all k ∈ D. Therefore, yk(t),
k ∈ D is monotonically decreasing and bounded from below by zero. This implies that for any initial time
t0 and initial state x(t0), there exists a constant y∗k, k ∈ D such that limt→∞ yk(t) = y∗k, ∀k ∈ D. Define
ik = lim supt→∞ yik(t) and ℓik = lim inf t→∞ yik(t), for all i ∈ V, and k ∈ D. Clearly, 0 ≤ ℓik ≤ ik ≤ y∗k.
We know that the componentwise absolute values of all the agents converges to the same if and only if
ik = ℓik = y∗k, ∀i ∈ V, ∀k ∈ D. The desired conclusion holds trivially if y∗k = 0, k ∈ D. Therefore, we
assume that y∗k > 0 for some k ∈ D without loss of generality.
Suppose that there exists a node i1 ∈ V such that 0 ≤ ℓi1k < i1k ≤ y∗k. Based on the fact that
limt→∞ yk(t) = y∗k, it follows that for any ε > 0, there exists a bt(ε) > t0 such that y∗k − ε ≤ yk(t) ≤ y∗k +
t ≥bt(ε). Take α1k =q 1
2 (ℓi1k + i1k). Therefore, there exists a time t1 ≥bt(ε) such that xi1k(t1) = α1k
based on the definitions of i1k and ℓi1k and continuousness of xi1k(t). This shows that
ε,
i1k(t1) = i1k − (i1k − α2
x2
1k) ≤ y∗k + ε − (i1k − α2
1k) = y∗k + ε − ε1,
(23)
where ε1 = i1k − α2
1k > 0 and the first inequality is based on the definition of i1k.
Since Gσ(t) is uniformly jointly strongly connected, there is a T > 0 such that the union graph G([t1, t1 +
T ]) is jointly strongly connected. Define T1 = T + 2τd, where τd is the dwell time. Denote κ1 = t1 + τd,
κ2 = t1 + T1 + τd, . . . , κn = t1 + (n − 1)T1 + τd. For each node i ∈ V, i has a path to every other nodes
jointly on time interval [κl, κl + T ], where l = 1, 2, . . . , n. Denote T = nT1.
Consider time interval [t1, t1 + T ]. Based on the fact that yk(x(t1)) ≤ y∗k + ε and considering y∗k + ε as
the role of y∗ in Lemma 9, it follows that yik(t) ≤ y∗k + ε − δ1, ∀t ∈ [t1, t1 + T ], where δ1 = e−L∗T ε1.
Since for each node i ∈ V, i has a path to every other nodes jointly on time interval [κl, κl + T ], where
l = 1, 2, . . . , n, there exists i2 ∈ V such that i1 is a neighbor of i2 during the time interval [κ1, κ1 +T ]. Based
i2k(t2) ≤ y∗k +ε−ε2,
on Lemma 10, it follows that there exists t2 ∈ [t1, τ 1 +τd] ⊂ [t1 +T, t1 +2T ] such that x2
i2k(t) ≤ y∗k + ε − δ2, ∀t ∈ [t2, t1 + T ], where
where ε2 =
δ2 = e−L∗T ε2. By repeating the above analysis, we can show that yik(t) ≤ y∗k + ε − δn, ∀t ∈ [tn, t1 + T ],
∀i ∈ V, where tn ∈ [t1, τ n + τd] ⊂ [t1 + (n − 1)T, t1 + nT ], and δn can be iteratively obtained as δn =
e−nL∗T
2(L+τd+γτd+1) δ1. This further implies that x2
2n−1(L+τd+γτd+1)n−1 . This is indeed true because δi ≤ δi−1, ∀i = 2, 3, . . . , n.
γn−1τ n−1
γτd
d
26
Meng et al. Multi-agent Systems with Compasses
This shows that yk(t1 + T ) = maxi∈V yik ≤ y∗k + ε − δn, which indicates a contradiction for sufficiently
small ε satisfying ε < δn/2. Therefore, ik = ℓik = y∗k, ∀i ∈ V, ∀k ∈ D. This proves that limt→∞(xik(t) −
py∗k) = 0, ∀i ∈ V and ∀k ∈ D, which shows the componentwise absolute values of all the agents converges
to the same values and the theorem holds.
(cid:3)
7 Conclusions
Agreement protocols for nonlinear multi-agent dynamics over cooperative or cooperative -- antagonistic
networks were investigated. A class of nonlinear control laws were introduced based on relaxed convexity
conditions. The price to pay was that each agent must have access to the orientations of a shared coordinate
system, similar to a magnetic compass. Each agent specified a local supporting hyperrectangle with the
help of the shared reference directions and the relative state measurements, and then a strict tangent cone
was determined. Under mild conditions on the nonlinear dynamics and the interaction graph, we proved
that for cooperative networks, exponential state agreement is achieved if and only if the interaction graph
is uniformly jointly quasi-strongly connected. For cooperative -- antagonistic networks, the componentwise
absolute values of all the agents converge to the same values if the time-varying interaction graph is
uniformly jointly strongly connected. The results generalize existing studies on agreement seeking of multi-
agent systems. Future works include higher-order agent dynamics, convergence conditions for bipartite
agreement, and the study on the case of mismatched shared reference directions.
References
[1] C. Altafini. Dynamics of opinion forming in structurally balanced social networks. PloS One,
7(6):e38135, 2012.
[2] C. Altafini. Consensus problems on networks with antagonistic interactions. IEEE Transactions on
Automatic Control, 58(4):935 -- 946, 2013.
[3] J.-P. Aubin. Viability Theory. Birkhauser Boton, Boston, 1991.
[4] D. Bauso, L. Giarre, and R. Pesenti. Non-linear protocols for optimal distributed consensus in
networks of dynamic agents. Systems & Control Letters, 55(11):918 -- 928, 2006.
[5] V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis. Convergence in multiagent
coordination, consensus, and flocking.
In 44th IEEE Conference on Decision and Control, pages
2996 -- 3000, 2005.
27
Meng et al. Multi-agent Systems with Compasses
[6] M. Cao, A. S. Morse, and B. D. O. Anderson. Agreeing asynchronously.
IEEE Transactions on
Automatic Control, 53(8):1826 -- 1838, 2008.
[7] M. Cao, A. S. Morse, and B. D. O. Anderson. Reaching a consensus in a dynamically changing
environment: convergence rates, measurement delays, and asynchronous events. SIAM Journal on
Control and Optimization, 47(2):601 -- 623, 2008.
[8] J. Cort´es, S. Mart´ınez, and F. Bullo. Robust rendezvous for mobile autonomous agents via proximity
graphs in arbitrary dimensions. IEEE Transactions on Automatic Control, 51(8):1289 -- 1298, 2006.
[9] J. M. Danskin. The theory of max-min, with applications. SIAM Journal on Applied Mathematics,
14(4):641 -- 664, 1966.
[10] D. Easley and J. Kleinberg. Networks, Crowds, and Markets. Reasoning About a Highly Connected
World. Cambridge Univ. Press, Cambridge, 2010.
[11] A. F. Filippov. Differential Equations with Discontinuous Righthand Sides. Norwell, MA: Kluwer,
1988.
[12] C. Godsil and G. Royle. Algebraic Graph Theory. New York: Springer-Verlag, 2001.
[13] J. Haverinen and A. Kemppainen. Global indoor self-localization based on the ambient magnetic
field. Robotics and Autonomous Systems, 57(10):1028 -- 1035, 2009.
[14] J. M. Hendrickx and J. N. Tsitsiklis. Convergence of type-symmetric and cut-balanced consensus
seeking systems. IEEE Transactions on Automatic Control, 58(1):214 -- 218, 2013.
[15] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of mobile autonomous agents using
nearest neighbor rules. IEEE Transactions on Automatic Control, 48(6):988 -- 1001, 2003.
[16] H. K. Khalil. Nonlinear Systems, Third Edition. Englewood Cliffs, New Jersey: Prentice-Hall, 2002.
[17] Z. Lin. Coupled dynamic systems: From structure towards stability and stabilizability. Ph. D. disser-
tation, University of Toronto, Toronto, 2006.
[18] Z. Lin, M. Broucke, and B. Francis. Local control strategies for groups of mobile autonomous agents.
IEEE Transactions on Automatic Control, 49(4):622 -- 629, 2004.
[19] Z. Lin, B. Francis, and M. Maggiore. State agreement for continuous-time coupled nonlinear systems.
SIAM Journal on Control and Optimization, 46(1):288 -- 307, 2007.
[20] N. G. Markley. Principles of Differential Equations. Wiley-Interscience, first edition, 2004.
28
Meng et al. Multi-agent Systems with Compasses
[21] I. Matei and J. S. Baras. The asymptotic consensus problem on convex metric spaces. IEEE Trans-
actions on Automatic Control. to appear.
[22] I. Matei, J. S. Baras, and C. Somarakis. Convergence results for the linear consensus problem under
markovian random graphs. SIAM Journal on Control and Optimization, 51(2):1574 -- 1591, 2013.
[23] Z. Meng, G. Shi, and K. H. Johansson. Multi-agent systems with compasses: cooperative and
cooperative-antagonistic networks. In Proceedings of 33rd Chinese Control Conference, pages 1430 --
1437, Nanjing, China, 2014.
[24] L. Moreau. Stability of continuous-time distributed consensus algorithms. In IEEE Conference on
Decision and Control, pages 3998 -- 4003, Atlantis, Bahamas, 2004.
[25] L. Moreau. Stability of multi-agent systems with time-dependent communication links. IEEE Trans-
actions on Automatic Control, 50(2):169 -- 182, 2005.
[26] R. Olfati-Saber, J. A. Fax, and R. M. Murray. Consensus and cooperation in networked multi-agent
systems. Proceedings of the IEEE, 95(1):215 -- 233, 2007.
[27] 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.
[28] A. Olshevsky and J. N. Tsitsiklis. Convergence speed in distributed consensus and averaging. SIAM
Journal on Control and Optimization, 48(1):33 -- 55, 2009.
[29] W. Ren and R. W. Beard. Consensus seeking in multiagent systems under dynamically changing
interaction topologies. IEEE Transactions on Automatic Control, 50(5):655 -- 661, 2005.
[30] G. Shi and Y. Hong. Global target aggregation and state agreement of nonlinear multi-agent systems
with switching topologies. Automatica, 45(5):1165 -- 1175, 2009.
[31] G. Shi and K. H. Johansson. Robust consensus for continuous-time multi-agent dynamics. SIAM
Journal on Control and Optimization, 51(5):3673 -- 3691, 2013.
[32] V. Y. Skvortzov, H. K. Lee, S. W. Bang, and Y. B. Lee. Application of electronic compass for mobile
robot in an indoor environment. In 2007 IEEE International Conference on Robotics and Automation,
pages 2963 -- 2970, Roma, Italy, 2007.
[33] A. M. Stoneham, E. M. Gauger, K. Porfyrakis, S. C. Benjamin, and B. W. Lovett. A new type of
radical-pair-based model for magnetoreception. Biophysical Journal, 102(5):961 -- 968, 2012.
29
Meng et al. Multi-agent Systems with Compasses
[34] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. Flocking in fixed and switching networks. IEEE
Transactions on Automatic Control, 52(5):863 -- 868, 2007.
[35] J. N. Tsitsiklis, D. Bertsekas, and M. Athans. Distributed asynchronous deterministic and stochastic
gradient optimization algorithms. IEEE Transactions on Automatic Control, 31(9):803 -- 812, 1986.
[36] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen, and O. Schochet. Novel type of phase transitions in a
system of self-driven particles. Physical Review Letters, 75(6):1226 -- 1229, 1995.
[37] S. Wasserman and K. Faust. Social Network Analysis: methods and applications. Cambridge Univ.
Press, Cambridge, 1994.
Appendices
Note that a function h(·) is called Lipschitz continuous on a set U if there exists a constant LU such
that kh(a) − h(b)k ≤ LUka − bk for all a, b ∈ U .
Lemma 11 Suppose Assumption 2 holds, i.e., fp, p ∈ P, is uniformly locally Lipschitz. Assume that there
exists a point z0 ∈ Rdn such that supp∈P fp(z0) (or inf p∈P fp(z0)) is finite. Then g(x) := supp∈P fp(x) (or
inf p∈P fp(x)) is well defined and is Lipschitz continuous on every compact set U .
Proof. Let U be a given compact set. Define Uz0 = co({z0}∪U ). Based on Theorem 1.14 of [20], a locally
Lipschitz function is Lipschitz continuous on every compact subset. Plugging in the fact that fp, p ∈ P is
> 0 such that kfp(a) − fp(b)k ≤ LUz0ka − bk for all a, b ∈ Uz0
uniformly locally Lipschitz, there is LUz0
and p ∈ P. It becomes straightforward that g(x) is finite at every point in Uz0 and LUz0
is a Lipschitz
constant of g on Uz0. Therefore, the lemma holds.
(cid:3)
The following lemma is originally from [17] and restated here.
Lemma 12 Suppose that f (x1, y) : R×M → R is locally Lipschitz with respect to [x1, y]T, where M ⊂ Rq
is compact. Then g(x1) = maxy∈M f (x1, y) is locally Lipschitz.
Lemma 13 Suppose that f (x1, y) : M1 × M → R is locally Lipschitz with respect to [x1, y]T, where
M1 ⊂ R and M ⊂ Rq are compact. Then g : M1 → R, x1 7→ maxy∈M x1f (x1, y) is Lipschitz continuous
with respect to x1 on M1.
Proof. Because f (x1, y) is locally Lipschitz with respect to [x1, y]T and M1 and M are compact, it
follows that f (x1, y) is Lipschitz continuous with respect to [x1, y]T. Therefore, there exists a constant L
such that kf (x1, y) − f (x1, y)k ≤ Lkx1 − x1k, ∀x1, x1 ∈ M1, ∀y ∈ M. Also, since f (x1, y) is continuous
and the continuous function on the compact set is compact, there exist constants Lx and Lf such that
Lx = maxx1∈M1 kx1k and Lf = maxx1∈M1,y∈M kf (x1, y)k.
30
Meng et al. Multi-agent Systems with Compasses
Let yx and yx be the points satisfying g(x1) = maxy∈M{x1f (x1, y)} = x1f (x1, yx) and g(x1) =
maxy∈M{x1f (x1, y)} = x1f (x1, yx). It is trivial to show that x1f (x1, yx) ≥ x1f (x1, yx) and x1f (x1, yx) ≥
x1f (x1, yx). Therefore, there exists ex = (1− λ)x1 + λx1, where 0 ≤ λ ≤ 1 such that exf (ex, yx) = exf (ex, yx).
Thus, kg(x1)−g(x1)k = kx1f (x1, yx)−exf (ex, yx)+exf (ex, yx)−x1f (x1, yx)k ≤ kx1f (x1, yx)−x1f (ex, yx)k+
kx1f (ex, yx) −exf (ex, yx)k + kexf (ex, yx) −exf (x1, yx)k + kexf (x1, yx) − x1f (x1, yx)k. It then follows that
kg(x1) − g(x1)k ≤ Lkx1kkx1 −exk + kf (ex, yx)kkx1 −exk + Lkexkkx1 −exk + kf (x1, yx)kkx1 −exk
≤ (LLx + Lf )kx1 −exk + (LLx + Lf )kx1 −exk
= (LLx + Lf )kx1 − x1k.
(24)
(cid:3)
Therefore, g(x1) is Lipschitz continuous with respect to x1 on M1.
Lemma 14 Suppose that f (x) : M → R is locally Lipschitz with respect to x, where M ⊂ Rq is compact.
Then g(x) : M → R, x 7→ x1f (x) is Lipschitz continuous with respect to x on M, where x1 denotes an
element of x.
Proof. Because f (x) is locally Lipschitz with respect to x and M is compact, it follows that f (x) is
Lipschitz continuous with respect to x. Therefore, there exists a constant L such that kf (x) − f (x)k ≤
Lkx − xk, ∀x, x ∈ M. Also, since f (x) is continuous and the continuous function on the compact set is
still compact, there exist constants Lx and Lf such that Lx = maxx∈M kxk and Lf = maxx∈M kf (x)k.
It then follows that ∀x, x ∈ M, kg(x)−g(x)k = kx1f (x)−x1f (x)k ≤ LLxkx−xk+Lfkx−xk = (LLx+
Lf )kx − xk. Therefore, g(x) is Lipschitz continuous with respect to x on M.
(cid:3)
31
|
1503.00237 | 1 | 1503 | 2015-03-01T08:33:28 | Task Allocation in Robotic Swarms: Explicit Communication Based Approaches | [
"cs.MA",
"cs.AI",
"cs.RO"
] | In this paper we study multi robot cooperative task allocation issue in a situation where a swarm of robots is deployed in a confined unknown environment where the number of colored spots which represent tasks and the ratios of them are unknown. The robots should cover this spots as far as possible to do cleaning and sampling actions desirably. It means that they should discover the spots cooperatively and spread proportional to the spots area and avoid from remaining idle. We proposed 4 self-organized distributed methods which are called hybrid methods for coping with this scenario. In two different experiments the performance of the methods is analyzed. We compared them with each other and investigated their scalability and robustness in term of single point of failure. | cs.MA | cs | Task Allocation in Robotic Swarms: Explicit
Communication Based Approaches
Aryo Jamshidpey1 and Mohsen Afsharchi2
1 Institute for Advanced Studies in Basic Sciences
[email protected]
2 Department of Computer Science, University of Zanjan
[email protected]
Abstract. In this paper we study multi robot cooperative task alloca-
tion issue in a situation where a swarm of robots is deployed in a confined
unknown environment where the number of colored spots which repre-
sent tasks and the ratios of them are unknown. The robots should cover
this spots as far as possible to do cleaning and sampling actions desir-
ably. It means that they should discover the spots cooperatively and
spread proportional to the spots area and avoid from remaining idle. We
proposed 4 self-organized distributed methods which are called hybrid
methods for coping with this scenario. In two different experiments the
performance of the methods is analyzed. We compared them with each
other and investigated their scalability and robustness in term of single
point of failure.
1 Introduction
Swarm robotics is a branch of robotic science which is inspired by social in-
sects and other nature colonies that show complex behavior although they have
simple members. Defining the roles of robots in society is one of the most im-
portant problems for simulating such behaviors. The action of assigning tasks
to agents for performing is called task allocation. Tasks are dependent to global
mission. If robots perform their tasks effectively this mission is performed ef-
fectively too. From a control architectural perspective Burger [3] distinguishes
between Heteronomous, Autonomous and Hybrid methods in task allocation. In
this paper, we introduce a practical scenario for the issue of task allocation in
swarm robotics and 4 hybrid methods for solving it in unknown environments
which, the number, locations and ratios of tasks are unknown to robots.
Market-based mechanism is one of the main approaches that tackle the task
allocation problem. TraderBots is one of the works in this subject that is pre-
sented by Dias [6]. In TraderBots the robots bid on tasks on the basis of cost and
it is given to a bidder with lowest cost. A comprehensive study of market-based
multi-robot coordination can be found in [7]. Most solutions in self-organized task
allocation is threshold-based that are inspired by models initially proposed to de-
scribe the behavior of insect societies [1]. In this case we can mention Krieger and
5
1
0
2
r
a
M
1
]
A
M
.
s
c
[
1
v
7
3
2
0
0
.
3
0
5
1
:
v
i
X
r
a
2
Aryo Jamshidpey and Mohsen Afsharchi
Billeter work [9] which benefits from a simple threshold-based model for task allo-
cation in a foraging scenario.Labella et al. [10] and Lui et al. [11,12] proposed two
probabilistic task allocation approaches which use adaptive thresholds.Brutschy
et al. in[2] presented a task allocation strategy in which robots specialize to per-
form tasks in the environment in a self-organized. Jones and Mataric [8] intro-
duced an adaptive distributed autonomous task allocation method for identical
robots. In this work a task allocation is called desirable in which the ratio of
robots that doing the same kind of task equal with the ratio of that task in the
environment. Dahl et al. [4] proposed a method that controls the group dynamics
of task allocation. In this work each robot can choose between two separated for-
aging cycles. Nouyan et al [13] presented autonomous rule-based task allocation
method in which robots attempting to identify and transfer a food to the nest
by using a set of rules and forming a chain in a self-organized way. Dasgupta [5]
presented a communication-based method for task allocation. Each task needs
multiple robots to be done. Robots can only partially complete tasks and one
after the other contribute to progressing them. Our proposed practical scenario
like Dasgupta's one is about unknown environments. The discussion part is ded-
icated for comparing these scenarios and their solutions and also stating their
differences.
2 Problem Definition
This scenario involves a colony of identical robots with limited energy levels that
are rechargeable and an environment full of obstacles and colored spots which
represent types of tasks.The individuals are unaware of the size of the population
and the distribution of other robots.At any moment, each robot is able to do
only one of the forage for green or forage for black subtasks.Depending on the
area of the spots, the number of robots to do cleaning and sampling actions in
them varies. Obviously it is not necessary to fill the whole capacity of spots with
robots. Even a robot is able to do cleaning and sampling actions on its own but
it takes much time and less desirable.In abstract the scenario can be defined
in terms of finding more colored spots with minimal energy waste. One of the
causes of this waste is unnecessary robot turns in the environment.Also there
are some other reasons for energy waste such as maximum spreading of robots
in different spots proportional to their area, avoiding robots from collision with
obstacles and finally preventing active robots from remaining idle.
3 Methods
In all proposed methods the robots are initially in random places. In each method
the energy level of each robot is divided into three parts. First part is for foraging.
Second part is to do cleaning and sampling actions and the third part is dedicated
for returning to the charging station, either nest or any other specified location.
All methods are organized based on transferring messages and all messages have
the same structure. Each message may have several rows that each of them
Task Allocation in Robotic Swarms
3
presents a distinct spots information. Information of each spot consists of 10
features which are as follows: A row is occupied or not, that is shown by 1 or
0 similarly. X, Y and Z coordinate of the current position of the robot in the
desired environment, which is calculated via GPS. The total capacity of the
existing spots (Number of robots required to do cleaning and sampling actions
in spots, which vary depending on the area of the discovered spot). The current
number of robots deployed in the desired spot. The number of robots needed
to complete the spot capacity. The number of hops (In order to estimate the
number of robots that have been informed of the discovered spot and to identify
the spots which expose to saturation and starvation). Time elapsed from the
beginning of the running of the code and finally color of the discovered spot
which are conventionally denoted by numbers (for example in our simple scenario
black and green are denoted by 0 and 1 respectively).
Message updating procedure is the same in all of proposed methods. Every
robot has a message that contains information about the spots, called private
message. This message is empty at first and then updated via occurrence of two
different events. First when a robot finds a non-observed spot for the first time,
and second when it receives a message from one of its neighbors:
1.when a robot explores a non-seen spot for the first time, it starts to calculate
the center of spot and also its area, computes its capacity based on the area and
calculate other features of message based on a pre-defined rule. It should be
considered that if a robot entered a spot under the guidance of another robot,
it would subtract the number of required robot field by one, add one unit to the
number of existing robots field and then set the current time as the row time
field.
2. In this situation two different tasks may exist for the receiver to do, one
is to add non existing rows to its message. The order of the updated message
is the same as sender's one and in every new row, hop count will be increased
by one. The other task is to replace existing rows according to the time field. If
the row associated with a specific spot that exists in both private and received
messages has greater time field in the received message, then it will be replaced
in the private message.
3.1 Static communication based method
This method is a hybrid method combining the autonomous and distributed
methods [3]. At the beginning, one of the two subtasks of exploring for green
or black spot is assigned to each robot which is called worker statically (With a
probability of 0.5). Static means that the assigned task will not change unless
the robot runs out of its searching energy.Each robot moves in the environment
by random walk and prevents collision with other robots, obstacles and walls
by the help of its distance computing sensors. Each robot updates its private
message periodically and floods it to every other robot in the coverage area of
its radio frequency transmitter. By this process, messages will propagate among
the robots (either deployed or searcher robots). If a robot could not find a match
(a spot with the same color of its associated subtask color) before running out
4
Aryo Jamshidpey and Mohsen Afsharchi
its searching energy, it would go to one of the spots in its private message. By
having this policy, our goals for preventing idle robots existence, maximum spot
coverage and also preventing spot starvation will be guaranteed in a desirable
level.
The use of desirable word is due to the energy limitation and also the heuristic
sense of the scenario. Therefore we proposed a method for trap prevention by
partitioning robot energy and then searching the environment for the best spot
while some amount of energy exists and also the robot reaches the best possible
spot (the first seen spot by a robot is not necessarily the best spot for it, so it
will search for a suitable spot until it runs out of searching energy). At the end
the robot reaches the decision step that is as follow:
Appropriate spot selection decision making. One of the advantages of the
proposed method is the way it decides to walk to the most appropriate spot.
As the searching energy of one robot reaches its threshold, the robot starts the
decision making step. At this time the robot should sort its private message
based on three fields; color, hop count and its current distance from the spots.
At first this sort is done based on color field, so that spots with the same color
as the robot current state will be placed at the top. Now the sorted list has two
parts, one for the same colors as the subtask color and another for opposite colors
to it. Then these lists are sorted based on the hop field. Each sorted list then will
be categorized according to sets of 5 hops (number of hops per category should
be set based on the spots area and count in the environment). The reason for
using this parameter is to prevent spot starvation and saturation. When a robot
finds out that the hop count of a spot is large, it estimates that at most one less
than the hop count robots are aware and can reach the spot before it. Hence the
robot considers the spot may get saturated. Robots prefer to go to spots with
low hop number to prevent starvation. In the third step the sorted list based on
color and hop count will be further sorted based on Euclidean distance. So in
each hop category, spots are sorted based on their distance from the robot.
If a robot in its decision step observes a spot with the same color of its
assigned color while transferring to its appropriate spot, it will explore the spots
features. If the spot is suitable for the robot and if saturation will not occur,
current spot will be accepted by the robot and necessary updates will be done. It
is possible that before a robot reaches to its highest priority spot, the capacity of
that becomes full. In this case the robot will be aware of this occurrence by the
messages received from other deployed robots in that spot. It will leave the spot
and will begin to make another decision from its private list. We can consider
an integer threshold in such cases. Robots will try to make decision until the
threshold. If a robot cannot find a suitable spot in its decision step before the
threshold, it will return to home or the nearest charge station to be recharged
and then starts to search again. The state diagram of the robots controller while
using this method is shown in figure 1.
Task Allocation in Robotic Swarms
5
Fig. 1. Method1 and Method 2 (by considering the dark green state which is shown
by dashed border) state diagram.
3.2 Dynamic communication based method
This method (figure 1) is similar to the first method in a way that it supports
dynamic task allocation in such a way that the robot will change its target color
probabilistically after a time step. For example consider that a robot has 4 spots
in its private list after 100 iterations containing 1 black and 3 green. Then the
robot will set its target color to black by probability and also will set it to
green by probability . This is why we expect this method to be better than
the previous one, because in this method the searching time for each robot to
find the appropriate spot may be decreased. if the robots are not distributed
uniformly in the environment at the beginning, it is possible that some spots
are not discovered after a small time steps (100 time steps). This is due to their
large distances from the initial places of the robots and also attitude changing
mechanism of the robots.
3.3 Decentralized Chapar Method
In this method we have some radio turrets called Chapar stations which are used
for radio communications and also we assume two groups of robots; workers (as
defined in the previously explained methods) and Chapars that transfer messages
between workers and Chapar stations. "Chapar Khaneh" or "Chapar-Khaneh" is
a term in Persian, meaning the "house of courier" as "Chapar" means "courier",
referring to the postal service used during the Achaemenid era [14]. Here we
used Chapar station instead of Chapar-khaneh.
High speed robots with simple structure are considered as Chapars. Chapars
flood their updated private message that is modified by some workers to the area
of their radio frequency coverage. Once a Chapar realizes a new row in its private
message (showing that a new spot has been discovered) it quickly goes to the
nearest Chapar station and sends its private message to it and also updates the
6
Aryo Jamshidpey and Mohsen Afsharchi
Fig. 2. Left: Method 4 State Diagram, Right: Method 3 State dDiagram.
message based on the content of the messages sent by the Chapar stations. In
addition each Chapar goes to the nearest Chapar station periodically (e.g. each
100 time steps) to update its private message and access to the most recently
information about the status of the spots.
Chapar stations are also in the coverage of each other and so they replicate
messages to keep the whole system up to date. In the simplest variation of this
model, there could be only one Chapar station whit limited coverage area. To
achieve high performance with the lowest cost, this method needs to coordinate
Chapar stations in a manner that their communication radiuses cover each other
sequentially. Although this method has a higher message transfer speed with
respect to the first and second methods, but it needs radio frequency transmitters
with higher coverage area and also in large environments it needs more Chapar
stations. Nevertheless, this method can perform fine even if Chapar stations
are not in the coverage area of each other. In this situation the messages may
not be similar in all Chapar stations but we can assume the environment as
some sub-environments each with a single Chapar station that can transform
messages with high speed. In a timely manner the information of one area may
be transferred to another area by Chapars. The state diagram of the Chapars
controller while using this method is shown in figure 2.
3.4 Centralized Chapar Method
This method (figure 2) is similar to the decentralized one in which there is
only one Chapar station that covers the entire environment. In this, method
Chapars have a single task to transfer messages from robots to the Chapar
station and they do it once they realize a new row in their private messages.
When the Chapar station receives a message that is not similar to its private
message, it updates the private message and sends it to all robots. To prevent
the Chapar station from single point of failure, in addition to message updating
by Chapars, they should transfer message in the environment quickly. In spite
of the centralized method, each Chapars only goes to the Chapar station when
Task Allocation in Robotic Swarms
7
it encounters a new spot in its private message. It should be reminded that the
robots used in the third and fourth methods, have the same behavior as the
robots that were used for one of the first or second method (optionally) and as
explained before, their behavior are the same as one of them.
4 Simulations
We have used e-puck robots in simulations which are extended by color detection
Infrared sensors, positioning module, compass and limited-range radio waves
emitter and receiver. Since the purpose of this article is to involve a wide range of
robots and the use of simpler hardware, we considered them without any camera.
In this way the methods can show their power in using of blind robots. All
experiments have been implemented in a 3m x 3m square environment enclosed
with walls by using of Webots as robotic simulation software. In two general
experiments, performance of the four proposed methods is evaluated individually
before and after energy consumption and in both of them 10 robots which are
called workers with IDs from 1 to 10 are used. Initially foraging mode of the
robots with IDs from 1 to 5 is adjusted to green and the robots with IDs from
6 to 10 is adjusted to black.
Further more, the speed of each method is determined by its foraging(searching)
energy threshold, hence setting the threshold is very important. It should not be
so low that the robot does not have enough time for foraging and also it should
not be so high that robots use their whole energy without any success or much
of it which in this case its performance will be reduced when it deploys in a
spot. Some factors such as area of environment, ratios of tasks, density of tasks
and the number of robots can help us to determine an appropriate value for this
parameter. In all of our experiments, the energy threshold level is considered as
250 iterates of each robots control code.
For simplicity, the details of cleaning and sampling operations are ignored,
the colored spots are considered as 30cm x 30cm squares and finally the length
of each robots communication radius is considered larger than the diameter of
each colored square. In both experiments, for each 300 square centimeters of
each spot one worker is sufficient for covering it desirably, which means that all
spots are covered desirably with 3 workers. Indeed as mentioned previously, with
less number of workers spot covering is also possible, but with lower speed.
4.1 First experiment: performance evaluation of proposed methods
before the threshold energy
Since the third and fourth methods use the worker robot which its controller is
that of the first or second method, it is not necessary to compare the performance
of all methods before foraging energy consumption (before the threshold).So we
have compared only the performance of the first and second methods before the
threshold. For this purpose four environments covered by green and black spots
and also containing obstacles are considered. The first environment has 3 green
8
Aryo Jamshidpey and Mohsen Afsharchi
Fig. 3. Right: Average number of successful robots before the threshold. Left:Average
number of discovered spots.
spots and 3 black ones, the second has 4 green spots and 2 black ones, the third
includes 5 green spots and 1 black ones and finally the fourth contains only 6
green spots. In each of the four areas, both the first and second methods are
tested 10 times separately. The average number of successful robots before the
energy threshold for the fist and second methods is shown in figure 3 (right).
It can be seen that, except for the first environment in which the average
number of successful robots before the threshold are equal for both methods,
in other environments, the average number of successful robots in the second
method is higher than the first one. This disparity grows by moving from the
first environment to the fourth and its reason is the changing attitude mechanism
which is used by second method's robots during their foraging operation. As
a result this leads to increasing in the number of robots which their attitude
changes when the number of green spots rise and the number of black ones
falls respectively. Subsequently this process will result in forming approximate
stability in the number of successful robots before the threshold. As it is shown
in figure 3 (right), we can conclude that in unknown environments where the
number of colored spots (black and green in our example) and the ratios of them
are unknown, the second method is more successful than the first one in terms
of the robots' attempts in finding spots by themselves before energy threshold.
In the next step the average number of green spots that have been found by
workers(by both the first and the second methods) before the threshold is shown
in figure 3 (left). In this figure the obtained curve from the second method is
steeper than the first methods one. Its primary reason is increasing the number
of green spots and decreasing the black ones which results in increasing the
probability of finding new green spots in both methods. But the reason that the
second methods curve is more steeper, is increasing of the number of workers
which search for green spots during the search time. As mentioned before this
is because of changing the robots attitudes during foraging operations. Figure 3
(left) shows that in unknown environments where the number of colored spots
(black and green in our example) and their ratios are unknown, the second
Task Allocation in Robotic Swarms
9
method is more successful than the first one in spot finding before the energy
threshold.
Figure 4 (right) shows the average number of robots deployed in the green
spots before the end of the foraging energy in both the first and second methods.
It can be observed that the steepness of the second methods curve is ascending
linear but the steepness of the first methods curve is sub linear. In the first
method the maximum number of robots in green spots is 5 because only 5
robots have green initial foraging modes and there is no any changing attitude
mechanism before threshold. But in the second method there is changing attitude
mechanism and so the maximum number of robots in green spots might be 10
and this means the whole colony. We can conclude that in unknown environments
where the number of colored spots (black and green in our example) and the
ratios of them are unknown, adaptability has a significant positive impact on
the performance of robots. As it can be observed, the second method is more
successful than the first one in the average number of robots which have deployed
in the green spots before the threshold. To sum up, from the above results in
unknown environments with the features which are mentioned above, the second
method is more efficient than the first one before the energy threshold.
4.2 Second experiment: performance evaluation of proposed
methods after the threshold energy
We use an environment consisting of 3 green spots and 3 black ones for evaluat-
ing the performance of the methods after foraging energy consumption. In this
area, all proposed methods are separately tested 10 times with random initial
distribution of robots. In the third and fourth methods, in addition to 10 workers,
another 3 Chapar robots that are faster than the workers are used too. It should
be mentioned about Chapar station that the third method is equipped with 3
Chapar stations which their communication radius cover each other sequentially
and the fourth method is equipped with one of them, which is omniscient. Figure
4 (left) shows the results of the second experiment in abstract. The absorption
percentage is the percentage of successfulness of finding spots by robots after
the threshold by applying the decision making mechanism.
As expected, the fourth method has the highest absorption percentage which
means 100%. This is due to the use of global message transferring system. The
third method is in second place with 87.09% and after it the first and the second
methods with approximate absorption 76% are both in third place. It should be
reminded that both of the first and second methods have the same communi-
cation mechanism. Higher speed in communicating will lead to making a better
task selection based on their colors, distances and number of hops by unsuccess-
ful workers. It results in a better robot distribution that reduces saturation and
starvation as far as possible. It also causes more unsuccessful robots attract to
the spots.
10
Aryo Jamshidpey and Mohsen Afsharchi
5 Discussion
Most studies in swarm robotic define their own test scenario, which is then
used to build a concrete swarm robotic system capable of solving the problem.
This leads to a huge amount of differently designed global missions and as a
result to many different solutions which are hard to compare[3]. Thus in most of
the proposed methods in this area, researchers have only introduced their own
methods and refrained from comparing with other methods. Our scenario also
possesses different features, goals and finally distinct global foraging mission
compared to previous scenarios in task allocation field. Accordingly applying
current approaches in our scenario and consequently comparing them with each
other is so difficult and in most of the times is impossible, except for changing the
scenario which in turn leads to changing both the problem and the solutions. As
pointed out before, Dasgupta's method is practical for unknown environments
with this difference that its global mission is different from what have been
proposed here. However our proposed methods have distinct advantageous over
Dasgupta's.
The proposed methods have the least waste time for robots as apposed to
that of Dasgupta in which, robots might be idle in environment for a long time.
These methods have been designed for the energy constrained scenario compared
to Dasgupta's. Consequently the priority of discovering new spots is higher than
accomplishing a task and then continuing unlimited foraging. Further more in
contrast with Dasgupta's method(which robots may be in idle mode for a long
time) our methods are close to optimal in the sense of idle mode.
On the other hand, our approaches are very practical for scenarios which
detecting a task should be done in a limited time. However in Dasgupta's ap-
proach, since after detecting a task other neighbor robots quit foraging and wait
for accomplishing the task sequentially, the probability of identifying other tasks
in a limited time will decrease. Moreover the proposed methods in this paper are
practical for blind robots while Dasgupta's method is practical for robots with
cameras which ignoring use of them obviously reduces cost in swarm robotics.
Also there is no more need to apply sophisticated techniques of machine vision
in order to determine starvation (which may have some errors).
5.1 Scalability and Robustness in term of single point of failure
In this section, we will analyze the scalability and robustness of the proposed
methods with respect to their efficiency. If by adding or removing robots the
performance of a method drops off, we will call it unscalable. Accordingly, both
the first and second methods are scalable because both are distributed, au-
tonomous and based on the local communication. So Adding or removing robots
will maintain their effectiveness. The decentralized Chapar method also has scal-
ability property. This is due to using of the limited range Chapar station, the
worker robots and the Chapar robots that are distributed and behave locally.
But the centralized Chapar method, due to using of omniscient Chapar station
with respect to its expected efficiency does not have scalability property.
Task Allocation in Robotic Swarms
11
Fig. 4. Left:Results for the second experiment.Right:Average number of robots de-
ployed in green spots.
With respect to robustness, the first and second methods are completely
robust against single point of failure because they behave in a distributed and
autonomous manner. If a robot fails outside a spot, others will consider it as an
obstacle. If it fails inside a spot and before announcing the center of it, there is
no problem because it looks like the situation in which the spot is not discovered
yet. But if it fails after the center announcement, it causes only decreasing in
speed of the operations in the spot, because they can be carried out by other
active robots on the spot yet.
The third method is less robust. The strength of the method for worker
robots is such as the first and second methods. In the case of Chapar robots
failures, there will be not any problem because in the worst case the speed of
communication will be reduced and other robots largely compensate this loss.
But about Chapar station as we have mentioned previously in describing the
method to benefit the high-speed informing relative to the cost, it is necessary
that their communication radius cover each other sequentially. Thus, if one of
Chapar stations fails, the connection between two parts of the environment will
be lost. Of course this problem can be solved by putting more than two Cha-
par stations in their each other communication ranges, but in this way costs
will increase. It can be said that this method is robust against single point of
failure. But this robustness is less than the first two methods (with respect to
maintaining the efficiency of the method). Centralized Chapar method has the
lowest robustness than the previous three methods because this method is only
have a wide range Chapar station. So if this Chapar station fails for any reason,
the efficiency of the method will be reduced so much.
6 Conclusions
In this paper a new practical scenario in swarm robotics is presented which is
about task allocation in unknown environments. Here we consider that there
12
Aryo Jamshidpey and Mohsen Afsharchi
is limited source of energy for each Robot. It is pointed out that energy man-
agement in the form of a 3 level structure is essential and four self-organized
threshold-based methods are proposed for solving the scenario.
Moreover, in two general experiments, the performance of them is analyzed.
From the results of simulations we can conclude that the second method is
more efficient in comparison to the first one before the threshold and so the
performance of the third and the fourth methods will increase by applying it
as their worker robot's algorithm. However, since the methods are using similar
communication mechanisms, they have almost the same performance after the
threshold.
As a result, from performance point of view, the fourth method has the
highest performance, followed by the third method which in turn, has higher
rank than the two others. The second method also is better than the first one
in this criteria. On the other hand, from scalability and robustness perspective,
both first and second methods have higher rank than the two others and further
more the third method has higher rank than the fourth.
References
1. Bonabeau, E., Dorigo, M., Theraulaz, G. (1999). Swarm intelligence: From natural
to artificial systems: Oxford university press.
2. Brutschy, A., Tran, N. L., Baiboun, N., Frison, M., Pini, G., Roli, A., Birattari,
M. 2012. Costs and benefits of behavioral specialization. Robotics and autonomous
systems, 60(11), 1408-1420.
3. Burger, M. 2012. Mechanisms for Task Allocation in Swarm Robotics (Doctoral
dissertation, Diploma thesis, Ludwig-Maximilians-Universitt Mnchen).
4. Dahl, T. S., Matari, M., Sukhatme, G. S. 2009. Multi-robot task allocation through
vacancy chain scheduling. Robotics and Autonomous Systems, 57(6), 674-687.
5. Dasgupta, P. 2011. Multi-robot task allocation for performing cooperative forag-
ing tasks in an initially unknown environment. In Innovations in Defence Support
Systems-2 (pp. 5-20). Springer Berlin Heidelberg.
6. Dias, M. B. 2004. Traderbots: A new paradigm for robust and efficient multirobot
coordination in dynamic environments (Doctoral dissertation, Carnegie Mellon Uni-
versity).
7. Dias, M. B., Zlot, R., Kalra, N., Stentz, A. 2006. Market-based multirobot coordi-
nation: A survey and analysis. Proceedings of the IEEE, 94(7), 1257-1270.
8. Jones, C., Mataric, M. J. 2003. Adaptive division of labor in large-scale minimalist
multi-robot systems. In Intelligent Robots and Systems, 2003.IROS 2003. Proceed-
ings. 2003 IEEE/RSJ International Conference on (Vol. 2, pp. 1969-1974). IEEE.
Integration. Technical report, Global Grid Forum (2002)
9. Krieger, M. J., Billeter, J. B. 2000. The call of duty: Self-organised task allocation
in a population of up to twelve mobile robots. Robotics and Autonomous Systems,
30(1), 65-84.
10. Labella, T. H. 2007. Division of labour in groups of robots (Doctoral dissertation,
Ph. D. thesis, Universite Libre de Bruxelles).
11. Liu, W., Winfield, A., Sa, J., Chen, J., Dou, L. 2007. Strategies for energy opti-
mization in a swarm of foraging robots. In Swarm robotics (pp. 14-26). Springer
Berlin Heidelberg.
Task Allocation in Robotic Swarms
13
12. Liu, W., Winfield, A. F., Sa, J., Chen, J.,Dou, L. 2007. Towards energy optimiza-
tion: Emergent task allocation in a swarm of foraging robots. Adaptive Behavior,
15(3), 289-305.
13. Nouyan, S., Gro, R., Bonani, M., Mondada, F., Dorigo, M. 2009. Teamwork in self-
organized robot colonies. Evolutionary Computation, IEEE Transactions on, 13(4),
695-711.
14. Wikipedia, Chapr Khaneh, http://www.en.wikipedia.org/wiki/Chapar_Khaneh
|
1309.3197 | 2 | 1309 | 2013-10-22T13:39:51 | Inducing Honest Reporting Without Observing Outcomes: An Application to the Peer-Review Process | [
"cs.MA",
"cs.AI",
"cs.DL",
"math.ST",
"math.ST"
] | When eliciting opinions from a group of experts, traditional devices used to promote honest reporting assume that there is an observable future outcome. In practice, however, this assumption is not always reasonable. In this paper, we propose a scoring method built on strictly proper scoring rules to induce honest reporting without assuming observable outcomes. Our method provides scores based on pairwise comparisons between the reports made by each pair of experts in the group. For ease of exposition, we introduce our scoring method by illustrating its application to the peer-review process. In order to do so, we start by modeling the peer-review process using a Bayesian model where the uncertainty regarding the quality of the manuscript is taken into account. Thereafter, we introduce our scoring method to evaluate the reported reviews. Under the assumptions that reviewers are Bayesian decision-makers and that they cannot influence the reviews of other reviewers, we show that risk-neutral reviewers strictly maximize their expected scores by honestly disclosing their reviews. We also show how the group's scores can be used to find a consensual review. Experimental results show that encouraging honest reporting through the proposed scoring method creates more accurate reviews than the traditional peer-review process. | cs.MA | cs |
Inducing Honest Reporting Without Observing Outcomes:
An Application to the Peer-Review Process
Arthur Carvalho
University of Waterloo
[email protected]
Stanko Dimitrov
University of Waterloo
[email protected]
Kate Larson
University of Waterloo
[email protected]
May 11, 2014
Abstract
When eliciting opinions from a group of experts, traditional devices used to promote hon-
est reporting assume that there is an observable future outcome. In practice, however, this as-
sumption is not always reasonable. In this paper, we propose a scoring method built on strictly
proper scoring rules to induce honest reporting without assuming observable outcomes. Our
method provides scores based on pairwise comparisons between the reports made by each
pair of experts in the group. For ease of exposition, we introduce our scoring method by illus-
trating its application to the peer-review process. In order to do so, we start by modeling the
peer-review process using a Bayesian model where the uncertainty regarding the quality of
the manuscript is taken into account. Thereafter, we introduce our scoring method to evaluate
the reported reviews. Under the assumptions that reviewers are Bayesian decision-makers and
that they cannot influence the reviews of other reviewers, we show that risk-neutral reviewers
strictly maximize their expected scores by honestly disclosing their reviews. We also show
how the group’s scores can be used to find a consensual review. Experimental results show
that encouraging honest reporting through the proposed scoring method creates more accurate
reviews than the traditional peer-review process.
1 Introduction
In the absence of a well-chosen incentive structure, experts are not necessarily honest when re-
porting their opinions. For example, when reporting subjective probabilities, experts who have a
reputation to protect might tend to produce forecasts near the most likely group consensus, whereas
experts who have a reputation to build might tend to overstate the probabilities of outcomes they
feel will be understated in a possible consensus [Nakazono, 2013]. Hence, an important question
when eliciting experts’ opinions is how to incentivize honest reporting.
Proper scoring rules [Winkler and Murphy, 1968] are traditional devices that incentivize hon-
est reporting of subjective probabilities, i.e., experts maximize their expected scores by honestly
reporting their opinions. However, proper scoring rules rely on the assumption that there is an ob-
servable future outcome, which is not always a reasonable assumption. For example, when market
analysts provide sales forecasts on a potential new product, there is no guarantee that the product
will ever be produced. Hence, the actual number of sales may never be observed.
In this paper, we propose a scoring method for promoting honest reporting amongst a group
of experts when future outcomes are unobservable.
In particular, we are interested in settings
where experts observe signals from a multinomial distribution with an unknown parameter. Honest
reporting then means that experts report exactly the signals that they observed. Our scoring method
is built on proper scoring rules. However, different than what is traditionally assumed in the proper
1
scoring rules literature, our method does not assume that there is an observable future outcome.
Instead, scores are determined based on pairwise comparisons between experts’ reported opinions.
The proposed method may be used in a variety of settings, e.g., strategic planning, reputation
systems, peer review, etc. When applied to strategic planning, the proposed method may induce
honest evaluation of different strategic plans. A strategic plan is a systematic and coordinated
way to develop a direction and a course for an organization, which includes a plan to allocate the
organization’s resources [Argenti, 1968]. After a candidate strategic plan is discarded, it becomes
nearly impossible to observe what would be the consequences of that plan because strategic plans
are long-term in nature. Hence, a method to incentivize honest evaluations of candidate strategic
plans cannot assume that the result of a strategic plan is observable in the future.
Our method can also be applied to reputation systems to elicit honest feedback. In reputation
systems, individuals rate a product/service after experiencing it, e.g., customer product reviews on
Amazon.com are one such reputation system. Due to the subjective nature of this task, incentives
for honest feedback should not be based on the assumption that an absolute rating exists.
For ease of exposition, we introduce our scoring method by illustrating its application to a
domain where traditionally there are no observable outcomes: the peer-review process. Peer re-
view is a process in which an expert’s output is scrutinized by a number of other experts with
relevant expertise in order to ensure quality control and/or to provide credibility. Peer review is
commonly used when there is no objective way to measure the output’s quality, i.e., when qual-
ity is a subjective matter. Peer review has been widely used in several professional fields, e.g.,
accounting [AICPA: American Institute of CPAs, 2012], law [LSC: Legal Services Commission,
2005], health care [Dans, 1993], etc.
Currently, a popular application of the peer-review process is in online education. Recent years
have seen a surge of massive online open courses, i.e., free online academic courses aimed at large-
scale participation. Some of these courses have attracted tens of thousands of students [Pappano,
2012]. One of the biggest challenges faced by online educators brought by this massive number
of students is the grading process since the available resources (personnel, time, etc.) is often
insufficient. Auto-grading by computers is not always feasible, e.g., courses whose assignments
consist of essay-style questions and/or questions that do not have clear right/wrong answers. Peer
review has been used by some companies like Coursera1 as a way to overcome this issue.
For simplicity’s sake, we focus on peer review as used in modern scientific communication.
The process, as we consider in this paper, can be described as follows: when a manuscript first
arrives at the editorial office of an academic journal, it is first examined by the editor, who might
reject the manuscript immediately because either it is out-of-scope or because it is of unacceptable
quality. Manuscripts that pass this first stage are then sent out to experts with relevant expertise
who are usually asked to classify the manuscript as publishable immediately, publishable after
some revisions, or not publishable at all. Traditionally, the manuscript’s authors do not know the
reviewers’ identities, but the reviewers may or may not know the identity of the authors.
In other words, peer review can be seen as a decision-making process where the reviewers
serve as cognitive inputs that help a decision maker (chair, editor, course instructor, etc.) judge
the quality of a peer’s output. A crucial point in this process is that it greatly depends on the
reviewers’ honesty.
In the canonical peer-review process, reviewers have no direct incentives
for honestly reporting their reviews. Several potential problems have been discussed in different
research areas, e.g., bias against female authors, authors from minor institutions, and non-native
English writers [Bornmann et al., 2007, Wenneras and Wold, 1997, Primack and Marrs, 2008,
Newcombe and Bouton, 2009].
In order to illustrate the application of our method to peer review, we start by modeling the
peer-review process as a Bayesian model so as to take the uncertainty regarding the quality of
1http://www.coursera.org/
2
the manuscript into account. We then introduce our scoring method to evaluate reported reviews.
We assume that the scores received by reviewers are somehow coupled with relevant incentives,
be they social-psychological, such as praise or visibility, or material rewards through prizes or
money. Hence, we naturally assume that reviewers seek to maximize their expected scores and
that there are no external incentives. We show that reviewers strictly maximize their expected
scores by honestly disclosing their reviews under the additional assumptions that they are Bayesian
decision-makers and that they cannot influence the reviews of other reviewers.
Honesty is intrinsically related to accuracy in our peer-review model: as the number of honest
reviews increases, the distribution of the reported reviews converges to the probability distribution
that represents the quality of the manuscript. We performed peer-review experiments to validate
the model and to test the efficiency of the proposed scoring method. Our experimental results cor-
roborate our theoretical model by showing that the act of encouraging honest reporting through the
proposed scoring method creates more accurate reviews than the traditional peer-review process,
where reviewers have no direct incentives for expressing their true reviews.
In addition to our method for inducing honest reporting, we also propose a method to aggregate
opinions that uses information from experts’ scores. Our aggregation method is general in a sense
that it can be used in any decision-making setting where experts report probability distributions
over the outcomes of a discrete random variable. The proposed method works as if the experts
were continuously updating their opinions in order to accommodate the expertise of others. Each
updated opinion takes the form of a linear opinion pool, where the weight that an expert assigns to
a peer’s opinion is inversely related to the distance between their opinions. In other words, experts
are assumed to prefer opinions that are close to their own opinions, where closeness is defined
by an underlying proper scoring rule. We provide conditions under which consensus is achieved
under our aggregation method and discuss a behavioral foundation of it. Using data from our peer-
review experiments, we find that the consensual review resulting from the proposed aggregation
method is consistently more accurate than the canonical average review.
2 Related Work
In recent years, two prominent methods to induce honest reporting without the assumption of
observable future outcomes were proposed: the Bayesian truth serum (BTS) method [Prelec, 2004]
and the peer-prediction method [Miller et al., 2005].
The BTS method works on a single multiple-choice question with a finite number of alterna-
tives. Each expert is requested to endorse the answer mostly likely to be true and to predict the
empirical distribution of the endorsed answers. Experts are evaluated by the accuracy of their pre-
dictions as well as how surprisingly common their answers are. The surprisingly common criterion
exploits the false consensus effect to promote truthfulness, i.e., the general tendency of experts to
overestimate the degree of agreement that the others have with them.
The score received by an expert from the BTS method has two major components. The first
one, called the information score, evaluates the answer endorsed by the expert according to the
log-ratio of its actual-to-predicted endorsement frequencies. The second component, called the
prediction score, is a penalty proportional to the relative entropy between the empirical distribu-
tion of answers and the expert’s prediction of that distribution. Under the BTS scoring method,
collective honest reporting is a Bayes-Nash equilibrium.
The BTS method has been used to promote honest reporting in many different domains, e.g.,
when sharing rewards amongst a set of experts [Carvalho and Larson, 2011] and in policy anal-
ysis [Weiss, 2009]. However, the BTS method has two major drawbacks. First, it requires the
population of experts to be large. Second, besides reporting their opinions, experts must also
make predictions about how their peers will report their opinions. While the artificial intelligence
3
community has recently addressed the former issue [Witkowski and Parkes, 2012, Radanovic and
Faltings, 2013], the latter issue is still an intrinsic requirement for using the BTS method.
The drawbacks of the BTS method are not shared by the peer-prediction method [Miller et al.,
2005]. In the peer-prediction method, a number of experts experience a product and rate its quality.
A mechanism then collects the ratings and makes payments based on those ratings. The peer-
prediction method makes use of the stochastic correlation between the signals observed by the
experts from the product to achieve a Bayes-Nash equilibrium where every expert reports honestly.
A major problem with the peer-prediction method is that it depends on historical data. For
example, when applied to a peer-review setting, after a reviewer i reports his review, say ri, the
mechanism then estimates reviewer i’s prediction of the review reported by another reviewer j,
P (rjri), which is then evaluated and rewarded using a proper scoring rule and reviewer j’s actual
reported review. The mechanism needs to have a history of previously reported reviews for com-
puting P (rjri), which is not always a reasonable assumption, e.g., when the evaluation criteria
may change from review to review and when the peer-review process is being used for the first
time. In other words, the peer-prediction method is prone to cold-start problems.
Carvalho and Larson [2012] addressed this issue by making the extra assumption that experts
have uninformative prior knowledge about the distribution of the observed signals. Given this
assumption, honest reporting is induced by simply making pairwise comparisons between reported
opinions and rewarding agreements. In this paper, we extend the method by Carvalho and Larson
[2012] in several ways. First, we show that the assumption of uninformative priors is unnecessary
as long as experts have common prior distributions and this fact is common knowledge. Moreover,
we provide stronger conditions with respect to the underlying proper scoring rule under which
pairwise comparisons induce honest reporting.
Another contribution of our work is a method to aggregate the reported opinions into a sin-
gle consensual opinion. Over the years, both behavioral and mathematical methods have been
proposed to establish consensus [Clemen and Winkler, 1999]. Behavioral methods attempt to
generate agreement through interaction and exchange of knowledge. Ideally, the sharing of infor-
mation leads to a consensus. However, behavioral methods usually provide no conditions under
which experts can be expected to reach an agreement. On the other hand, mathematical aggre-
gation methods consist of processes or analytical models that operate on the reported opinions in
order to produce a single aggregate opinion. DeGroot [1974] proposed a model which describes
how a group of experts can reach agreement on a consensual opinion by pooling their individual
opinions. A drawback of DeGroot’s method is that it requires each expert to explicitly assign
weights to the opinions of other experts. In this paper, we propose a method to set these weights
directly which takes the scores received by the experts into account. We also provide a behavioral
interpretation of the proposed aggregation method.
A related method for finding consensus was proposed by Carvalho and Larson [2013]. Under
the assumption that experts prefer probability distributions close to their own distributions, where
closeness is measured by the root-mean-square deviation, the authors showed that a consensus
is always achieved. Moreover, if risk-neutral experts are rewarded using the quadratic scoring
rule, then the assumption that experts prefer probability distributions that are close to their own
distributions follows naturally. The approach in this paper is more general because the underlying
proper scoring rule can be any bounded proper scoring rule.
From an empirical perspective, we investigate the efficiency of both our scoring method and
our method for finding consensus in a peer-review experiment. Formal experiments involving
peer review are still relatively scarce. Even though the application of the peer-review process to
scientific communication can be traced back almost 300 years, it was not until the early 1990s that
research on this matter became more intensive and formalized [van Rooyen, 2001]. Scientists in
the biomedical domain have been in the forefront of research on the peer-review process due to
4
the fact that dependable quality-controlled information can literally be a matter of life and death
in this research field.
In particular, the staff of the renowned BMJ, formerly British Medical
Journal, have been studying the merits and limitations of peer review over a number of years
[Lock, 1985, Godlee et al., 2003]. Most of their work has focused on defining and evaluating
review quality [van Rooyen et al., 1999], and examining the effect of specific interventions on the
quality of the resulting reviews [van Rooyen, 2001].
One mechanism used to prevent bias in the peer-review process is called double-blind review,
which consists of hiding both authors and reviewers’ identities. Indeed, it has been reported that
such a practice reduces bias against female authors [Budden et al., 2008]. However, it can be
argued that knowing the authors’ identities makes it easier for the reviewers to compare the new
manuscript with previously published papers, and it also encourages the reviewers to disclose con-
flicts of interest. Another argument that undermines the benefits of double-blind reviewing is that
the authorship of the manuscript is often obvious to a knowledgeable reader from the context, e.g.,
self-referencing, research topic, writing style, working paper repositories, seminars, etc. [Falagas
et al., 2006, Justice et al., 1998, Yankauer, 1991]. Furthermore, this mechanism does not prevent
against certain types of bias, e.g., when a reviewer rejects new evidence or new knowledge because
it contradicts established norms, beliefs or paradigms.
Some work has focused on the calibration aspect of peer review. Roos et al. [2011] proposed a
maximum likelihood method for calibrating reviews by estimating both the bias of each reviewer
and the unknown ideal score of the manuscript. Bias is treated as the general rigor of a reviewer
across all his reviews. Hence, Roos et al.’s method does not attempt to prevent bias by rewarding
honest reporting. Instead, it adjusts reviews a posteriori so that they can be globally comparable.
Instead of calibrating reviews a posteriori, Robinson [2001] suggested to “calibrate” review-
ers a priori. Reviewers are first asked to review short texts that have gold-standard reviews, i.e.,
reviews of high quality provided by experts with relevant expertise. Thereafter, they receive cali-
bration scores, which are later used as weighting factors to determine how well their future reviews
will be considered. This approach, however, does not guarantee that reviewers will report honestly
after the calibration phase, when gold-standard reviews are no longer available.
To the best of our knowledge, our peer-review experiments are the first to investigate the use of
incentives for honest reporting in a peer-review task. When objective verification is not possible,
as in the peer-review process, economic measures may be used to encourage experts to honestly
disclose their opinions. The proposed scoring method does so by making pairwise comparisons
between reported reviews and rewarding agreements.
Rewarding experts based on pairwise comparisons has been empirically proven to be an ef-
fective incentive technique in other domains. Shaw et al. [2011] measured the effectiveness of a
collection of social and financial incentive schemes for motivating experts to conduct a qualitative
content analysis task. The authors found that treatment conditions that provided financial incen-
tives and asked experts to prospectively think about the responses of their peers produced more
accurate responses. Huang and Fu [2013] showed that informing the experts that their rewards will
be based on how similar their responses are to other experts’ responses produces more accurate
responses than telling the experts that their rewards will be based on how similar their responses
are to gold-standard responses. Our work adds to the existing body of literature by theoretically
and empirically showing that pairwise comparisons make the peer-review process more accurate.
5
3 The Basic Model
In our proposed peer-review process, a manuscript is reviewed by a set of reviewers N = {1, . . . , n},
with n ≥ 2. The quality of the manuscript is represented by a multinomial distribution2 Ω with
unknown parameter ω = (ω0, . . . , ωv), where v ∈ N+ represents the best evaluation score that the
manuscript can receive and ωk is the probability assigned to the evaluation score being equal to k.
Each reviewer is modeled as possessing a privately observed draw (signal) from Ω. Hence,
our model captures the uncertainty of the reviewers regarding the quality of the manuscript. We
extend the model to multiple observed signals in Section 5. We denote the honest review of each
reviewer i ∈ N by ti ∼ Ω, where ti ∈ {0, . . . , v}. Honest reviews are independent and identically
distributed, i.e., P (titj) = P (ti). We say that reviewer i is reporting honestly when his reported
review ri is equal to his honest review, i.e., ri = ti.
Reviews are elicited and aggregated by a trusted entity referred to as the center3, which is also
responsible for rewarding the reviewers. Let si be reviewer i’s review score after he reports ri.
We discuss how si is determined in Section 4. Review scores are somehow coupled with relevant
incentives, be they social-psychological, such as praise or visibility, or material rewards through
prizes or money. We make four major assumptions in our model:
1. Autonomy: Reviewers cannot influence other reviewers’ reviews, i.e., they do not know each
other’s identity and they are not allowed to communicate to each other during the reviewing
process.
2. Risk Neutrality: Reviewers behave so as to maximize their expected review scores.
3. Dirichlet Priors: There exists a common prior distribution over ω, i.e., P (ω). We assume
that this prior is a Dirichlet distribution and this is common knowledge.
4. Rationality: After observing ti, every reviewer i ∈ N updates his belief by applying Bayes’
rule to the common prior, i.e., P (ωti).
The first assumption describes how peer review is traditionally done in practice. The sec-
ond assumption means that reviewers are self-interested and no external incentives exist for each
reviewer. The third assumption means that reviewers have common prior knowledge about the
quality of the manuscript, a natural assumption in the peer-review process. We discuss the formal
meaning of such an assumption in the following subsection. The fourth assumption implies that
the posterior distributions are consistent with Bayesian updating, i.e.:
P (ωti) =
P (tiω)P (ω)
P (ti)
The last three assumptions imply that reviewers are Bayesian decision-makers. We note that
different modeling choices could have been used, e.g., models based on games of incomplete
information. Unlike our model, an incomplete-information game is often used when experts do not
know each other’s beliefs. To find strategic equilibria in such incomplete-information models, one
would need information about experts’ beliefs about each other’s private information. A Bayesian
structure could be used to model each expert’s beliefs about the others, and it would permit the
calculation of experts’ expected scores, which are maximized at equilibrium. However, the natural
autonomy assumption makes such a Bayesian structure unrealistic.
2We use the term multinomial distribution to refer to the generalization of the Bernoulli distribution for discrete
random variables with any constant number of outcomes. The parameter of this distribution is a probability vector that
specifies the probability of each possible outcome.
3We refer to a single reviewer as “he” and to the center as “she”.
6
3.1 Dirichlet Distributions
An important assumption in our model is that reviewers have Dirichlet priors over distributions of
evaluation scores. The Dirichlet distribution can be seen as a continuous distribution over param-
eter vectors of a multinomial distribution. Since ω is the unknown parameter of the multinomial
distribution that describes the quality of the manuscript, then it is natural to consider a Dirichlet
distribution as a prior for ω. Given a vector of positive integers, α = (α0, . . . , αv), that determines
the shape of the Dirichlet distribution, the probability density function of the Dirichlet distribution
over ω is:
where:
P (ωα) =
1
β(α)
v
Yk=0
ωαk−1
k
(1)
β(α) = Qv
(Pv
k=0(αk − 1)!
k=0 αk − 1)!
Figure 1 shows the above probability density when v = 2 for some parameter vectors α.
k=0 αk. The
probability vector E[ωα] = (E[ω0α], . . . , E[ωvα]) is called the expected distribution regarding
ω.
For the Dirichlet distribution in (1), the expected value of ωj is E[ωjα] = αi/Pv
An interesting property of the Dirichlet distribution is that it is the conjugate prior of the
multinomial distribution [Bernardo and Smith, 1994], i.e., the posterior distribution P (ωα, ti)
is itself a Dirichlet distribution. This relationship is often used in Bayesian statistics to estimate
hidden parameters of multinomial distributions. To illustrate this point, suppose that reviewer i
observes the signal ti = x, for x ∈ {0, . . . , v}. After applying Bayes’ rule, reviewer i’s posterior
distribution is P (ωα, ti = x) = P (ω(α0, α1, . . . , αx + 1, . . . αv)). Consequently, the new
expected distribution is:
, . . . ,
αx + 1
k=0 αk
1 +Pv
αv
k=0 αk(cid:19) (2)
1 +Pv
E[ωα, ti = x] =(cid:18)
α0
k=0 αk
,
1 +Pv
1 +Pv
α1
k=0 αk
, . . . ,
We call the probability vector in (2) reviewer i’s posterior predictive distribution regarding
ω because it provides the distribution of future outcomes given the observed data ti. With this
perspective, we regard the values α0, . . . , αv as “pseudo-counts” from “pseudo-data”, where each
αk can be interpreted as the number of times that the ωk-probability event has been observed
before.
Throughout this paper, we assume that reviewers have common prior Dirichlet distributions
and this fact is common knowledge, i.e., the value of α is initially the same for all reviewers. A
practical interpretation of this assumption is that reviewers have common prior knowledge about
the quality of the manuscript, i.e., reviewers have a common expectation regarding the quality of
arriving manuscripts.
By using Dirichlet distributions as priors, belief updating can be expressed as an updating
of the parameters of the prior distribution4. Furthermore, the assumption of common knowledge
4We note that other priors could have been used. However, the inference process would not necessarily be analyti-
cally tractable. In general, tractability can be obtained through conjugate distributions. Hence, another modeling choice
is to consider that evaluation scores follow a normal distribution with unknown parameters. Assuming exchangeability,
we can then use either the normal-gamma distribution or the normal-scaled inverse gamma distribution as the conjugate
prior [Bernardo and Smith, 1994]. The major drawback with this approach is that continuous evaluation scores might
bring extra complexity to the reviewing process.
7
Figure 1: Probability densities of Dirichlet distributions when v = 2 for different parameter
vectors. Left: α = (1, 1, 1). Center: α = (2, 1, 1). Right: α = (2, 2, 2).
allows the center to estimate reviewers’ posterior distributions based solely on their reported re-
views, a point which is explored by our proposed scoring method. Due to its attractive theoretical
properties, the Dirichlet distribution has been used to model uncertainty in a variety of different
scenarios, e.g, when experts are sharing a reward based on peer evaluations [Carvalho and Larson,
2012] and when experts are grouped based on their individual differences [Navarro et al., 2006].
4 Scoring Method
In this section, we propose a scoring method to induce honest reporting of reviews. The proposed
method is built on proper scoring rules [Winkler and Murphy, 1968].
4.1 Proper Scoring Rules
Consider an uncertain quantity with possible outcomes o0, . . . , ov, and a probability vector z =
(z0, . . . , zv), where zk is the probability value associated with the occurrence of outcome ok. A
scoring rule R(z, e) is a function that provides a score for the assessment z upon observing the
outcome oe, for e ∈ {0, . . . , v}. A scoring rule is called strictly proper when an expert receives his
maximum expected score if and only if his stated assessment z corresponds to his true assessment
q = (q0, . . . , qv) [Winkler and Murphy, 1968]. The expected score of z at q for a real-valued
scoring rule R(z, e) is:
Eq [R(z, e)] =
qe R(z, e)
v
Xe=0
8
Proper scoring rules have been used as a tool to promote honest reporting in a variety of do-
mains, e.g., when sharing rewards amongst a set of experts based on peer evaluations [Carvalho
and Larson, 2010, 2012], to incentivize experts to accurately estimate their own efforts to accom-
plish a task [Bacon et al., 2012], in prediction markets [Hanson, 2003], in weather forecasting
[Gneiting and Raftery, 2007], etc. Some of the best known strictly proper scoring rules, together
with their scoring ranges, are:
logarithmic: R(z, e) = log ze
(−∞, 0]
v
quadratic: R(z, e) = 2ze −
z2
k
[−1, 1]
(3)
spherical: R(z, e) =
[0, 1]
Xk=0
ze
k=0 z2
k
qPv
All the above scoring rules are symmetric, i.e., R((z0, . . . , zv), e) = R((zπ0, . . . , zπv ), πe),
for all probability vectors z = (z0, . . . , zv), for all permutations π on v + 1 elements, and for
all outcomes indexed by e ∈ {0, . . . , v}. We say that a scoring rule is bounded if R(z, e) ∈ R,
for all probability vectors z and e ∈ {0, . . . , v}. For example, the logarithmic scoring rule is
not bounded because it might return −∞ whenever the probability vector z contains an element
equal to zero, whereas both the quadratic and spherical scoring rules are always bounded. A
well-known property of strictly proper scoring rules is that they are still strictly proper under
positive affine transformations [Gneiting and Raftery, 2007], i.e., argmaxz
Eq [γR(z, e) + λ] =
argmaxz
Eq [R(z, e)] = q, for a strictly proper scoring rule R, γ > 0, and λ ∈ R.
Proposition 1. If R(z, e) is a strictly proper scoring rule, then a positive affine transformation of
R, i.e., γR(z, e) + λ, for γ > 0 and λ ∈ R, is also strictly proper.
4.2 Review Scores
If we knew a priori reviewers’ honest reviews, we could then compare the honest reviews to
the reported reviews and reward agreement. However, due to the subjective nature of the peer-
review process, we are facing a situation where this objective truth is practically unknowable.
Our solution is to induce honest reporting through pairwise comparisons of reported reviews.
The first step towards computing each reviewer i’s review score is to estimate his posterior pre-
dictive distribution E[ωα, ti] shown in (2) based on his reported review ri. Let E[ωα, ri] =
(E[ω0α, ri], . . . , E[ωvα, ri]) be such an estimation, where:
E[ωkα, ri] =(
αk+1
1+Pv
x=0 αx
αk
1+Pv
x=0 αx
if ri = k,
otherwise.
(4)
Recall that the elements of reviewer i’s true posterior predictive distribution are defined as:
E [ωkα, ti] =(
αk+1
1+Pv
x=0 αx
αk
1+Pv
x=0 αx
if ti = k,
otherwise.
Clearly, E [ωkα, ri] = E [ωkα, ti] if and only if reviewer i is reporting honestly, i.e., when
he reports ri = ti. The review score of reviewer i is determined as follows:
(γR (E [ωα, ri] , rj) + λ)
si =Xj6=i
9
(5)
where γ and λ are constants, for γ > 0 and λ ∈ R, and R is a strictly proper scoring rule. Scoring
rules require an observable outcome, or a “reality”, in order to score an assessment. Intuitively,
we consider each review reported by every reviewer other than reviewer i as an observed outcome,
i.e., the evaluation score deserved by the manuscript, and then we score reviewer i’s estimated
posterior predictive distribution in (4) as an assessment of that value.
Proposition 2. Each reviewer i ∈ N strictly maximizes his expected review score if and only if
ri = ti.
Proof. Let Θi = E [ωα, ti] and Φi = E [ωα, ri]. By the autonomy assumption, reviewers
cannot affect their peers’ reviews. Hence, we can restrict ourselves to show that each reviewer
i ∈ N strictly maximizes EΘi [γR (Φi, rj) + λ], for j 6= i, if and only if ri = ti.
(If part) Since R is a strictly proper scoring rule, from Proposition 1 we have that:
arg max
Φi
EΘi [γR (Φi, rj) + λ] = Θi
If ri = ti, then by construction Φi = Θi, i.e., the estimated posterior predictive distribution
in (4) is equal to the true posterior predictive distribution in (2). Consequently, honest reporting
strictly maximizes reviewers’ expected review scores.
(Only-if part). Using a similar argument, given that R is a strictly proper scoring rule, from
Proposition 1 we have that:
arg max
Φi
EΘi [γR (Φi, rj) + λ] = Θi
By construction, Φi = Θi if and only if ri = ti (see equations (2) and (4)). Thus, reviewers’
expected review scores are strictly maximized only when reviewers are honest.
Another way to interpret the above result is to imagine that each reviewer is betting on the
review deserved by the manuscript. Since the most relevant information available to him is the
observed signal, then the strategy that maximizes his expected review score is to bet on that signal,
i.e., to bet on his honest review. When this happens, the true posterior predictive distribution in
(2) is equal to the estimated posterior predictive distribution in (4) and, consequently, the expected
score resulting from a strictly proper scoring rule is strictly maximized when the expectation is
taken with respect to the true posterior predictive distribution.
It is important to observe that by incentivizing honest reporting, the scoring function in (5)
also incentivizes accuracy since honest reviews are draws from the distribution that represents the
true quality of the manuscript. In other words, the center is indirectly observing these draws when
reviewers report honestly. Consequently, due to the law of large numbers, the distribution of the
reported reviews converges to the distribution that represents the true quality of the manuscript as
the number of honestly reported reviews increases. Our experimental results in Section 7 show
that there indeed exists a strong correlation between honesty and accuracy.
Different interpretations of the scoring method in (5) arises depending on the underlying
strictly proper scoring rule and the hyperparameter α. In the following subsections, we discuss
two different interpretations: 1) when R is a symmetric and bounded strictly proper scoring rule
and reviewers’ prior distributions are non-informative; and 2) when R is a strictly proper scoring
rule sensitive to distance.
4.3 Rewarding Agreement
Assume that reviewers’ prior distributions are non-informative, i.e., all the elements making up
the hyperparameter α have the same value. This happens when reviewers have no relevant prior
10
knowledge about the quality of the manuscript. Consequently, the elements of reviewers’ true and
estimated posterior predictive distributions can take on only two possible values (see equations (2)
and (4) for α0 = α1 = · · · = αv ).
Moreover, if R is a symmetric scoring rule, then the term R (E [ωα, ri] , rj) in (5) can
take on only two possible values because a permutation of elements with similar values does
not change the score of a symmetric scoring rule. When R is also strictly proper, it means that
R (E [ωα, ri] , rj) = δmax, when ri = rj, and R (E [ωα, ri] , rj) = δmin, when ri 6= rj, where
δmax > δmin. Consequently, each term of the summation in (5) can be written as:
γR (E [ωα, ri] , rj) + λ =(cid:26) γδmax + λ if ri = rj,
γδmin + λ otherwise.
When R is also bounded, we can then set γ =
, and the
above values become, respectively, 1 and 0. Hence, the resulting review scores do not depend on
parameters of the model. Moreover, we obtain an intuitive interpretation of the scoring method in
(5): whenever two reported reviews are equal to each other, the underlying reviewers are rewarded
by one payoff unit. Thus, in practice, our scoring method works by simply comparing reported
reviews and rewarding agreements whenever R is a symmetric and bounded strictly proper scoring
rule and reviewers have no informative prior knowledge about the quality of the manuscript.
δmax−δmin
1
and λ = −δmin
δmax−δmin
Another interesting point is that the center can reward different agreements in different ways,
i.e., reviewers are not necessarily equally valued. For example, if the center knows a priori that
a particular reviewer j is reliable (respectively, unreliable), then she can increase (respectively,
decrease) the reward of reviewers whose reviews are in agreement with reviewer j’s reported
review. Formally, this means that for different reviewers i and j, the center can use different
values for γ and λ in (5). Proposition 2 is not affected by this as long as γ > 0, λ ∈ R, and
their values are independent of the reported reviews. Hence, by having a few reliable reviewers,
this approach might help to eliminate the hypothetical scenario where a set of reviewers learn over
time to report similar reviews. A similar idea was proposed by Jurca and Faltings [2009] to prevent
collusions in reputation systems.
4.4 Strictly Proper Scoring Rules Sensitive to Distance
Pairwise comparisons, as defined in the previous subsection, might work well for small values of
v, the best evaluation score that the manuscript can receive, but it can be too restrictive and, to
some degree, unfair when the best evaluation score is high. For example, when v = 10 and the
review used as the observed outcome is also equal to 10, a reported review equal to 9 seems to be
more accurate than a reported review equal to 1. One effective way to deal with these issues is by
using strictly proper scoring rules in (5) that are sensitive to distance.
Using the notation of Section 4.1, recall that z = (z0, . . . , zv) is some reported probability
distribution. Given that the outcomes are ordered, we denote the cumulative probabilities by capital
letter: Zk = Pj≤k zj. We first define the notion of distance between two probability vectors as
proposed by Stael von Holstein [1970]. We say that a probability vector z′ is more distant from
the jth outcome than a probability vector z 6= z′ if:
Z ′
Z ′
k ≥ Zk, for k = 0, . . . , j − 1
k ≤ Zk, for k = j, . . . , v
Intuitively, the above definition means that z can be obtained from z′ by successively moving
probability mass towards the jth outcome from other outcomes [Stael von Holstein, 1970]. A
scoring rule R is said to be sensitive to distance if R(z, j) > R(z′, j) whenever z′ is more distant
11
0.2
0
−0.2
−0.4
e
r
o
c
S
−0.6
−0.8
−1
−1.2
−1.4
−1.6
0
j = 0
j = 1
j = 2
j = 3
j = 4
1
2
Reported Review
3
4
Figure 2: Scores returned by R (E [ωα, ri] , j) for different reported reviews when v = 4, R is the
RPS rule, and α = (1, 1, 1, 1, 1). Each line represents a different value for j (observed outcome).
from z for all j. Epstein [1969] introduced the ranked probability score (RPS), a strictly proper
scoring rule that is sensitive to distance. Using the formulation of Epstein’s result proposed by
Murphy [1970], we have for a probability vector z and an observed outcome j ∈ {0, . . . , v}:
j−1
v
RP S(z, j) = −
Z 2
k −
(1 − Zk)2
(6)
Xk=0
Xk=j
Figure 2 illustrates the scores returned by (6) for different reported reviews and values for j
when reviewers’ prior distributions are non-informative. When using RPS as the strictly proper
scoring rule in (5), reviewers are rewarded based on how close their reported reviews are to the
reviews taken as observed outcomes. For example, when the review used as the observed outcome
is equal to 0 (see the dotted line with squares in Figure 2), the returned score monotonically
decreases as the reported review increases. Since RPS is strictly proper, Proposition 2 is still
valid for any hyperparameter α, i.e., each reviewer strictly maximizes his expected review score
by reporting honestly. The scoring range of RPS is [−v, 0]. Hence, review scores are always
non-negative when using γ = 1 and λ = v in (5).
4.5 Numerical Example
Consider four reviewers (n = 4) and the best evaluation score being equal to four (v = 4). Suppose
that reviewers have non-informative Dirichlet priors with α = (1, 1, 1, 1, 1), and that reviewers
1, 2, 3, and 4 report, respectively, r1 = 0, r2 = 0, r3 = 1, and r4 = 4. From (4), the resulting
estimated posterior predictive distributions are, respectively, E[ωα, r1 = 0] = (cid:0) 2
E[ωα, r2 = 0] = (cid:0) 2
(cid:0) 1
6(cid:1),
6(cid:1), and E[ωα, r4 = 4] =
6(cid:1). In what follows, we illustrate the scores returned by (5) when using a symmetric
6 , 1
6 , 1
6 , 2
6 , 2
6 , 1
6 , 1
6 , 1
and bounded strictly proper scoring rule and when using RPS.
6(cid:1), E[ωα, r3 = 1] = (cid:0) 1
6 , 1
6 , 1
6 , 1
6 , 1
6 , 1
6 , 1
6 , 1
6 , 1
6 , 1
12
4.5.1 Rewarding Agreements
Assume that R in (5) is the quadratic scoring rule shown in (3), which in turn is symmetric,
bounded, and strictly proper. Consequently, as discussed in Section 4.3, the term γR (E[ωα, ri], rj )+
λ in (5) can take on only two values:
γ(cid:18) 4
γ(cid:18) 2
v+2(cid:17)2
v+2 −(cid:16) 2
v+2(cid:17)2
v+2 −(cid:16) 2
−Pv−1
−Pv−1
e=0(cid:16) 1
e=0(cid:16) 1
v+2(cid:17)2(cid:19) + λ if ri = rj,
v+2(cid:17)2(cid:19) + λ otherwise.
1
δmax−δmin
Hence, by setting γ =
2v+4 , the above values are
equal to, respectively, 1 and 0. Using the scoring method in (5), we obtain the following review
scores: s1 = s2 = 1 and s3 = s4 = 0. That is, the review scores received by reviewers 1 and 2
are similar due to the fact that r1 = r2. Reviewer 3 and 4’s review scores are equal to 0 because
there is no match between their reported reviews and others’ reported reviews.
and λ = −δmin
= v+2
2
= −v
δmax−δmin
4.5.2 Taking Distance into Account
Now, assume that R in (5) is the RPS rule shown in (6). In order to ensure non-negative review
scores, let γ = 1 and λ = v = 4 . Using the scoring method in (5), we obtain the following review
scores: s1 = s2 = (−0.8333γ + λ)+ (−0.5γ + λ)+ (−1.5γ + λ) = 9.1667, s3 = 2·(−1.0833γ +
λ) + (−1.4167γ + λ) = 8.4167, and s4 = 2 · (−1.5γ + λ) + (−0.8333γ + λ) = 8.1667. The
review score of reviewer 4 is the lowest because his reported review is the most different review,
i.e., it has the largest distance between it and all of the other reviews.
5 Multiple Criteria
In our basic model, reviewers observe only one signal from the distribution that represents the
quality of the manuscript. However, manuscripts are often evaluated under multiple criteria, e.g.,
relevance, clarity, originality, etc., meaning that in practice reviewers might observe multiple sig-
nals and report multiple evaluation scores. Under the assumption that these signals are indepen-
dent, each reported evaluation score can be scored individually using the same scoring method
proposed in the previous section. Clearly, Proposition 2 is still valid, i.e., honest reporting still
strictly maximizes reviewers’ expected review scores.
The lack of relationship between different criteria is not always a reasonable assumption. A
modeling choice that takes the relationship between observed signals into account, which is also
consistent with our basic model, is to assume that the quality of the manuscript is still represented
by a multinomial distribution, but now reviewers may observe several signals from that distribu-
tion. Formally, let ρ ∈ N+ be the number of draws from the distribution that represents the quality
of the manuscript, where each signal represents an evaluation score related to a criterion. Instead
of a single number, each reviewer i’s private information is now a vector: ti = (ti,1, . . . , ti,ρ),
where ti,k ∈ {0, . . . , v}, for k ∈ {1, . . . , ρ}. The basic assumptions (autonomy, risk neutrality,
Dirichlet priors, and rationality) are still the same. For ease of exposition, we denote reviewer i’s
(ρ)
true posterior predictive distribution in this section by Θ
i = E [ωα, ti]. Under this new model,
each reviewer i’s posterior predictive distribution is now defined as:
Θ
(ρ)
i =(cid:18) α0 +Pρ
ρ +Pv
k=1 H(0, ti,k)
x=0 αx
k=1 H(1, ti,k)
, . . . ,
x=0 αx
,
α1 +Pρ
ρ +Pv
13
k=1 H(v, ti,k)
x=0 αx
αv +Pρ
ρ +Pv
(7)
(cid:19)
where H(x, y) is an indicator function:
H(x, y) =(cid:26) 1 if x = y,
0 otherwise.
Assuming that each reported review is a vector of ρ evaluations scores, i.e., ri = (ri,1, . . . , ri,ρ),
where ri,k ∈ {0, . . . , v}, for i ∈ N and k ∈ {1, . . . , ρ}, the center estimates each reviewer i’s pos-
terior predictive distribution by applying Bayes’ rule to the common prior. The resulting estimated
posterior predictive distribution E [ωα, ri], referred to as Φ
for ease of exposition, is:
(ρ)
i
Thereafter, the center rewards Φ
by using a strictly proper scoring rule R and other review-
k=1 H(1, ri,k)
, . . . ,
x=0 αx
αv +Pρ
ρ +Pv
k=1 H(v, ri,k)
x=0 αx
(cid:19) (8)
Φ
(ρ)
i =(cid:18) α0 +Pρ
ρ +Pv
k=1 H(0, ri,k)
x=0 αx
,
α1 +Pρ
ρ +Pv
(ρ)
i
ers’ reported reviews as observed outcomes:
si =Xj6=i(cid:16)γR(cid:16)Φ
(ρ)
i
, G(rj )(cid:17) + λ(cid:17)
(9)
where G is some function used by the center to summarize each reviewer j’s reported review
in a single number, and whose image is equal to the set {0, . . . , v}. For example, G can be a
function that returns the median or the mode of the reported evaluation scores. Honest reporting,
i.e., ri = ti, maximizes reviewers’ expected review scores under this setting.
Proposition 3. When observing and reporting multiple signals, each reviewer i ∈ N maximizes
his expected review score when ri = ti.
Proof. Due to the autonomy assumption, we restrict ourselves to show that each reviewer i ∈ N
maximizes E
proper scoring rule, from Proposition 1 we have that:
, G(rj )(cid:17) + λi, for j 6= i, when ri = ti. Given that R is a strictly
i hγR(cid:16)Φ
(ρ)
i
(ρ)
Θ
arg max
E
Φ
(ρ)
i
Θ
(ρ)
i hγR(cid:16)Φ
(ρ)
i
, G(rj )(cid:17) + λi = Θ
(ρ)
i
When ri = ti, we have by construction that Φ
(ρ)
(ρ)
i = Θ
i
(see equations (7) and (8)). Thus,
reviewers’ expected review scores are maximized when reviewers report honestly.
When observing and reporting multiple evaluation scores, a reviewer can weakly maximize
his expected review score by reporting a review different than his true review as long as the es-
timated posterior predictive distributions are the same. For example, when reviewer i reports
ri = (1, 2, 3), the resulting estimated posterior predictive distribution is the same as when he re-
ports ri = (3, 1, 2), and, consequently, reviewer i receives the same review score in both cases.
This implies that the scoring method in (9) is more suitable to a peer-review process where all
criteria are equally weighted since honest reporting weakly maximizes expected review scores.
5.1 Summarizing Signals when Prior Distributions are Non-Informative
When reviewers report multiple evaluation scores, the intuitive interpretation of review scores as
rewards for agreements that arises when using symmetric and bounded strictly proper scoring rules
(see Section 4.3) is lost because the elements of the estimated posterior predictive distribution in
(8) can take on more than two different values.
14
A different approach that preserves the aforementioned intuitive interpretation when reviewers’
prior distributions are non-informative is to ask the reviewers to summarize their observed signals
into a single value before reporting it, instead of the center doing it on their behalf. Hence, each
reviewer i is now reporting honestly when ri = G(ti), where G is some function suggested by
the center whose image is equal to the set {0, . . . , v}. This new model can be interpreted as if
the reviewers were reviewing the manuscript under several criteria and reporting the manuscript’s
overall evaluation score by reporting the value G(ti).
Since reviewers are reporting only one value, we can use the original scoring method in
(5) to promote honest reporting. We prove below that for any symmetric and bounded strictly
proper scoring rule, honest reporting strictly maximizes reviewers’ expected review scores un-
der the scoring method in (5) if and only if G is the mode of the observed signals, i.e., when
k=1 H(x, ti,k). Ties between observed signals are broken randomly.
Proposition 4. When observing multiple signals and reporting ri = G(ti), each reviewer i ∈
N with non-informative prior strictly maximizes his expected review score under the scoring
method in (5), for a symmetric and bounded strictly proper scoring rule R, if and only if ri =
G(ti) = arg maxx∈{0,...,v}Pρ
arg maxx∈{0,...,v}Pρ
k=1 H(x, ti,k).
Proof. Recall that since each reviewer i ∈ N observes multiple signals, his true posterior predic-
(ρ)
tive distribution is equal to E [ωα, ti] = Θ
i = (θ0, θ1, . . . , θv) as shown in (7). Due to Proposi-
tion 1 and since reviewers cannot affect their peers’ reviews because of the autonomy assumption,
we restrict ourselves to show that each reviewer i ∈ N maximizes E
(1)
i = E [ωα, ri = G(ti)].
Let z ∈ {0, . . . , v} be the most common signal observed by reviewer i. Hence, reviewer i’s
subjective probability associated with z is greater than his subjective probability associated with
any other signal y ∈ {0, . . . , v}, i.e., θi,z > θi,y.
j 6= i, if and only if ri = arg maxx∈{0,...,v}Pρ
k=1 H(x, ti,k), where Φ
, rj(cid:17)i, for
i hR(cid:16)Φ
(1)
i
(ρ)
Θ
(If part) Given that R is a symmetric and strictly proper scoring rule and that each reviewer
(1)
i
two possible values: δmax, if ri = rj, and δmin otherwise (see discussion in Section 4.3). When
k=1 H(x, ti,k) = z, reviewer i’s expected review score is
, rj(cid:17) can take on only
is reporting only one evaluation score, the resulting score from R(cid:16)Φ
reporting ri = arg maxx∈{0,...,v}Pρ
θi,zδmax +Py6=z θi,yδmin. Given that θi,z > θi,y, for any y 6= z, and δmax > δmin, this ex-
pected review score is maximized. Thus, reporting ri = arg maxx∈{0,...,v}Pρ
(Only-if part) Recall that all the elements making up the hyperparameter α have the same
(1)
i = (φi,0, . . . , φi,v) be
value because reviewers’ prior distributions are non-informative. Let Φ
reviewer i’s estimated posterior predictive distribution computed according to the original scoring
maximizes reviewer i’s expected review score.
k=1 H(x, ti,k) = z
method in (5) when reviewer i is reporting ri = arg maxx∈{0,...,v}Pρ
αk+1
k=1 H(x, ti,k) = z, i.e.:
φi,k =(
1+Pv
x=0 αx
αk
1+Pv
x=0 αx
= αk+1
αk
=
(v+1)·αk +1
(v+1)·αk +1
if k = z,
otherwise.
For contradiction’s sake, suppose that reviewer i maximizes his expected review score by
i = ( φi,0, . . . , φi,v) be reviewer i’s
misreporting his review and reporting ri = y 6= z. Let Φ
estimated posterior predictive distribution when he is misreporting his review, i.e.:
(1)
φi,k =(
αk+1
1+Pv
x=0 αx
αk
1+Pv
x=0 αx
= αk+1
αk
=
(v+1)·αk +1
(v+1)·αk +1
if k = y,
otherwise.
15
As discussed in Section 4.3, the term R(cid:16)Φ
ever R is a symmetric scoring rule. Consequently, R(cid:16) Φ
consequence of our assumption that reviewer i maximizes his expected review score by misre-
porting his review is that E
symmetric and bounded proper scoring rule, this inequality becomes:
, rj(cid:17)i ≥ E
i hR(cid:16) Φ
, rj(cid:17) can take on only two possible values when-
, k(cid:17) for k 6= z, y. A
, rj(cid:17)i. Assuming that R is a
, k(cid:17) = R(cid:16)Φ
i hR(cid:16)Φ
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(ρ)
(ρ)
Θ
Θ
v
v
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
(1)
i
=⇒
=⇒
Xk=0
, k(cid:17)
θi,zR(cid:16) Φ
θi,kR(cid:16)Φ
θi,kR(cid:16) Φ
Xk=0
, z(cid:17) + θi,yR(cid:16) Φ
, k(cid:17) ≥
, y(cid:17) ≥ θi,zR(cid:16)Φ
R(cid:16)Φ
θi,y ≥ θi,z
R(cid:16) Φ
The second line follows from the fact that R(cid:16) Φ
, z(cid:17) + θi,yR(cid:16)Φ
, z(cid:17) − R(cid:16) Φ
, y(cid:17) − R(cid:16)Φ
, k(cid:17) = R(cid:16)Φ
, y(cid:17)
, z(cid:17)
, y(cid:17)
, k(cid:17) for k 6= z, y. Re-
, y(cid:17) = δmax, and
, y(cid:17) = δmin. Consequently, we obtain that θi,y ≥ θi,z. As we stated
, rj(cid:17)i, i.e., reviewer i maximizes his
i hR(cid:16) Φ
, z(cid:17) = R(cid:16) Φ
garding the last line, we have by construction that R(cid:16)Φ
R(cid:16) Φ
, z(cid:17) = R(cid:16)Φ
before, since z is the most common signal observed by reviewer i, then θi,z > θi,y. Thus, we have
a contradiction. So, E
, rj(cid:17)i < E
i hR(cid:16)Φ
expected review score only if he reports ri = arg maxx∈{0,...,v}Pρ
In other words, the above proposition says that each reviewer should report the evaluation score
most likely to be deserved by the manuscript when their prior distributions are non-informative and
they are rewarded according to the scoring method in (5). Any other evaluation score has a lower
associated subjective probability and, consequently, reporting it results in a lower expected review
score. To summarize, Proposition 4 implies that the scoring method proposed in (5) induces honest
reporting by rewarding agreements whenever reviewers’ prior distributions are non-informative
and the center is interested in the mode of each reviewer’s observed signals. It is noteworthy that
Proposition 4 does not assume that the center knows a priori the number of observed signals ρ,
thus providing more flexibility for practical applications of our method.
k=1 H(x, ti,k) = z.
(1)
i
(1)
i
(ρ)
Θ
(1)
i
(1)
i
(1)
i
(ρ)
Θ
6 Finding a Consensual Review
After reviewers report their reviews and receive their review scores, there is still the question of
how the center will use the reported reviews in making a suitable decision. Since reviewers are not
always in agreement, belief aggregation methods must be used to combine the reported reviews
into a single representative review. The traditional average method is not necessarily the best
approach since unreliable reviewers might have a big impact on the aggregate review. Moreover,
a consensual review is desirable because it represents a review that is acceptable by all.
In this section, we propose an adaptation of a classical mathematical method to find a consen-
sual review. Intuitively, it works as if reviewers were constantly updating their reviews in order
to aggregate knowledge from others. The scoring concepts introduced in previous sections are
incorporated by the reviewers when updating their reviews. In what follows, for the sake of gener-
ality, we assume that reviewers evaluate the manuscript under ρ ∈ N+ criteria, i.e., each reviewer
i observes ρ signals from the underlying distribution that represents the quality of the manuscript
16
and report a vector ri = (ri,1, . . . , ri,ρ) of evaluation scores, where ri,k ∈ {0, . . . , v} for all k.
The center then estimates reviewers i’s posterior predictive distribution E [ωα, ri], referred to as
for ease of exposition, as in (8). We relax our basic model by allowing the evaluation scores
Φ
in the aggregate review to take on any real value between 0 and the best evaluation score v.
(ρ)
i
6.1 DeGroot’s Model
DeGroot [1974] proposed a model that describes how a group might reach a consensus by pooling
their individual opinions. When applying this model to a peer-review setting, each reviewer i is
first informed of others’ reported reviews. In order to accommodate the information and expertise
of the rest of the group, reviewer i then updates his own review as follows:
n
r
(1)
i =
wi,jrj
Xj=1
this update. It is assumed that wi,j ≥ 0, for every reviewer i and j, and Pn
where wi,j is a weight that reviewer i assigns to reviewer j’s reported review when he carries out
j=1 wi,j = 1. In this
way, each updated review takes the form of a linear combination of reported reviews, also known
as a linear opinion pool. The weights must be chosen on the basis of the relative importance that
reviewers assign to their peers’ reviews. The whole updating process can be written in a more
general form using matrix notation: R(1) = WR(0), where:
W =
w1,1 w1,2
w2,1 w2,2
...
...
wn,1 wn,2
· · · w1,n
· · · w2,n
...
.. .
· · · wn,n
and R(0) =
r1
r2
...
rn
=
r1,1
r2,1
...
rn,1
r1,2
r2,2
...
rn,2
· · ·
· · ·
. . .
· · ·
r1,ρ
r2,ρ
...
rn,ρ
Since all the original reviews have changed, the reviewers might wish to update their new
reviews in the same way as they did before. If there is no basis for the reviewers to change their
assigned weights, the whole updating process after t revisions can then be represented as follows:
R(t) = WR(t−1) = WtR(0)
(10)
Let r
i,1, . . . , r(t)
(t)
i = (cid:16)r(t)
i,ρ(cid:17) be reviewer i’s review after t revisions, i.e., it denotes the ith row
(t)
j , for every reviewer
(t)
i = r
of the matrix R(t). We say that a consensus is reached if and only if r
i and j, when t → ∞.
6.2 Review Scores as Weights
The original method proposed by DeGroot [1974] does not encourage honesty in a sense that
reviewers can assign weights to their peers’ reviews however they wish so as long as the weights
are consistent with the construction previously defined. Furthermore, it requires the disclosure of
reported reviews to the whole group when reviewers are weighting others’ reviews, a fact which
might be troublesome when the reviews are of a sensitive nature.
A possible way to circumvent the aforementioned problems is to derive weights from the
original reported reviews by taking into account review scores. In particular, we assume the weight
that a reviewer assigns to a peer’s review is directly related to how close their estimated posterior
predictive distributions are, where closeness is defined by an underlying proper scoring rule. We
provide behavioral foundations for such an assumption in the following subsection. Formally, the
weight that reviewer i assigns to reviewer j’s reported review is computed as follows:
17
wi,j =
(11)
E
Φ
j=1
Pn
(ρ)
(ρ)
j
i hγR(cid:16)Φ
i hγR(cid:16)Φ
, e(cid:17) + λi
, e(cid:17) + λi
(ρ)
j
(ρ)
Φ
E
E
that is, the weight wi,j is proportional to the expected review score that reviewer i would receive if
he had reported the review reported by reviewer j, where the expectation is taken with respect to re-
viewer i’s estimated posterior predictive distribution. Consequently, the weight that each reviewer
indirectly assigns to his own review is always the highest because R is a strictly proper scoring rule,
i.e., arg max
0, for every reviewer i and j. As long as R is bounded, this assumption can be met by appropriately
j=1 wi,j = 1
i hγR(cid:16)Φ
setting the value of λ. Consequently, 0 < wi,j < 1, for every i, j ∈ N. Moreover,Pn
because the denominator of the fraction in (11) normalizes the weights so they sum to one.
In the interest of reaching a consensus, DeGroot’s method in (10) is applied to the original
reported reviews using the weights as defined in (11). We show that a consensus is always reached
under this proposed method whenever the review scores are positive.
, e(cid:17) + λi = Φ
i hγR(cid:16)Φ
, e(cid:17) + λi >
. We assume that E
(ρ)
j
(ρ)
i
(ρ)
j
(ρ)
j
(ρ)
(ρ)
Φ
Φ
Φ
(t)
i = r
(t)
j
Proposition 5. If E
when t → ∞.
Φ
(ρ)
i hγR(cid:16)Φ
(ρ)
j
, e(cid:17) + λi > 0, for every reviewer i, j ∈ N , then r
i hγR(cid:16)Φ
(ρ)
j
Φ
(ρ)
, e(cid:17) + λi > 0, for every reviewer i and j,
Proof. Due to the assumption that E
all the elements of the matrix W in (10) are strictly greater than zero and strictly less than one.
Moreover, the sum of the elements in any row is equal to one. Consequently, W can be regarded
as a n × n stochastic matrix, or a one-step transition probability matrix of a Markov chain with
n states and stationary probabilities. Furthermore, the underlying Markov chain is aperiodic and
irreducible. Therefore, a standard limit theorem of Markov chains applies in this setting, namely
given an aperiodic and irreducible Markov chain with transition probability matrix W, every row
of the matrix Wt converges to the same probability vector when t → ∞ [Ross, 1995].
(t)
(0)
j
(t−1)
j
(t−2)
k
j=1 βj r
j=1 wi,jr
k=1 wj,kr
Recall that r
= Pn
i = Pn
j=1 wi,jPn
= · · · = Pn
,
where β = (β1, β2, . . . , βn) is a probability vector that incorporates all the previous weights. This
equality implies that the consensual review can be represented as an instance of the linear opinion
pool. Hence, an interpretation of the proposed method is that reviewers reach a consensus regard-
ing the weights in (10). When β = (1/n, 1/n, . . . , 1/n) in the above equality, the underlying
linear opinion pool becomes the average of the reported evaluation scores. A drawback with an
averaging approach is that it does not take into account the scoring concepts introduced in the pre-
vious sections, a fact which might favor unreliable reviewers. Moreover, disparate reviews might
have a big impact on the resulting aggregate review. On the other hand, under our approach to find
β, reviewers weight down reviews far from their own reviews, which implies that the proposed
method might be less influenced by disparate reviews. A numerical example in subsection 6.4
illustrates this point. The experimental results discussed in Section 7 show that our method to find
a consensual review is consistently more accurate than the traditional average method.
6.3 Behavioral Foundation
The major assumption regarding our method for finding a consensual review is that reviewers
assign weights according to (11). An interesting interpretation of (11) arises when the proper
scoring rule R is effective with respect to a metric M. Formally, given a metric M that assigns a
real number to any pair of probability vectors, which can be seen as the shortest distance between
18
the two probability vectors, we say that a scoring rule R is effective with respect to M if the
following relation holds for all probability vectors Φ
[Friedman, 1983]:
, and Φ
, Φ
(ρ)
i
(ρ)
j
(ρ)
k
M(cid:16)Φ
(ρ)
i
, Φ
(ρ)
j (cid:17) < M(cid:16)Φ
(ρ)
i
, Φ
(ρ)
k (cid:17) ⇐⇒ E
Φ
(ρ)
i hγR(cid:16)Φ
(ρ)
j
, e(cid:17) + λi > E
Φ
(ρ)
i hγR(cid:16)Φ
(ρ)
k , e(cid:17) + λi
Thus, when R is effective with respect to a metric M, the higher the weight one reviewer
assigns to a peer’s review in (11), the closer their estimated posterior predictive distributions are
according to the metric M. In other words, when using effective scoring rules, reviewers naturally
prefer reviews close to their own reported reviews, and the weight that each reviewer assigns to
his own review is always the highest one. Hence, in spirit, the resulting learning model in (10) can
be seen as a model of anchoring [Tversky and Kahneman, 1974] in a sense that the review of a
reviewer is an “anchor”, and subsequent updates are biased towards reviews close to the anchor.
Friedman [1983] discussed some examples of effective scoring rules. For example, the quadratic
scoring rule in (3) is effective with respect to the root-mean-square deviation, the spherical scoring
rule is effective with respect to a renormalized L2-metric, whereas the logarithmic scoring rule is
not effective with respect to any metric [Nau, 1985].
6.4 Numerical Example
Consider a peer-review process where the best evaluation score is four (v = 4), three reviewers
(n = 3) observe three signals (ρ = 3), and they report the following reviews: r1 = (0, 1, 3),
r2 = (0, 2, 3), and r3 = (4, 4, 4). Consequently, the matrix R(0) in (10) is:
R(0) =
0 1 3
0 2 3
4 4 4
Consider the hyperparameter α = (1, 1, 1, 1, 1). Hence, the estimated posterior predictive
(3)
distributions are Φ
3 =
(1/8, 1/8, 1/8, 1/8, 4/8). Assume that R in (11) is the quadratic scoring rule in (3), and let γ = 1
and λ = 1 in order for the resulting expected values in (11) to be always positive. We obtain:
(3)
2 = (2/8, 1/8, 2/8, 2/8, 1/8), and Φ
(3)
1 = (2/8, 2/8, 1/8, 2/8, 1/8), Φ
W =
0.3545 0.3455 0.3000
0.3455 0.3545 0.3000
0.3158 0.3158 0.3684
Focusing on the main diagonal of W, we notice that each reviewer always assigns the highest
weight to his own review. From the first row of W, we can see that reviewer 1 assigns a high
weight to his review and to reviewer 2’s review, and a lower weight to reviewer 3’s review. This
happens because reviewer 3’s review is very distant from the others’ reported reviews. We can
draw similar conclusions from the other rows. We then obtain the following weights when carrying
out DeGroot’s method with the weights calculated according to (11):
lim
t→∞
Wt =
0.3390 0.3390 0.3220
0.3390 0.3390 0.3220
0.3390 0.3390 0.3220
and, consequently, the consensual review is represented by any row of the matrix limt→∞ WtR(0),
which results in the vector (1.288, 2.305, 3.322). It is worthwhile to discuss an interesting point
regarding the above example. The aggregate review would be (1.333, 2.333, 3.333) if it was equal
19
to the average of the reported reviews. Hence, reviewer 3’s review would have more impact on
the aggregate review because the evaluation scores in the average review are all greater than the
corresponding evaluation scores in the consensual review. In our proposed method, the influence
of reviewer 3 on the aggregate review is diluted because his review is very different from the
others’ reviews. More formally, reviewer 3’s estimated posterior predictive distribution is very
distant from the others’ estimated posterior predictive distributions when measured according to
the root-mean-square deviation, the metric associated with the quadratic scoring rule.
7 Experiments
In this section, we describe a peer-review experiment designed to test the efficacy of both the
proposed scoring method and the proposed aggregation method.
In the following subsections,
we discuss Amazon’s Mechanical Turk, the platform used in our experiments, the experimental
design, and our results.
7.1 Amazon’s Mechanical Turk
Amazon’s Mechanical Turk5 (AMT) is an online labor market originally developed for human
computation tasks, i.e., tasks that are relatively easy for human beings, but nonetheless challeng-
ing or even currently impossible for computers, e.g., audio transcription, filtering adult content,
extracting data from images, etc. Several studies have shown that AMT can effectively be used as
a means of collecting valid data in these settings [Snow et al., 2008, Marge et al., 2010].
More recently, AMT has also been used as a platform for conducting behavioral experiments
[Mason and Suri, 2012]. One of the advantages that it offers to researchers is the access to a large,
diverse, and stable pool of people willing to participate in the experiments for relatively low pay,
thus simplifying the recruitment process and allowing a faster iteration between developing theory
and executing experiments. Furthermore, AMT provides an easy-to-use built-in mechanism to pay
workers that greatly reduces the difficulties of compensating individuals for their participation in
the experiments, and a built-in reputation system that helps requesters distinguish between good
and bad workers and, consequently, to ensure data quality. Numerous studies have shown that
results of behavioral studies conducted on AMT are comparable to results obtained in other online
domains as well as in offline settings [Buhrmester et al., 2011, Horton et al., 2011], thus providing
evidence that AMT is a valid means of collecting behavioral data.
7.2 Experimental Design
We designed a task on AMT that required workers, henceforth referred to as reviewers, to review 3
short texts under three different criteria: Grammar, Clarity, and Relevance. The first two texts were
extracts from published poems, but with some original words intentionally replaced by misspelled
words. The third text contained random words presented in a semi-structured way. All the details
regarding the texts are included in the appendix. For each text, three questions were presented to
the reviewers, each one having three possible responses ordered in decreasing negativity order:
• Grammar: does the text contain misspellings, syntax errors, etc.?
– A lot of grammar mistakes
– A few grammar mistakes
– No grammar mistakes
5http://www.mturk.com
20
• Clarity: does the text, as a whole, make any sense?
– The text does not make sense
– The text makes some sense
– The text makes perfect sense
• Relevance: could the text be part of a poem related to love?
– The text cannot be part of a love poem
– The text might be part of a love poem
– The text is definitely part of a love poem
Words with subjective meaning were intentionally used so as to simulate the subjective nature
of the evaluation scores in a review, e.g., “a lot”, “a few”, etc. Each individual response was
translated into an evaluation score inside the set {0, 1, 2}. The most negative response received
the score 0, the middle response received the score 1, and the most positive response received the
score 2. Thus, each reviewer reported a vector of 9 evaluation scores (3 texts times 3 criteria).
We recruited 150 reviewers on AMT, all of them residing in the United States of America
and older than 18 years old. They were required to accomplish the task in at most 20 minutes.
Reviewers were split into 3 groups of equal size. After accomplishing the task, every reviewer in
every group received a payment of 20 cents. A study done by Ipeirotis [2010] showed that more
than 90% of the tasks on AMT have a baseline payment less than $0.10, and 70% of the tasks have
a baseline payment less than $0.05. Thus, our baseline payment was much higher than the average
payment from other jobs posted to the AMT marketplace.
Each reviewer was randomly assigned into one of the three groups. Reviewers in two of the
groups, the treatment groups, could earn an additional bonus of up to 10 cents. Reviewers in the
first treatment group, referred to as the Bonus Group (BG), were informed that their bonuses would
be proportional to the number of reviews similar to their reported reviews. Reviewers in the second
treatment group, the Bonus and Information Group (BIG), received similar information, but they
also received a short summary of some theoretical results presented in this paper:
“A group of researchers from the University of Waterloo (Canada) formally showed
that the best strategy to maximize your expected bonus in this setting is by being
honest, i.e., by considering each question thoroughly and deciding the best answers
according to your personal opinion”.
Members of the third group, the Control Group (CG), neither received extra explanations nor
bonuses. Their reported reviews were used as the control condition. Bonuses were computed by
rewarding agreements as described in Section 4.3. Due to the one-shot nature of this peer-review
task, we assumed that Dirichlet priors were non-informative with hyperparameter α = (1, 1, 1).
For each reported evaluation score, there could be at most 49 similar reported evaluation scores
because each group had 50 members. We then used the formula 10
to calculate the
reward for an individual evaluation score. Given that each reviewer reported 9 evaluation scores,
if the evaluation scores reported by all members of a group were the same, then all group members
would received the maximum bonus of 10 cents. The provided bonuses can be seen as review
scores. Our primary objective when performing this experiment was to empirically investigate the
extent to which providing review scores affects the quality of the reported reviews.
9 × #agreements
49
21
7.3 Gold-Standard Evaluation Scores
Since the source and original content of each text were known a priori, i.e., before the experiments
were conducted, we were able to derive gold-standard reviews for each text. In order to avoid
confirmation bias6, we asked five professors and tutors from the English and Literature Department
at the University of Waterloo to provide their reviews for each text. We set the gold-standard
evaluation score for each criterion in a text as the median of the evaluation scores reported by
the professors and tutors. Coincidentally, each median value was also the mode of the underlying
evaluation scores. All the evaluation scores reported by the professors and tutors as well as the
respective gold-standard evaluation scores are in the appendix.
7.4 Hypotheses
Our first research question was whether or not providing review scores through pairwise compar-
isons makes the reported reviews more accurate, i.e., closer to the gold-standard reviews. Based
on our theoretical results, our hypothesis was:
Hypothesis 1. The average accuracy of group BIG is greater than the average accuracy of group
BG, which in turn is greater than the average accuracy of group CG.
In other words, the resulting reviews would be on average more accurate when reviewers re-
ceived review scores, and the extra explanation regrading the theory behind the scoring method
would provide more credibility to it, thus making the reviews more accurate. Regarding the result-
ing bonuses, since honest reporting maximizes reviewers’ expected review scores in our model,
our second hypothesis was:
Hypothesis 2. The average bonus received by members of group BIG is greater than the average
bonus received by members of group BG, which in turn is greater than the average bonus received
by members of group CG.
In order to test whether or not Hypothesis 2 was true, we used the bonus the members of group
CG would have received had they received any bonus. It is important to note that Hypothesis 1
was measured by comparing how close the reported reviews were to the gold-standard reviews,
whereas Hypothesis 2 was measured by making pairwise comparisons between reported reviews:
the higher the number of agreements, the greater the resulting bonus.
Another metric used to compare groups’ performance was the task completion time. The
amount of time spent by reviewers on the reviewing task can be seen as a proxy for the effort
they exerted to complete the task. Regarding this metric, we expected reviewers who received
review scores to be more cautious when completing their tasks. Moreover, the extra explanation
regrading the theory behind the scoring method would provide more credibility to it, thus making
the members of group BIG work harder on the task. Hence, our third hypothesis was:
Hypothesis 3. The average task completion time of group BIG is greater than the average task
completion time of group BG, which in turn is greater than the average task completion time of
group CG.
Finally, we believed that the consensual review, computed as described in Section 6, would be
more accurate than the average review since disparate reviews are less likely to have a big influence
on the consensual review than on the average review. Hence, our fourth hypothesis was:
Hypothesis 4. The average accuracy of the consensual review is greater than the average accu-
racy of the average review.
6The tendency to interpret information in a way that confirms one’s preconceptions [Plous, 1993].
22
Table 1: Accuracy of each group on individual criteria. The average of the absolute difference
between the reported evaluation scores and the corresponding gold-standard evaluation scores
is shown below each group. For each criterion, the lowest average is highlighted in bold. The
standard deviations are in parenthesis. One-tailed p-values resulting from rank-sum tests are
shown in the last three columns. Given the notation A-B, the null hypothesis is that the outcome
measures resulting from groups A and B are equivalent, and the alternative hypothesis is that the
outcome measure resulting from group A is less than the outcome measure resulting from group B.
BG
BIG
CG
BIG-BG BIG-CG BG-CG
p-values
0.5000
(0.5051)
0.8200
(0.6606)
0.2200
(0.5067)
0.4400
(0.5014)
0.5000
(0.6468)
0.4400
(0.5014)
0.7600
(0.8466)
0.1400
(0.4046)
0.1200
(0.4352)
0.3200
(0.4712)
0.6200
(0.6024)
0.2000
(0.4518)
0.3600
(0.4849)
0.3800
(0.6024)
0.6400
(0.4849)
0.7800
(0.8640)
0.0000
(0.0000)
0.1000
(0.3642)
0.4400
(0.5014)
0.8600
(0.7287)
0.3000
(0.5803)
0.3800
(0.4903)
0.5400
(0.6131)
0.6600
(0.4785)
1.0200
(0.8449)
0.1600
(0.3703)
0.2000
(0.4949)
0.035**
0.110
0.726
0.065*
0.052*
0.413
0.484
0.213
0.230
0.209
0.420
0.729
0.155
0.067*
0.325
0.977
0.419
0.014**
0.539
0.077*
0.061*
0.006**
0.002**
0.301
0.491
0.112
0.122
Grammar
Text 1
Clarity
Relevance
Grammar
Text 2
Clarity
Relevance
Grammar
Text 3
Clarity
Relevance
* p ≤ 0.1
** p ≤ 0.05
7.5 Experimental Results
7.5.1 Accuracy on Individual Criteria
In our first analysis, we computed the absolute difference between each reported evaluation score
and the corresponding gold-standard evaluation score. Thus, the outcome measure was an integer
with a value between zero and two, and the closer this value was to zero, the better the resulting
accuracy. Table 1 shows the average accuracy of each group on individual criteria.
Focusing first on the groups BG and BIG, the group BIG is the most accurate group on all
criteria, except for the criterion Relevance in Text 2 and the criterion Grammar in Text 3. This
result is statistically significant with p-value ≤ 0.1 in three out of the seven cases in which BIG is
more accurate than BG. In two out of these three statistically significant cases, this result is also
statistically significant with p-value ≤ 0.05. BG is more accurate than BIG in only two criteria.
This result is only statistically significant for the criterion Relevance in Text 2 (p-value ≤ 0.05).
The group CG, the control condition that involved no incentives beyond the baseline compen-
sation offered for completing the task, never outperforms both BG and BIG at the same time, and
23
Table 2: Aggregate accuracy of each group. The average of the sum of the absolute difference
between the reported evaluation scores and the corresponding gold-standard evaluation scores is
shown below each group. For each text and for the whole task, the lowest average is highlighted
in bold. The standard deviations are in parenthesis. One-tailed p-values resulting from rank-sum
tests are given in the last three columns. Given the notation A-B, the null hypothesis is that the
outcome measures resulting from groups A and B are equivalent, and the alternative hypothesis is
that the outcome measure resulting from group A is less than the outcome measure resulting from
group B.
p-values
BG
BIG
CG
BIG-BG BIG-CG BG-CG
1.5400
(1.1287)
1.3800
(1.0669)
1.0200
(1.1865)
3.9400
(2.2352)
1.1400
(1.0304)
1.3800
(0.9666)
0.8800
(0.9179)
3.4000
(1.6903)
1.6000
(1.4142)
1.5800
(0.9916)
1.3800
(1.1933)
4.5600
(2.1301)
0.043**
0.085*
0.588
0.547
0.163
0.148
0.394
0.020**
0.052*
0.110
0.002**
0.064*
Text 1
Text 2
Text 3
Overall
* p ≤ 0.1
** p ≤ 0.05
it is the less accurate group in seven out of nine criteria. In two (respectively, four) occasions, CG
is statistically significantly less accurate than BG (respectively BIG) with p-value ≤ 0.1.
Giving these results, we conclude that Hypothesis 1 is true for individual criteria, i.e., the
resulting reviews are on average more accurate when using review scores, and the extra explana-
tion regrading the theory behind the scoring method seems to provide more credibility to it, thus
improving the accuracy of the reported reviews.
7.5.2 Aggregate Accuracy
We also computed the aggregate accuracy of each group for each text as well as for the whole
task. In the former case, the outcome measure was the sum of the absolute difference between
each reported evaluation score for a given text and the corresponding gold-standard evaluation
score. For example, given (0, 1, 2) as the reported evaluation scores for Text 1, and (1, 2, 2) as the
corresponding gold-standard evaluation scores, the outcome measure for Text 1 would be 0 − 1 +
1 − 2 + 2 − 2 = 2. For the whole task, we summed the absolute differences across all criteria
and texts. Table 2 shows the aggregate accuracy of each group.
For every single text as well as for the overall task, members of the group CG report less
accurate reviews than members of the group BG and the group BIG. For the group BG, this result
is statistically significant for Text 3 and for the overall task (p-value ≤ 0.1). For the group BIG,
this result is statistically significant for Text 1 (p-value ≤ 0.1), Text 3 (p-value ≤ 0.05), and for the
whole task (p-value ≤ 0.05). Thus, the experimental results suggest that providing review scores
produces a significant improvement in quality over the control condition. Moreover, providing
an extra explanation about the theory behind the scoring method improves the final quality of
the reviews because, on average, the reviews from group BIG are more accurate than the reviews
from group BG. This result is statistically significant for Text 1 (p-value ≤ 0.05). Therefore, we
conclude that Hypothesis 1 is also true on the aggregate level.
24
Table 3: Average bonus and completion time per group. The highest average values are high-
lighted in bold. The standard deviations are in parenthesis. One-tailed p-values resulting from
rank-sum tests are given in the last three columns. Given the notation A-B, the null hypothesis is
that the outcome measures resulting from groups A and B are equivalent, and the alternative hy-
pothesis is that the outcome measure resulting from group A is greater than the outcome measure
resulting from group B.
p-values
BG
0.053
(0.0086)
178.66
BIG
0.058
(0.0073)
215.90
CG
BIG-BG
BIG-CG
BG-CG
0.050
(0.0078)
196.36
< 0.0005** <0.0005**
0.0025**
0.0232**
0.0257**
0.4208
(87.4495)
(127.7471)
(149.0788)
Bonus
Time
* p ≤ 0.1
** p ≤ 0.05
7.5.3 Bonus
The average bonus per group is shown in the first row of Table 3. From it, we conclude that
Hypothesis 2 is true, i.e., the average bonus received by members of BIG is greater than the average
bonus received by members of BG, which in turn is greater than the average bonus hypothetically
received by members of CG. All these results are statistically significant with p-value ≤ 0.05. In
other words, providing review scores and informing reviewers about the theory behind the scoring
method do indeed increase the number of reported reviews that are similar.
Interestingly, there is a strong negative correlation between bonuses and the aggregate absolute
error for the whole task shown in the fourth row of Table 2, even though the former is computed
by making pairwise comparisons between reported reviews, whereas the latter is computed by
comparing reported reviews with gold-standard reviews. The Pearson correlation coefficients for
BG, BIG, and CG are, respectively, −0.73, −0.79, and −0.72. This result implies that there exists
a strong positive correlation between honest reporting and accuracy in this task, a fact which is in
agreement with our theoretical model.
7.5.4 Completion Time
The average completion time per group is shown in the second row of Table 3. We start by noting
that Hypothesis 3 is not true. Surprisingly, the average time spent on the task by members of the
group BG is statistically equivalent to the average time spent by members of the group CG since
the null hypothesis cannot be rejected. The average completion time by members of the group
BIG is the highest one amongst the three groups, and this result is statistically significant with
p-value ≤ 0.05. A possible explanation for this result is that reviewers work on the reviewing task
more seriously by taking more time to complete it when they receive a brief explanation regarding
some theoretical results of the proposed scoring method, whereas they could be quickly guessing
how their peers would review the texts when the extra explanation about the theoretical results is
not provided.
It is noteworthy that even though the average values might suggest that spending more time
reviewing the texts results in higher bonuses and lower overall absolute errors, we do not find any
significant correlation between these variables at an individual level.
25
Table 4: Accuracy of the average review (AVG) and the consensual review (CR) per group. The
average of the absolute difference between the aggregate evaluation scores and the corresponding
gold-standard evaluation scores is shown per group and criteria. For each criterion, the lowest
average absolute difference in each group is highlighted in bold. The standard deviations are in
parenthesis. A total of 1000 bootstrap resamples were used.
Grammar
Text 1 Clarity
Relevance
Grammar
Text 2 Clarity
Relevance
Grammar
Text 3 Clarity
Relevance
BG
BIG
CG
AVG
CR
AVG
CR
AVG
CR
0.2622
(0.0920)
0.8219
(0.0903)
0.2188
(0.0676)
0.1643
(0.0819)
0.4965
(0.0879)
0.3590
(0.0804)
0.7598
(0.1204)
0.1379
(0.0565)
0.1197
(0.0619)
0.2781
(0.1015)
0.8159
(0.0963)
0.1462
(0.0536)
0.1667
(0.0855)
0.4314
(0.0970)
0.3508
(0.0933)
0.6976
(0.1473)
0.0859
(0.0400)
0.0699
(0.0400)
0.1238
(0.0680)
0.6211
(0.0850)
0.2020
(0.0622)
0.0754
(0.0571)
0.3812
(0.0858)
0.4360
(0.0948)
0.7792
(0.1222)
0.0000
(0.0000)
0.1019
(0.0521)
0.1086
(0.0631)
0.6078
(0.0987)
0.1382
(0.0512)
0.0698
(0.0541)
0.3036
(0.0860)
0.4957
(0.1061)
0.7198
(0.1505)
0.0000
(0.0000)
0.0597
(0.0335)
0.1285
(0.0800)
0.8633
(0.1031)
0.2996
(0.0783)
0.0716
(0.545)
0.5405
(0.0857)
0.4996
(0.0909)
1.0231
(0.1161)
0.1615
(0.0522)
0.2022
(0.0696)
0.1307
(0.0829)
0.8477
(0.1156)
0.2156
(0.0696)
0.0679
(0.0531)
0.5022
(0.0987)
0.5642
(0.0999)
1.0288
(0.1455)
0.1115
(0.0453)
0.1320
(0.0537)
7.5.5 Consensus
Lastly, we tested the accuracy of the method proposed in Section 6 to find a consensual review.
We compared the resulting consensual review with the average review by using a bootstrapping
technique. For each group of reviewers, we randomly resampled with replacement reviews from
the original dataset so as to obtain bootstrap resamples. The size of each bootstrap resample was
equal to the size of the original dataset, i.e., each bootstrap resample contained 50 data points (the
original number of reviewers), each one consisting of 9 evaluation scores.
For each bootstrap resample, we aggregated evaluation scores individually using both the pro-
posed method for finding a consensual review and the average method. For each evaluation score,
the weight that each reviewer i assigned to reviewer j’s reported evaluation score was computed
according to equation (11). The proper scoring rule R and the constants γ and λ were set so as to
reward agreement as in Section 4.3. Given that the best evaluation score in this task was v = 2 and
α = (1, 1, 1), each element of a reviewer’s estimated posterior predictive distribution in (4) could
take on only two values: 0.25 and 0.5. Consequently, the numerator in (11) could take on only two
values: 0.5, if reviewer i and j’s reported evaluation scores were the same, and 0.25 otherwise.
After aggregating evaluation scores, we computed the accuracy of each aggregation method.
The outcome measure was the absolute difference between each aggregate evaluation score and
the corresponding gold standard score. Thus, the outcome measure was an integer with a value
26
between zero and two, and the closer this value was to zero, the better the resulting accuracy.
Table 4 shows the average accuracy by group resulting from a total of 1000 bootstrap resamples.
Table 4 shows that consensual evaluation scores are more accurate than average evaluation
scores in 20 out of 27 cases, and equally accurate in one case. It comes as no surprise that con-
sensual evaluation scores are more accurate in groups where review scores were provided since
these groups reported more accurate reviews and their reported reviews were more similar, as
previously discussed. We performed a statistical analysis to investigate whether or not these dif-
ferences in accuracy are statistically significant. Since we used the same bootstrap resamples for
both aggregation methods, the Wilcoxon signed-rank test was used. The null hypothesis was that
the outcome measures resulting from both aggregation methods are equivalent. The alternative
hypothesis was that the outcome measure resulting from the consensual method was less than the
outcome measure resulting from the average method, which implies that the former is more ac-
curate than the latter. All the resulting 27 p-values are extremely small (cid:0)< 10−10(cid:1). Therefore,
we conclude that Hypothesis 4 is indeed true, i.e., the proposed method for finding a consensual
review is, on average, more accurate than the average approach in this experiment. As discussed
in Section 6, we believe this result happens because disparate reported reviews are less likely to
have a big influence on the consensual review than on the average review.
8 Conclusion
We proposed a scoring method built on strictly proper scoring rules that induces honest reporting
when outcomes are not observable. We illustrated the mechanics behind our scoring method by
applying it to the peer-review process. In order to do so, we modeled the peer-review process
using a Bayesian model where the uncertainty regarding the quality of the manuscript is taken into
account. The main assumptions in our model are that reviewers cannot be influenced by other
reviewers, and reviewers are Bayesian decision-makers.
We then showed how our scoring method can be used to evaluate reported reviews and to
encourage honest reporting by risk-neutral reviewers. The proposed method assigns scores based
on how close reported reviews are, where closeness is defined by an underlying proper scoring rule.
Under the aforementioned assumptions, we showed that risk-neutral reviewers strictly maximize
their expected scores by honestly disclosing their reviews. We also proposed an extension of
our model and scoring method to scenarios where reviewers evaluate a manuscript under several
criteria. We discussed how honest reporting is related to accuracy in our model: when reviewers
report honestly, the distribution of reported reviews convergences to the distribution that represents
the quality of the manuscript as the number of reported reviews increases.
Since all reviews are not always in agreement, we suggested an adaptation of the method
proposed by DeGroot [1974] to find a consensual review. Intuitively, the proposed method works
as if the reviewers were going through several rounds of discussion, where in each round they are
informed about others’ reported reviews, and they update their own reviews using our predefined
method in order to reach a consensus. Formally, each updated review is a convex combination of
reported reviews, where review scores are used as part of the weights that reviewers assign to their
peers’ reviews. We showed that the resulting method always converges to a consensual review
when reviewers’ expected review scores are positive, and we provided behavioral foundations for
the aggregation method.
We tested the efficacy of both the proposed scoring method and the proposed aggregation
method on a peer-review experiment using Amazon’s Mechanical Turk. Our experimental results
corroborated the relationship between honest reporting and accuracy in our model. We empiri-
cally showed that providing review scores through pairwise comparisons results in more accurate
reviews than the traditional peer-review process, where reviewers have no direct incentives for ex-
27
pressing their true reviews. Moreover, reviewers tended to agree more with each other when they
received review scores. In addition, our method for finding a consensual review outperformed the
traditional average method in our peer-review experiments.
For ease of exposition, our discussion on peer review was focused on scientific communica-
tion. However, our model and scoring method are readily applied to most peer-review settings,
e.g., academic courses, clinical peer review, etc. Moreover, our proposed method to incentivize
honest reporting is readily applied to different domains. For example, our method can be used to
incentivize honest feedback in reputation systems, where individuals rate a product/service after
experiencing it, and to induce honest evaluation of different strategic plans in an organization’s
strategic planning process. The proposed aggregation method is also general in a sense that it can
be applied to any decision analysis process where experts express their opinions through probabil-
ity distributions over a set of exhaustive and mutually exclusive outcomes.
Given the positive results obtained in our peer-review experiments, an interesting open ques-
tion is whether or not the methods proposed in this paper would perform as well in other domains,
such as in the aforementioned reputation systems and strategic planning in organizations. Another
question worth contemplating is whether or not incentives other than from the received scores play
a role in our scoring method. For example, one can conjecture that altruism may play an important
role in our scoring method. In our peer-review experiments, the performance of the reviewers not
only affect their own review scores, but also the review scores of their peers. In other words, if
reviewers do not put enough effort into reporting high-quality reviews, not only might they re-
ceive low review scores, but other reviews evaluated based on those erroneous reviews might also
receive low review scores. Thus, an interesting future work is to investigate whether or not ex-
perts, in general, have an altruistic motive to put more effort into the underlying task in order to
maximize the potential payoffs of their peers.
Acknowledgments
The authors thank Carol Acton, Katherine Acheson, Stefan Rehm, Susan Gow, and Veronica
Austen for providing gold-standard reviews for our experiments.
Appendix
In this appendix, we describe the texts used in our experiments as well as the gold-standard reviews
reported by five professors and tutors from the English and Literature Department at the University
of Waterloo, henceforth referred to as the experts.
Text 1
An excerpt from the “Sonnet XVII” by Neruda [2007]. Intentionally misspelled words are high-
lighted in bold.
“I do not love you as if you was salt-rose, or topaz,
or the arrown of carnations that spread fire:
I love you as certain dark things are loved,
secretly, between the shadown and the soul”
Table 5 shows the evaluation scores reported by the experts. The gold-standard evaluation
score for each criterion is the median/mode of the reported evaluation scores.
28
Table 5: Evaluation scores reported by the experts for Text 1.
Criterion
Expert 1 Expert 2 Expert 3 Expert 4 Expert 5 Median/Mode
Grammar
Clarity
Relevance
1
2
2
0
2
2
1
2
2
0
1
2
1
2
2
1
2
2
Text 2
An excerpt from “The Cow” by Taylor et al. [2010]. Intentionally misspelled words are highlighted
in bold.
“THANK you, prety cow, that made
Plesant milk to soak my bread,
Every day and every night,
Warm, and fresh, and sweet, and white.”
Table 6 shows the evaluation scores reported by the experts. The gold-standard evaluation
score for each criterion is the median/mode of the reported evaluation scores.
Table 6: Evaluation scores reported by the experts for Text 2.
Criterion
Expert 1 Expert 2 Expert 3 Expert 4 Expert 5 Median/Mode
Grammar
Clarity
Relevance
1
2
1
1
2
0
1
2
0
1
1
1
1
2
1
1
2
1
Text 3
Random words in a semi-structured way. Each line starts with a noun followed by a verb in a
wrong verb form. All the words in the same line start with a similar letter in order to mimic a
poetic writing style.
“Baby bet binary boundaries bubbles
Carlos cease CIA conditionally curve
Daniel deny disease domino dumb
Faust fest fierce forced furbished”
Table 7 shows the evaluation scores reported by the experts. The gold-standard evaluation
score for each criterion is the median/mode of the reported evaluation scores.
Table 7: Evaluation scores reported by the experts for Text 3.
Criterion
Expert 1 Expert 2 Expert 3 Expert 4 Expert 5 Median/Mode
Grammar
Clarity
Relevance
0
0
0
1
0
1
0
0
0
0
0
0
0
0
0
0
0
0
29
References
AICPA: American Institute of CPAs. Peer Review Program Manual. 2012. Retrieved from
http://www.aicpa.org/InterestAreas/PeerReview.
J. Argenti. Corporate Planning: a Practical Guide. Number 2. Routledge, 1968.
D. F. Bacon, Y. Chen, I. Kash, D. C. Parkes, M. Rao, and M. Sridharan. Predicting Your Own
Effort. In Proceedings of the 11th International Conference on Autonomous Agents and Multi-
agent Systems, pages 695–702, 2012.
J. M. Bernardo and A. F. M. Smith. Bayesian Theory. John Wiley & Sons, 1994.
L. Bornmann, R. Mutz, and H. D. Daniel. Gender Differences in Grant Peer Review: A Meta-
Analysis. Journal of Informetrics, 1(3):226 – 238, 2007.
A. E. Budden, T. Tregenza, L. W. Aarssen, J. Koricheva, R. Leimu, and C. J. Lortie. Double-Blind
Review Favours Increased Tepresentation of Female Authors. Trends in Ecology and Evolution,
23(1):4–6, 2008.
M. D. Buhrmester, T. Kwang, and S. D. Gosling. Amazon’s Mechanical Turk: A New Source of
Inexpensive, Yet High-Quality, Data? Perspectives on Psychological Science, 6(1):3–5, 2011.
A. Carvalho and K. Larson. Sharing a Reward Based on Peer Evaluations.
In Proceedings of
the 9th International Conference on Autonomous Agents and Multiagent Systems, pages 1455–
1456, 2010.
A. Carvalho and K. Larson. A Truth Serum for Sharing Rewards.
In Proceedings of the 10th
International Conference on Autonomous Agents and Multiagent Systems, pages 635–642, 2011.
A. Carvalho and K. Larson. Sharing Rewards Among Strangers Based on Peer Evaluations. De-
cision Analysis, 9(3):253–273, 2012.
A. Carvalho and K. Larson. A Consensual Linear Opinion Pool.
In Proceedings of the 23rd
International Joint Conference on Artificial Intelligence, pages 2518–2524, 2013.
R. T. Clemen and R. L. Winkler. Combining Probability Distributions From Experts in Risk
Analysis. Risk Analysis, 19:187–203, 1999.
P. E. Dans. Clinical Peer Review: Burnishing a Tarnished Icon. Annals of Internal Medicine, 118
(7):566–568, 1993.
M. H. DeGroot. Reaching a Consensus. Journal of the American Statistical Association, 69(345):
118–121, 1974.
E. S. Epstein. A Scoring System for Probability Forecasts of Ranked Categories. Journal of
Applied Meteorology, 8(6):985–987, 1969.
M. E. Falagas, G. M. Zouglakis, and P. K. Kavvadia. How Masked Is the Masked Peer Review of
Abstracts Submitted to International Medical? Mayo Clinic Proceedings, 81(5):705, 2006.
D. Friedman. Effective Scoring Rules for Probabilistic Forecasts. Management Science, 29(4):
447–454, 1983.
T. Gneiting and A. E. Raftery. Strictly Proper Scoring Rules, Prediction, and Estimation. Journal
of the American Statistical Association, 102(477):359–378, 2007.
30
F. Godlee, T. Jefferson, M. Callaham, J. Clarke, D. Altman, H. Bastian, C. Bingham, and J. Deeks.
Peer Review in Health Sciences. BMJ books London, 2003.
R. Hanson. Combinatorial Information Market Design. Information Systems Frontiers, 5(1):107–
119, 2003.
J. J. Horton, D. G. Rand, and R. J. Zeckhauser. The Online Laboratory: Conducting Experiments
in a Real Labor Market. Experimental Economics, 14(3):399–425, 2011.
S.-W. Huang and W.-T. Fu. Enhancing Reliability Using Peer Consistency Evaluation in Human
In Proceedings of the 2013 Conference on Computer Supported Cooperative
Computation.
Work, pages 639–648, 2013.
P. G. Ipeirotis. Analyzing the Amazon Mechanical Turk Marketplace. XRDS Crossroads: The
ACM Magazine for Students, 17(2):16–21, 2010.
R. Jurca and B. Faltings. Mechanisms for Making Crowds Truthful. Journal of Artificial Intelli-
gence Research, 34:209–253, 2009.
A. C. Justice, M. K. Cho, M. A. Winker, J. A. Berlin, D. Rennie, and The PEER Investigators.
Does Masking Author Identity Improve Peer Review Quality?: A Randomized Controlled Trial.
Journal of the American Medical Association, 280(3):240–242, 1998.
S. Lock. A Difficult Balance: Editorial Peer Review in Medicine. Nuffield Provincial Hospitals
Trust, 1985.
LSC: Legal Services Commission.
Independent Peer Review.
2005.
Retrieved from
http://www.legalservices.gov.uk/civil/how/mq_peerreview.asp.
M. Marge, S. Banerjee, and A. I. Rudnicky. Using the Amazon Mechanical Turk for Transcription
of Spoken Language. In Proceedings of the 2010 IEEE International Conference on Acoustics
Speech and Signal Processing, pages 5270–5273, 2010.
W. Mason and S. Suri. Conducting Behavioral Research on Amazon’s Mechanical Turk. Behavior
Research Methods, 44(1):1–23, 2012.
N. Miller, P. Resnick, and R. Zeckhauser. Eliciting Informative Feedback: The Peer-Prediction
Method. Management Science, 51(9):1359–1373, 2005.
A. H. Murphy. A Note on the Ranked Probability Score. Journal of Applied Meteorology, 10(1):
155–156, 1970.
Y. Nakazono. Strategic Behavior of Federal Open Market Committee Board Members: Evidence
from Members’ Forecasts. Journal of Economic Behavior & Organization, 93:62–70, 2013.
R. F. Nau. Should Scoring Rules Be “Effective”? Management Science, 31(5):527–535, 1985.
D. J. Navarro, T. L. Griffiths, M. Steyvers, and M. D. Lee. Modeling Individual Differences with
Dirichlet Processes. Journal of Mathematical Psychology, 50(2):101–122, 2006.
P. Neruda. 100 Love Sonnets. Exile, Bilingual edition, 2007.
N. S. Newcombe and M. E. Bouton. Masked Reviews Are Not Fairer Reviews. Perspectives on
Psychological Science, 4(1):62–64, 2009.
L. Pappano. The Year of the MOOC. New York Times, page ED26, November 4th, 2012.
31
S. Plous. The Psychology of Judgment and Decision Making. Mcgraw-Hill Book Company, 1993.
D. Prelec. A Bayesian Truth Serum for Subjective Data. Science, 306(5695):462–466, 2004.
R. B. Primack and R. Marrs. Bias in the Review Process. Biological Conservation, 141(12):
2919–2920, 2008.
G. Radanovic and B. Faltings. A Robust Bayesian Truth Serum for Non-Binary Signals.
In
Proceedings of the 27th AAAI Conference on Artificial Intelligence, 2013.
R. Robinson. Calibrated Peer Review. The American Biology Teacher, 63(7):474–480, 2001.
M. Roos, J. Rothe, and B. Scheuermann. How to Calibrate the Scores of Biased Reviewers by
Quadratic Programming. In Proceedings of the Twenty-Fifth Conference on Artificial Intelli-
gence, pages 255–260, 2011.
S. M. Ross. Stochastic Processes (Wiley Series in Probability and Statistics). Wiley, 2 edition,
1995.
A. D. Shaw, J. J. Horton, and D. L. Chen. Designing Incentives for Inexpert Human Raters. In
Proceedings of the ACM 2011 Conference on Computer Supported Cooperative Work, pages
275–284, 2011.
R. Snow, B. O’Connor, D. Jurafsky, and A. Y. Ng. Cheap and Fast—But is it Good? Evaluating
In Proceedings of the Conference on
Non-Expert Annotations for Natural Language Tasks.
Empirical Methods in Natural Language Processing, pages 254–263, 2008.
C.-A. S. Stael von Holstein. A Family of Strictly Proper Scoring Rules Which Are Sensitive to
Distance. Journal of Applied Meteorology, 9(3):360–364, 1970.
J. Taylor, A. Taylor, and K. Greenaway. Little Ann and Other Poems. Nabu Press, 2010.
A. Tversky and D. Kahneman. Judgment under Uncertainty: Heuristics and Biases. Science, 185
(4157):1124–1131, 1974.
S. van Rooyen. The Evaluation of Peer-Review Quality. Learned Publishing, 14(2):85–91, 2001.
S. van Rooyen, N. Black, and F. Godlee. Development of the Review Quality Instrument (RQI) for
Assessing Peer Reviews of Manuscripts. Journal of Clinical Epidemiology, 52(7):625 – 629,
1999.
R. R. J. Weiss. Optimally Aggregating Elicited Expertise: A Proposed Application of the Bayesian
Truth Serum for Policy Analysis. PhD thesis, Massachusetts Institute of Technology, 2009.
C. Wenneras and A. Wold. Nepotism and Sexism in Peer-Review. Nature, 387:341–343, 1997.
R. L. Winkler and A. H. Murphy. “Good” Probability Assessors. Journal of Applied Meteorology,
7(5):751–758, 1968.
J. Witkowski and D. C. Parkes. A Robust Bayesian Truth Serum for Small Populations. In Pro-
ceedings of the 26th AAAI Conference on Artificial Intelligence, 2012.
A. Yankauer. How Blind is Blind Review? American Journal of Public Health, 81(7):843–845,
1991.
32
|
1003.3784 | 1 | 1003 | 2010-03-19T12:31:55 | Simulating Customer Experience and Word Of Mouth in Retail - A Case Study | [
"cs.MA",
"cs.CE"
] | Agents offer a new and exciting way of understanding the world of work. In this paper we describe the development of agent-based simulation models, designed to help to understand the relationship between people management practices and retail performance. We report on the current development of our simulation models which includes new features concerning the evolution of customers over time. To test the features we have conducted a series of experiments dealing with customer pool sizes, standard and noise reduction modes, and the spread of customers' word of mouth. To validate and evaluate our model, we introduce new performance measure specific to retail operations. We show that by varying different parameters in our model we can simulate a range of customer experiences leading to significant differences in performance measures. Ultimately, we are interested in better understanding the impact of changes in staff behavior due to changes in store management practices. Our multi-disciplinary research team draws upon expertise from work psychologists and computer scientists. Despite the fact we are working within a relatively novel and complex domain, it is clear that intelligent agents offer potential for fostering sustainable organizational capabilities in the future. | cs.MA | cs | Simulating Customer Experience and Word-Of-Mouth in
Retail - A Case Study
Peer-Olaf Siebers
Uwe Aickelin
University of Nottingham, School of Computer Science
Nottingham, NG8 1BB, UK
[email protected]
Helen Celia
Chris W. Clegg
University of Leeds, Centre for Organisational Strategy, Learning & Change (LUBS)
Leeds, LS2 9JT, UK
Agents offer a new and exciting way of understanding the world of work. In this paper
we describe the development of agent-based simulation models, designed to help to
understand
the relationship between people management practices and retail
performance. We report on the current development of our simulation models which
includes new features concerning the evolution of customers over time. To test the
features we have conducted a series of experiments dealing with customer pool sizes,
standard and noise reduction modes, and the spread of customers’ word of mouth. To
validate and evaluate our model, we introduce new performance measure specific to
retail operations. We show that by varying different parameters in our model we can
simulate a range of customer experiences leading to significant differences in
performance measures. Ultimately, we are interested in better understanding the
impact of changes in staff behavior due to changes in store management practices.
Our multi-disciplinary research team draws upon expertise from work psychologists and
computer scientists. Despite the fact we are working within a relatively novel and
complex domain, it is clear that intelligent agents offer potential for fostering
sustainable organizational capabilities in the future.
Keywords: agent-based modeling, agent-based simulation, retail performance,
management practices, shopping behavior, customer satisfaction, word of mouth
1. Introduction
The retail sector has been identified as one of the biggest contributors to the productivity gap, whereby the
productivity of the UK lags behind that of France, Germany and the USA [1, 2]. A recent report into UK
productivity asserted that, ‘...the key to productivity remains what happens inside the firm and this is
something of a ‘black box’…’ [3]. A recent literature review of management practices and organizational
productivity concluded that management practices are multidimensional constructs that generally do not
demonstrate a straightforward relationship with productivity variables [4], and that both management
practices and productivity measures must be context specific to be (respectively) effective and meaningful.
Many attempts have been made to link management practices to an organization’s productivity and
performance (for a review, see [5]), however research findings have been so far been mixed, creating the
opportunity for the application of different techniques to advance our understanding.
Simulation can be used to analyze the operation of dynamic and stochastic systems showing their
development over time. There are many different types of simulation, each of which has its specific field of
application. Agent-Based Simulation (ABS) is particularly useful when complex interactions between
system entities exist such as autonomous decision making or proactive behavior. Agent-Based Modeling
1
(ABM) shows how micro-level processes affect macro level outcomes; macro level behavior is not
explicitly modeled, it emerges from the micro-decisions made by the individual entities [6].
There has been a fair amount of modeling and simulation of operational management practices, but people
management practices have often been neglected although research suggests that they crucially impact
upon an organization’s performance [7]. One reason for this relates to the key component of people
management practices, an organization’s people, who may often be unpredictable in their individual
behavior. Previous research into retail productivity has typically focused on consumer behavior and
efficiency evaluation (e.g. [8, 9]), and we seek to build on this work and address the neglected area of
people management practices in retail [10].
The overall aim of our project is to investigate the link between different management practices and
productivity. This strand of work focuses on simulating various in-store scenarios grounded in empirical
case studies with a leading UK retailer. In this paper we aim to understand and predict how Agent-Based
Modeling and Simulation (ABMS) can assess and optimize the impact of people management practices on
customer satisfaction and Word-Of-Mouth (WOM) in relation to the performance of a service-oriented
retail department. To achieve this aim we have adopted a case study approach using applied research
methods to collect both qualitative and quantitative data. In summary, we have worked with a leading UK
retail organization to conduct four weeks’ of informal participant observations in four departments across
two retail stores, forty semi-structured interviews with employees including a sixty-three-item
questionnaire on the effectiveness of retail management practices, and drawn upon a variety of established
information sources internal to the company. This approach has enabled us to acquire a valid and reliable
understanding of how the real system operates, revealing insights into the working of the system as well as
the behavior of and interactions between the different individuals and their complementary roles within the
retail department. Using this knowledge and data, we have applied ABM to devise a functional
representation of the case study departments. By executing the developed ABS models we can run
experiments to investigate the effects of different management scenarios.
So far we have only studied the impact of people management practices (e.g. training and empowerment)
on a customer base that is not influenced by any external or internal stimuli, and hence does not evolve [11,
12]. In order to be able to investigate the impact of management practices on customer satisfaction in a
more realistic way we need to consider the cues that stimulate customers to respond to these practices.
Therefore our focus is currently on building the capabilities to model customer evolution as a consequence
of the implementation of management practices. Consumer decision making, and consequently changes in
the behavior of customers over time, are driven by an integration of each consumer’s cognitive and
affective skills [13]. At any given point in time a customer’s behavior, as the product of an individual’s
cognitions, emotions, and attitudes, may be attributable to an external social cue such as a friend’s
recommendation or in-store stimuli [14], or an internal cue such as memory of one’s own previous
shopping experiences [15]. Changing customer requirements may in turn alter what makes a successful
management practice (because these are context specific, and customers are a key component of any retail
context). In order to enable such studies we enhanced our original Management Practices Simulation
(ManPraSim) model and introduced some new features.
In this paper we discuss the key features we have implemented in order to allow the investigation of these
kinds of behavioral dynamics. What we are learning here about modeling human behavior has implications
for modeling any complex system that involves many human interactions and where the people function
with some degree of autonomy.
2. Related Work
When investigating the behavior of complex systems the choice of an appropriate modeling technique is
very important [16]. To inform the choice of modeling technique, the relevant literature spanning the fields
of Economics, Social Science, Psychology, Retail, Marketing, OR, Artificial Intelligence, and Computer
Science has been reviewed. Within these fields a wide variety of approaches are used which can be
classified into three main categories: analytical methods, heuristic methods, and simulation. Often a
2
combination of these are used within a single model (e.g. [17, 18]. After a thorough investigation of the
relevant literature we have identified simulation as being the most appropriate approach for our purposes.
Simulation introduces the possibility of a new way of thinking about social and economic processes, based
on ideas about the emergence of complex behavior from relatively simple activities [19]. The purpose of
simulation is either to better understand the operation of a target system, or to make predictions about a
target system’s performance [20]. It allows the testing and evaluation of a theory, and investigation of its
implications. Whereas analytical models typically aim to explain correlations between variables measured
at one single point in time, simulation models are concerned with the development of a system over time
[21]. They allow us to study the transient conditions during the process, along the way to the eventual
outcome [22]. The outcome of simulation study is the acquisition of more knowledge about the dynamic
behavior of a system, i.e. the state changes of the model as time advances [23]. There are many different
approaches to OR simulation, amongst them Discrete-Event Simulation (DES), System Dynamics (SD),
and ABS, which is sometimes also referred to as individual-based simulation [24]. The choice of the most
suitable approach depends on the issues investigated, the input data available, the required level of analysis,
and the type of answers that are sought [16].
Although computer simulation has been used widely since the 1960s, ABMS only became popular in the
early 1990s [25]. It is described by [26] as a mindset as much as a technology: ‘It is the perfect way to view
things and understand them by the behavior of their smallest components’. In agent-based models a
complex system is represented by a collection of agents that are programmed to follow simple behavioral
rules. Agents can interact with each other and with their environment to produce complex collective
behavioral patterns. The main characteristics of agents are their autonomy and their ability to take flexible
action in reaction to their environment, both determined by motivations generated from their internal states.
They are designed to mimic the attributes and behaviors of their real-world counterparts.
ABS is still a relatively new simulation technique and its principal application has been in academic
research. There are, however, a wide range of existing application domains that are making use of the agent
paradigm and developing agent-based systems, for example in software technology, robotics, and complex
systems, and an ever increasing number of computer games use the ABMS approach. [27] draw a
distinction between two central Multi-Agent System (MAS) paradigms: multi-agent decision systems and
multi-agent simulation systems. In multi-agent decision systems, agents participating in the system must
make joint decisions as a group. Mechanisms for joint decision-making can be based on economic
mechanisms, such as an auction, or alternative mechanisms, such as argumentation. Multi-agent simulation
systems are used as a model to simulate some real-world domain. The typical application is in domains
which involve many different components, interacting in diverse and complex ways and where the system-
level properties are not readily inferred from the properties of the components. ABMS is extensively used
by the game and film industry to develop realistic simulations of individual characters and societies. It is
used in computer games, for example The SIMS™ [28], or in films when diverse heterogeneous characters
animations are required, for example the Orcs in Lord of the Rings™ [29].
Due to the characteristics of the agents, the ABMS approach (which is often implemented as a process
interaction DES) appears to be more suitable than the event scheduling modeling approach (which is
typically associated with the term DES) when it comes to modeling human-centered systems. This is due to
the fact that it supports the modeling of autonomous behavior. ABMS seems to promote a natural form of
modeling these systems [22]. There is a structural correspondence between the real system and the model
representation, which makes them more intuitive and easier to understand than for example a system of
differential equations as used in SD. [30] emphasizes that one of the key strengths of ABMS is that the
system as a whole is not constrained to exhibit any particular behavior as the system properties emerge
from its constituent agent interactions. Consequently, assumptions of linearity, equilibrium and so on, are
not needed.
On the other hand, some difficulties have been linked to use of ABMS. One disadvantage that is often
mentioned is the high demand in computational power compared to other simulation techniques, most
notably in that all of the interactions between agents and between the agent and the environment have to be
considered during the simulation run. However, rapid growth in affordable computer power makes this
3
problem less significant [31]. There is consensus in the literature that it is difficult to evaluate agent-based
models, because of the heterogeneity of the agents and because the behavior of the system emerges from
the interactions between the individual entities. [32] argues that for ‘intellective models’ (e.g. models that
illustrate the relative impact of basic explanatory mechanisms) validation is somewhat less critical and the
most important thing is to maintain a balance between keeping a model simple and maintaining veridicality
(the extent to which a knowledge structure accurately reflects the information environment it represents).
More recently some attempts have been made to develop a common approach to assuring the credibility
(verification and validation) of ABS models (see for example [33 - 35].
A two stage approach for empirical validation has been proposed by [36]. The first stage comprises micro-
validation of the behavior of the individual agents and the authors recommend this is conducted with
reference to real data on individual behavior. The second stage covers macro-validation of the aggregate or
emergent behavior of the model when individual agents interact and is advised to be conducted with
reference to aggregate time series data. Further, [36] remark that at the macro level only qualitative
validation judgments may be possible. Also, problems often occur through the lack of adequate empirical
data. [37] state that most quantitative research has concentrated on ‘variable and correlation’ models that do
not cohere well with process-based simulation that is inherent in agent-based models. Finally, [37] point
out the danger that people new to ABMS may expect too much from the models, particularly regarding
their predictive ability. To mitigate this problem it is important to be clear with individuals about what this
modeling technique can really offer, in order to guide realistic expectations.
3. Model Design and Data Collection
Before building a simulation model one needs to understand the particular problem domain [38]. In order to
gain this understanding we have conducted some case studies. What we have learned during these case
studies is reflected in the conceptual models presented in this chapter. Furthermore we explain how we
intend to use the data we have gathered during our case studies. Throughout the rest of the paper we will
use the term 'actor' to refer to a person in the real system, whereas the term 'agent' will be reserved for their
counterparts in the simulation model.
3.1 Knowledge Gathering
Case studies were undertaken in four departments across two branches of a leading UK retailer. The case
study work involved extensive data collection techniques, spanning: participant observation, semi-
structured interviews with team members, management and personnel, completion of survey questionnaires
and the analysis of company data and reports (for further information, see [39]). Research findings were
consolidated and fed back (via report and presentation) to employees with extensive experience and
knowledge of the four departments in order to validate our understanding and conclusions. This approach
has enabled us to acquire a valid and reliable understanding of how the real system operates, revealing
insights into the working of the system as well as the behavior of and interactions between the different
actors within it.
In order to make sure that our results regarding the application of management practices are applicable for a
wide variety of departments we have chosen two types of case study departments which are substantially
different not only in their way of operating but also their customer type split and staff setup. We collected
our data in the Audio & Television (A&TV) and the WomensWear (WW) departments of the two case
study branches.
The key differences between these two department types can be summarized as follows. The average
customer service time in A&TV is significantly longer, and the average purchase is significantly more
expensive than in WW. The likelihood of a customer seeking help in A&TV is also much higher than in
WW. Out of customers who have received advice, those in WW have a higher likelihood of making a
purchase (indeed customers’ questions tend to be very specific to a desired purchase) than in A&TV.
Considering customer types, A&TV tends to attract more solution demanders and service seekers, whereas
WW customers tend to be shopping enthusiasts. Finally, it is important to note that the conversion rate (the
likelihood of customers making a purchase) is higher in WW than in A&TV.
4
3.2 Conceptual Modeling
A conceptual model captures only the essential and relevant characteristics of a system. It is formulated
from the initial problem statement, informal user requirements, and data and knowledge gathered from
analysis of previous development models [23]. We have used the knowledge gained from the case studies
to inform our conceptual models of the system to be investigated, the actors within the system, and their
behavioral changes due to certain stimuli.
3.2.1 Main Concepts for the Simulation Model
Our initial ideas for the simulation model and its components are display below in Figure 1. Here we can
see a bird’s eye view of a shopping centre with customers (red dots) visiting a range of shops. Based on this
conceptual overview, we determined that we required three types of agents (customers, sales staff and
managers), each with a different set of relevant attributes. Global parameters can influence any aspect of
the system. The core of our conceptual simulation model consists of an ABS model with a user interface to
allow some form of user interaction (change of parameters) before and during runtime. Regarding system
outputs, we aim to find some emergent behavior on a macro level. Visual representation of the simulated
system and its actors allows us to monitor and better understand the interactions of entities within the
system. Coupled with the standard DES performance measures, we can then identify bottlenecks to assist
with optimization of the modeled system.
<Insert Figure 1 somewhere here>
3.2.2 Concepts for the Agents
We have used state charts for the conceptual design of our agents. State charts show the different states an
entity can be in and define the events that cause a transition from one state to another. This is exactly the
information we need in order to represent our agents at a later stage within the simulation environment. We
have found this form of graphical representation a useful part of the agent design process because it is
easier for an expert in the real system (who is not an expert in ABMS) to quickly take on board the model
conceptualization and provide useful validation of the model structure and content.
Designing and building a model is to some extent subjective, and the modeler has to selectively simplify
and abstract from the real scenario to create a useful model [40]. A model is always a restricted copy of the
real system, and an effective model consists of only the most important components of the real system. In
our case, the case studies indicated that the key system components take the form of the behaviors of an
actor and the triggers that initiate a change from one behavior to another. We have developed state charts
for all of the agents in our simulation model. Figure 2 shows as an example the conceptual template model
of a customer agent. The transition rules have been omitted to keep the chart comprehensible. They are
explained in detail in the Section 3.3.
<Insert Figure 2 somewhere here>
3.2.3 Concepts for a Novel Performance Measure
We have introduced a customer satisfaction level index as a novel performance measure using satisfaction
weightings. This new measure is required because existing indices such as queuing times or service times
are less useful in modeling services than manufacturing activities. In essence, purely quantitative measures
fail to capture the quality of service, which is arguably the most important potential trade-off with retail
productivity. The inevitable trade-off between quality and quantity is particularly salient when customers
and staff come face-to-face and therefore we consider this measure of quality in conjunction with others of
quantity.
Historically customer satisfaction has been defined and measured in terms of customer satisfaction with a
purchased product [41]. The development of more sophisticated measures has moved on to incorporate
customers’ evaluations of the overall relationship with the retail organization, and a key part of this is the
service interaction. Indeed, empirical evidence suggests that quality is more important for customer
5
satisfaction than price or value-for-money [42], and extensive anecdotal evidence indicates that customer-
staff service interactions are an important determinant of quality as perceived by the customer.
The index allows customer service satisfaction to be recorded throughout the simulated lifetime. The idea is
that certain situations can have a bigger impact on customer satisfaction than others, and therefore weights
can be assigned to events to account for this. Satisfaction indices do not change the likelihood that a
customer revisits the department, and although this may at first appear counter-initiative, empirical
evidence supports the notion that regardless of a customer’s changing shopping intentions he or she will
tend to repeat shopping habits [43]. Applied in conjunction with an ABS approach, we expect to observe
interactions with individual customer differences; variations which have been empirically linked to
differences in customer satisfaction (e.g. [44]). This helps the analyst to find out to what extent customers
underwent a positive or negative shopping experience and it also allows the analyst to put emphasis on
different operational aspects and try out the impact of different strategies.
3.2.4 Concepts for Modeling Customer Evolution
There are two different ways in which we consider customer evolution: external stimulation attributable to
the WOM and internal stimulation triggered by memory of one’s own previous shopping experiences (this
is still work in progress). Sharing information with other individuals (referred to as WOM), significantly
affects the performance of retail businesses [45]. An important source of WOM results from customer
experiences of retail outlets, and a customer’s judgment about whether or not the experience left them
feeling satisfied (or dissatisfied). We incorporate WOM in our simulation model by using the number of
satisfied customers at the end of the day to calculate the number of additional customers visiting on the
next day. The calculation takes into account that only a fraction of people act upon received WOM.
Our concept for representing internal stimulation comprises the exertion of influence on picking certain
customer types more often than others. An enthusiastic shopper with a high satisfaction score is much more
likely to go shopping more frequently than a disinterested shopper. Therefore, we have introduced some
constraints (e.g. out of all customers picked 50% have to be enthusiastic shoppers, 30% normal shoppers,
and 20% disinterested shoppers, or, customers with a higher satisfaction score are more likely to be picked
to revisit the department). [15] identified different types of internal states and environmental stimuli which
act as consumer cues for triggering impulse buying. Internal cues include respondents’ positive and
negative feeling states, and environmental cues include atmospheric cues in retail settings, marketer-
controlled cues, and marketing mix stimuli. When modeling internal stimulation we specifically examine
the impact of a consumer’s memory of his or her own shopping experience and the impact on department
performance measures.
3.3 Empirical Data
Often agents are based on analytical models or heuristics and, in the absence of adequate empirical data,
theoretical models are employed. However, we use frequency distributions to model state change delays
and probability distributions to model decision making processes because statistical distributions are the
best way in which we can represent the numerical data we have gathered during our case study work. In
this way a population is created with individual differences between agents, mirroring the variability of
attitudes and behaviors of their real human counterparts.
The frequency distributions are modeled as triangular distributions defining the time that an event lasts,
using the minimum, mode, and maximum duration and these figures are based on our own observations and
expert estimates in the absence of objective numerical data.
The probability distributions are partly based on company data (e.g. the rate at which each shopping visit
results in a purchase) and partly on informed estimates (e.g. the patience of customers before they leave a
queue). Table 1 and 2 show some of the distributions we have defined for our simulation models. We also
gathered some company data about work team numbers and work team composition, varying opening hours
and peak times, along with other operational details.
Table 1. Sample frequency distribution values
6
event
someone makes a purchase after browsing
someone requires help
someone makes a purchase after getting help
probability it occurs
0.37
0.38
0.56
Table 2. Sample probability values
situation
leave browse state after …
leave help state after …
leave pay queue (no patience) after …
m in mode max
15
7
1
30
15
3
5
12
20
4. Implementation
4.1 Implementation of the Main Concepts
Our ManPraSim model has been implemented in AnyLogic™ (version 5.5) which is a Java™ based multi-
paradigm simulation software [46]. During the implementation we have applied the knowledge, experience
and data accumulated through our case study work. Within the simulation model we can represent the
following actors: customers, service staff (with different levels of expertise) and different types of
managers. Figure 3 shows a screenshot of the customer and staff agent logic in AnyLogic™ as it has been
implemented in the latest version of our ManPraSim model (v4). Boxes represent states, arrows transitions,
arrows with a dot on top entry points, circles with a B inside branches, and numbers satisfaction weights.
<Insert Figure 3 somewhere here>
At the beginning of each simulation run the main customer pool is created which represents a population of
potential customers who visit the simulated department on an unspecified number of occasions. Once the
simulation has started customers are chosen at a specified rate (customer arrival rate) and released into the
simulated department. Currently two different customer types are implemented: customers who want to buy
something and customers who require a refund. If a refund is granted, a customer decides whether his or
her goal changes to leaving the department straight away, or to making a new purchase. The customer agent
template consists of four main blocks which all use a very similar logic. In each block, in the first instance,
customers try to obtain service directly from an employee and if they cannot obtain it (i.e. no suitable staff
member is available) he or she has to queue. The customer is then either served as soon as the suitable staff
member becomes available, or leaves the queue if they do not want to wait any longer (an autonomous
decision). A complex queuing system has been implemented to support different queuing rules. Once
customers have finished their shopping (either successfully or not) they leave the simulated department and
are returned back to the main customer pool where he or she rests until they are picked again.
While the customer is in the department a satisfaction score is calculated by summing the satisfaction
weights attached to the transitions that take place during the customer’s visit. For example, a customer
starts browsing and then requires some help. If he or she obtains help immediately his or her satisfaction
score goes up (+2) and after he or she received the help the score goes up again (+4). He or she then moves
to the till. If he or she has to wait for help and then leaves the queue because he or she is fed up waiting his
or her score goes down (-2). Upon leaving the department he or she will end up in this example with an
overall satisfaction score of +4 due to the good and prompt advisory service. The actual customer
satisfaction weights shown in Figure 3 were derived based on our experience during the data collection
period and validated via interviews with the store management staff.
In comparison to the customer agent state chart, the staff agent state chart is relatively simple. Whenever a
customer requests a service and the staff member is available and has the right level of expertise for the
task requested, the staff member commences this activity until the customer releases the staff member.
Whereas the customer is the active component of the simulation model, the staff member is currently
passive, simply reacting to requests from the customer (although both are technically implemented as active
objects).
7
4.2 Key Features
There are some additional key features that the simulation model possesses which we describe in the
following two sections. First we provide an overview of some important features which have already been
implemented in an older version of our simulation model (ManPraSim model v2) but are still relevant to
modeling the evolution of customers. This sets the scene for the subsequent description of the new features
we have implemented in the latest two versions of our simulation model (ManPraSim model v3 + v4).
4.2.1 Key Features of the ManPraSim Model v2
In the ManPraSim model v2 we introduced realistic footfall, customer types, a finite population of
customers, and a quick exit of customers when the department is closing. In this paper we only provide an
overview of these features because they are described in more detail elsewhere (see [47]).
Realistic footfall (introduced in v2): There are certain peak times where a relatively high number of
customers are in the department and consequently the demands placed on staff members are greater. In the
ManPraSim model v2 we have implemented hourly fluctuations through the addition of realistic footfall
(based on some automatically recorded sales transaction data) reflecting different patterns (peaks and
troughs) of customer footfall during the day and across different days of the week. In addition we can
model the varying opening hours on different days of the week.
Customer types (introduced in v2): In real life customers display certain shopping behaviors which can
be categorized. Hence we enhance the realism of our agents’ behavior in the ManPraSim modelv2 by
introducing customer types to create a heterogeneous customer base, thereby allowing us to test customer
populations closer to what we would find in reality. Customer types have been introduced based on three
identified by the case study organization’s own market research (shopping enthusiasts, solution demanders,
service seekers) and have been expanded by the addition of two further types (disinterested shoppers, and
internet shoppers who are customers that only seek advice but are likely to buy only from the cheapest
source i.e. the internet). The three customer types have been identified by the case study organization as the
customers who make biggest contribution to their business, in terms of both value and frequency of sales.
In order to avoid artificially inflating the amount of sales that we model we have introduced the two
additional types which use services but do not tend to make purchases. The definition of each type is based
on the customer’s likelihood to perform a certain action, classified as either: low, moderate, or high. The
definitions can be found in Table 3.
Table 3. Definitions for each type of customer
Customer type
Shopping enthusiast
Solution demander
Service seeker
Disinterested shopper
Internet shopper
buy
high
high
moderate
low
low
Likelihood to
wait
ask for help
moderate
moderate
low
low
high
high
low
low
high
high
ask for refund
low
low
low
high
low
Distribution adaptation (introduced in v2): In the ManPraSim model v2 we have two algorithms which
have been developed to imitate the influence of the customer type attributes mentioned above on customer
behavior. They have been implemented as methods that are invoked when defining the state change delays
modeled by triangular frequency distributions, and when supporting decision making modeled by
probability distributions. Basically the methods define new threshold values for the distributions based on
the likelihood values mentioned above. Program 1 shows an example of the pseudo code for the probability
distribution threshold correction algorithm. If the customer is a shopping enthusiast and is about to make
the decision whether to make a purchase or to leave the department directly (see Figure 3, customer state
chart, second branch after leaving the browse state) a corrected threshold value (probability) for this
decision is calculated. For this calculation the original threshold of 0.37 (see Table 2) is taken into account
and, for a shopping enthusiast where there is a high likelihood to buy (see Table 3), the corrected threshold
value is calculated as follows: 0.37+0.37/2 = 0.56. Consequently the likelihood that a shopping enthusiast
precedes to the checkout rather than leaving the department without making any purchase has risen by
18.5%.
8
Program 1. Pseudo code for the probability distribution threshold correction algorithm
for
(
each
threshold
to
be
corrected
)
do
{
}
if
OT
(
<
lim)5.0
it
=
OT
2/
else
(lim
it
1
−=
OT
2/)
if
likelihood
(
=
)0
CT
=
OT
−
lim
it
if
likelihood
(
=
)1
CT
=
OT
if
likelihood
(
=
)2
CT
=
OT
+
lim
it
where
:
OT
=
original
threshold
CT
=
corrected
threshold
likelihood
0:
=
low
1;
=
mod
erate
2;
=
high
Finite customer population (introduced in v2): A key aspect to consider is that the most interesting
system outcomes evolve over time and many of the goals of the retail company (e.g. service standards) are
planned strategically over the long-term. In the ManPraSim model v2 we have therefore introduced a finite
population of customers (main customer pool) where each customer agent is assigned certain characteristics
based on the customer types mentioned above. The customer type split in the main customer pool can be
defined via an initialization file before the execution of the simulation. The shopping experience of each
visit (satisfaction index) is stored in the long term memory of the agent after he or she has left the
department. In this way the service a customer experiences can be evaluated over the complete simulated
time span.
Quick exit at closing time (introduced in v2): In the ManPraSim model v2 we have added transitions that
emulate the behavior of customers when the store is closing. These transitions result in an immediate exit
of each customer for his or her current state (i.e. the equivalent to a customer running out of shopping time
and promptly exiting the store). Not all customer states have these additional transitions because it is for
example very unlikely that a customer will leave the store immediately when they are already queuing to
pay. Advanced model development means that now the simulated department empties within a ten to
fifteen minute period, which conforms to what we have observed in the real system.
4.2.2 Key Features of the ManPraSim Models v3 + v4
In the ManPraSim model v3 we have introduced a staff pool with an additional staff type, a new operation
mode to support sensitivity analyses and the first implementation of customers’ WOM featuring a static
main customer pool. Furthermore, we have created some new performance measures that allow us to
measure and record the shopping experience of individual visits to the department as well as the daily
performance of the department. Finally, in the ManPraSim model v4 we have introduced the second
implementation of customers’ WOM featuring a dynamic main customer pool.
Staff pool and additional staff types (introduced in v3): Retail trends reflect that shops are now open for
longer hours over more days of the week, and our case study organization is no exception. To accurately
incorporate this source of system variability, we have introduced a staff pool in the ManPraSim model v3
to allow different staffing on different days of the week. The simulation uses Full Timers (FT) to cover all
staff shifts required during weekdays. Additional staff who are required to cover busy weekend shifts are
modeled by Part Timers (PT). The maximum number of staff required of each type is calculated during the
simulation initialization. A staff pool is then created to include weekday FT staff of different types, and the
required generic PT to fill the gaps left in the staff shifts which need to be covered. At the beginning of
each day the simulation checks how many workers are required and picks the required amount of staff out
of the pool at random. The selection process is ordered as follows: FT first and then PT, if required. PT
workers have been defined as a generic staff type and can take over any required role. We have tried to
model a staff rota with more complex constraints, for example FT staff working five days followed by two
days off, however this has as yet proved unsuccessful. Therefore the currently modeled shifts do not
incorporate days off work for FT staff.
9
Noise reduction mode (introduced in v3): In the ManPraSim model v3 a noise reduction mode has been
implemented which allows us to conduct a sensitivity analysis with constant customer arrival rates,
constant staffing throughout the week and constant opening hours. We have used the average values of real
world case study data to define the constant values. This has resulted in different values for the different
case study department types. As this is a simulation model we cannot take out all the stochasticity, but this
way at least we can reduce the system noise to clearly see the impact of the parameter under investigation.
When we progress to reintroduce the system noise, the knowledge we have accumulated when
experimenting in noise reduction mode helps us to better understand patterns in systems outcomes, and be
better able to attribute causation to the introduction of a particular variable (bearing in mind that in the end
we are particularly interested in patterns between variables when they are interacting with one another and
not in isolation).
Word of mouth (introduced in v3+v4): We have developed two different strategies for modeling WOM
(which have been implemented in the ManPraSim model v3 (static main customer pool) and in v4
(dynamic main customer pool). Both strategies use the same algorithm to calculate the additional customers
attracted to the department through positive WOM and those lost customers attributable to negative WOM.
They differ in the way they pick these customers. In the first case we have used a static main customer pool
which is defined with the customer split at the beginning of the simulation run. During the simulation run,
the population of the main customer pool stays constant and all customers in the pool are available to be
randomly picked to go shopping. In the second case we allow the main customer pool size to change; rather
than picking customers from the existing pool we create new customers and add them to the main customer
pool. When the main customer pool grows or declines the customer arrival rate is also adjusted, i.e. if the
pool grows by 10% due to positive WOM so does the customer arrival rate and vice versa for negative
growth.
The algorithm to calculate the number of additional new customers and the lost customers is the same for
both strategies. We count the number of customers from the previous day who were satisfied and those who
were dissatisfied with the service provided during their shopping visit(s). Satisfied customers are more
likely to recommend shopping in the department to others whereas dissatisfied customers are more likely to
advise others not to visit. The adoption rate (i.e. the success rate of convincing others to commit to either
action) depends on the proportion of people who act upon the received WOM (the adoption fraction) and
how many contacts a customer has (the contact rate). The equation behind this calculation is shown below
in Equation 1.
Equation 1. Calculation of adoption rate
n
additional
customers(
d)
=
(n
satisfied(
d
1)
−
−
n
dissatisfi
ed(d
1)
−
*)
adoption
fraction
*
contact
rate
where
:
d
=
current
day
1d
=−
y
previousda
n
=
number
Kof
At the beginning of each trading day we create a daily customer pool that contains randomly chosen
customers from the main customer pool according to the required number of customers defined by the
customer arrival rates for the day which might differ throughout the day (e.g. considering peak times).
For the first strategy (static customer pool) we simply pick the core number of customers (defined by the
original customer arrival rates) plus an additional number of customers which is calculated through
Equation 1. These customers are also selected from the main customer pool at random. If the number is
negative we release customers from our daily customer pool back into the main customer pool. If we have
to release all of our customers (due to customer dissatisfaction, which may result from bad service due to a
large number of customers visiting on the previous day) the simulation terminates. This means that in a
sense that the store is closed for good because it cannot recover from this situation in the current
implementation.
10
The second strategy (dynamic customer pool) is slightly more complex. Here we expand or decrease the
main customer pool by the number of additional or lost customers (respectively) calculated at the beginning
of each trading day. This differs from the static strategy because we do not use the existing customers to
model direct WOM influence (i.e. the ones resulting from the service satisfaction of the previous day). This
seems to be closer to reality because the WOM in most cases carries positive messages and therefore
attracts new visitors rather than motivating existing customers to come more often. Negative WOM
dissuades potential customers from visiting the department in the first place. Our daily customer pool is
created considering the hourly customer arrival rates (derived from our case study) as well as the actual size
of the main customer pool (see Equation 2).
Equation 2. Calculation of core customers per day
core
customers
per
day
=
dynamic
pool
size
*
known
customers
per
day
static
pool
size
where
:
customers
per
day
daily
pool
not
including
WOM
additional
customers
customer
=
study
per
day
derived
from
hourly
customer
rate
case
study
data
arrival
customers
case
=
dynamic
pool
size
actual
main
pool
size
=
pool
size
starting
main
pool
size
static
=
Once the daily pool size has been established, customers are chosen from the main customer pool according
to the requested customer split and customer pool size. Then additional customers are created as new
agents. The type of each new customer is determined at random. They are created with a positive
satisfaction score of +1 because they have heard only positive things about the department and therefore
will already have a positive attitude toward the shop (otherwise they would not have been attracted to it).
Through these additional (or lost) agents the daily customer pool size can expand (or contract).
If the negative WOM causes an overall reduction in customers then we choose customers at random from
the daily customer pool. Whether or not a particular customer is permanently removed from the main
customer pool is determined as follows, depending on the satisfaction score of the chosen customer. If he or
she has a positive satisfaction score then the score is neutralized and the customer is returned to the daily
customer pool (i.e. he or she can come back for further shopping visits but the shop must start to satisfy the
customer’s needs for good service to secure further visits from the customer). If he or she has a neutral or
negative score he or she is permanently eliminated from the simulation (i.e. he or she will not come back
for any more shopping trips). In the second case the main customer pool size will shrink. If the negative
WOM results in more agents being deleted than there are customers available in the daily customer pool the
simulation will terminate (and the department cannot recover from this in the current implementation).
For both strategies, when it reaches shop closing time all customers from the daily customer pool are
released back into the main customer pool where they are available for the next day. For the second
strategy (dynamic customer pool) it is very likely that the main pool size will then change due to WOM-
triggered additions or losses of customers whereas for the first strategy (static customer pool) the main pool
size remains constant.
New performance measures: With the introduction of a finite population (represented by our customer
pool) we have had to rethink the way in which we collect statistics about the satisfaction of customers.
Previously, the life span of a customer has been a single visit to the department. At the end of his or her
visit, the individual’s satisfaction score (direction and value) has been recorded. Now the life span of a
customer lasts the full runtime of the simulation and he or she can be picked several times to visit the
department during that period. Our previous performance measures now collect different information:
satisfaction scores considering customers’ satisfaction history. These measures do not reflect individuals’
satisfaction with the current service experience but instead the satisfaction with the overall service
experience during the lifetime of the agent. Furthermore, they are biased to some extent in that an
indifferent rating quickly shifts into satisfaction or dissatisfaction (arguably this is realistic because most
people tend to make a judgment one way or the other).
11
Whilst the above is still a valuable piece of information we would also like to know how current service is
perceived by each customer. For this reason we have introduced a new set of performance measures to
record the experience of each customer’s individual visit (referred to as Customer Satisfaction Measure –
Experienced Per Visit or CSM-EPV). These are the same measures as before but on a day-to-day basis they
are not anchored by the customer’s previous experiences. We also examine the sum of these ‘per visit’
scores across the lifetime of all customers (referred to as Customer Satisfaction Measure – Accumulated
Historical Data or CSM-AHD).
Another new measure tracks the satisfaction growth for customers’ current and overall service experience.
With the incorporation of varying customer arrival rates, opening hours and staffing we have brought in a
set of performance measures that capture the impact of these variations on a daily basis. These measures
are particularly useful for optimizing departmental performance throughout the week. We also record the
satisfaction growth per each individual customer visit. At the end of the simulation run the simulation
model produces a frequency distribution which informs us about how satisfied or dissatisfied individual
customers have been with the service provided. Furthermore, all forms of customer queue (cashier, normal
help, expert help, and refund decision) are now monitored through new performance measures that record
how many people have been queuing in a specific queue, and how many of these lost their patience and left
the queue prematurely. This measure helps us to understand individual customers’ needs because it tells us
what individual customers think about the service provided. Finally, we have added some methods for
writing all parameters and performance measures into files to support documentation and analysis of the
experiments.
4.3 Model Validation
Validation ensures that the model meets its intended requirements in terms of the methods employed and
the results obtained. In order to test the operation of the ManPraSim model and ascertain face validity we
have completed several experiments.
We have completed validation exercises on the micro and macro level, as proposed by [36]. Micro
simulation comprises tracing an individual entity through the system by building a trajectory of an
individual’s behavior throughout the runtime. We have conducted this test for our two different agent types
(customer and staff agents) considering all of the different roles these can take. Furthermore we have
validated who triggers which process and for what reason. We have then extensively tested the underlying
queuing system. On the macro level we have tried to establish simulation model performance outputs that
are sound. It has turned out that conducting the experiments with the input data we collected during our
case studies did not provide us with a satisfactory match to the performance output data of the real system.
We identified the staffing setup used in the simulation models as the root cause of the problems. The data
we use here have been derived from real staff rotas. On paper these real rotas suggested that all workers are
engaged in exactly the same (one) type of work throughout the day but we know from working with, and
observing workers in, the case study organization that in reality each role includes a variety of activities.
Staff members in the real organization allocate their time proactively between competing tasks such as
customer service, stock replenishment, and processing purchases. Proactive behavior refers to the
individual staff members’ self-started, long term oriented, and persistent service behavior which goes
beyond explicit job demands [48].
So far our simulation models incorporate only one type of work per staff member. For example, the A&TV
staff rota indicates that only one dedicated cashier works on weekdays. When we have attempted to model
this arrangement, customer queues have become extremely long, and the majority of customers ended up
losing their patience and leaving the department prematurely with a high level of dissatisfaction. In the real
system we observed other staff members working flexibly to meet the customer demand, and if the queue
of customers grew beyond a certain point then one or two would step in and open up further tills to take
customers’ money before they became dissatisfied with waiting. Furthermore, we observed that a service
staff member, when advising a customer, would often continue to close the sale (e.g. filling in guarantee
forms and receiving payment from the customer) rather than asking the customer to queue at the till for a
cashier whilst moving on to the next customer.
12
This means that currently our abstraction level is too high and we do not model the real system in an
appropriate way. We hope to be able to fix this in a later version. For now we do not consider this to be a
big problem so long as we are aware of it. We model as an exercise to gain insights into key variables and
their causes and effects and to construct reasonable arguments as to why events can or cannot occur based
on the model; we model for insights, not precise numbers.
In our experiments we have modulated the staffing levels to allow us to observe the effects of changing key
variables but we have tried to maintain the main characteristic differences between the departments (i.e.,
we still use more staff in the WW department compared to the A&TV department, only the amount has
changed).
5. Experiments
The purpose of the experiments described below is to further test the behavior of the simulation models as
they become increasingly sophisticated, rather than to investigate management practices per se. We have
defined a set of standard settings (including all probabilities, staffing levels and the customer type split) for
each department type which we use as a basis for all our experiments. The standard settings exclude the
main customer pool sizes for the different departments which are determined in the first experiment and
added to the set of standard settings for subsequent experiments. For the experiments that involved only a
static main customer pool the run length has been 10 weeks. For the experiments that involved a dynamic
main customer pool or both types the run length has been 52 weeks. We have conducted 20 replications (in
most cases) to allow the application of rigorous statistical tests.
5.1 Comparing Customer Pool Sizes
The first experiment is a sensitivity analysis. We want to find out what impact the customer pool size has
on the simulation results and which of our performance measures are affected by it. For this experiment we
have used our new noise reduction mode (described in Section 4.2.2). It allows us to focus our attention on
the impact of the variable to be investigated. We have run the experiments separately for both case study
department types using the standard settings and a pool size ranging from 2,000 to 10,000 customers in
increments of 2,000. We will interpret the results in terms of customer-related performance measures,
which can be separated out between the standard measures which are defined explicitly in terms of at which
stage of their visit a customer leaves the department, and those related to customer satisfaction and
dissatisfaction.
Hypothesis 1: We predict that varying the customer pool size will not significantly change standard
customer measures nor those customer satisfaction measures related measured in terms of experience per
visit (CSM-EPV).
Hypothesis 2: We predict that the customer satisfaction measures which accumulate historical data (CSM-
AHD) will result in significantly more customers being categorised as either ‘satisfied’ or ‘dissatisfied’
when compared to the number of customers in these categories for the customer satisfaction measure CSM-
EPV.
Hypothesis 3: We predict that the customer pool size will significantly impact on all CSM-AHD measures,
specifically the greater the customer pool size, the higher the count of neutral customers, and the lower the
count of satisfied and dissatisfied customers.
Our results in percentage terms (see Table 4) appear to provide support for all hypotheses. In order to
rigorously test our hypotheses, statistical analyses were applied to the raw customer counts as follows. To
investigate Hypothesis 1, a series of one-way between groups ANOVAs were conducted separately for
A&TV and WW. Levene’s test for homogeneity of variances was violated by CSM-EPV for satisfied
customers in A&TV only, therefore a more stringent significance value was set for this variable (p<.01).
Given that 8 simultaneous comparisons were to be drawn, a Bonferroni adjustment was made to the alpha
level, tightening the significance levels to p<.00625 (from p<.05) and p<.00125 (from p<.01). For A&TV,
no significant differences were found in the standard performance measures or CSM-EPV between
13
different customer pool sizes. For WW, no significant differences were found apart from the number of
customer leaving before receiving normal help (p=.0059). Inspecting mean differences revealed that the
mean customer count with a customer pool of 8,000 was higher than for 4,000, 6,000, and 10,000
customers. However, upon examination of supporting data, no single paired comparison was significantly
different. Across customer pools, the mean customer count varied between just 4.6 and 7.5 customers
leaving before receiving normal help, and so we can explain this pattern in terms of the relative infrequency
of this event and conclude that the difference is of little practical significance. This leads us to conclude
that overall the evidence supports Hypothesis 1. We predicted this pattern because customer decisions are
driven by customer types and the proportional mix of each type is kept constant throughout the experiment.
To test Hypothesis 2, a series of independent samples T-tests were conducted to evaluate the differences
between the two methods of measuring customer satisfaction. For A&TV, a significantly higher count of
customers were categorized as satisfied on CSM-AHD (M=18,031.18, SD=824.11) as compared to their
counterparts on the CSM-EPV measure (M=15,071.12, SD=108.47), t(102.43)=35.61, p<.000 (equal
variances not assumed). The same pattern was reflected in a significantly higher count of dissatisfied
customers (M=17,134.13, SD=895.06) on CSM-AHD as compared to CSM-EPV (M=11,404.65,
SD=221.34), t(111.06)=62.14, p<.000 (equal variances not assumed). The effect sizes are very large in both
cases, yielding .86 and .95 respectively.
For WW, a significantly higher count of customers were categorized as satisfied on CSM-AHD
(M=54,161.66, SD=3,559.94) as compared to on CSM-EPV (M=33,508.33, SD=82.76), t(99.11)=58.00,
p<.000 (equal variances not assumed). The inverse occurred for the count of dissatisfied customers,
whereby a smaller number were counted on CSM-AHD (M=3,447.49, SD=1,049.66) than CSM-EPV
(M=4,626.00, SD=128.59), t(101.97)=-11.14, p<.000 (equal variances not assumed). The effect size was
very large for A&TV at .94, and medium-to-large at .39 for WW.
Overall as expected CSM-AHD generally results in a significantly higher count of satisfied and dissatisfied
customers than CSM-EPV. This phenomenon has been discussed in Section 4.2.2. The evidence for A&TV
unequivocally supports this hypothesis, whereas it is mixed for WW. The surprisingly relatively low count
of dissatisfied of CSM-AHD in WW recurs in evaluation of the next hypothesis (see below).
Table 4. Descriptive statistics for Experiment 1 (all to 2 d.p.)
A&TV
Customer pool size
Customers …
… leaving after making a purchase
… leaving before receiving normal help
… leaving before receiving expert help
… leaving whilst waiting to pay
… leaving before f inding anything
… leaving satisf ied (accumulated historical data)
… leaving neutral (accumulated historical data)
… leaving dissatisf ied (accumulated historical data)
… leaving satisf ied (experience per visit)
… leaving neutral (experience per visit)
… leaving dissatisf ied (experience per visit)
Customer pool size
Customers …
… leaving after making a purchase
… leaving before receiving normal help
… leaving before receiving expert help
… leaving whilst waiting to pay
… leaving before f inding anything
… leaving satisf ied (accumulated historical data)
… leaving neutral (accumulated historical data)
… leaving dissatisf ied (accumulated historical data)
… leaving satisf ied (experience per visit)
… leaving neutral (experience per visit)
… leaving dissatisf ied (experience per visit)
2,000
4,000
6,000
8,000
10,000
Mean SD
Mean SD
Mean SD
Mean SD
Mean SD
29.5% 0.1% 29.5% 0.1% 29.6% 0.1% 29.5% 0.1% 29.4% 0.1%
2.7% 0.2% 2.7% 0.2% 2.8% 0.2% 2.7% 0.2% 2.8% 0.2%
1.1% 0.1% 1.1% 0.1% 1.1% 0.0% 1.1% 0.0% 1.1% 0.0%
17.5% 0.3% 17.4% 0.3% 17.3% 0.4% 17.4% 0.3% 17.4% 0.3%
49.3% 0.4% 49.3% 0.4% 49.3% 0.3% 49.3% 0.3% 49.3% 0.3%
46.5% 1.9% 45.3% 1.2% 43.8% 0.8% 42.9% 0.6% 42.0% 0.9%
8.7% 0.2% 12.1% 0.3% 14.6% 0.2% 16.5% 0.2% 18.1% 0.3%
44.8% 2.2% 42.5% 1.4% 41.6% 1.1% 40.6% 1.0% 39.9% 1.1%
36.8% 0.3% 36.9% 0.2% 36.9% 0.2% 36.8% 0.2% 36.7% 0.4%
35.3% 0.3% 35.3% 0.3% 35.3% 0.3% 35.2% 0.3% 35.2% 0.2%
27.9% 0.6% 27.8% 0.5% 27.8% 0.6% 27.9% 0.5% 28.0% 0.6%
WW
2,000
4,000
6,000
8,000
10,000
Mean SD
Mean SD
Mean SD
Mean SD
Mean SD
46.5% 0.1% 46.5% 0.1% 46.5% 0.1% 46.6% 0.1% 46.5% 0.1%
0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
0.1% 0.0% 0.1% 0.0% 0.1% 0.0% 0.1% 0.0% 0.1% 0.0%
10.0% 0.2% 10.0% 0.3% 10.0% 0.3% 10.0% 0.2% 10.0% 0.2%
43.4% 0.3% 43.4% 0.3% 43.3% 0.3% 43.3% 0.2% 43.3% 0.2%
93.4% 0.3% 88.3% 0.3% 84.0% 0.3% 80.7% 0.3% 78.0% 0.3%
3.9% 0.1% 7.2% 0.2% 10.2% 0.2% 12.6% 0.2% 14.7% 0.2%
2.7% 0.2% 4.5% 0.3% 5.8% 0.2% 6.7% 0.3% 7.3% 0.3%
52.4% 0.2% 52.5% 0.1% 52.5% 0.1% 52.5% 0.1% 52.5% 0.1%
40.3% 0.3% 40.3% 0.2% 40.2% 0.3% 40.2% 0.2% 40.2% 0.2%
7.3% 0.2% 7.2% 0.3% 7.2% 0.2% 7.2% 0.2% 7.3% 0.2%
14
To investigate Hypothesis 3, a series of one-way between groups ANOVAs were conducted. Levene’s test
for homogeneity of variances was violated by CSM-AHD for satisfied and dissatisfied customers in A&TV
and only dissatisfied customers in WW, therefore a more stringent significance value was set for these
variables (p<.01). Further to this a Bonferroni adjustment was implemented because of the multiple
comparisons, tightening the significance value to p< .0167 (from p<.05) and p<.0033 (from p<.01).
Separately for A&TV and WW, all 3 ANOVAs for each CSM-AHD variable exhibited significant
differences (p<.000). Examining mean differences for A&TV: for satisfied customers, significant
differences were present between all paired pool sizes except for between 2,000 and 4,000; 6,000 and
8,000; and 8,000 and 10,000. The mean customer count uniformly decreased as the customer pool size
increased, with significant differences focused on comparisons between paired customer pool sizes with a
greater size differential. For neutral customers, all paired comparisons were significantly different (p<.000),
and examination of mean differences reveals a uniformly increasing trend in the customer count as the
customer pool size increases. For dissatisfied customers, significant differences persisted (at p<.0033)
between the following pairs: 2,000 and all other pool sizes; 4,000 and 8,000; and 4,000 and 10,000. As
customer pool size increased, the count of dissatisfied customers uniformly decreased.
Looking at mean differences for WW: for satisfied customers, all paired comparisons were significantly
different (p<.000), and an examination of mean differences revealed a uniformly decreasing trend in the
count of satisfied customers as customer pool size increases. For neutral customers, again all paired
comparisons exhibited significant differences (p<.000), and this time the mean customer counts revealed a
uniformly increasing trend as customer pool size increases. Finally, for dissatisfied customers, all paired
comparisons were significantly different (p<.000), and the mean customer counts behaved according to a
uniformly increasing trend as customer pool size increases. These results follow a similar pattern to that
presented in A&TV, apart from the increasing number of dissatisfied customers. We expect that this
finding can be explained in terms of an accumulation of the impact of a number of subtle differences in the
ability of the department to cope with greatly increased numbers of customers (which one-by-one did not
produce significant effects, see evaluation of Hypothesis 1).
Looking at the percentages presented in Table 4, we can see that the bigger the customer pool size the more
the CSM-AHD values tend to the CSM-EPV values. This can be explained by the fact that with a larger
population the likelihood that a specific customer enters the department repeatedly (and therefore
accumulates some historical data) reduces. If we sufficiently increase the customer population we expect
both customer satisfaction measures to show similar results because the majority of customers are only
picked once during the simulation runtime and therefore most customers will enter the department with a
neutral satisfaction score rather than with an accumulated score. In summary for Hypothesis 3 we can
conclude that the evidence largely supports our prediction, with the exception of dissatisfied customer
counts in WW for which we have presented an explanation.
The results for Experiment 1 demonstrate that it is therefore important to select and maintain one customer
pool size to ensure that all performance measures are providing comparable information for different
experiments. Our case study organization does not collect or hold data on customer pool size so we have
instead calculated a suitable value based on the average numbers of customers who visit the department per
day (585 for A&TV and 915 for WW) and an estimate of customers’ average inter-arrival time (two weeks
for A&TV and one week for WW). These values have been estimated considering the standard customer
type split for the corresponding department as well as customer demand for the items sold in that
department. They do not necessarily apply to other customer type splits. Using these values we have
calculated an appropriate customer pool size for each department (8,000 for A&TV and 6,500 for WW).
We use these customer pool sizes for all subsequent experiments.
5.2 Comparing Normal and Noise Reduction Mode
In our second experiment we want to investigate the importance of considering hourly differences in
customer arrival rates and daily differences in staffing and opening hours (normal mode). These features
have been added to investigate hypotheses related to the real performance of the case study departments
and specific patterns that occur on a day-to-day basis. Modeling this level of detail can be problematic
15
when we conduct a sensitivity analysis where we want to be able to attribute causation to the introduction
of a particular variable. For this kind of experiment we prefer to control some of the system noise to clearly
see the impact of the parameter under investigation (noise reduction mode).
Hypothesis 4: We predict that the standard customer performance measures will not significantly differ
across the two modes.
Hypothesis 5: We predict the two modes will produce significantly different values when looking at the
daily performance measures on a day-to-day basis (rather than using averages).
<Insert Figure 4 somewhere here>
Table 5. Descriptive statistics for Experiment 2 (all to 2 d.p.)
Adoption fraction
Customers …
V
T
&
A
W
W
Overall customer count [per day]
… leaving after making a purchase [runtime]
… leaving before receiving normal help [runtime]
… leaving before receiving expert help [runtime]
… leaving whilst waiting to pay [runtime]
… leaving before finding anything [runtime]
Overall customer count [per day]
… leaving after making a purchase
… leaving before receiving normal help
… leaving before receiving expert help
… leaving whilst waiting to pay
… leaving before finding anything
T-Test
AF = 1
AF = 0.5
AF = 0
t-value p-value Eta 2
Mean
Mean
Mean
SD
SD
SD
-7.10
0.00
0.04
29.09
609.38
8.27
601.09
6.04
584.17
0.00
0.99
-0.01
50.29
46.75 12,063.05
36.97 12,094.30
12,062.95
0.14
0.00
-21.39
99.01
1,706.55
63.92
1,343.30
67.36
1,133.80
0.00
0.87
0.17
463.55
457.00
464.65
22.27
18.25
18.47
0.02
0.00
-16.05
7,048.70 105.59
7,633.60 124.13
7,520.30 116.47
0.00
0.00
-12.25
20,182.00 134.56 20,661.65 145.08 20,789.75 176.33
-65.00
0.00
0.66
910.58
6.25
1,093.24
22.44
1,224.79
39.98
0.03
0.00
-39.43
45.29
31.34 30,256.45
44.48 30,097.55
29,696.75
0.81
0.00
-19.66
17.71
84.60
7.24
29.55
3.30
5.40
0.67
0.00
83.60
-22.68
9.04
130.75
12.48
170.40
14.53
0.68
0.00
6,291.95 165.73 13,148.10 200.16 18,050.20 164.96 -224.88
27,662.80 166.79 33,120.80 198.49 37,173.90 196.79 -164.89
0.00
0.30
Looking at Figure 4 it is clear that there is very little variation between the two different operation modes
across the range of standard performance measures. Examining the descriptives (see Table 5), in most cases
the runtime performance measures in both modes are approximately the same with a small number of
exceptions. Contrary to hypothesis 4, in WW approximately 25% more customers leave whilst waiting to
pay in the normal mode (which consequently influences both customer satisfaction measures), as opposed
to the noise reduction mode. This apparent cashier bottleneck appears to be exacerbated by any
combination of the three factors which are held constant in noise reduction mode. Further analysis is
required to isolate the precise cause. We can also see that, as predicted, the smaller the values the more they
differ (on an absolute basis) between the two modes as they are accumulated from a smaller number of
events and therefore the influence of different random number streams is more apparent.
<Insert Figure 5 somewhere here>
Examining the daily measures on a day-to-day basis for A&TV in Figure 5, we can observe clear
differentiation between the number of customers and transactions across weekdays, Saturdays and Sundays
in normal mode whereas the noise reduction mode, as expected, shows no clear patterns. This additional
information can be very useful for optimizing the system. For example, we have the lowest number of
transactions on Sundays (day 1, 8, etc.) although we do not have the lowest number of customers on this
day, therefore we must have a problem with the optimizing the staffing arrangement on Sundays, because
on average more customers leave the shop on Sundays without buying anything, despite the fact that the
probability of this remains the same for all days of the week.
Overall the experiment has shown that it is legitimate to use the noise reduction mode within the scope of a
sensitivity analysis to test the impact of a specific factor on system behavior. It is only when we are
interested in specific features that are not available in noise reduction mode (e.g. when we want to study
differences between particular days of the week or when we need to obtain quantitative data to optimize a
real system) that we need to choose the normal mode.
5.3 Evaluation the Implementation of WOM
16
In the first part of this experiment we want to investigate how customer satisfaction can be influenced by
using different ways of representing inter-linkages within a customer population. We compare increasingly
sophisticated forms of implementing WOM. Firstly we have used a static customer pool where customer
shopping behavior is random (i.e. not influenced by customer’s word of mouth). Secondly, we have
introduced a static customer pool where shopping behavior is influenced by word of mouth. Finally, we
employ a dynamic customer pool to allow us to investigate growth (or reduction) in the size of the overall
customer pool (or customer base).
We have chosen a runtime of 52 weeks because we expect to see model developments further into the
model runtime, in particular for the dynamic WOM implementation. For this experiment we have used our
new noise reduction mode (described in Section 4.2.2) and the following WOM settings: an adoption
fraction of 0.5 and contact rate of 2. We expect there to be a link between customers’ shopping behavior
and customer satisfaction scores, moderated by the type of retail department (i.e. an interaction effect).
Specifically, we predict that:
Hypothesis 6: In the dynamic conditions, customer pool size will increase during the simulation run for
each department up to a maximum value, and maintain this level for the run duration.
Hypothesis 7: Average daily customer numbers will be significantly different for each of the 6 WOM
conditions. In particular, we predict that WW will experience a higher number of customer visits than
A&TV, the more advanced the WOM implementation the higher the number of customer visits, and we
expect to observe a positive moderating effect whereby an interaction between the two can explain further
increases in average daily customer numbers.
Hypothesis 8: Within each department the average number of satisfied customers will significantly
decrease, whereas the average number of dissatisfied customers will significantly increase (both CSM-
EPV), with the sophistication of the WOM implementation.
The results are presented in multiple forms to allow detailed analysis. Figures 6 to 9 show the results for a
single simulation run on a day to day basis, which are integral to revealing day-to-day fluctuations. Due to
concerns about the quality of presentation, this will be restricted to the first 26 weeks of each simulation
run. The authors have examined graphs for the full runtime period and variables’ trends continued as would
be expected based on the first 26 weeks. Table 6 shows the descriptive statistics of the average results of
the experiment; tabulated values are daily averages across all replications.
Regarding Hypothesis 6, Figure 6 clearly displays that under dynamic conditions the customer pool size
grows over the early stages of the simulation until it reaches a maximum value and then stabilizes around
this level (for averages see Table 6). Interestingly, the graph displays that WW reaches a notably higher
value than A&TV, which we can understand in terms of the tighter staffing constraints in A&TV restricting
the availability of staff to serve customers (and hence constrain the growth of the customer population). In
WW, the growth in customer numbers can be alikened to the diffusion of customers’ word-of-mouth in the
real world; as the numbers grow, the rate of growth also increases. In contrast, the growth in customer
numbers triggered by WOM in A&TV has a steadily slowing rate of growth, which may be related to the
department’s limited capacity to satisfy growing customer numbers.
Table 6. Descriptive statistics for Experiment 3 (all to 2 d.p.)
17
A&TV
static pool, no WOM
Mean
SD
62,679.65
4,727.70
2,245.15
37,010.95
88.11
105.62
34.55
195.20
static pool, WOM
Mean
SD
63,224.00
10,760.25
2,635.05
47,887.25
69.60
134.67
58.72
270.52
dynam ic pool, WOM
Mean
SD
63,711.70
25,787.85
2,969.45
57,511.75
98.50
273.77
55.50
426.80
105,929.80
311.08
119,330.35
268.35
144,318.30
153,231.20
17,211.90
73,393.80
76,344.00
122,440.55
45,052.35
243,836.90
30.50
722.13
240.75
803.15
171.84
333.06
232.58
286.48
5.28
124,970.40
24,612.60
144,716.05
69,782.75
156,304.80
68,211.50
294,299.05
3.03
629.55
971.28
267.97
851.11
130.40
640.76
213.43
750.04
0.70
165,413.55
13,501.05
33,678.65
81,322.90
100,754.15
30,516.20
212,593.25
657.24
209.10
740.80
219.71
430.03
205.27
447.40
5.16
26.80
WW
static pool, no WOM
Mean
SD
152,433.10
0.40
112.15
26,591.25
131.28
0.60
11.87
400.65
static pool, WOM
Mean
SD
156,100.00
2.40
217.80
83,259.55
83.91
1.93
13.26
361.07
dynam ic pool, WOM
Mean
SD
159,585.35
17,590.80
1,870.45
199,493.30
91.58
600.72
35.82
1,027.04
152,807.05
318,549.55
8,495.55
4,898.85
164,114.20
146,188.90
21,640.85
331,943.95
51.10
432.68
629.39
159.34
275.81
151.98
435.49
342.18
601.55
7.23
204,547.10
409,429.25
12,185.85
22,511.75
180,283.15
196,015.65
67,828.05
444,126.85
68.50
601.35
896.42
144.87
686.50
183.70
594.29
277.59
616.05
7.27
373,335.45
1,121.25
382,430.25
41,828.45
2,072.47
571.67
327,616.65
1,494.42
175,356.25
405,963.70
404.53
1,179.03
170,555.40
309.05
751,875.35
1,296.40
5.87
0.48
Customer pool type
Customers …
… leaving after making a purchase
… leaving before receiving normal help
… leaving before receiving expert help
… leaving whilst waiting to pay
… leaving before f inding anything
… leaving satisf ied (accumulated historical data)
… leaving neutral (accumulated historical data)
… leaving dissatisfied (accumulated historical data)
… leaving satisf ied (experience per visit)
… leaving neutral (experience per visit)
… leaving dissatisfied (experience per visit)
Overall no. of customers
Average no. of visits per customer
Customer pool type
Customers …
… leaving after making a purchase
… leaving before receiving normal help
… leaving before receiving expert help
… leaving whilst waiting to pay
… leaving before f inding anything
… leaving satisf ied (accumulated historical data)
… leaving neutral (accumulated historical data)
… leaving dissatisfied (accumulated historical data)
… leaving satisf ied (experience per visit)
… leaving neutral (experience per visit)
… leaving dissatisfied (experience per visit)
Overall no. of customers
Average no. of visits per customer
A two-way between groups ANOVA was conducted to test Hypothesis 7. Levene’s test for homogeneity of
variances was violated, and so we introduced a more stringent significance level (p<.01). The ANOVA
revealed a statistically significant main effect for department type [F(1,114)=3,698,165, p=.000] and WOM
implementation [F(2,114)=1,226,034, p=.000]. The effect sizes were extremely high in both cases (partial
eta-squared = 1.00). Examining mean differences by department type, WW’s customer count was
significantly higher than for A&TV. Tukey’s post-hoc comparisons were applied to the WOM
implementation variable, and all paired comparisons were significant. Results indicated that the more
sophisticated the WOM implementation, the higher the customer count. There was also a significant
interaction effect between the two independent variables [F(2,114)=572858.8, p=.000] with a very high
effect size (partial eta-squared = 1.00). This finding suggests that department type moderates the effect of
the WOM implementation, resulting in increasingly higher customer counts for WW (as compared to
A&TV) with increases in the sophistication of WOM implementation. These results provide support for
Hypothesis 7.
Differences in the daily count of customers visiting the department are presented in Figure 7. We expected
these relationships because WW has an inherently greater capacity to serve an increased number of
customers using the same amount of resources due to the relatively low service requirement of the average
WW customer. Although A&TV has some scope to increase its customer base, customer requirements for
advice and attention constraint the department’s ability to deal with an increased volume of customers. In
terms of different WOM implementations, the more sophisticated the implementation, the more
opportunities that customers have to spread the word and encourage other customers to visit. Consequently
we expected that more sophisticated versions would continue to grow customer numbers, and the
interaction with department type builds on the arguments already put forward.
It appears that the restricted availability of staff to serve customers is preventing a commensurate increase
in sales. Key bottlenecks occur in A&TV with the provision of normal help and cashier availability to take
payment, and the bottlenecks in WW occur with the same variables and also the provision of expert help
(see Table 6). All of these bottlenecks result in a significant rise in customers who leave without finding
anything to buy in either department.
<Insert Figure 6 somewhere here>
18
<Insert Figure 7 somewhere here>
Daily differences in the daily count of CSM-EPV satisfied and dissatisfied customers visiting the
department are presented in Figures 8 and 9. A series of one-way between-groups ANOVAs were used to
test Hypothesis 8 and evaluate the impact of different WOM implementations on average measures of
CSM-EPV. A Bernoulli correction was applied to tighten the significance value to p<.025. For A&TV,
tests revealed a statistically significant difference between the three WOM implementations for the number
of satisfied customers [F(2, 59)=21,201.95, p<.000] and the number of dissatisfied customers [F(2,
59)=152,952.60, p<.000]. The effect sizes, calculated using eta-squared, were both very high at 1.00.
Tukey’s post-hoc comparisons identified significant differences between every single paired comparison in
the directions predicted (p<.000).
For WW, Levene’s test for homogeneity of variances was violated for satisfied customers, and so the p-
value for this variable was further restricted to p<.005. The ANOVA indicated a statistically significant
difference between the three WOM implementations for the number of satisfied customers [F(2,
59)=18,689.89, p<.000] and the number of dissatisfied customers [F(2, 59)=1,203,536, p<.000]. The effect
sizes were very high (eta-squared 1.00 for both). Tukey’s post hoc tests revealed significant differences
between all paired comparisons (p<.000). All were in the predicted direction, with the exception of a large
increase in the number of satisfied customers between the first and second implementations of WOM. This
exception may be attributable to some slack in the WW staffing being taken up as the level of WOM
sophistication (and hence potential customer pool) increased. However, with further WOM developments
(dynamic customer pool) the staffing constraints start to take effect and the number of satisfied customers
decreases. Looking at Figure 8, it is clear that on a day-to-day basis there are great fluctuations in customer
satisfaction. In Figure 9, again much day-to-day variation can be seen, and the significant increase in the
average number of dissatisfied customers in the dynamic WW model can be clearly observed.
In general, Hypothesis 8 has been supported, in that evidence suggests that the average number of satisfied
customers (CSM-EPV) will decrease with a more sophisticated WOM implementation, and that the number
of dissatisfied customers will increase (CSM-EPV). The exceptional behavior of the count of satisfied
customers in WW has been explained in terms of slack in the staffing of the department, and with further
increases to customer number we would expect the corresponding decrease in the number of satisfied
customers to continue.
<Insert Figure 8 somewhere here>
<Insert Figure 9 somewhere here>
In the second part of the experiment we want to explore daily fluctuations in customer numbers. We want
to find out if there is a set threshold value for WOM, beyond which we would on the next day create an
empty department (and hence cause its permanent closure). This could be triggered by the great variation in
daily fluctuations, whereby a very high amount of satisfied customers causes an overwhelming volume of
customers on the subsequent day. This could result in crucially high levels of dissatisfaction and result in
no customers visiting the department on the day after.
For this experiment we have used our new noise reduction mode (described in Section 4.2.2). We vary the
adoption fraction using a step interval of 0.2 whilst keeping the contact rate constant at 5. The high contact
rate value has been selected to reach the upper boundary of customer numbers in a reasonable amount of
trials.
Hypothesis 9: We predict that the higher the WOM adoption fraction, the greater the increase and
corresponding subsequent decrease in the number of customer visits. We expect this relationship to be
linked to a reduction in satisfaction per customer. We hypothesize that this pattern will vary over time
because the impact of WOM will vary on a daily basis.
19
Hypothesis 10: We predict that there is a threshold upon which the daily number of customers shrinks to 0
and consequently the department ceases trading. We expect that this threshold will occur at an earlier stage
for A&TV than WW.
Given that we are looking at daily differences, it is important to examine what is going on graphically.
Therefore hypotheses will be tested through evaluation of simulation output in terms of daily fluctuations.
Comparing average values dampens this variability which we are interested in (due to model stochasticity),
and so we will be assess hypotheses using figures and not statistical tests.
Figure 10 presents the customer count relating to different adoption fractions across WW and A&TV (static
model). Comparing the two departments, a greater increase in customer numbers can be seen between
experimental conditions in WW than in A&TV as the customer pool grows, which is to be expected. This is
because in addition to the customer service limitations of A&TV, the starting customer pool size in WW is
23% greater than that of A&TV.
Assessing Hypothesis 9, Figures 10 (and also Figures 11 and 12) suggests that the higher the adoption
fraction the greater the day-to-day fluctuations in customer numbers. Each individual line tends to vary a
roughly equal amount above and below its middle, adding weight to the hypothesis that each increase in
customer numbers is accompanied by a subsequent decrease. Further evidence supporting this hypothesis
is covered in assessment of Hypothesis 10.
Evaluating Hypothesis 10, we can see in Figures 10, 11 and 12 that the results for A&TV suggest that it is
not able to meet the increased customer demands which the highest adoption fraction (0.6) places on it. The
overall shapes of the extreme cases (adoption fraction 0.6) behave in a pronounced cyclic manner for both
static and dynamic WOM implementations, with relatively large customer peaks and troughs compared to
the other adoption fractions in A&TV. Due to a day with no customers, the A&TV department ceases
business. This does appear to be caused by a sharp increase in one day’s customers triggering a sharp
decrease in the following day’s customer numbers. We predicted this relationship because the department’s
staffing resources remain the same regardless of customer pool growth, and therefore when the daily
customer pool grows, the same amount of staffing resources are shared between a larger group of
customers. The lower adoption fractions (0.2 and 0.4) result in a relatively restricted variation in customer
numbers after approximately 6 weeks in the static implementation, and 15 weeks in the dynamic
implementation. It is interesting to compare adoption fractions of 0.4 and 0.6 for the static model because
the 0.6 model only attracts a few more customers to trigger its downfall, whereas 0.4 recovers from this
near-miss.
Looking at WW in Figures 10 and 11, there are great variations in customer numbers in particular for the
highest adoption fraction (0.6) where we can see areas of pronounced oscillating peaks and troughs
alternating with less pronounced areas of daily fluctuations. For the dynamic model (Figure 12) at 50 weeks
this adoption fraction results in no customers. Given the large amount of variation in customer numbers
throughout this simulation, it is not surprising that this phenomenon occurs. Also as observed for A&TV,
all adoption fractions produce phases of relatively intense variation in customer numbers alternate with
relatively stable phases.
The figures presented support Hypothesis 10 in principle. A&TV the static WOM implementation resulted
in zero customers at week 3, whereas in WW the dynamic WOM implementation resulted in zero
customers at week 50. This provides some supports the prediction that this would happen at an earlier stage
in A&TV than in WW. Further experimentation (not documented here) investigated for each department
types we could not finish the simulation run for an adoption fraction value of 0.6 (dynamic pool). This
work identified that the boundary adoption fraction to allow the experiment to finish (with the complete
runtime of 52 weeks) occurred for the static implementation with a value of up to 0.500 for A&TV, and
0.791 for WW. The higher threshold for WW concords with what we have seen in daily differences. For
the dynamic implementation, the adoption fraction can have a value of up to 0.485 for A&TV and up to
0.599 for WW. Any adoption fraction values beyond these result in a premature termination of the
department, due to no customers.
20
In summary, we can conclude that comparing and contrasting different implementations of WOM can
provide fruitful insights into department performance. The current models help us to better understand
system outcomes where there is a finite staffing resource in a particular department, and the initial customer
demands are already ensuring there is little slack in the system (NB we observed this to a greater extent
with A&TV than for WW). Investigating a different scenario such as when a retailer has the investment to
introduce further resource with increasing customer demand would require a different configuration.
It is clear that the current implementation of a dynamic customer pool results in WOM only directly
affecting the growth or decline of the customer pool population on the following day. This means that the
behavior of customer count variable results in an unbelievable variation in day-to-day customer figures. In
the real world we would expect word of mouth to impact on an individual’s behavior with a short delay,
perhaps of a couple of days, or until the next opportunity that individual has to visit the department. Not
everyone runs to the shop immediately when he or she has received a recommendation, (NB we do not
consider that some customers can only shop on certain days of the week, or at certain times of the day), and
even if this did happen it is unlikely that a customer passing on WOM would recommend the department to
everyone immediately. There is a real delay in the spread of the WOM. Our findings so far suggest that it
would be better modeled using a positively skewed distribution showing a large immediate effect and also
an ongoing subtler effect which would be strong the more that WOM has accumulated. We will test this in
the next version of our simulation model.
<Insert Figure 10 somewhere here>
<Insert Figure 11 somewhere here>
<Insert Figure 12 somewhere here>
6. Conclusion
We have presented the conceptual design, implementation and operation of simulation models to help
inform the way we which we can understand the impact of people management practices on retail
performance. As far as we are aware this is the first time researchers have tried to use an agent-based
approach to simulate people management practices such as training and empowerment. Although our
simulation model uses specific case studies as source of information, we believe that the general model
could be adapted to other retail companies and areas of management practices that have a lot of human
interaction.
To respond to our key research questions, we have modeled and simulated customer satisfaction and the
effects of WOM diffusion on customer satisfaction. We have compared and analyzed data resulting from
two contrasting methods of modeling customer satisfaction: CSM-AHD and CSM-EPV. Likewise, we have
adopted a two-pronged approach to customers’ WOM and present comparative analysis of different
strategies to implement the WOM (using a static and a dynamic main customer pool). The evaluation of
daily trends has proved particularly valuable in understanding what causes a department to run out of
customers, and further model development is planned to implement a different approach to this.
In this paper we have focused in particular on the capabilities required to model customer evolution as a
consequence of the implementation of people management practices. We have discussed conceptual ideas
about how to consider external and internal stimuli and have presented an implementation of WOM as one
form of external stimuli. We are still testing and calibrating the new features we have implemented. The
validation experiments so far have shown that we need to improve our simulation model in order to be able
to model the real system in an appropriate way. In particular our current abstraction level with regards to
how staff spend their time is much too high. If we want to use real staffing data we need to model how staff
allocate their tasks between competing activities rather than focusing on one type of work (i.e. make our
staff members proactive). We have already modeled one form of proactive staff behavior (empowerment)
but that was only for specific experiments, e.g. (1) a normal sales staff member decides by himself or
herself whether to stay or not to stay with a customer when additional expert advice is given which allows
them to learn on the job if they wish to do so, or (2) an expert staff member helps out in the provision of
21
normal advice if no customer requires expert help (see [11]). Further development such as self-initiated till
manning needs to be implemented in order to mimic the behavior of the staff members in the real system in
a more appropriate way.
In addition to continuing to validate our current simulation model we have also planned to experiment with
different strategies of modeling customer evolution. We plan to implement and test our conceptual ideas
regarding customers’ memory of their own shopping experiences (internal stimuli). Furthermore, we have
planned to enhance the flexibility of staff members (i.e. empower them) to allow them to respond to
customer demand. This will help to solve the staffing problem we have discussed above. In the long term
we want to develop our simulation model to support testing the impact of team work related management
practices. This looks like an interesting but challenging task because we first need to come up with a way
to represent the effects of team work. Furthermore, we would like to enhance the capabilities of our agents,
giving them skills in reasoning, negotiation, and co-operation.
In conclusion, we can say that the new features appear promising and we are convinced they have already
improved our insights into the operation of the departments within a department store. In particular the new
performance measures we collect on a daily basis will be very useful in future for balancing the provision
of customer services throughout the week.
Overall, we believe that researchers should become more involved in this multi-disciplinary kind of work
to gain new insights into the behavior of organizations. In our view, the main benefit from adopting this
approach is the improved understanding of and debate about a problem domain. The very nature of the
methods involved forces researchers to be explicit about the rules underlying behavior and to think in new
ways about them. As a result, we have brought work psychology and agent-based modeling closer together
to form a new and exciting research area.
7. Acknowledgements
This work is part of an IDEAS Factory Network Project funded by the UK government through the
Engineering and Physical Sciences Research Council (EPSRC) under grant EP/D503949/1. We would also
like to thank the reviewers for their fruitful comments on earlier drafts of this article.
8. References
[1] Department of Trade and Industry. 2003. UK productivity and competitiveness indicators. DTI
Economics Paper No. 6. London, UK: DTI. http://www.berr.gov.uk/files/file14767.pdf
[2] Reynolds, J., E. Howard, D. Dragun, B. Rosewell, and P. Ormerod. 2005. Assessing the productivity of
the UK retail sector. International Review of Retail, Distribution and Consumer Research 15(3):237-
280.
[3] Delbridge, R., P. Edwards, J. Forth, P. Miskell, and J. Payne. 2006. The organisation of productivity:
Re-thinking skills and work organisation. London, UK: Advanced Institute of Management Research.
http://www.aimresearch.org/publications/orgprod.pdf.
[4] Siebers, P.O. et al. 2008. The role of management practices in closing the productivity gap. AIM
Working Paper Series 065. London, UK: Advanced
Institute of Management Research.
http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1309605.
[5] Wall, T.D. and S.J. Wood. 2005. Romance of human resource management and business performance
and the case for big science. Human Relations 58(5):429-462.
[6] Pourdehnad, J., K. Maani, and H. Sedehi. 2002. System dynamics and intelligent agent-based
simulation: Where is the synergy? In Proceedings of the 20th International Conference of the System
Dynamics Society, Palermo, Italy.
[7] Birdi, K. et al. 2008. The impact of human resource and operational management practices on company
productivity: A longitudinal study. Personnel Psychology (in press).
22
[8] Patel, S. and A. Schlijper. 2004. Models of consumer behaviour. 49th European Study Group with
Industry (ESGI 2004), Oxford, UK.
[9] Nicholson, M., I. Clarke, and M. Blakemore. 2002. One brand, three ways to shop: Situational variables
and multichannel consumer behaviour. International Review of Retail, Distribution and Consumer
Research 12:131-148.
[10] Keh, H.T., S. Chu, and J. Xu. 2006. Efficiency, effectiveness and productivity of marketing in
services. European Journal of Operational Research 170(1):265-276.
[11] Siebers, P.O., U. Aickelin, H. Celia, and C. Clegg. 2007a. A multi-agent simulation of retail
management practices. In Proceedings of the Summer Computer Simulation Conference (SCSC 2007),
San Diego, USA, pp 959-966.
[12] Siebers, P.O., U. Aickelin, H. Celia, and C. Clegg. 2007b. Using intelligent agents to understand
management practices and retail productivity. In Proceedings of the Winter Simulation Conference
(WSC 2007), Washington DC, USA, pp 2212-2220.
[13] Hansen, T. 2005. Perspectives on consumer decision making: An integrated approach. Journal of
Consumer Behaviour 4(6):420-437.
[14] Tai, S.H.C. and A.M.C. Fung. 1997. Application of an environmental psychology model to in-store
buying behaviour. International Review of Retail, Distribution and Consumer Research 7(4):311-337.
[15] Youn, S. and R.J. Faber. 2000. Impulse buying: Its relation to personality traits and cues. Advances in
Consumer Research 27(1):179-185.
[16] Robinson, S. 2004. Simulation: The practice of model development and use. Chichester, UK: John
Wiley & Sons
[17] Greasley, A. 2005. Using DEA and simulation in guiding operating units to improved performance.
Journal of the Operational Research Society 56(6):727-731.
[18] Schwaiger, A. and B. Stahmer. 2003. SimMarket: Multi-agent based customer simulation and decision
support for category management. In Applying Agents for Engineering of Industrial Automation
Systems, edited by M. Shillo et al., LNAI 2831. Berlin: Springer.
[19] Simon, H.A. 1996. The sciences of the artificial (3rd Ed.). Cambridge, MA: MIT Press.
[20] Gilbert, N. and K.G. Troitzsch. 1999. Simulation for the social scientist. Open University Press, UK:
Buckinghamshire.
[21] Law, A.M., and W.D. Kelton. 1991. Simulation modeling and analysis (2nd Ed.). New York, NY:
McGraw-Hill
[22] North, M.J. and C.M. Macal. 2007. Managing business complexity: discovering strategic solutions
with agent-based modeling and simulation. New York, NY: Oxford University Press.
[23] Garrido, J.M. 2001. Object-oriented discrete-event simulation with java: A practical introduction.
New York, NY: Kluwer Academic Publishers
[24] Borshchev, A. and A. Filippov. 2004. From system dynamics and discrete event to practical agent
based modeling: reasons, techniques, tools. In Proceedings of the 22nd International Conference of the
System Dynamics Society, Oxford, UK.
[25] Epstein, J.M. and R. Axtell. 1996. Growing artificial societies: Social science from the bottom up.
Cambridge, MA: MIT Press.
[26] Jeffrey, R. 2008. Expert voice:
Icosystem's Eric Bonabeau on agent-based modeling.
http://www.cioinsight.com/article2/0,3959,1124316,00.asp.
[27] Luck, M., P. McBurney, O. Shehory, and S. Willmott. 2005. Agent technology: Computing as
interaction (a roadmap for agent based computing). Liverpool, UK: AgentLink.
23
[28] ZDNet. 2000. ZDNet Complexity Digest 2000.10: The Sims and agent based modeling.
http://www.comdig.com/index.php?id_issue=2000.10.
[29] BBC. 2008. BBC h2g2: Computer animation. http://www.bbc.co.uk/dna/h2g2/A342 1045/.
[30] Hood, L. 1998. Agent-based modeling. In Conference Proceedings: Greenhouse Beyond Kyoto,
Issues, Opportunities and Challenges, Canberra, Australia.
[31] Macal, C.M. and M.J. North. 2007. Agent-based modeling and simulation: Desktop ABMS, In
Proceedings of the Winter Simulation Conference (WSC 2007), Washington, USA, pp95-106.
[32] Carley, K.M. 1996. Validating computational models. CASOS Working Paper. Pittsburgh, PA:
Carnegie Mellon University.
[33] Fagiolo, G., P. Windrum, and A. Moneta. 2006. Empirical validation of agent-based models: A critical
survey. LEM Working Paper 2006/14. Pisa, Italy: Sant’Anna School of Advanced Studies.
[34] Leombruni, R., M. Richiardi, N.J. Saam, and M. Sonnessa. 2006. A common protocol for agent-based
social simulation. Journal of Artificial Societies and Social Simulation 9(1):15.
[35] Midgley D.F., R.E. Marks, and D. Kunchamwar. 2007. The building and assurance of agent-based
models: An example and challenge to the field. Journal of Business Research 60(8):884-893.
[36] Moss, S. and B. Edmonds. 2005. Sociology and simulation: Statistical and quantitative cross-
validation. American Journal of Sociology 110(4):1095-1131.
[37] Twomey, P. and R. Cadman. 2002. Agent-based modelling of customer behaviour in the telecoms and
media markets. info 4:56-63.
[38] Chick, S.E. 2006. Six ways to improve a simulation analysis. Journal of Simulation 1:21-28.
[39] Celia, H. 2007. Retail management practices and performance: On the shop floor. Master Thesis.
University of the West of England.
[40] Shannon, R.E. 1975. Systems simulation: The art and science. Englewood Cliffs, NJ: Prentice-Hall.
[41] Yi, Y. 1990. A critical review of consumer satisfaction, In Review of Marketing, edited by V.A.
Zeithaml. Chicago, IL: American Marketing Association.
[42] Fornell, C., M.D. Johnson, E.W. Anderson, J. Cha, and B.E. Bryant. 1996. The American customer
satisfaction index: Nature, purpose, and findings. Journal of Marketing 60(4):7-18.
[43] Ji, M.F. and W. Wood. 2007. Purchase and consumption habits: Not necessarily what you intend.
Journal of Consumer Psychology 17(4):261-276.
[44] Simon, F. and J.C. Usunier. 2007. Cognitive, demographic and situational determinants of service
customer preference for personnel-in-contact over self-service technology’. International Journal of
Research in Marketing 24(2):163-173.
[45] Marsden, P., A. Samson, and N. Upton. 2005. The economics of buzz - word of mouth drives business
growth
finds
LSE
study.
http://www.lse.ac.uk/collections/pressAndInformationOffice/newsAndEvents/archives/2005/Word_of
Mouth.htm.
[46] XJ Technologies. 2008. Official AnyLogic website. http://www.xjtek .com/.
[47] Siebers, P.O., U. Aickelin, H. Celia, and C. Clegg. 2007c. Understanding retail productivity by
simulating management practices. In Proceedings of the EUROSIM Congress on Modelling and
Simulation (EUROSIM 2007), Ljubljana, Slovenia.
[48] Rank, J., J.M. Carsten, J. Unger, and P.E. Spector. 2007. Proactive customer service performance -
relationships with individual, task, and leadership variables. Journal of Human Performance
20(4):363-390
24
Peer-Olaf Siebers is a research fellow in computer science at The University of Nottingham (UK), where
he works in the Intelligent Modelling & Analysis Group.
Uwe Aickelin is a full Professor at The University of Nottingham (UK), where he leads the Intelligent
Modelling & Analysis Group.
Helen Celia is a visiting researcher at the Centre for Organisational Strategy, Learning & Change at Leeds
University Business School (UK).
Chris W. Clegg is Professor of Organisational Psychology and Director of the Centre for Socio-Technical
Systems Design at Leeds University Business School (UK).
25
Figures:
Figure 1. Initial ideas for the simulation model and its components
Figure 2. Conceptual model of our customer agents (transition rules have been omitted for
clarity)
26
Figure 3. Customer (left) and staff (right) agent logic implementation in AnyLogic™
27
t
n
u
o
C
r
e
m
o
t
s
u
C
35000
30000
25000
20000
15000
10000
5000
0
e
s
a
h
c
r
u
p
r
e
t
f
A
p
l
e
h
l
a
m
r
o
N
p
l
e
h
t
r
e
p
x
E
A&TV
y
a
p
o
t
g
n
i
t
i
a
W
g
n
i
d
n
i
f
e
r
o
f
e
B
g
n
i
h
t
y
n
a
e
s
a
h
c
r
u
p
r
e
t
f
A
p
l
e
h
l
a
m
r
o
N
Performance Measure
Noise Reduction Mode Normal Mode
p
l
e
h
t
r
e
p
x
E
WW
y
a
p
o
t
g
n
i
t
i
a
W
g
n
i
d
n
i
f
e
r
o
f
e
B
g
n
i
h
t
y
n
a
Figure 4. Average performance measures, by department and operation mode.
s
r
e
m
o
t
s
u
c
f
o
r
e
b
m
u
n
1200
1000
800
600
400
200
0
1 2 3 4 5 6 7 8 9
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
1
0
2
1
2
2
2
3
2
4
2
5
2
6
2
7
2
8
2
9
2
0
3
1
3
2
3
3
3
4
3
5
3
6
3
7
3
8
3
9
3
0
4
1
4
2
4
3
4
4
4
5
4
6
4
7
4
8
4
9
4
0
5
1
5
2
5
3
5
4
5
5
5
6
5
7
5
8
5
9
5
0
6
1
6
2
6
3
6
4
6
5
6
6
6
7
6
8
6
9
6
0
7
day
Noise reduction mode: overal l customers
Noise reduction mode: transactions
Normal mode: overall customers
Normal mode: transactions
Figure 5. Daily customer count and number of transactions in A&TV.
28
l
o
o
p
n
i
s
r
e
m
o
t
s
u
c
f
o
r
e
b
m
u
n
16000
14000
12000
10000
8000
6000
4000
2000
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV; static pool; no WOM
A&TV; static pool; WOM
A&TV; dynamic pool; WOM
WW ; static pool; no WOM
WW ; static pool; WOM
WW ; dynamic pool; WOM
Figure 6. Customer pool growth over time.
]
y
a
d
r
e
p
[
s
r
e
m
o
t
s
u
c
2400
2200
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV; static pool; no WOM
A&TV; static pool; WOM
A&TV; dynamic pool; WOM
WW ; static pool; no WOM
WW ; static pool; WOM
WW ; dynamic pool; WOM
Figure 7. Daily count of customers entering the department.
29
]
y
a
d
r
e
p
[
s
r
e
m
o
t
s
u
c
600
500
400
300
200
100
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV; static pool; no WOM
A&TV; static pool; WOM
A&TV; dynamic pool; WOM
WW ; static pool; no WOM
WW ; static pool; WOM
WW ; dynamic pool; WOM
Figure 8. Daily count of satisfied customers CSM-EPV.
]
y
a
d
r
e
p
[
s
r
e
m
o
t
s
u
c
600
500
400
300
200
100
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV; static pool; no WOM
A&TV; static pool; WOM
A&TV; dynamic pool; WOM
WW ; static pool; no WOM
WW ; static pool; WOM
WW ; dynamic pool; WOM
Figure 9. Daily count of dissatisfied customers CSM-EPV.
30
]
y
a
d
r
e
p
[
s
r
e
m
o
t
s
u
c
3600
3400
3200
3000
2800
2600
2400
2200
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV af=0.2
A&TV af=0.4
A&TV af=0.6
WW af=0.2
WW af=0.4
WW af=0.6
WW af=0.8
Figure 10. Static pool - daily customer count.
]
y
a
d
r
e
p
[
s
r
e
m
o
t
s
u
c
3600
3400
3200
3000
2800
2600
2400
2200
2000
1800
1600
1400
1200
1000
800
600
400
200
0
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 shown due to space restriction)]
A&TV af=0.2
A&TV af=0.4
A&TV af=0.6
WW af=0.2
WW af=0.4
WW af=0.6
Figure 11. Dynamic pool - daily customer count.
31
l
o
o
p
n
i
s
r
e
m
o
t
s
u
c
f
o
r
e
b
m
u
n
18000
17000
16000
15000
14000
13000
12000
11000
10000
9000
8000
7000
6000
1
k
e
e
W
n
i
g
e
B
2
k
e
e
W
n
i
g
e
B
3
k
e
e
W
n
i
g
e
B
4
k
e
e
W
n
i
g
e
B
5
k
e
e
W
n
i
g
e
B
6
k
e
e
W
n
i
g
e
B
7
k
e
e
W
n
i
g
e
B
8
k
e
e
W
n
i
g
e
B
9
k
e
e
W
n
i
g
e
B
0
1
k
e
e
W
n
i
g
e
B
1
1
k
e
e
W
n
i
g
e
B
2
1
k
e
e
W
n
i
g
e
B
3
1
k
e
e
W
n
i
g
e
B
4
1
k
e
e
W
n
i
g
e
B
5
1
k
e
e
W
n
i
g
e
B
6
1
k
e
e
W
n
i
g
e
B
7
1
k
e
e
W
n
i
g
e
B
8
1
k
e
e
W
n
i
g
e
B
9
1
k
e
e
W
n
i
g
e
B
0
2
k
e
e
W
n
i
g
e
B
1
2
k
e
e
W
n
i
g
e
B
2
2
k
e
e
W
n
i
g
e
B
3
2
k
e
e
W
n
i
g
e
B
4
2
k
e
e
W
n
i
g
e
B
5
2
k
e
e
W
n
i
g
e
B
6
2
k
e
e
W
n
i
g
e
B
time [days (only the first 26 weeks shown due to space restrictions)]
A&TV af=0.2
A&TV af=0.4
A&TV af=0.6
WW af=0.2
WW af=0.4
WW af=0.6
Figure 12. Dynamic pool - customer pool size.
32
|
1712.00190 | 1 | 1712 | 2017-12-01T04:44:14 | Modeling the Multiple Sclerosis Brain Disease Using Agents: What Works and What Doesn't? | [
"cs.MA",
"cs.SE",
"cs.SI",
"nlin.AO",
"q-bio.NC"
] | The human brain is one of the most complex living structures in the known Universe. It consists of billions of neurons and synapses. Due to its intrinsic complexity, it can be a formidable task to accurately depict brain's structure and functionality. In the past, numerous studies have been conducted on modeling brain disease, structure, and functionality. Some of these studies have employed Agent-based approaches including multiagent-based simulation models as well as brain complex networks. While these models have all been developed using agent-based computing, however, to our best knowledge, none of them have employed the use of Agent-Oriented Software Engineering (AOSE) methodologies in developing the brain or disease model. This is a problem because without due process, developed models can miss out on important requirements. AOSE has the unique capability of merging concepts from multiagent systems, agent-based modeling, artificial intelligence, besides concepts from distributed systems. AOSE involves the various tested software engineering principles in various phases of the model development ranging from analysis, design, implementation, and testing phases. In this paper, we employ the use of three different AOSE methodologies for modeling the Multiple Sclerosis brain disease namely GAIA, TROPOS, and MASE. After developing the models, we further employ the use of Exploratory Agent-based Modeling (EABM) to develop an actual model replicating previous results as a proof of concept. The key objective of this study is to demonstrate and explore the viability and effectiveness of AOSE methodologies in the development of complex brain structure and cognitive process models. Our key finding include demonstration that AOSE methodologies can be considerably helpful in modeling various living complex systems, in general, and the human brain, in particular. | cs.MA | cs | Modeling the Multiple Sclerosis Brain Disease Using
Agents: What Works and What Doesn't?
Ayesha Muqaddas
Muaz A. Niazi*
*Corresponding author
Department of Computer Science,
COMSATS Institute of IT,
Islamabad, Pakistan
[email protected]
Abstract:
The human brain is one of the most complex living structures in the known Universe. It consists of
billions of neurons and synapses. Due to its intrinsic complexity, it can be a formidable task to accurately
depict brain's structure and functionality. In the past, numerous studies have been conducted on modeling
brain disease, structure, and functionality. Some of these studies have employed Agent-based approaches
including multiagent-based simulation models as well as brain complex networks. While these models
have all been developed using agent-based computing, however, to our best knowledge, none of them
have employed the use of Agent-Oriented Software Engineering (AOSE) methodologies in developing
the brain or disease model. This is a problem because without due process, developed models can miss
out on important requirements. AOSE has the unique capability of merging concepts from multiagent
systems, agent-based modeling, artificial intelligence, besides concepts from distributed systems. AOSE
involves the various tested software engineering principles in various phases of the model development
ranging from analysis, design, implementation, and testing phases. In this paper, we employ the use of
three different AOSE methodologies for modeling the Multiple Sclerosis brain disease – namely GAIA,
TROPOS, and MASE. After developing the models, we further employ the use of Exploratory Agent-
based Modeling (EABM) to develop an actual model replicating previous results as a proof of concept.
The key objective of this study is to demonstrate and explore the viability and effectiveness of AOSE
methodologies in the development of complex brain structure and cognitive process models. Our key
finding include demonstration that AOSE methodologies can be considerably helpful in modeling various
living complex systems, in general, and the human brain, in particular.
1. Introduction:
Cai et al. note that the human brain can be considered as one of the most complex living structures of the
known world [1]. Forstmann and Wagenmakers note the complexity of the brain because of its
composition of billions of neurons, synapsis, blood vessels, glial cells, neural stem cells, and layered
tissues [2]. For many decades, researchers have been attempting to model the human brain. These studies
have been conducted primarily to understand the structure, function, connection, dynamics of neurons and
overall brain at multiple spatial-temporal scales. Understanding and modeling the brain is extremely
important because common brain diseases such as multiple sclerosis, dementia, cancer, Alzheimer, and
epilepsy are often caused by a minor distraction of neurons or severe injury.
Due to the its inherent complexity, it is formidable to effectively model and depict brain's structure and
functionality at all scales. Numerous studies have previously been conducted on brain disease, structure,
and functionality modeling by applying ABM (Agent-Based Modeling), MAS (Multi-Agent systems) and
Complex Networks (CN). Few of them have modeled the overall brain structure or functions simply
while others have modeled the brain disease. According to researchers' best knowledge, there is no single
research that could claim that their findings are complete and absolutely flawless. These studies achieved
beneficial results for disease cure and prevention, brain structure and function understanding. However,
these marvelous systems have flaws, as these models are developed by using Agent-Oriented technology,
without following any AO methodology. To our best knowledge, none of them have used AO
methodologies for modeling brain, brain disease, function, and structure.
AOSE (Agent-Oriented Software Engineering) combines MAS, ABM, AI, and distributed systems, and
demands the application of AI and software engineering principles in the analysis, design, and
implementation phases of a software systems development process. AOSE technology has the broad
capability of autonomy, proactivity, reactivity, robustness, and social ability. Due to these capabilities
nowadays, AOSE is becoming popular in the development of the distributed and complex application
(such as e-commerce, healthcare systems, and social systems). Lucena and Nunes and R. Cunha et al,.
state in their study that, AOSE promises to deal with complex and distributed systems[3][4]. For 2000,
researchers are making a great effort to produce AOSE techniques for agent-based systems for the
guidance of design, development and maintenance process. Now at present, this is a mature technology
and has methodologies, architectures, methods, and developmental tools.
As system engineering methodologies focus on technical issues of system development. AOSE
methodologies guide step by step development of agent modeling from system requirement to system
implementation. When developers jump into system development without any guided methodology, then
they ignore the most important aspect of the system to be implemented. Moreover, developers need the
high-level expertise of system development, and they face difficulties throughout the development
process, and the ongoing project takes more time, which ultimately leads to cost exceeding without any
remarkable achievement. To overcome all these mentioned problems we are proposing an idea to follow
AOSE methodologies in brain modeling scenarios. Because of neurons in the brain interact in a
heterogeneous way, that's why we should use AOSE.
We will model MAS of Multiple Sclerosis disease by following AOSE methodologies GAIA, TROPOS,
and MaSE. After applying these methodologies, we will model exploratory agent based model as a proof
of concept. The Multiple Sclerosis ( MS) is considered as an inflammatory, autoimmune and
demyelination disease of central nervous as stated in articles [5] [6] [7] [8]. In this disease myelin sheath
on axonal part of neuron destroys with time. This destruction became an important cause of
neurodegeneration which results in impaired muscular performance [9]. It can significant cause of
physical and mental disabilities, irreversible neurologic deficits, including paralysis, muscle weakness,
ataxia tremor, spasticity, cognitive impairment, body balance disorder, vision loss, double vision, vertigo,
pain, fatigue, and depression [10].The dark side of this disease is it is becoming the main cause of
neurological disability in young adults[11] that's why intensive research is needed on this disease to save
youngsters life. The autoimmune system plays an important role in the disease progression and the body's
own immunity assaults the myelin sheath causing injury. However, a satisfactory explanation for the
origins and mechanism for MS is lacking in the literature.
In our work, we are taking MS disease as a case study for simulation system development. This case
study is presented by Pennisi et al in the article [12]. First of all, we will develop the MS disease model
by AOSE methodologies GAIA V.2, TROPOS, MaSE then, we will perform a comparative analysis of
these methodologies to find out which methodology is best in a specific scenario. The purpose of AOSE
methodologies implementation and comparison is to explore, if agent-based modeling with AOSE
methodologies is a suitable paradigm for modeling human brains structure, cognitive process, and disease.
And what kind of data would be needed for the sake of validation without spending a considerable
amount of time on these models. Moreover, we will evaluate does the AOSE technology have worth in
brain modeling. Our findings will prove that the AOSE methodologies are mature enough, that easily can
implement the complex real system of any domain without any remarkable background knowledge.
2. Related Work:
The ever-growing use of agent-oriented technology demonstrates that it is used in almost every field of
science. In this section, we will briefly present the well-known brain models, who have used agent
oriented-technology However, the analysis of literature unveils that, the wide use of agent technology in
modeling the human brain, leads the idea to use AOSE methodologies, that could aid the development
team throughout the system development project. Such as in early requirement analysis phase, design,
development and deployment phases. The AOSE methodologies work same like conventional software
engineering methodologies, which have proved that methodologies are fundamentally needed for
traditional software projects.
Y. Mansury and T. S. Deisboeck have proposed agent-based model for Spatial-temporal progression of
tumor cells, according to environmental heterogeneities in mechanical confinement toxic metabolites.
Results reveal that tumor cells follow each other along preform pathways. However, this model can be
extended to other cancers by incorporating real data[13].
C. A. Athale and T. S. D. Ã, have developed an agent-based system, to simulate progression dynamics of
the tumor. This model integrates Transforming Growth Factor α (TGFα) and induces EGFR gene-protein
interaction network. Results show the progression rate of tumor cells according to Spatio-temporal and
progression speed up when increasing EGFR density per cell [14].
Y.mansuray and .S. Deisboeck have generated a 2D agent-based system, that performs a simulation of
two different gene expression, which are Tenascin C and Proliferating-Cell-Nuclear-Antigen for brain
tumor progression. Moreover, this model investigates the effect of an environmental factor in gene
expression changes. The results reveal that Tenascin C plays a crucial role in the migration of Glioma cell
phenotype, and the expression of Proliferating-Cell-Nuclear-Antigen is responsible for proliferating
behavior of tumor cell [15].
Y.Mansury and T.S.Deisboeck have proposed a 2D agent-based system, to examine the ongoing
progressive performance of multiple tumor cells in human brain. The results demonstrate that these
dangerous tumor cells proliferate and migrate on an adaptive grid lattice. And there is acorrelation among
the progression dimension of the tumor and the velocity of tumor expansion[16].
Zhang et al. have developed a 3D multi-scale agent-based system, to perform a simulation of the cancer
cell decision process. The results reveal that tumor cells do not only oscillate between migration and
proliferation area, besides this these dangerous cells directly disturb the entire SpatioTemporal expansion
patterns of brain area which is affected by cancer[17].
Zhang et al. have proposed a hybrid agent-based model that predict severe cancer areas in the brain. This
model simulates cancer cells, either active or inactive cluster [18].
Germond et al. have proposed a multi-agent model, with remarkable features of deformation and edge
detector. This model performs segmentation on MRI data and clearly shows each dynamic dimension of a
scanned brain. When the results compared with real brain image then a real brain phantom is
reported[19].
Richard et al have advocated multi-agent framework for brain MR image segmentation. That focus on
radiometry tissue interpretation. This framework performs segmentation on a complete volume in less
than 5 min with about 0.84% accuracy[20].
Vital-Lopez et al. have developed a 3D mathematical agent-based system, that performs a simulation of
tumor progression as a collective behavior of individual tumor cells. The simulation results conclude that
vascular network damage is one of the main reasons for tumor growth and invasiveness[21].
Soc,R has proposed an agent-based model for action selection in autonomous systems. He incorporated
biological details in this Agent-Based system, in order to generate a set of predictions for decision
selection. This research concludes that agent-based technology has the broader implication of action
selection mechanism for both natural and artificial sciences related systems[22].
Signaling et al. have expanded their already proposed 3D Agent-Based and multi-scale system to perform
simulation on the brain tumor. Which examine the simplified progression of Glioma pathway. Moreover,
intracranial pressure is introduced to examine the influence of clonal heterogeneity on the human tumor
growth. This research concludes that the brain regions that are near to blood vessels and are affected by
tumor cells, these regions are the best nutrient source for tumor[23].
The scientist of Health Agent project González-Vélez et al have proposed an Agent-Based decision
support model, for the detection and the diagnosis of brain tumor classifications. This study determines
that now prediction of tumor classification is more accurate and optimize because reasoned argument is
performed among intelligent agents[24].
Zhang et al., have developed ABM for cancer simulation according to multiple dynamic scales in space-
time. This model proposes an idea of incorporating macroscopic expression patterns and microscopic cell
behavior with molecular pathway dynamics. The results show that at the same time this 3D model
efficiently model five different cancer cell clones[25].
Zhang et al., have proposed a Multiscale Agent-Based Model system to perform the simulation of
Glioblastoma Multiforme cancer. This simulation model considers the progression and proliferation of
cancer in the real time. The results conclude that this model is 30 times faster than the previous models.
Because this model's extracellular matrix have large fine grids[26].
Haroun et al., have proposed a simulation model, based on multi-agents for brain MRI segmentation. The
results show a global and local view of te image and these images are more clear than the other
competitive models result. This research also demonstrates that an appropriate quantity of agents in the
simulation is important to improve the segmentation quality of MRI images[27].
Pennisi et al., have proposed an ABM to formulate medication for MS disease and to reveal the
mechanism of distraction and new potential formulation of the blood-brain barrier (BBB). The results
suggest that vitamin D is effective for blood blockage [12].
Koutkias and Jaulent have developed ABM for Pharmacovigilance to detect complete potential signals
associated with drugs and adverse effects. This model enables clinicians to detect timely and accurate
drug signals effect[28].
Leistritz et al. have proposed an agent-based framework for agents, that have Belief, Desire, and Intention
capabilities. As a result, this model satisfies a key integration challenge, of coupling time stepped ABM
with event-based BDI systems[29].
Barrah has developed a MAS for magnetic resonance imaging (MRI) with a median filter to decrease
computational time and save data details. The proposed FRFCM with MAS is fastidious, accurate and
take less computational time[30].
Pennisi et al., have developed ABM for RRMS (Relapsing Remitting Multiple Sclerosis) disease. The
outcome results show that the occurrence of genetic disposition in neurons is one cause of MS disease.
However, the main cause is a breakdown of peripheral tolerance mechanism.[31].
MAS ABM
✓
S.E
✓
✓
✓
AOSE
Ref
[13]
Title
"The impact of ''search precision'' in
an agent-based tumor model"
Author Year
Mansury and
Deisboeck
2003
[14]
"The effects of EGF-receptor density
on multiscale tumor growth patterns"
Athale and Ã
2006
Journal
Journal
Theoretical
biology
Journal
Theoretical
Biology
of
of
"Simulating the time series of selected
gene expression profile in an agent –
based tumor model"
"Simulating
patterns of malignant brain tumors"
'structure-function'
Mansury and
Deisboeck
Mansury and
Deisboeck
2004
Physica D
2004
Physica A
[15]
[16]
[17]
"Development of a three-dimensional
multiscale agent-based tumor model:
interaction
Simulating gene-protein
profiles,
and
multicellular patterns in brain cancer"
phenotypes
cell
2007
Zhang,
Athale
Deisboeck
and
Journal
Theoretical
Biology
of
✓
[18]
"Multi-scale, multi-resolution brain
cancer modeling"
2009
Zhang, Chen
and
Deisboeck
[19]
cooperative
framework
"A
segmentation of MRI brain scans"
or
Germond et
al.,
2000
MATHEMATICS
AND
COMPUTERS IN
SIMULATION
Artificial
Intelligence
Medicine
in
✓
✓
[20]
"Automated segmentation of human
brain MR images using a multi-agent
approach"
Richard,
Dojat
Garbay,
and
[21]
"Modeling the Effect of Chemotaxis on
Glioblastoma Tumor Progression"
Vital-lopez,
Armaou and
Hutnik,
2004
2011
Artificial
Intelligence
Medicine
in
American Institute
of
Chemical
Engineers
"Introduction. Modeling natural action
selection"
Brain
"Simulating
Tumor
Heterogeneity with a Multiscale Agent-
Based Model
: Linking molecular
signatures, phenotypes and expansion
rate"
Soc,R
2007
The royal society
2009
Signaling,
Bias
Rate
and
Mathematical and
Computer
Modeling
"HealthAgents: distributed multi-agent
brain tumor diagnosis and prognosis"
González-
Vélez et al
2009
"Multiscale
modeling"
agent-based
cancer
Zhang et al.,
2009
a multiscale, multi-
tumor
"Developing
resolution agent-based brain
model by graphic processing"
Zhang et al.,
2011
Applied
Intelligence
Journal
Mathematical
Biology
Theoretical
Biology
Medical
Modelling
of
and
[22]
[23]
[24]
[25]
[26]
[27]
[12]
✓
✓
✓
✓
✓
✓
✓
"A Massive Multi-Agent System for
Brain MRI Segmentation"
Haroun
al.,
et
2005
Massively Multi-
Agent Systems
✓
"Agent-based modeling of the effects of
potential treatments over the blood–
brain barrier in multiple sclerosis"
Pennisi,
Marzio, et al
2015
Journal
immunological
methods
of
✓
[28]
"A multi-agent system for integrated
detection of pharmacovigilance signals"
[29]
"Time-variant modeling
processes"
of brain
[30]
"MAS based on a Fast and Robust
FCM Algorithm for MR Brain Image
Segmentation"
Koutkias, V,
and Jaulent,
M. C
Leistritz, L.,
Schiecke, K.,
Astolfi, L.,
and Witte, H.
Barrah, H. et
al.
2016
2016
2016
[31]
"Agent based modeling of Treg-Teff
cross regulation in relapsing-remitting
multiple sclerosis"
M. Pennisi et
al.
2013
Journal of medical
systems
Proceedings of the
IEEE
✓
✓
of
. International
Journal
Advanced
Computer Science
& Applications
BMC
Bioinformatics
✓
✓
3. Material and Methods
This section describes in detail the case study of multiple sclerosis (MS) disease and the implementation
of the MS model by using AOSE methodologies like Gaia V.2, Tropos, and MASE. According to our
best knowledge that AOSE methodology has never been used in any biological model. First time we tried
to model biological problem, according to well defined agent methodologies. To do this, here we are
copying same rules and regulations that are used to model other systems in engineering field. We
implemented MS model in Net-Logo environment. This model is already implemented by Pennisi et al. in
[12].
3.1. Case study:
MS is one of the central nervous system diseases. In which myelin sheath from the axonal part of the
neurons in the brain and spinal cord is removed and the communication between neurons is
disconnected.In this disease, different factors and cells are involved. The overall mechanism of MS
disease is given below.
The thymus gland is a lymphoid organ that produces T-cells (T-reg and T-eff). The T-reg and T-eff both
cells have two states, "active and resting". EBV (virus) is an external factor that latent infection. EBV
cause activation of both T-eff and T-reg. The active T-eff (A.T-eff) attacks the myelin of axons and cause
the neural communication damage and in return duplicate itself. A.T-eff produce one cytokine against one
attack. Active T-reg (A.T-reg) try to catch A.T-eff and suppress it. In return, A.t-reg receive a positive
feedback and will duplicate. After the attack of A.T-eff the amount of myelin in BWM will be lowered. If
the damaged portion still has any amount of myelin then it is recoverable. If myelin amount reaches zero,
then it is unrecoverable. A.T-eff produce cytokines after damage myelin. Cytokine attacks BBB and
damages it. The damaged BBB allows other T-cells and virus to enter into the brain. The main purpose of
the BBB is blocked and bounce back to all T-reg, A.T-reg, T-eff, A.T-eff, Virus, Cytokines.
At the start of the simulation, all agents are introduced in the model in the resting state. The agents are
considered as brain cells. All agents have life counter that decrement by one at every step of the agent. If
the life counter of agent reaches to zero, then that agent will be removed from the simulation
automatically. In running simulation, all agents will move randomly in the environment (actually brain)
and new agents will be introduced at random time.
3.2. GAIA V.2:
The original GAIA methodology was presented by Wooldridge in 2000, It was the first complete AOSE
methodology developed for large-scale real-world applications [32]. GAIA V.2 focuses on two main
development phases of a system, analysis phase, and development phase. Moreover, these models are
considered as a guideline for developing the real models of the to-be-developed complex system.
This methodology is applied to ABM and MAS modeling after requirements are gathered and specified.
After adoption of Gaia methodology for complex systems development, researchers realize that it is easy
to use and implement. However, it is not much suitable for complex systems design [33]. Because
modeling notations are poor for expressing complex problems, such as complex and multiphase
interaction protocols. Moreover, Gaia has a deficiency of requirement phase, environment model, and
does not provide an appropriate domain knowledge. To overcome these limitations Zambonelli, Jennings
and Wooldridge proposed Gaia V.2 agent methodology in [34]. In our MS model we are using Gaia V.2,
same like Gaia it has two development phases 1) analysis and 2) design as shown in the following picture.
Figure 1: GAIA V.2 methodology [35]
3.2.1 Analysis Phase:
Generally, the analysis phase emphasizes on understanding what the MAS and ABM system will have to
be as their properties or specifications (without reference to any implementation detail). The output of the
analysis phase became the base input for the design phase. In actual, in this phase analyst decide about the
functional and non-functional characteristics of the to-be-system. The output of this phase is five basic
models such as 1) sub-organizations, 2) environmental model, 3) preliminary role model, 4) preliminary
interaction model, and 5) organizational rules model.
1) System Sub-Organizations
This model tries to identify all the involved organizations, and used a fruitful way to categorize the whole
system into loosely coupled sub-organizations. The identification of sub-organization is easy if
1) The Sub-organizations are already identified through system specifications.
2) The system itself mimics the overall structure of the real world, in which multiple organizations
interact.
3) Apply the modularity technique. This technique splits the overall complexity of the to-be-system
into a set of smaller and more manageable components.
According to our experience, the sub-organizations of the developing system can be found easily, when
we divide the overall system into portions. And each portion exhibit a specific behavior and they interact
with each other to achieve a subgoal. Any portion of the system which performs any task or interacts with
other to perform a specific task, then we can take it as a sub-organization.
Sub-organizations
Description
T-eff
T-reg
Virus
A.T-eff
A.T-reg
Cytokines
BWM
BBB
The goal to achieve is to "become active" by catching the virus. It will interact
with the virus for activation and BBB to enter into the brain.
The goal to achieve is to "become active" by catching the virus. It will interact
with the virus for activation and BBB to enter into the brain.
The goal to achieve is to "enter into the brain" and make " T-reg and T-eff
active". It interacts with T-reg, T-eff to make them active and BBB to enter into
the brain.
The goal to achieve is to "attack on myelin" to damage brain, "duplicate itself"
and "produce cytokines". It interacts with BBB to enter into the brain.
The goal to achieve is to "attack on " A.T-eff to kill them and "duplicate itself".
It interacts with A.T-eff to kill them and interact with BBB to enter into the
brain.
The goal to achieve is "attack on BBB" to damage it. It interacts with only BBB.
The goal to achieve is "recover damaged myelin" in the brain to maintain neural
communication.
The goal to achieve is "stop the entrance" of T-eff, T-reg, A.T-eff. A.T-
reg,virus, Cytokine into the Brain.
Table 1: Sub-Organizations and their goals
In our's MS disease model, we identified eight sub-organizations T-reg, T-eff, virus, A.T-eff, A.T-reg,
Cytokines, BBB and BWM as showed in table 1. After identification of sub-organizations, we divided
them into two main groups by arranging roles into groups with logical or physical similarities as Silva has
done in his work [35]. The first group is brain organs and the second one is neuron cells. The Brain organ
group consist of two types of agents (BBB, and BWM). On the other hand cell group contain EBV ,T-reg,
T-eff, A.T-reg, A.T-eff and Cytokines. According to AOSE term, in general, we would call "agents" to all
neuron cells and brain organs. The agents in the second group (neuron cells) are tightly connected and
they also interact with the second group (brain organ) agents. Although GAIA is lenient with complex
systems development, however, due to the lack of requirement analysis phase, the accurate identification
of sub-organizations is not feasible from GAIA, as it also lacks of established hierarchy model and
organizational structure[35] [36].
2) Environment Model
The environmental model describes the real world in which the system operates with all its variables,
resources, and uncertainties. There is not a single easy way, to provide a general modeling abstraction and
the general modeling techniques for the development of the system environment. Since the environments
of the applications are divergent in nature because in most cases systems suffers from compatibility issues
with the technology. As Passos stated in his study that the development of the environmental model is a
separate one type of agent-oriented methodology [37]. That's why we can not model complete
environmental model of any system.
Figure 2 Environmental Model of MS disease simulation
The environment of our model is Netlogo development tool and resources are Netlogo's buttons, sliders,
plots, switches eg. We divide this model into two environments, static and dynamic according to Gaia
methodology. The static part of our model is buttons, sliders, switches that do not change frequently. And
the dynamic part is disply monitor in which agents move randomly in environment to perform complex
operation and to access resources. The discription of the environmental model of MS disease is given in
below table.
Environment Resources
Description
Display Screen
Buttons
Sliders
Switches
Plots
Agents
The screen shows all the time movement of agents and changing in the
model. It also displays what is going on among agents.
The buttons stop or start the simulation and initialize special characteristics
of agents, as creation or deletion agents etc.
Sliders assign special values to agents in the start or during the simulation to
observe effects of changes.
Switches on/off characteristics of agents in simulation.
The Plot shows the behavior of agent interaction.
Agents representative of characters of one type that are involved in the
Patches
simulation to perform a specific task.
The patches are also representative of agents of another type that are
involved in the simulation to perform a specific task.
Table 2: Description of Environmental Model
3) Preliminary Role Model
A role model provides an abstraction of the agent with a set of its expected behavior[34]. The final goal of
the analysis phase is not to identify the all involving entities and model them into the real system
organization. The main purpose is to identify the active roles and their interaction with others roles at a
very abstract level. Particularly, the role abstraction leads to the identification of basic specifications for
the developing system, to achieve its main goal. As well as the important interactions that are required for
the completion of system's specification. However, this type of identification can be performed without
knowing that, what the actual structure of the developing system will be. There is a chance to identify
some characteristics that will remain same independently throughout the organizational structure. This
identification can be more beneficial if the analysis phase carefully models the system specifications in
terms of involved characters and their responsibilities at a very beginning stage. In our model, we
identified 8 roles T-reg, T-eff, A.T-reg, A.T-eff, Virus, BBB, Cytokines and BWM (Brain White Matter)
which are shown in figure 2.
4) Preliminary Interaction or protocol Model
The interaction model identifies all dependencies which show the relationship among roles by defining
the protocol[37]. Protocols are a request for resources or to complete a task, that a role does to interact
with other roles. In protocol development, more focus is on the purpose of interaction and the nature of
interaction, than to the type of message exchange and sequence of execution steps. A standard protocol
model of GAIA V.2 consists of these points:
1) Protocol title/name: The protocol name is a description of the roles interaction nature. For
example, it specifies that a request is for resource share or to assign a task.
2) Initiator: The role who starts the conversation.
3) Partner: that role will be partner which responds to initiator during conversation.
4) Input: The information used by the initiator role as a reason to initiate the protocol.
5) The final action that will take the responder role. It can be an information, resource or a request
for anything.
6) Description: The description explains the whole scenario of protocol processing in detail. e.g
which roles will involve and what will be their intentions and what will be the consequences.
In MS disease system five protocols are identified. That are Attack, Catch, Produce, MakeActive, and
stop. At the end, the general model of preliminary interaction and role express the overall scenario of
protocols among the roles of MS disease. The interaction model is also called a protocol model. In the
below tables, all protocols are described in detail.
Protocol Name: catch
Initiator:
T-reg
Partner:
Virus
Input:
Information about
virus existence
Description:
When any T-reg knows, that there is virus
around it then, it tries to catch the virus to
become active
Output:
A.T-reg
Table 3: T-reg's Protocol
Protocol Name: catch
Initiator:
T-eff
Partner:
Virus
Input:
Information about
virus existence
Description:
When any T-eff knows, that there is virus
around it then, it tries to catch the virus to
become active
Output:
A.T-eff
Table 4: T-eff's Protocol
Protocol Name: makeActive
Initiator:
Virus
Partner:
T-reg, T-eff
Description:
When any virus knows, that there is T-reg, T-
eff around it. Then, virus tries to catch the T-
reg, T-eff to make it active
Input:
Information about
T-reg, T-eff
existence
Output:
A.T-reg
Table 5: Virus's Protocol
Protocol Name: Attack
Initiator:
A.T-eff
Partner:
BWM
Description:
When any A.T-eff finds, that there is myelin
in the brain, then it attacks on myelin and
damage it.
Input:
Information about
myelin existence
Output:
Damaged myelin
Table 6:A.T-eff's Protocol
Protocol Name: Attack
Initiator:
A.T-reg
Partner:
A.T-eff
Input:
Information about
A.T-eff existence
Description:
When any A.T-reg finds, that there is A.T-eff
in brain then, it attacks on A.T-eff and kill it.
Output:
A.T-eff killed
Table 7: A.T-reg's Protocol
Protocol Name: Attack
Initiator:
Cytokine
Partner:
BBB
Input:
Information about
BBB existence
Description:
When any Cytokine finds, that there BBB,
then Cytokine attacks on BBB and damage it.
Output:
Damaged BBB
Table 8: Cytokine's Protocol
Protocol Name: Produce
Initiator:
A.T-reg
Partner:
A.T-eff
Input:
Information about
A.T-eff existence
Description:
When any A.T-reg finds, that there is A.T-eff
in brain then, it kills A.T-eff and produce
itself duplicate.
Output:
A.T-reg
Table 9: A.T-reg's Protocol
Protocol Name: Produce
Initiator:
A.T-eff
Partner:
Cytokine
Input:
Information about
myelin existence
Description:
When any A.T-eff finds, that there is myelin
in the brain, then it attacks on myelin and
duplicate itself.
Output:
A.T-reg
Table 10A.T-eff's Protocol
Protocol Name: Stop
Initiator:
BBB
Partner:
T-reg, T-eff, Virus,
A.T-reg, A.T-eff,
Cytokines,
Input:
Information about
enterance of cells
into brain
Description:
When ever BBB knows that any cell is trying
to enter into brain then BBB stop its
entrance.
Output:
A.T-reg
Table 11:BBB's Protocol
Figure 3: Preliminary Role and Interaction Model of MS
This is a combined figure of the preliminary role and their interaction model of MS disease.
In this figure, there are all identified role and protocols. T-reg and T-eff roles interact with the virus by
using catch protocol. These both roles aim is to become active by catching the virus. In general, the
protocol defines the purpose of communication among agents. In the same way, the virus role interacts
with T-reg and T-eff by using "makeActive" protocol to make them active. A.T-reg (Active T-reg) and
A.T-eff (Active T-eff) are active agents of T-reg and T-eff community. A.T-eff interacts with BWM by
using "attack" protocol. The "attack" protocol is used to damage the BWM. A.T-eff also interact with
Cytokine by using "produce" protocol. A.T-reg role interacts with A.T-eff by using "attack" protocol. The
purpose of this protocol or communication is to kill the A.T-eff role to stop its dangerous activities in the
brain. In the same way, the Cytokine interacts with BBB by using "attack" protocol to damage BBB. The
BBB role interacts with almost all other roles except BWM by using "stop" protocol. The purpose of this
protocol is to stop all other roles, to entering the brain.
5) Organizational Rules
The organizational rules define whether a new agent can be added into the organization. If new agents
added then what would be their position in the organization and which type of behavior would be
expected from the added agents. In simple words, the organizational rule defines the overall responsibility
of the concerned organization in an abstract way. There are two types of organizational rule:
(1) Liveness: The liveness rules take concern about the evolution of system dynamics according to time.
It also takes concern about that a specific agent will play a specific role and allow the agent to play next
task if it has played the previous role. Same like roles, a specific protocol may execute only after the
execution of the other specific protocol. Moreover, liveness organizational rule decides that which role
would be played by which agent.
(2) Safety: The safety rules consider all unexpected events that can occur during processing of a specific
task. These events are considered as time-independent events. To overcome all the unexpected faults, the
organization rule force to apply the concept that a single role must play by a distinct entity or agent and
two concurrent tasks should not be played by a single entity.
By following these conditions, in our MS disease system sub-organizations define their rules separately.
Organization Name: T-eff
Liveness: This role would become active if only it catches a virus, or it would transform from T-eff
to A.T-eff if only virus attacks on it.
Safety: Only this agent can transform from T-eff to A.T-eff agent.
Table 12: T-eff Organizational Rules
Organization Name: T-reg
Liveness: This role would become active if only it catches a virus, or it would transform from T-reg
to A.T-reg if only virus attacks on it.
Safety: Only this agent can transform from T-reg to A.T-reg agent.
Table 13: T-reg Organizational Rules
Organization Name: Virus
Liveness: This agent can enter into the brain if BBB is broken.
Safety: Only this agent trigger the MS disease.
Table 14: Virus's Organizational Rules
Organization Name: A.T-eff
Liveness: This agent can damage myelin if the only virus makes it active.
Safety: Only this agent can damage the myelin.
Table 15: A.T-eff Organizational Rules
Organization Name: A.T-reg
Liveness: This agent can kill the A.T-eff if only it is active.
Safety: Only this agent can kill the dangerous agent A.T-eff.
Table 16:A.T-reg Organizational Rules
Organization Name: Cytokine
Liveness: These agents born after myelin damage.
Safety: Only these agents can damage BBB.
Table 17: Cytokine's Organizational Rules
Organization Name: BBB
Liveness: This agent stops the other agent who tries to enter into the brain.
Safety: This agent stops all agents to enter into the brain and recover damaged BBB.
Table 18: BBB Organizational Rules
Organization Name: BWM
Liveness: This agent recovers myelin after myelin damage occurs.
Safety: This agent recovers myelin for neural communication.
Table 19: BWM's Organizational Rules
3.2.2. Design Phase
The design phase is a much important phase in system engineering, where the final decisions take about
what will be the specifications of the developed system, what will be the operational environment. The
main purpose of this phase is to transform the analysis models into low-level design models. Low-level
design models are abstract level models which can be easily implemented in the development phase.
This phase is also important in this way that, it identifies missing and conflicting requirements and helps
developers when system specifications should be considered mature to develop a complex system. The
design phase of Gaia V.2 is divided into two sub design phases, first one is architectural design and the
second one is detailed design.
1. Architectural Design
1.1 Organizational structure
The selection of organizational structure of a system is a very crucial decision in ABM and MAS
development since, it affects all subsequent phases. The one important benefit of an organizational
structure is, it organizes roles as a topology, which is easily understandable by inexperienced persons.
And the other main objective of this structure is to explicate the inter-role relationship type among
roles/agents. In general, there are three types of relationships:
I.
II.
III.
In the control relationship, one role has authority over the other role, in this relationship a role
can partially or fully control the action of the other role.
The peer relationship defines that, in the organization, the involved roles have equal status.
In the dependency relationship, one role depends on the other role for resources or
knowledge, which is compulsory for its accomplishment.
However, the organizational structure same like sub-organizations model does not accurately be
implemented, because there is no fixed hierarchical structure exists. The MS disease system is
implementing the generic architectural model by considering whole MS circumstances as a hierarchy, and
dependency between agents as a control structure. In our model the organizational structure manifest
roles, sub-organizations and the relations and association between them.
Figure 4: GAIA's Organizational Structure
The MS model defines brain organs (BBB, BWM), T-cells, Cytokines, and virus as sub-organizations.
The arrows show organization's communication with other organizations. This communication can be
direct organizations' communication or with the role of other organizations. One organization can
communicate with the roles of other organizations.
IN MS model T-cells, Virus, BWM, and BBB are sub-organizations because they all are different in
structure and nature. T-reg and T-eff cells belong to the same cell category because they are produced by
the same body organ. That's why they belong to the same sub-organization. However, after activation,
they behave differently according to their builtin nature. In the same way, BBB organization
communicates with all others organizations' agents to stop their entrance into the brain. Further, A.T-reg
develops hierarchy structure. In this structure, A.T-reg is organized according to its goals. A.T-eff also
develops a hierarchy structure. Which, is organized according to roles.
1.2 Role and interaction model
The design phase transforms the preliminary role model and the role interaction model into absolute
detailed role and interaction model to form an organizational structure. Which clearly defines the roles of
each agent and interaction type among them. To construct detailed role and interaction model a developer
should consider these mentioned instructions:
Identify the new roles and organizations which was not identified by analysis phase.
I.
II. Developers should identify all the possible activities in which a role can be involved, as
well as classify role's safety and liveness responsibilities contribute to overall
organizational rules.
Identify all the possible protocols, needed by the developing system. As well as classify
which roles will be involved in the relationship to complete the protocol execution.
III.
IV. Moreover, developers should identify the protocols for the organizational level
relationship.
In short, for absolute role model, we identify the complete role's activities and services. And for the
absolute interaction model, we identify and model interaction of all roles involved.
Figure 5: Role and Iteraction Model
In this model, we defined roles with their responsibilities, services, and protocols. The difference between
preliminary phase models and detailed design models is, in the design phase, we identify all possible and
complete information about the roles characteristics and protocols.
As already defined that 8 MS disease roles and two types of agents are identified. BBB and myelin belong
to brain organ type and virus, T-cells belong to cells agent type. T-cells (T-reg and T-eff) belong to the
same community of cells, but they react to brain according to their built-in nature. In other words, some
of T-cells became dangerous as T-eff and some became part of immune system T-reg. The T-reg role is
responsible to catch viruses to get active. The active T-reg (A.T-reg) support immune system by killing
dangerous Active T-eff (A.Teff) cells and it produce duplicate A.Treg cell. The duplicate T-cell perform
functionality same like original T-cell. Virus role is responsible for T-reg and T-eff cells activation. The
virus searches T-cells and try to attack them after that it died. A.T-eff role is responsible to attack myelin
and damage it. It also duplicates A.T-eff and produces cytokines. Cytokine roles, responsibility is to
attack the BBB to damage it and allow viruses and all other agents to move toward the brain. BBB role
belongs to brain organ agent type and is responsible to block unwanted molecules and cells to enter into
the brain. It also manages to allow wanted molecules (water, glucose, and water) and cells, T-reg to enter
into the brain to protect the brain. BWM or myelin is responsible to maintain communication between
neuron cells and if any damage occurs, then regrow myelin to maintain communication.
In this MS model, there are five types of protocols "stop, catch, attack, produce, and makeActive". BBB
initiates stop protocol to stop the virus, T-reg, T-eff, A.T-reg, A.T-eff, Cytokine, to enter the brain. Virus
initiates "makeActive" protocol with T-reg, T-eff to make them active. T-reg and T-eff agents initiate
catch protocol with a virus to get active. A.T-reg agent initiates "attack" protocol with A.T-eff agent to
stop their unhealthy activities. On the other side, A.T-eff initiate "attack" protocol for demyelinate BWM.
Cytokines initiate "attack" protocol with BBB to destroy the BBB. Moreover, "produce" protocol is
initiated by A.T-reg, A.T-eff to produce their duplicates. The "produce" protocol is also initiated by BBB
and BWM to recover themselves.
2. Detailed Design Phase
2.1 Agent Model
In the detailed design phase, developers are more experienced about the developing system. This phase
helps developers to identify the actual role model and the actual interaction model which in return will
assist the developers in the implementation phase. The Agent model captures all agent types and agent's
roles that will be implemented in the system. This phase implements agent model same like class model
in UML. As Castro and Oliveira state that GAIA does not provide modeling notations for agent modeling.
It simply suggests to adopting UML class diagram in article[38]. Zambonelli, Jennings and Wooldridge
states in article [34], according to GAIA "An agent is a software that plays a set of roles of a certain
type". Thus the agent model carefully identifies, what type of agent classes should be involved to play the
specific roles and how many instances should be in each class of the actual system.
2.2 Service Model
The service model identifies all the possible services associated with each agent class and consistently
with the roles. The service model can be applied in both cases: 1) the static assignment of roles to agent
classes, 2) the assignment of dynamic roles to agent classes[34]. The service is a function of the
considered agent, and it is derived from the agent's protocols, liveness and responsibilities, and the
activities of the role that each agent implements. This fig is representing service and agent model.
Services express what agents are contributing for the beneficial of the overall models working.
In our MS disease model, we are combining agent and role model in a single model.
Figure 6: Agent and Service Model
In the MS disease model, there are two types of agents patches and turtles. The patches represent to brain
organ (BBB, BWM) and turtles represent to T-Cells (T-eff, T-reg, A.T-reg, A.T-eff), Virus, Cytokines. In
figure 6, all agents are defined with precise services.
3.3 TROPOS
The TROPOS is an agent-oriented software engineering methodology, which was proposed by Castro et
al. in 2002. The name was derived from a Greek word "trope" which means that it is easily adaptable and
modifiable[39]. The complete methodology and its implementation proposed for Media shop in the article
[40]. The principle aims to propose this methodology was to overcome the semantic gap between the
operational environment and the developed system. Since the structured and object-oriented software
development methodologies only have the programming concepts, not organizational ones [40].
Basically, this methodology was founded on two features [41]:
1. This methodology was developed according to the agent's concept and the mentalistic notions of
the agent such as goal, plan, and resource. Then these notions help the developers throughout the
development process of the system.
2. This methodology introduces legitimate requirement specifications to give a crucial role to
requirement analysis phase.
This methodology adopts ⅈ* model, which was proposed by Yu in [42]. The i* model explicitly describes
to the actor(can be role, position, and agent), actor dependencies and goals as primitive concepts for all
models which will be developed in different phases of software development. Tropos more specifically
covers four phases of software development 1) Early Requirement Analysis, 2) LateRequirement
Analysis, 3) Architectural Design and 4) Detailed design. And in somehow way it also supports
implementation phase [43][44][39].
3.3.1 Early Requirements
In the early requirement analysis phase, the developers focus on the identification of the problem domain.
After the identification of problem, they study the existing organizational setting where the system to-be
operated. The Study reveals that the core intention of requirement analysis phase is, derive a set of
functional and nonfunctional requirements for the system to-be[40][43]. Both the early and late
requirement phases share the same methodological and conceptual approach. The early requirement
analysis phase analyzes and identifies the involved stakeholders and their intentions The stakeholders of
the system to-be is modeled as social actors or agents, that depend on one another for plans to be
performed, goals to be achieved, and resources to be shared. In MS disease model the stakeholders are
cells who participate in disease cause. And the agent's intentions are modeled as the goals to be achieved.
In this phase, we will implement i* model, that consist of 1) strategic dependency model as actor diagram
and 2) strategic rationale model. The rational model supports and describes the reason to fulfill its goals
and to make a relationship with the other agents. The outcome of this phase is an organizational model. The
organizational model includes all the involving actors, their dependency relationship with other actors and
their particular goals[40]. However, the ultimate goal of this phase is to clearly understand the
environment and context of the organization, where the system to-be will perform[41].
In the depth of modeling, Tropos depend on ⅈ* notations as described in below.
Figure 7: Notation Diagram
Actor: An actor is a representative of physical entity or position in the organization, a software entity, a
social role. Who is responsible for goals to be achieved, resources to be shared, and tasks to be performed.
Goal: Dam and Winikoff states that goals represent actor's strategic interest in [45]. There are two types
of goal: hard goal and soft goal. The hard goal represents functional requirement or specification of the
system. However, soft goal have no clear definition for deciding whether they are achieved or not.
Plan: A strategy that is adopted to fulfill a specific goal or soft goal.
Resource: Resources are something such as physical or informational entity, which required for agent to
fulfill specific tasks.
AND/OR Decomposition: The AND decomposition divides the main goal or task into more than one
sub goal. In this case, all sub goals must be achieved to fulfill the root goal. In the OR decomposition, the
main goal is divided into other alternative ways. In this case, the root goal can be achieved by any sub
goal's completion.
Mean-end Analysis: In mean-end relationship a mean (in term of goal, resource, and plan) completely
satisfy the root goal.
Social Dependency: The social dependency is an agreement between two actors: the depender and the
dependee, to achieve their common goal. Their common goal is called the dependum.
In TROPOS implementation, we are considering each neural cell and brain organ as an agent.
Figure 8: Actor Diagram
In the early requirement analysis phase of TROPOS, we identified 8 actors T-reg,T-eff, Virus, A.T-reg
(active T-reg), A.T-eff (active T-eff), BBB (Blood Brain Barrier), BWM (Brain White Matter or myelin)
and cytokine. Each role has its own goals, soft goals, resources, and plans. Actor T-reg has a goal "catch
virus" and "get active" to perform its responsibility such as protect myelin by killing A.T-eff cells. To
achieve goals T-reg depends on the virus. If it detects any virus in the brain, then it became active. As
actors can depend on other actors by soft goals, resources and plans. The T-eff also depend on virus to
complete its goal which is "catch the virus" and "become active". The Virus agent depends on T-eff and
T-reg agents to make them active without knowledge that T-ref will slow down damage rate. A.T-reg has
goals attack injurious agents as A.T-eff and kill them and in return produce A.T-reg. To fulfill its goals, it
depends on A.T-eff agent. And the soft goal is to increase the amount of T-reg agents to slow down
damage rate. A.T-eff's goals are "attack on myelin" to damage brain as well as "produce A.T-eff and
cytokines". To achieve its all goals it depends on BWM. After the attack on myelin then, it is able to
produce A.T-eff and Cytokines. Its soft goal is increase brain damage rate by producing more A.T-eff and
Cytokines. In fact, these agents play injurious activities in the brain. The BBB's goals are "stop agents
entering into brain" and "repair damage part of the BBB". The intention and goal of BBB is to protect the
brain by blocking all minacious agents.To fulfill goal this agent depends on all agents except BWM. The
BWM agents' goal is to repair damaged brain by regrowing myelin. Its soft goal is cover axons to
maintain neural communication between neurons. The Cytokine agents have a threatening behavior for
BBB. It depends on BBB to achieve goal, its goals are "damage BBB" and "allow all minacious agents to
enter into brain".
Rational models
The strategic rational model is a balloon like circle contain on four types of nodes task, goal, resource,
soft goal, and two types of links task decomposition link and mean-end link. The rational model captures
how an agent makes plan to fulfill its root goal and how it makes relation with other agents to fulfill the
system's goal.
Figure 9: T-eff Rationale model Figure 10:T-eff Rationale model
The goal of T-eff agent is to become active to perform its fundamental activities. For this, it plans either
active by catching a virus or be exposed to the virus. OR decomposition represent to an option to fulfill a
single aim and these options positively effect the plan. As a consequence plan positively supports the
main goal. This is the same scenario in the case of T-reg agent.
Figure 11 Virus Rationale Model Figure 12: A.T-reg Rationale Model
In Rationale model virus have goals made T-cells active and damage brain. To achieve damage the brain
goal, it has two options, either make more T-eff active or make less T-reg active. These options provide
alternative ways to fulfill a single goal. The Active T-Reg goal has a negative effect on soft goal and plan.
In A.T-Reg Rationale Model the soft goal of the agent is to support the immune system. This goal is
achieved by OR decomposition of killing A.T-eff or producing A.T-reg agents. The plan to kill A.T-eff
can succeed by producing more A.T-reg agents or by searching more A.T-eff agents.
Figure 13: T-eff Rationale Model
The overall objective of A.T-eff is to damage the brain. To achieve this goal A.T-eff agent set a hard goal
attack on myelin (BWM). For this goal this agent plans for increase damage rate by producing more
A.T-eff and cytokine agents. The Plan is decomposed into two goals. These goals have a positive effect
on the plan.
Figure 14: BBB Rationale Model
The main objective of the BBB is to protect brain from minacious molecules and cells. To achieve this
goal this agent set a hard goal to stop entrance of minacious cells into the brain. To fulfill this goal, it
plans to enter only T-helper cells T-reg and block T-eff, virus and dangerous molecules.
Figure 15: BWM (Myelin) Rationale Model
BWM or myelin maintains neural communication between neurons that would be its soft goal. To fulfill
its duty It regrows myelin if any damage occurs.
Figure 16: Cytokine Rationale Model
Cytokines are produced from A.T-eff agents. So they have the same goal as T-eff have "damage brain".
But they achieve this goal from another way by damage BBB. To achieve this goal cytokine try to reach
at BBB and hit it.
3.3.2 Late Requirements
In this phase, the developers identify all possible specifications of the system to-be and its operational
environment. The output of this phase is the final specifications of the system, in the form of functional
and nonfunctional requirement. This phase introduces the operational view of the system as an actor
model. To iterate, the conceptual model of the system to-be is extended by introducing system actor in the
overall system model and shows the dependency relationship between system actor and the other involved
actors. These dependencies show the main reason of interaction of system agent with the other involved
agents[43].
As Bertolini et al. state that, the system actor has the same features, which the other social actors have, in
terms of social dependency and goals to analyze[46]. The addition of system actor helps the developers to
recheck and conform that all involved actors and their dependencies are identified and the system actor
can be more than one actor. To summarize that, the actual system actor comes into a picture of one or
more actors, who helps the other actors to fulfill their goals.
In MS disease case MS is the system actor which describes the overall phenomena, how this disease
appear and how other actors contribute to proliferate this disease.
Figure 17: System Rationale Diagram
The system model controls overall system's environment and the involvement of actors. To iterate,
system actor is responsible for the entire behavior of the system. The behavior of system actor is, how it
responds and fulfill the requirements of other actors and the users of the system. In MS model the user
actor will be, who will run the simulation against particular factors or agents attributes. The hard goal of
the MS system actor is to show an accurate simulation of MS disease. This simulation shows, the varying
behavior of agents, for example, it shows at which point, which agents interact, how they interact, the
reason of agent's interaction and what was the effect of their interaction. To achieve this goal MS actor
plan for behavior space and plot. This plan uses behavior space as a resource. Actually, system actor
shows how and why other actors depend on it. The other hard goal of this system actor is to show a
simulation of varying scenario. Many other sub-goals (like show myelin recovery, show agent's
interaction, show agent's movement, show interaction b/w agents and BBB, show interaction b/w agents
and BWM) participate to complete this goal. System actor also plans for this goal and plan succeed by
Buttons, sliders, and switches. The soft goals are increasing understandability and usability.
Understandability support users to operate the whole system easily understand the output results of the
simulation.
3.3.3 Architectural Design
The architectural design phase expresses the overall architecture of the developing system. In architecture
sub-system or agents are connected by data flow, control flow, or other dependencies[40]. In actual, this
phase is an extended version of the actor diagram. In which each actor is introduced in detail according to
its dependencies with the other actors.
Moreover, the architectural design phase performs a mapping on the system actors, and the outcome is a
set of software agents. And each agent is characterized according to its specific abilities[43]. The system
architecture helps the designers to understand how system components work together and how to
constitute a relatively small, intellectually manageable model of system structure. The main objective of
this phase is, to make a check that either all involved agents are identified? If new agent is identify then it
should be incorporated in the system and also identify its control interconnections in the form of
dependencies.
In MS disease model, new actors are not identified in architectural design phase; however the overall
organizational structure is presented. This model includes rational diagram of each actor and the
dependency between actors through goal, soft goal, plans, and resources.
Create rationale diagram of each agent and link them to show architectural design.
3.3.4 Detailed design
The detailed design phase further elaborates the architectural design and the behavior of its interconnected
components[40]. The main concern of this phase is, actively deals the requirement specification
document, and all the involved agents at micro and macro level. Bresciani and Perini suggest that mostly
we use AUML activity diagram during detailed design for representing capabilities, plans and interaction
of agents in the article [43].
Activity diagram captures the dynamic process of the overall system and each involving agent. The agent
dynamics are the behavior of participating agents in the protocol and internal plans to achieve a specific
goal. For capability modeling, the UML's activity diagram assist to model the capability from the stand
point of a specific agent. For this purpose, the external events stars up the starting state of the activity
diagram, action state models plan, transition arcs model finishing condition of action state and interaction,
and the beliefs are models as object. Furthermore, for the plan and interaction modeling, the AUML's
sequence diagram can be exploited.
Figure 18: Activity Diagram
The activity diagram shows the general scenario of the MS disease model. In our developed model the
start button is "go" button which act as external event to start the simulation. When users, press start
button, then simulation starts. Go button checks the condition either BBB is broken, then agents are
allowed to go into the brain. In brain T-cells (T-reg, and T-eff) search for viruses. If they find a virus in
their radius, then they change their state from inactive to active else they search for viruses. In the same
way, virus search for T-cells to make them active and die. In the case of A.T-reg, it searches for A.T-eff.
If it finds any A.T-eff agents in its radius, then it kills them and produces a duplicate of A.T-reg agent.
The A.T-eff agents search for BWM, if they find BWM then, they damage BWM and produce duplicate
of A.T-eff, Cytokines and at the end they die . The cytokine agent moves randomly and if it finds BBB
then it damages BBB and die. The BBB agent checks a condition that if it is damaged, then allow agents
to move towards brain otherwise block them. And with time BBB try to recover from damage. The BWM
checks a condition if any A.T-eff attacks on myelin then it became damaged and checks another condition
if the damaged portion is recoverable then recovered otherwise remain damaged. The damaged portion
shows the severity of MS in the model.
3.4 MaSE
At the very first time, MaSE methodology was proposed by Deloach in 1999 in the article [47] and after
that, the improved version of MaSE methodology was proposed by Deloach, Matson and Li in 2003 for
the development of a team of rescue robots which are autonomous, heterogeneous searcher, and rescuer
[48]. Several attempts have been made at creating methodologies and their tools for building autonomous,
heterogeneous, distributed and complex dynamic systems. However, mostly the proposed methodologies
and their tools have focused on either the agent architecture, or the methodology lack of sufficient details
to adequately support the designing of complex systems[49].
The significant purpose of the development of this methodology was, MaSE should be independent of
any particular agent-based system architecture, specific agent architecture, programming language, and
precise agent communication framework. Then the developed methodology proved that, the MaSE
seemed a good fit for the cooperative robotic systems[48].
The main objective of this methodology is, guide a system developer throughout the system development
process, by following the set of interrelated system model[48][50]. The previous research on intelligent
agent has focused, on the structure and the capabilities development of an individual agent. Now,
researchers have realized that to solve the complex system's problems, agents coordination is mandatory
for the heterogeneous environment.
The one another constructive aspect of this methodology is, it has its own developmental tool
"agentTool". Same as the MaSE methodology, the agentTool is also independet of the particular agent
architecture, the agent's programming language, and the agent communication language. The analysis and
design phases perform a transformation that shows how to derive new models from the existing models.
The MaSE is implemented after gathering requirements, which further divided into sub-goals as defined
in the article[49][51][50].
Figure 19: MaSE Methodology Models [49].
3.4.1 Analysis Phase:
The analysis phase is further divided into three precise phases, such as
1. Capturing Goals:
This phase develops a goal hierarchy of the system to-be, in which each goal has system level objectives.
The developers take system requirements and organize them into a sequential set of system goals. Simply,
The developers collect all the requirement and set them into hierarchy of basic goals of the system to-be.
We developed the goal hierarchy model of MS disease case study, as researchers implemented MaSE in
different scenarios. [48][49]. Goals must be identified through initial system context, which gives a
starting point to the analyst for system analysis.
Figure 20: Goal Diagram
The figure 20 shows that, the main goal of the MS disease model is to simulate the severity of disease.
This cental goal is further decomposed into sub-goals. However, each goal should always be defined as a
system level goal.
2. Applying Use Cases:
This phase develops Use Cases and sequence diagram. The use cases of a developing system explain the
complete scenario that a system would perform in the real operating environment. To put it simply, a use
case of a system explains the sequence of all working events that must be performed by the developed
system, these events may be its failure and hanging events. For the development of a flawless system, the
analyst should develop enough use case which covers every possible event that can occur in the system by
using different data and event scenario.
The sequence diagram helps the developer to understand the overall system scenario and find out all
involved agents with their communication paths. Moreover, the sequence diagrams provide assistance in
capturing the use cases of the system. Furthermore, these use cases would use later in the analysis phase,
in which a particular role would be assigned to a specific goal. In general, a single sequence diagram is
considered as a representative of each use case.
Figure 21: Sequence Diagram
In the sequence diagram, normally roles are represented by rectangular boxes which are placed at the top
place of the diagram and the connecting arrows of the roles represent to ongoing events among the roles.
And the time is assumed as a sequence of events from the top of the sequence diagram to the bottom.
In MS disease' sequence diagram we identified nine agents. To precisely represent the sequence, we used
BBB two time as an agent. If the BBB is broken then, it informs to T-Cells and virus. When virus finds T-
cells around then it makes them active. An A.T-eff damage the BWM and in return produce Cytokine and
duplicate A.T-eff. When A.T-reg finds an A.T-eff in around then it attacks and kills. In return A.T-reg
duplicate itself. In the case of damage, BWM and BBB recover themselves. Cytokine search BBB and
damage it.
3. Refining Roles:
The main objective of refining role model is, transform the structured goals and sequence diagram into
the final rules of the system to-be. Furthermore, these roles provide a foundation the agent class model
and became system goals during the design phase. Refining role is one of the important steps of the MaE
methodology because, the system goals would be satisfied with the only one way, if every single goal is
considered as an identical role, and every role is played by an identical agent class. This phase develops
concurrent tasks and goal hierarchy models. In general, the transformation of goals to roles is a one-to-
one mapping. However, a single role may have multiple goals.
Figure 22: Refined Role Model
In the MS disease model we identified eight roles, which are shown in rectangle boxes. According to
Agent definition we select those entities as an agent who have some responsibilities or activities to fulfill
the system requirement. All selected roles have multiple goals as shown in figure 22.
3.4.2 Design Phase:
1. Creating Agent Classes:
This phase develops agent class diagram, which consists of agent classes and the conversation between
them. The agent class model clearly depicts the overall organization of agents. Additionally, the agent is
an actual instance of an agent class, and class is a template for a single type of agent in the system.
Additionally, the agent is an actual instance of the agent class, and class is just a template for a single type
of active role in the system to-be. In this phase, the agent classes are developed in terms of the roles, that
must be played. And the conversation is developed as protocols, in which they must participate.
Figure 23: Agent classes Interaction
2. Constructing Conversation:
This phase develops conversation diagram of each software agent. The MaSE conversation model defines
the conversation of two software agents as a coordination protocol[1][2]. In a conversational event, two
agent classes participate, one is an initiator and the second is the responder. The actual purpose of the
conversational model is to define the purpose and the detail of the conversation. When a software agent
wants to communicate with the other agent then it sends the initiator message to the partner agent to start
the communication.
When the responder agent received a conversation request, then it compares it with all its permitted
conversations. If it finds a match, then it performs the required task that can be a request of resource, data
share, or a coordination request. Otherwise, the responder agent assumes that the conversation message is
a request to start a new conversation. In this case, the responder compares the request with all its possible
conversations in which this agent can participate and all the agents' with whom it can make the
conversation. If it finds a match, it begins a new conversation.
Figure 24: T-reg Conversation model. This conversation take place between T-reg and Virus.
Figure 25: T-eff conversational model with Virus.
Figure 26: A-T-eff Conversational Model
Figure 27:A.T-reg Conersational Model
Figure 28: Virus makes conversation with T-cells to make them active.
Figure 29: BWM makes conversation with itself.
Figure 30: BBB's conversational Model.
Figure 31: Cytokines Conversational Model.
3. Assembling Agent Classes:
This phase develops the architecture diagram of agent software by integrating different agent classes
into a single model. This development phase is accomplished by accomplishing two sub-steps: 1) by
defining the agent architecture, 2) by defining the all components that makeup the agent architecture
Figure 32: Assembling agent classes.
4. System Design:
Generally, this phase develops the deployment diagrams. This phase receives the agent classes of each
involved agents and introduces them as actual agents of the system to-be. And go through a data flow
diagram to demonstrate the location and control flow of each agent.
Moreover, this final step of MaSE performs the configuration on the involved agents and their
responsibilities by developing deployment diagrams. Which shows, the total numbers of agents, their
location and agent's type within the system.
4. Comparative Analysis
As biological researchers have no clear idea about engineering methodologies. And the other problem is,
at present, more than two dozen AOSE methodologies exist. Therefore, the process of choosing the right
methodology for their specific problem agonize them. To escape from this tension they just model their
problem without any guided methodology. So this hastens development create some serious problems
such as missing essential information about the problem, the wrong decision of choosing development
tool, the doubt on systems authenticity. To overcome all these mentioned nervousness, in this work, we
applied three well known AOSE methodologies: GIA, TROPOS, and MASE on MS brain model. In this
work we performed a comparative analysis of these methodologies to analyze that, which methodology is
more suitable for biological models development.
For comparative analysis, we used a framework that focuses on four major facets of methodologies: 1)
notations and modeling techniques, 2) concept and properties, 3) development process and 4) pragmatics.
To be very clear, that our work would not attempt to state that which the right methodology is. Rather, it
examines the existing methodologies to advise researchers to choose right methodology according to their
specific scenario. A lot of comparative analysis has been done in literature as shown in the below table.
Ref Paper Title
Author
Journal
Year
I.F
2017
Compared
methodologies
[52]
"Multi-agent approach
automated registration"
for cancer
Sanislav et
al
Control Engineering and
Applied Informatics
2010
0.695
[53]
"A Methodology to Evaluate Agent
Oriented
Engineering
Techniques"
Software
Lin et al.
[54]
"A framework for the evaluation of
agent-oriented methodologies"
Abdelaziz,
Elammari
and
Unland
[55]
"A comparative analysis of i* agent-
oriented modelling techniques"
Grau et al.
[56]
"Agent-Oriented Methodologies
Towards a Challenge Exemplar"
-
Yu
and
Cysneiros
[57]
"On the evaluation of agent oriented
modeling methods"
[56]
Proceedings of the 40th
International
Hawaii
Conference
on System
Sciences
4th
Innovations'07:
International Conference
on
in
Information Technology,
IIT
Innovations
In Proceedings of The
Eighteenth
International
Conference on Software
Engineering
and
Knowledge
Engineering
(SEKE'06)
Proceedings
of
the International Bi-
Conference Workshop on
Agent-Oriented
Information Systems
2007
2008
2006
2002
Proceedings
Oriented
Workshop
of Agent
Methodology
2002
[58]
"ASPECS:
An
software process
complex systems"
agent-oriented
for engineering
Cossentin
o et al
Autonomous Agents and
Multi-Agent Systems
2010
2.103
"A COMPARISON OF THREE
AGENT-ORIENTED
SOFTWARE
DEVELOPMENT
METHODOLOGIES : MASE , GAIA
, AND TROPOS"
Jia et al.
In Information,
Computing, and
Telecommunication,
2009. YC-ICT'09. IEEE
Youth Conference on
2009.
2009
"Evaluating how agent methodologies
support
the
normative environment through the
development process"
specification of
the
Garcia et
al.
Autonomous Agents and
Multi-Agent Systems
2015
2.103
[33]
[59]
[60] "METHODOLOGIES
AND F.
Springer
Science
\& 2009
1. GAIA
2. PASSI
3.
4. MASE
INGENIAS
1. Tropos
2. GAIA
3. MASE
1. GAIA
2. MASE
3. HLIM
1. Tropos
2. GBM
3. ATM
4. BPD
5. RiSD
6. PriM
just
They
question
to
methodology
proposed
evaluate
They just proposed terms
for
methodology
evaluation
1. PASSI
2.
INGENIAS
3. ANEMONA
4. GAIA
5. ROADMAP
6. TROPOS
7. PROMETHEUS
8. ADELFE
9. ASPECS
1. GAIA
2. TROPOS
3. MASE
1. OMASE
2. OPERA
3. TROPOS
4. GORMAS
1. GAIA
SOFTWARE ENGINEERING FOR
AGENT SYSTEMS"
Bergenti,
Federico
and
Gleizes,
Marie-
Pierre and
Zambonell
i
Business Media
2. TROPOS
3. MaSE
4.1 The Proposed Evaluation framework
The structure of the proposed analysis framework is taken from the study [60] however, the evaluation
attributes are taken from different literature. The framework is divided into four subparts. 1) concept and
properties, 2) Notations and Modeling Techniques, 3) Development Process, 4) Pragmatics. The proposed
evaluation framework is based on feature analysis attributes. These attributes evaluate feature of each
examined methodology from different aspects. Before implementing the evaluation framework on
methodologies, first of all, we will discuss in brief the main objective of each section of the framework.
4.1.1 Concepts and properties
The "Concepts and Properties" evaluation criteria are important for agent-oriented methodology
evaluation. "A concept is an abstraction or a notion derived from a specific instance within a problem
domain". And the property represents a special characteristic or capability of an agent. This facet is
concerned with the question whether a methodology addresses the basic notions such as concepts and
properties of agents in MAS or ABM. In the following, there are concepts and properties according to
which methodologies should be evaluated.
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[33]
[59]
[60] Gaia V.2
TROPOS
MaSE
✓
✓
✓ ✓ ✓
✓ ✓
✓
✓
✓ ✓
✓
✓ ✓
✓ ✓ ✓
✓
✓
✓ ✓ ✓
✓
✓ ✓
✓ ✓
✓ ✓ ✓
✓
✓ ✓ ✓ ✓
✓ ✓ ✓
✓ ✓
✓
✓
✓ ✓ ✓ ✓
✓
✓
✓
✓
✓
✓
✓ ✓
✓
✓ ✓ ✓ ✓
✓
✓
✓
✓
✓ ✓ ✓
✓
✓ ✓ ✓
✓
✓ ✓ ✓
✓
✓ ✓ ✓
✓
✓ ✓
✓
✓ ✓ ✓
✓
✓ ✓
✓
✓ ✓ ✓
✓ ✓
✓ ✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
and
Concepts
properties
Autonomy
Adaptability
Agent Abstraction
Belief
Communication
Concurrency
Collaboration
Cooperation
Desire
Problem
Decomposition
Events
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Multiple Interest
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Intention
System
Guidance
Interface
Pro-activity
Protocol
Organization
Message
Reactivity
Role
Tasks
Norms
Society
Soci-ality
Service
Agent- Oriented
5. Results and Discussion
In this section we present ……
1. Autonomy: The ability of an agent to perform its tasks without any supervision.
In GAIA analysis phase, each role has some responsibilities, and they are independent to perform their
responsibilities. In the design phase roles are replaced with agents. So autonomy is expressed in the way
that roles encapsulate its functionality.
The requirement analysis phase of TROPOS defines the autonomy through actor's individual goals, the
association between agents, and their plans. In the requirement analysis phase, the agent's autonomy is
expressed by its goal and plan to achieve goals. In the architectural design, the autonomy is stated by
elaborating actors agenda for achieving the system goal. In the detail design phase, the activity diagram,
which expresses autonomy.
In the analysis phase of MaSE, autonomy is expressed by tasks. And agents are responsible to execute
their tasks on its own responsibility. In the design phase autonomy is expressed through agent classes, in
which roles encapsulate their functionality. And this functionality is internal states of agents that not
affected by the environment. The autonomy is followed by each phase of each methodology, therefore;
we can say that autonomy is high in these three methodologies.
2. Adaptability: The ability of an agent or methodology to deal with the variety of computing
environments, and adjust itself according to changing settings.
In GAIA the adaptability is expressed by environmental model. That deal with its internal states and
expresses how other agents and environment effect on it. In TROPOS, activity diagram express the
adaptability and deals with the variety of computing environments, and changing settings. The MaSE
methodology does not support adaptability. The environmental model of GAIA does not explicitly define
the negative or alternate responses to change, however, the activity diagram in TROPOS explicitly
explains both worse and better responses to change. In that way, adaptation is stronger in TROPOS as
compare to GAIA.
3. Agent Abstraction: The ability of a methodology to describe agents using high-level abstraction.
In GAIA, the preliminary role model support agent abstraction by extraction roles from requirement
specification document. And In the design phase the agent model expresses the agent abstraction
property. In TROPOS the actor model in the early requirement analysis phase and the architectural design
phase define the actor abstraction property. In MaSE the role model in the analysis phase and, agent
classes in the design phase expresses the agent abstraction property. Almost all phases of each
methodology support to agent abstraction, therefore agent abstraction is strong in these three
methodologies.
4. Belief: The belief is a faith of an agent, which believes that it is always true about the world.
GAIA does not have a belief. The goals of the actor in TROPOS express the belief concept. In MaSE
methodology, belief is expressed by goals, tasks, and states. MaSE strongly support belief property as
compare to TROPOS and GAIA.
5. Communication:
The interaction model in the analysis and design phase of GAIA defines communication. The activity
diagram explicitly expresses the communication in TROPOS. In MaSE methodology conversation model
defines communication. Due to activity diagram and conversation model the communication is strong in
TROPOS and MASE as compared to GAIA.
6. Concurrency:
In GAIA the service model, in the detailed design phase support concurrency. The activity diagram in the
detailed design phase of TROPOS defines concurrency. In MaSE the role model in the refining role
model explicitly explains concurrency.
7. Collaboration: An agent has methods to cooperate with other agents to achieve goals.
In GAIA's analysis and design phase, the interaction model expresses collaboration between agents. In
the detailed design phase of the TROPOS sequence diagram defines collaboration between agents. The
sequence diagram in the analysis phase and a conversation model in the design phase of MaSE express
collaboration between agents. Due to sequence diagram, collaboration is more precise and strong in
TROPOS and MaSE as compare to GAIA.
8. Cooperation: The cooperation is a collaborative activity with one objective, but it is distributed
among several actors. In cooperation, each agent performs actions according to the shared objectives.
The organizational structure of GAIA expresses the cooperation of agents. The architectural design in
TROPOS expresses cooperation of agents to perform a specific task. The agent class model in the design
phase of MaSE expresses the cooperation of agents. Due to the class diagram, cooperation is strongest in
MaSE as compared to GAIA and TROPOS.
9. Desire: A goal of an agent to be achieved.
GAIA does not support Desire property of the agents. The goal of the agent in TROPOS expresses the
desire of the agent. The goal, task models and state in the conversation model of MaSE express the desire
of agents. Desire is much stronger in the MaSE methodology after that TROPOS have strong desire
property. And GAIA does not have desire property.
10. Problem Decomposition: The ability of a methodology to divide the large problem into smaller and
more manageable parts. Basically, this property tackle complexity.
The analysis phase, architectural design, and detailed design phases of GAIA support problem
decomposition property. The earlier requirement, late requirement, architectural design and detailed
design phases of TROPOS support problem decomposition property. The analysis and design phases of
MaSE support problem decomposition property. The problem decomposition is stronger in these three
methodologies since each phase of these methodologies is divided into subparts. The division of phases
helps to decompose and understand complex problems.
11. Events: The ability of a methodology to control event triggering. The events trigger the interaction
and agents become responsible for a new goal.
In GAIA the analysis phase, the sub-organization model interacts with the environmental model, role
model, and interaction model. And these models interact with the organizational rules model. The
interaction generates events in models to interact with other models to perform a required task. The same
event phenomena happen in the design phase. In this way, GAIA expresses event generation property. In
TROPOS early requirement or any other phase does not interact with late requirement phase and with
others phases. Since TROPOS phases do not generate events or interaction, that's why this methodology
does not express an event concept. MaSE methodology's phases interact with other phases and models in
a phase interact with each other, this interaction expresses an event concept. Only TROPOS does not
support to event concept.
12. Multiple Interests: At a time, an agent may have multiple tasks such as co-operate with other agents,
be independent, or help the other agents to achieve a goal in the environment.
GAIA does not support this concept. TROPOS express this concept through goals, plans, and resources.
Since, TROPOS agents interact with other agents through goals, plans, and resources. Since they have
multiple interests to achieve their goals or to interact with other agents. The MaSE methodology does not
support multiple interests concept. Only TROPOS have a strong concept of multiple interests.
13. Intention: A fact that represents the way of realizing a desire, sometimes referred to as a plan.
The GAIA does not support intention concept. The agents of TROPOS express intention through its
plans. The goals, task, and states in a conversation model of MaSE express the intention concept. The
Intention is stronger in MaSE since three different models of MaSE express this concept. After the MaSE
intention is strong in TROPOS.
14. System Interface Guidance:
The GAIA does not support interface to the external world. The TROPOS methodology does not offer an
interface for the external world. The MaSE methodology does not support or offer an interface for the
external world. These three methodologies lack an interface to the external world. This flaw should
improve in all these methodologies.
15. Pro-activity: The ability of an agent and methodology to pursue new goals.
The service model in the detailed design phase of GAIA expresses pro-activeness. The plans of an agent
to achieve a goal express the pro-activeness in TROPOS. Any phase of MaSE methodology does not
support pro-activeness. The pro-activity is high in GAIA and TROPOS and MaSE lack of this property.
16. Protocol: A set of messages that defines the purpose and detail of a particular interaction among the
agents.
The role and interaction model in the architectural design phase and agent and service model in the detail
design phase of GAIA explicitly define protocol concept. The sequence diagram in the detailed design
phase of TROPOS somehow defines protocols, however; this methodology does not clearly define this
concept. The conversation model in the analysis phase and a conversation model in the design phase of
MaSE explicitly expresses protocol concept. GAIA and MaSE methodologies strongly support protocol
concept. On the other side, TROPOS does not express this concept explicitly. TROPOS lack conversation
model.
17. Organization: A group of agents working together to achieve a common purpose. An organization
consists of roles that characterize the agents, which are members of the organization.
The Sub-organization model in the analysis phase and the organizational structure model in the design
phase of GAIA express organization concept. The architectural design phase of TROPOS expresses
organization concept. The agent architecture model in the detailed design phase of MaSE expresses
organization concept. GAIA strongly expresses an organization model as compared to TROPOS and
MaSE.
18. MESSAGE: The message is a request for making conversation between agents for resources and task
completion.
The protocols in the interaction model, agent model and services model of GAIA express this concept.
The activity diagram in TROPOS expresses message concept explicitly. The conversation model in the
MaSE defines message concept. The message concept is defined by all these methodologies.
19. Reactivity: The ability of an agent and methodology to respond to changes in environment on time.
The responsibilities of a role in the role model and agent model show liveness. This liveness expresses
reactivity in GAIA. The state change in activity diagram and sequence diagram of TROPOS methodology
expresses reactivity. The conversation model and sequence diagram in MaSE express reactivity. All these
methodologies somehow meet the requirement of reactivity. There is no explicit or clear model to support
reactivity.
20. Role: An abstract level description of the agent's function, its services or its specific identification
within a group.
The role models in the analysis and design phases of GAIA define the role concept explicitly. TROPOS
does not express role concept. The role model in the analysis phase of MaSE expresses the role concept.
The role concept is strong in the GAIA methodology as compared to MASE. In TROPOS role abstraction
concept is weak.
21. Task: A precise piece of work that is assigned to the agent of the system to be in the form of its
function.
Task represents to agent responsibilities. In GAIA the role model and service model express task concept.
In TROPOS methodology the task concept represents to a capability of the agent. And capability is
expressed by activity diagram. The concurrent task model in MaSE expresses a task concept explicitly. In
MaSE the task concept is stronger as compared to TROPOS and GAIA.
22. Norms: A set of rules that characterize a society and the agents of this society are bound to follow all
the mentioned norms.
GAIA: The norms defined by the organizational rules model in the analysis phase of GAIA. The norms
are defined by the organizational structure in TROPOS. The agent architecture in the design phase
expresses norm concept in MaSE. In GAIA norm concept is stronger as compared to TROPOSE and
MaSE methodology.
23. Society: A collection of agents and organizations that collaborates to promote their individual goals.
The organizational rules and organizational structure models define society in GAIA. The organizational
structure in TROPOS somehow way defines society. The agent architecture in MaSE somehow defines
society. GAIA methodology has a strong society's concept as compared to TROPOS and MaSE.
24. Sociality: The capability of an agent to communicate with the other agents of the system by sending
and receiving messages and cooperate with them to perform a specific task.
In GAIA the sociality is expressed within the interaction model that defines the communication links
among agent types. In TROPOS the sociality is expressed by the system model in the late requirement
phase. The sociality is somehow expressed by agent architecture in MaSE. All methodologies weakly
support sociality concept.
25. Service: The service is a "knowledge level analogue" of an agent's operation to achieve a specific
goal.
The service is expressed by service model in the detailed design phase of GAIA. The TROPOS
methodology does not support service concept. The MaSE methodology does not support service concept.
Only GAIA methodology explicitly expresses service feature.
26. Agent-Oriented: The agent-oriented features focus on whether the methodology addresses Agent-
based features during the analysis and design.
GAIA methodology somehow defines all attributes of an agent such as pro-activity, autonomy, reactivity,
and sociality so, it is an agent-oriented methodology. The TROPOS methodology somehow defines all
attributes of an agent such as pro-activity, autonomy, reactivity, and sociality so, it is an agent-oriented
methodology. The MaSE methodology somehow defines all attributes of an agent such as pro-activity,
autonomy, reactivity, and sociality so, it is an agent oriented methodology. All these methodologies are
agent-oriented.
4.1.2 Notations and modeling techniques
Notations are a set of symbols that technically represent agents and their functional goal in a system to-be.
These modeling techniques collectively build a precise model that represents developing system at
different levels of abstraction and express their different facets such as structural and behavioral sides.
This section deals with the properties of notions which a modeling methodology should have. The list of
properties is given below:
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[33]
[59]
[60] GAIA V.2
TROPOS MaSE
and
Notation
Mdeling
Technique
✓ ✓
✓
✓
Agent Attributes
✓ ✓
✓ ✓
✓
✓
✓ ✓ ✓
✓
✓
✓
✓ ✓ ✓
✓ ✓
✓ ✓
✓
✓
✓ ✓
✓ ✓
✓ ✓
✓
✓
✓
✓ ✓
✓ ✓ ✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Accessibility
Analyze-ability
Complexity
Management
Dynamic Structure
Expressiveness
Consistency
Open system
Execute-ability
Unambiguity
Consistency
Trace-ability
Preciseness
System View
Modularity
1. Agent Attributes: The ability of a methodology that concern about the description of the agent's parts
that make up the internal structure.
The internal structure of the agent is strong in the GAIA as the role model of the analysis phase
transforms to agent model in the design phase. This strength is due to that every agent plays a specific
role and is independent in making decisions. In TROPOS the actor diagram of the requirement phase
transforms to the activity diagram in the design phase. So in this sense, the internal structure of an agent is
strong in TROPOS. In MaSE the role model transforms to agent model in the design phase. All these
methodologies explicitly manifest agent attribute for system development.
2. Accessibility: The ability of a methodology that assist the developers to easily adapt it, understand it
and implement it.
All GAIA models and the phase distribution are simple and understandable. Engineers can use easily
these models and can develop complex systems. Gaia does not support accessibility when developers
transform its models to the overall system model. The notation understandability and modeling in
TROPOS is easy. However, the transformation of models from rational models to activity diagram and
sequence diagram is difficult. Same like GAIA, MaSE methodology has simple models that can be
understood by developers easily, but the transformation of these models into the overall system model is a
difficult task. All these methodologies in some way are accessible; however, some improvement is
required in this area.
3. Analyze-ability: The capability of a methodology to check the internal consistency of each agent. And
also point out the unclear aspects of models and agents.
GAIA does not deal with the analyze-ability. TROPOS check analyze-ability of models through the agent
plan, resource, and goals. However, the analyze-ability compromise between rational models and
architectural structure. The MaSE methodology provides analyze-ability within models through the
transformation of models from the analysis phase to design phase. However, analyze-ability compromise
when transforms activity diagram into a sequence diagram. All these methodologies need a development
in this area.
4. Complexity Management: An ability of the agent-oriented methodology to deal with the different
levels of complexity. For example, sometimes only high-level requirements are needed while in the
other situation more details are required. In this case, the agent oriented methodology should provide
all level information.
GAIA gives somehow favor to manage complexity by giving different levels of abstraction by different
models. However, it does not have a hierarchical structure to manage system's complexity. The TROPOS
controls complexity by goal model in the early requirement phase, but it cannot control the details within
it. The hierarchical structure somehow controls complexity. The MaSE methodology provides an
abstraction of almost every concept of the goal, agent, conversation model, in this way it controls a
limited extent of complexity. However, it does not have any hierarchy model to control complexity and
detail of complex tasks. Generally, GAIA and MaSE are suffering from this deficiency. TROPOS is
somehow better on this property.
5. Dynamic Structure: The ability of a methodology to provide support for the dynamic structural
reconfiguration of the system.
The GAIA does not deal with a dynamic structure. The TROPOS somehow deal with the dynamic
structure of activity diagram. The MaSE does not deal with the dynamic structure. The future work should
be done on the dynamic structure feature of all methodologies.
6. Expressiveness: The ability of a methodology of presenting system concepts such as system structure,
encapsulated knowledge of models, ontology structure, data flow, control flow, and concurrent
activities of the involving agents.
Due to the generic structure, GAIA can handle and model a large variety of systems. However, the system
structure of the system to-be is not presented explicitly. Usually, TROPOS helps to develop BDI (Belief,
Desire, and Interaction) systems. However, the system structure of the system to-be is not presented
explicitly. The MaSE methodology explicitly describes the system structure by using an agent
architecture and system design diagrams. However, the encapsulated knowledge of the system is not
presented explicitly. All methodologies express expressiveness in a limited way.
7. Consistency: The ability of methodology to maintain the quality of a system throughout the
development process.
Each model in the GAIA explicitly defines each concept clearly, means that agent model, role model,
environmental model, service model. There is no chance of ambiguity between models. In TROPOS the
rational model and the architectural design confuse the developers due to ambiguity. However, in detail
design phase it controls this lack through activity diagram. The MaSE explicitly express consistency
within each model through control flow and data flow. Moreover, each model clearly defines a single
concept such as goal model, role model, and agent model, etc. Each methodology tries to be consistent in
some way. However, MaSE strongly expresses the consistency as compared to GAIA and TROPOS. On
the second number, GAIA is stronger in consistency than TROPOS.
8. Open System: The ability of a methodology to add or remove new agents in the development system.
The GAIA methodology completely supports open system property because this methodology has
separate models for each concept. New agents, roles, and protocols, etc easily can be removed or add into
the system. TROPOS methodology does not support to open system property due to activity and sequence
diagrams. Same like GAIA, MaSE explicitly supports to open system property. GAIA and MaSE are
strong, according to this aspect.
9. Execute-ability: The ability of the agent-oriented methodology, to provide the facility of performing a
simulation to validate the system specification. Or at least generate a prototype of the overall system
to-be.
GAIA does not deal with the execute-ability issue. TROPOS dealt with the execute-ability by using
JACK. The MaSE methodology has a developmental tool "agentTool" which support partially to code
generation. MaSE methodology is stronger as compared to the GAIA, and TROPOS, according to this
aspect.
10. Unambiguity:
As GAIA has clear models for each concept, so in this way, it tries to overcome this issue. In TROPOS
the architectural design and rationale models lead to ambiguity and create confusion for developer's
understanding. Same like GAIA, MaSE has clear models for each concept. MaSE and GAIA
methodologies have less ambiguity as compared to TROPOS.
11. Trace-ability: The ability of a methodology to handle the main concept of the system to-be throughout
the system development.
The models in GAIA's analysis phase become input for models in the design phase. In this way we can
say it maintain trace-ability throughout the development phases. Same like GAIA, the TRPOS
methodology also maintains trace-ability. The models in the analysis phase of MaSE become the base for
the design phase models. Each mode can be traced by its base model. All these methodologies in
somehow way fulfill the trace-ability aspect.
12. Preciseness: The ability of the methodology to handle the ambiguity throughout the system
development. It assists the developers in avoiding the misinterpretation in system development.
Each model has a clear meaning and interpretation in GAIA methodology. This edge makes GAIA
accurate. As the TROPOS base on i* model that has clear notation and meaning of each symbol. Thus, it
prevents users from misinterpretation. Same like GAIA, MaSE model, symbols, and notations have a
clear meaning and interpretation. This advantage makes the MaSE accurate and prevent the developers
from misinterpretation. All these methodologies are precise.
13. System View: The ability of a methodology to provide a microscopic system oriented model, to
understand the whole working scenario of the developing system.
GAIA does not promote system view feature. Since, this methodology has a deficiency of hierarchical
model and overall system model. In TROPOS the activity diagram demonstrates the general view of the
system model. In MaSE, agent classes depict a vague concept of system view. Only TROPOS has
somehow clear abstraction of system view.
14. Modularity: The ability of a methodology to develop a system in an iterative way. That allows adding
new requirements without affecting the existing specifications.
GAIA is modular, due to its models (agent, service and, role). The changing in the role does not affect the
whole system. This change only influences the internal structure of an agent or role. Within the TROPOS
modularity is fully supported. Same like GAIA, MaSE support modularity because of models (task, agent,
goal). TROPOS strongly support modularity as compared to GAIA and MaSE.
4.1.3 Development Process
A development process is a step by step guideline for developing a system from scratch. This process
consists of a series of actions, functions, and models, that when performed, then the outcome is an
operational computerized system. Basically, this section of evaluation framework deals with the different
facets of system development process. The terminologies for checking the developing process are given
in the below table:
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[33]
[59]
[60] GAIA V.2 TROPOS MaSE
✓ ✓
✓
✓
✓ ✓ ✓
✓
✓ ✓ ✓
✓
✓ ✓ ✓
✓
✓
✓
✓
✓ ✓ ✓
✓ ✓
✓
✓
✓ ✓ ✓
✓ ✓ ✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓ ✓ ✓
✓ ✓ ✓
✓
✓
✓
Development
Process
Architecture
Design
Requirement
Norms
Legal Document
System Design
Development Life
Cycle
Development
Context
Implementation
SDLC coverage
Development
Approach
System
Specification
Project
Management
guideline
Verification
Validation
and
✓ ✓ ✓
Quality Assurance
1. Architecture Design: The ability of a methodology to facilitate design by using patterns or modules.
GAIA provides almost all modules of a system such as interaction, role, environmental, service, agent
models and also provides an architectural design of developing system. The architectural design phase of
TROPOS explicates the overall system's global architecture in terms of sub-systems, interconnected
through data and control flow. The MaSE have all architectural bricksin the orm of models such as agent
model, goal model, role model, conversation model and task model. The agent architecture model
represents the architectural design of the whole system. All these methodologies have architecture design
property for system development.
2. Requirement Norm: The ability of a methodology to identify organization's norm during
requirement analysis.
GAIA does not support the system requirement. This methodology performs development process after
requirement specification. During early and late requirement analysis phases of TROPOS, this
methodology identifies and formalize norms for the system to be developed. Same like GAIA, the MaSE
methodology have analysis and design phase. It does not support system requirement analysis phase.
Only TROPOS has a strong feature of building system norms.
3. System Design: The ability of a methodology to introduce the normative environment of the system
as an integral part of the development phase.
GAIA does not support system design. TROPOS does not support system design. MaSE expresses
system design through deployment diagrams. More work is required in the system design phase of all
methodologies.
4. SDLS Coverage: The ability of an agent-oriented methodology to include elements of software
development.
GAIA covers only two main phases of the development process such as analysis and design phase. This
coverage is not sufficient for developing an outstanding system. The TROPOS methodology covers
almost all phases of the development cycle. However, it does not deal with the testing stage. MaSE covers
development cycle from analysis to implementation phase. However, its goal model, use cases and
sequence diagrams in somehow manners support requirement phase. MaSE does not support only to
testing phase. Only GAIA methodology has a deficiency of DLC coverage.
5. Development Context: The ability of a methodology to be remolded according to users need such as
creating new software, reverse engineering, re-engineering systems by using reusability or creating
new system property.
GAIA has the ability of creating new software, designing and re-engineering systems by using reusability
propert or creating from scratch. It does not address implementation phase and does not support classical
reverse engineering . The TROPOS methodology can be used for creating software systems from scratch
and for prototyping, re-engineering and designing systems by using reusability property. However,
TROPOS does not support reverse engineering. Because when going from one stage to the next stage,
then several concepts undergo significant changes. The MaSE can be used in creating new systems,
designing systems from scratch and re-engineering with reuseability property. However, MaSE does not
support reverse engineering. All methodologies are incapable of reverse engineering.
6.
Implementation Guidance: The capability of a methodology to deal with coding issues, quality,
performance, libraries and debugging.
GAIA does not support implementation phase. TROPOS has implementation guidance property as it
suggests to use JACK toolkit since it easily maps a BDI architecture. The MaSE methodology has a
graphical agentTool, which is a fully human-interactive tool and guides each step of MaSE development.
The agentTool have the ability for automatic verification of inter-agent communication, semi-automated
design, and code generation for multiple MAS framework. These all methodologies are strong in
implementation guidance property except GAIA.
7. Development Approach: The attribute of a methodology to guide system development in a specific
way, such as top-down, bottom-up, mix.
GAIA follows a top-down development approach. TROPOS follows a top-down development approach.
MaSE follows a top-down development approach. All these three methodologies follow a top-down
approach.
8. System Specification: The ability of a methodology to accurately interpret the problem from
specification document. And confirm that this is the right problem to be solved.
The analysis phase of GAIA, manifest the requirement specification of the system in the form of sub-
organizations, role model, interaction model, and organizational rules. The actor model in the early
requirement analysis phase of TROPOS embodies the requirement of the system. In the analysis phase
the MaSE express system specification in the form of goal model, use cases, sequence diagram, roles and
task model. All these methodologies identify and specify system specifications.
9. Project Management guideline: The ability of a methodology to effectively and efficiently guide all
aspects of a project from conception through completion.
This issue is not dealt with in the GAIA. The TROPOS has lack to deal with project management issues.
This issue is not dealt with in the MaSE. ALL these methodologies are unable to deal with the project
management issues.
10. Verification and Validation: The ability of a methodology to provide a way for formal verification
and validation.
GAIA performs verification and validation during the transformation of preliminary roles to role model.
In TROPOS, there is no coverage checking with respect to the initial requirement. However, TROPOS
extension "formal TROPOS" can be used for verification and validation. MaSE performs verification
over its models by checking consistency, deadlocks and unused elements between the stages. ALL these
methodologies have no specific verification and validation method. However, TROPOS overcome this
lack by TROPOS formal method. GAIA and MaSE overcome this lack of transformation of models.
11. Quality Assurance: The ability of a methodology that ensures that the developed software meets
and compiles with defined or standardize quality specifications.
This issue is not dealt with GAIA. This issue is not dealt with TROPOS. This issue is not dealt with
MaSE. All methodologies have lack quality assurance property.
4.1.4 Pragmatics
The pragmatics of a methodology, determine the industrial success of a methodology. Pragmatics
determine that a methodology is applicable in the industry for developing complex and distributed
systems. Moreover, this section determines that is the considered methodology have the ability of project
management and determine that the considered methodology can be adapted within the organization
according to the organizational budget and experience. The pragmatics checking terminologies are given
below in the table.
[54]
[33]
[59]
[60] GAIA V.2
[55]
[56]
[53]
[52]
✓ ✓
✓ ✓ ✓
[57]
[58]
✓
TROPOS MaSE
✓
✓
✓
Pragmatics
Tools Available
Required Expetries
✓ ✓
✓ ✓ ✓
✓ ✓
✓
✓ ✓ ✓ ✓
✓ ✓ ✓
✓ ✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Modeling
suitability
Domain
Applicability
Scalability
Resources
Language
Suitability
1. Tools available: The ability of a methodology to guide about available tools and tools ready to use.
The GAIA methodology does not provide any automated Tool. TROPOS provides several automated
tools for animation, model checking, and reasoning. The famous one is JACK tool. The MaSE
methodology has agentTool, which represents the behavior of internal components of agents and
protocols. All these methodologies are somehow strong on this property except GAIA.
2. Required expertise: The ability of a methodology, be as simple as possible that users don't require
any background knowledge.
The GAIA required a strong background knowledge of logic and temporal logic. These logics reduce its
accessibility. Since many developers do not know or do not want to get familiar with logic. TROPOS
does not require strong background knowledge. Since, its modeling notations are very simple. Same like
GAIA, the MaSE methodology requires a strong background knowledge of logic and temporal logic.
TROPOS is easy to develop at the first time as compared to MaSE and TROPOS.
3. Modeling Suitability: The ability of a methodology, consisting of a specific architecture.
The GAIA methodology has no specific modeling architecture. So, designers have no need of
architectural information for development. TROPOS has BDI architecture, So designers have to explicitly
describe beliefs, desires, and intentions. MaSE has no specific architecture, So designers have no need of
architectural information for development. Only TROPOS methodology has architecture, and developers
need to search if their system has BDI properties then they use this methodology.
4. Domain Applicability: The ability of a methodology to be suitable for a variety of domains.
GAIA is suitable for dynamic-open (where the agents are not known) and dynamic-close (where the
agents are known) systems. However, this methodology is not suitable for developing applications with
dynamic characteristics such as goals generation. Generally, TROPOS is suitable for developing
componentized systems like e-business applications and BDI based systems. MaSE can be used for
various types of agents and systems. Only TROPOS and MaSE support to dynamic characteristics in
complex system development.
5. Scalability: The ability of a methodology to be adjusted to dealt with various application sizes.
GAIA has a simple structure, that's why it can be fitted in different sizes of applications. TROPOS does
not provide information for subsets and supersets. So, developers have no idea of exact size must be to
avoid complexity. MaSE also does not provide information for subsets and supersets. Thus, developers
have no idea of exact size must be to avoid complexity. GAIA has strong scalability characteristics as
compared to TROPOS and MaSE.
6. Resources: The ability of a methodology to be mature enough that, publish material, tools and
training groups are easy to find.
The GAIA methodology does not provide any automated Tool. However, published papers describe the
implementation of this methodology in detail. The TROPOS's Tools and published material are available.
It fully supports system development from scratch. The MaSE Tools, web site, and various published
studies are available. It fully supports system development from scratch. MaSE and TROPOS are well
built in resources as compared to GAIA.
7. Language Suitability: The ability of a methodology to be coupled with a particular implementation
language or a specific architecture.
As GAIA does not have any specific architecture and language. And it does not refer to the
implementation issues. Thus, the specifications made using GAIA can be implemented in any language.
As TROPOS based on the BDI concept, so its implementation will be biased towards BDI direction.
MaSe does not target to any specific architecture, language, and a specific framework. MaSE and GAIA
are more independent from any language suitability than the TROPOS.
Exploratory Agent Based Modeling (EABM):
Cognitive Agent-Based Computing (CABC) is a unified framework, proposed by Niazi and Hussain in
the article [61] for modeling the complex systems of agents. This framework has four levels for
developing complex systems. Exploratory Agent-Based Modeling (EABM) is the second level of CABC
framework. The EABM guides the researchers to develop a proof of concept model of the developed
system with the goal of performing multiple simulations for improving understanding about a particular
real-world complex system [62].
Breeds:
There are two types of breed.
1. T-Cells of immune system
a) Regulatory T-Cells (T-reg)
b) Effectors T-Cells (T-eff)
c) Active Regulatory T-Cells (A.T-reg)
d) Active Effectory T-Cells (A.T-eff)
e) External environmental virus (virus)
f) Cytokines
2. Brain organs
a. Blood Brain Barrier BBB
b. Brain White Matter (Myelin)
a) T-reg
This agent search for the virus, if virus found, then it changes its state from inactive T-reg to active A.T-
reg.
1. Move randomly
2. Search virus
3. Transform from T-reg to A.T-reg
b) T-eff
This agent search for the virus, if virus found, then it changes its state from inactive T-eff to Active A.T-
eff.
1. Move randomly
2. Search virus
3. Transform from T-eff to A.T-eff
c) A.T-reg
When a virus attacks any inactive regulatory cell or inactive regulatory cell became successful to catch
the virus , then T-reg regulatory cell changes its state from inactive to active A.T-reg and its color turns
from black to blue. When A.T-reg blue agent finds any active Effector A.T-eff in radius, then it kills
effector cell and does these jobs.
1. Move randomly
2. Kills active Effector T-cells
3. Gain energy
4. Duplicate A.T-reg
d) Effector T-Cells
When a virus attacks any Inactive Effector T-Cell T-eff or T-eff cell became successful to catch the virus,
then effector cell changes its state from resting T-eff to active A.-eff and color turns from white to red.
Active effector T-cell perform these tasks;
1. Move randomly
2. Search Myelin
3. Eat myelin
4. Gain energy
5. Duplicate A.T-eff
6. Produce Cytokines
e) Virus
The virus agent search for T-Cells T-reg and T-eff. If the virus finds any T-cell in its radius, then it makes
them active. It performs the following tasks.
1. Move randomly
2. Search T-Cells
3. Make T-Cells active
f) Cytokines
Cytokines search for BBB. If they become successful to find the BBB then they attack BBB and damage
it. After the attack, it died. It performs the following tasks.
1. Move randomly
2. Search BBB
3. Damage BBB
4. Die
2. Brain organs
a) Blood Brain Barrier BBB
BBB is a brain organ that stops all unwanted minerals, molecules, and cells to enter into the brain. If any
cell tries to enter the brain, then it bounces back that cell. If the BBB is damaged, then it repairs BBB. It
performs the following tasks:
1. Stop cells' entrance into the brain
2. Repair damaged BBB
b) Brain White Matter (Myelin)
Myelin covers the axonal part of the neuron cells and maintains communication between two neurons. If
any damaged occur due to effectory cells attack, then it repairs the myelin to maintain communication. It
performs the following tasks:
1. Maintain communication
2. Repair myelin
Global Variables:
Global variables are accessible anywhere in the simulation code and store particular value for agents
during the one continued simulation experiment. There are four preferred global variables.
1. The recoverable brain patches (recoverable)
2. The unrecoverable brain patches (unrecoverable)
3. The total myelin in each brain patch (initmyelin)
4. The total energy of each agent (total)
Besides these variables, there are many other input variables in the user interface which can be adjusted
by the user manually or else by using behavior space.
Procedures:
1. Setup
2. Setup-world
3. Reproduce-Tregs
4. reproduce-effectors
5. grow-myelin
6. eat-myelin
7. go
8. move
9. death
10. count-virus
11. count-Tegg
12. count-effectors
13. bounce
14. display-labels
15. catch-effectors
16. count unrecoverable
17. display label
18. do-ploting
Figure 33: These are various configurable values for the simulation
1. Setup:
First of all setup functions clears all variable, agents and environment of the simulation. This is important
because we need to run multiple simulations on a specific hypothesis. This procedure divides simulation
environment into three types of patches 1) Blood / LYMPHATIC system, 2) Brain axonal area and 3)
Blood Brain Barrier. After creating patches, the next step is to create three types of agents (regulatory,
effectors, viruses).
Create-Tregs:
This function creates regulatory T-cells T-reg according to its input variable and sets its attributes. The
input variable is adjustable from 1 to 2000 numbers. Finally, this procedure adjusts regulatory cells xy-
coordinate in simulation environment.
Create-effectors:
This function creates effectors T-cells T-eff according to its input variable and sets its attributes. The
input variable is adjustable from 1 to 2000 numbers. Finally, this setup procedure adjusts effector cells
xy-coordinate in simulation environment.
Create-viruses:
This function creates an external environmental virus, according to its input variable and sets its
attributes. The input variable is adjustable from 1 to 100 numbers. Finally, setup procedure adjusts virus
cells xy-coordinate on simulation environment.
At the end, setup function calls the display-labels function. The detailed description of each function
follows along with their function.
2. Setup-world:
This function divides 51*51 grid into patches of size 9. 0 to 13 y-coordinate patches are declared as blood
portion and from 13 to 15 y-coordinate are declared as BBB portion. 15 to 35 y-coordinates represent
brains white portion. 35 to 37 again represent BBB and 38 to 50 represent a blood portion of the brain.
Figure 34: model view after pressing the setup button
3. Reproduce-Tregs:
This procedure reproduces regulatory T-cell according to probability. If random float of 100 is less than
reg-reproduce the input variable, then it checks a condition. If the number of turtles on a single patch is
less than the patch density, then it cuts down turtle energy by half and produces another regulatory T-cell.
Otherwise, move forward according to move function.
4. Reproduce-effectors:
This procedure calculates the probability of reproduction of effector T-cells. If an effector successfully
eats myelin and causes axonal damage then, it will be simulated to duplicate. And the duplication is
modeled as a stochastic Bernoulli process. The duplication probability p is calculated for every
duplication process according to the following law:
P = effector-dupl × myelin2
/ init-mye2 × mean-Tregs / Treg-here + mean-Treg
Where myelin indicates the quantity of myelin in the current patch, effector-dupl is a duplication constant
representing the maximum duplication rate of Teff, mean-Treg is a given threshold and Treg-here is the
number of Tregs in a given radius Treg-radius. The term myelin2
/ init-mye2 gives higher probabilities to
duplicate if the patch has higher quantities of myelin, where the term mean-Tregs / Treg-here + mean-
Treg is used to model the down-regulation of Teff duplication rates by Treg actions.
5. Grow-myelin:
Each patch with color gray shows myelin amount. Initially, this procedure sets each patch myelin to 100.
For each time when an active effector eats myelin then, that's patched myelin amount decreased by -5 and
then this procedure again check condition if the myelin amount reaches to zero then, that patch color turns
to black. After each attack, the patch color turns from gray to black. After that, this black patch recovers
myelin according to rec-mye input variable.
6. Eat-myelin:
This procedure sets a condition that if any active A.T-eff is on any patch with color gray and the patch
myelin amount is greater than the eat-myelin input variable then minus -5 from that's patched myelin
amount. If myelin amount reaches zero then turn patch color from gray to red.
7. Go:
First of all, this procedure calculates the total number of unrecoverable brain patches. Unrecoverable
patches are those whose myelin amount reaches to 0 and their color turn from gray to black. After that, it
calculates myelin of each patch and subtract it from the total myelin amount and assign the total myelin
amount to each patch, which amount it has at this simulation time. This procedure also calculates the total
number of recoverable patches. If any patch has some amount of myelin then it can recover from multiple
sclerosis. Recoverable patches are calculated by subtracting unrecoverable patches from the total number
of patches.
I.
II.
III.
Create regulatory: This procedure creates regulatory T-cells according to a specific probability.
The probability to create a regulatory T-Cells is, if random float of 1 is less than 10/365 then
create a regulatory T-Cells according to the slider value.
Create effectors: This procedure calculates specific probability to create an effector T-cells. If
random float of 1 is less than 10/365 then create an effector T-cells according to the slider value.
Create viruses: This procedure also creates external environmental viruses, according to a specific
probability. The probability to create viruses is if random float of 1 is less than 10/365 then create
regulatory T-cells according to the slider value.
After creation agents, this procedure asks effectors call bounce and move function. After that, this
procedure asks active effectors eat myelin and produce 1 cytokine against one myelin attack. It also
instructs inactive effectors that try for activation. If any virus is in radius, then they change their state to
active else move forward and try to catch the virus.
After creation cytokines, this procedure instructs cytokines that if they are at patch with color green they
change the patch color to red. Means if any cytokine collides with the blood brain barrier then barrier
destroy and that patch start countdown. When the countdown reaches 0 then they again became a barrier.
Go function ask regulatory T-cells call to move and bounce function. They also instruct that if regulatory
cells are active then catch effectors and produce one another regulatory cell or else try to catch virus to
make himself active.
At the end, his function asks viruses call move, bounce and death function. And instruct patches call
grow-myelin function.
8. Move:
This procedure decreases turtle's energy at every single step and sets turtle heading random 360. After
that it forces turtle to take the next random step.
9. Death:
This procedure order to each turtle if its energy level is zero, then it must be die at any place in the
simulation environment.
10. Count-virus:
This procedure reports total number of virus for every moment in the running simulation.
11. Count-Teg:
This procedure reports a total number of active regulatory T-cells with the color blue for every moment in
the running simulation.
12. Count-effectors:
This procedure reports the total number of active effector T-cells with the color red for every moment in
the running simulation.
13. Bounce:
This procedure ask all turtles if the next move patch's color is green, then bounce back and set heading
random 360. This green patch is a blood Brain Barrier which blocks turtle's movement.
14. Display-labels:
This procedure asks turtles if the energy level switch is on then, they display their current energy level
during the current simulation experiment.
15. Catch-effectors:
This procedure asks active regulatory T-cell if any active effector T-cell in its radius, then kill him and
breed one regulatory T-cell. If there is no one effector in radius, then try to catch effectors with the color
red.
16. Count unrecoverable
This procedure counts all patches with color black. These black color patches show to damaged brain that
is unrecoverable from multiple sclerosis disease.
17. Display label
This procedure displays the current value of each patch and agent during running a simulation.
Figure 35: model view during simulation
Parameters
Init-Treg-n
Treg-life
Treg-repro
Treg-radius
Init-Teff-n
Teff-life
Teff-repro
Init-virus-n
V-energy
v-radius
Mimicry
Myseed
Simulation 1 Simulation 2 Simulation 3 Simulation 4 Simulation 5 Simulation 6
100
60
25%
3
100
60
25%
100
20
3
1.0
100
50
30
12 %
2
100
60
25%
100
20
3
1.0
100
100
60
25%
3
50
30
12%
100
20
3
1.0
100
100
60
25%
3
100
60
25%
100
20
3
1.0
100
50
30
12
2
100
60
25%
100
20
3
1.0
100
100
60
25%
3
100
60
25%
50
10
2
1.0
100
off
off
100
2
1.5
2
off
off
100
2
1.5
2
Show-enegy off
Disable-Treg off
100
Init-mye
Ate-mye
2
1.5
Rec-mye
Mye-rgrow-
2
time
BBB-coun-
down
Cytokine-
energy
Cytokine-n
Hill1
Hill2
Patch-
density
Table 20: values of all variables during simulation
50
25
50
25
1
2
1
3
1
2
1
3
off
off
100
5
1.5
0
50
25
1
2
1
3
off
off
100
5
1.5
0
50
25
1
2
1
3
off
off
100
2
1.5
2
50
12
1
2
1
1
50
25
1
2
1
3
First simulation:
In this simulation the T-regulatory and T-effectors production and the initial birth ratio are same. In this
case brain, axonal damage and brain recovery time are almost same. Sometimes T-effectors strongly
attack brain axonal, at this time brains immune system gives quick response to damage and regulatory
cells control T-effectors progression.
2nd simulation:
In this simulation, T-effectors cell production. Reproduction and initial number rate are strong. Less T-
regulatory strength shows weak immune system. In which recovery cells and immune cells are weak or
they are in very small amount. In this case, T-effectors strongly attack brain axonal area and cause severe
brain damage. Most of the time this damage is unrecoverable and growing with time.
3rd simulation:
This simulation is adjusted according to the strong immune system. In which T-regulatory cells are more
active than T-effector cells ant they give a quick response to external and internal disease attack. In this
simulation, T-regulatory sells swiftly controls T-effectors activity without any significant brain damage.
4th simulation (slow recovery time)
This simulation shows very slow healing power in patients. This simulation is adjusted as T-regulatory
and T-effector cells production and reproduction ratio are same. When an effector attack on axonal then
this damage goes to severe damage because of very slow recovery time. This type of brain is an easy
target for disease. The brain damage rate increase, alternatively recovery process leads to slow.
5th simulation (weak immune system)
This simulation is adjusted as when an active T-effector attack on axonal it would be a severe damage and
the production and reproduction of regulatory cells is half of the effector cells. This simulation shows a
rapid damage rate and rapid decline in brain recovery time. At the end, most of the brain damage and
became unrecoverable. This sign is a high severity of the disease.
6th simulation ()
This simulation is adjusted as low virus attack on the immune system and the immune system have low
cytokine energy. Cytokines support the immune system by decreasing activity of opposing cells. When T-
effector cells attack axonal then T-effector cells give quick a response to effectors by controlling their
activities. But alternatively, their produced cytokines have low energy. This case leads to increasing
damage rate and decreasing recovery time.
6. Conclusion:
In this study, we developed a simulation model for MS disease by using AOSE methodologies. The
novelty of our work is proposing an idea that, develop biological models with the help of AOSE
methodologies, that provide developer support in the analysis, design, and implementation phase. After
development, we have evaluated AOSE methodologies utilizing a framework that examines the various
facets of a methodology. The proposed framework consists of four phases. The analysis of methodologies
shows that although methodologies are mature, that can be used to develop biological or any other
complex system, however, there are still open issues because methodologies do not provide a solid rule
for agent identification from the requirement document and do not guide the transformation of roles to
agents. In general, some software engineering issues such as quality assurance, cost estimation, and
project management guidelines are not supported by any of the methodology. Besides all these
limitations, AOSE methodologies have shown the potential for the development of the MS disease model.
Moreover, this framework and comparison results can be utilized for selecting a methodology for
developing an agent-based application.
Reference
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
S. Cai, W. Chen, D. Liu, M. Tang, and X. Chen, "Complex network analysis of brain functional
connectivity under a multi-step cognitive task," Physica A, vol. 466, pp. 663–671, 2017.
B. U. Forstmann and E. J. Wagenmakers, "An introduction to model-based cognitive
neuroscience," An Introd. to Model. Cogn. Neurosci., no. August, pp. 1–354, 2015.
R. Cunha, D. Adamatti, and C. Billa, "Agent Oriented Software Engineering: A Comparative
Study between methodologies that support the Development of Multi-Agent Systems," Proc. 7th
Int. Conf. Manag. Comput. Collect. Intell. Digit. Ecosyst., pp. 48–52, 2015.
C. Lucena and I. Nunes, "Contributions to the emergence and consolidation of Agent-oriented
Software Engineering," J. Syst. Softw., vol. 86, no. 4, pp. 890–904, 2013.
V. Kannan, N. A. Kiani, F. Piehl, and J. Tegner, "A minimal unified model of disease trajectories
captures hallmarks of multiple sclerosis," Math. Biosci., vol. 289, pp. 1504–1514, 2017.
J. Komatsu, K. Sakai, M. Nakada, K. Iwasa, and M. Yamada, "A long spinal cord lesion in a
patient with biopsy proven multiple sclerosis," Neuropathology, vol. 32, no. 3, p. 338, 2012.
S. Afraei et al., "Therapeutic effects of D-aspartate in a mouse model of multiple sclerosis," J.
Food Drug Anal., vol. 25, no. 3, pp. 699–708, 2017.
T. Zrzavy, S. Hametner, I. Wimmer, O. Butovsky, H. L. Weiner, and H. Lassmann, "Loss of
'homeostatic' microglia and patterns of their activation in active multiple sclerosis," Brain, vol.
140, no. 7, pp. 1900–1913, 2017.
D. J. Arpin, E. Heinrichs-Graham, J. E. Gehringer, R. Zabad, T. W. Wilson, and M. J. Kurz,
"Altered sensorimotor cortical oscillations in individuals with multiple sclerosis suggests a faulty
internal model," Hum. Brain Mapp., vol. 38, no. 8, pp. 4009–4018, 2017.
[10] F. Halabchi, Z. Alizadeh, M. A. Sahraian, and M. Abolhasani, "Exercise prescription for patients
with multiple sclerosis; potential benefits and practical recommendations," BMC Neurol., vol. 17,
pp. 1–11, 2017.
[11] P. Sowa, G. O. Nygaard, A. Bjørnerud, E. G. Celius, H. F. Harbo, and M. K. Beyer, "Magnetic
resonance imaging perfusion is associated with disease severity and activity in multiple sclerosis,"
Neuroradiology, vol. 59, no. 7, pp. 655–664, 2017.
[12] M. Pennisi, G. Russo, S. Motta, and F. Pappalardo, "Agent based modeling of the effects of
potential treatments over the blood-brain barrier in multiple sclerosis," J. Immunol. Methods, vol.
427, pp. 6–12, 2015.
[13] Y. Mansury and T. S. Deisboeck, "The impact of '" search precision "' in an agent-based tumor
model," J. Theoratical bBology, vol. 224, pp. 325–337, 2003.
[14] C. A. Athale and T. S. D. Ã, "The effects of EGF-receptor density on multiscale tumor growth
patterns," J. Theoratical Biol., vol. 238, pp. 771–779, 2006.
[15] Y. Mansury and T. S. Deisboeck, "Simulating the time series of a selected gene expression profile
in an agent-based tumor model," Phys. D, vol. 196, pp. 193–204, 2004.
[16] Y. Mansury and T. S. Deisboeck, "Simulating ' structure – function ' patterns of malignant brain
tumors," Physica A, vol. 331, pp. 219–232, 2004.
[17] L. Zhang, C. A. Athale, and T. S. Deisboeck, "Development of a three-dimensional multiscale
agent-based tumor model : Simulating gene-protein interaction profiles , cell phenotypes and
multicellular patterns in brain cancer," J. Theoratical Biol., vol. 244, pp. 96–107, 2007.
[18] L. Zhang, L. L. Chen, and T. S. Deisboeck, "Multi-scale, multi-resolution brain cancer modeling,"
Math. Comput. Simul., vol. 79, pp. 2021–2035, 2009.
[19] L. Germond, M. Dojat, C. Taylor, and C. Garbay, "A cooperative framework for segmentation of
MRI brain scans," Artif. Intell. Med., vol. 20, no. 1, pp. 77–93, 2000.
[20] N. Richard, M. Dojat, and C. Garbay, "Automated segmentation of human brain MR images using
a multi-agent approach," Artif. Intell. Med., vol. 30, pp. 153–175, 2004.
[21] F. G. Vital-lopez, A. Armaou, and M. Hutnik, "Modeling the Effect of Chemotaxis on
Glioblastoma Tumor Progression," Am. Inst. Chem. Eng., vol. 57, no. 3, 2011.
[22] R. Soc, "Introduction . Modelling natural action selection," R. Soc., no. April, pp. 1521–1529,
2007.
[23] L. C. Signaling, M. Bias, and E. Rate, "Simulating Brain Tumor Heterogeneity with a Multiscale
Agent-Based Model :Linking molecular signatures, phenotypes and expansion rate," Math.
Comput. Model., vol. 49, no. 1–2, pp. 307–319, 2009.
[24] H. González-Vélez et al., "HealthAgents: Distributed multi-agent brain tumor diagnosis and
prognosis," Appl. Intell., vol. 30, no. 3, pp. 191–202, 2009.
[25] L. Zhang, Z. Wang, J. A. Sagotsky, and T. S. Deisboeck, "Multiscale agent-based cancer
modeling," J. Math. Biol., vol. 58, no. 4–5, pp. 545–559, 2009.
[26] L. Zhang et al., "Developing a multiscale , multi-resolution agent- based brain tumor model by
graphics processing units," Theor. Biol. Med. Model., vol. 8, no. 1, p. 46, 2011.
[27] R. Haroun, F. Boumghar, S. Hassas, L. Hamami, and C. B. University-lyon, "A Massive Multi- A
gent System for Brain MRI Segmentation," Massively Multi-Agent Syst., vol. 3446, pp. 174–186,
2005.
[28] V. Koutkias and M. C. Jaulent, "A Multiagent System for Integrated Detection of
Pharmacovigilance Signals," J. Med. Syst., vol. 40, no. 2, p. 37, 2016.
[29] L. Leistritz, K. Schiecke, L. Astolfi, and H. Witte, "Time-Variant Modelling of Brain Processes,"
Proc. IEEE, vol. 104, no. 2, pp. 262–281, 2016.
[30] H. Barrah, "MAS based on a Fast and Robust FCM Algorithm for MR Brain Image
Segmentation," vol. 7, no. 7, pp. 191–196, 2016.
[31] M. Pennisi, A.-M. Rajput, L. Toldo, and F. Pappalardo, "Agent based modeling of Treg-Teff cross
regulation in relapsing-remitting multiple sclerosis," BMC Bioinformatics, vol. 14, no. Suppl 16, p.
S9, 2013.
[32] M. Wooldridge, N. R. Jennings, and D. Kinny, "The Gaia Methodology for AgentOriented
Analysis and Design. Autonomous Agents and Multi-Agent Systems 3(3): 285–312," pp. 285–
312, 2000.
[33] Y. C. H. H. C. Jia, "A COMPARISON OF THREE AGENT-ORIENTED SOFTWARE
DEVELOPMENT METHODOLOGIES : MASE , GAIA , AND TROPOS Yubo Jia , Chengwei
Huang , Hao Cai School of Information and Electronics , Zhejiang Sci-Tech University ,
Hangzhou Zhejiang 310018 , China , software_wei@1," pp. 106–109, 2009.
[34] F. Zambonelli, N. R. Jennings, and M. Wooldridge, "Developing Multiagent Systems: The Gaia
Methodology," ACM Trans. Softw. Eng. Methodol., vol. 12, no. 3, pp. 317–370, 2003.
[35] D. C. Silva, R. A. M. Braga, L. P. Reis, and E. Oliveira, "Designing a meta-model for a generic
robotic agent system using Gaia methodology," Inf. Sci. (Ny)., vol. 195, pp. 190–210, 2012.
[36] D. C. Silva, R. A. M. Braga, L. P. Reis, and E. Oliveira, "A generic model for a robotic agent
system using GAIA methodology: Two distinct implementations," 2010 IEEE Conf. Robot.
Autom. Mechatronics, RAM 2010, pp. 280–285, 2010.
[37] L. S. Passos, R. J. F. Rossetti, and J. Gabriel, "An Agent Methodology for Processes, the
Environment, and Services," Adv. Artif. Transp. Syst. Simul., pp. 37–53, 2014.
[38] A. Castro and E. Oliveira, "The rationale behind the development of an airline operations control
centre using Gaia-based methodology," Int. J. Agent-Oriented Softw. Eng., vol. 2, no. 3, p. 350,
2008.
J. Vilela, R. Ramos, and J. Castro, "Mastem: A mathematics tutoring multi-agent system," CEUR
Workshop Proc., vol. 1005, 2013.
J. Castro, M. Kolp, and J. Mylopoulos, "Towards requirements-driven information systems
engineering: The Tropos project," Inf. Syst., vol. 27, no. 6, pp. 365–389, 2002.
[39]
[40]
[41] P. Giorgini, J. Mylopoulos, and R. Sebastiani, "Goal-oriented requirements analysis and reasoning
in the Tropos methodology," Eng. Appl. Artif. Intell., vol. 18, no. 2, pp. 159–171, 2005.
[42] E. S. K. Yu, "Towards modelling and reasoning support for early-phase\nrequirements
engineering," Proc. ISRE '97 3rd IEEE Int. Symp. Requir. Eng., pp. 226–235, 1997.
[43] P. Bresciani and A. Perini, "Tropos : An Agent-Oriented Software Development Methodology,"
pp. 203–236, 2004.
[44] S. A. DeLoach, L. Padgham, A. Perini, A. Susi, and J. Thangarajah, "Using three AOSE toolkits to
develop a sample design," Int. J. Agent-Oriented Softw. Eng., vol. 3, no. 4, pp. 416–476, 2009.
[45] H. K. Dam and M. Winikoff, "Towards a next-generation AOSE methodology," Sci. Comput.
Program., vol. 78, no. 6, pp. 684–694, 2013.
[46] D. Bertolini, A. Perini, A. Susi, and H. Mouratidis, "The Tropos visual modeling language. A
MOF 1.4 compliant meta-model.," Second Agentlink Iii Tech. Forum, no. January, 2005.
[47] S. A. Deloach, "Multiagent Systems Engineering: A Methodology and Language for Designing
Agent Systems," Aois'99, 1999.
[48] S. A. Deloach, E. T. Matson, and Y. Li, "EXPLOITING AGENT ORIENTED SOFTWARE
ENGINEERING IN COOPERATIVE ROBOTICS SEARCH AND RESCUE," Syst. Eng., vol. 17,
no. 5, pp. 1–19, 2003.
[49] S. DeLoach, "Analysis and Design using MaSE and agentTool," Proc. 12 Midwest Artif. Intell.
Cogn. Sci. Conf. (MAICS 2001), Miami Univ. Oxford, Ohio, no. Maics, pp. 1–7, 2001.
[50] DELOACH et al., "MULTIAGENT SYSTEM ENGINEERING," Int. J. Softw. Eng. Knowl. Eng.,
vol. 11, no. 3, pp. 231–258, 2001.
[51] A. K. Pandey and R. Pandey, "Role of Multi Agent System Methodology in," 2015 2nd Int. Conf.
Comput. Sustain. Glob. Dev., pp. 438–441, 2015.
[52] T. Sanislav, D. Cǎpǎţnǎ, A. Stoian, and I. Stoian, "Multiagent approach for cancer automated
registration," Control Eng. Appl. Informatics, vol. 12, no. 3, pp. 41–46, 2010.
[53] C. Lin, K. M. Kavi, F. T. Sheldon, K. M. Daley, and R. K. Abercrombie, "A Methodology to
Evaluate Agent Oriented Software Engineering Techniques University of North Texas," pp. 1–10,
2007.
[54] T. Abdelaziz, M. Elammari, and R. Unland, "A framework for the evaluation of agent-oriented
methodologies," Innov. 4th Int. Conf. Innov. Inf. Technol. IIT, pp. 491–495, 2008.
[55] G. Grau, C. Cares, and X. Franch, "A Comparative Analysis of i*Agent-Oriented Modelling
Techniques," 18th Int. Conf. Softw. Eng. Knowl. Eng., pp. 657–663, 2006.
[56] E. Yu and L. M. Cysneiros, "Agent-Oriented Methodologies - Towards a Challenge Exemplar,"
pp. 1–13, 2002.
[57] L. Cernuzzi, G. Rossi, and L. Plata, "On the evaluation of agent oriented modeling methods,"
Proc. Agent Oriented Methodol. Work., pp. 21–30, 2002.
[58] M. Cossentino, N. Gaud, V. Hilaire, S. Galland, and A. Koukam, "ASPECS: An agent-oriented
software process for engineering complex systems," Auton. Agent. Multi. Agent. Syst., vol. 20, no.
2, pp. 260–304, 2010.
[59] E. Garcia, S. Miles, M. Luck, and A. Giret, "Evaluating how agent methodologies support the
specification of the normative environment through the development process," Auton. Agent.
Multi. Agent. Syst., vol. 29, no. 6, pp. 1041–1060, 2015.
[60] F. Bergenti, Federico and Gleizes, Marie-Pierre and Zambonelli, METHODOLOGIES AND
SOFTWARE ENGINEERING FOR AGENT SYSTEMS. 2009.
[61] M. A. Niazi, "Towards A Novel Unified Framework for Developing Formal, Network and
Validated Agent-Based Simulation Models of Complex Adaptive Systems," 2017.
[62] M. A. Niazi, "Emergence of a snake-like structure in mobile distributed agents: An exploratory
agent-based modeling approach," Sci. World J., vol. 2014, 2014.
|
1008.4115 | 1 | 1008 | 2010-08-24T19:24:24 | Noise in Naming Games, partial synchronization and community detection in social networks | [
"cs.MA",
"cs.SI"
] | The Naming Games (NG) are agent-based models for agreement dynamics, peer pressure and herding in social networks, and protocol selection in autonomous ad-hoc sensor networks. By introducing a small noise term to the NG, the resulting Markov Chain model called Noisy Naming Games (NNG) are ergodic, in which all partial consensus states are recurrent. By using Gibbs-Markov equivalence we show how to get the NNG's stationary distribution in terms of the local specification of a related Markov Random Field (MRF). By ordering the partially-synchronized states according to their Gibbs energy, taken here to be a good measure of social tension, this method offers an enhanced method for community-detection in social interaction data. We show how the lowest Gibbs energy multi-name states separate and display the hidden community structures within a social network. | cs.MA | cs |
Noise in Naming Games, partial synchronization and community detection in social
networks
Weituo Zhang1 and Chjan C. Lim1
1Mathematical Sciences, RPI, 110 8th St., Troy, NY, 12180
The Naming Games (NG) are agent-based models for agreement dynamics, peer pressure and
herding in social networks, and protocol selection in autonomous ad-hoc sensor networks. By intro-
ducing a small noise term to the NG, the resulting Markov Chain model called Noisy Naming Games
(NNG) are ergodic, in which all partial consensus states are recurrent. By using Gibbs-Markov equiv-
alence we show how to get the NNG's stationary distribution in terms of the local specification of
a related Markov Random Field (MRF). By ordering the partially-synchronized states according to
their Gibbs energy, taken here to be a good measure of social tension, this method offers an enhanced
method for community-detection in social interaction data. We show how the lowest Gibbs energy
multi-name states separate and display the hidden community structures within a social network.
PACS numbers: 89.65.Ef, 05.65.+b, 02.50.Ga
Keywords: Naming Game , Social network , Noise in Markov Chain
I.
INTRODUCTION
The possibility of using the Naming Game (NG)[5, 11],
a family of agent-based models, for modeling agreement
dynamics, leader election and the effects of peer pres-
sure in social networks [4, 6, 9], protocol and security
key selection in encrypted communication in autonomous
sensor networks[10], vocabulary selection in linguistics
[19, 22] and the role of herding in market crashes in finan-
cial networks have recently attracted great interest in the
scientific community. Formation of opinions and agree-
ment dynamics in complex financial and social networks
are relevant examples of collective / cooperative behav-
iors that are driven by the individual's innate propensity
to imitate and hence herd in the absence of reliable infor-
mation or when there is information asymmetry. Finite-
size effects and the underlying network's topology such
as scale-free [2] and small world [3], have significant im-
pact on the expected time to consensus (single-name or
total-synchrony state) and other related critical expo-
nents.
Indeed simulations showed departure from the
values obtained by classical scaling arguments based on
the emergence of the giant cluster in percolation theory
[1, 20, 21] and on the dynamical theory of coarsening in
high-q Pott's models [13, 14, 17, 23, 24], highlighting the
need for more simulation studies and also Markov Chain
analysis [7].
A recent application of the Naming Games is to find
hidden communities in social networks [16, 17], [31]. The
main objectives of this paper are to sharpen the commu-
nity detection capability of the original Naming Games
by the explicit and purposeful addition of rare noise
events, and to analyze rigorously the consequences of
adding this noise. Community structure is extremely
difficult to define and uncover and remains one of the
outstanding open problems in network science. Amongst
existing methods, the most powerful algorithms are based
on the notion of modularity [32] which provides an order-
ing of the closeness between the given network's commu-
nity structure and a large family of motifs or partitions.
The optimization problem involved in finding the motif
with the largest modularity grows rapidly with the size
of the network. While several working heuristic methods
have been proposed, they are not better than a spectral
method based on the eigenvectors of a Laplacian-like ma-
trix [32]. This procedure of splitting the network into two
parts is repeated in a binary tree way until the subgraphs
are indivisible.
In particular, direct simulated-annealing of the modu-
larity optimization problem will be computationally ex-
pensive, just as other Monte-Carlo methods applied to
the original Naming Games have yielded good results
only after costly long simulations [18].
In this article
we indicate how Monte-Carlo methods and other impor-
tance sampling algorithms can be numerically efficient on
the Noisy Naming Games but not on the original Naming
Games. Comparing the Noisy Naming Game method to
the modularity optimization procedure, it is worthwhile
noting that these two methods start from opposite ends
of the community structure: the modularity - binary tree
method finds first, the optimal division of the whole net-
work into two parts but the Noisy Naming Games, begin-
ning from random initial sublists of the allowed words,
first finds the (last) optimal division into small communi-
ties. The modularity increases as the binary tree method
proceeds towards the indivisible units but the Gibbs po-
tential decreases as the Noisy Naming Game proceeds
towards the low-lying states near to its ground state of
total consensus. In this way, the Noisy Naming Games
are complementary to the modularity based spectral bi-
nary tree method. If one is looking for the finer divisions
of the network into small communities then the Noisy
Naming Games will find them provably faster than other
methods including the modularity-based methods.
But before the Naming Games can provide an applica-
ble procedure for community detection, three problems
need to be addressed. The first is the transient nature of
its multi-name partial consensus states that correspond
to possible community structures. The second concerns
the existence and computability of a stationary measure
that can be used -- like the notion of modularity -- to rank
the community test-patterns in the form of multi-name
partial consensus states. The third issue concerns the
computational efficiency of the Naming Games viewed
herein as an inverse method for finding communities in
networks. It turns out that these problems can be over-
come by adding rare noise events to the standard stochas-
tic framework provided by the Markov chain model of
the original Naming games. Unlike the original Naming
Games which have been compared to q-Potts models [24],
the Noisy version introduced here is more closely related
to spin-glass models with an additional layer of random-
ness [35]. As is well-known certain aspects of informa-
tion processing and combinatorial optimization (such as
the graph-partitioning problem that is closely related to
community detection) can be addressed from the vantage
point of the spin-glass framework.
In many instances,
the ground-state of the spin-glass model provides infor-
mation for the optimization problem. But computing or
finding the ground-state is often an NP-hard problem,
of an equivalent degree of difficulty as the optimization
problem itself. The spin-glass framework offers the alter-
native path -- often seen to work well in explicit posi-
tive usage of noise to solve hard deterministic problems
-- of exploiting not only the inherent spin entropy of the
model (when the interactions are quenched) but also the
additional randomness and corresponding entropy arising
from the ensemble of quenched interactions for which the
replica method [35] is designed to average over. We ex-
ploit these relationships in discussing each of these three
issues in formulating an efficient procedure based on the
Noisy Naming Games for finding community structures.
First, because the total consensus states of the orig-
inal Naming Games are the only absorbing states, the
detection of communities or partial consensus states is
complicated by their transient nature. Meta-stable states
representing multi-names or partial consensus configura-
tions arise in the original Naming Games on networks
with strong community-structure but have been shown
numerically to evolve into other partial-synchrony states
after an initial phase [16, 17], according to a time hierar-
chy in general. Complete agreement on leader selection
or total consensus occurs on time-scales that are rela-
tively short after a long meta-stable stage, with negative
consequences on the design of algorithms for detection
of hidden communities, for example, in social networks
with Small World topologies [29], [31]. Related results
on the synchronization of dynamical networks and dis-
crete time maps on networks that highlight the roles of
the network topology and the coupling weights have been
discovered [26], [30]. In other significant contexts such as
the modeling of brain functions by coupled oscillators on
a network, the transient nature of partially synchronous
states play significant positive roles.
The first issue that has to be solved is to find a version
of the Naming Games where all partial consensus states
are recurrent. This can be done if we come up with a ver-
2
sion of the Naming Games which is an ergodic Markov
chain. A nice property of ergodic Markov chains is the
existence and uniqueness of an invariant measure with
positive weights on all the recurrent states. Then the
next issue in designing an efficient procedure for commu-
nity detection starting from the original Naming Games
is finding a closed-form expression for this invariant mea-
sure and proving that given the network topology (such
as neighborhood structure) and a reasonably small set of
allowed words in the original game, this invariant mea-
sure ranks the partial consensus states in an averaged
sense (macrostates) incorporating all entropic contribu-
tions in the Noisy Naming Games. One key property of
this invariant measure is the free energy gaps between
the low-lying states near the ground-state -- larger gaps
lead to better resolution of the community structure. We
show by exact calculations on small cliques that the in-
variant measure of the Noisy Naming Game inherits this
property from its closed-form Gibbs potential. Although
the large-scale implementation of the Monte-Carlo sim-
ulation of the Noisy Naming Game on a real world so-
cial network remains to be done, our small-scale testing
of this method on computer-generated randomized net-
works of about 60 nodes allow us to infer that the method
is computationally efficient. In the context herein, rig-
orous results stating new conditions for the existence
and rank-ordering of recurrent multi-names or partial-
consensus states in the Naming Games are useful in view
of potential applications to detecting hidden communi-
ties in interaction data such as Twitter and Facebook.
In this article, we show that a class of arbitrarily small
noisy perturbations of the Naming Games, called Noisy
Naming Games (NNG), satisfies these conditions. The
main results here are analytical ones: First, the Noisy
Naming Games differing from the NG by arbitrarily small
stochastic perturbation, has a unique invariant distribu-
tion with positive weights for partial consensus states.
Second, we construct the NNG's invariant distribution as
the Gibbs potential of a related Markov Random Field
(NNGGS) called the Noisy Naming Game with saturated
training from its local specification. Third, we show the
Gibbs energy (which is taken to be a reasonably efficient
measure of social tension in this article) provides an or-
dering of the recurrent multi-name or partial-consensus
states of the NNG whereby the lowest energy and thus
most probable ones, also have single-name or single-color
cliques. This last result has significant impact on the de-
tection of (hidden) community structure in a social net-
work by Monte-Carlo simulations. After the initial equili-
bration process, the fraction of time spent by the Monte-
Carlo simulator in any given partial-consensus state is
proportional to the probability it is assigned by the Gibbs
invariant measure. Thus, besides identifying the low-
lying Gibbs energy states which are the most indicative of
underlying community structure - near-cliques or tightly
connected clusters have a common single-name or color --
the simulation when run long enough will also provide an
ordering of these states. We will show with explicit calcu-
lations on simple examples, that the exact Gibbs energy
of the multi-name states not only orders them but have
the essential property of significant energy gaps between
the low-lying states.
In brief, the introduction of additional randomness
in the form of rare noisy events to the original Nam-
ing Games provides the key to successfully enhancing
community-detection through the existence and rank-
ordering of a large family of stable partial-consensus
states where tightly-connected clusters in the network
show up as single-named or single-colored subgraphs
in the social network. Moreover, these partial consen-
sus states can be found and ranked by simulating the
Noisy Naming Games using an equivalent Gibbs sampler
method which was used to produce the figures in the
last section. In contrast to the generic need to avoid or
prevent noise in most technological systems, the precise
and explicit use of noise to achieve a positive aim in this
project has few precedents. Significant examples are the
classical Parrondo games and the Brownian Ratchet [25].
II. MARKOV CHAIN MODEL
We construct a Markov chain model for the NG [5]
where a transition consists of the change of local state
at only one site in the network and the set Name of all
allowed words is fixed at the outset. It is easier to an-
alyze by the Gibbs Sampler method than the usual NG
to which it is equivalent. Consider a network based on a
connected graph containing N sites S = {si}. The neigh-
borhood structure of the network is given by {Nsi}si∈S.
Each site si will have a word list chosen from the finite
set N ame = {A, B, C...}. Set Γ = {γk}, which consists
of all non-empty subsets γk of N ame, represents all pos-
sible "word lists" of a site. The configuration function
X(si) : S → Γ mapping each site to its word list gives
the local state of si. A configuration restricted to a subset
of sites Σ ⊂ S is denoted by X(Σ). Therefore the net-
work state G- a word list assigned to each site- is given
by the configuration G = {X(si)}si∈S = X(S). In the
Naming Game (NG), we change the state of the network
G as follows:
1. In this step we randomly choose a site si ∈ S as a
"listener", with probability q(si).
2. Next, choose a site sj ∈ Nsi with equal probabil-
ity as a "speaker". The "speaker" will randomly
choose a word W from its word list with equal prob-
ability and send it to the "listener". The latter will
change its state X(si) into X ′(si) by the following
way:
X ′(si) =(X(si) ∪ {W } W /∈ X(si)
W ∈ X(si)
{W }
We call step (2) the local transition step and the pro-
cess (1),(2) together comprises a transition between G =
3
{X(si), X(sj)}sj ∈S\{si} and its neighboring state G′ =
{X ′(si), X(sj)}sj ∈S\{si}. The global transition proba-
bility P (G, G′) of the NG depends on the probability
for choosing si and the local transition probability from
state X0(si) to X ′(si) under neighborhood configuration
X0(Nsi ) is P (G, G′) = q(si)P (X0(si), X ′(si)X0(Nsi )).
By randomly choosing an initial state G0 ∈ ΓS and
applying steps (1) and (2) in each time period, we ob-
tain a homogeneous Markov chain of the Naming Game
{G0, G1...Gn...}, where the transition matrix from Gn to
Gn+1 is given by the formula above.
Next, we show that any invariant measure of this
Markov chain is a linear combination of 1GW where 1
is the indicator function and GW are the "single-name
states" in which every site has only W in its word list.
Let Gγk for γk ∈ Γ be the set of network states that sat-
isfies the conditionSsi∈S X(si) = γk. It is clear from the
NG Markov chain above that single-name states are the
only absorbing ones.
To proceed, every network state G ∈ Gγk with γk ≥ 2 is
accessible to at least one absorbing state GW , i.e. it will
have a path {G, G1, G2...Gn, GW } with nonzero probabil-
ity p = (G, G1)P (G1, G2)...P (Gn, GW ). Since this prob-
ability p should be counted in the probability of leaving
the state G and never returning to it, any state G which
is not absorbing is transient.
In an invariant measure
of a Markov chain, the weight of any transient state is
zero; thus, the only states with positive weights in an in-
variant measure of the NG Markov chain are the single-
name states GW (W ∈ N ame). Since 1GW itself is al-
ways an invariant measure, any invariant measure of the
Naming Game is a linear combination of 1GW over all
W ∈ N ame. In other words, the Naming Game Markov
Chain, starting from any initial state will eventually go
to a single-name state.
III. NOISY NAMING GAMES (NNG)
In view of the above Markov chain analysis, we need
to perturb the original Naming Game to enhance its
community-detection capabilities. It turns out that the
class Noisy Naming Game (NNG) of arbitrarily small
random perturbations of NG, given next, has the required
key property of persistent multi-name steady states. In
each step in the NNG Markov chain, the listener has a
very small probability ǫ to receive a noise word rather
than the speaker's message where the noise word is a
random word chosen uniformly from N ame, and a prob-
ability of 1 − ǫ of proceeding as in the original NG. Then
given two arbitrary network states, one can construct a
path from one to the other with nonzero probability by
forcing the network to receive a sequence of noise words
as required. In this way, the NNG Markov chain is forced
to be communicative and therefore ergodic, by arbitrarily
small ǫ. On the other hand, since ǫ is arbitrarily small,
the event of receiving a noise word rarely happens. So the
noise will essentially change the long term behavior but
can be ignored in finite time. From the NNG Markov
chain's ergodicity, its invariant measure is positive and
unique up to a scalar multiple; hence the first main re-
sult:
Result : The NNG has a unique invariant probabil-
ity distribution π(G) > 0 directly related to its global
transition probability P (G, G′). Moreover, each multi-
name state Gm is recurrent and has positive weight, i.e.,
π(Gm) > 0.
transition step (2)
In both the original NG and the NNG, the lo-
is given by X0(si) →
cal
X ′(si) with corresponding local transition probability
P (X0(si), X ′(si)X0(Nsi ))1X(Nsi )=X0(Nsi ). By repeat-
ing the NNG's local transition step (2) many times
at fixed site si ∈ S with the same neighborhood
state X0(Nsi ), we generate a sequence of local states
{X0(si), X ′(si), ..., X (n)(si)},
such that the marginal
probability distribution of X (n)(si) converges to a l ocal
conditional probability distribution fi(XX0(N (si)) as n
goes to infinity. Thus, fi(XX0(N (si)) over all si ∈ S
is a well-defined limiting distribution of a local Markov
chain with one site and fixed neighborhood state.
To calculate the NNG's invariant distribution, we first
construct a related Markov chain (NNGGS) verified to be
a Gibbs Sampler for a Markov Random Field [7]. Keep-
ing step (1) the same, the NNGGS replaces the local
transition step (2) in the NNG by a "training step":
(2') X0(si) → X ∗(si) which is the value of a random
variable distributed according to the local conditional dis-
tribution fi(X ∗X0(N (si)).
The local training step (2') is thus equivalent to repeat-
ing the NGG's local transition step (2) infinite number
of times at fixed site with the same neighborhood state.
Its global transition probability is given by
′
P
(G, G′) = q(si)fi(X
′
(si)X0(N (si))1X(Nsi )=X0(Nsi )
Since only the local state at a single random site is
changed according to a local specification, the NNGGS
is indeed a Gibbs Sampler on a Markov Random Field
(MRF) with local specification given by the family
fi(XX(N (si)) for si ∈ S [7]. Here, the NNGGS is more
an indirect analytical method to derive the stationary
distribution of the NNG in closed form than a method
for numerical simulation of the NNG, for which there are
better Markov chain Monte-Carlo methods [15].
IV.
STATIONARY DISTRIBUTION OF THE
GIBBS SAMPLER
Using the Gibbs-Markov equivalence [7], we construct
in terms of the NNGGS's local specification, a Gibbs po-
tential which is the stationary distribution π
(G) of the
(G, G′), the NNGGS
NNGGS. In other words, under P
generates a Markov chain of realized states and a distri-
bution, that converge to π
(G). First, we need to assign a
0 local state for each site. Then 0(E) for E ⊂ S denotes
′
′
′
4
the 0 configuration state where all sites in E are at the 0
local state. The choice of the 0 local state is not unique --
for convenience we choose the whole word-list γ = N ame
as the 0 local state. For a clique L ⊂ S of the network
graph, let x(L) be a certain configuration on L. Then
the clique potential of this configuration is[12]:
V (x(L)) = − XE⊂L
(−1)L−E ln[F (x(E)0(N (E)))].
Here the summation is over al, subsets of the clique L, the
neighborhood of a subset E = {sii = 1, ..., M } is defined
given by:
by N (E) = Ssi⊂E N (si) \E and F (x(E)0(N (E))) is
Yi=1
fi(x(si)X(s1, ..., si−1), 0(si+1, ..., sM ), 0(N (E)))
fi(0(si)X(s1, ..., si−1), 0(si+1, ..., sM ), 0(N (E)))
M
Then the Gibbs potential is given by the sum (over all
cliques in the network graph) of clique potentials:
H(x) = XL⊂S
V (x(L)).
The stationary distribution of the NNGGS is thus:
′
π
(x) = Z −1 exp(−H(x)).
where Z = PG exp(−H(G))) is the Gibbs partition
function[7]. It remains to show that π
the NNG, that is, π
is invariant under
= π.
′
′
V.
INVARIANT DISTRIBUTION OF THE NNG
′
By the existence and uniqueness of the invariant dis-
tribution of the NNG, any distribution that is invariant
under the NNG must be equal to π(G). Here, we ver-
ify that the Gibbs distribution π
(G) of the NNGGS
is invariant under the NNG, thence the explicit con-
struction of π(G) as the second main result. We will
do this by showing and using detailed balance sev-
eral times. Consider a chosen network state G0 =
{X0(s1), ..., X0(si), ...X0(sN )} and any neighboring net-
work state Gik = {X(s1), ..., X(si), ...X(sN )} where all
but site si have the same local states as G0 and the dif-
fering local state is the word list γk ∈ Γ. By detailed
balance of the NNGGS:
π′(G0)q(si)fi(γkX0(N (si))) = π′(Gik)q(si)fi(X0(si)X0(N (si))).
In terms of the global transition probability P (G, G′) of
the NNG, the change of measure under a single step of
the NNG ∆π′(G0) is
[P (Gik, G0)π′(Gik) − P (G0, Gik)π′(G0)]
Xi Xk
= Xi
−P (X0(si), γkX0(N (si)))π′(Gik)i
q(si)Xk hP (γk, X0(si)X0(N (si)))π′(G0)
where the internal sum is taken over all word lists γk ∈ Γ.
Each term of the internal sum is zero by the NNG's local
detailed balance at fixed site si because P (·, ·X0(N (si)))
is the local transition probability of the NNG condi-
tioned on fixed neighborhood state X0(N (si)), and the
quotient π′(Gik)/π′(G0) has been shown to equal the
quotient fi(γkX0(N (si)))/fi(X0(si)X0(N (si))) where
fi(·X0(N (si))) are the values of the local specification
at site si of the NNGGS. In other words, the NNGGS'
local specification, by construction (repeating the local
transition step (2) in the NNG at fixed site si, under
the same neighborhood state X0(N (si))), is the station-
ary distribution of an auxillary Markov chain consisting
of varying local states at the single site si and a fixed
boundary state at the sites in N (si). We deduce that
∆π′(G0) = 0, i.e. the Gibbs distribution π
(G) of the
NNGGS is also the invariant distribution of the NNG.
′
VI. CLIQUE POTENTIALS AND
COMMUNITY STRUCTURE - ANALYSIS
The first example concerns a 2-clique as in figure 1.
With the local 0 state chosen to be N ame = {A, B}, and
applying the above expression for F (x(E)0(N (E))) in
terms of the local specification which are in turn calcu-
lated using the stationary distribution of the correspond-
ing local Markov chains (each with one site), we derived,
in the vanishing noise limit, the probabilities tabulated
below. This table gives a 2-clique potential for all pos-
sible configurations with its neighborhood fixed at the
local 0 state, after using the natural symmetry in the
problem. It shows this 2-clique has lowest energy when
the two sites has the same single name.
AB
AB
A
A
FIG. 1: 2-Clique
AB
AB
x(L)
A-AB
A-A
A-B
F (x(L)0(N (L)))
1
2
0.5
V (x(L))
0
-0.6931
0.6931
A more interesting example in the following graph has
two 3-cliques and a 2-clique that bridges them. Using the
above procedure (and labeling the sites starting from the
site on the bridge, e.g. B-A-AB means the site on the
bridge has word list B), we calculate the clique potential
for the 3-clique and tabulate its values below.
5
A
A
B
B
A
B
FIG. 2: Example 2
TABLE I: Clique Potential for 3-Clique
x(L)
A-A-A
AB-A-A
B-A-A
F (x(L)0(N (L)))
15
3
3/5
V (x(L))
-2.7080
-1.0986
-0.5108
In this way, we calculate the Gibbs potential for each
network state and show that multi-name states are or-
dered by their Gibbs energy which we take to be a good
measure of social tension in a particular state. After the
single-name ground state, the state in figure 2 has the sec-
ond lowest energy. Significantly, as the third main result
and primary focus of this letter, the naming or coloring
that reveals the underlying cliques in the graph also has
the least energy amongst all multi-name states for given
community (clique) structure in a network. Moreover, it
is a local minimum, i.e. any one-step change of this con-
figuration will increase its energy: H(AAA − BBB) =
−4.7230 and H(AAA − ABB) = −2.8904. These lowest
energy multi-name states are therefore the most likely
ones to be found by simulated annealing of the NNG.
This is consistent with the results on meta-stable states
in [18], but have the advantage, in the NNG, of being
persistent (recurrent), thence, enhancing its community-
detection capability over the original NG.
VII. STATIONARY PROBABILITY
DISTRIBUTIONS OF THE NNG ON SMALL
RANDOMIZED NETWORKS - SIMULATIONS
We conclude with a brief discussion of the method used
to compute the invariant probability distributions of the
NNG on a family of small randomized networks consist-
ing of 60 nodes - these network graphs are computer gen-
erated. Using a total word list of cardinality two, we show
that the NNG efficiently finds the 40:20 splits into com-
munity structures where two of the communities or sub-
graphs (by design of about 20 nodes each) have one name
/ color and the remaining subgraph of 20 nodes have the
second name / color In addition, it also produces an in-
variant measure which, after discounting highest values
achieved naturally at the total consensus or single-name
states at the 0 and 60-nodes peaks, gives relative weights
on the meta-stable states that provide information on
community structure. For example, we refer to the sym-
metric 20 and 40 - nodes peaks in the following figures.
About 100 million steps of the NNG was used to produce
the invariant measure depicted.
FIG. 3: Blue triangle for A, red circle for B and black square
for AB
x 106
Stationary Measure
T
14
12
10
8
6
4
2
0
given by:
6
pA(si) = ǫ
pB(si) = ǫ
2 + (1 − ǫ) #{sj ∈N (si)X(sj )=A}+ 1
2 + (1 − ǫ) #{sj ∈N (si)X(sj)=B}+ 1
#{sj ∈N (si)}
2 #{sj∈N (si)X(sj )=AB}
2 #{sj∈N (si)X(sj )=AB}
#{sj ∈N (si)}
Given the noise level ǫ is arbitrarily small, we can take
the limit ǫ → 0:
pA(si) = #{sj ∈N (si)X(sj )=A}+ 1
#{sj ∈N (si)}
2 #{sj ∈N (si)X(sj )=AB}
pB(si) = #{sj ∈N (si)X(sj )=B}+ 1
#{sj ∈N (si)}
2 #{sj ∈N (si)X(sj )=AB}
Leaving aside the transitions involving a single-name
state, that is, as long as one only considers the transi-
tion between two multi-name states, taking this limit is
valid. From detailed balance we get
fi(Ax(N (si)))pB(si) = fi(ABx(N (si)))pA(si)
fi(ABx(N (si)))pB(si) = fi(Bx(N (si)))pA(si)
from which we find the local specification:
fi(Ax(N (si))) = p2
A(si)/Zl(si)
fi(Bx(N (si))) = p2
B(si)/Zl(si)
fi(ABx(N (si))) = pA(si)pB(si)/Zl(si)
Zl(si) = p2
A(si) + p2
B(si) + pA(si)pB(si)
10
20
30
40
50
60
Number of nodes in group A (blue triangle)
2. Properties of F (x(E)0(N (E)))
FIG. 4: Invariant measure for NNG
By definition,
F (x(E)0(N (E))) is given by:
fi(x(si)X(s1,...,si−1),0(si+1,...,sM ),0(N (E)))
fi(0(si)X(s1,...,si−1),0(si+1,...,sM ),0(N (E)))
Several facts will help us to simplify this formula.
i=1
QM
[Gibbs potential from local specification]
1. Local specification
In general, it's not easy to calculate the invariant mea-
sure of a local Markov chain, namely, its local specifica-
tion fi(X(si)X(N (si))). However for small cliques and
when there are only a small number of allowed words
such as two in the three local states (A, B, AB) in the
above examples, the transition in the Markov chain is
totally decided by the message that si received from its
neighbors. Let pA(si),pB(si) be the probability for si to
receive A, B respectively. Their values can be calculated
from the neighborhood condition X(N (si)); more pre-
cisely the number of neighbors in state A, B and AB, are
1. In each factor of the above product, the condi-
tionals in the numerator and denominator are
the same. So if x(si) = 0(si), the factor is 1.
Therefore only the sites not in 0 state counts in
the product. A direct consequence of this fact is
F (0(E)0(N (E))) = 1.
2. For any 1-Clique in this example:
F (AB(si)0(N (si))) = 1 according to the last para-
graph, and
F (A(si)0(N (si))) =
fi(A0(N (E)))
fi(AB0(N (E)))
=
pA(si)
pB(si)
= 1.
Similarly F (B(si)0(N (si))) = 1, thus, the func-
tion F on any 1-Clique has 0 value.
From A-B symmetry and graph symmetry, we can get
other function values from these results.
7
3. Consider the clique E = {s1, ..., sM } and its
two distinguished configurations which only dif-
fer at one site sm. By relabeling the sites we
can write the two configurations as X(E) =
{x(s1), ..., x(sm−1), x(sm), 0(sm+1), ..., 0(sM )} and
Y (E) = {x(s1), ..., x(sm−1), y(sm), 0(sm+1), ..., 0(sM )}.
In the ratio F (Y (E)0(N (E)))/F (X(E)0(N (E))),
the factors for 0 states do not appear and the
factors for s1, ..., sm−1 cancel.
So we have the
recursive relationship:
F (Y (E)0(N (E)))
F (X(E)0(N (E))) = fm(y(sm)X(s1,...,sm−1),0(sm+1,...,sM ),0(N (E)))
fm(x(sm)X(s1,...,sm−1),0(sm+1,...,sM ),0(N (E))) .
Using this relationship we calculate the function
value of F recursively.
5. Gibbs Potential
3.
2-Clique
In the example above, the network have one 2-Clique
and two similar 3-Cliques. We need to calculate the func-
tion value of F for certain configurations on the given
Cliques. In the following calculations, pA(si) and pB(si)
are always calculated according to the current neighbor
configuration. For the 2-Clique shown in Figure 1, we
have
The 3-Clique in Figure 2 has three 2-Cliques and three
1-Cliques embedded in it. Adding them, calling the sum
"net clique potential" for the 3-Clique "L", and using the
fact that the terms for 2-Cliques cancel, the result turns
out to be: Vnet(L) = − ln(F (LN (L))).
Finally, the total Gibbs potential includes the clique po-
tential of the 2-Clique and the net clique potential of two
3-Cliques,
F (AB − AB0(N (E)))
= F (0(E)0(N (E))) = 1
F (A − AB0(N (E)))
= f1(A0(s2),0(N (E)))
f1(AB0(s2),0(N (E)))
= pA(s1)
pB (s1) = 1
F (A − A0(N (E)))
= F (A − AB0(N (E))) f2(AA(s1),0(N (E)))
f2(ABA(s1),0(N (E)))
= pA(s2)
pB (s2) = 2
F (A − B0(N (E)))
= F (A − AB0(N (E))) f2(BA(s1),0(N (E)))
f2(ABA(s1),0(N (E)))
= pB (s2)
pA(s2) = 1
2 .
4.
3-Clique
For the 3-Clique on the left in Figure 2,
F (AB − AB − AB0(N (E))) = F (0(E)0(N (E))) = 1
F (AB − AB − A)0(N (E)))
f3(AB0(s1,s2),0(N (E))) = pA(s3)
F (AB − A − A)0(N (E)))
pB (s3) = 1
= f3(A0(s1,s2),0(N (E)))
= F (AB − AB − A0(N (E))) f2(A0(s1),A(s3),0(N (E)))
f2(AB0(s1),A(s3),0(N (E)))
= pA(s2)
pB (s2) = 3
F (A − A − A)0(N (E)))
= F (AB − A − A0(N (E))) f1(AA(s2,s3),0(N (E)))
f1(ABA(s2,s3),0(N (E)))
= 3 pA(s1)
pB (s1) = 15
F (B − A − A)0(N (E)))
= F (AB − A − A0(N (E))) f1(BA(s2,s3),0(N (E)))
f1(ABA(s2,s3),0(N (E)))
H(AAA − BBB)
= V (A − B) + Vnet(A − A − A) + Vnet(B − B − B)
= −{ln(F (A − B0(N (E)))) + 2 ln(F (A − A − A0(N (E)))}
= −4.7230
H(AAA − ABB)
= V (A − A) + Vnet(A − A − A) + Vnet(A − B − B)
= −{ln(F (A − A0(N (E)))) + ln(F (A − A − A0(N (E)))
+ ln(F (A − B − B0(N (E)))}
= −2.8904
Acknowledgments: Research was sponsored by the
Army Research Laboratory and was accomplished under
Cooperative Agreement Number W911NF-09-2-0053 to
the Social Cognitive Networks Acad Res Center at RPI.
= 3 pB (s1)
pA(s1) = 3
5
8
[1] R. Albert and A-L. Barabasi, Statistical Mechanics of
Complex Networks, Rev. Mod. Phys., 74, 47-97, 2002.
[18] Qiming Lu. PhD thesis, R.P.I., NY, 2009.
[19] F.A. Matsen and M.A. Nowak. Proc. Natl. Acad. Sci.
[2] R. Albert H. Jeong and A-L. Barabasi Nature, 401, 130,
U.S.A., 101(18053), 2004.
1999.
[3] D. Watts and S. Strogatz, Nature, 393, 440, 1998.
[4] D. Angluin. In Proceedings of the 12th ACM Symposium
on Theory of Computing, page 82-93, New York, 1980.
ACM.
[20] R. Meester and R. Roy. Continuum Percolation. Cam-
bridge University Press, Cambridge, England, 1996.
[21] M. Penrose, Random Geometric Graphs. Oxford Univer-
sity Press, New York, 2003.
[22] M.A. Nowak and N.L. Komarova. Science, 291(114),
[5] A. Baronchelli, M. Felici, E. Caglioti, V. Loreto, and L.
2001.
Steels J. Stat. Mech.: Theory Exp., (P06014), 2006
[6] E. Ben-Naim, L. Frachebourg, and P. L. Krapivsky. Phys.
[23] C. Roland and M. Grant. Phys. Rev. B, 41(4663), 1990.
[24] H.E. Stanley, Intro to Phase transitions and Critical phe-
Rev. E, 53(3078), 1996.
nomena, Oxford U. Press, 1982.
[7] P. Bremaud. Markov chains: Gibbs fields, Monte Carlo
[25] Y. Li, A. Allison, D. Abbott, H.E. Stanley, PRL, 91(22),
simulation, and queues. Springer, 1999.
220601-1, 2003.
[8] G. Casella and E. I. George. The American Stat., 46(3),
[26] C. Zhou, A.E. Motter and J. Kurths Phys. Rev. Lett., 96,
1992.
034101, 2006. 164101, 2006.
[9] C. Castellano, V. Loreto, A. Barrat, F. Cecconi, and D.
[27] A. Arenas, A. Diaz-Guilera, C.J. Perez-Vicente Phys.
Parisi. Phys. Rev. E, 71(066107), 2005.
Rev. Lett., 96, 114102, 2006.
[10] T. C. Collier and C. E. Taylor. J. Parallel Distrib. Com-
[28] M. Chavez, D.U. Hwang, A. Amann, H.G.E. Hentschel
put., 64(866), 2004.
and S. Boccaletti Phys. Rev. Lett., 94, 218701 ,2005.
[11] L. Dall'Asta, A. Baronchelli, A. Barrat, and V. Loreto.
[29] L. Amaral, A. Scala, M. Barthelemy and H.E. Stanley,
Phys. Rev. E, 74(036105), 2006.
[12] G. R. Grimmett. Bull. London Math. Soc., 5, 1973.
[13] K. K. Kaski, J. Nieminen, and J. D. Gunton. Phys. Rev.
Proc. Natl. Acad. Sci. U.S.A., 97(21), 11149, 2000.
[30] D. J. D. Earn and S. A. Levin, Proc. Natl. Acad. Sci.
U.S.A., 103 , 11, 2006.
B, 31(2998), 1985.
[31] M. Girvan and M. E. J. Newman Proc. Natl. Acad. Sci.
[14] S. Kumar, J. D. Gunton, and K. K. Kaski. Phys. Rev. B,
U.S.A. 99, 12, 2002.
35(8517), 1987.
[32] M. E. J. Newman Proc. Natl. Acad. Sci. U.S.A. 103,
[15] C. Lim and J. Nebus, Vorticity, Stat Mech and Monte-
8577, 2006.
Carlo simulations, Springer-Verlag 2006.
[33] S.Chauhan, M.Girvan and E. Ott, arXiv:0911.2735v1,
[16] Q. Lu, G. Korniss, and B.K. Szymanski, The Naming
Game in social networks: community formation and con-
sensus engineering, J. Economic Interaction and Coordi-
nation, 4, 221-235 (2009).
2009.
[34] M. Thattai and A. Oudenaarden Proc. Natl. Acad. Sci.
U.S.A. 98,15,2001.
[35] H. Nishimori, Stat Phys of Spin-Glass and Info processing
[17] Q. Lu, G. Korniss and, and B. K. Szymanski. Phys. Rev.
: an Intro., Oxford U. Press.
E, 77(016111), 2008.
|
1009.2997 | 1 | 1009 | 2010-09-15T18:55:48 | Sensor Scheduling for Energy-Efficient Target Tracking in Sensor Networks | [
"cs.MA"
] | In this paper we study the problem of tracking an object moving randomly through a network of wireless sensors. Our objective is to devise strategies for scheduling the sensors to optimize the tradeoff between tracking performance and energy consumption. We cast the scheduling problem as a Partially Observable Markov Decision Process (POMDP), where the control actions correspond to the set of sensors to activate at each time step. Using a bottom-up approach, we consider different sensing, motion and cost models with increasing levels of difficulty. At the first level, the sensing regions of the different sensors do not overlap and the target is only observed within the sensing range of an active sensor. Then, we consider sensors with overlapping sensing range such that the tracking error, and hence the actions of the different sensors, are tightly coupled. Finally, we consider scenarios wherein the target locations and sensors' observations assume values on continuous spaces. Exact solutions are generally intractable even for the simplest models due to the dimensionality of the information and action spaces. Hence, we devise approximate solution techniques, and in some cases derive lower bounds on the optimal tradeoff curves. The generated scheduling policies, albeit suboptimal, often provide close-to-optimal energy-tracking tradeoffs. | cs.MA | cs |
Sensor Scheduling for Energy-Efficient Target
1
Tracking in Sensor Networks
George K. Atia, Member, IEEE, Venugopal V. Veeravalli, Fellow, IEEE, and
Jason A. Fuemmeler Member, IEEE
Abstract
In this paper we study the problem of tracking an object moving randomly through a network of
wireless sensors. Our objective is to devise strategies for scheduling the sensors to optimize the tradeoff
between tracking performance and energy consumption. We cast the scheduling problem as a Partially
Observable Markov Decision Process (POMDP), where the control actions correspond to the set of
sensors to activate at each time step. Using a bottom-up approach, we consider different sensing, motion
and cost models with increasing levels of difficulty. At the first level, the sensing regions of the different
sensors do not overlap and the target is only observed within the sensing range of an active sensor. Then,
we consider sensors with overlapping sensing range such that the tracking error, and hence the actions
of the different sensors, are tightly coupled. Finally, we consider scenarios wherein the target locations
and sensors' observations assume values on continuous spaces. Exact solutions are generally intractable
even for the simplest models due to the dimensionality of the information and action spaces. Hence,
we devise approximate solution techniques, and in some cases derive lower bounds on the optimal
tradeoff curves. The generated scheduling policies, albeit suboptimal, often provide close-to-optimal
energy-tracking tradeoffs.
I. INTRODUCTION
In large networks of inexpensive sensors with small batteries, the sensor nodes are required to operate
on limited energy budgets. Sensor management can prolong the lifetime of a sensor network and conserve
This work was funded in part by a grant from the Motorola corporation, a U.S. Army Research Office MURI grant W911NF-
06-1-0094 through a subcontract from Brown University at the University of Illinois, a NSF Graduate Research Fellowship, and
by a Vodafone Fellowship.
This work was done at the Coordinated Science Laboratory (CSL), University of Illinois at Urbana Champaign, Urbana
IL, and was submitted in part
in June 2010 to the Asilomar conference on Signals, Systems, and Computers. Emails:
{atia1,vvv,fuemmele}@illinois.edu
DRAFT
scarce energy resources. However, inefficient management could result in severe performance degradation.
In this paper, we consider a network of n sensors tracking a single object. The sensors can be turned
on or off at consecutive time steps and the goal is to select the subset of sensors to activate at each time
step. This problem is challenging due to the inherent tradeoff between the value of information in the
sensor measurements and the energy cost, combined with the combinatorial complexity of the decision
space.
In previous work [1], two of the authors considered approximate strategies for sensor sleeping, where
the sensors are put to sleep to save energy and decisions are made concerning their sleep duration (in
time slots). Once in a sleep mode, a sensor would only wake up after its own sleep timer expires. Here,
we consider a scheduling variant of the problem which can be thought of as a sleeping problem with an
external wake-up mechanism, i.e., sensors can be woken up by external means (e.g. a low-power wake-up
radio). At time k, the permissible control actions for an n-sensor scheduling problem are n-dimensional
binary vectors, i.e., vectors in {0, 1}n (corresponding to set sensor nodes to activate at each time step),
for the sleeping problem (corresponding to the sleep durations of awake
in contrast to vectors in N
na(k)
0
sensors), where N0 is the set of non-negative integers and na(k) the number of awake sensors at time
k. While this does not address the combinatorial nature of the control space, the simpler structure of the
control space for the scheduling problem enables efficient approximate solution methodologies for the
more realistic models that we study in this paper.
A significant body of related research work considers sensor management for tasking sensors in
dynamically evolving environments. Castanon [2] has developed an approximate dynamic programming
approach for dynamic scheduling of multi-mode sensor resources for the classification of a large number of
unknown objects. The goal is to achieve an accurate classification of each object at the end of a fixed finite
horizon by assigning different sensor modes to different objects subject to periodic or total resource usage
constraints. Mode allocation strategies are computed based on Lagrangian relaxation for an approximate
optimization problem wherein sample-path resource constraints are replaced by expected value constraints.
In the context of sensor scheduling for target tracking, information-based approaches [3], [4] have been
developed for optimizing tracking performance subject to an explicit constraint on communication costs
in a decentralized setting. Williams et al. [3] also adopt a Lagrangian relaxation approach to solve a
constrained dynamic program over a rolling horizon. There, the combinatorial complexity of the decision
space is avoided by first selecting one leader node, followed by greedy sensor subset selection. Other
related work on sensor scheduling include leader-based distributed tracking schemes [5], [6], where at
any time instant there is only one sensor active, namely, the leader sensor which changes dynamically
2
as a function of the object state, while the rest of the network is idle.
While previous work focused on developing distributed implementations of efficient sensor scheduling
strategies, our goal here is to study the fundamental theory of sensor scheduling for tracking and
surveillance applications. Specifically, to explicitly study the fundamental tradeoff between tracking
performance and energy expenditure, we define a unified objective function combining tracking and
energy costs trading-off the complexity of per-stage costs to better capture the inherent energy-tracking
tradeoff. We adopt a bottom-up approach where we consider a range of sensing, motion and cost models
with increasing levels of difficulty and devise suboptimal scheduling policies to balance the tradeoff
between energy expenditure and tracking performance. In some cases we are also able to derive lower
bounds on the optimal energy-tracking tradeoff.
Due to noise and model uncertainties, natural limitations of the measurement devices, or incomplete
data about the surroundings, we need to design scheduling policies when the system's state is only partially
observable to the controller. Partially-Observable Markov Decision Processes (POMDPs) provide a natural
framework for addressing sequential decision problems where the goal is to find a policy (strategy) for
selecting actions based on the information available to the controller while addressing both short-term and
long-term benefits and costs. Solving POMDPs optimally is generally intractable. For example, the value
function for a POMDP with a finite state space depends on information states consisting of conditional
probability vectors of dimension equal to the number of states. This has led to a number of POMDP
approximations and we refer the reader to Monahan [7] and Hauskrecht [8] for excellent surveys on
approximate methods for stochastic dynamic programming. Usually, no single approximation can be
prescribed for all POMDPs, rather approximations can be judiciously used to exploit specific problem
structures. In this paper, we use a subset of these approximate solution techniques, including reduced-
uncertainty and point-based approximations [9] -- [12]. The former assumes that more information would be
available to the controller at future time steps, and the latter solves a reduced optimization problem based
on a relatively small subset of sampled beliefs about the object's state. We devise different approaches to
deal with the aforementioned computational complexity of the decision space. In one approach, instead
of solving one large combinatorial problem, we solve a set of simpler subproblems based on the intuition
gained from a simplistic sensing model. In another approach, we iteratively sample control actions from
a reduced control space based on the sparsity of a reachable belief set combined with point-based value
updates.
The remainder of this paper is organized as follows. In Section II, we describe the tracking problem
and define the sensing, transition and cost models, as well as the optimization problem, for each of the
3
considered models. In Section III we describe approximate strategies to generate suboptimal scheduling
policies. In Section IV, we present some experimental results, and finally, in Section V, we provide some
concluding remarks.
II. SCHEDULING PROBLEM
In the following we consider different models with increasing level of difficulty. Depending on the
structure of the model, we devise approximate methods to address the associated difficulties and generate
efficient scheduling policies. For notation, vectors are denoted by bold lower-case letters. Superscript T
denotes transposition and the indicator function is written as 1I{.}.
A. Simple sensing, observation and cost models
In this model, the network is divided into n distinct cells, one for each sensor. In other words, each cell
corresponds to the sensing range of one particular sensor and sensors' ranges do not overlap. A Markov
chain with an (n + 1)× (n + 1) probability transition matrix P describes the motion of the target through
the field of interest. The extra state is for an absorbing termination state of the Markov chain which is
reached when the object leaves the network. It is further assumed that all information about the object
trajectory is stored at some central unit and is used to determine the scheduling actions for the different
sensors.
We let uk,ℓ denote the action for sensor ℓ at time k; uk,ℓ = 1 if sensor ℓ is activated at time k + 1 and
0 if the decision is to turn it off. The action vector at time k, denoted uk, is a binary vector of size n× 1,
one decision per sensor. In this simplistic model, we assume that the target is perfectly observable within
the cell of an awake sensor or if it reaches the terminal state τ , otherwise it is unobservable. Thus, the
observation sk at time k is defined according to:
bk,
ε,
τ,
if bk 6= τ and uk−1,bk = 1;
if bk 6= τ and uk−1,bk = 0;
if bk = τ .
(1)
sk =
where ε stands for erasure. The observation model in (1) induces a well-defined probabilistic observation
model p(skbk, uk−1) such that the current observation depends on that actual target location and the
scheduling action for the n sensors.
At each time step, the incurred cost is the sum of the energy and the tracking costs. An energy cost
of c ∈ (0, 1] per unit time is incurred for every active sensor and a tracking cost of 1 for each time unit
that the object is not observed. Once state τ is reached the problem terminates and no further cost is
4
incurred. In other words, τ is an absorbing cost-free state; all n states are transient so that τ is the only
recurrence class of the Markov chain. Hence,
g(bk, uk−1) = 1I{bk 6= τ} 1I{uk−1,bk = 0} +
c1I{uk−1,ℓ = 1}!
(2)
n
Xℓ=1
The parameter c is thus used to tradeoff energy consumption and tracking errors.
B. Overlapping sensors with discrete observations models
In this model, we continue to use a discrete model for the target transition but we redefine a new
sensing model and cost structure to account for the fact that sensors could have overlapping visibility
regions. Within that model we further consider simple and probabilistic sensing.
1) Overlapping sensors with simple sensing: In this case, the target is perfectly observed within the
visibility region of any active sensor. Denote by Rℓ the set of locations in the visibility region of sensor
ℓ and by Bi the set of sensors that observe location i. The observation at time k is as follows:
bk,
ε,
τ,
if bk 6= τ and ∃j ∈ Bbk : uk−1,j = 1;
if bk 6= τ and uk−1,j = 0, ∀j ∈ Bbk;
if bk = τ .
(3)
sk =
Therefore, a tracking error is incurred if none of the sensors observing the current target location is
active. Redefining the cost structure for this model:
g(bk, uk−1) = 1I{bk 6= τ} 1I{uk−1,j = 0,∀j ∈ Bbk} +
c1I{uk−1,ℓ = 1}!
(4)
n
Xℓ=1
2) Overlapping sesnors with probabilistic sensing: By probabilistic sensing we account for observation
uncertainty even if the target is within the visibility region of one or more active sensors. We assume,
where
p(skbk,∃j ∈ Bbk : uk−1,j = 1) =
Rj(cid:15) [i /∈Bbk ,
R = \j∈Bbk ,
1−q
uk−1,j=1
uk−1,i=1
Ri.
sk = bk;
q,
R−1, sk = i, ∀i ∈ R
(5)
That is, the observation is uniformly distributed over the remaining locations (other than the true target
location) that belong to the visibility regions of the set of awake sensors monitoring the true location
bk. If the true target location does not belong to the visibility region of an awake sensor, we naturally
exclude the visibility region of that sensor since no measurement is received from such a sensor. When
5
R is a singleton {bk}, we set q = 1. A tracking error is incurred if the target is not directly observed
and the uncertainty in the target location cannot be resolved.
C. Continuous observation, continuous state and arbitrary cost models
In this class of models, the object sensing model allows for an arbitrary distribution for the observations
given the current object location. Tracking cost is modeled as an arbitrary distance measure between the
actual and the estimated object location. If we denote the set of possible object locations B, we have
B = m + 1. Note that, in contrast to the simplistic model in II-A, m is different from n since object
locations are arbitrary and we no longer assume one location corresponds to the sensing range of one
particular sensor. The (m + 1)-th state again corresponds to a termination state. Furthermore, the target
can be moving on a continuous state space in which case m is ∞.
If the state space is discrete, then conditioned on the object state bk at time k, bk+1 has a probability
mass function that is given by the bk-th row of the transition matrix P . If the state space is continuous,
P is a kernel such that P (x,Y) is the probability that the next object location is in the set Y ⊂ B given
the current object location is x. For simplicity of exposition, we focus on discrete state spaces. Also, we
omit indexing time whenever the time evolution is well-understood to avoid cumbersome notation. We
consider the following observation model for illustration; however, our approach is fairly general:
n
p(sb, u) =
Yi=1( 1
√2π
exp −
1
2(cid:18)si −
10
(b − pi)2 + 1(cid:19)2! 1I{ui = 1} + δ(si − ε)1I{ui = 0})
(6)
where s is an n× 1 continuous observation vector with the i-th entry, si, representing the observation of
sensor i, pi, i = 1, . . . , n, is the position of the i-th sensor, b is the target state, and ε stands for erasure.
δ(.) is the Dirac Delta function. In (6), the observation of an active sensor is Gaussian with a mean
received signal strength inversely proportional to the square of the distance between the sensor and the
actual target location. The observation of an inactive sensor is just an erasure.
The estimated target location (given the entire history) is denoted by b. We define the tracking error
through an arbitrary bounded distance function d(b, b) between the actual and the estimated object
locations, which can be the Hamming distance d(b, b) = 1I{b 6= b} or the Euclidean distance for discrete
and continuous state spaces, respectively. The control at each time step is the tuple (bk, uk). Since b does
not affect the state evolution, the optimal value for bk is the value that minimizes the tracking cost over
a single time step given history up to time k, i.e.,
bk = arg min
b
E[d(bk, bk)Ik]
(7)
6
where, Ik denotes the information state, i.e., the total information available to the central controller at
time k which is given by
In the case of Hamming cost, it follows that b is simply the MAP decision, i.e., b = arg maxb pk(b).
Ik = {s0, s1 . . . , sk, u0, u1 . . . , uk−1}
D. Optimal scheduling policy
The design of an optimal scheduling policy depends on the history up to time k, i.e., the information
state Ik. However, the posterior probability distribution, pk = Pr[bkIk], of the target's state given Ik is
a sufficient statistic for this class of partially observable processes. The distribution pk, also known as
belief, summarizes all the information needed for optimal control. The sufficient statistic itself forms a
Markov process whose evolution can be obtained through Bayes' rule updates 1. For example, the belief
update equation for the simplistic model in Section II-A can be written as:
eτ ,
ebk+1,
[pkP ]{j:uk,j=0} ,
if sk+1 = τ ;
if uk,bk+1 = 1;
if uk,bk+1 = 0.
(8)
pk+1 =
where ei is a row vector with a 1 at the i-th entry and 0 elsewhere. The vector [pkP ]S is the probability
vector formed by setting the i-th entry [pkP ]i of the vector pkP to zero, ∀i /∈ S, and then normalizing
the vector into a probability distribution. The set {j : uk,j = 0} signifies the set of deactivated sensors.
In other words, the updated belief for the model in II-A, is a point mass distribution concentrated at τ
if the object exits the network, and concentrated at bk+1 if the object is observed. When the object is
unobservable, we eliminate the probability mass at all sensors that are awake, since the object cannot be
at these locations, and normalize. The multi-valued function in (8), and equivalent Bayes' updates for
the other models, define a transformation pk+1 = φ(pk, sk+1, uk), mapping the current belief pk, the
current control vector uk, and the future observation sk+1, to a future belief.
The policy uk = µk(Ik) is defined as a mapping from information states Ik to control actions uk. The
goal is to design a policy that minimizes the expected sum of costs J, where,
J(I0, µ0, µ1, . . .) = E" ∞
Xk=1
g(bk)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
I0# .
(9)
1Equivalently, for a continuous state space, a sufficient statistic would be pk(X ) = Pr[bk ∈ X Ik]. The updated belief pk+1
can be computed using standard Bayesian non-linear filtering as the posterior measure resulting from prior measure pP and
observation sk+1.
7
J is well-defined since g is upper bounded by cn + 1 (regardless of the model) and the expected time
till the object exits the network is finite. Note that the termination is inevitable, thus the objective is to
reach the termination state with minimal expected cost. Hence, the scheduling policy is the solution of
the minimization problem,
J ∗ = min
µ0,µ1,...
J(I0, µ0, µ1, . . .)
(10)
This POMDP problem falls within the class of infinite horizon stochastic shortest path problems.
Noting that the termination state is observable, cost-free and absorbing, and that every policy is proper2,
a stationary policy µ∗(.), i.e., one which does not depend on k, is optimal in the class of all history-
dependent policies and pk is a sufficient statistic for control [13], i.e., u∗
k = µ∗(pk), is defined through
a time-invariant mapping from the belief space to the action space. J can be written in terms of the
sufficient statistic and the optimal policy can be obtained from the solution of the Bellman equation:
J(p) = min
u∈{0,1}n
E[g(b′, u)p, u] +Xs
p(sp, u)J(φ(p, s, u))
(11)
such that J(eτ ) = 0, where J(.) is the value function for the POMDP, and the expectation is taken
over the future state b′ which is distributed according to pP . Note that we removed the time dependence
due to the aforementioned time invariance property. For continuous observations, summation over s is
replaced by an integration.
III. APPROXIMATE SOLUTIONS AND LOWER BOUNDS
There are a number of algorithms for solving POMDPs exactly [14] -- [16]. These algorithms rely on the
powerful result of Sondik that the optimal value function for any POMDP can be approximated arbitrarily
closely using a set of hyper-planes (α-vectors) defined over the belief simplex [14]. This fact is the basis
for exact value iteration based algorithms, such as the Witness algorithm [17] for computing the value
function. The result is a value function parameterized by a number of hyper-planes (or vectors) whereby
the belief space is partitioned into a finite number of regions. Each vector minimizes the value function
over a certain region of the belief space and has a control action associated with it, which is the optimal
control for beliefs in its region.
To clarify, in value iteration we generally start with some initial estimate for J ∗ and repeatedly apply
the transformation defined by the right hand side of Bellman equation (11) until the sequence of cost
2A proper policy is a policy that leads to the termination state with probability one regardless of the initial state. In our
problem, the scheduling policy does not affect the target motion and all policies are proper in the sense that there is a positive
probability that the target will reach the termination state after a finite number of stages.
8
functions converges. Let {α
after k iterations, where J (k) is the total number of hyper-planes, and α
in the belief space, represents the value of executing the k-step policy associated with the i-th vector
denote the set of vectors parameterizing the value function J (k)
(b), which is a hyperplane
(k)
i
(k)
i }J (k)
i=1
starting from a state b. Hence, the value of executing the i-th hyperplane policy starting from a belief
state p is simply the dot product of αi and p:
J (k)
i
(p) =Xb
p(b)α
(k)
i
(b) = p · α
(k)
i
.
Therefore, the value of the optimal k-step policy starting at p is simply the minimum dot product over
all hyperplanes, i.e.,
J ∗(k)(p) = min
{α(k)
i }
(k)
i
p · α
.
Hence, J ∗(k)(p) is piecewise linear and concave. Some of the vectors (also known as policy trees)
may be dominated by others in the sense that they are not optimal at any region in the belief simplex.
Thus, many exact algorithms devise pruning mechanisms whereby a parsimonious representation with a
minimal set of non-dominated hyper-planes is maintained [7].
Even though the aforementioned linearity/concavity property makes the policy search a great deal
simpler, the exact computation is generally intractable except for relatively small problems. The two
major difficulties for exact computation arise from the exponential growth of the vectors with the planning
horizon and with the number of observations, and the inefficiencies related to identification of such vectors
and subsequently pruning them. Namely, the number of hyper-planes grows double exponentially such
that after k steps the number of hyperplanes is O(cid:16)USk(cid:17), where U and S denote the cardinality of
the control and observation spaces, respectively. Equivalently, the number of hyperplanes per iteration
grows as:
This has led to a number of approximations and suboptimal solutions techniques trading off solution
J (k+1) = O(cid:16)UJ (k)S(cid:17) .
quality for speed.
Remark III.1. The intractability of the optimal solution for our problem is primarily due to the following
reasons:
(i) The cost function is minimized over the simplex of probability distributions, i.e., the (m − 1)-
dimensional belief simplex for m-state discrete state-space models, and the space of probability
density functions for continuous state-space models.
9
(ii) The exponential explosion of the action space with the number of sensors (2n actions).
(iii) The exponential growth of the α-vectors with the planning horizon and with the number of obser-
vations, especially for continuous observation models.
A. Approximate solutions
In this section, we outline our approximate solution methodologies for the different models introduced
in Section II. First, we consider approximations where it is assumed that more information becomes
available to the controller at future time steps. Policies based on the assumption that uncertainty in the
current belief state will be gone after the next action were first introduced within the artificial intelligence
community and known as QMDP policies
assumption, our sensor scheduling problem decomposes into n simpler subproblems, one subproblem
[10], [17]. We show that under an observable-after-control
per sensor, for the simplistic model of II-A. These subproblems can then be solved exactly using policy
iteration [13]. Furthermore, in this case, the QMDP solution gives us a lower bound on the optimal
tracking-energy tradeoff. Unfortunately, this natural decomposition does not extend to the other class of
models due to the inherent coupling of their tracking errors. However, based on intuition gained from
the simplistic model, we artificially decouple the scheduling problem for those models and individually
learn the tracking costs corresponding to each subproblem under the aforementioned QMDP assumption.
This approach combines QMDP with reinforcement learning [18].
Second, we develop sensor scheduling strategies based on point-based approximations. Despite the
fact that the generated QMDP based policies perform reasonably well, generally the resulting policies
would not take actions to gain information (an effect of the observable-after-control assumption), leading
to situations wherein the belief state does not get updated appropriately. Furthermore, while decoupling
the scheduling problem provides close-to optimal performance for uncoupled or lightly-coupled sensing
and tracking models (see Section IV), it might come at the expense of reduction in solution quality for
more realistic or heavily-coupled models. To that end, we develop point-based approximate scheduling
policies. While our previous approach reduced complexity via decoupling and learning, the key idea here
is to optimize the value function only for a small set of reachable beliefs P and not over the entire
belief simplex. Point-based methods have shown great potential for solving large scale POMDPs mostly
for robotic applications [8], [9], [11], [19]. Pineau et al. [9] proposed point-based value iteration (PBVI)
which performs point-based backups only at a discrete set of reachable belief points, that can be actually
encountered by interacting with the environment. Developing a class of point-based algorithms, which
mostly differ in the way the subset of belief points is chosen and the execution order of the backup
10
Belief point (p)
Reduce
scheduling
control space
U(p)
Point-based
update
new policy
Aggregated
measurements
weights
Aggregate
multi-dimensional
sensor
measurements
past policy
Fig. 1: Structure of the point-based scheduling approximation
operations over the selected belief points, has been the focus of recent algorithm-development research
targeting large scale POMDPs. Perseus [11] is one such randomized point-based algorithm that maintains
a fixed set of belief points. There, backup speedups can be obtained by exploiting the key observation
that a single backup may improve the value of many belief points simultaneously. These algorithms were
designed to deal with large state spaces, yet, two extra difficulties in the scheduling problem arise from
the size of the action space 2n (for all models) and the observation space (for the models in Sections II-C).
Regarding the dimensionality of the action space, we devise a strategy to sample actions based on the
support of the beliefs and the sparse structure of the transition models. Intuitively speaking, an object can
only move from one side of the network to the other side within time constraints rendering exponentially
many scheduling actions irrational at certain times. Hence, instead of performing full updates including
2n actions, we perform the minimization over a reduced control space U (p) for every p ∈ P (see Section
III-C1). When dealing with continuous or large observations, we combine that with a methodology that
aggregates observations and uses aggregate observations for value iteration updates (Section III-C2). At
the core of the algorithm we use Perseus [11], a variant of PBVI [9], whereby value iteration updates
are not carried out for every sampled belief. Instead, the values for many belief points are improved
simultaneously in one update. Fig. 1 depicts the structure of our point-based approximation, combining
control space reduction and observation aggregation with point-based updates.
B. QMDP based scheduling policies
Next, we consider our first class of policies based on the QMDP reduced future uncertainty assumption.
First, we consider the simplistic model in Section II-A, then we use the intuition we developed from this
model to devise similar policies for the other models. Since the POMDP is a stochastic shortest path
problem with an absorbing cost-free termination state, and the expected termination time is finite, the
11
cost-to-go function for a given belief can be written as the minimum of the dot product of the belief
vector and a set of hyper-planes (α vectors):
J(p) = min
= min
αi(b)p(b)
{αi}Xb
u∈{0,1}n( n
[pP ]i 1I{ui = 0} +
Xi=1
αi Xb′
+ Xs∈{1...n,ε}
min
n
c1I{uℓ = 1}!
Xℓ=1
p(su, b′)Xb
p(b′b)p(b)αi(b′))
(12)
where {αi} is the set of hyperplanes constituting the value function J. In essence, the complexity of the
Bellman equation (12) stems from the evolution of the belief pk in (8). We can see why (12) is hard
to analyze if we further divide the second term in the summation into two terms depending on whether
there is observability or there is an erasure,
J(p) = min
u∈{0,1}n( n
[pP ]i 1I{ui = 0} +
Xi=1
+Xb′
1I{ub′ = 1}[pP ]b′ min
{αi}
n
Xℓ=1
c1I{uℓ = 1}!
{αi}Xb′
αi(b′) + min
1I{ub′ = 0}[pP ]b′ αi(b′)).
(13)
To further clarify we observe that:
n
n
p(su, p)J(p1) =
1I{ui = 1}[pP ]iJ(ei) +
1I{ui = 0}[pP ]iJ([pP ]{j:uj=0})
(14)
Xs
Xi=1
Xi=1
and the minimization problem is coupled across the sensors as the second term in (14), which is due
to non-observability, depends on the action vector u. The action of one sensor affects belief evolution
therefore coupling the problem across sensors. Now, if we make the assumption that perfect observations
would be available to the controller after taking a scheduling action, we obtain an approximate surrogate
function which can be used to generate a suboptimal scheduling policy. Namely, we replace p(su, b′) =
δ(s − b′) in (12). We get
J(p) = min
u∈{0,1}n( n
Xi=1
u∈{0,1}n( n
Xi=1
= min
[pP ]i 1I{ui = 0} +
[pP ]i 1I{ui = 0} +
n
n
Xℓ=1
Xℓ=1
c1I{uℓ = 1}! +Xb′
c1I{uℓ = 1}! +Xb′
[pP ]b′ min αi · eb′)
[pP ]b′J(eb′)).
(15)
The terms in the summation in (15) only depend on the control action for each sensor. Furthermore,
the belief evolution is independent of the scheduling action, wherefore the approximate recursion in (15)
decomposes into separable terms, one per sensor. Hence, the value function and the scheduling policy for
12
sensor ℓ, under the observable-after-control assumption, can be obtained from the solution of per-sensor
Bellman equation:
J (ℓ)(p) = min
uℓ∈{0,1}( n
Xi=1
[pP ]i 1I{ui = 0} +
n
Xℓ=1
c1I{uℓ = 1}! +Xb′
[pP ]b′J (ℓ)(eb′)).
(16)
The POMDP problem is now decomposed into n separate simpler subproblems such that the total cost
function is the sum of the per-sensor cost function while the overall scheduling policy is the per-sensor
policies applied in parallel. Each subproblem can be easily solved using standard policy iteration [13]
with a simple minimization over a binary control action.
Fundamentally, for the simplistic model, we were able to decompose the problem into n simpler
subproblems due to the separability of the tracking cost into per-sensor costs. Note that the problem is
still coupled due to the belief evolution in (8) yet that coupling is resolved under the observable-after-
control assumption.
While separability holds for the simplistic model, this is not the case for the other models. Hence, we
devise a strategy where we artificially decouple the problem into n simpler subproblems. To this end,
we perform Monte Carlo simulations to determine appropriate values for the per-sensor tracking cost
corresponding to each subproblem. For example, consider the continuous observation model of Section
II-C. For simplicity of exposition, assume a discrete state space model with m possible object locations.
In this case, we define a surrogate value function for the ℓ-th subproblem as follows:
J ℓ(p) = min
u (1I{u = 0}
m
m
p(i)T (i, ℓ)+1I{u = 1}
c[pP ]i +
Xi=1
m
[pP ]iJ ℓ(ei)) ℓ = 1, . . . , n (17)
Xi=1
Xi=1
where T (i, ℓ) captures the contribution of the ℓ-th sensor to the total tracking error when the target's
previous state is i and is obtained via Monte Carlo simulations. Namely, the expected tracking cost can
be evaluated by repeatedly simulating our system from time k − 1 to time k while changing the state of
the ℓ-th sensor. Similarly, (17) can be generalized for continuous state spaces.
Even though the QMDP assumption leads to a separable problem and provides a lower bound on the
optimal energy-tracking tradeoff for the simplistic model as we elaborate in Section III-D, the resulting
scheduling policies are myopic, unlike the sleeping policies in [1]. This follows from the fact that under an
observable-after-control assumption, the future cost term is independent of the control vector u. Therefore,
we consider more efficient, albeit more difficult, point-based approximations in the next section.
C. Point-based approximate policies
In the previous section, we described QMDP based policies, whereby issues (i) and (iii) in Remark III.1
are resolved since we only needed to solve the underlying Markov Decision Process to describe the full
13
approximate surrogate function. Decoupling the problem into one-per-sensor subproblems (naturally or
artificially) further enabled us to address issue (ii). Yet, we just argued in Sections III-A and III-B that the
resulting scheduling policies are myopic and generally do not take control actions to gain information.
To that end, we develop point-based approximate scheduling policies. Instead of reducing complexity
via artificial decoupling and learning, the key idea here is to optimize the value function only for
specific reachable sampled beliefs and not over the entire belief simplex (addressing issue (i) in Remark
III.1). Such techniques have shown great potential for solving large scale POMDPs while significantly
reducing complexity. Due to the large size of the control space, we also devise strategies to sample
actions exploiting the sparsity of the beliefs and the problem structure (to address issue (ii)). Moreover,
observation aggregation is used for continuous observation models. Furthermore, since Perseus updates
are not carried out for every sampled belief and multiple belief points are improved simultaneously, the
number of α vectors grows modestly with the number of iterations. This addresses issue (iii) in Remark
III.1.
For completeness we first briefly outline the steps of Perseus and refer the reader to [11], [12] for
further details. Later, we discuss specific variations to the algorithm to address the dimensionality of the
action and the observation spaces.
One iteration of Perseus
1) Sample a set of belief points P. We obtain these beliefs by simulating the target motion through
the field taking random actions and generating observation according to the observation models in
(1), (3), (5), and (6)
2) Sample a belief point p ∈ P at random and compute the backup using (18a) and (18b),
α = arg min
{αp
u}u∈U
p
u
p · α
where
α
p
u = g(b, u) +Xs
p(su, p) min
α(k)
i
φ(p, u, s) · α
(k)
i
(18a)
(18b)
3) If Pb p(b)α(b) ≤ J (k)(p) then add new α to J (k+1) otherwise keep old hyperplane
4) If {p ∈ P : Jk+1(p) > J (k)(p)} = ∅, i.e., the empty set, iteration is complete otherwise repeat from
step 1
Fig.2 illustrates the progress of one iteration of Perseus. The x-axis represents the belief space with
circles representing the sampled belief set P = {p1, . . . , p7}. The y-axis is the value function at
consecutive iterations, i.e. J k−1 (solid lines) and J k (dashed lines). The figure displays the α vectors and
14
(a)
(b)
(c)
(d)
(e)
Fig. 2: One iteration of Perseus illustrating the progress of the algorithm. The x-axis represents the
belief space with circles representing the sampled belief set P = {p1, . . . , p7}. The y-axis is the value
function at consecutive iterations, i.e. J k−1 and J k. Solid lines represent the hyper-planes in the (k−1)-th
iteration and dashed lines represent the newly added hyper-planes during the k-th iteration. (a) The initial
value function J k−1; (b) p1 is randomly selected and a new α vector is added to J k. This update step
only happens to improve p1. Dark circles represent belief points which did not yet improve; (c) p3 is
sampled and a new hyperplane is added which improves the value for p2 through p6; (d) Only p7 did
not improve, thus p7 is sampled and a new hyperplane is added to J (k); (e) All belief points improved,
J (k) is computed, the iteration ends.
different steps illustrating the progress of the algorithm. The algorithm selects a belief point at random
and updates the value function for that belief. Then a new update is carried out for a belief point randomly
selected from the set of remaining beliefs, i.e., beliefs which did not improve in the previous step. The
algorithm repeats till all belief points are updated. Solid lines represent the hyper-planes in the (k− 1)-th
iteration and dashed lines represent the newly added hyper-planes during the k-th iteration. In a way, the
Perseus updates in POMDPs are the counterpart of asynchronous dynamic programming for MDPs [13]
since the order of backup of the belief points is arbitrary and does not require full sweeps over the entire
sampled belief set.
15
1) Sampling actions based on the support of the belief: Note that the update equation (18) involves
a minimization over all control actions in U. Even though one iteration of the algorithm is linear in
the cardinality U of the control space, U itself is exponential in the number of sensors rendering the
minimization infeasible for a relatively large sensor network.
The idea here is to exploit the structure of the scheduling/tracking problem. Since the target transition
model is naturally sparse, we predict relatively small uncertainty regions for the target state at future time
steps. More specifically, for every belief point in P, we use prior information about the target transition
model to project the future state of the target. This is particularly useful when the current belief vector
is sparse leading to more restricted uncertainty regions. Subsequently, we restrict our attention to a
significant subset of sensors, that is, sensors of relevance to the particulars of the uncertainty region.
Hence, we only consider scheduling actions involving scheduling different combinations of a reduced
number of sensors which considerably reduces the control space for every belief in P. If the number
of significant sensors is still large, we randomly sample actions from the reduced control space. Note
that the same intuition extends to more complex motion models wherein information about target speed,
maneuver, and acceleration can be factored in to define the future uncertainty regions. Hence, instead of
performing full updates including 2n actions, we perform the minimization over a reduced control space
for every p ∈ P. Specifically, we redefine the point update equation as:
α = arg min
{αp
u}u∈U(p)
p
u
p · α
(19)
where U (p) designates the reduced control space for the belief vector p.
Note that, future iterations of the algorithm involving a particular belief point, ensure sufficient sampling
to relevant control actions in the reduced control space. This approach is well suited to Perseus wherein
the value for every belief point is guaranteed to improve over consecutive stages of the algorithm. It
is worth mentioning that the observation and the cost models need to be computed on the fly for each
sampled control action during the algorithm implementation.
2) Observation aggregation: The point update equation (18) involves back-projecting all hyper-planes
in the current iteration one step from the future and returning the vector that minimizes the value of
the belief. Since this involves computing a cross sum by enumerating all possible combinations of
alpha vectors for the different observations, a number of vectors which is exponential in the number
of the observations is generated at each stage. The recursion has to be redefined to address continuous
observation models. Looking carefully at (18), it is not hard to see that if different observations map
16
to the same minimizing hyperplane, then they can be aggregated [20]. Hence, if we can partition the
observation space into regions that map to the same hyperplane (possibly non contiguous), the continuous
model is reduced to a corresponding discrete model. Integration is replaced by a summation over these
partitions and the weighing probabilities are obtained by integrating the conditional density over these
partitions. This is clarified in the following:
Zs
min
αi Xb′
p(su, b′)Xb
p(b′b)p(b)αj(b′) ds
p(b′b)p(b)αi(b′) ds =Xj ZSjXb′
=Xj Xb′
=Xj Xb′
p(su, b′)Xb
[pP ]b′ αj(b′)ZSj
[pP ]b′ Pr[Sju, b′]αj(b′).
p(su, b′) ds
To find the regions of aggregate observations, we need to solve for the boundaries, i.e., for each pair
(20)
(21)
(22)
(23)
(24)
(i, j) of α vectors we need to solve for s:
αi · φ(p, u, s) = αj · φ(p, u, s)
Hence, we need to solve:
where φ(p, u, s) = p1(b′) ∝Pb p(b)p(sb′, u)p(b′b)
(αi(b′) − αj(b′))[pP ]b′ exp(−
1
Xb′
After solving for the boundaries, we can readily define the regions:
2 Xi:ui=1
(si −
10
(b′ − pi)2 + 1
)2) = 0
Sj ∗ = {sj∗ = arg max
j
αj · φ(p, u, s)}
Now the update step is simply:
J(p) = g(p, u
where
[pP ]b′ Pr[Sju
∗, b′]αj(b′)
∗) +Xj Xb′
∗, b′] =Zs∈Sj
Pr[Sju
∗, b′)ds.
p(su
Finding a closed form analytical solution for (22) is not feasible. Instead, we use Monte-Carlo simulations
to solve for the boundaries and get estimates of the weighing probabilities by sampling observations from
p(su, b′) for different combinations of actions and target states.
17
D. Lower bounds
We are able to derive lower bounds on the energy-tracking tradeoff for the simple as well as the
continuous Gaussian observation models. For the simple model, the QMDP value function is itself a lower
bound on the expected total cost since more information is available to the controller at future time steps
given the reduced uncertainty assumption. To further clarify, observe that if we interchange the order of
minimization and summation in the last term of (13), we obtain a lower bound on the optimal cost to go
function. Hence, a lower bound can be obtained from the solution of the following equation:
1I{ub′ = 0}[pP ]b′ min
{αi}
αi(b′))
[pP ]b′ min αi · eb′)
[pP ]b′J(eb′))
(25)
J(p) = min
Xℓ=1
1I{ub′ = 1}[pP ]b′ min
u∈{0,1}n( n
[pP ]i 1I{ui = 0} +
Xi=1
+Xb′
u∈{0,1}n( n
[pP ]i 1I{ui = 0} +
Xi=1
u∈{0,1}n( n
[pP ]i 1I{ui = 0} +
Xi=1
Xℓ=1
Xℓ=1
c1I{uℓ = 1}!
αi(b′) +Xb′
c1I{uℓ = 1}! +Xb′
c1I{uℓ = 1}! +Xb′
{αi}
n
n
n
= min
= min
Interchanging the order of the summation and minimization corresponds to a fully observable state after
the next scheduling action, i.e., that the future belief is eb′. Hence, the QMDP value function is a lower
bound on the cost function of the original problem.
Unfortunately, this is only true for the simplistic model and does not extend to the coupled models
since the factored tracking cost in (17) need not be a lower bound on the true tracking cost.
To obtain a lower bound on the optimal energy-tracking tradeoff for such models, we combine the
observable-after-control assumption with a decomposable lower bound on the tracking cost which we
derive next. Consider the continuous observation model with discrete state space. Given the current
belief pk and a control vector uk the expected tracking cost can be written as:
E[d(bk+1, bk+1)pk, uk] =
=
m
m
Xj=1
Xi=1
Pr[bk+1 6= jpk, uk, bk+1 = j] Pr[bk+1 = jpk, uk]
pk(i)
m
Xj=1
p(bk+1 = jbk = i) Pr[bk+1 6= jpk, uk, bk+1 = j]
(26)
Defining
P (EHj) , Pr[bk+1 6= jpk, uk, bk+1 = j]
18
which is a conditional error probability for a multiple hypothesis testing problem with m hypotheses,
each corresponding to a different mean vector contaminated with white Gaussian noise. Conditioned on
Hj the observation model is:
Hj : s(ℓ) = (mj(ℓ) + w(ℓ))1I{uk,ℓ = 0} + ε1I{uk,ℓ > 0}
(27)
where s(ℓ) is the ℓ-th entry of an n×1 vector s denoting the received signal strength at the n sensors, mj
is the mean received signal strength when the target is at state j (j-th hypothesis), and w is a zero mean
white Gaussian Noise, i.e. w ∼ N (0, σ2I). According to (27), sensor ℓ gets a Gaussian observation,
which depends on the future target location, if activated at the next time step, and an erasure, otherwise.
Since the current belief is pk, the prior for the j-th hypothesis is πj = [pkP ]j. The error event E can
be written as the union of pairwise error regions as
where
p(EHj) = Pr[∪k6=jζkj]
ζkj = {s : Lkj(s) >
πj
πk }
(28)
is the region of observations for which the k-th hypothesis Hk is more likely than the j-th hypothesis
Hj and where
denotes the likelihood ratio for Hk and Hj. Using standard analysis for likelihood ratio tests [21], [22],
it is not difficult to show that:
(29)
Lkj ,
f (sHk)
f (sHj)
p(ζkjHj) = Q dkj
2
+
ln πj
πk
dkj !
∆mkj
∆mT
kj
σ2
kj =
where, d2
, ∆mkj = mk − mj, and Q(.) is the normal distribution Q-function. The
quantity dkj plays the role of distance between the two hypothesis and hence depends on the difference
of their corresponding mean vectors and the noise variance σ2. Note that, for different values of k and
j, ζkj are not generally disjoint but allow us to lower bound the error probability in terms of pairwise
error probabilities, namely, a lower bound can be written as:
p(EHj) ≥ max
k6=j
p(ζkjHj).
(30)
19
And we can readily lower bound the expected tracking error:
m
m
E[d(bk+1, bk+1)pk, uk] ≥
=
pk(i)
pk(i)
m
Xi=1
Xi=1
m
Xj=1
Xj=1
p(bk+1 = jbk = i) max
k6=j
p(bk+1 = jbk = i) max
k6=j
p(ζkjHj)
Q dkj
2
+
ln πj
πk
dkj !
(31)
Next we separate out the effect of each sensor on the tracking error:
E[d(bk+1, bk+1)pk, uk] ≥ 1I{uk,ℓ = 1}E[d(bk+1, bk+1)pk, uk = 1]
+ 1I{uk,ℓ = 0}E[d(bk+1, bk+1)pk, uk,i = 0 ∀i 6= ℓ]
for every ℓ
(32)
where 1 is the vector of all ones designating that all sensors will be active at the next time slot. The
inequality in (32) follows from the fact that if we separate out the effect of the ℓ-th sensor we get a better
tracking performance when all the remaining sensors are awake. Since this holds for every ℓ, a lower
bound on the expected tracking error can be written as a convex combination of all sensors contributions:
E[d(bk+1, bk+1)pk, uk] ≥
n
Xℓ=1
λℓ(pk)n1I{uk,ℓ = 1}E[d(bk+1, bk+1)pk, uk = 1]
+ 1I{uk,ℓ = 0}E[d(bk+1, bk+1)pk, uk,i = 0 ∀i 6= ℓ]o
(33)
Let 1−ℓ denote a vector of length n with all entries equal to one except for the ℓ-th entry being zero.
where, Pℓ λℓ(pk) = 1.
Then replacing from (31),
n
E[d(bk+1, bk+1)pk, uk] ≥
λℓ(pk)(1I{uk,ℓ = 1}
+1I{uk,ℓ > 0}
Xℓ=1
m
m
pk(i)
pk(i)
m
Xi=1
Xi=1
m
Xj=1
Xj=1
p(bk+1 = jbk = i) max
k6=j
p(bk+1 = jbk = i) max
k6=j
+
2
Q dkj(1)
Q dkj(1−ℓ)
2
ln πj
πk
dkj(1)!
dkj(1−ℓ)!) (34)
ln πj
πk
+
To simplify notation, we define the following 2 quantities:
T1(p; i, ℓ) ,
T (p; i, ℓ) ,
m
k6=j
Xj=1
Xj=1
p(bk+1 = jbk = i) max
Q
p(bk+1 = jbk = i) max
Q
k6=j
m
dkj(1)
2
dkj(1−ℓ)
2
+
+
ln [pP ]j
[pP ]k
dkj(1)
dkj(1−ℓ)
ln [pP ]j
[pP ]k
20
J (ℓ)(p)
J(p) =Xℓ
uℓ∈{0,1} 1I{uℓ = 1} Xb
+ 1I{uℓ = 0}Xb
p(b)λℓT (p; b, ℓ) +
m
[pP ]i!
Xi=1
[pP ]iJ (ℓ)(ei)!
Xi=1
m
(35)
(36)
Intuitively, T1(p; i, ℓ) represents the contribution of sensor ℓ to the total expected tracking cost when the
underlying state is i, the belief is p, and when all sensors are awake. On the other hand T (p; i, ℓ) is the
ℓ-th sensor contribution when it is inactive and all the other sensors are awake.
Now if we assume that the target will be perfectly observable after taking the scheduling action, a lower
bound on the total cost can be readily obtained from the solution of the following Bellman equation:
where
J (ℓ)(p) = min
p(b)λℓT1(p; b, ℓ) + c
Note that if we can solve the equation above for p = ei for all i ∈ {1, . . . , m}, then it is straightforward
to find the solution for all other values of p. We therefore focus on specifying the value function at those
points. Since this is the case, we further simplify our notation and use T (i, ℓ) and λ(i, ℓ) as shorthand
for T (ei; i, ℓ) and λℓ(ei), respectively. We can see that a lower bound on the value function of sensor ℓ
can be obtained as a solution to the following minimization problem over u:
m
m
J (ℓ)(eb) = min(λ(b, ℓ)T (b, ℓ); λ(b, ℓ)T1(b, ℓ) + c
[ebP ]i) +
Xi=1
Xi=1
[ebP ]iJ (ℓ)(ei)
(37)
Equation (37) together with (35) define a lower bound on the total expected cost. To further tighten
the bound we can now optimize over a matrix Λ for every value of c, where Λ(c) is an m × n matrix
with the (i, ℓ) entry equal to λ(i, ℓ), i.e., Λ(c) = {λ(i, ℓ)}. Hence
Xi=1
Xℓ=1 min(λ(b, ℓ)T (b, ℓ); λ(b, ℓ)T1(b, ℓ) + c
[ebP ]iJ (ℓ)(ei)! (38)
[ebP ]i) +
J(eb) = max
Λ(c)
Xi=1
m
m
n
subject to
Λ1n = 1m
where 1m is a column vector of all ones of length m.
The inner recursion can be solved to obtain a closed form solution for J (ℓ)(eb) as:
J ℓ(eb) =
∞
m
Xj=0
Xi=1
minn[ebP j]iλ(i, ℓ)T1(i, ℓ) + c
[ebP j+1]k ; [ebP j]iλ(i, ℓ)T (i, ℓ)o
(39)
m
Xk=1
Since the problem is only constrained across the different sensors, we obtain a lower bound from the
solution of the following optimization problem,
m
Xi=1
max
λ(i,ℓ)
n
∞
Xℓ=1
Xj=0
[ebP j]i min λ(i, ℓ)T1(i, ℓ) + c
[eiP ]k ; λ(i, ℓ)T (i, ℓ)!
m
Xk=1
(40)
21
subject to
n
We observe that for every i we are maximizing a concave piecewise linear function in λ(i, ℓ). We pose
λ(i, ℓ) = 1 ∀i = 1, . . . , m.
Xℓ=1
an equivalent convex optimization problem by realizing that the minimum of a set of concave functions
is also concave. Since affine functions are concave, we can apply the technique here. Since the problem
is unconstrained across the i dimension we focus on solving the max-min problem for a fixed i. The
final solution can then be obtained by summing the objective function for m subproblems.
For each ℓ = 1, . . . , n add a variable tℓ to the optimization problem. Also for every ℓ append 2
constraints to the optimization problem. The constraints state the minimization over uℓ implicitly, by
k=1[eiP ]k ≥ tℓ and λ(i, ℓ)T (i, ℓ) ≥ tℓ. The modified problem is
requiring that λ(i, ℓ)T1(i, ℓ) + cPm
therefore:
n
maximizeλ(i,ℓ),tℓ;ℓ=1,...,n
subject to
n
Xℓ=1
Xl=1
tℓ,
λ(i, ℓ) ≤ 1,
m
λ(i, ℓ)T1(i, ℓ) + c
Xk=1
λ(i, ℓ)T (i, ℓ) ≥ tℓ,
(41)
[eiP ]k ≥ tℓ,
ℓ = 1, . . . , n.
which can be readily solved using standard convex optimization techniques [23].
IV. RESULTS AND SIMULATIONS
In this section, we show experimental results illustrating the performance of the proposed scheduling
policies for the different models considered in this paper. In each simulation run, the object was initially
placed at the center of the network and the simulation run concluded when the object reached the
absorbing state τ . We perform Monte Carlo runs to compute the average tracking and energy costs for
different values of the energy parameter c. For the planning phase in case of point-based policies, beliefs
are sampled by simulating multiple object trajectories through the sensor network. Each trajectory starts
from a random state sampled from the initial belief, picking actions at random, until the target leaves
the network.
First, we consider the simple model in Section II-A with a linear network of 41 sensors. Figure 3(a)
shows the tradeoff curve between the number of active sensors per unit time and the tracking error
per unit time using the point-based and the QMDP policies. The figure also shows a lower bound on
22
Tradeoff between the number of awake sensors and tracking cost (simplistic model)
1
Point−based
QMDP
LB
s
r
o
r
r
e
i
g
n
k
c
a
r
T
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.5
1
1.5
2.5
2
3
Sensors awake
3.5
4
4.5
5
50
40
30
20
10
s
f
e
i
l
e
b
l
l
a
r
o
f
l
e
u
a
v
g
v
A
0
0
30
25
20
15
10
5
0
0
l
s
e
n
a
p
r
e
p
y
h
f
o
r
e
b
m
u
n
500
1000 1500 2000 2500
time
t
s
o
c
g
v
A
24
22
20
18
16
14
0
800
600
400
200
s
e
g
n
a
h
c
y
c
i
l
o
P
500
1000 1500 2000 2500
time
500
1000 1500 2000 2500
0
0
500
1000 1500 2000 2500
time
200 beliefs, 16 sampled actions
time
(a) Energy-tracking tradeoff
(b) Convergence results
Fig. 3: Simplistic model
Fig. 4: A sensor network with overlapping sensing ranges (12 sensors and 20 object locations). An edge
connects a sensor to a given location if this location falls within the sensing range of that sensor.
the optimal performance (see Section III-D). It is clear that both policies lead to tradeoffs that closely
approach the lower bound. The QMDP policy gets even closer to the lower bound at small tracking errors
since the observable-after-control assumption is more meaningful in this regime. In Fig. 3(b) we show
convergence results for the point-based algorithm with reduced control space minimization. The top left
subplot displays the convergence of the sum cost of all the belief points in P; the top right shows the
expected cost averaged over many trajectories; the bottom left subplot shows the number of hyper-planes
constituting the value function as a function of time; the bottom right subplot shows the number of
policy changes versus time, i.e., the number of belief points for which the optimal action changed over
2 consecutive iterations of the algorithm.
Figure 5 displays the tradeoff curves for the network in Fig. 4 with a probabilistic observation model.
The network is composed of 12 sensors and 20 object locations with the shown connectivity such that the
23
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
e
m
T
i
t
i
n
U
r
e
p
t
s
o
C
Q
MDP
Point−based
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
e
m
T
i
t
i
n
U
r
e
p
s
r
o
r
r
i
E
g
n
k
c
a
r
T
0
0.1
0.2
c
0.3
0.4
0.5
0
0
1
Fig. 5: Overlap model
4
Sensors Awake per Unit Time
2
3
5
observation range for the different sensors overlap. Since the tracking error for this model is inherently
coupled across sensors, the global point-based policy clearly outperforms the learning-based QMDP policy.
Next, we consider a network of 10 sensors where object locations are located on integers from 1 to
21. The observation for each sensor is continuous as in (6). For every object state and every scheduling
action in the reduced control space, we sample 50 observations to construct estimates of the weight
probabilities and compute the aggregate observation boundaries. Up to 32 actions are sampled from the
reduced control space. In this setup, the belief set consists of 500 sampled belief vectors and we assume a
Hamming error cost. Fig. 6 shows the performance of the different policies for the continuous observation
model. It is shown that the point-based scheduling policy outperforms the QMDP policy. We further show
a lower bound on the optimal performance tradeoff. The lower bound is loose especially in the high
tracking error regime since the derived bound on per-sensor tracking errors assumes all other sensors are
awake. However, we can exactly compute the saturation point for the optimal scheduling policy since
every policy has to eventually meet the all-asleep performance curve, shown in Fig. 6a, when the energy
cost per sensor is high. At that point, all sensors are inactive and hence the target estimate can only be
based on prior information.
V. CONCLUSIONS
In this paper we studied the problem of tracking an object moving randomly through a dense network
of wireless sensors. We devised approximate strategies for scheduling the sensors to optimize the tradeoff
24
e
m
i
t
t
i
n
u
r
e
p
t
s
o
c
l
t
a
o
T
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
Total cost versus c (continuous observation model)
Lower bound
Q
MDP
Point−based
All asleep
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Energy cost per sensor (c)
e
m
i
t
t
i
n
u
r
e
p
t
s
o
c
i
g
n
k
c
a
r
T
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
Tradeoff between number of awake sensors and tracking cost (continuous observation model)
Lower bound
Q
Point−based
MDP
1
2
3
4
5
6
7
8
9
10
Average number of awake sensors
(a)
(b)
Fig. 6: Continuous observation model: (a) Total cost versus energy cost per sensor, (b) Energy-tracking
tradeoff
between tracking performance and energy consumption for a wide range of models. First, we proposed
policies that rely on an observable-after-control assumption (QMDP policies). Key to this solution is
the decoupling of the optimization problem into per-sensor subproblems combined with simulation-
based learning of individual tracking costs for each subproblem. Second, we developed point-based
sensor scheduling strategies which optimize the value function over a small set of reachable beliefs
within the belief simplex. Based on the belief support and the sparsity of the transition models, we
developed a methodology to sample actions from reduced control spaces. This was combined with
observation aggregation to address the complexity of the observation space for continuous observations
models. In some cases we derived lower bounds on the optimal tradeoff curves. While being suboptimal,
the generated scheduling policies often provide close-to-optimal energy-tracking tradeoffs. Developing
distributed scheduling strategies when no central controller is available is an area for future research.
Another interesting challenge is when the statistics for object movement are unknown or partially known.
[1] J. A. Fuemmeler and V. V. Veeravalli, "Smart sleeping policies for energy efficient tracking in sensor networks," IEEE
Trans. Signal Processing, vol. 56, no. 5, pp. 2091 -- 2101, May 2008.
REFERENCES
25
[2] David A. Castanon, "Approximate dynamic programming for sensor management," in 36th conference on decision and
control (CDC), 1997, pp. 1202 -- 1207.
[3] Jason L. Williams, John W. Fisher, and Alan S. Willsky,
"Approximate dynamic programming for communication
constrained sensor network management," IEEE Transactions on Signal Processing, vol. 55, 2007.
[4] C. Kreucherm, K. Kastella, and A. Hero, "Sensor management using an active sensing approach," IEEE Transactions on
Signal Processing, vol. 85, no. 3, pp. 607 -- 624, March 2005.
[5] Juan Liu, Reich, and Feng Zhao, "Collaborative in-network processing for target tracking," Journal on Applied Signal
Processing, vol. 4, pp. 378 -- 391, 2002.
[6] Ying He and Edwin K.P. Chong, "Sensor scheduling for target tracking: A monte carlo sampling approach," Digital Signal
Processing, vol. 16, no. 5, pp. 533 -- 545, 2006, Special Issue on DASP 2005.
[7] G.E. Monahan, "A survey of partially observable markov decision processes: theory, models, and algorithms," Management
Science, vol. 28, pp. 1 -- 16, 1982.
[8] M. Hauskrecht, "Value-function approximations for partially observable markov decision processes," Journal of Artificial
Intelligence Research (JAIR), vol. 13, pp. 33 -- 94, 2000.
[9] J. Pineau, G. Gordon, and S. Thrun, "Point-based value iteration: An anytime algorithm for POMDPs," in the International
Joint Conference on Artificial Intelligence (IJCAI), 2003, pp. 1025 -- 1032.
[10] M. L. Littman, A. R. Cassandra, and L. P. Kaelbling, "Learning policies for partially observable environments: scaling
up," in Twelfth International Conference on Machine Learning, 1995, pp. 362 -- 370.
[11] M. T. J. Spaan and N. Vlassis, "Perseus: Randomized point-based value iteration for POMDPs," Journal of Artificial
Intelligence Research (JAIR), vol. 24, pp. 195 -- 220, 2005.
[12] J. M. Porta, N. Vlassis, M. T. J. Spaan, and P. Poupart, "Point-based value iteration for continuous POMDPs," Journal of
Machine Learning Research, vol. 7, pp. 2329 -- 2367, 2006.
[13] Dimitri P. Bertsekas, Dynamic programming and optimal control, Athena Scientific, 2001.
[14] E. J. Sondik, The Optimal Control of Partially Observable Markov Processes, Ph.D. thesis, Stanford University, 1971.
[15] H. T. Cheng, Algorithms for partially observable Markov decision processes, Ph.D. thesis, University of British Columbia,
1988.
[16] L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, "Planning and acting in partially observable stochastic domains,"
Artificial Intelligence, vol. 101, pp. 99 -- 134, 1998.
[17] A. Cassandra, M. L. Littman, and N. L. Zhang,
"Incremental pruning: A simple, fast, exact algorithm for partially
observable markov decision processes," in Thirteenth Annual Conference on Uncertainty in Artificial Intelligence. 1997,
pp. 54 -- 61, Morgan Kaufmann.
[18] L. P. Kaelbling, M. L. Littman, and A. Moore, "Reinforcement learning: A survey," Journal of Artificial Intelligence
Research, vol. 4, pp. 237 -- 285, 1996.
[19] N. Roy and G. Gordon, "Exponential family PCA for belief compression in POMDPs," In Advances in Neural Information
Processing Systems, vol. 15, 1995.
[20] Jesse Hoey and Pascal Poupart, "Solving pomdps with continuous or large discrete observation spaces," in IJCAI'05:
Proceedings of the 19th international joint conference on Artificial intelligence, San Francisco, CA, USA, 2005, pp.
1332 -- 1338, Morgan Kaufmann Publishers Inc.
[21] H. Vincent Poor, An Introduction to Signal Detection and Estimation (2nd ed.), Springer-Verlag New York, Inc., New
York, NY, USA, 1994.
26
[22] Bernard C. Levy, Principles of Signal Detection and Parameter Estimation, Springer Publishing Company, Incorporated,
2008.
[23] Stephen Boyd and Lieven Vandenberghe, Convex Optimization, Cambridge University Press, New York, NY, USA, 2004.
27
|
1205.0561 | 7 | 1205 | 2013-11-15T14:44:00 | Multi-level agent-based modeling - A literature survey | [
"cs.MA"
] | During last decade, multi-level agent-based modeling has received significant and dramatically increasing interest. In this article we present a comprehensive and structured review of literature on the subject. We present the main theoretical contributions and application domains of this concept, with an emphasis on social, flow, biological and biomedical models. | cs.MA | cs |
Multi-level agent-based modeling
A literature survey
Gildas Morvan
http://www.lgi2a.univ- artois.fr/~morvan/
gildas.morvan@univ- artois.fr
Univ Lille Nord de France, F-59000 Lille, France
UArtois, LGI2A, F-62400, Béthune, France
January 7, 2014
1
Absract
During last decade, multi-level agent-based modeling has received significant and dra-
matically increasing interest. In this article we conduct a comprehensive and struc-
tured review of literature on this emerging research domain that aims at extending
the classical agent-based modeling paradigm to overcome some of its limitations. We
present the main theoretical contributions and applications with an emphasis on
social, flow, biological and biomedical models.
Main acronyms used in this paper
ABM According to the context, Agent-Based Model or Agent-Based Modeling
DEVS Discrete Event System Specification
EBM Equation-Based Model
ML -ABM According to the context, Multi-Level Agent-Based Model or Multi-
Level Agent-Based Modeling
Contents
Introduction
1
. . . . . . . . . . . . . . . . . . . . . . . . .
1.1 Agent-based modeling
1.2 Multi-level agent-based modeling . . . . . . . . . . . . . . . . . . . .
1.3 Terminology issues . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Theoretical issues
2.1 Meta-models, simulation engines and platforms . . . . . . . . . . . .
2.2 Observation, detection and agentification of emergent phenomena
.
2.3 Representation of aggregated entities . . . . . . . . . . . . . . . . . .
3 Application domains
3.1
Social simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Flow modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 Biological and biomedical models
. . . . . . . . . . . . . . . . . . .
3.4 Ecology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Discussion
4.1 Level integration . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Multi-level technical tools
. . . . . . . . . . . . . . . . . . . . . . .
5 Conclusion
References
3
3
3
4
5
6
6
8
9
10
11
12
14
16
18
18
19
20
20
2
1
Introduction
1.1 Agent-based modeling
Agent-based modeling (ABM) is a computational modeling paradigm that allow to
simulate the interactions of autonomous agents in an environment.
It has been
widely used to study complex systems in various domains (Epstein, 2006; Ferber,
1999; Gilbert, 2007; Railsback and Grimm, 2011; Resnick, 1994; Treuil et al., 2008).
However, it suffers from important known limitations that reduce its scope (Dro-
goul et al., 2003; Scerri et al., 2010). First, ABM is purely bottom-up: a microscopic
knowledge, i.e., related to system components, is used to construct models while a
macroscopic knowledge, i.e., related to global system properties, is used to validate
models.Therefore, it is not straightforward to explicitly introduce bidirectional rela-
tions between these two points of view in the general case. It becomes even harder
when different spatial or temporal scales or domains of interest are involved in a same
simulation. Moreover, agent-based models do not scale easily and generally require
large computational resources since many agents are simulated. Finally, most agent-
based simulation platforms lack tools to reify complex singular emergent properties:
human observation often remains the most efficient way to capture multi-level pat-
tern formation or crowd behavior.
1.2 Multi-level agent-based modeling
1.2.1 Definitions
The works surveyed in this paper aim at extending the classical ABM paradigm to
overcome these limitations. While they can be very different (in terms of goals,
technical approaches or application domains), these woks share a common idea: in-
troducing more levels of description in agent-based models. Therefore, we group
them under a common term: Multi-Level Agent-based Modeling (ML -ABM). In the
following we propose definitions of the concepts of level and multi-level agent-based
model that seem broad enough to encompass the different surveyed approaches (see
section 1.3).
Definition 1 A level is a point of view on a system, integrated in a model as a specific
abstraction.
The name refers to the familiar expressions levels of organization, observation,
analysis, etc.
Definition 2 A multi-level agent-based model integrates heterogenous (agent-based)
models, representing complementary points of view, so called levels, of the same sys-
tem.
In this definition, three concepts are highlighted and should themselves be de-
fined formally.
Definition 3 Integration means that the ABMs within a ML-ABM can interact and
share entities such as environments and agents.
Definition 4 Heterogeneity means that the ABMs integrated in a ML-ABM can be based
on different modeling paradigms (differential equations, cellular automata, etc.), use dif-
ferent time representation (discrete events, step-wise) and represent processes at different
spatio-temporal scales.
3
Definition 5 Points of view are complementary for a given problem since they can
not be taken in isolation to address it.
This idea is very important in the literature on complex systems (Morin, 1992).
Indeed, as Müller and Aubert (2011) note, "the global behavior of a complex system
cannot be understood without letting a set of points of view interact".
1.2.2 Types of problems solved using multi-level agent-based approaches
ML -ABM is mainly used to solve three types of modeling problems:
• the modeling of cross-level interactions, e.g., an explicit top-down feedback
control,
• the coupling of heterogeneous models,
• the (dynamic) adaptation of the level of detail of simulations, e.g., to save
computational resources or use the best available model in a given context.
In the first case, the different points of view always co-exist, as they integrate interde-
pendent models, while in the last ones, levels are (de)activated at run-time according
to the context, as they represent independent models designed for specific situations.
For instance, in flow hybrid models areas with simple topologies are handled
with an equation-based model (EBM) while others are handled with an ABM.
1.3 Terminology issues
Different terms have been used to describe what we call here a level: e.g., perspective,
interaction, layer or view (Desmeulles et al., 2009; Parunak et al., 2009; Torii et al.,
2005; White, 2007). Some are domain-specific; thus, in the flow modeling domain,
when two levels with static relations are considered, models are often described as
hybrid as shown in the section 3.2 (Burghout et al., 2005; El hmam et al., 2008,
2006a,b,c; Marino et al., 2011; Mathieu et al., 2007a; Nguyen et al., 2012b; Rejniak
and Anderson, 2011; Wakeland et al., 2007).
The term multi-scale — or multi-resolution (Jeschke and Uhrmacher, 2008; Zhang
et al., 2009a, 2011) — is often used but has a more restrictive meaning as it focuses
on the spatial and temporal extents of levels and not on their interactions and orga-
nization. Gil-Quijano et al. (2012, p. 622–623) pointed that the term multi-scale can
be misleading and advocated for using multi-level instead. Let take as an example the
Simpop3 model, described by their authors as multi-scale (Pumain and Louail, 2009).
Two levels are considered: the city level, representing the internal dynamics of a city,
and the system of cities level, representing the interactions between cities. However,
the idea of scale does not fit to describe the relation between them: one can easily
figure a city bigger (in terms of population, spatial extent, economic exchanges, etc.)
than a system of cities. In contrast, the idea of levels of organization in interaction
seems more appropriate. Furthermore, to extend the definitions 1 and 2 such mod-
els could be more precisely denoted as nested or hierarchical multi-level agent-based
models.
4
1.4 Bibliography
During last decade, ML -ABM has received significant and dramatically increasing
interest (fig. 1). In this article we present a comprehensive and structured review of
literature on the subject1 .
Another survey on the subject has been previously conducted by Gil-Quijano
et al. and published in different versions2 . While the present article aims at providing
an overview of the literature, Gil-Quijano et al. performed a comparative study of
three models (Gil-Quijano et al., 2008; Lepagnot and Hutzler, 2009) and (Pumain
and Louail, 2009). A similar survey, comparing four models, can be found in Vo
(2012, p. 28–34).
s
n
o
i
t
a
c
i
l
b
u
p
per year
cumulated
160
140
120
100
80
60
40
20
0
1996 2000 2004 2008 2012
year
s
n
o
i
t
a
t
i
c
per year
cumulated
1200
1000
800
600
400
200
0
1996 2000 2004 2008 2012
year
(a) publications
(b) citations
10
2006
2012
2011
.
.
.
8
6
4
s
n
o
i
t
a
c
i
l
b
u
p
/
s
n
o
i
t
a
t
i
c
s
n
o
i
t
a
t
i
c
(cid:80)
1200
1000
800
600
400
200
(cid:80) publications
0
40
80
120
160
(d) (cid:80) citations vs. (cid:80) publications
Figure 1: Bibliographical statistics on ML -ABM computed from author’s biblio-
graphic database and google scholar data on the Nov. 15, 2013
(c) ratio of citations to publications
2008
year
2010
2012
The paper is structured as follows. Section 2 introduces the main theoretical
contributions and section 3 presents the different application domains of ML -ABM,
with an emphasis on social, flow, biological and biomedical models. In section 4,
some issues about ML -ABM are discussed. Finally, the remaining sections conclude
1The bibliographic database is available at the following URL: http://www.lgi2a.univ- artois.
fr/~morvan/Gildas_Morvan/ML- ABM_files/mlbib.bib.
2This work was first published in french (Gil-Quijano et al., 2009), then extended (Gil-Quijano et al.,
2010),(Louail, 2010, p. 185–204) and finally translated into english (Gil-Quijano et al., 2012)
5
this paper by an analysis of the benefits, drawbacks and current limitations of the
existing approaches.
2 Theoretical issues
In the surveyed literature, three main theoretical issues have been addressed so far:
• the definition and implementation of meta-models and simulation engines,
• the detection and reification of emergent phenomena,
• and the definition of generic representations for aggregated entities.
They are described in the following sections.
2.1 Meta-models, simulation engines and platforms
Many meta-models and simulation engines dedicated to ML -ABM have been pro-
posed in the literature. They are are briefly presented in the following, in a chrono-
logical order.
Approaches based on DEVS have also been included (Zeigler et al., 2000). Indeed,
DEVS, as a generic event-based simulation framework, has been extended to support
ABM (Duboz, 2004; Duboz et al., 2003; Müller, 2009). A comprehensive survey of
the literature on multi-level DEVS extensions can be found in Duhail (2013).
GEAMAS (Marcenac et al., 1998a,b; Marcenac and Giroux, 1998) (GEneric Ar-
chitecture for MultiAgent Simulation) is a pioneering ML -ABM framework inte-
grating three levels of description (micro, meso, macro). Micro and macro levels
represent respectively agent and system points of view while the meso (or middle)
level represents an aggregation of agents in a specific context. Communication be-
tween levels is asynchronous. GEAMAS-NG is a newer version of the framework
providing tools to detect and reify emergent phenomena (David et al., 2011).
tMans3 (Scheutz et al., 2005) is a multi-scale agent-based meta-model and plat-
form. Unfortunately, the project seems to have died in the bud.
ML -DEVS (Steiniger et al., 2012; Uhrmacher et al., 2007) is an extension of
DEVS that allows the simulation of multi-scale models (and not only coupled models
in which the behavior of a model is determined by the behaviors of its sub-models).
Two types of relation between levels are defined: information propagation and event
activation. However, ML -DEVS focuses on multi-scale modeling and therefore, only
supports pure hierarchies of models: interaction graphs are viewed as trees (Maus et al.,
2008).
CRIO (Gaud, 2007; Gaud et al., 2008a,b) (Capacity Role Interaction Organiza-
tion) is an organizational meta-model dedicated to ML -ABM based on the concept
of holon (Koestler, 1967, 1978). It has been used to develop multi-scale simulations
of pedestrian flows (cf. section 3.2).
SPARK4 (Solovyev et al., 2010) (Simple Platform for Agent-based Representation
of Knowledge) is a framework for multi-scale ABM, dedicated to biomedical research.
Scerri et al. proposed a technical architecture to distribute the simulation of
complex agent-based models (Scerri et al., 2010). This approach aims at integrating
3 http://tmans.sourceforge.net/
4 http://www.pitt.edu/~cirm/spark/
6
multiple ABMs in a single simulation, each ABM representing a specific aspect of the
problem. In their article, authors focus on the management of time in such simula-
tions. Therefore, the proposed platform provides two main technical services that
ensure the consistency of simulations: (1) a time manager that ensures that integrated
ABMs advance in time in a consistent way and (2) a conflict resolver that manages the
problematic interactions between agents and shared data (such as the environment).
Authors evaluate their approach on a modified version of Repast5 and show it can
scale large-scale models easily.
IRM4MLS6 (Morvan and Jolly, 2012; Morvan et al., 2011) (Influence Reaction
Model for Multi-Level Simulation) is a multi-level extension of IRM4S (Influence
Reaction Model for Simulation) (Michel, 2007), an ABM meta-model based on the
Influence Reaction model which views action as a two step process: (1) agents pro-
duce "influences", i.e., individual decisions, according to their internal state and per-
ceptions, (2) the system "reacts", i.e., computes the consequences of influences, ac-
cording to the state of the world (Ferber and Müller, 1996). The relations of per-
ception and influence between levels are specified with digraphs. IRM4MLS relies
on a generic vision of multi-level modeling (see section 1). Therefore, interactions
between levels are not constrained. It has been applied to simulate and control intel-
ligent transportation systems composed of autonomous intelligent vehicles (Morvan
et al., 2012, 2009; Soyez et al., 2013, 2011) (see section 3.2).
ML -Rules (Maus et al., 2011) is a rule-based multi-scale modeling language ded-
icated to cell biological systems. Rules, describing system dynamics, are described
in a similar way as in chemical reaction equations. ML -Rules has been implemented
within the simulation framework JAMES II7 . This approach does not refer explicitly
to ABM; however, multi-level rule-based languages seem a promising way to engineer
complex individual-based models.
Müller et al. developed an approach that consists in decomposing a problem
according to the complementary points of view involved in the modeling (Müller
and Aubert, 2011; Müller and Diallo, 2012; Müller et al., 2011). For instance in
their case study the problem is the relation between residential and scholar segrega-
tion. Three points of view are considered: the geographer, the sociologist and the
economist. Then, independent conceptual agent-based models are defined for each
point of view. As models share agents and concepts, the conceptual models can-
not be merged without some processing. Indeed, a same concept can have different
meanings according to the point of view. To solve this issue, Müller et al. adapt
a technique described in the modular ontology literature: defining bridge rules that
explicit the relations between concepts.
PADAWAN (Picault and Mathieu, 2011) (Pattern for Accurate Design of Agent
Worlds in Agent Nests) is a multi-scale ABM meta-model based on a compact ma-
tricial representation of interactions, leading to a simple and elegant simulation
framework. This representation is based on the meta-model of IODA (Interaction-
Oriented Design of Agent simulations) dedicated to classical (1-level) ABM (Kubera
et al., 2008).
GAMA8 (Drogoul et al., 2013; Taillandier et al., 2010, 2012) is an ABM plat-
form with a dedicated modeling language, GAML, that offers multi-level capabili-
ties. Moreover, it includes a framework (a set of predefined GAML commands) to
5 http://repast.sourceforge.net
6 http://www.lgi2a.univ- artois.fr/~morvan/Gildas_Morvan/IRM4MLS.html
7 http://www.jamesii.org
8 http://code.google.com/p/gama- platform/
7
Figure 2: The Seck and Honig (2012) approach
It is certainly the most advanced
agentify emerging structures (Vo et al., 2012b).
platform, from an end-user point of view, that integrates a multi-level approach. The
multi-scale meta-model focuses on the notion of situated agent and therefore, top class
abstractions include geometry and topology of simulated entities (Vo et al., 2012a).
The notion of level does not appear explicitly but the concept of species defines at-
tributes and behaviors of a class of same type agents and the multi-scale structure of
the model, i.e., how species can be nested within each other.
Seck and Honig developed an extension of DEVS that allows the simulation of
multi-level (i.e., non hierarchically coupled) models (Seck and Honig, 2012). The
coupling between levels is done through regular DEVS models, named bridge mod-
els (fig. 2).
AA4MM (Camus et al., 2013, 2012; Siebert, 2011; Siebert et al., 2010) (Agent and
Artifact for Multi-Modeling) is a multi-modeling (or model coupling) meta-model
applied to ML -ABM. Levels are reified by agents that interact trough artifacts. This
meta-model extends existing ones, see e.g., Bonneaud (2008); Bonneaud et al. (2007),
distributing the scheduling between levels.
2.2 Observation, detection and agentification of emergent phe-
nomena
An important issue concerning ML -ABM is to observe, detect and possibly reify (or
more precisely agentity) phenomena emerging from agent interactions. Of course,
the question is not to detect any emergent phenomenon but those of interest, in
order to e.g., adapt the level of detail of simulations, model cross-level interactions or
observe multi-level behaviors.
8
Multi-perspectivemodellingofcomplexphenomenaFig.1ModellingsimplephenomenaFig.2ModellingcomplexphenomenathroughmultipleperspectivesNowifthephenomenawetrytomodelarecomplex,areductionistformalsys-temcanonlybepartiallysuccessfulindescribingthenaturalsystem(Agazzi1991;Mikulecky2001).Bydescribinganaturalsystemasacollectionofperspectives,though,whereeachperspectiveisassociatedwithauniqueformalsystem(havingauniquedecomposition)asshowninFig.2,wecanmodelasysteminaninher-ently‘richer’waybyhavingmultiplenon-isomorphicdecompositionsthatmayin-fluenceeachother.Suchmulti-perspectivemodelscanindeedcapturethetanglednessofthesystemsthatresultwhenweobservetheworldfromdifferentperspectives.AsMorinputsit(Morin1990),“wemustfoundtheideaofacomplexsystemonanon-hierarchicalconceptofthewhole”(Morin1990).Inasimilarway,Levins(2006)pro-posestherobustnessmethodology,which,inasortoftriangulation,invitestoanalyzeandmodelsystemswithmultipleconceptuallyindependenttools,thusimprovingac-curacyofthemodelsbyrelatingtheoutcomesobtainedfromdifferentperspectives.Therelationbetweencomplexityandmultipleperspectiveshasbeenacknowl-edgedbyvariousauthors.Kaufmannhasstatedthatthenumberofpossibletheo- Very different approaches have been proposed to solve this problem. The first
ones were of course exploratory. Therefore, they rely on dedicated methods related
to specific models. Newer works focus on generic methodologies and frameworks.
They are briefly presented in a chronological order.
2.2.1 Dedicated clustering methods
The pioneering RIVAGE project (Servat, 2000; Servat et al., 1998a,b) aimed "at mod-
eling runoff, erosion and infiltration on heterogeneous soil surfaces" (Servat et al.,
1998a, p. 184). At the microscopic level, water is viewed as a set of interacting water-
balls. An indicator characterizes waterball movements to detect two types of remark-
able situations: straight trajectories (corresponding to the formation of ravines) and
stationary particles (corresponding to the formation of ponds). Close agents sharing
such properties are aggregated in ravine or pond macroscopic agents.
Bertelle et al. (2002); Tranouez (2005); Tranouez et al. (2001, 2006); Tranouez
and Dutot (2009); Tranouez et al. (2003) aimed at changing the level of detail of
fluid flow simulations using the vortex method (Leonard, 1980). The goal was, as
in the RIVAGE model, to detect complex structures, i.e., clusters of particles shar-
ing common properties, and aggregate them. However, the detection of emergent
phenomena relies on graph-based clustering methods. Moncion et al. (2010) used a
similar approach to detect aggregations of agents in flocking simulations.
Gil-Quijano (2007); Gil-Quijano and Piron (2007) and Gil-Quijano et al. (2008)
used various clustering algorithms such as self-organizing maps, K-Means and particle
swarm algorithms, to detect group formations.
2.2.2 Generic frameworks
Chen (2009, 2013); Chen et al. (2010); Chen and Hardoon (2010); Chen et al. (2008a,b,
2009) proposed a formalism, named complex event types to describe multi-level behav-
iors in ABMs. "Conceptually, complex events are a configuration of simple events
where each component event can be located in a region or point in a hyperspace
that includes time, physical space and any other dimensions" (Chen et al., 2008b, p.
4). Using this approach, it is then possible to formally define and observe complex
phenomena at different levels.
David and Courdier (2008, 2009) developed a conceptual and technical frame-
work to handle emergence reification. It is implemented in the GEAMAS-NG plat-
form (cf. GEAMAS paragraph in the previous section) and has been used in a popu-
lation model of the Reunion Island, to detect and reify new urban areas (David et al.,
2012, 2011).
A similar framework has been integrated in the GAMA platform (see GAMA
paragraph in the previous section). It includes various clustering methods developed
in the literature (Vo et al., 2012b).
SimAnalyzer is a general-purpose tool, to detect and describe group dynamics in
simulations (Caillou and Gil-Quijano, 2012; Caillou et al., 2012).
2.3 Representation of aggregated entities
While developing generic representations for aggregated entities (also called group
abstractions) seems an important issue, to the best of our knowledge a few publica-
tions are available on the subject. The common idea in these works is to delegate
9
the computation of agent behavioral functions to less detailed agents, representing
groups of aggregated agents, in other levels in order to reduce the complexity of
interaction computing.
Sharpanskykh and Treur (2011a,b) proposed two approaches to group abstrac-
tion dedicated to models where agent state are represented by variables taking values
in {0, 1} or [0, 1]:
• weighting averaging: an aggregated state of a group is estimated by averaging
agent states, with a weighting factor related to the strength of influence of an
agent in the group (the stronger, the more important).
• invariant-based abstraction: this approach consists in determining an invariant
in a group of agents, i.e., a property that does not change in time and using it
as a conservation law.
Sharpanskykh and Treur applied these methods to a collective decision making model
of social diffusion. They performed a comparative study of the methods, focusing
on computational efficiency and approximation error.
Parunak (2012) introduced the notion of pheromone field (refering to the concept
of mean field in statistical physics) that "gives the probability of encountering an
agent of the type represented by the field at a given location" (Parunak, 2012, p.
115). In this approach, agents act according to their perceptions of pheromone fields
(but not of agents).
Navarro et al. (2012) proposed a generic approach based on the notion of meso-
scopic representation: agents sharing common properties (related to their physical
or mental states) delegate the computation of behavioral functions to a mesoscopic
agent. Authors developed this approach to reduce the computational cost of simula-
tions while guaranteeing accurate results.
3 Application domains
ML -ABM has been used in various fields such as
• biomedical research
– cancer modeling (Andasari et al., 2012; Brown et al., 2011; Deisboeck and
Stamatakos, 2010; Lepagnot and Hutzler, 2009; Olsen and Siegelmann,
2013; Paiva et al., 2009; Rejniak and Anderson, 2011; Sun et al., 2012;
Wang et al., 2013, 2008; Wang and Deisboeck, 2008; Wang et al., 2007;
Zhang et al., 2007, 2009a, 2011, 2009b),
– inflammation modeling (An, 2008; An et al., 2013; An and Wilensky,
2009; Kim et al., 2012; Scheff et al., 2012; Vodovotz et al., 2008; Wakeland
et al., 2007),
– arterial adaptation (Hayenga et al., 2011; Thorne et al., 2011),
– stent design (Tahir et al., 2011),
– vascular tissue engineering (Zahedmanesh and Lally, 2012),
– bone remodeling (Cacciagrano et al., 2010),
10
• flow modeling of walking (and running) (Gaud et al., 2008a; Navarro et al.,
2011; Nguyen et al., 2011, 2012b; Xi and Son, 2012), driving (Bourrel, 2003;
Bourrel and Henn., 2002; Bourrel and Lesort, 2003; Burghout et al., 2005;
El hmam et al., 2008, 2006a,b,c; Espié et al., 2006; Magne et al., 2000; Mam-
mar and Haj-Salem, 2006; Morvan et al., 2012, 2009; Poschinger et al., 2002;
Sewall et al., 2011; Soyez et al., 2013, 2011; Wedde and Senge, 2012) or stream-
ing (Servat et al., 1998a,b; Tranouez et al., 2006) agents,
• biology (Adra et al., 2010; Biggs and Papin, 2013; Christley et al., 2007a,b;
Jeschke and Uhrmacher, 2008; Marino et al., 2011; Montagna et al., 2010a,b;
Seal et al., 2011; Shimoni et al., 2011; Smallwood et al., 2010; Smallwood and
Holcombe, 2006; Stiegelmeyer et al., 2013; Sun et al., 2009),
• social simulation (Conte et al., 2007; Conte and Castelfranchi, 1996; Das-
calu et al., 2009, 2011; Gil-Quijano et al., 2008; Hassoumi et al., 2012; Laper-
rière, 2012; Louail, 2010; North et al., 2010; Ozik et al., 2008; Parry and
Bithell, 2012; Pumain and Louail, 2009; Sawyer, 2001, 2003; Schaller et al.,
2012; Schillo et al., 2001; Seck and Honig, 2012; Squazzoni, 2008),
• ecology (Belem, 2009; Belem and Müller, 2009; Belem and Müller, 2013; Cheong
et al., 2012; Duboz, 2004; Duboz et al., 2003; Le et al., 2011; Marilleau et al.,
2008; Morvan et al., 2008, 2009; Prévost et al., 2004; Ratzé et al., 2007; Roun-
sevell et al., 2012; Schmidt et al., 2011; Seidl et al., 2012, 2010; Semeniuk et al.,
2011; Vincenot et al., 2011),
• military simulation (Mathieu et al., 2007a,b; Parunak et al., 2009).
An interesting comparative analysis of three of these models can be found in Gil-
Quijano et al. (2009, 2010); Gil-Quijano et al. (2012) and Louail (2010, p. 185–204).
3.1 Social simulation
Social simulation is defined by Squazzoni (2008, p. 4) as "the study of social out-
comes, let us say a macro regularity, by means of computer simulation where agents’
behavior, interactions among agents and the environment are explicitly modeled to
explore those micro-based assumptions that explain the macro regularity of interest".
Major social theories developed in the second half of the twentieth century, e.g.,
structuration (Giddens, 1987) and habitus (Bourdieu, 1994) theories9 , share a common
ambition: solving the micro/macro (so called agency/structure) problem that can
be summarized by the following question: To understand social systems, should we
observe agent interactions (micro level) or structures emerging from these interactions
(macro level)? Such theories tend to consider altogether agent positions in the social
space (objective facts) and goals (subjective facts) to explain their beliefs and actions.
Their answer to the previous question could be: social systems can only be understood
by considering simultaneously agent interactions and structures in which they occur :
social structures
social practices
agent interactions.
9These theories are described by some sociologists as hybrid (Sawyer, 2001).
11
A key concept used by social theorists and modelers to understand downward (or
top-down) causation in social systems, i.e., how social structures influence agents, is
reflexivity. It can be defined as the "regular exercise of the mental ability, shared by all
normal people, to consider themselves in relation to their (social) contexts and vice
versa" (Archer, 2007, p. 4). Thus, social systems differ from other types of systems,
by the reflexive control that agents have on their actions: "The reflexive capacities of
the human actor are characteristically involved in a continuous manner with the flow
of day-to-day conduct in the contexts of social activity" (Giddens, 1987, p. 22). Two
very different approaches, both from technical and methodological perspectives, can
be considered to simulate systems composed of reflexive agents:
• a purely emergentist approach, only based on the cognitive capabilities of
agents to represent and consider themselves in relation to the structures emerg-
ing from their interactions — e.g., Conte and Paolucci (2002); Gilbert (2002),
• a multi-level approach based on the cognitive capabilities of agents and the
dynamic reification of interactions between social structures and agents, i.e.,
processes that underlie social practices — e.g., Gil-Quijano et al. (2009); Pumain
et al. (2009).
According to Giddens (1987), two forms of reflexivity can be distinguished: practical
(agents are not conscious of their reflexive capabilities, and therefore, are not able to
resonate about them) and discursive (agents are conscious of their reflexive capabili-
ties) reflexivity. These two forms are respectively related to the ideas of immergence
in which agent interactions produce emergent properties that modify the way they
produce them (Conte et al., 2007) and second order emergence in which agent interac-
tions produce emergent properties that are recognized (incorporated) by agents and
influence their actions (Gilbert, 2002).
ML -ABM can also be viewed as a way to link independent social theories (and
therefore concepts) defined at different levels (fig. 3) (Sawyer, 2001, 2003; Seck and
Honig, 2012). Thus, Seck and Honig (2012) proposed a model of social conflicts
integrating agent behaviors and social laws. Gil-Quijano (2007); Gil-Quijano et al.
(2007, 2008) developed a multi-scale model of intra-urban mobility. The microscopic
level represents households and housing-units, the mesoscopic one, groups of micro-
agents and urban-sectors and the macroscopic one, the city itself. A clustering algo-
rithm is applied to detect and reify groups of households and housing-units. Pumain
and Louail (2009) developed Simpop3, a multi-scale model based on two previously
developed single-scale models: Simpop nano, simulating the internal dynamics of a
city and Simpop2, simulating city interactions.
Readers interested in a more comprehensive presentation of these questions may
refer to Sawyer (2003); Schillo et al. (2001); Squazzoni (2008) and Raub et al. (2011) .
3.2 Flow modeling
A flow of moving agents can be observed at different scales. Thus, in traffic modeling,
three levels are generally considered: the micro, meso and macro levels, representing
respectively the interactions between vehicles, groups of vehicles sharing common
properties (such as a common destination or a common localisation) and flows of
vehicles. Each approach is useful in a given context: micro and meso models allow
to simulate road networks with complex topologies such as urban area, while macro
12
macrosociology
large-scale social structures
mesosociology
community
organization
microsociology
agent
Figure 3: ML -ABM in social simulation as a link between concepts defined at differ-
ent levels
models allow to develop control strategies to prevent congestion in highways. How-
ever, to simulate large-scale road networks, it can be interesting to integrate these
different representations (fig. 4). The main problem is to determine an appropriate
coupling between the different representations, i.e., that preserves the consistency of
simulations (Davis and Hillestad, 1993).
simulated entities
control strategies
macro
flow of vehicles
ramp metering
meso
group of vehicles
variable-message panels
micro
vehicle
vehicle instrumentation
Figure 4: ML -ABM and control in traffic simulation
3.2.1 Micro-macro models
An interesting comparison of existing micro-macro traffic models can be found in El
hmam (2006, p. 31–44) (table 1). The coupling methods rely on the following idea:
creating a virtual section which is both macro and micro at level connections. This
virtual section is used as a buffer to generate appropriate micro or macro data ac-
cording to the type of connection. On this basis, El hmam (2006); El hmam et al.
(2008, 2006a,b,c) proposed a generic coupling method between agent-based micro-
scopic models and widely used macroscopic models such as LWR, ARZ and Payne.
While micro-macro flow models were essentially developed in the traffic domain,
other applications such as crowd simulation emerged in recent years (Nguyen et al.,
2011, 2012b).
However, all the surveyed hybrid models share the same limitation: connections
between levels are fixed a priori and cannot be changed at runtime. Therefore, to
13
be able to observe some emerging phenomena such as congestion formation or to
find the exact location of a jam in a large macro section, a dynamic hybrid modeling
approach is needed (Sewall et al., 2011).
model
Magne et al. (2000)
Poschinger et al. (2002)
Bourrel and Lesort (2003)
Mammar and Haj-Salem (2006)
Espié et al. (2006)
El hmam (2006)
Sewall et al. (2011)
micro model
SITRA-B+
IDM
optimal velocity
ARCHISM
generic ABM
macro model
SIMRES
Payne
LWR
ARZ
SSMT
LWR, ARZ, Payne
ARZ
Table 1: Main micro-macro traffic flow models, adapted from El hmam (2006, p. 42)
3.2.2 Micro-meso models
This kind of models is often used to reduce the complexity of agent interactions.
Agents sharing common properties can be aggregated to form up a higher level
(mesoscopic) agent and then, save computer resources or describe group dynamics
such as in the already mentioned RIVAGE (Servat, 2000; Servat et al., 1998a,b) and
DS (David et al., 2011) models (cf. section 2.2). Conversely, mesoscopic agents can
be broken up into lower level agents if related structures vanish.
Morvan et al. (2012, 2009) introduced an multi-level approach to solve the dead-
lock problem in field-driven autonomous intelligent vehicle systems. These sys-
tems generally rely on self-organization to achieve their goals, but AIVs can remain
trapped into dead-locks. When such a situation is detected (using a similar approach
than Servat (2000); Servat et al. (1998a,b)), it is agentified to solve the problem using
hierarchical control.
Flacher et al. (2012); Navarro et al. (2012, 2013, 2011) proposed an innovative
framework for such models: (de)aggregation functions rely not only on the observ-
able state of simulations (the environment) but also on the internal state of agents. It
has been applied to pedestrian flow simulation. The proximity between agent states
(external and internal) is computed by an affinity function.
Soyez et al. (2013, 2011) extended this framework on the basis of IRM4MLS.
Agents are "cut" into a a set of physical parts (bodies), situated in different levels,
and a non-situated part (mind) (see fig. 5). Therefore, these different parts can be
(de)aggregated independently. This approach has been applied to dynamically adapt
the level of detail in a port operations simulator.
3.3 Biological and biomedical models
A biological system can be considered at different levels of organization:
... → molecule → cell → tissue → organ → ... ,
that basically correspond to the segmentation of biological research into specialized
communities:
... → molecular biology → cell biology → histology → physiology → ... .
14
conceptualAgent
environment
1
spiritAgent
1
1
1
1..n
1..n
0..n
bodyAgent
1
1
0..n
0..n
1
level
Figure 5: Mind/bodies separation in the Soyez et al. (2013) model
Each research area has developed its own ontologies and models to describe the
same reality observed at different levels. However, this reductionist approach fails
when addressing complex issues (Schnell et al., 2007). Thus, it has been shown that
living systems are co-produced by processes at different levels of organization (Mat-
urana and Varela, 1980). Therefore, an explanatory model of such systems should
account for the interactions between levels.
3.3.1 Cell biology
At least two levels are explicitly represented in cell biology models: the macroscopic
one, representing the extracellular environment and the interactions between cells,
and the microscopic one, representing the intracellular environment and the inter-
actions between cell components such as signaling pathways and gene regulatory
networks. A major modeling issue is that these two levels continuously influence
each other. It leaded to the development of multi-scale models.
Montagna et al. (2010a,b) developed a model of morphogenesis in biological sys-
tems, in particular for the Drosophila Melanogaster species. Maus et al. (2011) pro-
posed a model of Schizosaccharomyces pombe (a species of yeast) cell division and
mating type switching based on the ML -Rules approach (cf. section 2.1).
3.3.2 Cancer modeling
Cancer is a complex spatialized multi-scale process, starting from genetic mutations
and potentially leading to metastasis. Moreover, it has multi-scale (from from genetic
to environmental) causes. Therefore, it can be studied from various perspectives
from the intracellular (molecular) to the population levels.
ML -ABM is a promising paradigm to model cancer development (Wang and Deis-
boeck, 2008). Indeed, as Schnell et al. (2007, p. 140) note, "a multi-scale model would
allow us to explore the effect of combination therapies, approaches that attempt to
stop cancer in its tracks by barricading multiple pathways. Most present models,
focusing on processes at a single scale, cannot provide this comprehensive view."
Zhang et al. developed an ML -ABM of a brain tumor named Glioblastoma Multi-
forme (GBM) (Zhang et al., 2007, 2009a, 2011, 2009b). This model explicitly defines
the relations between scales and uses different modeling approaches: ordinary dif-
ferential equations (ODE) at the intracellular level, discrete rules typically found in
ABM at the cellular level and partial differential equations (PDE) at the tissue level
(see fig. 6).
Moreover, this model also relies on a multi-resolution approach: heterogenous
clusters, i.e., composed of migrating and proliferating cells are simulated at a high
15
resolution while homogenous clusters of dead cells are simulated at a lower resolu-
tion. In short, "more computational resource is allocated to heterogenous regions of
the cancer and less to homogenous regions" (Zhang et al., 2011, p. 6). This model
has been implemented on graphics processing units (GPU), leading to an efficient
parallel simulator (Zhang et al., 2011).
tissue
cellular
tissue evolution (PDE)
diffusion of the chemoattractants
cell — cell interactions (discrete rules)
cell’s phenotype
intracellular
gene — protein interactions (ODE)
r
o
t
p
e
c
e
r
y
a
w
h
t
a
p
s
’
l
l
e
c
Figure 6: ML -ABM in brain tumor modeling (Zhang et al., 2007, 2009a, 2011, 2009b)
Sun et al. (2012) also developed a brain tumor ML -ABM available as a MATLAB
library called ABM-TKI10 . It is based on a 4 level architecture (tissue, microenviron-
mental, cellular, modelcular).
Lepagnot and Hutzler (2009) model the growth of avascular tumors to study the
impact of PAI-1 molecules on metastasis. To deal with the problem complexity (a
tumor may be composed of millions of cells) two levels are introduced: the cell and
the tumor’s core levels (fig. 7). Indeed, such cancers are generally structured as a ker-
nel of necrosed or quiescent cells surrounded by living tumor cells. As necrosed and
quiescent cells are mostly inactive, tumor’s core is reified as a single upper-level agent,
interacting with cells and PAI-1 molecules at its boundary. A more comprehensive
analysis of this model can be found in Gil-Quijano et al. (2012).
meso
tumor’s core
micro
cell
PAI-1 molecules
Figure 7: ML -ABM in avascular tumor growth modeling (Lepagnot and Hutzler,
2009)
3.4 Ecology
Ecologists study processes that can have very different spatio-temporal dynamics.
Then, characterizing their interactions is a complicated problem and traditional
bottom-up or top-down approaches do not seem relevant: ABMs tend to be too
10 https://sites.google.com/site/agentbasedtumormodeling/home
16
complex, requiring a lot of computational resources11 while EBMs cannot deal with
complex heterogenous environments (Shnerb et al., 2000).
Ecological systems are generally described as hierarchies (Müller et al., 2005;
Ratzé et al., 2007). Thus the hierarchy theory is "a view of ecological systems, which
takes the scales of observation explicitly into account and which tries to conceptual-
ize the phenomena at their proper scale" (Ratzé et al., 2007, p. 14). ML -ABM seems
a interesting way to implement this concept. Different modeling issues in Ecology
have been solved by ML -ABM.
Duboz (2004); Duboz et al. (2003) proposed the scale transfer approach to link
microscopic and macroscopic models: the state of the system is computed by an
ABM and is used to parametrize an EBM describing population dynamics. This
EBM can then be used to parametrize the ABM environment (fig. 8).
emergent computation
ABM
EBM
environment parametrization
Figure 8: The scale transfer approach (Duboz, 2004; Duboz et al., 2003)
Marilleau et al. (2008) introduced a efficient method to represent complex soils,
named APSF (Agent, Pores, Solid and fractal). Traditionally, the environment is
viewed as a regular grid, discretized into cells. A cell can represent a pore, i.e., a part
of a soil cavity, a solid or a fractal. The idea is that a cell is not necessarily an atomic
element describing an homogenous area but can be fractal, i.e., composed of smaller
pore, solid or fractal cells with a self similar structure. Fractal cells are instantiated
at run time, generating finer representations of the environment when it is needed.
Thus, this approach based on self-generation allows to represent complex multi-scale
environments at a minimal computational cost. It has been used in the SWORM
(Simulated WORMS) model that studies the relation between earthworm activity
and soil structure (Blanchart et al., 2009; Laville et al., 2012).
Diptera larvae have a complex gregarious behavior that lead to the formation of
large groups in which individuals regulate the temperature to optimize their devel-
opment speed. This phenomenon can be described by a mesoscopic equation-based
model (knowing the mass of the group and and the external temperature), while the
crowding behavior of larvae can be modeled by an ABM. Moreover, the thermal
dynamics of the cadaver can be modeled by a Cellular Automaton (CA). Morvan
et al. (2008, 2009) integrated these different models in a ML -ABM to perform more
accurate forensic entomology expertises. In this model, the EBM is parametrized ac-
cording to the ABM state. It computes the increase of temperature caused by Diptera
interactions at the group level and send it to the CA model that is used as an envi-
ronment for the ABM (fig. 9). The environment can thus be viewed as an artifact,
used to synchronize the different models.
11An interesting solution to this problem is to reduce the complexity of agent interactions using esti-
mation algorithms such as the fast multipole method (Razavi et al., 2011).
17
meso
thermal dynamics (CA)
maggot mass effect (EBM)
micro
Diptera larvae (ABM)
Figure 9: Levels of organization in a ML -ABM of necrophagous Diptera develop-
ment (Morvan et al., 2008, 2009)
4 Discussion
In this section two issues are discussed:
• the different forms of level integration,
• the use of ML -ABM to solve technical problems.
4.1 Level integration
In the introduction, ML -ABM has been defined as integrating heterogenous ABMs
in a single model. Following the approach of Michel et al. (2003) on interaction, one
can distinguish at least two forms of integration:
• weak integration: levels share objects, e.g., environment properties, but not
agents,
• strong integration: levels share objects and agents.
Weak integration can be regarded as a form of multi-modeling (or model cou-
pling) where levels represent different models interacting through shared variables
called artifacts (Camus et al., 2012; Seck and Honig, 2012; Siebert et al., 2010).
Weaker forms of integration are not regarded as multi-level modeling. Thus, in
the SWARM platform (Minar et al., 1996), integration can be described as bottom-up
or isotropic (information flows in one direction). An agent is designed as a russian
doll and its behavior at a given level depends on the lower ones.
An simple example of strong integration is given by Picault and Mathieu (2011,
p. 334): "a membrane protein, which has an end inside the cell, and the other end
outside".
In the meta-models presented in section 2.1, IRM4MLS and PADAWAN are,
to the best of our knowledge, the only ones able to simulate strongly integrated
levels. It is not surprising as they are based on formal interaction models (respectively
IRM4S and IODA) that differentiate between agent influences and level reaction and
therefore, are able to represent strong interaction (Michel et al., 2003). In short,
agents are strongly interacting if the interaction output depends on the influences of
each agent. Thus, STRIPS-like action models (i.e., that view action as a change of
the state of the world), used in most of agent-based simulation platforms, are unable
to represent such interactions. Yet, (Michel et al., 2003) showed that modeling a
strong interaction as a weak one leads to arbitrary implementation choices and result
interpretation issues.
18
In ML -ABM, the problem is similar since levels can be seen are strongly interact-
ing entities. Thus, we can conclude that a modeling formalism capable of represent-
ing strong interaction12 can be extended to a multi-level one.
4.2 Multi-level technical tools
Although considering cross-level interactions is usually related to the application
domain as shown previously in this article, it can also be viewed as a technical tool:
• an ABM (microscopic level) can be used to parametrize an equation based
model (macroscopic level) (Duboz, 2004; Duboz et al., 2003; Nguyen et al.,
2012a),
• levels can be created at run-time by other levels to generate fractal environ-
ments (Marilleau et al., 2008),
• a mesoscopic level can be viewed as a controller (in the control theory mean-
ing) of group-related properties (Morvan and Jolly, 2012).
• automated observation and analysis tools can be introduced at levels not ex-
plicitly present in the model to detect and study (multi-level) emergent phe-
nomena13 (cf. section 2.2). However, reified emergent phenomena cannot be
considered as model entities since they are not re-injected in the simulation (see
fig. 10),
• a radical interpretation of ML -ABM is the concept of multi-future (Parunak
et al., 2010; Parunak and Brueckner, 2010). The possible trajectories of agent
actions are computed by "ghosts" as a pheromone field and agents act according
to it, selecting the most probable one.
As these ideas are domain-independent, they could be implemented as a generic
library, providing such services to a classical ABM or ML -ABM framework or plat-
form.
Inputs
Simulation
emergent phenomena
Detection
n
o
i
t
a
c
fi
i
t
n
e
g
A
Visualization tools
Figure 10: Two main uses of detected emergent phenomena: visualization or re-
injection in the simulation as agents
12 Some authors described these approaches as "interaction-based modeling", by opposition to the
term "individual-based modeling", as that they focus on interactions rather than on individual behav-
iors (Desmeulles et al., 2009; Kubera et al., 2008).
13 Indeed, as An (2008, p. 4) notes about ABM, "since the models rely on an ill-defined principle of
’emergence’ in order to transcend the epistemological boundaries represented by the multiple hierarchies
of system organization, their behavior is difficult to characterize analytically".
19
5 Conclusion
An important challenge for the scientific community is to find ways to deal with
the complexity of natural and artificial complex systems. This issue led to the devel-
opment of dedicated modeling paradigms and engineering principles that focus on
interaction and organization. We strongly believe that such techniques will play an
important role in the future.
In this article, we conducted a comprehensive analysis of the bibliography avail-
able on one of them: multi-level agent-based modeling.
As we shown, many papers focus on the application of this technique and then,
are published in domain-related journals and conferences. However, a dedicated
venue for theoretical or methodological papers is lacking. Such forum would al-
low to unify the vocabulary and concept definitions, discuss the main issues of this
approach and more generally, share ideas with the interested communities.
Acknowledgments
I am particularly grateful to my col-
leagues Shahin Gelareh, Yoann Kubera
and Jean-Baptiste Soyez for their help
and support. I also would like to thank
Chih-Chun Chen,
Jean-Pierre Müller
and Sebastian Senge for suggesting new
entries in the bibliographical database.
Belem, M. and Müller, J.-P. (2009). Toward a conceptual
framework for multi-points of view analysis in com-
plex system modeling: Orea model.
In 7th Interna-
tional Conference on Practical Applications of Agents and
Multi-Agent Systems (PAAMS 2009), volume 55 of Ad-
vances in Intelligent and Soft Computing, pages 548–556.
Springer.
Belem, M. and Müller, J.-P. (2013). An organizational
model for multi-scale and multi-formalism simulation:
Application in carbon dynamics simulation in west-
african savanna. Simulation Modelling Practice and The-
ory, 32:83–98.
References
Adra, S., Sun, T., MacNeil, S., Holcombe, M., and Small-
wood, R. (2010). Development of a three dimensional
multiscale computational model of the human epider-
mis. PloS one, 5(1):e8511.
An, G. (2008). Introduction of an agent-based multi-scale
modular architecture for dynamic knowledge represen-
tation of acute inflammation. Theoretical Biology and
Medical Modelling, 5(11).
An, G., Wandling, M., and Christley, S. (2013). Agent-
based modeling approaches to multi-scale systems bi-
ology: An example agent-based model of acute pul-
monary inflammation. In Systems Biology - Integrative
Biology and Simulation Tools, pages 429–461. Springer.
An, G. and Wilensky, U. (2009). Artificial life models in
software.
In From Artificial Life to In Silico Medicine:
NetLogo as a Means of Translational Knowledge Represen-
tation in Biomedical Research, pages 183–212. Springer.
Andasari, V., Roper, R., and ans M.A.J. Chaplain, M. S.
(2012).
Integrating intracellular dynamics using com-
pucell3d and bionetsolver: Applications to multiscale
modelling of cancer cell growth and invasion. PLoS
ONE, 7(3):e33726.
Archer, M. (2007). Making our Way through the World: Hu-
man Reflexivity and Social Mobility. Cambridge Univer-
sity Press.
Bertelle, C., Jay, V., Lerebourg, S., Olivier, D., and
Tranouez, P. (2002). Dynamic clustering for auto-
organized structures in complex fluid flows.
In ESS
2002 Conference.
Biggs, M. B. and Papin, J. A. (2013). Novel multiscale mod-
eling tool applied to pseudomonas aeruginosa biofilm
formation. PloS one, 8(10):e78011.
Blanchart, E., Marilleau, N., Chotte, J., Drogoul, A.,
Perrier, E., and Cambier, C. (2009).
SWORM: an
agent-based model to simulate the effect of earthworms
on soil structure. European Journal of Soil Science,
60(1):13–21.
Bonneaud, S. (2008). Des agents-modèles pour la modélisa-
tion et la simulation de systèmes complexes: application
à l’écosystémique des pêches. PhD thesis, Université de
Bretagne occidentale, Brest.
Bonneaud, S., Redou, P., and Chevaillier, P. (2007). Pat-
tern oriented agent-based multi-modeling of exploited
ecosystems.
In Proceedings of the 6th EUROSIM
congress.
Bourdieu, P. (1994). Raisons pratiques:
sur la théorie de
l’action, volume 331 of Points. Éditions du Seuil.
Bourrel, E. (2003). Modélisation dynamique de l’écoulement
du trafic routier : du macroscopique au microscopique.
PhD thesis, Institut National des Sciences Appliquées
de Lyon, Lyon.
Belem, M. (2009). A conceptual model for multipoints of
view analysis of complex systems. Application to the analy-
sis of the carbon dynamics of village territories of the West
African savannas. PhD thesis, AgroParisTech.
Bourrel, E. and Henn., V. (2002). Mixing micro and macro
representations of traffic flow: a first theoretical step.
Proceedings of the 9th Meeting of the Euro Working Group
on Transportation.
20
Bourrel, E. and Lesort, J. (2003). Mixing micro and macro
representations of traffic flow: a hybrid model based
on the LWR theory. 82th Annual Meeting of the Trans-
portation Research Board.
Brown, M., Bowring, E., Epstein, S., Jhaveri, M., Mah-
eswaran, R., Mallick, P., Mumenthaler, S., Povinelli,
M., and Tambe, M. (2011). Applying multi-agent tech-
niques to cancer modeling. In Sixth Annual Workshop
on Multiagent Sequential Decision Making in Uncertain
Domains (MSDM-2011).
Burghout, W., Koutsopoulos, H., and Andreasson, I.
(2005). Hybrid mesoscopic-microscopic traffic simu-
lation. Transportation Research Record: Journal of the
Transportation Research Board, 1934(1):218–255.
Cacciagrano, D., Corradini, F., and Merelli, E. (2010).
Bone remodelling: A complex automata-based model
running in bioshape. Cellular Automata, pages 116–
127.
Caillou, P. and Gil-Quijano, J. (2012). Simanalyzer : Au-
tomated description of groups dynamics in agent-based
simulations. In Proc. of 11th Int. Conf. on Autonomous
Agents and Multiagent Systems (AAMAS 2012).
Caillou, P., Gil-Quijano, J., and Zhou, X. (2012). Auto-
mated observation of multi-agent based simulations: a
statistical analysis approach. to appear in Studia Infor-
matica Universalis, 10(3):62–86.
Camus, B., Bourjot, C., and Chevrier, V. (2013). Multi-
level modeling as a society of interacting models.
In
SpringSim’13, ADS Symposium-Spring Simulation Multi-
Conference, Agent-Directed Simulation Symposium-2013,
volume 1, pages 15–22.
Camus, B., Siebert, J., Bourjot, C., and Chevrier, V. (2012).
Modélisation multi-niveaux dans AA4MM.
In Actes
des 20èmes Journées Francophones sur les Systèmes Multi-
Agents ( JFSMA).
Chen, C. (2009). Complex Event Types for Agent-Based Sim-
ulation. PhD thesis, University College London.
Chen, C. (2013). A theoretical framework for con-
ducting multi-level studies of complex social systems
with agent-based models and empirical data. CoRR,
abs/1302.4774.
Chen, C., Clack, C., and Nagl, S. (2010). Identifying multi-
level emergent behaviors in agent-directed simulations
using complex event type specifications. Simulation,
86(1):41–51.
Chen, C. and Hardoon, D. (2010). Learning from multi-
level behaviours in agent-based simulations: a systems
biology application.
Journal of Simulation, 4(3):196–
203.
Chen, C., Nagl, S., and Clack, C. (2008a). A method for
validating and discovering associations between multi-
level emergent behaviours in agent-based simulations.
In N.T. Nguyen, G.S. Jo, R. H. and Jain, L., editors,
Agent and Multi-Agent Systems: Technologies and Appli-
cations, volume 4953 of Lecture Notes in Computer Sci-
ence, pages 1–10. Springer.
Chen, C., Nagl, S., and Clack, C. (2008b). Multi-level be-
haviours in agent-based simulation: colonic crypt cell
populations. In Proceedings of the Seventh International
Conference on Complex Systems.
Chen, C., Nagl, S., and Clack, C. (2009). A formal-
ism for multi-level emergent behaviours in designed
component-based systems and agent-based simulations.
In Aziz-Alaoui, M. and Bertelle, C., editors, From
System Complexity to Emergent Properties, volume 12
of Understanding Complex Systems, pages 101–114.
Springer.
Cheong, S., Brown, D., Kok, K., and Lopez-Carr, D.
(2012). Mixed methods in land change research: to-
wards integration. Transactions of the Institute of British
Geographers, 37(1):8–12.
Christley, S., Newman, S. A., and Alber, M. S. (2007a).
Agent-based model for developmental pattern forma-
tion with multiscale dynamics and varying cell geome-
try. In Mathematical Modeling of Biological Systems, vol-
ume 1, pages 149–161. Springer.
Christley, S., Zhu, X., Newman, S. A., and Alber, M. S.
(2007b). Multiscale agent-based simulation for chon-
drogenic pattern formation in vitro. Cybernetics and
Systems: An International Journal, 38(7):707–727.
Conte, R., Andrighetto, G., Campennì, M., and Paolucci,
M. (2007). Emergent and immergent effects in com-
plex social systems. In Proceedings of AAAI Symposium,
Social and Organizational Aspects of Intelligence, Wash-
ington DC.
Conte, R. and Castelfranchi, C. (1996). Simulating multi-
agent interdependencies. a two-way approach to the
micro-macro link.
In Social Science Microsimulation,
pages 394–415. Springer.
Conte, R. and Paolucci, M. (2002). Reputation in Artificial
Societies. Kluwer Academic Publishers.
Dascalu, M., Gheorghe, S., and Zafiu, A. (2009). Multi-
level simulator for artificial societies. Romanian Journal
of Information Science and Technology, 12(4):466–484.
Dascalu, M., Stefan, G., Zafiu, A., and Plavitu, A. (2011).
Applications of multilevel cellular automata in epi-
demiology. In Proceedings of the 13th WSEAS interna-
tional conference on Automatic control, modelling & sim-
ulation, pages 439–444. World Scientific and Engineer-
ing Academy and Society (WSEAS).
David, D. and Courdier, R. (2008). Emergence as meta-
knowledge: refining simulation models through emer-
gence reification. In Proceedings of ESM’2008, Le Havre,
France, pages 25–27.
David, D. and Courdier, R. (2009). See emergence as a
metaknowledge. a way to reify emergent phenomena in
multiagent simulations? In Proceedings of ICAART’09,
pages 564–569, Porto, Portugal.
David, D., Gangat, Y., Payet, D., and Courdier, R. (2012).
Reification of emergent urban areas in a land-use sim-
ulation model in reunion island.
In ECAI workshop
on Intelligent Agents in Urban Simulations and Smart
Cities (IAUSSC2012), pp. 28-32, Montpellier, pages 28–
32, Montpellier, France.
David, D., Payet, D., and Courdier, R. (2011). Réifica-
tion de zones urbaines émergentes dans un modèle sim-
ulant l’évolution de la population à la réunion. In Actes
des 19èmes Journées Francophones sur les Systèmes Multi-
Agents ( JFSMA’11), pages 63–72, Valenciennes, France.
Cépaduès.
21
Davis, P. K. and Hillestad, R. (1993). Families of models
that cross levels of resolution: issues for design, calibra-
tion and management. In Proceedings of the 25th confer-
ence on Winter simulation, WSC ’93, pages 1003–1012,
New York, NY, USA. ACM.
Deisboeck, T. and Stamatakos, G. (2010). Multiscale cancer
modeling, volume 34. CRC Press.
Desmeulles, G., Bonneaud, S., Redou, P., Rodin, V., and
Tisseau, J. (2009). In virtuo experiments based on the
multi-interaction system framework: the réiscop meta-
model. Computer Modeling in Engineering and Sciences,
47(3):299–329.
Drogoul, A., Amouroux, E., Caillou, P., Gaudou, B., Grig-
nard, A., Marilleau, N., Taillandier, P., Vavasseur, M.,
Vo, D.-A., and Zucker, J.-D. (2013). GAMA: A spa-
tially explicit, multi-level, agent-based modeling and
simulation platform. In Advances on Practical Applica-
tions of Agents and Multi-Agent Systems, pages 271–274.
Springer.
Drogoul, A., Vanbergue, D., and Meurisse, T. (2003).
Multi-agent based simulation: Where are the agents?
In Sichman, J., Bousquet, F., and Davidsson, P., editors,
Multi-Agent-Based Simulation II, volume 2581 of Lecture
Notes in Computer Science, pages 1–15. Springer.
Duboz, R. (2004). Intégration de modèles hétèrogenes pour la
modélisation et la simulation de systèmes complexes. PhD
thesis, Université du Littoral.
Duboz, R., Ramat, É., and Preux, P. (2003). Scale transfer
modeling: using emergent computation for coupling
an ordinary differential equation system with a reac-
tive agent model. Systems Analysis Modelling Simula-
tion, 43(6):793–814.
Duhail, N. (2013). DEVS et ses extensions pour des simula-
tions multi-modèles en interaction multi-échelles. Mas-
ter’s thesis, Université de Rennes 1 — Telecom Bre-
tagne.
El hmam, M. (2006). Contribution à la modélisation et à la
simulation hybride du flux de trafic. PhD thesis, Univer-
sité d’Artois.
El hmam, M., Abouaissa, H., and Jolly, D. (2008). Con-
tribution à la modélisation et la simulation du flux de
trafic: approche hybride basée sur les systèmes multi-
agents. Actes INRETS, pages 93–111.
El hmam, M., Abouaissa, H., Jolly, D., and Benasser, A.
(2006a). Macro-micro simulation of traffic flow.
In
Proceedings of the12th IFAC Symposium on Information
Control Problems in Manufacturing (INCOM06), pages
351–356, Saint Etienne FRANCE.
El hmam, M., Abouaissa, H., Jolly, D., and Benasser, A.
(2006b). Simulation hybride de flux de trafic basée sur
les systèmes multi-agents. In 6e Conférence Francophone
de MOdélisation et SIMulation-MOSIM.
El hmam, M., Jolly, D., Abouaissa, H., and Benasser, A.
(2006c). Modélisation hybride du flux de trafic. Revue
électronique Sciences et Technologies de l’Automatique.
Epstein, J. (2006). Generative social science: Studies in agent-
based computational modeling. Princeton University
Press.
Espié, S., Gattuso, D., and Galante, F. (2006). Hybrid traf-
fic model coupling macro- and behavioral microsimula-
tion. In 85th Annual Meeting of Transportation Research
Board, Washington D.C.
Ferber, J. (1999). Multi-Agent Systems: An Introduction
to Distributed Artificial Intelligence. Addison-Wesley
Longman Publishing Co., Inc.
Ferber, J. and Müller, J.-P. (1996). Influences and reaction:
a model of situated multiagent systems. In 2nd Inter-
national Conference on Multi-agent systems (ICMAS’96),
pages 72–79.
Flacher, F., Navarro, L., and Corruble, V. (2012). Multi-
agent simulation system. European Patent EP2518711,
Thalès S.A.
Gaud, N. (2007).
Systèmes multi-agent holoniques : De
l’analyse à l’implantation. Méta-modèle, méthodologie, et
simulation multi-niveaux. PhD thesis, Université de
Technologie de Belfort-Montbéliard.
Gaud, N., Galland, S., Gechter, F., Hilaire, V., and
Koukam, A. (2008a). Holonic multilevel simulation of
complex systems : Application to real-time pedestrians
simulation in virtual urban environment. Simulation
Modelling Practice and Theory, 16:1659–1676.
Gaud, N., Galland, S., and Koukam, A. (2008b). Towards a
multilevel simulation approach based on holonic mul-
tiagent systems. In UKSIM, pages 180–185, Cambridge,
UK.
Giddens, A. (1987). La constitution de la société: éléments de
la théorie de la structuration. Presses Universitaires de
France.
Gil-Quijano, J. (2007). Modèles d’auto-organisation pour
l’émergence de formes urbaines à partir de comportements
individuels à Bogota. PhD thesis, Université Paris VI.
Gil-Quijano, J., Hutzler, G., and Louail, T. (2009). De
la cellule biologique à la cellule urbaine : retour sur
trois expériences de modélisation multi-échelles à base
d’agents. In Actes des 17èmes Journées Francophones sur
les Systèmes Multi-Agents ( JFSMA’09).
Gil-Quijano, J., Hutzler, G., and Louail, T.
(2010).
Accroche-toi au niveau,
j’enlève l’échelle: Éléments
d’analyse des aspects multiniveaux dans la simula-
tion à base d’agents. Revue d’Intelligence Artificielle,
24(5):625–648.
Gil-Quijano, J., Louail, T., and Hutzler, G. (2012). From
biological to urban cells: Lessons from three multi-
level agent-based models.
In Desai, N., Liu, A., and
Winikoff, M., editors, Principles and Practice of Multi-
Agent Systems, volume 7057 of Lecture Notes in Com-
puter Science, pages 620–635. Springer.
Gil-Quijano, J. and Piron, M. (2007). Formation au-
tomatique de groupes d’agents sociaux par techniques
d’apprentissage non supervise. In Actes de EGC’07 Ate-
lier Fouille de Données et Algorithmes Biomimétiques.
(2007).
Gil-Quijano, J., Piron, M., and Drogoul, A.
Vers une simulation multi-agent de groupes d’individus
résidentielles
les mobilités
pour modéliser
intra-
urbaines.
géomatique,
Revue
internationale de
17(2):161–181.
Gil-Quijano, J., Piron, M., and Drogoul, A. (2008). Social
Simulation: Technologies, Advances and New Discover-
ies, chapter Mechanisms of Automated Formation and
Evolution of Social-Groups: A Multi-Agent System to
Model the Intra-Urban Mobilities of Bogota City. IGI
Global.
22
Gilbert, N. (2002). Varieties of emergence. In Agent 2002
Conference: Social agents: ecology, exchange, and evolu-
tion, Chicago.
Gilbert, N. (2007). Agent-Based Models. Number 153 in
Quantitative Applications in the Social Sciences. Sage
Publications.
Hassoumi,
I., Lang, C., Marilleau, N., Temani, M.,
Ghedira, K., and Zucker, J. D. (2012). Toward a
spatially-centered approach to integrate heterogeneous
and multi-scales urban component models. In Advances
on Practical Applications of Agents and Multi-Agent Sys-
tems, volume 155 of Advances in Intelligent and Soft
Computing Volume, pages 81–86. Springer.
Hayenga, H., Thorne, B., Peirce, S., and Humphrey,
J. (2011). Ensuring congruency in multiscale mod-
eling:
towards linking agent based and continuum
biomechanical models of arterial adaptation. Annals
of biomedical engineering, pages 1–14.
Jeschke, M. and Uhrmacher, A. (2008). Multi-resolution
spatial simulation for molecular crowding.
In Proc.
of the Winter Simulation Conference, pages 1384–1392.
IEEE.
Kim, M., Christley, S., Alverdy, J., Liu, D., and An, G.
(2012). A multi-scale agent-based model of necrotizing
enterocolitis integrating oxidative stress, inflammation,
and microvascular thrombosis. Journal of Surgical Re-
search, 172(2):196.
Koestler, A. (1967). The Ghost in the Machine. Hutchinson.
Koestler, A. (1978). Janus: A summing up. Random House.
Kubera, Y., Mathieu, P., and Picault, S. (2008). Interaction-
oriented agent simulations: From theory to implemen-
tation. In Proc. of the18th European Conf. on Artificial
Intelligence (ECAI’08), pages 383–387.
Laperrière, V. (2012). Un cadre de modélisation dy-
namique multiniveau pour la peste bubonique des
hautes terres de madagascar. Bulletin de l’Association
de géographes français, 89(2):278–288.
Laville, G., Mazouzi, K., Lang, C., Marilleau, N., and
Philippe, L. (2012). Using gpu for multi-agent multi-
scale simulations. In Distributed Computing and Arti-
ficial Intelligence, volume 151 of Advances in Intelligent
and Soft Computing, pages 197–204. Springer.
Le, Q., Seidl, R., and Scholz, R. (2011). Feedback loops
and types of adaptation in the modelling of land-use
decisions in an agent-based simulation. Environmental
Modelling & Software.
Leonard, A. (1980). Vortex methods for flow simulation.
Journal of Computational Physics, 37(3):289–335.
Mammar, S.and Lebacque, J. and Haj-Salem, H. (2006).
Hybrid model based on second-order traffic model.
85th Annual Meeting of Transportation Research Board,
1(06-2160).
Marcenac, P., Courdier, R., and Calderoni, S. (1998a). To-
wards a toolkit to study emergence in complex systems
simulations.
http://personnel.univ- reunion.
fr/courdier/pdfs/int_iea1998_MC98.pdf.
Marcenac, P., Courdier, R., Calderoni, S., and Soulie,
J. (1998b). Towards an emergence machine for com-
plex systems simulations. In Tasks and Methods in Ap-
plied Artificial Intelligence, volume 416, pages 785–794.
Springer.
Marcenac, P. and Giroux, S. (1998). Geamas: A generic
architecture for agent-oriented simulations of complex
processes. Applied Intelligence, 8(3):247–267.
Marilleau, N., Cambier, C., Drogoul, A., Chotte, J., Per-
rier, E., and Blanchart, E. (2008). Multiscale mas mod-
elling to simulate the soil environment: Application to
soil ecology. Simulation Modelling Practice and Theory,
16(7):736–745.
Marino, S., El-Kebir, M., and Kirschner, D. (2011). A hy-
brid multi-compartment model of granuloma forma-
tion and t cell priming in tuberculosis. Journal of Theo-
retical Biology.
Mathieu,
J., Mahoney, P., Boiney, L.,
James,
J.,
and White, B.
Hubbard, R.,
(2007a).
Hy-
brid systems dynamic, petri net, and agent-based
modeling of
the air and space operations cen-
ter. http://www.mitre.org/work/tech_papers/
tech_papers_07/07_0455/07_0455.pdf.
J., Mahoney, P.,
Mathieu,
J., Melhuish,
J.,
James,
and White, B.
Boiney, L.,
Multi-
(2007b).
scale modeling of
center.
air operations
the
http://www.mitre.org/work/tech_papers/
tech_papers_07/06_1497/06_1497.pdf.
Maturana, H. R. and Varela, F. J. (1980). Autopoiesis and
Cognition The Realization of the Living, volume 42 of
Boston Studies in the Philosophy of Science.
Springer
Berlin Heidelberg.
Maus, C., John, M., Röhl, M., and Uhrmacher, A. (2008).
Hierarchical modeling for computational biology.
In
Bernardo, M., Degano, P., and Zavattaro, G., editors,
Formal Methods for Computational Systems Biology, vol-
ume 5016 of Lecture Notes in Computer Science, pages
81–124. Springer.
Maus, C., Rybacki, S., and Uhrmacher, A. (2011). Rule-
based multi-level modeling of cell biological systems.
BMC systems biology, 5(1):166–186.
Lepagnot, J. and Hutzler, G. (2009). A multiscale agent-
based model for the simulation of avascular tumour
growth.
Journal of Biological Physics and Chemistry,
9(1):17–25.
Michel, F.
the influ-
The IRM4S model:
(2007).
ence/reaction principle for multiagent based simula-
tion. In Proc. of 6th Int. Conf. on Autonomous Agents
and Multiagent Systems (AAMAS 2007), pages 1–3.
Louail, T. (2010). Comparer les morphogénèses urbaines
en Europe et aux États-Unis par la simulation à base
d’agents – Approches multi-niveaux et environnements de
simulation spatiale. PhD thesis, Université d’Evry-Val
d’Essonne.
Michel, F., Gouaïch, A., and Ferber, J. (2003). Weak in-
teraction and strong interaction in agent based simu-
lations.
In Multi-Agent-Based Simulation III, volume
2927 of Lecture Notes in Computer Science, pages 43–56.
Springer.
Magne, L., Rabut, S., and Gabard, J. (2000). Towards an
hybrid macro-micro traffic flow simulation model. Pro-
ceedings of the INFORMS Salt Lake City String 2000 Con-
ference.
Minar, N., Burkhart, R., Langton, C., and Askenazi, M.
(1996). The swarm simulation system: A toolkit for
building multi-agent simulations. Technical Report 96-
06-042, Santa Fe Institute.
23
Moncion, T., Amar, P., and Hutzler, G. (2010). Automatic
characterization of emergent phenomena in complex
systems.
Journal of Biological Physics and Chemistry,
10:16–23.
Montagna, S., Donati, N., and Omicini, A. (2010a).
An agent-based model for the pattern formation in
drosophila melanogaster. Artificial Life XII, pages 110–
117.
Montagna, S., Omicini, A., and Ricci, A. (2010b). A multi-
scale agent-based model of morphogenesis in biological
systems. In Omicini, A. and Viroli, M., editors, Pro-
ceedings of 11 ◦ Workshop nazionale "Dagli Oggetti agli
Agenti" WOA 2010, Rimini, Italy.
Morin, E. (1992). Method: Towards a Study of Humankind,
volume 1. Peter Lang Pub Inc.
Morvan, G., Dupont, D., Soyez, J.-B., and Merzouki, R.
(2012). Engineering hierarchical complex systems: an
agent-based approach – the case of flexible manufactur-
ing systems. In Service Orientation in Holonic and Multi
Agent Manufacturing Control, volume 402 of Studies in
Computational Intelligence. Springer.
Morvan, G. and Jolly, D. (2012). Multi-level agent-based
modeling with the Influence Reaction principle. CoRR,
abs/1204.0634.
Morvan, G., Jolly, D., Veremme, A., Dupont, D., and
Charabidze, D. (2008). Vers une méthode de modéli-
sation multi-niveaux.
In Actes de la 7ème Conférence
de Modélisation et Simulation MOSIM, Paris, France, vol-
ume 1, pages 167–174.
Morvan, G., Veremme, A., and Dupont, D.
(2011).
IRM4MLS: the influence reaction model for multi-level
simulation.
In Bosse, T., Geller, A., and Jonker, C.,
editors, Multi-Agent-Based Simulation XI, volume 6532
of Lecture Notes in Artificial Intelligence, pages 16–27.
Springer.
Morvan, G., Veremme, A., Dupont, D., and Jolly, D.
(2009). Modélisation et conception multiniveau de sys-
tèmes complexes : stratégie d’agentification des organi-
sations. Journal Européen des Systèmes Automatisés, 43(4-
5):381–406.
Müller, J.-P. (2009). Towards a formal semantics of event-
based multi-agent simulations. In David, N. and Sich-
man, J., editors, Multi-Agent-Based Simulation IX, vol-
ume 5269 of Lecture Notes in Computer Science, pages
110–126. Springer.
Müller, J.-P. and Aubert, S. (2011). L’ontologie pour con-
struire une représentation multi-niveau de et par les
systèmes sociaux.
In Rencontres interdisciplinaires de
Rochebrune sur les systèmes complexes naturels et artifi-
ciels.
Müller, J.-P. and Diallo, A. (2012). Vers une méthode
multi-point de vue de modélisation multi-agent.
In
Actes des 20èmes Journées Francophones sur les Systèmes
Multi-Agents, pages 33–42. Cepadues Editions.
Müller, J.-P., Diallo, A., François, J.-C., Sanders, L.,
Waldeck, R., and Mathian, H. (2011). Building ontolo-
gies from a variety of points of view. In Proceedings of
the 7th European Social Simulation Association Confer-
ence.
Müller, J.-P., Ratzé, C., Gillet, F., and Stoffel, K. (2005).
Modeling and simulating hierarchies using an agent-
based approach. In Proceedings of the MODSIM 2005 In-
ternational Congress on Modelling and Simulation, pages
1631–1638.
Navarro, L., Corruble, V., Flacher, F., and Zucker, J.-D.
(2012). Mesoscopic level: A new representation level
for large scale agent-based simulations. In SIMUL 2012,
The Fourth International Conference on Advances in Sys-
tem Simulation, pages 68–73.
Navarro, L., Corruble, V., and Zucker, J.-D. (2013). A
flexible approach to multi-level agent-based simulation
with the mesoscopic representation.
In Proc. of 12th
Int. Conf. on Autonomous Agents and Multiagent Systems
(AAMAS 2013).
Navarro, L., Flacher, F., and Corruble, V. (2011). Dynamic
level of detail for large scale agent-based urban simula-
tions. In Proc. of 10th Int. Conf. on Autonomous Agents
and Multiagent Systems (AAMAS 2011), pages 701–708.
Nguyen, N., Taillandier, P., Drogoul, A., and Auger, P.
(2012a).
Inferring equation-based models from agent-
based models: a case study in competition dynamics.
In Principles and Practice of Multi-Agent Systems, volume
7455 of Lecture Notes in Artificial Intelligence, pages
413–427. Springer.
Nguyen, T., Zucker, J.-D., Nguyen, D., Drogoul, A.,
and Vo, A. (2011). Hybrid equation-based and agent-
based modeling of crowd evacuation on road network.
In Eighth International Conference on Complex Systems
(ICSS), pages 558–570.
Nguyen, T., Zucker, J.-D., Nguyen, D., Drogoul, A., and
Vo, A. (2012b). A hybrid macro-micro pedestrians
evacuation model to speed up simulation in road net-
works.
In Dechesne, F., Hattori, H., ter Mors, A.,
Such, J., Weyns, D., and Dignum, F., editors, Advanced
Agent Technology, volume 7068 of Lecture Notes in Com-
puter Science, pages 371–383. Springer.
North, M., Macal, C., Aubin, J., Thimmapuram, P., Bra-
gen, M., Hahn, J., Karr, J., Brigham, N., Lacy, M., and
Hampton, D. (2010). Multiscale agent-based consumer
market modeling. Complexity, 15(5):37–47.
Olsen, M. M. and Siegelmann, H. T. (2013). Multiscale
agent-based model of tumor angiogenesis. In Interna-
tional Conference on Computational Science, ICCS 2013,
volume 18 of Procedia Computer Science, pages 1026–
1035. Elsevier.
Ozik, J., Sallach, D. L., and Macal, C. M. (2008). Model-
ing dynamic multiscale social processes in agent-based
models. IEEE Intelligent Systems, 23(4):36–42.
Paiva, L., Binny, C., Ferreira Jr., S., and Martins, M.
(2009). A multiscale mathematical model for oncolytic
virotherapy. Cancer Research, 69:1205–1211.
Parry, H. and Bithell, M. (2012). Agent-based models of geo-
graphical systems, chapter Large scale agent-based mod-
elling: A review and guidelines for model scaling, pages
271–308. Springer.
Parunak, H. (2012). Between agents and mean fields.
In Multi-Agent-Based Simulation XII, volume 7124 of
Lecture Notes in Artificial Intelligence, pages 113–126.
Springer.
Parunak, H., Bisson, R., and Brueckner, S. (2010). Agent
interaction, multiple perspectives, and swarming simu-
lation. In Proc. of 9th Int. Conf. on Autonomous Agents
and Multiagent Systems (AAMAS 2010), pages 549–556.
24
Parunak, H. and Brueckner, S. (2010). Multi-perspective,
multi-future modeling and model analysis. In Proceed-
ings of ICCDM, Miami, FL.
Sawyer, R. (2003). Artificial societies, multiagent systems
and the micro-macro link in sociological theory. Socio-
logical Methods & Research, 31(3):325–363.
Parunak, H., Sauter, J., and Crossman, J. (2009). Multi-
layer simulation for analyzing IED threats. In Technolo-
gies for Homeland Security, 2009. HST’09. IEEE Confer-
ence on, pages 323–330. IEEE.
Picault, S. and Mathieu, P. (2011). An interaction-oriented
model for multi-scale simulation.
In Walsh, T., ed-
itor, Twenty-Second International Joint Conference on
Artificial Intelligence, pages 332–337, Barcelona, Spain.
AAAI Press.
Poschinger, A., Kates, R., and Keller, H. (2002). Coupling
of concurrent macroscopic and microscopic traffic flow
models using hybrid stochastic and deterministic disag-
gregation. Transportation and Traffic Theory for the 21st
century.
Prévost, G., Tranouez, P., Lerebourg, S., Bertelle, C., and
Olivier, D. (2004). Ecosystem complexity described
with ontological tool for a multi-scale, multi-model ap-
proaches in distributed environment. In Jordan Inter-
national Conference on Computer Sciences and Engineer-
ing, Al-Salt, Jordan.
Pumain, D. and Louail, T. (2009). Interaction des ontolo-
gies informatique et géographique pour simuler les dy-
namiques multiscalaires.
In XVIèmes rencontres inter-
disciplinaires de Rochebrune, Ontologies et dynamique des
systèmes complexes, perspectives interdisciplinaires.
Pumain, D., Sanders, L., Bretagnolle, A., Glisse, B., and
Mathian, H. (2009). The future of urban systems: Ex-
ploratory models. In Lane, D., Pumain, D., der Leeuw,
S. V., and West, G., editors, Complexity Perspectives in
Innovation and Social Change, volume 7 of Methodos Se-
ries, pages 331–360. Springer.
Railsback, S. and Grimm, V. (2011). Agent-Based and
Individual-Based Modeling: A Practical Introduction.
Princeton University Press.
Ratzé, C., Gillet, F., Müller, J.-P., and Stoffel, K. (2007).
Simulation modelling of ecological hierarchies in con-
structive dynamical systems. Ecological Complexity,
4(1–2):13–25.
Raub, W., Buskens, V., and van Assen, M. A. (2011).
Micro-macro links and microfoundations in sociology.
The Journal of Mathematical Sociology, 35(1-3):1–25.
Razavi, S., Gaud, N., Mozayani, N., and Koukam, A.
(2011). Multi-agent based simulations using fast mul-
tipole method: application to large scale simulations of
flocking dynamical systems. Artificial Intelligence Re-
view, 35(1):53–72.
Rejniak, K. and Anderson, A. (2011). Hybrid models of
tumor growth. Wiley Interdisciplinary Reviews: Systems
Biology and Medicine, 3(1):115–125.
Resnick, M. (1994). Turtles, termites, and traffic jams: Ex-
plorations in massively parallel microworlds. Mit Press.
Rounsevell, M., Pedroli, B., Erb, K., Gramberger, M.,
Busck, A., Haberl, H., Kristensen, S., Kuemmerle, T.,
Lavorel, S., Lindner, M., et al. (2012). Challenges for
land system science. Land Use Policy.
Sawyer, R. (2001). Simulating emergence and downward
causation in small groups.
In Multi-agent-based simu-
lation, Lecture Notes in Artificial Intelligence, pages
49–67. Springer.
Scerri, D., Drogoul, A., Hickmott, S., and Padgham, L.
(2010). An architecture for modular distributed simula-
tion with agent-based models. In Proc. of 9th Int. Conf.
on Autonomous Agents and Multiagent Systems (AAMAS
2010), pages 541–548. International Foundation for Au-
tonomous Agents and Multiagent Systems.
Schaller, N., Lazrak, E., Martin, P., Mari, J., Aubry, C.,
and Benoît, M. (2012). Combining farmers’ decision
rules and landscape stochastic regularities for landscape
modelling. Landscape Ecology, 27(3):1–14.
Scheff, J., Mavroudis, P., Foteinou, P., An, G., Calvano, S.,
Doyle, J., Dick, T., Lowry, S., Vodovotz, Y., and An-
droulakis, I. (2012). A multiscale modeling approach
to inflammation: A case study in human endotoxemia.
Journal of Computational Physics.
Scheutz, M., Madey, G., and Boyd, S. (2005).
tMANS -
the multi-scale agent-based networked simulation for
the study of multi-scale, multi-level biological and so-
cial phenomena. In Spring Simulation Multiconference.
Schillo, M., Fischer, K., and Klein, C. (2001). The micro-
macro link in DAI and sociology. In Multi-Agent Based
Simulation, volume 1979 of Lecture Notes in Artificial
Intelligence, pages 133–148. Springer.
Schmidt, S., Picioreanu, C., Craenen, B., Mackay, R.,
Kreft, J., and Theodoropoulos, G. (2011). A multi-
scale agent-based distributed simulation framework for
groundwater pollution management. In 15th Interna-
tional Symposium on Distributed Simulation and Real
Time Applications (DS-RT), pages 18–27. IEEE.
Schnell, S., Grima, R., and Maini, P. (2007). Multiscale
modeling in biology. new insights into cancer illustrate
how mathematical tools are enhancing the understand-
ing of life from the smallest scale to the grandest. Amer-
ican Scientist, 95(2):134–142.
Seal,
J., Zaborina, O., An, G., et al.
J., Alverdy,
(2011). Agent-based dynamic knowledge representa-
tion of pseudomonas aeruginosa virulence activation in
the stressed gut: Towards characterizing host-pathogen
interactions in gut-derived sepsis. Theoretical Biology
and Medical Modelling, 8(1):33.
Seck, M. and Honig, H. (2012). Multi-perspective mod-
elling of complex phenomena. Computational & Math-
ematical Organization Theory, 18(1):128–144.
Seidl, R., Rammer, W., Scheller, R., and Spies, T.
(2012). An individual-based process model to simulate
landscape-scale forest ecosystem dynamics. Ecological
Modelling, 231:87–100.
Seidl, R., Rammer, W., Scheller, R., Spies, T., and Lexer,
M. (2010). A mechanistic, individual-based approach
to modeling complexity and dynamics of forest ecosys-
tems across scales. The 95th ESA Annual Meeting.
Semeniuk, C., Musiani, M., and Marceau, D. (2011). Biodi-
versity, chapter Integrating Spatial Behavioral Ecology
in Agent-Based Models for Species Conservation, pages
3–26. InTech.
Servat, D. (2000). Modélisation de dynamiques de flux par
agents. Application aux processus de ruissellement, infil-
tration et érosion. PhD thesis, Université Paris VI.
25
Servat, D., Perrier, E., Treuil, J.-P., and Drogoul, A.
(1998a). When agents emerge from agents: Introduc-
ing multi-scale viewpoints in multi-agent simulations.
In Multi-Agent Systems and Agent-Based Simulation, vol-
ume 1534 of Lecture Notes in Computer Science, pages
183–198. Springer.
Soyez, J.-B., Morvan, G., Dupont, D., and Merzouki, R.
(2013). A methodology to engineer and validate dy-
namic multi-level multi-agent based simulations.
In
Multi-Agent-Based Simulation XIII, volume 7838 of Lec-
in Artificial Intelligence, pages 130–142.
ture Notes
Springer.
Servat, D., Pierrer, E., Treuil, J., and Drogoul, A. (1998b).
Towards virtual experiment laboratories : How multi-
agent simulations can cope with multiple scales of anal-
ysis and viewpoints. In Virtual Worlds, volume 1434
of Lecture Notes in Computer Science, pages 205–217.
Springer.
Sewall, J., Wilkie, D., and Lin, M. C. (2011). Interactive
hybrid simulation of large-scale traffic. ACM Trans-
action on Graphics (Proceedings of SIGGRAPH Asia),
30(6).
Sharpanskykh, A. and Treur, J. (2011a). Group abstrac-
tion for large-scale agent-based social diffusion models.
In International conference on Privacy, security, risk and
trust and international conference on social computing,
pages 830–837. IEEE.
Sharpanskykh, A. and Treur, J. (2011b). Group abstrac-
tion for large-scale agent-based social diffusion models
with unaffected agents.
In Agents in Principle, Agents
in Practice, volume 7047 of Lecture Notes in Artificial
Intelligence, pages 129–142. Springer.
Shimoni, Y., Nudelman, G., Hayot, F., and Sealfon, S.
(2011). Multi-scale stochastic simulation of diffusion-
coupled agents and its application to cell culture simu-
lation. PloS one, 6(12):e29298.
Shnerb, N., Louzoun, Y., Bettelheim, E., and Solomon,
S. (2000). The importance of being discrete: Life al-
ways wins on the surface. Proceedings of the National
Academy of Sciences, 97(19):10322–10324.
Siebert, J. (2011). Approche multi-agent pour la multi-
modélisation et le couplage de simulations. Application à
l’étude des influences entre le fonctionnement des réseaux
ambiants et le comportement de leurs utilisateurs. PhD
thesis, Université Henri Poincaré-Nancy I.
Siebert, J., Ciarletta, L., and Chevrier, V. (2010). Agents
and artefacts for multiple models co-evolution. build-
ing complex system simulation as a set of interacting
models. In Proc. of 9th Int. Conf. on Autonomous Agents
and Multiagent Systems (AAMAS 2010), pages 509–516.
Smallwood, R., Adra,
and Khodabakshi, G.
S.,
Multiscale, multi-paradigm modelling
(2010).
embedding development
of
in tissue be-
tissues:
haviour.
http://www.dcs.shef.ac.uk/~rod/
integrative_biology/Smallwood_Cellular_
dynamics_and_tissue_behaviour.pdf.
Smallwood, R. and Holcombe, M. (2006). The epithe-
liome project: multiscale agent-based modeling of ep-
ithelial cells.
In Biomedical Imaging: Nano to Macro,
2006. 3rd IEEE International Symposium on, pages 816–
819. IEEE.
Solovyev, A., Mikheev, M., Z., L., Dutta-Moscato, J.,
Ziraldo, C., An, G., Vodovotz, Y., and Mi, Q.
(2010). SPARK: A framework for multi-scale agent-
based biomedical modeling.
International Journal of
Agent Technologies and Systems, 2(3):18–30.
Soyez, J.-B., Morvan, G., Merzouki, R., Dupont, D., and
Kubiak, P. (2011). Multi-agent multi-level modeling – a
methodology to simulate complex systems. In Proceed-
ings of the 23rd European Modeling & Simulation Sympo-
sium.
Squazzoni, F. (2008). The micro-macro link in social sim-
ulation. Sociologica, 1.
Steiniger, A., Kruger, F., and Uhrmacher, A. M. (2012).
Modeling agents and their environment in multi-level-
DEVS.
In Proc. of the Winter Simulation Conference,
pages 1–12. IEEE.
Stiegelmeyer, S. M., Giddings, M. C., et al. (2013). Agent-
based modeling of competence phenotype switching in
bacillus subtilis. Theoretical Biology and Medical Mod-
elling, 10(1):23.
Sun, T., Adra, S., Smallwood, R., Holcombe, M., and Mac-
Neil, S. (2009). Exploring hypotheses of the actions of
tgf-β 1 in epidermal wound healing using a 3d computa-
tional multiscale model of the human epidermis. PloS
one, 4(12):e8515.
Sun, X., Zhang, L., Tan, H., Bao, J., Strouthos, C., and
Zhou, X. (2012). Multi-scale agent-based brain cancer
modeling and prediction of tki treatment response: In-
corporating egfr signaling pathway and angiogenesis.
BMC bioinformatics, 13(1):218.
Tahir, H., Hoekstra, A., Lorenz, E., Lawford, P., Hose, D.,
Gunn, J., and Evans, D. (2011). Multi-scale simulations
of the dynamics of in-stent restenosis: impact of stent
deployment and design. Interface Focus, 1(3):365–373.
Taillandier, P., Vo, D.-A., Amouroux, E., and Drogoul, A.
(2010). GAMA: bringing GIS and multi-level capabil-
ities to multi-agent simulation. In European Workshop
on Multi-Agent Systems.
Taillandier, P., Vo, D.-A., Amouroux, E., and Drogoul,
A. (2012). GAMA: A simulation platform that inte-
grates geographical information data, agent-based mod-
eling and multi-scale control. In Desai, N., Liu, A., and
Winikoff, M., editors, Principles and Practice of Multi-
Agent Systems, volume 7057 of Lecture Notes in Com-
puter Science, pages 242–258. Springer.
Thorne, B., Hayenga, H., Humphrey, J., and Peirce, S.
(2011). Toward a multi-scale computational model of
arterial adaptation in hypertension: verification of a
multi-cell agent based model. Frontiers in Physiology,
2.
Torii, D., Ishida, T., Bonneaud, S., and Drogoul, A. (2005).
Layering social interaction scenarios on environmental
simulation. In Multi-Agent and Multi-Agent-Based Simu-
lation, volume 3415 of Lecture Notes in Artificial Intelli-
gence, pages 78–88. Springer.
Tranouez, P. (2005). Contribution à la modélisation et à la
prise en compte informatique de niveaux de descriptions
multiples. Application aux écosystèmes aquatiques (Peni-
cillo haere, nam scalas aufero). PhD thesis, Université
du Havre.
Tranouez, P., Bertelle, C., and Olivier, D. (2001). Chang-
ing the level of description of a fluid flow in agent-based
simulation. In ESS Conference.
26
Tranouez, P., Bertelle, C., and Olivier, D. (2006). Emer-
gent properties in natural and artificial dynamical sys-
tems, chapter Changing levels of description in a fluid
flow simulation, pages 87–99. Springer.
Wang, Z., Birch, C., and Deisboeck, T. (2008). Cross-
scale sensitivity analysis of a non-small cell lung cancer
model: linking molecular signaling properties to cellu-
lar behavior. Biosystems, 92(3):249–258.
Tranouez, P. and Dutot, A. (2009). Different goals in mul-
tiscale simulations and how to reach them. In Complex
Systems and Self-organization Modelling, Understanding
Complex Systems, pages 29–39. Springer.
Wang, Z. and Deisboeck, T. (2008). Computational mod-
eling of brain tumors: discrete, continuum or hybrid?
Scientific Modeling and Simulation, 15(1):381–393.
Tranouez, P., Lerebourg, S., Bertelle, C., and Olivier, D.
(2003). Contribution à la représentation multi-échelle
des écosystèmes aquatiques. Technique et Science Infor-
matiques, 22(4):255–259.
Wang, Z., Zhang, L., Sagotsky, J., and Deisboeck, T.
(2007). Simulating non-small cell lung cancer with a
multiscale agent-based model. Theoretical Biology and
Medical Modelling, 4.
Treuil, J.-P., Drogoul, A., and Zucker, J.-D. (2008). Modéli-
sation et simulation à base d’agents. Dunod.
Uhrmacher, A. M., Ewald, R., John, M., Maus, C.,
Jeschke, M., and Biermann, S. (2007). Combining mi-
cro and macro-modeling in DEVS for computational
biology. In Proceedings of the 39th conference on Winter
simulation, pages 871–880.
Vincenot, C., Giannino, F., Rietkerk, M., Moriya, K., and
Mazzoleni, S. (2011). Theoretical considerations on the
combined use of system dynamics and individual-based
modeling in ecology. Ecological Modelling, 222(1):210–
218.
Vo, A. (2012). An operational architecture to handle mul-
tiple levels of representation in agent-based models. PhD
thesis, Université Paris VI.
Vo, D.-A., Drogoul, A., and Zucker, J.-D. (2012a). An
operational meta-model for handling multiple scales in
agent-based simulations. In International Conference on
Computing and Communication Technologies, Research,
Innovation, and Vision for the Future (RIVF), pages 1–6,
Ho Chi Minh City. IEEE.
Vo, D.-A., Drogoul, A., Zucker, J.-D., and Ho, T.-V.
(2012b). A modelling language to represent and spec-
ify emerging structures in agent-based model. In Desai,
N., Liu, A., and Winikoff, M., editors, Principles and
Practice of Multi-Agent Systems, volume 7057 of Lecture
Notes in Computer Science, pages 212–227. Springer.
Vodovotz, Y., Csete, M., Bartels, J., Chang, S., and An, G.
(2008). Translational systems biology of inflammation.
PLoS Comput Biol, 4(4):e1000014.
Wakeland, W., Macovsky, L., and An, G. (2007). A hybrid
simulation model for studying acute inflammatory re-
sponse. In Proceedings of the 2007 spring simulation mul-
ticonference, volume 2 of SpringSim ’07, pages 39–46.
Society for Computer Simulation International.
Wang, J., Zhang, L., Jing, C., Ye, G., Wu, H., Miao,
H., Wu, Y., and Zhou, X. (2013). Multi-scale agent-
based modeling on melanoma and its related angiogen-
esis analysis. Theoretical Biology and Medical Modelling,
10(1):41.
Wedde, H. and Senge, S. (2012).
2-way evaluation of
the distributed BeeJamA vehicle routing approach. In
Proceedings of the IEEE Intelligent Vehicles Symposium,
pages 205–210.
White, B. (2007). On interpreting scale (or view) and emer-
gence in complex systems engineering. In Systems Con-
ference, 2007 1st Annual IEEE, pages 1–7. IEEE.
Xi, H. and Son, Y.-J. (2012). Two-level modeling frame-
work for pedestrian route choice and walking behav-
iors. Simulation Modelling Practice and Theory, 22:28–
46.
Zahedmanesh, H. and Lally, C. (2012). A multiscale
mechanobiological modelling framework using agent-
based models and finite element analysis: application
to vascular tissue engineering. Biomechanics and Model-
ing in Mechanobiology, 11:363–377.
Zeigler, B., Kim, T., and Praehofer, H. (2000). Theory of
Modeling and Simulation. Academic Press, 2nd edition.
Zhang, L., Athale, C., and Deisboeck, T. (2007). Devel-
opment of a three-dimensional multiscale agent-based
tumor model: Simulating gene-protein interaction pro-
files, cell phenotypes and multicellular patterns in
brain cancer. Journal of Theoretical Biology, 244(1):96–
107.
Zhang, L., Chen, L., and Deisboeck, T. (2009a). Multi-
scale, multi-resolution brain cancer modeling. Mathe-
matics and computers in simulation, 79(7):2021–2035.
Zhang, L., Jiang, B., Wu, Y., Strouthos, C., Sun, P., Su, J.,
and Zhou, X. (2011). Developing a multiscale, multi-
resolution agent-based brain tumor model by graphics
processing units. Theoretical Biology and Medical Mod-
elling, 8(1):46.
Zhang, L., Wang, Z., Sagotsky, J., and Deisboeck, T.
(2009b). Multiscale agent-based cancer modeling. Jour-
nal of Mathematical Biology, 48(4–5):545–559.
27
|
1207.0405 | 1 | 1207 | 2012-07-02T14:34:05 | MITRA: A Meta-Model for Information Flow in Trust and Reputation Architectures | [
"cs.MA"
] | We propose MITRA, a meta-model for the information flow in (computational) trust and reputation architectures. On an abstract level, MITRA describes the information flow as it is inherent in prominent trust and reputation models from the literature. We use MITRA to provide a structured comparison of these models. This makes it possible to get a clear overview of the complex research area. Furthermore, by doing so, we identify interesting new approaches for trust and reputation modeling that so far have not been investigated. | cs.MA | cs |
MITRA: A Meta-Model for Information Flow in
Trust and Reputation Architectures
netRapid GmbH & Co. KG (http://www.netrapid.de/)
Eugen Staab
[email protected]
Guillaume Muller
Work partly done at:
Escola Polit´ecnica de Sao Paulo (http://www.pcs.usp.br/~lti/) and
Commissariat `a l'´Energie Atomique (http://www-list.cea.fr/)
[email protected]
August 12, 2010
Abstract
We propose MITRA, a meta-model for the information flow in
(computational) trust and reputation architectures. On an abstract
level, MITRA describes the information flow as it is inherent in promi-
nent trust and reputation models from the literature. We use MITRA
to provide a structured comparison of these models. This makes it
possible to get a clear overview of the complex research area. Further-
more, by doing so, we identify interesting new approaches for trust and
reputation modeling that so far have not been investigated.
Keywords: Computational Trust, Reputation Systems, Meta Model.
1 Introduction
Open and decentralized systems are vulnerable to buggy or malicious agents.
To make these systems robust against malfunction, agents need the ability
to assess the reliability and attitudes of other agents in order to choose
trustworthy interaction partners. To this end, a multitude of trust and
reputation models have been and still are being proposed in the literature.
In short, each such model generally defines all or some of the three following
processes: how evidence about the trustworthiness of an agent is gathered,
how this evidence is combined into a final assessment, and how this final
assessment is used in decision-making.
Currently, it is difficult to get an overview of what has been done in
the area, and what needs to be done. The main reasons for this are: that
the proposed trust and reputation models use no common terminology; that
1
they are not compatible in their basic structure; and that their respective
contributions are evaluated against different metrics. While not considering
the last point in this article, the first two questions motivate the introduction
of an abstract model that allows researchers to organize their models in a
unified way.
To this end, we propose MITRA, a meta-model for the information flow
in computational trust and reputation architectures. MITRA formalizes and
organizes the flow of information inside and between agents. More precisely,
it describes the top-level processes to gather evidence, and to combine it
with information exchanged with other agents. MITRA makes use of four
simple concepts of information processing, namely the observation, the eval-
uation, the fusion and the filtering of information, and abstracts away from
numerical computations. Although being an abstract model, MITRA cap-
tures important concepts used by existing trust and reputation models. This
way, MITRA provides a big-picture of the trust and reputation domain and
paves the way for a structured survey of the domain.
The model is useful for the community in at least four respects. First, it
serves as a terminological and structural framework to describe new models.
Secondly, it provides a means for researchers to classify and compare existing
approaches in this domain. In addition to this, MITRA helps to identify new
approaches to model trust and reputation, as we will show in this article.
Finally, it helps newcomers to get a concise overview on the structure of
computational models of trust and reputation.
The remainder of the article is organized as follows.
In Sect. 2, we
introduce basic concepts of trust and reputation modeling. We describe
the MITRA model in Sect. 3. Following this, in Sect. 4, we use MITRA
to classify existing models and identify what has not yet been done in the
research field of trust and reputation. Finally, we review related work in
Sect. 5 and we draw conclusions in Sect. 6.
2 Basic Concepts
In this section, we describe basic concepts and notations that are used in
this article.
2.1 Trust Beliefs/Intentions/Acts and Reputation
Following [MC01], we distinguish three concepts that are often confused in
the literature: the trust belief, the trust intention and the trust act. In the
same spirit as for the BDI architecture [Rao96], raw observations are at first
evaluated and then used to form trust beliefs, which in turn, are used to
build trust intentions. Finally, these intentions can be used as one criterion
in the decision-making process, eventually leading to a trust or distrust act.
Figure 1 provides a schematic view of the trust information chain.
2
observations (obs)
evaluations (eval )
trust beliefs (tb)
trust intentions (ti )
trust acts (ta)
Figure 1: Information Chain in Trust Reasoning.
A basic trust belief, which we denote with tbΓ
θ , reflects the view of an
individual agent θ on the trustworthiness of the agents Γ. In the common
case, the set of agents Γ contains only one agent. However, trust beliefs can
also reflect how an agent θ thinks about a group of agents (see also [Hal02,
FC08]); the agents belonging to the group Γ have to be similar in some regard
and so, experiences with any of them may to some extent be evidence for the
trustworthiness of the group as a whole. For example, consider several agents
being employed by a certain company; in such a situation, the agents can be
judged in their roles as employees of this company, and their characteristics
can be "generalized" to other agents in the same company [FC08].
There is a second type of trust beliefs, which is also called "reputation":
a collective assessment of a group of agents about other agents [MFT+08].
Here, the corresponding trust beliefs are the estimate by an individual of
what could be such a shared opinion of a group of agents Θ about other
agents Γ. We denote this collective trust believe by tbΓ
Θ.
Each trust belief is relating to a certain context, for which it is valid.
The context includes a particular task, and the environmental conditions,
under which the trustee is believed (or not) to successfully carry out this
task on behalf of the trustor. We will discuss this concept of context in
more detail later. Now, for given trustor(s) Θ, trustee(s) Γ and a certain
context, there can only be one unique trust belief. This makes sure that the
trustor cannot believe that trustee(s) Γ are at the same time trustworthy and
untrustworthy concerning a context. Nevertheless, the various kinds of trust
beliefs can still be contradictory. For example, an agent can believe that the
agents that belong to a certain group are usually untrustworthy, but that
a specific agent in this group is well known and believed to be trustworthy.
The reasoning about how to deal with such situations is typically what
occurs in another process, that does not fall into the scope of this paper.
While trust beliefs are solely estimates about the trustworthiness of other
agents, the process of forming trust intentions incorporates also strategic
3
considerations or characteristics of the trustee. For example, although a
trustee believes another agent to be trustworthy, he might still be very
pessimistic about relying on this agent. When forming trust intentions, an
agent actually transforms trust beliefs, which are based on the past behavior
of other agents, into its own intended future behavior towards them. This
typically corresponds to computing the "shadow of the future" [Axe84].
Trust intentions of an agent α towards an agent γ, derived from sets of trust
beliefs, are denoted by tiγ
α.
2.2 Acquisition of Evidence
Two types of information exist that can be used as input for a trust reasoning
process. An agent can make direct observations about and evaluations of
the behavior of other agents, and it can also receive messages from other
agents that contain observations, evaluations or trust beliefs.
2.2.1 Direct Observations and Evaluations
A trust belief about an agent γ is derived from information that provides
evidence for γ's trustworthiness. We call such evidence an evaluation (see
Fig. 1). An evaluation is a subjective interpretation of a set of direct ob-
servations about γ's behavior towards some other agent δ. In other words,
the process of evaluation decides whether the observations are evidence for
(cid:104)δ,γ(cid:105)
trustworthiness of γ. We write eval
to denote an evaluation that is done
α
by agent α concerning the behavior of agent γ towards agent δ. Analogously,
the direct observations made by α on an interaction between γ and δ are
(cid:104)δ,γ(cid:105)
denoted by obs
α
.
2.2.2 Communicated Evidence
Besides directly acquired information, an agent can use information received
by messages from other agents. We introduce the notation (x)α
[θ,...,β] for a
message x that is sent by agent β to agent α through a path of transmit-
ters β (direct sender to α), . . . , θ (initial sender). We call such a message
communicated evidence. To ensure the correctness of the indicated path, ei-
ther the final receiver can assess the correctness of this chain of transmitters
(e.g., by spot-checking agents on the path and asking them whether they
have sent the message as is), or mechanisms are put in place to prove that
the message has indeed taken the indicated path and which intermediary
has made which modifications (by means of cryptography, e.g., public-key
infrastructure and signing).
Direct observations that an agent makes on its own can only be incorrect
if the sensors with which the observations were made are faulty. Commu-
nicated evidence additionally can be wrong when the sender is dishonest or
incompetent.
4
2.3 Context and Uncertainty
Each observation, evaluation, trust belief or trust intention is in general only
valuable if two pieces of information are attached to it:
1. the context to which the information applies;
2. a measure of uncertainty of the information itself.
The concept of context is a vital part for evidence-based trust reasoning.
[MC96] give the example that "one would trust one's doctor to diagnose and
treat one's illness, but would generally not trust the doctor to fly one on a
commercial airplane". If some direct observations are made in a certain con-
text, then this context should be annotated also to the evaluations (and the
trust beliefs and trust intentions) that are based on these observations. As
a result, context annotations should be propagated through the information
chain shown in Fig. 1.
A measure of uncertainty is needed to express how uncertain a piece of in-
formation is believed to be. This is especially important when a trust model
incorporates communicated evidence that may be biased or wrong. Simi-
larly to the context information, uncertainty should be propagated through
the whole information chain.
3 The Meta-Model
In this section we present MITRA, illustrated in Fig. 2. This meta-model
organizes the different ways of how the information chain of Fig. 1 can
be realized in an agent α's trust model. The agent can send all pieces of
information that occur in this information chain (the square boxes in the
Figure) to other agents; for the sake of clarity, we did not illustrate such
actions in Fig. 2.
On the top level, MITRA divides the trust modeling process into four
consecutive sub-processes:
1. observation;
2. evaluation;
3. fusion;
4. decision-making.
Let us first exemplify these sub-processes from the perspective of an
agent α, which reasons about its trust in an agent γ. In the observation
process, agent α tries to collect any form of evidence it can get to assess γ's
trustworthiness. Agent α judges the direct observations (either its own or
the communicated ones) about γ's behavior in the evaluation process. In the
5
Figure 2: Structure of MITRA for an agent α. For the sake of clarity what
is not shown in this illustration: agent α can decide to send any piece of
information occurring in the illustration to other agents.
fusion process, α can use its own evaluations, and the filtered evaluations
received from other agents, to form its trust beliefs about γ and an image of
other agents' trust beliefs about γ. Finally, in the decision-making process,
α builds its trust intentions and applies them in the respective situations.
6
actuatorsAgent αderiving trust intentionsmaking decisionsevaluating (θ)OBSERVATIONEVALUATIONFUSIONDECISION MAKINGobservingevaluating (α)forming others'trust beliefs (θ)...decision-contextavailabilityforming owntrust beliefs (α)forming collectivetrust beliefs (Θ)sensorscommunicated evidencedirectobs.information flowprocesscredibility filterinformationLegendpersonality filtersubjectivity filterinfluence Below, we describe each sub-process in greater detail and then have a closer
look at the issue of context-sensitivity.
3.1 Observation
During observation, agent α uses its sensors to capture information about
other agents and the environment.
In network settings for instance, sen-
sors could be network cards, that can receive or overhear packets from the
network.
In the general case, the process of direct observation is subjec-
tive, since different agents may use sensors that differ in certain respects,
for instance in quality.
As mentioned earlier, information can result from direct observation of
the environment or be received as communicated evidence. For an observa-
tion obs
, α may be the same agent as δ; in this case, α observes its own
interaction with some other agent γ, otherwise α is observing the interac-
tions of other agents.
(cid:104)δ,γ(cid:105)
α
Because communicated evidence is received from other agents, α needs to
filter the information. Communicated evidence might be incorrect for differ-
ent reasons [CP02, BS08]. First, the communicator can intentionally provide
wrong information, i.e., lie. For instance, the communicator might want to
increase its own reputation or the reputation of an acquaintance; this is
usually called "misleading propaganda". It can also try to decrease the rep-
utation of another competing agent, which is usually called "defamation".
Secondly, the communicator can provide information that is not wrong but
leads the receiver intentionally to wrong conclusions. For instance, it can
hide information, give partial information or give out-of-context informa-
tion. Finally, the communicator can unwittingly communicate untrue facts.
Note that agent β, the last agent who sent the information, needs not to be
the originator of the information [CP02], and therefore might alter the infor-
mation if it is not signed. The credibility filter should take all these aspects
into account and filter the information according to how much the sources
of this information are intended to be trusted, more precisely, whether they
are decided to be trusted. In Figure 2, this is indicated by the "influence"-
arrows from "making decisions" to the credibility filters, which exactly try
to filter out information that does not seem to be credible.
3.2 Evaluation
During the process of evaluation, an agent evaluates sets of direct obser-
vations about the behavior of other agents. Such an evaluation estimates
whether the set of observations provides evidence for the agent in ques-
tion being trustworthy or untrustworthy. At this stage, it is not yet decided
whether an agent is actually believed to be trustworthy or not; sets of obser-
vations are only examined for their significance in respect to an agent's trust-
7
worthiness. In many models, this is done by comparing the actual behavior
of an agent to what its behavior was expected to be like. This expected be-
havior can, for instance, be determined by formal contracts [Sab02] or social
norms [COTZ00, VM10]. Various representations are used to describe eval-
uations: binary (e.g., [SS02]), more fine-grained discrete (e.g., {−1, 0, +1}
in eBay [eBa09]), or continuous assessments (e.g., [Sab02, HJS04, VM10]).
As the sets of norms and established (implicit) contracts can be subjec-
tive, an evaluation can be subjective too. Therefore, different agents might
contradictorily interpret the same observations as evidence for trustworthy
and untrustworthy behavior. This is evident in the case of eBay [eBa09]
where each human does the evaluation along his own criteria. As a con-
sequence, an agent can try to emulate how other agents would evaluate
observations, to eventually emulate their trust beliefs. Therefore, MITRA
contains two different ways of evaluation (see Fig. 2): using α's criteria,
which results in evaluations eval
; and the way α thinks θ would do the
.
evaluation, which results in eval
(cid:104)δ,γ(cid:105)
α
(cid:104)δ,γ(cid:105)
θ
For evaluation, it can be vital to know about the causality behind what
happened. Indeed, if the causal relationships are not clear to the evaluating
agent, it can wrongly evaluate a failure to be evidence for untrustworthiness,
although it is actually an excusable failure -- or vice versa [CF00, SE07].
3.3 Fusion
In a third step, the different evaluations are fused into trust beliefs. Trust
beliefs can be formed based on evaluations of individual agents or groups
of agents, and can be about individuals or groups. Figure 2 shows that an
agent α can fuse evaluations to get its own trust beliefs tbΓ
α; or to emulate
the trust modeling of another agent θ, in order to estimate this agent's θ's
trust beliefs tbΓ
received from
another agent β, it first needs to filter out the "incompatible" subjectivity
inherent in the evaluations: For evaluation, β may have applied criteria that
α does not agree with, or -- when α is emulating the trust beliefs of θ -- where
α thinks that θ would not agree with. In the figure, the check for the match
of the applied criteria, and the potential adjustment of the evaluations, is
named subjectivity filtering.
(cid:104)··· (cid:105)
θ . In any case, if α uses evaluations eval
β
The emulation of another agent θ's trust beliefs is for instance needed
when forming reputation, i.e., trust beliefs that a certain group of agents Θ
(with θ ∈ Θ) would associate to a given trustee. For this, also trust beliefs
received from other agents can be incorporated. A subjectivity filter should
not be applied to these received trust beliefs, because reputation reflects the
subjectivity of different agents. Still, a credibility filter is applied to avoid
a biased reputation estimate.
Furthermore, if an agent forms trust beliefs based on interactions where
itself was not involved, it has to account for the relation between the inter-
8
(cid:104)δ,γ(cid:105)
θ
acting agents. Assume, an agent α receives from another agent the evalua-
tion eval
, where all four agents α, θ, δ, and γ are distinct agents. If α
wants to reason about δ's trustworthiness towards itself, then it first needs
to apply a personality filter. Here, information is filtered out that contains
no evidence for the behavior of δ towards α, because it is specific for in-
teractions between δ and γ. For example, imagine that the evaluation is
about the behavior of a mother (δ) towards her child (γ), and she behaved
very trustworthy. Then this behavior does not say much about how the
mother will behave towards another unrelated person (e.g., α). This shows
that trustworthiness is directed, and that this direction has to be taken into
account in trust reasoning.
Whenever fusing sets of evaluations or trust beliefs into a single trust
belief, it is particularly important to account for the "correlated evidence"
problem [Pea88]. This problem arises when different evaluations or trust
beliefs are based on the same observed interactions of agents.
In other
words, the evidence expressed by the evaluations/trust beliefs "overlaps".
If the reasoning agent is not aware of this overlapping, certain parts of the
evidence will wrongly be amplified and the resulting trust belief be biased.
Many different approaches for fusing evaluations into trust beliefs (e.g.,
[WS07, TPJL06a, RRRJ07]), and trust beliefs into community models (e.g.,
[KSGM03, SFR99]) have been proposed in the literature. Most importantly,
each trust belief should incorporate the two properties mentioned in Sect.
2.3: the uncertainty, i.e., how strong the belief is, and in which context(s)
it applies.
3.4 Decision-Making
The last component in MITRA is the decision-making process, which con-
sists of two steps:
1. fix trust intentions based on trust beliefs;
2. apply the trust intentions to make the final decision to act (or not) in
trust.
In the first step, a set of individual and/or collective trust beliefs are used
to derive one or several trust intentions for different contexts. Trust beliefs
can, for instance, be aggregated into trust intentions by simply averaging
over them, or by taking the most "pessimistic" or "optimistic" trust belief,
etc. But also, it is possible to ignore available negative trust beliefs about
another agent γ, and to decide to act in trust with γ in order to give this
agent the opportunity to rethink its behavior [FC04]. This "advance" in
trust, which can in some cases be forgiveness, accounts for the dynamics of
trust such as "trust begets trust" [BE89].
9
A trust intention tiγ
α has to reflect two things:
in which situations α
actually intends to act in trust with γ (context), and how strong the intention
is (uncertainty). Trust intentions with these two properties can be used in
decision-making in the same way as other criteria. Although many trust
and reputation models from the literature do not separate the derivation
of trust intentions from the process of decision-making, we strongly argue
for a separation of the two processes. The reason is that the derivation of
trust intentions is specific to trust research, while a trust intention plays in
decision making the role of a context-sensitive criterion in the same way as
many other criteria (like availability of a potential partner). This problem is
however studied in depth in a research area called Multiple Criteria Decision
Making (MCDM) [Kal06] and is not specific to trust.
3.5 Context in MITRA
In general, everything that can impact the behavior of a trustee and is not
part of the trustee itself, is said to belong to the context [Dey00]. The more
of the available context information is considered by a trust and reputation
model, the better the final decision can be. We propose to arrange the
different facets of context into five classes:
1. time (points in time or time intervals);
2. external conditions:
(a) physical conditions,
(b) laws/norms,
(c) other agents nearby;
3. type of the delegated task;
4. contract;
5. information source.
The information source is the one that provides the information, on
which the reasoning is based, e.g., a sensor or another agent. This context
facet is a special case, because it is not linked to a trustee's behavior. It
is important though, since it allows an agent to have several trust beliefs
about the same trustee, based on different information sources.
Two different contexts can be distinguished: the one in which the in-
formation was taken, and the one in which a decision is made (decision
context). To fuse information that comes from different contexts, or use it
in decision-making for different contexts, an agent needs to know how simi-
lar the different contexts are. In the literature, many ways for representing
context information, and similarities between contexts, have been proposed.
10
[KR03] represent context relations in form of a weighted directed graph.
[S¸Y07] use ontologies and a set of rules to represent context information.
In a more general way, [RP07] represent context information as points in a
multi-dimensional space, where each dimension represents one characteris-
tic of the context (e.g., the point in time, the dollar exchange rate, etc.).
The distance between two points in the context space, which is determined
by some distance metric, states the similarity of the two contexts. These
approaches have in common that an agent was given some similarity metric
about different contexts in advance. MITRA does not assume that, and so,
if the similarity metric cannot be known in advance, an agent needs to learn
the metric.
At which stage information that belongs to different contexts is eventu-
ally fused, is up to the concrete model. However, since during this fusion
process some information is lost, it should be done as late as possible in
the information chain. As a consequence, that principle should generally be
used when processing information.
4 Organizing Existing Models
In this section, we exemplify how existing trust and reputation (T&R) mod-
els can be classified by means of MITRA. The approach we take here is to
investigate which types of data are used in the considered trust models. For
a small selection of trust models, this is shown in tables 1 and 2. Table 1 lists
types of data that belong to the observation process. Table 2 lists the re-
maining data types, i.e. those in the evaluation, fusion and decision-making
processes. A check-mark indicates whether a trust model accounts for the
respective kind of data. Still, not every model accounts for the context and
uncertainty of the processed information.
Additionally to this classification approach, T&R models can also be
described by specifying:
1. which filters are applied,
2. where uncertainty is considered, and
3. which facets of context are accounted for at which stage.
Since the focus of this article lies on the meta model MITRA itself, we
leave it to future work to provide an extensive and comprehensive classifica-
tion using MITRA. However, as it can already be seen in the resulting tables,
a clear picture emerges of what has been done (columns with check-marks)
and what needs to be done in the T&R modeling domain -- for example,
there is no model with no empty columns, i.e., that uses all available types
(cid:104)δ,γ(cid:105)
of information. Furthermore, the column eval
is always empty. None of
θ
11
]
.
.
.
,
β
α [
(cid:1)
ΓΘ
b
t
(cid:0)
]
.
.
.
,
β
α [
(cid:1)
Γθ
b
t
(cid:0)
d
b
(cid:88)
b
(cid:88)
b
(cid:88)
b
(cid:88)
b
(cid:88)
b
a
(cid:88)
b
(cid:88)
b
(cid:88)
b
(cid:88)
b
(cid:88)
b
(cid:88)
(cid:88)
θ
(cid:88)
c
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
]
.
.
.
,
β
α [
n
o
i
t
a
v
r
e
s
b
O
.
)
n
o
i
(cid:17)
t
a
(cid:105)
v
γ
,
r
δ
e
(cid:104)
s
b
O
(cid:16)
(
A
R
T
I
M
(cid:17)
h
(cid:105)
γ
t
,
i
δ
w
(cid:104)
n
o
(cid:16)
i
t
a
c
fi
i
s
s
a
l
C
:
1
e
l
b
a
T
l
a
v
e
]
.
.
.
,
β
α [
θ
s
b
o
>
γ
,
δ
<
α
s
b
o
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
a
(cid:88)
]
4
9
r
a
M
[
h
s
r
a
M
]
9
9
R
F
S
[
.
l
a
t
e
o
l
l
i
h
c
S
]
9
9
P
M
Z
[
]
9
9
P
M
Z
[
s
o
t
s
i
H
s
a
r
o
p
S
]
2
0
I
J
[
m
e
t
s
y
S
.
p
e
R
-
a
t
e
B
]
2
0
S
S
[
a
j
j
a
S
d
n
a
n
e
S
]
3
0
G
S
K
[
t
s
u
r
T
n
e
g
E
i
]
3
0
M
S
[
t
e
r
g
e
R
]
3
0
+
S
G
C
[
e
r
u
c
e
S
]
3
0
V
W
[
a
v
e
l
i
s
s
a
V
d
n
a
g
n
a
W
]
4
0
L
X
[
t
s
u
r
T
r
e
e
P
]
6
0
M
C
[
i
s
e
l
o
s
u
M
d
n
a
a
r
p
a
C
]
b
6
0
L
J
P
T
[
s
o
v
a
r
T
]
7
0
J
R
R
R
[
.
l
a
t
e
e
c
e
e
R
]
7
0
Y
S¸
[
m
u
l
o
Y
d
n
a
y
o
s
n
e
S¸
]
7
0
S
W
[
h
g
n
i
S
d
n
a
g
n
a
W
]
8
0
C
F
[
i
h
c
n
a
r
f
l
e
t
s
a
C
d
n
a
e
n
o
c
l
a
F
]
0
1
M
V
[
r
a
L
i
]
0
1
C
M
V
[
.
l
a
t
e
s
i
z
t
a
i
g
o
V
1
=
Γ
f
o
e
s
a
c
e
h
t
g
n
i
r
e
d
i
s
n
o
c
y
l
n
o
b
y
l
t
i
c
i
l
p
m
i
a
r
e
t
n
e
c
g
n
i
t
a
r
l
a
r
t
n
e
c
a
o
t
d
e
t
t
i
m
b
u
s
c
r
e
t
n
e
c
g
n
i
t
a
r
l
a
r
t
n
e
c
a
y
b
d
e
d
i
v
o
r
p
d
12
.
)
g
n
i
k
a
M
-
n
o
i
s
i
c
e
D
,
n
o
i
s
u
F
,
n
o
i
t
a
u
l
a
v
E
(
A
R
T
I
M
h
t
i
w
n
o
i
t
a
c
fi
i
s
s
a
l
C
:
2
e
l
b
a
T
γα
i
t
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
a
(cid:88)
a
(cid:88)
g
n
i
k
a
M
-
n
o
i
s
i
c
e
D
n
o
i
s
u
F
ΓΘ
b
t
Γθ
b
t
Γα
b
t
a
(cid:88)
a
(cid:88)
a
(cid:88)
b
a
(cid:88)
a
a
(cid:88)
(cid:88)
a
(cid:88)
a
(cid:88)
a
(cid:88)
a
a
a
a
a
a
a
a
a
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
a
a
(cid:88)
(cid:88)
(cid:105)
γ
,
δ
(cid:104)
θ
l
a
v
e
n
o
i
t
a
u
l
a
v
E
(cid:105)
γ
,
δ
(cid:104)
α
l
a
v
e
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
(cid:88)
]
4
9
r
a
M
[
h
s
r
a
M
]
9
9
R
F
S
[
.
l
a
t
e
o
l
l
i
h
c
S
]
9
9
P
M
Z
[
]
9
9
P
M
Z
[
s
o
t
s
i
H
s
a
r
o
p
S
]
2
0
I
J
[
m
e
t
s
y
S
.
p
e
R
-
a
t
e
B
]
2
0
S
S
[
a
j
j
a
S
d
n
a
n
e
S
]
3
0
G
S
K
[
t
s
u
r
T
n
e
g
E
i
]
3
0
M
S
[
t
e
r
g
e
R
]
3
0
+
S
G
C
[
e
r
u
c
e
S
]
3
0
V
W
[
a
v
e
l
i
s
s
a
V
d
n
a
g
n
a
W
]
4
0
L
X
[
t
s
u
r
T
r
e
e
P
]
6
0
M
C
[
i
s
e
l
o
s
u
M
d
n
a
a
r
p
a
C
]
b
6
0
L
J
P
T
[
s
o
v
a
r
T
]
7
0
J
R
R
R
[
.
l
a
t
e
e
c
e
e
R
]
7
0
Y
S¸
[
m
u
l
o
Y
d
n
a
y
o
s
n
e
S¸
]
7
0
S
W
[
h
g
n
i
S
d
n
a
g
n
a
W
]
8
0
C
F
[
i
h
c
n
a
r
f
l
e
t
s
a
C
d
n
a
e
n
o
c
l
a
F
]
0
1
M
V
[
r
a
L
i
]
0
1
C
M
V
[
.
l
a
t
e
s
i
z
t
a
i
g
o
V
r
e
t
n
e
c
g
n
i
t
a
r
l
a
r
t
n
e
c
y
b
d
e
m
r
o
f
r
e
p
n
o
i
t
a
t
u
p
m
o
c
b
1
=
Γ
f
o
e
s
a
c
e
h
t
g
n
i
r
e
d
i
s
n
o
c
y
l
n
o
a
13
the here considered trust models1 uses the intermediate step of simulating
the evaluation of another agent to eventually form a collective trust belief
(reputation). However, we believe that this is something what humans do
regularly; for example in form of questions like "What would my mother
think about his behavior?" or "How would my best friend judge this agent's
behavior?". This issue deserves more attention as it makes it possible to
evaluate the behavior of other agents in cases where an own opinion on
their behavior is lacking.
5 Related Work
Numerous survey or overview papers on trust and reputation models have
been published. Many works start with a basic classification and then
enumerate and describe existing models [SMS05, AG07, JIB07, RHJ04,
MHM02]. Opposed to that, we tried in our work to use a rigorous method-
ology for the organization of the state of the art: extract the core structure
of prominent trust and reputation models, in order to get clear and simple
differentiation criteria for the models of the literature.
[KBR05] proposed a generic trust model that integrates several existing
trust models. They show how to map each of these models to their model.
However, they focus on the fusion of evaluations, whereas we examined the
overall structure of trust and reputation modeling.
[CS05] define a functional ontology of reputation, that is used in different
systems [VCSB07, NBSV08] in order to help agents using different trust
and reputation models inter-operate. These approaches try to cope with a
situation where different models exist, whereas we try to propose a unified
meta-model of trust and reputation.
6 Conclusion
In this article, we presented MITRA, a meta-model for trust and reputation.
The model structures many essential concepts found in the literature on
evidence-based T&R models. The simple and generic structure of MITRA
makes it suitable both for experts to organize their models in a common
way, and for newcomers to easily enter the domain. Although there is the
possibility that a specific T&R model does not comply with MITRA, the
fact that the latter was derived from the study of many existing models
argues for its comprehensiveness, and we believe that its modularity should
make it simple to modify in order to encompass new elements.
Finally, by using MITRA, we classified existing T&R models from the
literature. This classification revealed which kinds of information are com-
1To our knowledge there is no such trust model in the literature.
14
monly used by these models, and which kinds of information are often ne-
glected. In this way, we found that none of the considered models actually
tries to emulate another agent's evaluation of a trustee; this emulation would
make it possible to form the reputation of an agent in a new way.
References
[AG07]
D. Artz and Y. Gil, A survey of trust in computer science and
the semantic web, Web Semant. 5 (2007), no. 2, 58 -- 71.
[Axe84]
R. Axelrod, The evolution of cooperation, Basic Books, 1984.
[BE89]
[BS08]
[CF00]
J. L. Bradach and R. G. Eccles, Price, authority, and trust:
From ideal types to plural forms, Annu. Rev. Sociol. 15 (1989),
97 -- 118.
P. Barreira-Avegliano and J. S. Sichman, Reputation based part-
nership formation: some experiments using the repart simulator,
Proc. of the Wksh. on Trust in Agent Societies (at AAMAS'08),
2008.
C. Castelfranchi and R. Falcone, Trust is much more than sub-
jective probability: Mental components and sources of trust,
Proc. of the 33rd Hawaii
Int. Conf. on System Sciences
(HICSS'00), 2000.
[CGS+03] V. Cahill, E. Gray, J.-M. Seigneur, C. D. Jensen, Y. Chen,
B. Shand, N. Dimmock, A. Twigg, J. Bacon, C. English,
W. Wagealla, S. Terzis, P. Nixon, G. di Marzo Serugendo,
C. Bryce, M. Carbone, K. Krukow, and M. Nielsen, Using trust
for secure collaboration in uncertain environments, IEEE Per-
vasive Computing 02 (2003), no. 3, 52 -- 61.
[CM06]
L. Capra and M. Musolesi, Autonomic trust prediction for per-
vasive systems, Proc. of the 20th Int. Conf. on Advanced Infor-
mation Networking and Applications (Vol. 2) (AINA'06), 2006,
pp. 481 -- 488.
[COTZ00] C. Castelfranchi, A. Omicini, R. Tolksdorf, and F. Zambonelli,
Engineering social order, Proc. of Engineering Societies in the
Agents World (ESAW'00), LNCS, vol. 1972, 2000, pp. 1 -- 18.
[CP02]
R. Conte and M. Paolucci, Reputation in artificial societies. so-
cial beliefs for social order, Kluwer Academic Publishers, 2002.
15
[CS05]
[Dey00]
[eBa09]
[FC04]
[FC08]
[Hal02]
[HJS04]
[JI02]
[JIB07]
[Kal06]
S. Casare and J. Sichman, Towards a functional ontology of rep-
utation, Proc. of Autonomous Agents and Multi-Agent Systems
(AAMAS'05), 2005, pp. 505 -- 511.
A. K. Dey, Providing architectural support for building context-
aware application, Ph.D. thesis, College of Computing, Georgia
Institute of Technology, 2000.
eBay, eBay auction website, accessed April, 15, 2009, http:
//www.ebay.com.
R. Falcone and C. Castelfranchi, Trust dynamics: How trust is
influenced by direct experiences and by trust itself, Proc. of Au-
tonomous Agents and Multi-Agent Systems (AAMAS'04), 2004,
pp. 740 -- 747.
Rino Falcone and Cristiano Castelfranchi, Generalizing trust:
Inferencing trustworthiness from categories, Trust in Agent So-
cieties, LNCS (LNAI), vol. 5396, 2008, pp. 65 -- 80.
David Hales, Group reputation supports beneficent norms, Jour-
nal of Artificial Societies and Social Simulation 5 (2002), no. 4.
T. D. Huynh, N. R. Jennings, and N. Shadbolt, FIRE: an in-
tegrated trust and reputation model for open multi-agent sys-
tems, Proc. of the 16th European Conf. on Artificial Intelligence
(ECAI'04), 2004.
A. Jøsang and R. Ismail, The beta reputation system, Proc. of
the 15th Bled Conf. on Electronic Commerce, 2002, pp. 324 -- 337.
A. Jøsang, R. Ismail, and C. Boyd, A survey of trust and repu-
tation systems for online service provision, Decis. Support Syst.
43 (2007), no. 2, 618 -- 644.
I. Kaliszewski, Soft computing for complex multiple criteria de-
cision making, International Series in Operations Research &
Management Science, vol. 85, Springer Verlag, 2006.
[KBR05] M. Kinateder, E. Baschny, and K. Rothermel, Towards a generic
trust model -- comparison of various trust update algorithms,
Proc. of the 3rd Int. Conf. on Trust Management (iTrust'05),
2005, pp. 177 -- 192.
[KR03]
M. Kinateder and K. Rothermel, Architecture and algorithms
for a distributed reputation system, Proc. of the 1st Int. Conf.
on Trust Management (iTrust'03), 2003, pp. 1 -- 16.
16
[KSG03]
S.D. Kamvar, M.T. Schlosser, and H. Garcia-Molina, The eigen-
trust algorithm for reputation management in p2p networks,
Proc. of the 12th Int. World Wide Web Conference, 2003,
pp. 640 -- 651.
[KSGM03] S. D. Kamvar, M. T. Schlosser, and H. Garcia-Molina, The
eigentrust algorithm for reputation management in p2p net-
works, Proc. of the 12th Int. World Wide Web Conf. (WWW'03),
2003, pp. 640 -- 651.
[Mar94]
[MC96]
[MC01]
S. P. Marsh, Formalising trust as a computational concept, Ph.D.
thesis, Department of Computing Science and Mathematics,
University of Stirling, 1994.
D.H. McKnight and N.L. Chervany, The meanings of trust, Tech.
report, University of Minnesota, 1996.
, Trust and distrust definitions: One bite at a time, Proc.
of the 4th Wksh. on Deception, Fraud, and Trust in Agent So-
cieties (at AAMAS'01), 2001, pp. 27 -- 54.
[MFT+08] G. A. Miller, C. Fellbaum, R. Tengi, P. Wakefield, H. Langone,
and B. R. Haskell, Wordnet 3.0, "reputation" definition #3,
Princeton University, accessed 12/08 2008.
[MHM02] L. Mui, A. Halberstadt, and M. Mohtashemi, Notions of repu-
tation in multi-agent systems: A review, Proc. of Autonomous
Agents and Multi-Agent Systems (AAMAS'02), 2002, pp. 280 --
287.
[NBSV08] L. G. Nardin, A. A. F. Brandao, J.S. Simao Sichman, and Lau-
rent Vercouter, A service-oriented architecture to support agent
reputation models interoperability, Proc. of WONTO'08, 2008.
[Pea88]
[Rao96]
J. Pearl, Probabilistic reasoning in intelligent systems: networks
of plausible inference, Morgan Kaufmann Publishers Inc., 1988.
A. S. Rao, AgentSpeak(L): BDI agents speak out in a logi-
cal computable language, Proc. of the 7th Wksh. on Modelling
Autonomous Agents in a Multi-Agent World (MAAMAW'96),
LNAI, 1996, pp. 42-55.
[RHJ04]
S. D. Ramchurn, T. D. Huynh, and N. R. Jennings, Trust in
multi-agent systems, Knowl. Eng. Rev. 19 (2004), no. 1, 1 -- 25.
[RP07]
M. Reh´ak and M. Pechoucek, Trust modeling with context repre-
sentation and generalized identities, Proc. of the 11th Int. Wksh.
on Cooperative Information Agents (CIA'07), 2007, pp. 298 -- 312.
17
[RRRJ07] Steven Reece, Stephen Roberts, Alex Rogers, and Nicholas R.
Jennings, A multi-dimensional trust model for heterogeneous
contract observations, Proc. of the 22nd Conf. on Artificial In-
telligence (AAAI'07), AAAI Press, 2007, pp. 128 -- 135.
[Sab02]
[SE07]
J. Sabater-Mir, Trust and reputation for agent societies, Ph.D.
thesis, Artificial Intelligence Research Institute, Universitat
Aut`onoma de Barcelona, Spain, 2002.
E. Staab and T. Engel, Formalizing excusableness of failures in
multi-agent systems, Proc. of the 10th Pacific Rim International
Wksh. on Multi-Agents (PRIMA'07), 2007, pp. 124 -- 135.
[SFR99] M. Schillo, P. Funk, and M. Rovatsos, Who can you trust:
Dealing with deception, Proc. of the 2nd Wksh. on Deception,
Fraud, and Trust in Agent Societies (at AA'99) (C. Castel-
franchi, Y. Tan, R. Falcone, and B. S. Firozabadi, eds.), May
1999, pp. 81 -- 94.
[SM03]
J. Sabater-Mir, Trust and reputation for agent societies, Ph.D.
thesis, Institut d'Investigaci´o en Intel·lig`encia Artificial, 2003.
[SMS05]
J. Sabater-Mir and C. Sierra, Review on computational trust and
reputation models, Artif. Intell. Rev. 24 (2005), no. 1, 33 -- 60.
[SS02]
[S¸Y07]
S. Sen and N. Sajja, Robustness of reputation-based trust:
boolean case, Proc. of Autonomous Agents and Multi-Agent Sys-
tems (AAMAS'02), 2002, pp. 288 -- 293.
M. S¸ensoy and P. Yolum, Ontology-based service representation
and selection, IEEE Trans. Knowl. Data Eng. 19 (2007), no. 8,
1102 -- 1115.
[TPJL06a] W. T. L. Teacy, J. Patel, N. R. Jennings, and M. Luck,
TRAVOS: Trust and reputation in the context of inaccurate in-
formation sources, Auton. Agents Multi-Agent Syst. 12 (2006),
no. 2, 183 -- 198.
[TPJL06b] W.T.L. Teacy, J. Patel, N.R. Jennings, and M. Luck, Travos:
Trust and reputation in the context of inaccurate information
sources, Auton. Agents Multi-Agent Syst. 12 (2006), no. 2, 183 --
198.
[VCSB07] L. Vercouter, S. J. Casare, J. S. Sichman, and A. A. F. S.
Brandao, An experience on reputation models interoperability
based on a functional ontology, Proc. of the 20th Int. Joint Conf.
on Artificial Intelligence (IJCAI'07), 2007, pp. 617 -- 622.
18
[VM10]
L. Vercouter and G. Muller, L.I.A.R.: Achieving social control
in open and decentralised multi-agent systems, Applied Artificial
Intelligence (2010).
[VMC10] George Vogiatzis, Ian MacGillivray, and Maria Chli, A prob-
abilistic model for trust and reputation, Proc. of 9th Int. Conf.
on Autonomous Agents and Multi-Agent Systems (AAMAS'10),
2010, pp. 225 -- 232.
[WS07]
[WV03]
Yonghong Wang and Munindar P. Singh, Formal trust model
for multiagent systems, Proc. of the 20th Int. Joint Conf. on
Artificial Intelligence (IJCAI'07), 2007, pp. 1551 -- 1556.
Yao Wang and Julita Vassileva, Bayesian network-based trust
model, Proceedings of the 2003 IEEE/WIC International Con-
ference on Web Intelligence (WI'03) (Washington, DC, USA),
IEEE Computer Society, 2003, p. 372.
[XL04]
Li Xiong and Ling Liu, Peertrust: Supporting reputation-
based trust for peer-to-peer electronic communities, IEEE Trans.
Knowl. Data Eng. 16 (2004), no. 7, 843 -- 857.
[ZMP99] G. Zacharia, A. Moukas, and P. Paes, Collaborative reputa-
tion mechanisms in electronic marketplaces, Proceedings of the
Thirty-second Annual Hawaii International Conference on Sys-
tem Sciences-Volume 8, IEEE Computer Society, 1999.
19
|
0705.3050 | 1 | 0705 | 2007-05-22T14:26:05 | A competitive multi-agent model of interbank payment systems | [
"cs.MA"
] | We develop a dynamic multi-agent model of an interbank payment system where banks choose their level of available funds on the basis of private payoff maximisation. The model consists of the repetition of a simultaneous move stage game with incomplete information, incomplete monitoring, and stochastic payoffs. Adaptation takes place with bayesian updating, with banks maximizing immediate payoffs. We carry out numerical simulations to solve the model and investigate two special scenarios: an operational incident and exogenous throughput guidelines for payment submission. We find that the demand for intraday credit is an S-shaped function of the cost ratio between intraday credit costs and the costs associated with delaying payments. We also find that the demand for liquidity is increased both under operational incidents and in the presence of effective throughput guidelines. | cs.MA | cs |
A competitive multi-agent model of interbank
payment systems
Marco Galbiati† and Kimmo Soramaki‡
14 May 2007
Abstract
We develop a dynamic multi-agent model of an interbank payment
system where banks choose their level of available funds on the basis of
private payoff maximisation. The model consists of the repetition of a
simultaneous move stage game with incomplete information, incomplete
monitoring, and stochastic payoffs. Adaptation takes place with bayesian
updating, with banks maximizing immediate payoffs. We carry out nu-
merical simulations to solve the model and investigate two special sce-
narios: an operational incident and exogenous throughput guidelines for
payment submission. We find that the demand for intraday credit is an
S-shaped function of the cost ratio between intraday credit costs and the
costs associated with delaying payments. We also find that the demand for
liquidity is increased both under operational incidents and in the presence
of effective throughput guidelines.
† Bank of England. E-mail: [email protected].
The views expressed in this paper are those of the authors, and not nec-
essarily those of the Bank of England.
‡ Helsinki Univ. of Technology. E-mail: [email protected].
The authors thank W. Beyeler, M. Manning, S. Millard, M. Willison
for important insights. Useful comments came from the participants of a
number of seminars: FS seminar (Bank of England, May 3, 2007), Central
Bank Policy Workshop (Basel, March 13, 2007), University of Liverpool,
Seimnar at the Computer Science Department (Liverpool, April 24, 2007).
The usual disclaimer applies.
1
1 Introduction
Virtually all economic activity is facilitated by transfers of claims towards public
or private financial institutions. The settlement of claims between banks takes to
a large extent place at the central bank, in central bank money. These interbank
payment systems transfer vast amounts of funds, and their smooth operation is
critical for the functioning of the whole financial system. In 2004, the annual
value of interbank payments made in the European TARGET was around $552
trillion, in the US Fedwire system $470 trillion, and in the UK CHAPS $59
trillion - tens of times the value of their respective gross domestic products (BIS
2006). These transfers originate from customer requests, and from the banks'
proprietary operations in e.g.
foreign exchange, securities and the interbank
money market. The sheer size of these transfers, and their centrality for the
functioning of a number of markets, make the mechanisms that regulate these
fluxes and the incentives that generate them interesting to policy makers and
regulators.
At present most payment systems work on a real-time gross settlement
(RTGS) or equivalent modality.
In RTGS payments are settled continuously
and individually throughout the day with immediate finality. To cover the pay-
ments banks generally use their reserve balances, access intraday credit from
the central bank or use incoming funds from payments from other banks. The
first two sources carry an (opportunity) cost which gives banks incentives to
economize on their use. We call these funds liquidity. The third source, on the
other hand, is dependent by the liquidity decisions of other banks. The less
liquidity a bank commits for settlement, the more dependent it is from incom-
ing payments - and may thus need to delay its own payments until these funds
arrive, causing the receivers of its payments to receive funds later. If also delays
are costly, each bank faces a trade-off between liquidity costs and delay costs.
Both aspects are dependent on the banks own liquidity decision, but the latter
is also dependent on the liquidity decisions by other banks.
This paper develops a dynamic model to study this trade-off. The model
consists of a sequence of independent settlement days where a set of homogenous
banks make payments to each other. Each of these days is a simultaneous-move
game (or a stage game) in which banks choose their level of liquidity for payment
processing. At the end of the day they receive a stochastic payoff determined
by the amount of liquidity they committed and delays they experienced. Due
to the nature of the settlement process, the payoff function is a random variable
unknown to the banks. In this context, a reasonable assumption is that banks
use heuristic, bounded-rational like rules to adapt their behaviour over time.
Hence, we simulate a learning process taking place over many days, until banks
settle down in equilibria. We are interested in the properties of the equilib-
ria in aggregate terms, i.e.
in the behaviour of the system as the product of
independent, single agents' private payoff maximization.
Given its game-theoretic approach, this paper is related to recent work by
Angelini (1998), Bech and Garratt (2003, 2006), Buckle and Campbell (2003)
and Willison (2004). These study various "liquidity management games" with
2
few (typically, two) agents and few (typically, three) periods. There, however,
the payoff function is common knowledge. Due to the complex mechanics taking
place in real payment systems this is likely to be unrealistic. Recent work
by Beyeler et al.
(2007) on the relationship between instruction arrival and
payment settlement in a similar setting shows that with low liquidity, payment
settlement gets coupled across the network and is governed by the dynamics
of the queue - and largely unpredictable when a large number of payments
are made. The present paper makes an effort to model this complexity; in a
similar spirit, it also considers a large number of banks, which settle payments
in a continuous-time day, and which interact over a long sequence of settlement
days.
Recently, a growing literature has used simulation techniques to investigate
the effects e.g. of failures in complex payment systems (see eg. BoE (2004),
Leinonen (2005), Devriese and Mitchell (2005)). These studies generally use
historical payment data and simulate banks' risk exposures under alternative
scenarios, or ways to improve liquidity efficiency of the systems. The shortcom-
ing of this approach has been that the behaviour of banks is not endogenously
determined. It is either assumed to remain unchanged or to change in a prede-
termined manner.
The present paper tries to overcome some of the shortcomings of both "game
theoretic" and "simulation" approaches by modelling banks as learning agents.
Agents who learn about each others' actions through repeated interaction is a
recurring theme in evolutionary game theory. In one strand of the literature1 the
agents know their payoff function, and learn about others' behaviour. They do
so playing the stage game repeatedly, while choosing their actions on the basis
of adaptive rules of the type "choose a best reply to the current strategy profile"
or "choose a best reply to the next expected strategy profile". Results obtained
in this strand cannot be immediately applied here: banks cannot choose best
replies as they do not know their payoff function. A second research line does
not require knowledge of the payoff function on the part of the learners; they
are instead of the kind "adopt more frequently an action that has produced
a high payoff in the past". The main results of this literature are about the
convergence (or non-convergence) of actions to equilibria of the stage game.
The approach adopted here is close to the latter. However, because the pay-
offs are calculated on the basis of a settlement algorithm, we cannot analytically
calculate the equilibria ex-ante, and then demonstrate convergence (or the lack
of it). Instead, we show convergence by means of simulations, inferring then
that the attraction points are equilibria of the stage game - in a sense that
we make precise. Because the payoff function is stochastic and unknown, the
problem of each optimizing bank lends itself to a heuristic approach. From this
perspective, our work bears strong links to the reinforcement learning litera-
ture2. From an individual agent's perspective it relates it relates to operations
research, where a typical problem is that of maximizing an unknown function.
1E.g. fictitious play, following Brown (1951)
2See Sutton and Barto (1998) for an overview. For Q-Learning, a common reinforcement
learning technique, see Watkins and Dayan (1992).
3
However, in our setting the environment is not static: through time, actions
yield different payoffs both because the payoff function is random, and because
the other agents change their behaviour.
The model is rich enough to investigate a number of policy issues; here, we
focus on the aggregate liquidity of the system. As a first result we derive a
liquidity demand function, relating total funds to the ratio of delay to liquidity
costs. This function is found to be increasing to the relative cost of delay,
and S-shaped. Then, we look at the effect of operational incidents affecting
random participants of the system. We find that banks would generally prefer
to commit more liquidity in case the disruption were known - except from the
extreme cases of very low and very high delay costs. Throughput guidelines
for payment submission are a common used by system-designers to reduce risk
in payment systems;3 we look at the effect of one such rule on liquidity usage
and find that at sufficiently low delay costs banks would increase their liquidity
holdings to contain delays. Finally, we explore some efficiency issues, namely
whether smaller systems are more or less liquidity efficient than large ones. We
find that a system with a smaller number of banks uses less liquidity for a given
level of payment activity.
The paper is organised as follows: Section 2 develops a formal description of
the model and the agents' learning process, and describes the payoff function.
Section 3 presents the results of the experiments and section 4 concludes.
2 Description of the model
2.1 Stage game and its repetition
The model consists of N agents indexed by i = 1...N , who repeatedly play a
stage game Γ = hA, π1, π2, ...πNi. Here A = {0, 1, ...K} is the (common) finite
action set for each agent, and πi : AN → F is i's outcome function, which
maps the set of action profiles into a set of payoff distribution functions. That
is, given the action profile a ∈ AN , agent i receives a stage-game payoff drawn
from a univariate distribution πi (a), whose shape depends on N parameters -
the stage game action profile. To keep the exposition uncluttered, we leave the
precise form of the outcome function πi (.) undefined at the stage. Details are
given in Section 3.1, where we also give a precise economic interpretation to the
abstract entities introduced here. Information in the game is incomplete as the
outcome function π (.) is unknown to the agents. Agents are risk-neutral, so
they care about the expected payoff. Hence, bank i will only be concerned with
its payoff functions fi (a) = E (πi (a)).
The stage game Γ is repeated through discrete time, running from t = 0 to
(potentially) infinity. The action profile chosen in stage game t is denoted by
a (t) = {a1 (t) , a2 (t) , ..., aN (t)}. A particular realization of the payoff vector
3A throughput guideline is a constraint imposed on banks' behaviour by the system regu-
lator; typically, it demands that certain percentages of the total daily payments be executed
by given deadlines within the day.
4
drawn from π (a (t)), is indicated by y (t) ∈ RN , which is therefore also called
the "game-t payoff".
Monitoring is incomplete. At the beginning of (stage-) game t, each agent i
knows the following: all its own previous choices and realized payoffs, and some
statistics of other's past choices a−i (k). A (observed) history (by i at time
i = {ai (k) , yi (k) , a−i (k)}k=0..t−1. Let us call H t the
t) is thus denoted by ht
set of all possible histories that i may observe up to t, and let us define H =
∪H t. Differently from the literature on repeated games, but more in line with
that on evolutionary game theory, we assume that agents aim at maximizing
immediate payoffs (instead of e.g. the discounted stream of payoffs). That
is, histories are essential to learn about payoffs and about others' actions, but
agents disregard strategic spillover effects between stage games. This seems a
sensible assumption here: the complexity of the environment makes it unlikely
that agents anticipate all interactions.4
2.2
Information, learning and strategies
Agent i faces two forms of uncertainty: uncertainty about the payoff function
given others' actions, and uncertainty about other's actions. The first element
gives to our model a flavour of decision theory, the second one is a game theory
issue.
2.2.1 Information
As time goes by and histories are updated, agents can be seen to accumulate
information. More formally, we posit that of the whole history observed up to t
each i retains some multi-dimensional statistics, say ℘i (t). These are the beliefs
about the state of the environment that i is learning, and it constitutes the basis
for the definition of strategies. Here, ℘i (t) is composed of two parts:
i (.) of the payoff function fi (a) = E [yi a ];5
i (.) of probabilities for other agents' actions in the next
stage game.
a) an "estimate" ef t
b) an "estimate" ept
a−i which correlates with a−i. Hence, we assume that the estimate ef t
Of a whole action profile a, i only observes its own action ai and a statistic
i (.) assigns
to each (ai, a−i) an expected payoff. As for the estimate b), each i is assumed
to maintain static expectations about others' actions. That is, i believes that
a−i is drawn from a time-invariant distribution, as if other agents were adopting
a constant mixed strategy. We adopt this assumption, a classic in evolutionary
game theory, because it is simple and because it yielded the same results as some
4In the realm of reinforcement learning, immediate payoff maximisation where actions are
associated with situations is referred after Barto, Sutton and Brouwer (1981) as associative
search.
5Risk-neutral players are only interested in the expected value of payoffs, so there is no
gain in assuming that ef instead maps actions into payoff probability distributions.
5
more sophisticated forms of beliefs.6 Because action profiles that generate the
i (.) also refers to a−i
instead of ai. In the simulation, we posit that a−i is the average action of the
"other" agents, which clearly takes values in [0, K]. So, approximating to the
i (a−i) is a vector with K entries, collecting the probabilities
same statistic a−i are indistinguishable to i, the estimate ept
nearest integer, ept
of any of the (other agents') average action being played.
2.2.2 Learning
The information stored in ef t
i (.) is updated as time goes by, according
to learning rules. A learning rule for agent i, denoted by Λi, assigns to each
observed history ht−1
i (.) and ept
an updated ℘i (t).
i
Define Ik (ai, a−i) as the indicator function equal to 1 if action profile (ai, a−i)
appears at time k and zero otherwise. We use the following learning rule:
i (ai, a−i) = Pk=0...t−1 yi (k) Ik (ai, a−i)
ef t
Pk=0...t−1 Ik (ai, a−i)
1 +Pk=1...t−1 Ik (a−i)
ept
i (a−i) =
t + K
(1)
(2)
In words, ef t
i (ai, a−i) is the average payoff obtained under action profile
(ai, a−i) up until time t excluded. Similarly, the components of the vector
i (a−i) are calculated according to the observed frequencies, starting from an
initial estimate 1/N . This is known as the "fictitious play" updating rule start-
ing from a uniform estimate; it corresponds to Bayesian updating of beliefs
about a constant, unknown distribution over the other agents' actions.7
ept
2.2.3 Strategies
A strategy for i is a map assigning to each ℘i(t) an action to be taken, i.e.
some ai (t). A particular strategy can be seen as motivated by some "rationale",
resting in turn on the basis of a learning process which we now describe.
Each i is risk-neutral and aims at maximizing the expected immediate payoff.
Because i believes that the opponents play a particular a−i with probability
i (a−i), its strategy dictates:
ai (t + 1) = arg max
= arg max
ai
Ehef t
ai Xa−i ef t
ii
i (ai, a−i)(cid:12)(cid:12)ept
i (ai, a−i)ept
i (a−i)
(3)
The fact that banks maximize their immediate payoff is only one of the
many possible preference specifications. Alternatively, agents might also be
6We explored in particular the possibility that players believe that the opponents' actions
follow a Markov process. In this case, the estimate under 2) is a transition matrix, containing
the probabilities of a particular a−i being played at t + 1, conditional on a = (ai, a−i) being
observed at t.
7See e.g. Fudenberg and Levine (1998) pg. 31 for details.
6
ept
0
1
2
t-2
t-1
t
t
Figure 1: Information, learning and strategies
taking into consideration future payoffs. In this case, however, optimal strate-
gies would be far more complex. Indeed, discounting expected future payoffs
would create an implicit trade-off between exploitation (the use of actions that
appear optimal in the light of the available information), and exploration (the
use of seemingly sub-optimal actions, which might appear such because of lack
of experimentation). Our preference specification severs this payoff-related link
between stage games, which nevertheless are interrelated because learning takes
place across them. This short-sighted maximization assumption is common in
the bounded-rationality and evolutionary game theory literature.8
Figure 1 illustrates the relationship between histories, information and strate-
gies. A history up to t− 2 is summarized in ℘i(t− 1). This, along with the new
data obtained in t − 1, is updated into ℘i (t) by the learning rule Λi. In turn
℘i (t), which is the information available at t, is mapped by a strategy si, to an
action ai (t).
It should be noted that in early repetitions of the stage game, ept
while ef t
i is heavily
influenced by the initial (arbitrary) estimate, for which we simply use 1/K,
i (.) is the average of a few observed payoffs only. Hence, strict adoption
of Eq. 3 would most likely yield, and possibly lock into, sub-optimal actions.
To avoid this, we suppose that agents first randomly choose a certain number of
actions to explore the environment, and then start making choices as in Eq. 3
- which we call "informed decisions". To ensure further exploration, each agent
also tries itself out at least once every a−i that it encounters. This models
8Fudenberg and Levine (1998) contains an authoritative review of models with myopic
agents. To quote only some of these seminal contributions, see i) the literature on Fictitious
Play by Brown (1951), Foster et al. (1998), and Krishna et al. (1998), ii) the literature on
learning and bounded rationality inc. Kandori et al. (1993), Young (1993), Ellison (2003),
and Blume (1993), iii) studies on Imitation and Social Learning by e.g. Schlag (1994).
7
learning from other agents.
These choices on the length of the "exploration phase" are evidently arbi-
trary; however, in the model there are clearly no exploration costs, so the length
of the exploration period can indeed be assumed exogenously. On the other
hand, some limit to exploration must be imposed, as the sheer size of the action
spaces inhibits a brute force approach, whereby i collects a very large sample
of all possible action profiles (and respective payoffs) before making informed
decisions.9
2.3 Specification of the payoff function
The model of learning about an unknown stochastic payoff function that is
determined party by the agent's own actions and partly by the actions of other
agents can lend itself to a number of applications. The specification of the payoff
function fi ties it to the problem of a payment system analyzed in this paper.
One possible specification of f could be a simple analytical function of the
players' actions. The problem would then become that of analyzing the lim-
iting behaviour of the learning rules and strategies, something that could be
done analytically, provided f is simple enough. However, in quest of increased
realism in payment system modelling we specify f via an algorithm represent-
ing a "settlement day" with a large number of daily payments. To understand
why realistic analytical functions are difficult to develop, consider the follow-
ing:
imagine first that banks have always enough liquidity to make payments
instantaneously. In this case payments flow undisturbed, delay costs are zero,
and only liquidity costs matter. Their calculation is trivial. However, if banks
commit less funds for settlement (as banks want to minimize costs), it becomes
more likely that the funds are at some point insufficient for banks to execute
payments immediately. As shown in Beyeler et al.. (2006), these liquidity short-
ages cause payments to occur in "cascades", whose length and frequency bears
no correlation with the instruction arrival process that regulates payment in-
structions as the settlement of payments becomes coupled across the network
when incoming funds allow the bank to release previously queued payment. As
a consequence, the flow of liquidity and thus delay times for individual banks
become largely unpredictable.
In the model ai represents any external funding decision by the bank. The
funds allow the bank to execute payment instructions, which the bank receives
throughout the day according to a random process. Banks have costs for both
committing liquidity for settlement, and from experiencing delays in payment
processing due to insufficient funds. The settlement day is modelled in contin-
uous time, with time indexed as t ∈ [0, T ].10 At any time interval dt, bank
9In the simulations, each i chooses among 40 possible actions, and 40 are the possible
average actions by the "others" (a−i). Thus, full exploration would require observing 402 =
1600 different action profiles (ai, a−i), each of which should be sampled enough times to obtain
a reliable estimate of f (.).
10In the previous section, t indicizes "days", but we feel there is no risk of confusion, as
Section 2 and the present are relatively independent.
8
1
i receives an instruction to pay 1 unit to any other bank j with probability
1
N −1 dt. Because there are N such banks i, and N − 1 "other" banks j, the
N
arrival of payment instructions in the whole system is a Poisson process with
parameter 1 so that, on the average day, T payment instructions are generated.
Payments are executed using available liquidity; i's available liquidity at
time t is defined as:
li (t) = ai +Z t
s=0
(yi (s) − xi (s)) ds
where xi (s) (viz. yi (s)) is the amount of i's sent (viz. received) payments
at time s. For simplicity, we assume that every i adopts the following payment
rule:11
at each t, execute instructions using First-in-First-out (FIFO) as long as li (t) > 0;
else, queue received instructions
(4)
We assume that a payment instruction received by bank at t and executed
at t′ carries a cost equal to
CD = κ
(t′ − t)
T
κ > 0
(5)
where κ is the "daily interest cost" of delaying payments. Similarly, liquidity
costs (e.g. opportunity cost of collateral) are linear:
CL = λai,
λ > 0
(6)
Finally, the stage game payoff is the sum of the costs in Eq. 5 and those
in Eq. 6, the former summed up over all i′s delayed payments.
3 Experiments
3.1 Parameters and equilibria
The continuous-time settlement day is modelled as a sequence of 104 time units
indexed t ∈ [0, 104]. Given that the arrival of payment instructions is a Poisson
process with parameter 1, on average banks receive a total of 104 payment
instructions per day. A sequence of days (stage games) is called a play. In the
simulations, we terminate a play when no bank changes its liquidity commitment
decision for 10 consecutive days (convergence). We run 30 plays for each set of
model parameters and find that convergence always occurs.
Banks start the adaptation process with random decisions for liquidity, and
gradually accumulate information on the shape of the payoff function. When
11The rule under (4) is evidently optimal for the cost specification given here. As banks
need to pay upfront for liquidity, they have no incentive to delay payments if liquidity is
available. Under other cost specifications (e.g. priced credit or heterogeneous payment delay
costs) this would, however, not be the case.
9
enough information has been collected, banks adopt the rule described in Eq.
3 for making decisions on liquidity to commit. A series of stage games is ended
in the simulations when no bank changes its collateral posting decision for 10
consecutive games. This means that at this point, no bank wishes to change its
action, given the information available and given other banks' actions.12
Suppose that the payoff function were known by the banks; given our spec-
ification of strategies, it would then be clear that the converged-to actions are
a Nash equilibrium of the stage game.13 We cannot quite draw the same con-
clusion in our setting: the payoff exploration is necessarily partial, so it might
be that some profitable actions were never tested enough to be recognized as
such. Hence, the equilibria converged to are only "partial"-Nash equilibria, or
Nash-equilibria conditional on the "partial" information that banks have about
the payoff function. However, as we discuss later, we observe a clear consistency
in learning. This suggests that the partiality of the information collected is suf-
ficient, and the equilibria reached are probably good approximations of the true
stage-game Nash equilibria.
The base system consists of 15 banks.
In section 3.2 we investigate the
impact of the system size. Banks choose their action, ai among forty different
levels, ranging from 0 to 80 in intervals of 2. The cost functions are as in Eq. 5
and Eq. 6. We normalize liquidity costs at λ = 1 and look at different values
of delay costs κ ranging from 1/8 to 512 in multiples of 4. We are interested in
the demand for liquidity (i.e. in the choices of ai) at different values of κ, and
in the resulting settlement delays and payoffs.
3.2 Base experiment
As expected, with low delay costs banks tend to commit low amounts of liquidity
(∼50 units) and delay payments instead, and at high cost of delay bank prefer
to commit plenty of liquidity (∼1044 units) in order to avoid expensive delays.
Figures 2 and 3 illustrate typical evolution of the simulations for two extreme
levels of delay costs. The sudden changes in liquidity correspond to the point
where banks start making informed decisions (see Section 2.2.3).
We find that convergence always occurs - on an aggregate level within a
narrow range. A priori, learning might be sensitive to initial observations, and
hence it might be subjected to drastic differences in the final "conclusions". The
consistency of the learning process is illustrated in Figure 4, where we plot the
different parameter specifications. Due to randomness - which makes histories
converged-to value of P ai (i.e. the total liquidity committed) across plays, for
necessarily different - the "learned" liquidity level P ai clearly differ, but they
do so within small ranges.
12While some changes in actions may occur due to the randomness of payoffs and learning,
these did not qualitatively change the results in simulations with longer convergence criteria.
13This simple property of Fictitious Play stems from the fact that, if for all t > t′ the action
profile is some constant a, then the estimates epi converge to the true value 1
N Σai. Strategies
prescribe playing a best reply to epi so, if a were not a Nash Equilibrium, sooner or later one
player would choose some a′
i 6= ai.
10
d
e
t
t
i
m
m
o
c
s
d
n
u
f
1200
1000
800
600
400
200
0
d
e
t
t
i
m
m
o
c
s
d
n
u
f
1
11
21
31
41
51
stage game
61
71
81
Figure 2: Total liquidity - low delay costs
1200
1000
800
600
400
200
0
1
51
101
151
201
stage game
251
301
Figure 3: Total liquidity - high delay costs
11
d
e
t
t
i
m
m
o
c
s
d
n
u
f
1200
1000
800
600
400
200
0
1
cost ratio
512
128
32
8
2
0.5
0.125
30
play
Figure 4: Total equilibrium liquidity across plays
It should be noted that while the system consistently "learns" the same level
of total liquidity, this can represent many configurations of single banks' liquidity
choices. Hence, our simulation don't show that always the same equilibrium
is reached; rather, that the equilibria that are reached are characterized by
a narrow span of total liquidity in the system. Given the symmetry of the
model, it is clear that for any equilibrium (i.e. any equilibrium profile of actions
(a1, a2...aN )), there are many other equilibria obtained via a permutation of
the actions between the players that yield to same total liquidity on the system
level.
Another interesting feature of the model is its ability to match a well known
empirical fact: a low ratio of available liquidity to daily payments ("netting
ratio"), which in turn implies high levels of liquidity recycling. Because the
system processes on average 10.000 payments a day, the above results imply
that the ratio in our simulation is between 0.5% and 10.4%. For comparison,
CHAPS Sterling's netting ratio is 15% (James 2004)14 and in Fedwire as low
as 2.2%.15 The real netting ratios are bound to be higher due to the fact that
payments in them are of varying sizes in contrast to the more fluid unit size
payments modelled here.
Figure 5 shows the equilibrium demand for intraday credit as a function of
the cost ratio. This function is S-shaped in the exponential delay cost scale, that
14Calculated as the ratio of collateral used for intraday credit to the value of payment
settled.
15in 2001.
Calculated as (balances + mean overdrafts) / total value.
Sources:
www.federalreserve.gov/paymentsystems/fedwire
12
1200
1000
800
600
400
200
0
1000
0.1
1
10
100
delay cost
s
d
n
u
f
r
o
f
d
n
a
m
e
d
Figure 5: Liquidity as a function of delay costs
is, it is relatively flat at both low and high levels of delay costs. At comparatively
low delay costs banks evidently commit little liquidity; hence the return for
increasing liquidity holdings are high, and so a little more liquidity suffices to
cope with increased delay costs. As a consequence, for low delay cost levels the
demand curve is flat. Consider now the situation with high delay costs. There,
the liquidity committed is high and returns to increasing liquidity are low, so
one might think that an increase in delay costs calls for high extra amounts
of liquidity. However, this is not the case, because gains from liquidity indeed
diminish above a certain level when all payments can be made promptly. Hence,
for high delay costs, liquidity demand is insensitive to further increases in delay
costs, and the demand curve is flat again. In between these two extremes, the
demand for liquidity increases exponentially with delay costs.
We find that delays in the system increase exponentially as banks reduce
the amount of liquidity when this is relatively expensive compared to delaying
payments. The phenomenon is known as "deadweight losses" (Angelini 1998) or
"gridlocks" (Bech and Soramaki 2002) in payment systems. Figure 6 shows the
relationship between system liquidity and payment delays. In intuitive terms,
the reason of this exponential pattern is the following. First, a bank that reduces
its liquidity holdings might have to delay its outgoing payments. Second, as a
consequence, the receivers of the delayed payments may in turn need to delay
their own payments, causing further downstream delays and so on. Hence, a
decrease of a unit of liquidity may cause multiple units to be delays. Third,
such a chain of delays - and hence this multiplicative effect - is more likely and
13
t
n
e
m
y
a
p
r
e
p
y
a
e
d
l
500
450
400
350
300
250
200
150
100
50
0
0
200
400
600
800
1000
1200
funds committed
Figure 6: Delays as a function of liquidity
longer, the lower the liquidity possessed by the banks. Thus the total effect of
liquidity reduction acts in a compounded (exponential) fashion.
An interesting question is how good the performance of the banks is in
absolute terms. To understand this we compare the payoffs received by the
banks through adaptation with two extreme strategies:
a) all banks delay all payments to the end of the day;
b) all banks commit enough liquidity to be able to process all payments
promptly.16
The comparison between the performance of these two pure strategies and
the learned strategy is shown in Figure 7. For any cost ratio, the adaptive
banks obtain better payoffs than any of the two extreme strategies - except for
the case with high liquidity costs when the costs are equal. Banks manage to
learn a convenient trade-off between delay and liquidity costs. On the contrary,
the strategy under a) becomes quickly very expensive as delay costs increase,
and the strategy under b) is exceedingly expensive when delays are not costly.
3.3
Impact of network size
In order to investigate the impact of system size on the results presented in the
previous section, we ran simulations varying the number of participants. To en-
sure comparability, we kept the number of payments constant across simulations
16In fact the liquidity committed in the simulation with the highest delay cost was used as
the scenario for prompt payment processing.
14
0.1
1
10
100
1000
a)
b)
funds committed
0.00
-0.02
-0.04
-0.06
-0.08
-0.10
-0.12
-0.14
-0.16
-0.18
-0.20
t
n
e
m
y
a
p
r
e
p
f
f
o
y
a
p
Figure 7: Comparison of strategies: a) min liquidity; b) min delays
and maintained the network complete.
We observe that liquidity demand increases with the system size, and the in-
crease is more pronounced the higher the delay costs. For example, the demand
function is unchanged at low delay cost while, for high delay costs, a 50-bank
system requires 215% the liquidity needed in a 15-bank system.
Similar results hold about delays. The relationship between liquidity com-
mitted and delays remains close to exponential irrespective of system size; how-
ever, larger systems experience more delays, for any level of initial liquidity (see
Figure 9).
An intuitive explanation of these phenomena could be the following. First,
note that if the number of participants is increased by a factor x (keeping
turnover constant), the volatility of the balance of each bank is multiplied by
a factor 1/x′ > 1/x - we show this in a moment. Second, suppose that i) the
optimal ai is proportional to the volatility of a bank's balance δ (i.e. ai = zδ)17
and that ii) banks post all the same amount of liquidity (i.e. ai = aj). It then
simply follows that the total amount of liquidity increases with the system's
size: (nx) z δ
x′ > nzδ (here nx is the number of banks in the larger system, and
δ
x′ the corresponding volatility of balances).
The key point is that, if the number of participants is increased (keeping
turnover constant), the volatility of banks balances rises more than proportion-
ally. To see why this is the case, consider the simplified but illustrative situation
17This is exactly the case if a bank chooses ai as to cover z "standard deviations" from the
average balance.
15
0.2500
0.2000
0.1500
0.1000
0.0500
0.0000
1000
0.1
1
10
100
2 banks
15 banks
5 banks
50 banks
10 banks
Figure 8: The effect of system size on the demand for liquidity
where liquidity is abundant, so there are no delays. In this case, a bank's net
position is the sum of a series of random perturbations (incoming and outgoing
payments), equally likely to affect it positively and negatively. In other words,
a bank's net position is a random walk, whose value after n perturbations av-
erages zero, with a standard deviation √n. By increasing the system size by
a factor x, the orders are distributed over more banks, so the average number
of perturbations for any given time interval is multiplied by 1/x < 1. Accord-
ingly, the standard deviation of the balances at the end of any time interval is
multiplied by 1/x′ =p1/x > 1/x.
3.4 Throughput guidelines
Some interbank payment systems have guidelines on payment submission jointly
agreed upon by the system participants18. The rationale for throughput guide-
lines is to induce early settlement in order to e.g. reduce operational risk or
perceived coordination failures among participant. For example, if a large chunk
of payments are settled late in the day, an operational incident would be more
severe as more payments could potentially remain unsettled before close of the
payment system and the financial markets where banks balance their end-of-day
liquidity positions.
18E.g. the FBE (Banking Federation of the European Union 1998) has set guidelines on
the timing of certain TARGET payments. In CHAPS Sterling, members must ensure that on
average (over a calendar month) 50% of its daily value of payments are made by 12 pm and
that 75%, by value, are made by 2.30 pm. (James 2004).
16
1200
1000
800
600
400
200
b)
a)
s
d
n
u
f
r
o
f
d
n
a
m
e
d
0.10
1.00
10.00
delay cost
100.00
0
1000.00
Figure 9: Impact of throughput guidelines (b) on liquidity
We simulate a particular realization of throughput guidelines by introducing
an additional penalty charge for delays that last longer than one tenth of the
settlement day (i.e. 1000 time units). The penalty charge is set to 64, in order
to sufficiently penalize non-compliance with the rules.
Figure 10 shows the impact of the throughput guidelines on the amount of
liquidity committed by the banks. When delay costs are high, banks already
commit enough liquidity to avoid long delays, so the throughput guidelines
have no effect. They do however, in the case of low delays costs, induce banks
to commit more liquidity. Not surprisingly, this comes at a cost to the banks,
which are forced away from their first-best choice. The increase in costs are
of the order of 70% at the lowest level of delay costs, and 20% at the second
lowest. The payoff comparison is shown in Figure 11.
3.5 Operational incident
Short term outages by banks in the payment system are rare in actual payment
systems, but do take place occasionally. In a typical scenario a participating
bank experiences problems connecting to the system due to temporal unavail-
ability of IT systems or telecommunication facilities. Due to the design of the
payment systems, a disconnected bank can in such situations generally still
receive payments to its account at the central bank, but cannot submit instruc-
tions to pay from its account. Unless other banks stop paying to the troubled
bank, it quickly becomes a liquidity sink, and the liquidity available for settling
17
0.10
1.00
delay cost
10.00
100.00
1000.00
0
b)
a)
f
f
o
y
a
p
-0.02
-0.04
-0.06
-0.08
-0.1
-0.12
-0.14
Figure 10: Impact of throughput guidelines (b) on payoffs
payments at other banks is reduced.19
In this set of simulations we ask the question of how much liquidity banks
would wish to commit in such a situation, i.e. what is the impact of an opera-
tional outage on the demand for intraday credit. The banks are assumed to be
unaware of the possible incident, and unable to discriminate among their coun-
terparts, so the intraday liquidity management rule under Eq. 4 is still adopted.
Under these assumptions, we simulated a scenario where a randomly selected
bank can receive, but cannot send payments for the first half of the settlement
day. On average, this means that up to T /2N = 10.000/(2 · 15) ≃ 333 liquidity
units cannot be used by other banks as a source of liquidity20. Depending on
the delay cost, this figure varies between 1900% and 30% of the average total
liquidity injected in the system at the beginning of the day (the first figure being
for the case when both delay cost and liquidity demand are low, and the second
for when costs liquidity demand are both high).
We found that the effect of operational incidents on the demand for liquidity
is highest at a relative delay/liquidity cost ratio of 2 - hence fairly low in the
range; at this point, the increase in intraday credit demand is 144 units, or
an increase of 85%.
It should be noted that banks do not compensate for
the full amount of liquidity "trapped" by the distressed bank, but prefer to
partly make up for that, and partly increase delays. For higher delay costs,
19see e.g. analysis on the impact of the 9/11 terrorist attacks in McAndrews and Potter
(2002), Lacker (2004), and Soramaki et al (2007)
20The shortage of liquidity equals the number of payment orders received by the distressed
bank, and not yet executed until the second half of the day.
18
1200
1000
800
600
400
200
0
1000
b)
a)
0.1
1
10
cost ratio
100
s
d
n
u
f
r
o
f
d
n
a
m
e
d
Figure 11: Liquidity - normal circumstances (a) and operational incidents (b)
delays remain approximately unchanged compared to the scenario without the
incident. Finally, when delay costs are lower than liquidity costs (i.e. for a cost
ratio <1), banks prefer to hold about the same amount of liquidity as without
the incident, and experience the delays caused by the reduced liquidity.
In
this case, the impact of an operational incident increases both the demand for
liquidity (but less than what was trapped) and delays - more so the one which
is less costly.
4 Summary and conclusions
In this paper we developed an agent-based, adaptive model of banks in a pay-
ment system. Our main focus is on the demand for intraday credit under alterna-
tive scenarios: i) a "benchmark" scenario, where payments flows are determined
by the initial liquidity, and by an exogenous arrival of payment instructions; ii)
a system where, in addition, throughput guidelines are exogenously imposed iii)
a system subject to operational incidents.
It is well known that the demand for intraday credit is generated by a trade-
off between the costs associated with delaying payments, and liquidity costs.
Simulating the model for different parameter values, we were able to draw with
some precision a liquidity demand function, which turns out to be is an S-shaped
function of the delay / liquidity cost ratio. We also looked at the costs expe-
rienced by the banks, as a function of the model's parameters. By the process
of individual payoff maximization, banks adjust their demand for liquidity up
19
b)
a)
0.1
1
10
100
funds committed
1800
1600
1400
1200
1000
800
600
400
200
t
n
e
m
y
a
p
r
e
p
y
a
e
d
l
0
1000
Figure 12: Delays as a function of cost ratio (b: incident)
(reducing delays) when delay costs increase, and down (increasing delays), when
they rise. Interestingly however, the absolute delay cost remains approximately
constant when the ratio delay/liquidity costs changes. As expected, the de-
mand for intraday credit is increased by an operational incident. However, this
effect is found to be important only if liquidity is costly compared to delaying
payments. Likewise, throughput guidelines increase the demand for intraday
credit - as banks try to avoid penalties for not adhering to them. In total this
reduces the payoffs of the banks. Nevertheless, throughput guidelines may be
beneficial when additional benefits that are not in the current model are taken
into account (among these, benefits related to reducing operational risk).
This model produces realistic behaviour, suggesting that it may be used to
investigate a wide array of issues in future applications. A number of extensions
are possible. First, alternative specifications for the instruction arrival process
may be applied (see e.g. Beyeler at al. (2006)). Alternatively, one could change
the assumptions on the banks' network: while the complete network assumption
implicitly adopted here fits well with e.g. the UK CHAPS system, an interesting
question is how other topologies such as a scale free network topology such as in
Fedwire (Soramaki et al.. 2007) would affect the results. Also, different individ-
ual preferences could be investigated. We assumed that banks are risk neutral
and interested in maximising their immediate payoffs; it would be interesting
to verify if the introduction of risk aversion and / or preferences over expected
stream of payoffs may change the results. Finally, more complex behaviour can
be easily studied within our model; for example, the "pay-as-much-as-you-can"
rule for queuing payments could be replaced by sender limits. Similarly, more
20
sophisticated strategies can be easily modelled, supposing e.g. that banks keep
constant their actions for a number of periods (to gather more data and explore
the environment), instead of exploiting after a fixed amount of time what seems
to be the best action.
References
[1] Angelini, P (1998), ''An Analysis of Competitive Externalities in Gross
Settlement Systems", Journal of Banking and Finance n. 22, pages 1-18.
[2] Bank for International Settlements (2006), Statistics on payment and
settlement systems in selected countries - Figures for 2005 - Preliminary
version. CPSS Publication No. 75.
[3] Barto, A G and Sutton, R S (1998), Reinforcement Learning: an
introduction. MIT Press, Cambridge, Massachusetts.
[4] Barto, A G, Sutton, R S and Brouwer, P S (1981), "Associative
search network: A reinforcement learning associative memory". Biological
Cybernetics, 40, pages 201-211.
[5] Bech, M L and Soramaki, K (2002), "Liquidity, gridlocks and bank
failures in large value payment systems". in E-money and payment systems
Review, Central Banking Publications. London.
[6] Bech, M L and Garratt, R (2003), "The intraday liquidity manage-
ment game", Journal of Economic Theory, vol. 109(2), pages 198-219.
[7] Bech, M L and Garratt, R (2006), "Illiquidity in the Interbank Pay-
ment System Following Wide-Scale Disruptions". Federal Rererve Bank of
New York Staff Report No. 239.
[8] Beyeler, W, Bech, M, Glass, R. and Soramaki K (2006), "Con-
gestion and Cascades in Payment Systems". Federal Reserve Bank of New
York Staff Report No. 259.
[9] Blume, L (1993), "The statistical mechanics of best-response strategy
revision". Games and Economic Behavior, 11, pages 111 -- 145.
[10] Brown, G W (1951), "Iterative Solutions of Games by Fictitious Play"
in Koopmans T C (Ed.), Activity Analysis of Production and Allocation,
New York: Wiley.
[11] Buckle, S and Campbell, E (2003), "Settlement bank behaviour in
real-time gross settlement payment system". Bank of England Working
paper No. 22.
[12] Devriese, J and Mitchell J (2006), "Liquidity Risk in Securities Set-
tlement", Journal of Banking and Finance, v. 30, iss. 6, pages 1807-1834.
21
[13] Ellison, G (1993), "Learning,
local
interaction, and coordination",
Econometrica, 61, pages 1047-1072.
[14] Foster, D P and Young H P (1998), "On the Nonconvergence of Fic-
titious Play in Coordination Games", Games and Economic Behavior, v.
25, iss. 1, pages 79-96.
[15] Fudenberg, D and Levine D K (1998), "The Theory of Learning in
Games". MIT Press., Cambridge, Massachusetts.
[16] James, K and Willison M. (2004), "Collateral Posting Decisions in
CHAPS Sterling", Bank of England Financial Stability Report, December
2006.
[17] Kandori. M, Mailath, G J and Rob, R (1993), "Learning, Mutation,
and Long Run Equilibria in Games", Econometrica, Vol. 61, No. 1. pages
29-56.
[18] Krishna, V and Sjostrom, T (1995), "On the Convergence of Fictitious
Play", Mathematics of Operations Research 23(2), pages 479 - 511.
[19] Lacker, J M (2004), "Payment system disruptions and the federal reserve
following September 11, 2001." Journal of Monetary Economics, Vol. 51,
No. 5, pages 935-965.
[20] McAndrews, J and Potter S M (2002), "The Liquidity Effects of the
Events of September 11, 2001", FRBNY Economic Policy Review, Vol. 8,
No. 2, pages 59-79.
[21] Soramaki, K, Bech, M L, Arnold, J, Glass R J and Beyeler, W
E (2006), "The Topology of Interbank Payment Flows". Physica A Vol.
379, pages 317-333.
[22] Schlag, K (1995), "Why Imitate, and if so, How? A Boundedly Rational
Approach to Multi-Armed Bandits", Journal of Economic Theory 78(1),
pages 130 -- 156.
[23] Watkins, C and Dayan P (1992), "Technical note: Q-learning", Ma-
chine Learning, 8(3/4), pages 279-292.
[24] Willison, M (2005), "Real-Time Gross Settlement and hybrid payments
systems: a comparison", Bank of England Working Paper No. 252.
[25] Young P (1993), "The evolution of conventions", Econometrica 61, pages
57-84.
22
|
1907.09110 | 1 | 1907 | 2019-07-22T03:17:19 | Strategic Voting Under Uncertainty About the Voting Method | [
"cs.MA",
"cs.GT"
] | Much of the theoretical work on strategic voting makes strong assumptions about what voters know about the voting situation. A strategizing voter is typically assumed to know how other voters will vote and to know the rules of the voting method. A growing body of literature explores strategic voting when there is uncertainty about how others will vote. In this paper, we study strategic voting when there is uncertainty about the voting method. We introduce three notions of manipulability for a set of voting methods: sure, safe, and expected manipulability. With the help of a computer program, we identify voting scenarios in which uncertainty about the voting method may reduce or even eliminate a voter's incentive to misrepresent her preferences. Thus, it may be in the interest of an election designer who wishes to reduce strategic voting to leave voters uncertain about which of several reasonable voting methods will be used to determine the winners of an election. | cs.MA | cs | Strategic Voting Under Uncertainty About the Voting Method∗
Wesley H. Holliday
Eric Pacuit
University of California, Berkeley
University of Maryland
[email protected]
[email protected]
Much of the theoretical work on strategic voting makes strong assumptions about what voters know
about the voting situation. A strategizing voter is typically assumed to know how other voters will
vote and to know the rules of the voting method. A growing body of literature explores strategic
voting when there is uncertainty about how others will vote. In this paper, we study strategic voting
when there is uncertainty about the voting method. We introduce three notions of manipulability
for a set of voting methods: sure, safe, and expected manipulability. With the help of a computer
program, we identify voting scenarios in which uncertainty about the voting method may reduce or
even eliminate a voter's incentive to misrepresent her preferences. Thus, it may be in the interest of
an election designer who wishes to reduce strategic voting to leave voters uncertain about which of
several reasonable voting methods will be used to determine the winners of an election.
1 Introduction
A well-known fact in the study of voting methods is that strategic voting cannot be avoided (see, e.g.,
[35, 21]). A voter has an incentive to vote strategically, or manipulate, if she will achieve a prefer-
able outcome by misrepresenting her true preferences about the candidates. There are two fundamental
results showing that every reasonable voting method is susceptible to strategic voting. The Gibbard-
Satterthwaite theorem [13, 32] shows that every voting method for three or more candidates that is reso-
lute (i.e., always elects a single winner), unanimous (i.e., never elects a candidate y if all voters rank an-
other candidate x above y), and non-dictatorial (i.e., does not always elect the favorite candidate of some
distinguished voter) can be manipulated. To study strategizing for irresolute voting methods, which may
select more than one candidate as a winner, additional assumptions are needed about the voters' rankings
of sets of candidates. The Duggan-Schwartz theorem [9] shows that every voting method -- whether res-
olute or irresolute -- for three or more candidates that is non-imposed (i.e., any candidate can be elected)
and has no nominator (i.e., no voter can always ensure that her top ranked candidate is among the set of
winners) can be manipulated by a optimist or pessimist, defined as a voter who compares sets of candi-
dates in terms of her highest (resp. lowest) ranked candidate in each set. Related results by Kelly [19],
Benoit [2], Feldman [11], and Gardenfors [12], among others, show that manipulation is unavoidable
under different assumptions about how to lift a voter's ranking of candidates to sets of candidates.
In the theorems mentioned above, a voter's decision to manipulate relies on strong assumptions about
what the voter knows about the voting situation. Two key assumptions are: (1) the strategizing voter
knows how the other voters in the population will vote or have voted; (2) the strategizing voter knows
and understands the rules of the voting method. There is a growing body of literature that explores
weakening assumption (1) [6, 27, 5, 22, 30, 36]. One reason voters may be uncertain about how other
voters will vote, even if they know other voters' true preferences, is because they think that other voters
might be voting strategically. There are sophisticated game-theoretic models that explore the implications
∗For helpful comments, we wish to thank Mikayla Kelley, the three anonymous referees for TARK, and an audience at the
Center for Human-Compatible AI's Interdisciplinary Insights Seminar on May 1, 2019.
L.S. Moss (Ed.): TARK 2019
EPTCS 297, 2019, pp. 252 -- 272, doi:10.4204/EPTCS.297.17
c(cid:13) Holliday and Pacuit
This work is licensed under the
Creative Commons Attribution License.
Holliday and Pacuit
253
of weakening assumption (1) for this reason (consult part II of [21] for an overview of this literature).
Whatever the reason, when voters have partial information about how other voters will vote, the decision
about whether to manipulate is more complicated, since voters are uncertain about which sets of winners
will result from their decision to manipulate. One natural assumption is that voters are risk averse and
only strategize when doing so cannot lead to a worse outcome (according to their true preferences) and
might lead to a better outcome [6] (cf. [30] for other assumptions about how voters decide to strategize).
The potentially negative aspects of strategic voting (see, e.g., [31, 7]) have motivated the search
for barriers against manipulation.1 One potential barrier, investigated mostly in the AI literature, is the
computational complexity of determining a profitable manipulation for a given voting method (see, e.g.,
[10, 7]).
In this paper, we will investigate another potential barrier against strategic voting: adding
uncertainty about the voting method that will be used to determine the set of winners. Thus, we weaken
assumption (2) above. Suppose that S is a set of voting methods. We assume that the voters know that
one of the methods from S will be used to select the winners, but they do not know which one -- hence
we call S the uncertainty set. To simplify the initial study, in this paper we assume that a strategizing
voter knows how the other voters will vote or have voted. This assumption is not unrealistic in certain
voting situations; for example, in a hiring committee meeting in which committee members are asked
to sequentially report their rankings of job candidates, the committee member last in the sequence will
already know how all the other members have ranked the candidates.
As in the situation where a voter has partial information about how other voters will vote, in the
situation where a voter has uncertainty about the voting method, the voter must take into account different
possible winning sets of candidates when deciding whether to strategize. Given a ranking of sets of
candidates, we consider three ways for a voter to solve this decision problem. From the most conservative
to the least conservative, the three types of manipulation are:
1. Sure manipulation: It is certain that submitting an insincere ranking will lead to a better outcome
no matter which voting method is used.
2. Safe manipulation: It is certain that submitting an insincere ranking will lead to an outcome that is
at least as good and might lead to a better outcome.2
3. Expected manipulation: Given a probability distribution on the set of voting methods, submitting
an insincere ranking is more likely to lead to a better outcome than to lead to a worse outcome.
We aim to find combinations of voting methods that reduce or even eliminate a voter's incentive to
strategize (given one of the above notions of manipulation and a ranking of sets of candidates). That is,
does uncertainty about the voting method reduce the chance that voters will strategize?
If so, it may be in the interest of an election designer who wishes to reduce strategic voting to
leave voters uncertain about which of several voting methods will be used to determine the winners of
an election. Though perhaps implausible in the case of a democratic political election, creating such
uncertainty does not seem an implausible choice by, e.g., the Chair of a hiring committee. Moreover, if
committee members judge each of the possible voting methods to be reasonable and also wish to reduce
strategic voting, they may well endorse the choice of the Chair to create such uncertainty.3
1It has been argued that strategic voting also has positive aspects (see, e.g., [8]). In this paper, we will not debate whether the
negative aspects outweigh the positive. Our point is about what barriers are available if one wishes to reduce strategic voting.
2In [17] and [34] the term 'safe manipulation' is used in a different sense and in a different context involving manipulation
by coalitions of voters.
3Naturally, the Chair should be required to fix the voting method to be used before seeing the votes, so that the Chair cannot
pick the voting method that produces his or her favored outcome based on the already submitted votes.
254
Strategic Voting Under Uncertainty About the Voting Method
Another way that voters may be uncertain about the outcome of an election is that a voting method
may use randomization to select the winners. A probabilistic voting method assigns a lottery over the set
of candidates to each collection of rankings of the candidates for the voters (see [4] for a recent survey
of results about probabilistic voting methods). Gibbard [14] noted that a random dictatorship is immune
to strategizing.4 Suppose that S is a set of dictatorships -- one for each voter. After the voters submit
their rankings, one of the dictatorships in S is chosen, and the candidate ranked first by the dictator is
selected as the winner. Assuming voters submit rankings of the candidates without ties, this method
always selects a single winner. It is clear that no voter has an incentive to misrepresent their top choice,
since either they will not be chosen as the dictator, in which case their ranking will be ignored, or they
will be chosen as the dictator, in which case their top choice is guaranteed to win. Of course, each
dictatorship is immune to strategizing by itself. Beyond this example of random dictatorship, there are
general results suggesting that randomization can be an effective barrier to manipulation [26, 1].
The relationship between our work and probabilistic voting methods is clarified at the end of this
In short, although any uncertainty set S of voting methods determines a probabilistic voting
paper.
method FS, none of our three notions of manipulation above with respect to S is equivalent to a standard
notion of manipulation with respect to the probabilistic voting method FS. In addition, since the mapping
S (cid:55)→ FS from uncertainty sets of voting methods to probabilistic voting methods is not one-to-one, there
is no obvious definition of sure, safe, or even expected manipulation for probabilistic voting methods
such that for any uncertainty set S of voting methods, S is susceptible to sure/safe/expected manipulation
if and only if FS is susceptible to sure/safe/expected manipulation. Thus, there is no obvious way to
reframe our investigation purely in the language of probabilistic voting methods.
Both our approach and that of probabilistic voting methods involve a loss of transparency at the time
of the vote. However, the approaches seem to differ in the explainability of the outcome after the vote.
In our approach, the election designer may simply inform voters after the vote of which voting method
f from S was used to determine the outcome of the election, so the algorithm for f can be used to
explain why the election had one outcome rather than another. This may be a more intelligible reason for
voters than "this was the outcome of the lottery." Abstractly, the difference is that in our setting, initially
the voters have subjective uncertainty about which deterministic mechanism will be used; but after this
subjective uncertainty is removed, there is an explanation of why the election had one outcome rather
than another in terms of the deterministic mechanism applied to the voters' inputs.5 By contrast, with a
probabilistic voting method, the only available "explanation" of the outcome is in terms of an inherently
stochastic mechanism applied to the voters' inputs, which may fail to explain why the election had one
outcome rather than another with non-zero probability (cf. [18, p. 24], [29, p. 238]).6 Of course, whether
such a contrastive explanation is desirable may vary from case to case.
Our findings in this paper are of two main types: analytic results with mathematical proofs and data
from computer searches. For these searches we generalize to sets of voting methods a standard index
of manipulability for a single voting method, known as the Nitzan-Kelly index [25, 20], which gives
the percentage of profiles with n candidates and m voters in which at least one voter in the profile has
4Cf. [16], where voters are assumed to submit utility functions over the set of candidates. In this paper, we restrict attention
to the case where voters submit rankings of the candidates.
5That the voters have subjective uncertainty about which method in S will be used does not imply that the election designer
uses a chance process to determine which method will be used. As in standard voting theory, we make no assumption about
how the election designer chooses the voting method, whether by randomization, consideration of axiomatic properties, etc.
6In the case where f is an irresolute voting method that outputs a winning set X of candidates in a given election, a single
winner may be chosen by lottery; but still we have an explanation in terms of the algorithm for f of why the field was narrowed
to X rather than some other winning set before applying tiebreaking.
Holliday and Pacuit
255
an incentive to manipulate. In probabilistic terms, assuming the Impartial Culture Model [15] in which
every profile is equally probable, this index gives the probability that in a randomly chosen profile, at
least one voter has an incentive to manipulate. Similarly, we report data on the percentage of profiles in
which at least one voter has an incentive to manipulate against a set S of voting methods. All of the data
in the paper were produced by a Python script available at https://github.com/epacuit/strategic-voting.
The rest of the paper is organized as follows. The next section introduces our formal framework,
including the definitions of the voting methods we study in this paper. The three notions of manipulation
are studied in Sections 3, 4, and 5, respectively. Section 6 discusses the connections with probabilistic
social choice. Finally, in Section 7 we conclude with some pointers to future work.
2 Preliminaries
Let C be a nonempty finite set of candidates and V a nonempty finite set of voters. We use lower case
letters from the beginning and end of the alphabet a,b,c, . . . ,x,y,z, . . . for elements of C and lower case
letters from the middle of the alphabet i, j,k, . . . for elements of V .
A voter's ranking of the set of candidates is a strict linear order P on C. Let L(C) be the set of all
strict linear orders on C. For P ∈ L(C) and X ⊆ C, let max(X,P) be the maximally ranked element of
X, i.e., max(X,P) = y where for all x ∈ X, if x (cid:54)= y, then y P x (such an element always exists since P
is assumed to be a strict linear order). Similarly, let min(X,P) be the minimally ranked element of X,
i.e., min(X,P) = y where for all x ∈ X, if x (cid:54)= y, then x P y. We say y ∈ C is ranked rth by P when
{x ∈ C x P y} = r− 1.7 So, for example, y is ranked 1st by P if and only if max(C,P) = y. To simplify
our notation, we specify a ranking by simply listing candidates from highest to lowest in the ranking,
e.g., abcd for the ranking a P b P c P d.
A profile P for (C,V ) is an element of L(C)V , i.e., a function assigning to each i ∈ V a relation
Pi ∈ L(C). If C = n and V = m, we call a profile for (C,V ) an (n,m)-profile. A pointed profile for
(C,V ) is a pair (P,i) where P is a profile and i ∈ V . For x,y ∈ C, let P(x,y) = {i ∈ V xPiy}. We write
NP(x,y) for the number of voters in P ranking x above y, i.e., NP(x,y) = P(x,y). We say that a majority
P y, when NP(x,y) > NP(y,x). Let NetP(x,y) = NP(x,y)− NP(y,x). So
prefers x to y in P, denoted x >M
x >M
P y if and only if NetP(x,y) > 0. Finally, from the strict linear order Pi we define the weak relation
Ri by xRiy iff xPiy or x = y.
A voting method for (C,V ) is a function assigning a nonempty subset of candidates, called the win-
ning set, to each profile, i.e., f : L(C)V → ℘(C)\{∅}. The following are the voting methods we will
discuss in this paper (also see, e.g., [28]).
(1) Positional scoring rules: Suppose (cid:104)s1,s2, . . . ,sm(cid:105) is a vector of numbers, called a scoring vector,
where for each l = 1, . . . ,m−1, sl ≥ sl+1. Suppose P∈ L(C). The score of x∈C given P is score(P,x) = sr
where r is the rank of x in P. For each profile P and x ∈ C, let score(P,x) = ∑n
i=1 score(Pi,x). A positional
scoring rule for a scoring vector(cid:126)s assigns to each profile P the set of candidates that maximize their score
according to (cid:126)s in P. That is, a voting method f is a positional scoring rule for a scoring vector (cid:126)s provided
that for all P ∈ L(C)V , f (P) = argmaxx∈Cscore(P,x). We study two such rules:
Borda: the positional scoring rule for (cid:104)n− 1,n− 2, . . . ,1,0(cid:105).
Plurality: the positional scoring rule for (cid:104)1,0, . . . ,0(cid:105).
(2) The Condorcet winner in a profile P is a candidate x ∈ C that is the maximum of the majority
ordering, i.e., for all y ∈ C, if x (cid:54)= y, then x >M
P y. The Condorcet voting method is:
7As usual, for a set A, A is the number of elements in A.
256
Strategic Voting Under Uncertainty About the Voting Method
(cid:40){x} if x is the Condorcet winner in P
C
if there is no Condorcet winner.
Condorcet(P) =
(3) The win-loss record for a candidate x ∈ C in a profile P is the number of candidates z such that
a majority prefers x to z in P minus the number of candidates z such that a majority prefers z to x in P.
Formally, for each P and x ∈ C, let wlP(x) = {z NetP(x,z) > 0}−{z NetP(z,x) > 0}. The Copeland
winners are the candidates with maximal win-loss records: Copeland(P) = argmaxx∈C(wlP(x)).
(4) The support for a candidate x ∈ C in a profile P is found by calculating for each candidate y (cid:54)= x
the number of voters who rank x above y and then taking the minimum of these values. Formally, for
each P and x ∈ C, let supp(x,P) = min({NP(x,y) y ∈ C,y (cid:54)= x}). The MaxMin (also known as Simpson's
Rule) winners are the candidates with maximal support: MaxMin(P) = argmaxx∈C(supp(x,P)).
(5) PluralityWRunoff: Calculate the plurality score for each candidate -- the number of voters
who rank the candidate first. If there are 2 or more candidates with the highest plurality score, remove
all other candidates and select the Plurality winners from the remaining candidates. If there is one
candidate with the highest plurality score, remove all candidates except the candidates with the highest
or second-highest plurality score, and select the Plurality winners from the remaining candidates.
(6) Hare: Iteratively remove all candidates with the fewest number of voters who rank them first,
until there is a candidate who is a majority winner, i.e., ranked first by a majority of voters. If, at some
stage of the removal process, all remaining candidates have the same number of voters who rank them
first (so all candidates would be removed), then all remaining candidates are selected as winners.
(7) Coombs: Iteratively remove all candidates with the most number of voters who rank them last,
until there is a candidate who is a majority winner. If, at some stage of the removal process, all remaining
candidates have the same number voters who rank them last (so all candidates would be removed), then
all remaining candidates are selected as winners.
(8) Baldwin: Iteratively remove all candidates with the smallest Borda score, until there is a single
candidate remaining. If, at some stage of the removal process, all remaining candidates have the same
Borda score (so all candidates would be removed), then all remaining candidates are selected as winners.
(9) Rather than removing candidates with the lowest Borda score, the next two methods remove all
candidates who have a Borda score below the average Borda score for all candidates. There are two
versions of this voting method [24]: StrictNanson iteratively removes all candidates whose Borda
score is strictly smaller than the average Borda score (of the candidates remaining at that stage), until
one candidate remains. WeakNanson iteratively removes all candidates whose Borda score is less than or
equal to the average Borda score (of the candidates remaining at that stage), until one candidate remains.
If, at some stage of the removal process, all remaining candidates have the same Borda score (so all
candidates would be removed), then all remaining candidates are selected as winners.
Definition 2.1. Let Methods be the set of 11 voting methods described above.
We are interested in situations in which the strategizing voter is uncertain about which voting method
will be used to determine the winner(s). As noted in Section 1, we represent such uncertainty by a set S
of voting methods, called the uncertainty set. This is the set of voting methods f such that it is consistent
with the strategizing voter's knowledge that f will be used.
Each of the above methods may select more than one winner for a given profile. Thus, to discuss
strategizing, one needs a notion of when one set of candidates is "preferable" to another for a particular
voter. The following are the standard notions from the literature on strategic voting (see, e.g., [35]).
Definition 2.2. Let P be a profile, i ∈ V , and X,Y ⊆ C. We define the following dominance notions, each
of which has a nonstrict (≥) and strict (>) version:
Y and ∃x ∈ X ∃y ∈ Y : xPiy.
weak: (a) X ≥weak
Y iff ∀x ∈ X ∀y ∈ Y : xRiy;
Y iff X ≥weak
(b) X >weak
Pi
Pi
Pi
Holliday and Pacuit
257
optimistic: (a) X ≥Opt
Y iff max(X,Pi) Ri max(Y,Pi);
Pi
pessimistic: (a) X ≥Pes
Pi Y iff min(X,Pi) Ri min(Y,Pi);
(b) X >Opt
Y iff max(X,Pi) Pi max(Y,Pi).
Pi
(b) X >Pes
Pi Y iff min(X,Pi) Pi min(Y,Pi).
3 Sure manipulation
If a voter is uncertain about which voting method from a set S of voting methods will determine the
winners of an election, the most conservative approach to strategic voting is to submit an insincere
ranking if and only if the voter is sure that by doing so, the set of winners will be strictly better from the
point of view of her true ranking (and the relevant dominance notion) no matter which voting method
from S is used. This approach is appropriate when there is a cost to submitting an insincere ranking
that a voter is only willing to incur if it will surely improve the set of winners. For example, in the
context of a hiring committee in which committee members know each other's true preferences over the
candidates, e.g., through deliberation, there may be a social cost in submitting an insincere ranking of the
candidates, which a committee member is willing to bear only if doing so is sure to result in a preferable
set of winners. These considerations motivate the following definition.
Definition 3.1. Let (P,i) be a pointed profile, ∆ a dominance notion, and S a set of voting methods. We
say that (P,i) witnesses sure ∆-manipulability for S if and only if there is a profile P(cid:48) differing from P only
in i's ranking such that ∀ f ∈ S : f (P(cid:48)) >∆
Pi f (P). We then say that (P,i) witnesses sure ∆-manipulability
for S by transitioning to P(cid:48) and that i has a sure ∆-manipulation incentive under S to transition to P(cid:48). A
profile P witnesses sure ∆-manipulability for S if and only if there is an i ∈ V such that (P,i) witnesses
sure ∆-manipulability for S. Finally, we say that S is susceptible to sure ∆-manipulation for (n,m) if and
only if there is an (n,m)-profile P that witnesses sure ∆-manipulability for S.
As an initial example, consider the Borda method. Recall that Borda is not resolute, so the Borda
winning set for a particular profile may contain multiple candidates. To obtain a resolute method from
an irresolute method f such as Borda, we may fix a tiebreaking mechanism, understood as a strict linear
order L on C, and define fL to be the resolute voting method defined by fL(P) = max( f (P),L). An
especially natural example of uncertainty about the voting method arises when there is uncertainty about
the tiebreaking mechanism to be used for a fixed voting method. Moreover, such uncertainty can be a
barrier to sure manipulation. For example, for n = 3, {Borda} is susceptible to sure weak dominance
manipulation, but this is not so when there is complete uncertainty about the tiebreaking mechanism.
Proposition 3.2. For any m ≥ 4, {BordaL L a linear order on C} is not susceptible to sure weak domi-
nance manipulation for (3,m).
Proof. Suppose for contradiction that there is a pointed profile (P,i) that witnesses sure weak dominance
manipulation for S = {BordaL L a linear order on C} by transitioning to P(cid:48). By the choice of S, it
follows that (i) every candidate in the Borda winning set for P(cid:48) is strictly preferred according to Pi to
every candidate in the Borda winning set in P. Let i's ranking of the candidates in P be αβγ.
First, the Borda winning set in P cannot contain α. For if it does, then there is no incentive to manip-
ulate under Bordaα>β >γ or Bordaα>γ>β , contradicting sure weak dominance manipulation. Second, the
Borda winning set in P cannot contain γ. For if it does, then by (i), the Borda winning set in P(cid:48) cannot
contain γ; but there is no P(cid:48) differing from P only in i's ranking such that γ is in the Borda winning set
in P but not in the Borda winning set in P(cid:48). Finally, we claim that the Borda winning set in P cannot be
{β}. Suppose it is. By (i), the Borda winning set in P(cid:48) cannot contain β or γ, so it must be {α}. But this
contradicts the fact that for 3 candidates, Borda is not single-winner manipulable [35, p. 57]. Having
ruled out all possible Borda winning sets in P, we obtain a contradiction.
258
Strategic Voting Under Uncertainty About the Voting Method
In this paper, we will focus on the case where S contains different voting methods, as in the following
example, rather than the same voting method with different tiebreaking rules.
Example 3.3. Consider the following (3,4)-profile P for C = {a,b,c} and V = {1,2,3,4}:
1
a
b
c
2
b
c
a
3
c
a
b
4
c
b
a
The winning sets for the voting methods are:
• {c} for Borda, Copeland, Hare, WeakNanson, Plurality,
PluralityWRunoff;
• {b,c} for Baldwin, Coombs, MaxMin, and StrictNanson.
If candidate 1 changes her ranking to bac, then all the methods select the winning set {b,c}. Suppose
P1 {c}, (P,1) witnesses sure ∆-dominance
that ∆ is either weak or optimistic dominance. Since {b,c} >∆
manipulability for any nonempty subset of methods from {Borda, Copeland, Hare, WeakNanson,
P1 {b,c}, this pointed profile (P,1) does not wit-
Plurality, PluralityWRunoff}. Since {b,c} (cid:54)>∆
ness sure ∆-dominance manipulability for any set of voting methods that contains one or more of the
following methods: Baldwin, Coombs, MaxMin, StrictNanson. Of course, there may be other pointed
profiles witnessing sure ∆-manipulation for sets of methods containing one of these methods.
the following profile P(cid:48) witnesses sure pessimist manipulation:
The above profile P does not witness sure pessimist manipulation for any set of voting methods. But
1
a
b
c
2
a
c
b
3
b
a
c
4
c
b
a
The winning sets for the voting methods are:
• {a} for Borda, Copeland, Hare, WeakNanson, Plurality, and
PluralityWRunoff;
• {a,b} for Baldwin, Coombs, MaxMin, and StrictNanson.
If candidate 1 changes her ranking to acb, then all the methods select the winning set {a}. Since
{a} >Pes
{a,b}, (P,1) witnesses sure pessimist manipulability for any nonempty
P(cid:48)
subset of methods from {Baldwin, Coombs, MaxMin, StrictNanson}, but no set containing any of the
1
following methods: Borda, Copeland, Hare, WeakNanson, Plurality, PluralityWRunoff.
{a,b} and {a,b} (cid:54)>Pes
P(cid:48)
1
3.1 Eliminating sure manipulation with a pair of voting methods
Our first question is whether uncertainty about the voting method may eliminate sure manipulation.
Definition 3.4. Let S be a set of voting methods. We say that S eliminates sure ∆-manipulation for (n,m)
iff S is not susceptible to sure ∆-manipulation for (n,m) but every nonempty S(cid:48) (cid:40) S is susceptible to sure
∆-manipulation for (n,m).
The numbers in Figure 1 represent the percentage of profiles that witness sure weak dominance
manipulation for (3,4) and (3,7). For example, 7.6% of (3,4)-profiles witness sure weak dominance
manipulation for {Plurality, Copeland}. The numbers along the diagonal give the percentage of
profiles witnessing weak dominance manipulation for a single voting method.8 The numbers highlighted
in red identify pairs of voting methods that eliminate sure weak dominance manipulation according to
Definition 3.4. For instance, 25% of the (3,4)-profiles witness weak dominance manipulation for Borda,
15% of the pointed (3,4)-profiles witness weak dominances manipulation for StrictNanson, but there
are no instances of sure weak dominance manipulation for {Borda, StrictNanson}.
8For 3 candidates, Hare and PluralityWRunoff always pick the same winners.
Holliday and Pacuit
259
Figure 1: Percentage of profiles witnessing sure weak dominance manipulation for (3,4) (left) and (3,7) (right).
To illustrate what happens with larger number of voters, we present in Figure 2 percentages for sure
weak dominance manipulation for Borda alone and Borda paired with several other voting methods.
The data for (4,4) was obtain by exhaustive search, while the data for (4,5) -- (4,70) was obtained by
generating 10,000 profiles for each (4,m) using the Impartial Culture model.
By exhaustive search from (3,4) -- (3,8),
we find that {Borda, Baldwin}, {Borda,
StrictNanson}, {WeakNanson, Baldwin},
and {WeakNanson, StrictNanson} are not
susceptible to sure weak dominance manip-
ulation, while each method individually is.
This raises the question of whether one can
prove that for all (3,m), those sets of meth-
ods eliminate sure weak dominance manipu-
lation. Indeed, we will prove this result.
Lemma 3.5. For any n ≥ 3 and m ≥
4,9 Baldwin, Borda, StrictNanson, and
WeakNanson are each susceptible to sure
weak dominance manipulation for (n,m).
Proof. Given a (3,m)-profile P, let P(cid:93)P2 be
the (3,m+2)-profile that results from adding
to P two fresh voters with rankings abc and
cba, respectively. The rankings of candidates by Borda score in P and P(cid:93) P2 are the same; the sets of
candidates with the lowest Borda scores in P and P(cid:93)P2 are the same; and the sets of candidates with less
than average (resp. less than or equal to average) Borda scores in P and P(cid:93)P2 are the same. Thus, the set
of Borda (resp. Baldwin, StrictNanson, WeakNanson) winners does not change from P and P(cid:93) P2.
It follows that if P witnesses ∆-manipulation for Borda (resp. Baldwin, StrictNanson, WeakNanson)
by transitioning to P(cid:48), then P(cid:93)P2 witnesses ∆-manipulation for Borda (resp. Baldwin, StrictNanson,
WeakNanson) by transitioning to P(cid:48) (cid:93) P2. Thus, to prove that one of these voting methods is susceptible
to ∆-dominance manipulation for (3,m) for any m ≥ 4, it suffices to show this for (3,4) and (3,5), as
9The lemma can also be proved by similar reasoning for n ≥ 4 and m ≥ 3. However, for n = 3 and m = 3, there are no
Figure 2: Percentage of (4,m)-profiles witnessing sure
weak dominance manipulation for Borda alone and
Borda paired with several other methods.
instances of weak dominance manipulation for these methods.
260
Strategic Voting Under Uncertainty About the Voting Method
susceptibility for all other pairs (3,m) then follows by the preceding observation about P(cid:93) P2. As we
verified using our Python script, there are indeed (3,4) and (3,5) profiles witnessing weak dominance
manipulability for all four methods individually.
For any (n,m)-profile P, let P+ be the (n+1,m)-profile that results from adding to P a fresh candidate
at the bottom of every voter's ranking. The set of Borda (resp. Baldwin, StrictNanson, WeakNanson)
winners does not change from P to P+. Thus, to prove that one of these methods is susceptible to
∆-dominance manipulation for (n,m) for any n ≥ 3, it suffices to show this for (3,m), as above.
The following main theorems, proved in the Appendix, show that uncertainty about the voting
method can entirely eliminate sure weak dominance manipulation.
Theorem 3.6. For any m ≥ 4, the sets {Borda, Baldwin} and {Borda, StrictNanson} eliminate sure
weak dominance manipulation for (3,m).
Theorem 3.7. For any m ≥ 4, the sets {WeakNanson, Baldwin} and {WeakNanson, StrictNanson}
eliminate sure weak dominance manipulation for (3,m).
By contrast, already for (3,5), {Baldwin, Borda, StrictNanson, WeakNanson} is susceptible to
sure optimistic dominance and sure pessimistic dominance. We give an example for optimistic domi-
nance, leaving pessimistic dominance as an exercise.
Example 3.8. Let P be the following (3,5) profile:
2
a
b
c
4
c
b
a
5
c
b
a
3
b
a
c
1
a
b
c
For another contrast to Theorems 3.6-3.7, when we increase to 4 candidates, {Baldwin, Borda,
Baldwin, Borda, StrictNanson, and WeakNanson choose {b}. If voter
1 changes her ranking to a c b, then {a,b,c} is the winning set for all four
methods. Since {a,b,c} >Opt
{b}, (P,1) witnesses sure optimist dominance
P1
manipulability for {Baldwin, Borda, StrictNanson, WeakNanson}.
StrictNanson, WeakNanson} is susceptible to sure weak dominance manipulation.
Example 3.9. Consider the following (4,3)-profile P:
1
a
b
c
d
2
b
d
c
a
3
c
a
b
d
is {b}, while the Baldwin, StrictNanson, and
The Borda winning set
WeakNanson winning sets are all {a,b,c}.
If voter 1 changes her rank-
is {a,b}, while the Baldwin,
ing to ad bc,
then the Borda winning set
StrictNanson, and WeakNanson winning sets are {a}. Since {a,b} >weak
{b}
P1
and {a} >weak
{a,b,c}, (P,1) witnesses sure weak dominance manipulability for
P1
{Baldwin, Borda, StrictNanson, WeakNanson}.
3.2 The failure of the Duggan-Schwartz theorem for sets of voting methods
A natural analogue of the Duggan-Schwartz theorem [9] for sure manipulation of sets of voting methods
would state that for any (n,m) with n ≥ 3 and set S of methods, each of which is non-imposed and
has no nominator, S is susceptible to sure optimistic or pessimistic dominance manipulation for (n,m).
However, this statement is false, as shown by the following example verified by our Python script.
Example 3.10. These sets of methods (which are non-imposed and have no nominator) eliminate sure
optimistic and pessimistic dominance manipulation for (3,6): {Baldwin, Condorcet}, {Condorcet,
Copeland}, {Condorcet, MaxMin}, {Condorcet, StrictNanson}, and {Condorcet, WeakNanson}.
Holliday and Pacuit
261
3.3 Cases where three methods are needed for elimination
So far we have only considered sets of two voting methods. But in some cases three voting methods are
needed to eliminate sure manipulation. We saw in Example 3.9 that {Borda, Baldwin} is susceptible to
sure weak dominance manipulation for (4,3). However, adding Coombs to {Borda, Baldwin} eliminates
sure weak dominance manipulation, as verified by our Python script (which also finds profiles witnessing
sure weak dominance manipulability for {Borda, Coombs} and {Coombs, Baldwin}).
Fact 3.11. {Borda, Coombs, Baldwin} eliminates sure weak dominance manipulation for (4,3).
3.4 (n,m) for which sure manipulation cannot be eliminated
So far we have focused on eliminating sure manipulation using uncertainty about the voting method.
However, as the following result shows, eliminating sure manipulation is not always possible.
Proposition 3.12. For every (n,m), there are n(cid:48) > n and m(cid:48) > m such that every nonempty subset of
Methods\{Condorcet} is susceptible to sure weak dominance manipulation for (n(cid:48),m(cid:48)).
Proof. First, we claim that if there is one (n,m)-profile that witnesses sure weak dominance manipula-
bility for every nonempty subset of Methods\{Condorcet}, then for all k ∈ N, there is an (n,m +24k)-
profile that does so. For any profile P, let P(cid:93)P24 be the result of adding 24 new voters to P, one with each
of the possible 24 rankings of {a,b,c,d}. It is easy to see that for any f ∈ Methods, f (P) = f (P(cid:93) P24).
It follows that if P witnesses sure ∆-dominance manipulability for S by transitioning to P(cid:48), then so does
P(cid:93) P24 by transitioning to P(cid:48) (cid:93) P24. In addition, using the construction from P to P+ in the proof of
Lemma 3.5, we can increase the number n of candidates, since for any f ∈ Methods, f (P) = f (P+).
Now consider the following (4,4)-profile P:
1
a
b
c
d
2
b
d
c
a
3
c
a
b
d
4
c
a
b
d
If voter 1 changes to bd ca,
Every method in Methods \ {Condorcet} chooses {c} as the set of win-
ners.
then the winning set for all methods
in Method\{Condorcet} becomes {b,c}.
{c}, (P,1)
witnesses sure weak dominance manipulability for any nonempty subset of
Method\{Condorcet}.
Since {b,c} >weak
P1
3.5 Reduction without elimination
Even when eliminating sure manipulation is not possible, one may hope to reduce it. For example,
in Figures 1 and 2, there are a number of cases in which a pair of methods does not eliminate sure
weak dominance manipulation but does reduce the number of profiles witnessing sure weak dominance
manipulation relative to either method individually. This motivates the following notions.
Definition 3.13. For any sets S and S(cid:48) of voting methods, S is less susceptible to sure ∆-manipulation
than S(cid:48) for (n,m)-profiles (resp. pointed (n,m)-profiles) iff there are fewer (n,m)-profiles (resp. pointed
(n,m)-profiles) witnessing sure ∆-manipulation for S than there are for S(cid:48).
Definition 3.14. A set S of methods improves on all its subsets with respect to sure ∆-manipulation for
(n,m) iff S is less susceptible to sure ∆-manipulation for (n,m)-profiles than any nonempty S(cid:48) (cid:40) S.
Such improvement is especially pronounced with optimistic and pessimistic dominance, as in Figure 3.
Even if a set S does not improve on all of its subsets, that S is less susceptible to sure ∆-manipulation
than one of its subset may still be significant. An election designer who intends to use method f may
wish to leave voters uncertain between f and f (cid:48) in order to reduce the chance that voters will surely
262
Strategic Voting Under Uncertainty About the Voting Method
Figure 3: Percentage of profiles witnessing sure optimistic dominance manipulation for (3,7) (left) and sure
pessimistic dominance manipulation for (3,6) (right).
manipulate, relative to what would happen if voters knew the method was f , even if there is no reduction
relative to what would happen if the planner intended to use f (cid:48) and voters knew this. For example, in
Figure 1 for (3,7), someone intending to use Plurality could reduce the percentage of profiles in which
a voter will surely manipulate from 29% to 9% by leaving the voter uncertain between Plurality and
Hare, even though an election designer intending to use Hare would have no incentive to do so, since
the percentage of profiles witnessing sure manipulation for Hare by itself is already 9%.
A striking pattern in Figures 1, 2, and 3 is that pairing Borda with another method leads to significant
reductions in sure dominance manipulation. Using one or more methods to help reduce manipulation for
a preferred method is even more important in the case of safe manipulation discussed next.
4 Safe manipulation
Pi f (P) and ∃ f ∈ S : f (P(cid:48)) >∆
We now turn to a less conservative approach to strategic voting under uncertainty about the voting
method: submit an insincere ranking whenever you know that doing so will lead to an outcome that
is at least as good and might lead to a better outcome.
Definition 4.1. Let (P,i) be a pointed profile, ∆ a dominance notion, and S a set of voting methods.
Then (P,i) witnesses safe ∆-manipulability for S iff there is a profile P(cid:48) differing from P only in i's ballot
such that: ∀ f ∈ S : f (P(cid:48)) ≥∆
Pi f (P). We then say that (P,i) witnesses safe
∆-manipulability for S by transitioning to P(cid:48). A profile P witnesses safe ∆-manipulability for S iff there
is an i ∈ V such that (P,i) witnesses safe ∆-manipulability for S.
Remark 4.2. A variant of safe manipulability, which we will call harmless manipulability, says that
you should submit an insincere ranking whenever you know that doing so will not lead to a worse
outcome and might lead to a better outcome: ∀ f ∈ S : f (P(cid:48)) (cid:54)<∆
Pi f (P). Since
f (P(cid:48)) (cid:54)<weak
f (P), harmless weak dominance manipulability does not
imply safe weak dominance manipulability for a given (P,i) transitioning to P(cid:48). But for pessimistic and
optimistic dominance, harmless and safe manipulability are equivalent.
Pi f (P) and ∃ f ∈ S : f (P(cid:48)) >∆
f (P) does not imply f (P(cid:48)) ≥weak
Pi
Pi
For certain profiles and uncertainty sets S, submitting an insincere ranking may lead to a better
outcome with one method in S but a worse outcome with another method in S, in which case it is not safe
to manipulate using that insincere ranking.
Holliday and Pacuit
263
Example 4.3. Let P be the following (3,5)-profile:
1
c
b
a
2
a
c
b
3
b
a
c
4
c
b
a
5
a
c
b
Here Hare(P) = {a}, Borda(P) = {c}, and MaxMin(P) = {a,c}. Suppose
voter 1 changes her ranking to b a c, resulting in P(cid:48). Then Hare(P(cid:48)) = {b},
Borda(P(cid:48)) = {a}, and MaxMin(P(cid:48)) = {a,b}. Since {b} >weak
{a}, voter
1 has an incentive to manipulate with Hare. But voter 1 does not have an
incentive to manipulate with Borda, since {c} >weak
{a}, or MaxMin, since
P1
{a,b} (cid:54)≥weak
P1
{a,c}.
P
Thus, (P,1) does not witness safe weak dominance manipulation for Borda together with any nonempty
subset of {Hare, MaxMin}.
It seems too much to hope to eliminate safe manipulation by adding reasonable methods to S; for this
would require that for every profile in which a manipulation results in a better outcome for one method
in S, it results in a worse outcome for another method in S, which seems unlikely to hold for a set of
reasonable methods. However, one can eliminate safe manipulation by adding methods that would be
considered unreasonable by themselves.
Example 4.4. For any distinct x,y ∈ C and i ∈ V , let fx,y,i be the method such that fx,y,i(P) selects as the
winner whichever of x and y is ranked higher according to Pi (cf. [26]). It is easy to see that if S contains
fx,y,i for each distinct x,y ∈ C, then i cannot safely manipulate with S.
Although eliminating safe manipulation with reasonable methods may be too much to hope for, one
can reduce safe manipulation. Thus, we are interested in the following analogue of Definition 3.13.
Definition 4.5. For any sets S and S(cid:48) of voting methods, S is less susceptible to safe ∆-manipulation
than S(cid:48) for (n,m)-profiles (resp. pointed (n,m)-profiles) iff there are fewer (n,m)-profiles (resp. pointed
(n,m)-profiles) witnessing sure safe ∆-manipulation for S than there are for S(cid:48).
Figure 4: Percentage of profiles witnessing safe weak dominance manipulation for (3,6) (left) and (3,7) (right).
Figure 4 shows the percentage of profiles witnessing safe weak dominance manipulation for (3,6)
and (3,7) for sets of two voting methods. All of the following can happen: (1) Unlike with sure
manipulation, with safe manipulation a voter who is uncertain between methods f and f (cid:48) may have
an incentive to manipulate on more profiles than a voter who knows the method is f and more pro-
files than a voter who knows the method is f (cid:48). E.g., this happens for (3,7) when f = Borda and
f (cid:48) = Hare. (2) A voter who is uncertain between methods f and f (cid:48) may have an incentive to manip-
ulate on fewer profiles than a voter who knows the method is f and fewer profiles than a voter who
264
Strategic Voting Under Uncertainty About the Voting Method
knows the method if f (cid:48). E.g., this happens for (3,6) when f = Coombs and f (cid:48) = Hare. In Figure 4,
all examples of this phenomenon are indicated with the blue boxes. For (3,7), there are no examples.
(3) A voter who is uncertain between meth-
ods f and f (cid:48) may have an incentive to ma-
nipulate on fewer profiles than a voter who
knows the method is f but more profiles than
a voter who knows the method is f (cid:48). E.g.,
this happens with (3,6) when f = Borda
and f (cid:48) = Hare. In case (3), an election de-
signer who intends to use method f may de-
cide to leave voters uncertain between f and
f (cid:48) in order to decrease the chance that voters
will safely manipulate (cf. the end of Section
3.5).10 E.g., Figure 5 shows how an election
designer intending to use Coombs could pair
Coombs with several other methods to form
an uncertainty set of two methods in order to
decrease the percentage of profiles in which
a voter will safely manipulate.
Figure 5: Percentage of (4,m)-profiles witnessing safe
weak dominance manipulation for Coombs alone and
Coombs paired with several other methods.
5 Expected manipulation
Our last approach to strategic voting under uncertainty about the voting method is the most liberal:
assuming one's uncertainty about the voting method is given by a lottery on the set of voting methods,
submit an insincere ranking if and only if doing so is more likely to lead to a better outcome than to lead
to a worse outcome. Recall that a lottery on a set Y is a function ν : Y → [0,1] such that ∑y∈Y ν(y) = 1.
For X ⊆ Y , let ν(X) = ∑x∈X ν(x).
Definition 5.1. Let (P,i) be a pointed profile, ∆ a dominance notion, S a set of voting methods, and ν a
lottery on S. Then (P,i) witnesses ν-expected ∆-manipulability for S if and only if there is a profile P(cid:48)
differing from P only in i's ballot such that ν({ f ∈ S f (P(cid:48)) >∆
Pi f (P)}).
Then we say (P,i) witnesses ν-expected ∆-manipulability for S by transitioning to P(cid:48).
Pi f (P)}) > ν({ f ∈ S f (P(cid:48)) <∆
For simplicity, here we focus on ν being the uniform lottery on S, in which case we simply speak
of 'expected ∆-manipulability' instead of 'ν-expected ∆-manipulability'. For ν uniform, this amounts to
simply counting the number of voting methods in S that lead to a better outcome vs. a worse outcome.
Example 5.2. The pointed profile (P,1) in Example 4.3 does not witness expected weak dominance
manipulability for {Borda, Hare} by transitioning to P(cid:48) since voter 1's manipulation results in a better
outcome according to one method (Hare) but a worse outcome according to another method (Borda).
However, like Hare, Baldwin chooses {a} as the set of winners in P and {b} as the set in P(cid:48). Hence
(P,1) witnesses expected weak dominance manipulability for {Baldwin, Borda, Hare} by transitioning
to P(cid:48), as two methods lead to a better outcome and only one leads to a worse outcome.
10In this case, a sophisticated voter with access to the manipulation data for f , f (cid:48), and { f , f (cid:48)} could infer that the election
designer intenders to use f . For example, if S = {Borda, Hare} for (3,6), then such a voter could infer that the designer intends
to use Borda, since a designer who intends to use Hare will not decrease manipulation with S = {Borda, Hare}. If the designer
anticipates that voters are sophisticated in this way, then she should move from a single method to a pair only in case (2).
(4, 4)(4, 5)(4, 6)(4, 7)(4, 10)(4, 25)(4, 40)(4, 55)(4, 70)Voting Sizes1020304050PercentagePercentage of profiles witnessing safe weak dominance manipulationUncertainty SetCoombsCoombs,WeakNansonCoombs,CopelandCoombs,BaldwinCoombs,MaxMin Holliday and Pacuit
265
The following fact is immediate from the definitions.
Fact 5.3. If (P,i) witnesses safe ∆-dominance manipulability for S, then (P,i) witnesses expected ∆-
dominance manipulability for S.
The converse of Fact 5.3 does not hold, as shown by Example 5.2. But for S = 2 expected weak
dominance manipulability is equivalent to harmless weak dominance manipulability (recall Remark 4.2).
Using the obvious notion of less susceptible to expected ∆-dominance manipulation analogous to
Definitions 3.13 and 4.5, we have the following result, inspired by [26], showing how adding a method
to the set S may reduce expected manipulation.
Proposition 5.4. Suppose S is a set of voting methods such that for some f ∈ S, (n,m), pointed (n,m)-
profile (P,i), and a,b ∈ C, (1) (P,i) witnesses weak dominance manipulability for f , and (2) for any
P(cid:48) differing from P only in i's ranking, if (P,i) witnesses weak dominance manipulability for f by
transitioning to P(cid:48), then P(cid:48) differs from P in i's ranking of a vs. b, and for all g ∈ S\{ f}, g(P) = g(P(cid:48)).
Then where fa,b,i is the method defined in Example 4.4, S∪{ fa,b,i} is less susceptible to expected weak
dominance manipulation than S for pointed (n,m)-profiles.
Proof. The key properties of fa,b,i are that (i) there are no profiles P and P(cid:48) differing only in i's rank-
ing such that we have fa,b,i(P(cid:48)) >weak
fa,b,i(P), and (ii) if P and P(cid:48) differ in i's ranking of a vs. b, then
fa,b,i(P(cid:48)) <weak
fa,b,i(P). It follows from (i) that any pointed profile witnessing expected weak dominance
manipulability for S∪{ fa,b,i} also witnesses expected weak dominance manipulability for S. Thus, to
prove the proposition, we need only find a pointed profile witnessing expected weak dominance manip-
ulability for S but not S∪{ fa,b,i}. By assumption, there is a pointed (n,m)-profile (P,i) and a,b ∈ C
satisfying items (1) and (2) of the proposition. Given (1), consider any P(cid:48) such that (P,i) witnesses
weak dominance manipulability for f by transitioning to P(cid:48). It then follows by (2) that (P,i) witnesses
expected weak dominance manipulability for S by transitioning to P(cid:48). However, we claim that (P,i) does
not witness expected weak dominance manipulability for S∪{ fa,b,i} by transitioning to P(cid:48). This follows
from the facts that f and fa,b,i are equally likely according to the uniform measure, {h ∈ S ∪{ fa,b,i}
h(P(cid:48)) <weak
h(P)} = { fa,b,i} by (2) and (ii), and {h ∈ S∪{ fa,b,i} h(P(cid:48)) >weak
Pi
h(P)} = { f} by (2).
Pi
Pi
Pi
We conclude this section with an example in which the conditions of Proposition 5.4 apply.
Example 5.5. Consider the following (3,5)-profile P:
1
c
a
b
2
c
a
b
3
a
c
b
4
a
c
b
5
a
c
b
Let S = {Borda, Coombs}.
(P,i) witnesses the weak manipulability of
Borda by transitioning to the profile P(cid:48) in which i's new ranking is cba,
as the Borda winning set in P is {a}, the Borda winning set in P(cid:48) is {a,c},
and {a,c} >weak
P1
{a}.
Moreover, this is the only P(cid:48) differing from P only in i's ranking such that (P,i) witnesses the weak
manipulability of Borda by transitioning to P(cid:48). Finally, the winning set for Coombs in both P and P(cid:48) is
{a}. Thus, the conditions of Proposition 5.4 are satisfied for f = Borda, so {Borda, Coombs, fa,b,i} is
less susceptible to expected weak dominance manipulation than {Borda, Coombs}.
6 Relation to probabilistic social choice
In this section, we briefly relate our work to strategic voting in the setting of probabilistic social choice.
Definition 6.1. A probabilistic social choice function (PSCF) is a function F assigning to each profile P
a lottery F(P) on C.
266
Strategic Voting Under Uncertainty About the Voting Method
To define manipulation of PSCFs, we need a notion of when a voter prefers one lottery to another.
Among many possible options (see, e.g., [4, Sec. 1.3.2]), the following is popular.
Definition 6.2. Let µ and µ(cid:48) be lotteries on C and (P,i) a pointed profile. We say that µ stochastically
dominates µ(cid:48) in (P,i) if and only if for every x ∈ C, the probability that µ selects a candidate ranked at
least as highly as x by i is greater than or equal to the probability that µ(cid:48) selects a candidate ranked at
least as highly as x by i:
∀x ∈ C : ∑
µ(y) ≥ ∑
µ(cid:48)(y).
y:yRix
y:yRix
We write µ (cid:37)Pi µ(cid:48) if µ stochastically dominates µ(cid:48) in (P,i) and µ (cid:31)Pi µ(cid:48) if µ (cid:37)Pi µ(cid:48) but µ(cid:48) (cid:54)(cid:37)Pi µ.
Note that µ (cid:37)Pi µ(cid:48) if and only if for every utility function on C that is compatible with Pi, the expected
utility of µ is at least as great as the expected utility of µ(cid:48) (see, e.g., [3, p. 302-3]).
Definition 6.3. Let F be a PSCF. A pointed profile (P,i) witnesses stochastic dominance manipulability
for F if and only if there is a profile P(cid:48) differing only in i's ranking such that F(P(cid:48)) (cid:31)Pi F(P).11 We then
say that (P,i) witnesses stochastic dominance manipulability for F by transitioning to P(cid:48).
For results on stochastic dominance manipulability, see [4].
To relate the above notions to this paper, we observe how any set S of voting methods gives rise to
a PSCF, assuming that (i) each method in S is equally likely to be used and (ii) each candidate in the
set of winners selected by a method is equally likely to be chosen as the unique winner by a tiebreaking
mechanism. Then the probability that a given candidate a ∈ C will be chosen as the winner is:
Pr(a wins) = ∑
f∈S
= ∑
(cid:16)
f∈S
a∈ f (P)
∑
f∈S
a∈ f (P)
Pr(a wins f is used)× Pr( f is used)
Pr(a wins f is used)× Pr( f is used) + ∑
f∈S
a(cid:54)∈ f (P)
= ∑
f∈S
a∈ f (P)
Pr(a wins f is used)× Pr( f is used)
f (P) × 1
S .
0× 1
S
1
f (P) × 1
S
∑
f∈S
a(cid:54)∈ f (P)
(cid:17)
(cid:16)
(cid:17)
1
=
+
Thus, for a profile P and set S of voting methods, we define the lottery µP
S on C by
S (a) = ∑
µP
f∈S
a∈ f (P)
1
f (P) × 1
S .
Finally, for any set S of voting methods, we define the PSCF FS by FS(P) = µP
S .
Remark 6.4. For any set S of voting methods and lottery ν on S, one can define a PSCF FS,ν in the
obvious way by weighting the methods in S according to ν (and one can modify assumption (ii) with a
non-uniform measure for tiebreaking). For simplicity, here we focus on the uniform measure on S.
The following fact can be verified from the definitions.
11Gibbard's [14] notion of srategyproofness for a PSCF F is equivalent to the condition (called strong SD-strategyproofness
in [4]) that for every pointed profile (P,i) and profile P(cid:48) differing only in i's ranking, F(P) (cid:37)Pi F(P(cid:48)), i.e., there is no profile
(P,i) witnessing manipulation in the sense that there exists a profile P(cid:48) differing only in i's ranking such that F(P) (cid:54)(cid:37)Pi F(P(cid:48))
(which Gibbard states in the equivalent form: there is some utility function on C such that the expected utility of F(P(cid:48)) is
greater than that of F(P)). Following Brandt [4], we prefer the notion of (P,i) witnessing stochastic dominance manipulation
in Definition 6.3, which is the notion used in Brandt's definition of (weak) SD-strategyproofness.
Holliday and Pacuit
267
Fact 6.5. If (P,i) witnesses safe weak dominance manipulation for S by transitioning to P(cid:48), then (P,i)
witnesses stochastic dominance manipulation for FS by transitioning to P(cid:48).
However, stochastic manipulation does not imply safe weak manipulation.
Example 6.6. Consider the following (3,4)-profile P for C = {a,b,c} and V = {1,2,3,4}:
1
a
b
c
2
a
c
b
3
b
a
c
4
b
a
c
The Coombs, Copeland, and Hare winning set is {a,b}.
If voter 1 changes
to the ranking cab, transitioning to the profile P(cid:48), then the new winning set for
Coombs and Copeland is {a} and the new winning set for Hare is {b}. Note that
{a} >weak
{b}, so voter 1's new ranking leads to a better
P1
weak dominance outcome according to Coombs and Copeland but a worse weak
dominance outcome according to Hare.
{a,b} and {a,b} >weak
P1
Pi
Pi
Thus, (P,1) does not witness safe weak dominance manipulability by transitioning to P(cid:48).
Let S = {Coombs, Copeland, Hare}. The lottery FS(P) is [a : 1/2, b : 1/2, c : 0]. The lottery FS(P(cid:48))
is [a : 2/3, b : 1/3, c : 0]. As voter 1's ranking in P is abc, FS(P(cid:48)) stochastically dominates FS(P), since
the probability according to FS(P(cid:48)) of selecting a candidate at least as good as b is equal to the probability
according to FS(P) (both are 1), and the probability according to FS(P(cid:48)) of selecting a candidate at least
as good as a (probability 2/3) is greater than the probability according to FS(P) (probability 1/2). So
(P,1) witnesses stochastic dominance manipulability for FS(P) by transitioning to P(cid:48).
It is easy to see abstractly that stochastic dominance manipulation also does not imply expected
weak dominance manipulation: e.g., if S = { f1, f2} and the transition from (P,i) to P is such that
f2(P(cid:48)), then the transition does not witness expected weak manip-
f1(P) <weak
ulation, but the amount by which f1 increases the probability of getting a preferred candidate may be
greater than the amount by which f2 decreases the probability of getting a preferred candidate, so that
the lottery Ff1, f2(P(cid:48)) stochastically dominates the lottery Ff1, f2(P). The same idea applies for more than
two methods. It remains to be seen whether an example of this kind exists with standard voting methods.
We can, however, use standard methods to show that expected weak dominance manipulation does
f1(P(cid:48)) and f2(P) >weak
not imply stochastic weak dominance manipulation.
Example 6.7. Let P and P(cid:48) be the profiles in Example 4.3 and S = {Baldwin, Borda, Hare}. In P,
Baldwin and Hare select a as the winner, while Borda selects c as the winner. Thus, the lottery FS(P)
is [a : 2/3, b : 0, c : 1/3]. In P(cid:48), Baldwin and Hare select b as the winner, while Borda selects a as the
winner. Thus, the lottery FS(P(cid:48)) is [a : 1/3, b : 2/3, c : 0]. As voter 1's ranking in P is cba, FS(P(cid:48)) does
not stochastically dominate FS(P), since the probability according to FS(P(cid:48)) of selecting a candidate at
least as good as c is not greater than or equal to the probability according to FS(P) of selecting a candidate
at least as good as c. Thus, (P,1) does not witness stochastic dominance manipulability for FS(P) by
transitioning to P(cid:48). However, as two of the three methods in S lead to a better set of winners (in the sense
of weak dominance) in P(cid:48) than in P according to voter 1's ranking in P, (P,1) does witness expected
weak dominance manipulability for {Baldwin, Borda, Hare} by transitioning to P(cid:48).
7 Conclusion
In this paper, we have shown that uncertainty about the voting method can be used as a barrier to ma-
nipulation. Considering such uncertainty led to three decision rules that voters may use to decide when
to manipulate: sure, safe, and expected manipulation. Related issues arise when studying probabilistic
voting methods, though Section 6 shows that our notions of sure, safe, and expected manipulation do not
collapse to a standard notion of stochastic manipulation of probabilistic voting methods.
268
Strategic Voting Under Uncertainty About the Voting Method
This initial study relied heavily on computer searches. Two natural next steps are (i) to prove addi-
tional possibility (or impossibility) theorems, like our Theorems 3.6-3.7, and (ii) to incorporate uncer-
tainty about voting methods into the asymptotic analysis of strategic voting as the number of candidates
or voters increases (see, e.g., [33, 23]). Finally, a full analysis should take into account both uncertainty
about the voting method and uncertainty about how others will vote.
References
[1] Haris Aziz, Florian Brandl, Felix Brandt & Markus Brill (2018): On the tradeoff between efficiency and
strategyproofness. Games and Economic Behavior 110, pp. 1 -- 18, doi:10.1016/j.geb.2018.03.005.
[2] Jean-Pierre Benoit (2002): Strategic Manipulation in Voting Games When Lotteries and Ties Are Permitted.
Journal of Economic Theory 102(2), pp. 421 -- 436, doi:10.1006/jeth.2001.2794.
[3] Anna Bogomolnaia & Herv´e Moulin (2001): A New Solution to the Random Assignment Problem. Journal
of Economic Theory 100(2), pp. 295 -- 328, doi:10.1006/jeth.2000.2710.
[4] Felix Brandt (2017): Rolling the dice: Recent results in probabilistic social choice. In Ulle Endriss, editor:
Trends in Computational Social Choice, AI Access, pp. 3 -- 26.
[5] Samir Chopra, Eric Pacuit & Rohit Parikh (2004): Knowledge-theoretic properties of strategic voting.
In: Proceedings of the 8th European Conference on Logics in Artificial Intelligence (JELIA), pp. 18 -- 30,
doi:10.1007/11493402 15.
[6] Vincent Conitzer, Tobby Walsh & Lirong Xia (2011): Dominating manipulations in voting with partial
information. In: Proceedings of AAAI 2011, pp. 638 -- 643.
[7] Vincent Conitzer & Toby Walsh (2016): Barriers to Manipulation in Voting. In: Handbook of Computational
Social Choice, Cambridge University Press, pp. 127 -- 145, doi:10.1017/CBO9781107446984.007.
[8] Keith Dowding & Martin Van Hees (2008): In Praise of Manipulation. British Journal of Political Science
38(1), pp. 1 -- 15, doi:10.1017/S000712340800001X.
[9] John Duggan & Thomas Schwartz (2000): Strategic manipulability without resoluteness or shared beliefs:
Gibbard-Satterthwaite generalized. Social Choice and Welfare 17(1), pp. 85 -- 93, doi:10.1007/PL00007177.
[10] Piotr Faliszewski & Ariel D. Procaccia (2010): AI's War on Manipulation: Are We Winning? AI Magazine
31(4), pp. 53 -- 64, doi:10.1609/aimag.v31i4.2314.
[11] Allan M. Feldman (1980): Strongly nonmanipulable multi-valued collective choice rules. Public Choice
35(4), pp. 503 -- 509, doi:10.1007/BF00128127.
[12] Peter Gardenfors (1976): Manipulation of social choice functions. Journal of Economic Theory 13, pp.
217 -- 228, doi:10.1016/0022-0531(76)90016-8.
[13] Allan Gibbard (1973): Manipulation of Voting Schemes: A General Result. Econometrica 41(4), pp. 587 --
601, doi:10.2307/1914083.
[14] Allan Gibbard (1977): Manipulation of Schemes that Mix Voting with Chance. Econometrica 45(3), pp.
665 -- 681, doi:10.2307/1911681.
[15] Georges-Th´eodule Guilbaud (1952): Les th´eories de l'int´eret gen´eral et le probl´emelogique de l'agr´egation.
Economie Appliqu´ee 5(4), pp. 501 -- 551.
[16] Aanund Hylland (1980): Strategy proofness of voting procedures with lotteries as outcomes and infinite sets
of strategies. University of Oslo.
[17] Egor Ianovski, Lan Yu, Edith Elkind & Mark C. Wilson (2011): The Complexity of Safe Manipulation under
Scoring Rules. In: Proceedings of the Twenty-Second International Joint Conference on Artificial Intelli-
gence, pp. 246 -- 251, doi:10.5591/978-1-57735-516-8/IJCAI11-052.
Holliday and Pacuit
269
[18] Richard Jeffrey (1971): Statistical Explanation vs. Statistical Inference.
In Wesley C. Salmon, ed-
itor: Statistical Explanation and Statistical Relevance, University of Pittsburgh Press, pp. 19 -- 28,
doi:10.2307/j.ctt6wrd9p.5.
[19] Jerry S. Kelly (1977): Strategy-Proofness and Social Choice Functions Without Single-valuedness. Econo-
metrica 45(2), pp. 439 -- 446, doi:10.2307/1911220.
[20] Jerry S. Kelly (1985): Minimal Manipulability and Local Strategy-Proofness. Social Choice and Welfare
5(1), pp. 81 -- 85, doi:10.1007/BF00435499.
[21] Reshef Meir (2018): Strategic Voting. Synthesis Lectures on Artificial Intelligence and Machine Learning,
Morgan & Claypool.
[22] Samuel Merrill (1982): Strategic voting in multicandidate elections under uncertainty and under risk. In:
Power, voting, and voting power, Springer, pp. 179 -- 187, doi:10.1007/978-3-662-00411-1 12.
[23] Elchanan Mossel & Mikl´os Z. R´acz (2015): A quantitative Gibbard-Satterthwaite theorem without neutrality.
Combinatorica 35(3), pp. 317 -- 387, doi:10.1007/s00493-014-2979-5.
[24] Emerson M. S. Niou (1987): A Note on Nanson's Rule.
doi:10.1007/BF00123006.
Public Choice 54(2), pp. 191 -- 193,
[25] Shmuel Nitzan (1985): The vulnerability of point-voting schemes to preference variation and strategic ma-
nipulation. Public Choice 47(2), pp. 349 -- 370, doi:10.1007/BF00127531.
[26] Matias Nunez & Marcus Pivato (2019): Truth-revealing voting rules for large populations. Games and
Economic Behaviour 113, pp. 285 -- 305, doi:10.1016/j.geb.2018.09.009.
[27] Martin Osborne & Ariel Rubinstein (2003): Sampling equilibrium, with an application to strategic voting.
Games and Economic Behavior 45(2), pp. 434 -- 441, doi:10.1016/S0899-8256(03)00147-7.
[28] Eric Pacuit (2019): Voting Methods. In Edward N. Zalta, editor: The Stanford Encyclopedia of Philosophy,
Metaphysics Research Lab, Stanford University.
[29] Peter Railton (1981): Probability, Explanation, and Information.
doi:10.1007/BF01063889.
Synthese 48(2), pp. 233 -- 256,
[30] Annemieke Reijngoud & Ulle Endriss (2012): Voter response to iterated poll information. In: Proceedings
of the 11th International Conference on Autonomous Agents and Multiagent Systems, pp. 635 -- 644.
[31] Mark Satterthwaite (1973): The Existence of a Strategy Proof Voting Procedure. Ph.D. thesis, University of
Wisconsin.
[32] Mark Satterthwaite (1975): Strategy-proofness and Arrow's Conditions: Existence and Correspondence The-
orems for Voting Procedures and Social Welfare Functions. Journal of Economic Theory 10(2), pp. 187 -- 217,
doi:10.1016/0022-0531(75)90050-2.
[33] Arkadii Slinko (2002): On Asymptotic Strategy-Proofness of Classical Social Choice Rules. Theory and
Decision 52(4), pp. 389 -- 398, doi:10.1023/A:1020240214900.
[34] Arkadii Slinko & Shaun White (2014): Is it ever safe to vote strategically? Social Choice and Welfare 43(2),
pp. 403 -- 427, doi:10.1007/s00355-013-0785-4.
[35] Alan D. Taylor (2005): Social Choice and the Mathematics of Manipulation. Cambridge University Press,
Cambridge, doi:10.1017/CBO9780511614316.
[36] Hans van Ditmarsch, Jerome Lang & Abdallah Saffidine (2013): Strategic voting and the logic of knowledge.
In: Proceedings of the 14th Conference on Theoretical Aspects of Rationality and Knowledge (TARK), pp.
196 -- 205.
A Proof of Theorem 3.6
In this appendix, we prove Theorem 3.6: for any m ≥ 4, the sets {Borda, Baldwin} and {Borda,
StrictNanson} eliminate sure weak dominance manipulation for (3,m).
270
Strategic Voting Under Uncertainty About the Voting Method
Proof. Given Lemma 3.5, we need only show that the two sets of methods are not susceptible to sure
weak dominance manipulation for (3,m). Toward a contradiction, suppose (P,i) is a pointed (3,m)-
profile witnessing sure weak dominance manipulation for either of the two sets by transitioning to a
profile P(cid:48). Suppose i's ballot in P is αβγ.
Claim 1: the Borda winning set in P(cid:48) does not contain all three candidates. This follows by analyzing
the possible sets of Borda winners in P. A three-way tie {α,β ,γ} does not weakly dominate any of the
following sets for i: {α}, {β}, {α,β}, {α,γ}, {α,β ,γ}. This leaves only the winning sets {β ,γ} and
{γ} to consider. But voter i cannot manipulate so as to change the set of Borda winners from {β ,γ}
to {α,β ,γ}; for the Borda scores of β and γ must still be tied in P(cid:48), so i's ranking in P(cid:48) must be βγα,
which decreases α's Borda score. In addition, voter i cannot manipulate so as to change the set of Borda
winners from {γ} to {α,β ,γ}. For if i ranks γ higher in P(cid:48), then γ's Borda score increases by at least
one, and no other candidate's Borda score increases by more than one, so the set of winners is still {γ};
hence i's ranking in P(cid:48) must be βαγ, but this fails to add α to the set of winners.
Label the candidates in order of ascending Borda score in P(cid:48) as ϕ, ψ, and χ. Thus, by Claim 1,
B(cid:48)(ϕ) < B(cid:48)(ψ) ≤ B(cid:48)(χ) or
B(cid:48)(ϕ) = B(cid:48)(ψ) < B(cid:48)(χ),
(1)
(2)
where B(cid:48) indicates the Borda score in P(cid:48). Let B(ϕ),B(ψ),B(χ) be the Borda scores of ϕ, ψ, and χ in P.
Claim 2: i's ranking changes from P to P(cid:48) either by switching her 1st and 2nd placed candidates or
by switching her 2nd and 3rd candidates; thus, one candidate's Borda score remains the same, and no
candidate's Borda score changes by more than one point. Suppose for a contradiction that i's 3rd place
candidate in P is her 1st place candidate in P(cid:48) or her 1st place candidate in P is her 3rd place candidate
in P(cid:48). Then since i's ranking in P is αβγ, we have that i's ranking in P(cid:48) is either γβα, γαβ , or βγα.
But we claim i does not have an incentive to transition to any of these rankings from the original ranking
αβγ. The first two transitions are such that the only candidate to increase in Borda score is i's last place
candidate, which never improves the winning set for i. The transition from αβγ to βγα does not improve
the winning set if the winning set in P contains α; and if the winning set in P is {β}, {γ}, or {β ,γ},
again the transition from αβγ to βγα does not change the winning set. Thus, we have a contradiction
with the assumption that (P,i) witnesses sure weak dominance manipulation for Borda.
Using Claim 2, we can rule out case (2) above. For (3,m), Borda is not single-winner manipulable
[35, p. 57, Exercise 10], which means that if {χ} is the set of Borda winners in P(cid:48), then the set of Borda
winners in P cannot be a singleton, so it must be one of {ϕ,ψ, χ}, {ψ, χ}, {ϕ, χ}, or {ϕ,ψ}. By Claim
2, there is no way i can change the Borda scores from B(ϕ) = B(ψ) = B(χ) to B(cid:48)(ϕ) = B(cid:48)(ψ) < B(cid:48)(χ),
so we can rule out {ϕ,ψ, χ}. Also by Claim 2, in order for i to change the Borda scores from B(ϕ) <
B(ψ) = B(χ) to B(cid:48)(ϕ) = B(cid:48)(ψ) < B(cid:48)(χ), her ranking must go from ψχϕ to χψϕ or from ϕψχ to ϕχψ,
but then the new winning set {χ} is worse for i than the original {ψ, χ}. Thus, we can rule out {ψ, χ},
and by the same reasoning, {ϕ, χ}. Finally, by Claim 2, there is no way for i to change the Borda scores
from B(χ) < B(ϕ) = B(ψ) to B(cid:48)(ϕ) = B(cid:48)(ψ) < B(cid:48)(χ), so we can rule out {ϕ,ψ}. Thus, (1) holds.
Claim 3: ϕ has the unique below average Borda score in P and P(cid:48). We argue by cases.
Case 1: B(cid:48)(ψ) = B(cid:48)(χ). Then it is immediate from (1) that ϕ has the unique below average Borda
score in P(cid:48). We now show that ϕ has the unique below average Borda score in P. Given B(cid:48)(ψ) = B(cid:48)(χ)
and B(cid:48)(ϕ) + B(cid:48)(ψ) + B(cid:48)(χ) = 3m, we have
We claim that
B(cid:48)(ϕ) + 2B(cid:48)(ψ) = 3m.
B(cid:48)(ψ)− B(cid:48)(ϕ) > 2.
(3)
(4)
Holliday and Pacuit
271
For B(cid:48)(ψ)− B(cid:48)(ϕ) (cid:54)= 0 since ϕ was chosen as the candidate with the lowest Borda score in P(cid:48), and if
B(cid:48)(ψ) − B(cid:48)(ϕ) = k for k ∈ {1,2}, so B(cid:48)(ϕ) = B(cid:48)(ψ) − k, then from (3) we have 3B(cid:48)(ψ) − k = 3m, a
contradiction. Now by Claim 2, we have B(ϕ)− B(cid:48)(ϕ) ≤ 1, B(ψ)− B(cid:48)(ψ) ≤ 1, and B(χ)− B(cid:48)(χ) ≤
1. It follows by (4) that B(ϕ) < B(ψ) and B(ϕ) < B(χ). Thus, ϕ has a below average Borda score in P.
Finally, we claim that ϕ is the only candidate with a below average Borda score in P. Since the average
Borda score is m, this means B(ψ) ≥ m and B(χ) ≥ m. Suppose for contradiction that B(ψ) < m or
B(χ) < m. Without loss of generality, suppose B(ψ) < m. By Claim 2 and the fact that B(cid:48)(ψ) = B(cid:48)(χ),
we have B(ψ)− B(χ) ≤ 2. But together B(ϕ) < B(ψ), B(ψ) < m, and B(ψ)− B(χ) ≤ 2 contradict
the fact that B(ϕ) + B(ψ) + B(χ) = 3m. Thus, ϕ has the unique below average Borda score in P.
Case 2: B(cid:48)(ψ) < B(cid:48)(χ). First, we claim it is not the case that B(ϕ) = B(ψ) = B(χ). It follows from
Claim 2 that there are only two ways to go from B(ϕ) = B(ψ) = B(χ) to B(cid:48)(ϕ) < B(cid:48)(ψ) < B(cid:48)(χ): i's
ranking goes from ϕχψ to χϕψ or from ψϕχ to ψχϕ. But in both cases the new Borda winning set
{χ} does not weakly dominate the old winning Borda set {ϕ,ψ, χ}, contradicting the assumption that
i had an incentive to manipulate. In addition, it is not the case that B(ϕ) = B(ψ) < B(χ) or B(ϕ) <
B(ψ) < B(χ), for then i would have no incentive to transition to B(cid:48)(ϕ) < B(cid:48)(ψ) < B(cid:48)(χ), since the
Borda winning set would not change. Thus, we have that B(ϕ) < B(ψ) = B(χ), so ϕ has the unique
below average Borda score in P. Now by Claim 2, the Borda score of one of ϕ, ψ, and χ must remain
the same from P to P(cid:48). But we cannot have B(ϕ) = B(cid:48)(ϕ), for then in order to go from B(ψ) = B(χ) to
B(cid:48)(ψ) < B(cid:48)(χ), by Claim 2 i's ranking must either change from ψχϕ to χψϕ or from ϕψχ to ϕχψ;
but in both cases the new Borda winning set {χ} does not weakly dominate the old winning Borda set
{ψ, χ}, contradicting the assumption that i had an incentive to manipulate. Thus, either B(ψ) = B(cid:48)(ψ)
or B(χ) = B(cid:48)(χ). But if B(ψ) = B(cid:48)(ψ) or B(χ) = B(cid:48)(χ), then together B(ψ) = B(χ) and B(cid:48)(ψ) < B(cid:48)(χ)
imply B(cid:48)(ψ)− B(cid:48)(χ) = 1 by Claim 2, in which case B(cid:48)(ϕ) < B(cid:48)(ψ) < B(cid:48)(χ) implies that ϕ has the
unique below average Borda score in P(cid:48).
Claim 4: the majority ordering between ψ and χ does not change from P to P(cid:48). This follows from
the claim that i's ordering of ψ and χ does not change from P to P(cid:48). We know ϕ is not in the set of
Borda winners in P or in P(cid:48) by Claim 3, and the set of Borda winners in P(cid:48) is not {ϕ,ψ, χ} by Claim 1.
Case 1: the Borda winning set in P(cid:48) is {ψ, χ}, so B(cid:48)(ψ) = B(cid:48)(χ). Suppose for contradiction that i's
ranking for ψ vs. χ in P is the reverse of i's ranking in P(cid:48). Suppose, without loss of generality, that i
ranks ψ over χ in P but i ranks χ over ψ in P(cid:48). Then since in P(cid:48), we have B(cid:48)(ψ) = B(cid:48)(χ), it follows that
B(ψ) > B(χ), which with Claim 3 implies that ψ is the unique Borda winner in P. But then there is no
incentive for i to manipulate, as i would do worse with the winning set {ψ, χ} in P(cid:48). This contradicts our
assumption that (P,i) witnesses manipulation for Borda by transitioning to P(cid:48).
Case 2: the Borda winning set in P(cid:48) is {χ}, so B(cid:48)(ψ) < B(cid:48)(χ). Since ϕ is not in the Borda winning
set in P by Claim 3, the Borda winning set in P is {ψ, χ} or {ψ}, but the latter is ruled out since Borda
is not single-winner manipulable. Since by assumption the winning set in P(cid:48) weakly dominates that in P
for i, it follows that i ranks χ above ψ in P. But then the transition from the Borda winning set {ψ, χ}
to {χ} cannot be the result of i switching her ranking of χ and ψ so that ψ is ranked above χ in P(cid:48).
For any (3,m)-profile, the set of Baldwin (resp. StrictNanson) winners may be obtained by first
eliminating the candidates (if there are any) with the strictly lowest Borda scores (resp. with below aver-
age Borda scores) and then selecting as winners from the remaining candidates those who are maximal in
the majority ordering. Thus, by Claim 3 and Claim 4, the sets of Baldwin winners and StrictNanson
winners do not change from P to P(cid:48). This contradicts the assumption that (P,i) witnesses sure weak
dominance manipulation for {Borda, Baldwin} or {Borda, StrictNanson} by transitioning to P(cid:48).
272
Strategic Voting Under Uncertainty About the Voting Method
B Proof of Theorem 3.7
In this appendix, we prove Theorem 3.7: for any m ≥ 4, {WeakNanson, Baldwin} and {WeakNanson,
StrictNanson} eliminate sure weak dominance manipulation for (3,m).
Proof. Given Lemma 3.5, we need only show that the two sets of methods are not susceptible to sure
weak dominance manipulation for (3,m). Consider an arbitrary pointed profile (P,i). We show that
for any profile P(cid:48) differing from P only in i's ranking, (P,i) does not witness sure weak dominance
manipulation for {WeakNanson, Baldwin} or {WeakNanson, StrictNanson} by transitioning to P(cid:48).
Let i's ranking in P be αβγ. Let B(α), B(β ), and B(γ) be the Borda scores of α, β , and γ, respectively,
in P, and likewise for B(cid:48)(α), B(cid:48)(β ), and B(cid:48)(γ) in P(cid:48). For any (3,m)-profile, the set of WeakNanson
(resp. StrictNanson, Baldwin) winners may be obtained by first eliminating the candidates (if there
are any) with less than or equal to average Borda scores (resp. with below average Borda scores, with
strictly lowest Borda scores) and then selecting as winners from the remaining candidates those who are
maximal in the majority ordering. Note that the average Borda score for any given (3,m)-profile is m.
We consider the following exhaustive list of cases, where 'WN' stands for WeakNanson:
1. B(β ),B(γ) ≤ m < B(α); or B(β ) ≤ m < B(α),B(γ) and α >M
P γ; or B(γ) ≤ m < B(α),B(β ) and
P β . Then {α} is the WN-winning set in P, so i has no incentive to transition from P to P(cid:48)
α >M
2. B(α) ≤ m < B(β ),B(γ) and β >M
P γ, so {β} is the WN-winning set in P. Then B(cid:48)(α) ≤ m < B(cid:48)(γ),
so α is not in the WN-winning set in P(cid:48). But only sets that contain α weakly dominate {β} for i.
3. B(α) ≤ m < B(β ),B(γ) and γ >M
P β , so {γ} is the WN-winning set in P. In this case B(cid:48)(α) ≤ m ≤
B(cid:48)(β ), m < B(cid:48)(γ), and γ >M
P(cid:48) β for any P(cid:48) differing only in i's ranking. So the winning set in P(cid:48) is again
{γ}, providing no incentive under WN to transition to P(cid:48).
P β , so {β ,γ} is the WN-winning set in P. In this case B(cid:48)(α) ≤
m ≤ B(cid:48)(β ), m < B(cid:48)(γ), and γ ≥M
P(cid:48) β for any P(cid:48) differ only in i's ranking. So the WN-winning set in P(cid:48) is
{γ} or {β ,γ}, neither of which weakly dominates {β ,γ} for i.
P α, so the WN-winning set in P is either {γ} or {α,γ}. It follows
that B(β ) ≤ m− 2 and hence B(cid:48)(β ) < m, and γ ≥M
P(cid:48) α for any P(cid:48) differ only in i's ranking. Thus, if {γ} is
the WN-winning set in P, it is the WN-winning set in P(cid:48); and if {α,γ} is the WN-winning set in P, then
the WN-winning set in P(cid:48) contains γ, but no set containing γ weakly dominates {α,γ}.
P α, so the WN-winning set is either {β} or {α,β}. It follows
that B(γ) < m and indeed B(γ) ≤ m− 2. Hence B(cid:48)(γ) ≤ m. Thus, γ is eliminated for WN in the first
round for P(cid:48). Since β ≥M
P(cid:48) α for any P(cid:48)
differing from P only in i's ranking. Thus, if {β} is the WN-wining set in P, it is still the WN-winning
set in P(cid:48); and if {α,β} is the WN-winning set in P, then the WN-winning set in P(cid:48) contains β , but no set
containing β weakly dominates {α,β}.
7. B(α),B(γ)≤ m≤ B(β ), so the WN-winning set in P is {β} or {α,β ,γ}. In this case B(cid:48)(α)≤ m, so
the WN-winning set in P(cid:48) is either {α,β ,γ} or a set not containing α, and in both cases the WN-winning
set in P(cid:48) does not weakly dominate the WN-winning set in P.
If
B(cid:48)(β ) ≤ m, then {γ} is still the WN-winning set in P(cid:48). So suppose m < B(cid:48)(β ), which implies B(β ) = m,
B(α) < m, and B(cid:48)(α) < m. There are only two post-manipulation rankings consistent with m < B(cid:48)(β ):
βαγ and βγα. Note that the majority ordering of β and γ has not changed from P to P(cid:48), and in both P
and P(cid:48), α is the unique candidate with below average Borda score. It follows that the StrictNanson
winning set in P and P(cid:48) is the same, and the Baldwin winning set in P and P(cid:48) is the same. Hence the
transition from P to P(cid:48) does not witness sure manipulation.
8. B(α),B(β ) ≤ m < B(γ), so {γ} is the WN-winning set in P. Hence B(cid:48)(α) ≤ m < B(cid:48)(γ).
P α with i having the ranking αβγ in P, it follows that β ≥M
4. B(α) ≤ m < B(β ),B(γ) and γ =M
5. B(β ) ≤ m < B(α),B(γ) and γ ≥M
6. B(γ) ≤ m < B(α),B(β ), and β ≥M
|
1905.05717 | 2 | 1905 | 2019-11-05T09:17:57 | CoLight: Learning Network-level Cooperation for Traffic Signal Control | [
"cs.MA",
"cs.GT",
"cs.LG"
] | Cooperation among the traffic signals enables vehicles to move through intersections more quickly. Conventional transportation approaches implement cooperation by pre-calculating the offsets between two intersections. Such pre-calculated offsets are not suitable for dynamic traffic environments. To enable cooperation of traffic signals, in this paper, we propose a model, CoLight, which uses graph attentional networks to facilitate communication. Specifically, for a target intersection in a network, CoLight can not only incorporate the temporal and spatial influences of neighboring intersections to the target intersection, but also build up index-free modeling of neighboring intersections. To the best of our knowledge, we are the first to use graph attentional networks in the setting of reinforcement learning for traffic signal control and to conduct experiments on the large-scale road network with hundreds of traffic signals. In experiments, we demonstrate that by learning the communication, the proposed model can achieve superior performance against the state-of-the-art methods. | cs.MA | cs |
CoLight: Learning Network-level Cooperation
for Traffic Signal Control
Hua Wei†∗, Nan Xu‡∗, Huichu Zhang‡, Guanjie Zheng†, Xinshi Zang‡, Chacha Chen‡,
Weinan Zhang‡, Yanmin Zhu‡, Kai Xu§, Zhenhui Li†
†Pennsylvania State University, ‡Shanghai Jiao Tong Univerisity, §Shanghai Tianrang Intelligent Technology Co., Ltd
†{hzw77, gjz5038, jessieli}@ist.psu.edu, ‡{xunannancy, chacha1997, wnzhang, yzhu}@sjtu.edu.cn,
‡{zhc,xszang}@apex.sjtu.edu.cn, §[email protected]
ABSTRACT
Cooperation among the traffic signals enables vehicles to move
through intersections more quickly. Conventional transportation
approaches implement cooperation by pre-calculating the offsets
between two intersections. Such pre-calculated offsets are not suit-
able for dynamic traffic environments.
To enable cooperation of traffic signals, in this paper, we pro-
pose a model, CoLight, which uses graph attentional networks to
facilitate communication. Specifically, for a target intersection in a
network, CoLight can not only incorporate the temporal and spatial
influences of neighboring intersections to the target intersection,
but also build up index-free modeling of neighboring intersections.
To the best of our knowledge, we are the first to use graph atten-
tional networks in the setting of reinforcement learning for traffic
signal control and to conduct experiments on the large-scale road
network with hundreds of traffic signals. In experiments, we demon-
strate that by learning the communication, the proposed model can
achieve superior performance against the state-of-the-art methods.
CCS CONCEPTS
• Computing methodologies → Artificial intelligence; Con-
trol methods; • Applied computing → Transportation.
KEYWORDS
Deep reinforcement learning, traffic signal control, multi-agent
system
ACM Reference Format:
Hua Wei, Nan Xu, Huichu Zhang, Guanjie Zheng, Xinshi Zang, Chacha
Chen, and Weinan Zhang, Yanmin Zhu, Kai Xu, Zhenhui Li. 2019. CoLight:
Learning Network-level Cooperation for Traffic Signal Control. In The 28th
ACM International Conference on Information and Knowledge Management
*Equal contribution.
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 profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
CIKM '19, November 3 -- 7, 2019, Beijing, China
© 2019 Association for Computing Machinery.
ACM ISBN 978-1-4503-6976-3/19/11...$15.00
https://doi.org/10.1145/3357384.3357902
(CIKM '19), November 3 -- 7, 2019, Beijing, China. ACM, New York, NY, USA,
10 pages. https://doi.org/10.1145/3357384.3357902.
1 INTRODUCTION
A key question often asked of traffic signal control is "How do traf-
fic signals cooperate between intersections?" Cooperation among
intersections is especially important for an urban road network
because the actions of signals could affect each other, especially
when the intersections are closely connected. Good cooperation
among the traffic signals enables the vehicles to move through
intersections more quickly.
In the transportation field, a typical way to solve the cooperation
problem between intersections is to formulate it as an optimization
problem and solve it under certain assumptions (e.g., uniform ar-
rival rate [21, 27] and unlimited lane capacity [24]). Such methods,
however, do not perform well because the assumptions do not hold
in the real world.
Recently, researchers start to investigate reinforcement learning
(RL) techniques for the cooperation of traffic signals. The most
common way is to have an RL agent control each intersection and
communication is achieved by sharing information among agents [1,
8, 9, 22]. At each time step, the agent observes the traffic condition
of the target intersection and its neighboring intersections, and
decides an action a to take. After the action is taken, a reward
r (often defined as a measure correlated with travel time) is fed
back to the agent indicating how good the action a is. Different
from conventional approaches, such RL methods avoid making
strong assumptions and directly learn good strategies from trials
and errors.
Existing RL-based methods still fail to communicate with neigh-
bors in the most efficient way. We propose CoLight that improves
communication of agents and is scalable to hundreds of intersec-
tions. In particular, our work makes the following key contributions:
• Cooperation through dynamic communication. Most methods
to achieve cooporation are through expanding the observation of
the target agent for more comprehensive information. Existing
studies [1, 5, 8, 9, 22] tend to select the traffic conditions from
adjacent intersections and directly concatenate them with the traffic
condition of the target intersection, neglecting the fact that the
traffic is changing both temporally and spatially. For example, if
intersections A and B are adjacent intersections on a major road,
but intersection C is on a side road linked with B. The information
justify their cooperation strategies on small road networks with
only fewer than 20 intersections [29]. In this paper, we conduct
comprehensive experiments on both synthetic and real-world data,
including a large-scale road network with 196 intersections derived
from Manhattan, New York. The experiments demonstrate that the
proposed model benefits from the dynamic communication and the
index-free modeling mechanism and significantly outperforms the
state-of-the-art methods.
2 RELATED WORK
Conventional coordinated methods [15] and systems [12, 13, 17] in
transportation usually coordinate traffic signals through modifying
the offset (i.e., the time interval between the beginnings of green
lights) between consecutive intersections and require the intersec-
tions to have the same cycle length. But this type of methods can
only optimize the traffic flow for certain pre-defined directions [10].
Actually, it is not an easy task to coordinate the offsets for traffic sig-
nals in the network. For network-level control, Max-pressure [24]
is a state-of-the-art signal control method which greedily takes
actions that maximizes the throughput of the network, under the
assumption that the downstream lanes have unlimited capacity.
Other traffic control methods like TUC [7] also use optimization
techniques to minimize vehicle travel time and/or the number of
stops at multiple intersections under certain assumptions, such as
the traffic flow is unform in a certain time period. However, such
assumptions often do not hold in the network setting and therefore
prevent these methods from being widely applied.
Recently, reinforcement learning techniques have been proposed
to coordinate traffic signals for their capability of online optimiza-
tion without prior knowledge about the given environment. One
way to achieve coordination is through centralized optimization
over the joint actions of multiple intersections. [20] directly trains
one central agent to decide the actions for all intersections but it
cannot learn well due to the curse of dimension in joint action
space. [16, 23] propose to jointly model two adjacent intersections
and then using centralized coordination over the global joint ac-
tions, but they require the maximization over a combinatorially
large joint action space and face scalability issues during deploy-
ment. Therefore, these centralized methods are hard to apply on
the large-scale road network.
To mitigate this issue, independently modelling RL methods [3,
28, 31] are proposed in which they train a bunch of RL agents
separately, one for each intersection. To avoid the non-stationary
impacts from neighboring agents, communication among agents [2,
19] are proposed in order to achieve coordination using neigh-
boring information: [5, 8, 9, 22, 31] add downstream information
into states, [1, 31] add all neighboring states, and [18] adds neigh-
bors' hidden states. However, in these methods, the information
from different neighbors is concatenated all together and treated
with equal importance, which leads to two major issues: 1). the
impacts from neighboring intersections are changing dynamically
with the traffic flow and should not be treated evenly. Even when
the traffic flow is static, Kinenmatic-wave theory [11] from the
transportation area shows that the upstream intersections could
have larger influence than downstream intersections; 2). simple
concatenation of neighboring information requires all agents to
have an extra indexing mechanism, which is usually unrealistic and
Figure 1: Illustration of index-based concatenation. Thick
yellow lines are the arterials and grey thin lines are the side
streets. With index-based concatenation, A and B's obser-
vation will be aligned as model inputs with an fixed order.
These two inputs will confuse the model shared by A and B.
from A is more useful to B compared with that from C to B. Besides,
the influences from intersection A could change temporally. For
example, during the morning rush hour, there are a large number
of vehicles moving from A to B, and the direction of vehicles is
reversed at night peak hours. Therefore, the effect of A on B is also
changing at different times of the day. In this paper, we propose
to use the graph attentional network [26] to learn the dynamics of
the influences, which shows superior performance against methods
without this mechanism in Section 5.5. To the best of our knowledge,
we are the first to use graph attentional network in the setting of
traffic signal control.
• Index-free model learning with parameter sharing. Agents often
share the same model parameters to enable efficient learning [1, 9].
However, if we simply take neighboring information as part of the
state for the communication purpose, the fixed indexing of neigh-
bors may cause conflicting learning problems. Take Figure 1 as an
example. Figure 1(a) shows two target intersections A and B, where
their neighbors A-W, A-E, B-N and B-S are on the major roads,
while A-N, A-S, B-W and B-E are on the side streets. Intuitively,
the two intersections on the major roads should relate more to the
target intersection than those on the side streets, e.g., the influence
from A-W to A should be greater than that from A-N to A. When
neighbors are indexed in the way as [E, W, N, S], agent A would
care more about the first and second neighbors (i.e., East and West)
and agent B would pay more attention to the third and fourth neigh-
bors (i.e., North and South). However, when two agents share the
model parameters, this will cause conflicts to learn the influences
of neighbors to the target intersection. In this paper, similar to the
"mean-field" idea [32], we tackle this problem by averaging over
the influences of all neighboring intersections with the learned
attention weights, instead of using a fixed indexing for neighbors.
This weighted average mechanism provides index-free modeling
of the influences of neighboring intersections, and along with the
parameter sharing strategy, the overall parameters of the learning
model can be largely reduced. We will show the parameter analysis
later in Section 4.4.
• Experiment on the large-scale road network. To the best of our
knowledge, none of the existing studies that use RL to cooperate
traffic signals have evaluated their methods in large-scale road
networks with hundreds of traffic signals. Instead, most of them
AA-NA-WA-SA-EBB-NB-WB-SB-E (b) Model input alignmentEastNorthWestSouthSelfA-SA-NA-WA-EAB-SB-NB-WB-EB(c) Actual influence to target intersectionEastNorthWestSouthA's NeighborB's Neighbor(a) Two target intersections in a real-world road networkN requires heuristic design. To address the shortcomings of previous
methods, our proposed method leverages the attention mechanism
to learn and specify different weights to different neighboring inter-
sections and directly models the overall influence of neighboring
intersections on the target intersection.
It is worth noting that most of joint modelling methods for the
traffic signal control problem only conduct experiments on simple
road networks with at most 20 intersections. To the best of our
knowledge, none of the individual modelling methods conduct
experiments on more than 70 signals [29]. In this paper, we conduct
experiments on the simulator under different scales, including a
real-world road network with about 200 intersections.
3 PROBLEM DEFINITION
In this section, we present the problem of traffic signal control as a
Markov Game. Each intersection in the system is controlled by an
agent. Given each agent observes part of the total system condition,
we would like to proactively decide for all the intersections in the
system which phases should they change to so as to minimize
the average queue length on the lanes around the intersections.
Specifically, the problem is characterized by the following major
components ⟨S, O, A, P, r, π , γ⟩:
• System state space S and observation space O. We assume
that there are N intersections in the system and each agent can
observe part of the system state s ∈ S as its observation o ∈ O. In
this work, we define ot
i for agent i at time t, which consists of its
current phase (which direction is in green light) represented by a
one-hot vector, and the number of vehicles on each lane connected
with the intersection.
• Set of actions A. In the traffic signal control problem, at time
t, an agent i would choose an action at
i from its candidate action
set Ai as a decision for the next ∆t period of time. Here, each inter-
section would choose a phase p as its action at
i from its pre-defined
phase set, indicating that from time t to t + ∆t, this intersection
would be in phase p.
• Transition probability P. Given the system state st and cor-
responding joint actions at of agents at time t, the system arrives
at the next state st +1 according to the state transition probability
P(st +1st , at) : S × A1 × · · · × AN → Ω(S), where Ω(S) denotes
the space of state distributions.
i from the
environment at time t by a reward function S×A1×· · ·×AN → R.
In this paper, we want to minimize the travel time for all vehicles
in the system, which is hard to optimize directly. Therefore, we
= −Σlut
define the reward for intersection i as rt
i,l is
i
the queue length on the approaching lane l at time t.
• Policy π and discount factor γ. Intuitively, the joint actions
have long-term effects on the system, so that we want to minimize
the expected queue length of each intersection in each episode.
Specifically, at time t, each agent chooses an action following a
certain policy O × A → π, aiming to maximize its total reward
Gt
= ΣT
i , where T is total time steps of an episode and
i
γ ∈ [0, 1] differentiates the rewards in terms of temporal proximity.
In this paper, we use the action-value function Qi(θn) for each
agent i at the n-th iteration (parameterized by θ) to approximate
• Reward r. Each agent i obtains an immediate reward rt
t =τ γ t−τ rt
i,l where ut
a′ Q(ot
i
′
, at
i
the total reward Gt
L(θn) = E[(rt
i with neural networks by minimizing the loss:
(1)
′; θn−1) − Q(ot
i ; θn))2]
i + γ max
i , at
′ denotes the next observation for ot
where ot
i . These earlier snap-
i
shots of parameters are periodically updated with the most recent
network weights and help increase the learning stability by decor-
relating predicted and target q-values.
4 METHOD
In this section, we will first introduce the proposed cooperated RL
network structure, as Figure 2 illustrates, from bottom to top layer:
the first observation embedding layer, the interior neighborhood
cooperation layers (shorted as GAT layers) and the final q-value
prediction layer. Then we will discuss its time and space complex-
ity compared with other methods of signal control for multiple
intersections.
4.1 Observation Embedding
Given the raw data of the local observation, i.e., the number of
vehicles on each lane and the phase the signal currently in, we first
embed such k-dimensional data into an m-dimensional latent space
via a layer of Multi-Layer Perceptron:
hi = Embed(ot
i ) = σ(oiWe + be),
(2)
i ∈ Rk is intersection i's observation at time t and k is
where ot
i , We ∈ Rk×m and be ∈ Rm are weight
the feature dimension of ot
matrix and bias vector to learn, σ is ReLU function (same denotation
for the following σ). The generated hidden state hi ∈ Rm represents
the current traffic condition of the i-th intersection.
4.2 Graph Attention Networks for Cooperation
Communication between agents is necessary for cooperation in
multi-agent reinforcement learning (MARL) environment, since
the evaluation of the conducted policy for each agent depends
not only on the observable surrounding, but also on the policies
of other agents [6, 19]. CoLight agent learns to communicate on
the representations of neighboring intersections by leveraging the
attention mechanism, a widely-used technique to boost model ac-
curacy [4, 14, 33, 34]. Then an overall summary of the agent's
neighborhood is generated, upon which the agent learns to model
the influence of neighborhoods.
4.2.1 Observation Interaction. To learn the importance of informa-
tion from intersection j (source intersection) in determining the
policy for intersection i (target intersection), we first embed the
representation of these two intersections from previous layer and
calculate ei, j, the importance of j in determining the policy for i,
with the following operation:
eij = (hiWt) · (hjWs)T ,
(3)
where Ws ,Wt ∈ Rm×n are embedding parameters for the source
and target intersection respectively.
Note that eij is not necessarily equal to eji. Take the scenario
in Figure 2(a) as an example, where the 9-th Avenue is a one-way
arterial. On one hand, since the traffic flow goes from Inter 9-50 to
Inter 9-49, the traffic condition from Inter 9-50 is important for Inter
Figure 2: Left: Framework of the proposed CoLight model. Right: variation of cooperation scope (light blue shadow, from
one-hop to two-hop) and attention distribution (colored points, the redder, the more important) of the target intersection.
(cid:1),
αij(hjWc) + bq
(5)
9-49 to decide for the future actions, thus, e9-49,9-50 should be quite
large. On the other hand, as the traffic condition of the downstream
Inter 9-49 is less helpful to Inter 9-50, e9-50,9-49 should be relatively
small.
4.2.2 Attention Distribution within Neighborhood Scope. To re-
trieve a general attention value between source and target inter-
sections, we further normalize the interaction scores between the
target intersection i and its neighborhood intersections:
αij = softmax(eij) =
exp(eij/τ)
j∈Ni exp(eij/τ) ,
(4)
where τ is the temperature factor and Ni is the set of intersections
in the target intersection's neighborhood scope. The neighborhood
of the target contains the top Ni closest intersections, and the
distance can be defined in multiple ways. For example, we can
construct the neighborhood scope for target intersection i through:
• Road distance: the geo-distance between two intersections'
• Node distance: the smallest hop count between two nodes over
Note that intersection i itself is also included in Ni to help the
agent get aware of how much attention should be put on its own
traffic condition.
the network, with each node as an intersection.
geo-locations.
The general attention score αij is beneficial not only for it applies
to all kinds of road network structures (intersections with different
numbers of arms), but also for it relaxes the concept of "neighbor-
hood". Without losing generality, the target can even take some
other intersections into Ni although they are not adjacent to them.
For instance, one four-way intersection can determine its signal
policy based on information from five nearby intersections, four of
which are the adjacent neighbors while the other is disconnected
but geographically close to the target intersection.
Index-free Neighborhood Cooperation. To model the overall
4.2.3
influence of neighborhoods to the target intersection, the repre-
sentation of several source intersections are combined with their
respective importance:
hsi = σ(cid:0)Wq ·
j∈Ni
where Wc ∈ Rm×c is weight parameters for source intersection
embedding, Wq and bq are trainable variables. The weighted sum of
neighborhood representation hsi ∈ Rc accumulates the key infor-
mation from the surrounding environment for performing efficient
signal policy. By summing over the neighborhood representation,
the model is index-free that does not require all agents to align the
index of their neighboring intersections.
The graph-level attention allows the agent to adjust their focus
according to the dynamic traffic and to sense the environment in
a larger scale. In Figure 2(b) and (c), the emphasizes of Inter 9-49
on four neighbors are quite distinct due to the uni-directional traf-
fic flow, i.e., a higher attention score for Inter 9-50 (upstream, red
marked) than for Inter 9-48 (downstream, green marked). The agent
for Inter 9-49 acquires the knowledge of adjacent intersections (In-
ter 9-48, Inter 9-50, Inter 10-49 and Inter 8-49) directly from the first
layer of Graph Attention Networks (GAT). Since the hidden states
of adjacent neighbors from the first GAT layer carry their respective
neighborhood message, then in the second GAT layer, the coop-
eration scope of Inter 9-49 expands significantly (blue shadow in
Figure 2(c)) to 8 intersections. Such additional information helps the
target Inter 9-49 learn the traffic trend. As a result, Inter 9-49 relies
more on the upstream intersections and less on the downstream to
take actions, and the attention scores on Inter 9-50 and Inter 8-49
grow higher while those on Inter 10-49 and Inter 9-48 become lower.
More GAT layers helps the agent detect environment dynamics
more hops away.
4.2.4 Multi-head Attention. The cooperating information hsi for
the i-th intersection concludes one type of relationship with neigh-
boring intersections. To jointly attend to the neighborhood from
different representation subspaces at different positions, we extend
the previous single-head attention in the neural network to multi-
head attention as much recent work did [25, 26]. Specifically, the
attention function (procedures including Observation Interaction,
10-th Ave.9-th Ave.8-th Ave.W47-th St.W48-th St.W49-th St.W50-th St.W51-st St.9-499-509-4810-498-49NN⇥⇥⇥⇥(cid:1)(cid:1)ihmLi1Ni(cid:1)(cid:1)(cid:1)L1. Observation embedding2. Graph Attentional Network(GAT)GATGAT1(cid:1)3. Q-value predictionObservationsActionPhase 1Phase p(cid:1)↵i,1...↵i,N 1↵i,2↵i,iH+ hihm1i1Ni(cid:1)(cid:1)......qiai(cid:1)(a) Model framework(b) 1st GAT layer (one-hop view)(c) 2nd GAT layer (two-hop view) Attention Distribution and Neighborhood Cooperation) with differ-
ent linear projections (multiple sets of trainable parameters {Wt ,
Ws, Wc}) is performed in parallel and the different versions of neigh-
borhood condition summarization hsi are averaged as hmi:
ij = softmax(eh
αh
ij = (hiW h
t ) · (hjW h
s )T
eh
exp(eh
ij/τ)
ij) =
j∈Ni exp(eh
ij(hjW h
αh
(cid:16)
Wq ·(cid:0) 1
h=H
ij/τ)
c )(cid:1) + bq
(cid:17)
(6)
(7)
(8)
hmi = σ
H
h=1
j∈Ni
where H is the number of attention heads. Besides averaging oper-
ation, concatenating the product of multi-head attention is another
feasible way to conclude multiple types of the neighborhood coop-
eration.
In this work, we investigate the effects of multi-head attention
on performance of the proposed model and find that 5 attention
heads achieve the best performance.
4.3 Q-value Prediction
As illustrated in Figure 2(a), each hidden layer of model CoLight
learns the neighborhood representation through methods intro-
duced in Section 4.2. We denote such layerwise cooperation proce-
dure by GAT, then the forward propagation of input data in CoLight
can be formatted as follows:
hm1
hi =
i =
· · ·
hmL
i =
i ) =
(cid:101)q(ot
Embed(ot
i ),
GAT 1(hi),
,
GAT L(hmL−1
),
hmL
i Wp + bp ,
i
(9)
where Wp ∈ Rc×p and bp ∈ Rp are parameters to learn, p is the
number of phases (action space), L is the number of GAT layers,(cid:101)q
is the predicted q-value.
According to Eq. (1), the loss function for our CoLight to optimize
the current policy is:
L(θ) = 1
t =T
i =N
(cid:0)q(ot
i ) −(cid:101)q(ot
i ; at
i , θ)(cid:1)2
i , at
T
t =1
i =1
,
(10)
where T is the total number of time steps that contribute to the
network update, N is the number of intersections in the whole road
network, θ represents all the trainable variables in CoLight.
As the importance of neighborhood to the target intersection
varies spatially and temporally, the proposed attention mechanism
is able to help the target agent distinguish among the complex
scenarios by considering the traffic condition of any source-target
intersection pair.
4.4 Complexity Analysis
Although CoLight spares additional parameters to learn the dy-
namic cooperation from neighborhood, owing to the index-free
parameter sharing mechanism, both the time and space it demands
are approximately equal to O(m2L), which is irrelevant to the num-
ber of intersections. Hence CoLight is scalable even if the road
network contains hundreds of or even thousands of intersections.
the number of layers (L), the phase space (p) and comparable to the
input dimension (k). Therefore, the space complexity of CoLight is
approximately equal to O(m2L).
Space Complexity. If there are L hidden layers and each layer
4.4.1
has m neurons, then the size of the weight matrices and bias vectors
in each component of CoLight is: 1) Observation Embedding layer:
m(4m + 1)L; 3) Q-value Prediction layer: mp + p. Hence the total
km + m; 2) interior Graph Attentional layers:(cid:0)3m2 + (m2 + m)(cid:1)L =
number of learnable parameters to store is O(cid:0)m(4mL + L + k + 1 +
p) +p(cid:1). Normally, the size of the hidden layer (m) is far greater than
(cid:17) ≈ O(m2L · N), which
(cid:16)(cid:0)(km + m) + (m2 + m)L + (mp + p)(cid:1) · N
If we leverage N separate RL models (without parameter sharing)
to control signals in N intersections, then the space complexity is
O
is unfeasible when N is extremely large for city-level traffic signal
control. To scale up, the simplest solution is to allow all the intersec-
tions to share parameters and maintain one model, in this case, the
space complexity is O(m2L), which is identical to that of CoLight.
4.4.2 Time Complexity. We assume that: 1) all the agents lever-
age CoLight to predict q-values for the corresponding intersections
concurrently; 2) the multiple heads of attention are independently
computed so that they are as fast as the single-head attention; 3)
the embeddings for either source or target intersection condition
via Ws, Wc and Wt are separate processes that can also be executed
at the same time, 4) for one target intersection, the interaction
with all the neighbors is computed simultaneously, then the time
complexity (only multiplication operations considered since the
addition procedures are relatively insignificant) in each compo-
nent of CoLight is: 1) Observation Embedding layer: km; 2) interior
Graph Attentional layers: (m2 + m2)L; 3) Q-value Prediction layer:
it is approximately equal to O(m2L).
mp. Hence the time complexity is O(cid:0)m(k + 2mL + p)(cid:1), and similarly,
signal control in multiple intersections requires O(cid:0)m(k +mL +p)(cid:1) ≈
Either the individual RL models or the shared single RL model for
O(m2L) computation, approaching that of CoLight.
5 EXPERIMENTS
We perform experiments on two synthetic datasets and three real-
world datasets to evaluate our proposed method, especially the
graph-level attention for neighborhood cooperation.
5.1 Settings
We conduct experiments on CityFlow1, an open-source traffic sim-
ulator that supports large-scale traffic signal control [35]. After the
traffic data being fed into the simulator, a vehicle moves towards its
destination according to the setting of the environment. The simu-
lator provides the state to the signal control method and executes
the traffic signal actions from the control method. Following the
tradition, each green signal is followed by a three-second yellow
signal and two-second all red time.2
In a traffic dataset, each vehicle is described as (o, t, d), where
o is the origin location, t is time, and d is the destination location.
Locations o and d are both locations on the road network.
1http://cityflow-project.github.io
2CoLight's codes, parameter settings, public datasets can be found at:
https://github.com/wingsweihua/colight. More datasets can be found at: http://
traffic-signal-control.github.io
(a) Upper East Side, Manhattan,
New York, USA
(b) Gudang Sub-district,
Hangzhou, China
(c) Dongfeng Sub-district,
Jinan, China
Figure 3: Road networks for real-world datasets. Red polygons are the areas we select to model, blue dots are the traffic signals
we control. Left: 196 intersections with uni-directional traffic, middle: 16 intersections with uni- & bi-directional traffic, right:
12 intersections with bi-directional traffic.
5.2 Datasets
Synthetic Data. In the experiment, we use two kinds of syn-
5.2.1
thetic data, i.e., uni- and bi-directional traffic, on the following
different road networks:
• Arterial1×3: A 1 × 3 arterial to show the spatial attention distri-
bution learned by CoLight.
• Grid3×3: A 3 × 3 grid network to show convergence speed of
different RL methods and the temporal attention distribution.
• Grid6×6: A 6 × 6 grid network to evaluate effectiveness and effi-
ciency of different methods.
Each intersection in the synthetic road network has four direc-
tions (West→East, East→West, South→North, North→South), and
3 lanes (300 meters in length and 3 meters in width) for each direc-
tion. In bi-directional traffic, vehicles come uniformly with 300 ve-
hicles/lane/hour in West↔East direction and 90 vehicles/lane/hour
in South↔North direction. Only West→East and North→South
directional flows travel in uni-directional traffic.
5.2.2 Real-world Data. We also use the real-world traffic data from
three cities: New York, Hangzhou and Jinan. Their road networks
are imported from OpenStreetMap1, as shown in Figure 3. And
their traffic flows are processed from multiple sources, with data
statistics listed in Table 1. The detailed descriptions on how we
preprocess these datasets are as follows:
only contains the origin and destination geo-locations of each trip,
we first map these geo-locations to the intersections and find the
shortest path between them. Then we take the trips that fall within
the selected areas.
• DH anдzhou: There are 16 intersections in Gudang Sub-district
with traffic data generated from the roadside surveillance cameras.
Each record in the dataset consists of time, camera ID and the
information about vehicles. By analyzing these records with camera
locations, the trajectories of vehicles are recorded when they pass
through road intersections. We use the number of vehicles passing
through these intersections as traffic volume for experiments.
• D J inan: Similar to DH anдzhou, this traffic data is collected by
roadside cameras near 12 intersections in Dongfeng Sub-district,
Jinan, China.
Transportation Methods:
5.3 Compared Methods
We compare our model with the following two categories of meth-
ods: conventional transportation methods and RL methods. Note
that all the RL models are learned without any pre-trained parame-
ters for fair comparison.
• Fixedtime [15]: Fixed-time with random offsets. This method uses
a pre-determined plan for cycle length and phase time, which is
widely used when the traffic flow is steady.
• MaxPressure [24]: A state-of-the-art network-level traffic signal
control method in the transportation field, which greedily chooses
the phase that maximizes the pressure (a pre-defined metric about
upstream and downstream queue length).
• CGRL [23]: A RL-based method for multi-intersection signal con-
trol with joint-action modelling [23]. Specifically, the cooperation
is achieved by designing a coordination graph and it learns to opti-
mize the joint action between two intersections.
• Individual RL [30]: An individual deep RL approach which does
not consider neighbor information. Each intersection is controlled
by one agent, and the agents do not share parameters, but update
their own networks independently.
RL Methods:
Table 1: Data statistics of real-world traffic dataset
# intersections Arrival rate (vehicles/300s)
Std Max Min
216
10.08
86.70
256
208
38.21
Mean
240.79
526.63
250.70
274
676
335
196
16
12
Dataset
DN ewY ork
DH anдzhou
D J inan
• DN ewY ork: There are 196 intersections in Upper East Side of
Manhattan with open source taxi trip data2. Since the taxi data
1https://www.openstreetmap.org
2http://www.nyc.gov/html/tlc/html/about/trip_record_data.shtml
Table 2: Performance on synthetic data and real-world data w.r.t average travel time. CoLight is the best.
Model
Fixedtime [15]
MaxPressure [24]
CGRL [23]
Individual RL [30]
OneModel [5]
Neighbor RL [1]
GCN [18]
CoLight-node
CoLight
1950.27
1633.41
2187.12
-∗
Grid6×6-Uni Grid6×6-Bi DN ewY ork DH anдzhou D J inan
869.85
361.33
1210.70
325.56
728.63
1168.32
625.66
340.70
291.14
728.79
422.15
1582.26
345.00
394.56
1053.45
768.43
331.50
297.26
209.68
186.07
1532.75
314.82
181.81
240.68
205.40
178.42
173.79
209.68
194.96
2884.23
261.60
242.63
248.11
272.14
176.71
170.11
1973.11
2280.92
1876.37
1493.37
1459.28
∗No result as Individual RL can not scale up to 196 intersections in New York's road network.
• OneModel [5]: This method uses the same state and reward as
Individual RL in its agent design, which only considers the traffic
condition on the roads connecting with the controlled intersection.
Instead of maintaining their own parameters, all the agents share
the same policy network.
• Neighbor RL [1]: Based on OneModel, agents concatenate their
neighboring intersections' traffic condition with their own and all
the agents share the same parameters. Hence its feature space for
observation is larger than OneModel.
• GCN [18]: A RL based traffic signal control method that uses a
graph convolutional neural network to automatically extract the
traffic features of adjacent intersections. This method treats each
neighboring traffic condition without difference.
• CoLight: The neighborhood scope of an intersection is constructed
through geo-distance.
• CoLight-node: The neighborhood scope is constructed through
node distance, i.e., the smallest hop count between two intersections
in the road network.
Variants of Our Proposed Method:
5.4 Evaluation Metric
Following existing studies [28, 30], we use the average travel time
to evaluate the performance of different models for traffic signal
control. It calculates the average travel time of all the vehicles spend
between entering and leaving the area (in seconds), which is the
most frequently used measure of performance to control traffic
signal in the transportation field [21, 29].
5.5 Performance Comparison
In this section, we investigate on the performance of CoLight w.r.t.
the travel time and compare it with state-of-the-art transportation
and RL methods.
5.5.1 Overall Analysis. Table 2 lists the performance of two types
of the proposed CoLight, classic transportation models as well as
state-of-the-art learning methods in both synthetic and real-world
datasets. We have the following observations:
1. CoLight achieves consistent performance improvements over
state-of-the-art transportation (MaxPressure) and RL (Individual
RL) methods across diverse road networks and traffic patterns: the
average improvement is 6.98% for synthetic data and 11.69% for
real-world data. The performance improvements are attributed to
the benefits from dynamic communication along with the index-
free modeling. The advantage of our model is especially evident
when controlling signals in real-world cities, where road structures
are more irregular and traffic flows are more dynamic. Specifically,
Individual RL can hardly achieve satisfactory results because it in-
dependently optimizes the single intersection's policy; Neighbor RL
and GCN do not work well for either DN ewY ork or DH anдzhou,
as the agent treats the information from the upstream and down-
stream intersections with static importance according to the prior
geographic knowledge rather than real-time traffic flows.
2. The performance gap between the proposed CoLight and the
conventional transportation method MaxPressure becomes larger
as the evaluated data change from synthetic regular traffic (aver-
age gap 8.08%) to real-world dynamic traffic (average gap 19.89%).
Such growing performance divergence conforms to the deficiency
inherent in MaxPressure, that it is incapable of learning from the
feedback of the environment.
3. Our method outperforms the joint-action modelling method
CGRL. In order to achieve cooperation, CGRL first builds up one
model to decide the joint actions of two adjacent intersections and
then uses centralized coordination over the global joint actions. It
requires the centralized maximization over a combinatorially large
joint action space and faces scalability issues. On the constrast,
our method achieves cooperation through communication between
decentralized agents, which has a smaller action space and shows
superior performances.
5.5.2 Convergence Comparison. In Figure 4, we compare CoLight's
performance (average travel time for vehicles evaluated at each
episode) to the corresponding learning curves for the other five RL
methods. Evaluated in all the listed datasets, the performance of
CoLight is better than any of the baselines by a large margin, both in
jumpstart performance (initial performance after the first episode),
time to threshold (learning time to achieve a pre-specified perfor-
mance level), as well as in asymptotic performance (final learned
performance). Learning the attention on neighborhood does not
slow down model convergence, but accelerates the speed of ap-
proaching the optimal policy instead.
From Figure 4(a), we discover that model Individual RL starts
with extremely huge travel time and approaches to the optimal per-
formance after a long training time. Such disparity of convergence
speed shown in Figure 4 agrees with our previous space complexity
analysis (in Section 4.4.1), that agents with shared models (CGRL,
(a) Grid3×3
(b) Grid6×6-Uni
(c) Grid6×6-Bi
(d) DN ewY or k
(e) DH anдzhou
(f) D J inan
Figure 4: Convergence speed of CoLight (red continuous curves) and other 5 RL baselines (dashed curves) during training.
CoLight starts with the best performance (Jumpstart), reaches to the pre-defined performance the fastest (Time to Threshold),
and ends with the optimal policy (Aysmptotic). Curves are smoothed with a moving average of 5 points.
Neighbor RL, OneModel, GCN and CoLight) need to learn O(m2L)
parameters while individual agents (Individual RL) have to update
O(m2L · N) parameters.
5.6 Scalability Comparison
In this section, we investigate on whether CoLight is more scalable
than other RL-based methods in the following aspects:
5.6.1 Effectiveness. As is shown in Table 2 and the convergence
curve in Figure 4, CoLight performs consistently better than other
RL methods on networks with different scales, ranging from 9-
intersection grid network to 196-intersection real-world network.
5.6.2 Training time. We compare CoLight's training time (total
clock time for 100 episodes) to the corresponding running time for
the other five RL methods on road networks with different scales.
All the methods are evaluated individually on the server for fair
comparison. As is shown in Figure 5, the training time for CoLight
is comparable to that of OneModel and GCN, which is far more
efficient than that of CGRL, Individual RL and Neighbor RL. This is
consistent with the time complexity analysis (in Section 4.4.2), as
most of the parallel computation assumptions are satisfied in our
experiments.
Note that the average travel time for Individual RL (in Table 2)
is missing and the bar of training time (in Figure 5) is estimated on
DN ewY ork setting. This is because Individual RL is non-scalable
because all the separate 196 agents cannot be trained and updated
simultaneously due to processor and memory limitation.
5.7 Study of CoLight
In this section, we investigate on how different components (i.e.,
neighborhood definition, number of neighbors, and number of
attention heads) affect CoLight.
Impact of Neighborhood Definition. As mentioned in Sec-
5.7.1
tion 4.2.2, the neighborhood scope of an intersection can be defined
Figure 5: The training time of different models for 100
episodes. CoLight is efficient across all the datasets. The bar
for Individual RL on DN ewY ork is shadowed as its running
time is far beyond the acceptable time.
in different ways. And the results in Table 2 show that CoLight (us-
ing geo-distance) achieves similar performance with CoLight-node
under synthetic data, but largely outperforms CoLight-node under
real-world traffic. The reason could be that under synthetic data,
since the lane lengths of all intersections are the same, the top
closest neighboring intersections set according to geo-distance is
identical to that based on node distance. In the following parts of
our experiments, we only compare CoLight with other methods.
Figure 6: Performance of CoLight with respect to different
numbers of neighbors (Ni) on dataset DH anдzhou (left) and
D J inan (right). More neighbors (Ni ≤ 5) for cooperation
brings better performance, but too many neighbors (Ni >
5) requires more time (200 episodes or more) to learn.
Impact of Neighbor Number. In Figure 6, we show how the
5.7.2
number of neighbors Ni influences the performance and also
shed lights on how to set it. As is shown in Figure 6, when the
020406080Episode010002000Travel time (s)CGRLNeighbor RLIndividual RLOneModelGCNCoLightJumpstartTime to ThresholdAsymptotic020406080Episode100020003000Travel time (s)CGRLNeighbor RLIndividual RLOneModelGCNCoLight020406080Episode100020003000Travel time (s)CGRLNeighbor RLIndividual RLOneModelGCNCoLight020406080100Episode1500175020002250Travel time (s)CGRLNeighbor RLOneModelGCNCoLight020406080Episode50010001500Travel time (s)CGRLNeighbor RLIndividual RLOneModelGCNCoLight0255075100125150175200Episode50010001500Travel time (s)CGRLNeighbor RLIndividual RLOneModelGCNCoLight2357913Number of neighbors400500Travel time (s)100 episodes200 episodes1000 episodes2357913Number of neighbors300400500Travel time (s)100 episodes200 episodes1000 episodes number of neighbors grows from 2 to 5, the performance of CoLight
achieves the optimal. Further adding nearby intersections into the
neighborhood scope Ni, however, leads to the opposite trend. This
is because including more neighbors in the neighborhood results in
learning more relations. To determine signal control policy for each
intersection, computing the attention scores only on four nearby
intersections and itself seems adequate for cooperation with both
time and performance guarantee.
Impact of Attention Head Number. To evaluate the effective-
5.7.3
ness of multi-head attention, we test different numbers of attention
heads and find that moderate numbers of heads are beneficial to
better control intersection signals. As shown in Table 3, drivers
spend less time as the number of attention heads grows. However,
the benefits of more types of attention disappear as H exceeds 5.
Similar conclusions can be made on other datasets with details
unshown due to space limitation.
Table 3: Performance of CoLight with respect to different
numbers of attention heads (H) on dataset Grid6×6. More
types of attention (H ≤ 5) enhance model efficiency, while
too many (H > 5) could distract the learning and deteriorate
the overall performance.
#Heads
1
3
5
7
9
Travel Time (s)
176.32
172.47
170.11
174.54
174.51
6 ATTENTION STUDY
To analyze how well the neighborhood cooperation is implemented
via the attention mechanism, we will study both the spatial and
temporal distribution of attention learned by CoLight on both syn-
thetic and real-world data.
6.1 Spatial Distribution
The spatial distribution of attention scores indicates the importance
of different neighbors to the target agent. In this section, we analyze
the learned average attention that CoLight learns from the real-
world data in each episode. We have the following observations:
•Upstream vs. Downstream. Figure 7(a) shows an intersection (green
dot) in New York, whose neighborhood includes four nearby inter-
sections along the arterial. Traffic along the arterial is uni-directional
(blue arrow). From the attention distribution learned by CoLight,
we can see that while the majority of attention is allocated to it-
self, the upstream intersections (orange and blue lines) have larger
scores than downstream intersections (red and purple lines).
•Arterial vs. Side Street. Figure 7(b) shows an intersection (green
dot) in Hangzhou, whose neighborhood includes two intersections
along the arterial and two intersections on the side street. In the left
part of Figure 7(b), the arterial traffic is heavy and uni-directional
(thick horizontal arrow), and the side-street traffic is light and bi-
directional (thin vertical arrow). From the right part of Figure 7(b),
we can see that the arterial intersections (orange and blue lines)
have larger scores than side-street intersections (red and purple
lines).
(a) Intersection A in New York
(b) Intersection B in Hangzhou
Figure 7: Spatial difference of attention distribution learned
by CoLight during training process in real-world traffic. Dif-
ferent colored lines in the right figures correspond with
the colored dots in the left figures. Up: For intersection
A in DN ewY ork , major concentration is allocated on up-
stream intersections and A itself. Down: For intersection B
in DH anдzhou, major concentration is allocated on arterial
intersections and B itself.
6.2 Temporal Distribution
The temporal distribution of neighbors' attention scores throughout
a certain time period indicates the temporal change of neighbors'
influence. In this section, we analyze the learned attention of a
converged CoLight model in a 3 × 3 network under a 3600-second
traffic. As is shown in the upper part of Figure 8(b), there is small
basic traffic on every direction; at the same time, similar to the traffic
change between morning peak hours and evening peak hours, the
traffic of Inter #4 experiences great changes, where the traffic from
Inter #3 to Inter #4 drops (blue arrows in Figure 8(a)) and the traffic
from Inter #1 to Inter #4 increases (green arrows in Figure 8(a)).
(a) Grid3×3 road network
(b) Traffic flows and attention
scores of neighbors to Inter#4
Figure 8: Temporal distribution of attention score learned
by CoLight corresponds with temporally changing traffic.
The attention scores of the neighboring four intersections of
Inter #4 are shown in the lower part of Figure 8(b). Firstly, we can
see that the score from Inter #4 to Inter #4 always occupies the
largest purple area, indicating that the traffic condition of Inter
Inter 0Inter 1Inter 2Inter 3Inter 4Inter 5Inter 8Inter 7Inter 6500100015002000250030003500Timestep (s)0.00.20.40.60.81.0Attention ScoreInter 3 to Inter 4Inter 5 to Inter 4Inter 1 to Inter 4Inter 7 to Inter 4Inter 4 to Inter 4500100015002000250030003500Timestep (s)0100300Number of Vehicles/hInter 3 to Inter 4Inter 1 to Inter 4 #4 is of a great importance. Secondly, the attention scores of Inter
#1 and #7 increase with more traffic on South↔North direction.
Similarly, the attention scores of Inter #3 and #5 decrease as less
traffic on West↔East direction. The temporal change of attention
scores of Inter #4's four neighbors matches exactly with the change
of traffic flow from the four directions, which demonstrates that the
proposed CoLight model is capable to capture the key neighborhood
information over time, i.e., the temporal attention.
7 CONCLUSION
In this paper, we propose a well-designed reinforcement learning
approach to solve the network-level traffic signal control problem.
Specifically, our method learns the dynamic communication be-
tween agents and constructs an index-free model by leveraging
the graph attention network. We conduct extensive experiments
using synthetic and real-world data and demonstrate the superior
performance of our proposed method over state-of-the-art meth-
ods. In addition, we show in-depth case studies and observations
to understand how the attention mechanism helps cooperation.
We would like to point out several important future directions
to make the method more applicable to the real world. First, the
neighborhood scope can be determined in a more flexible way. The
traffic flow information between intersections can be utilized to
determine the neighborhood. Second, the raw data for observation
only include the phase and the number of vehicles on each lane.
More exterior data like the road and weather condition might help
to boost model performance.
ACKNOWLEDGMENTS
The work was supported in part by NSF awards #1652525 and
#1618448. The views and conclusions contained in this paper are
those of the authors and should not be interpreted as representing
any funding agencies.
REFERENCES
[1] Itamar Arel, Cong Liu, T Urbanik, and AG Kohls. 2010. Reinforcement learning-
based multi-agent system for network traffic signal control. IET Intelligent Trans-
port Systems 4, 2 (2010), 128 -- 135.
[2] Christopher M Bishop et al. 2006. Pattern recognition and machine learning
(information science and statistics). (2006).
[3] Eduardo Camponogara and Werner Kraus. 2003. Distributed learning agents in
urban traffic control. In Portuguese Conference on Artificial Intelligence. Springer,
324 -- 335.
[4] Weiyu Cheng, Yanyan Shen, Yanmin Zhu, and Linpeng Huang. 2018. A neural at-
tention model for urban air quality inference: Learning the weights of monitoring
stations. In Thirty-Second AAAI Conference on Artificial Intelligence.
[5] Tianshu Chu, Jie Wang, Lara Codecà, and Zhaojian Li. 2019. Multi-Agent Deep
Reinforcement Learning for Large-Scale Traffic Signal Control. IEEE Transactions
on Intelligent Transportation Systems (2019).
[6] Denise de Oliveira and Ana LC Bazzan. 2009. Multiagent learning on traffic lights
control: effects of using shared information. In Multi-agent systems for traffic
and transportation engineering. IGI Global, 307 -- 321.
[7] Christina Diakaki, Markos Papageorgiou, and Kostas Aboudolas. 2002. A multi-
variable regulator approach to traffic-responsive network-wide signal control.
Control Engineering Practice 10, 2 (2002), 183 -- 195.
[8] Kurt Dresner and Peter Stone. 2006. Multiagent traffic management: Opportu-
nities for multiagent learning. In Learning and Adaption in Multi-Agent Systems.
Springer, 129 -- 138.
[9] Samah El-Tantawy, Baher Abdulhai, and Hossam Abdelgawad. 2013. Multiagent
reinforcement learning for integrated network of adaptive traffic signal con-
trollers (MARLIN-ATSC): methodology and large-scale application on downtown
Toronto.
IEEE Transactions on Intelligent Transportation Systems 14, 3 (2013),
1140 -- 1150.
[10] Nathan H Gartner, Susan F Assman, Fernando Lasaga, and Dennis L Hou. 1991.
A multi-band approach to arterial traffic signal optimization. Transportation
Research Part B: Methodological 25, 1 (1991), 55 -- 74.
[11] Wallce D Hayes. 1970. Kinematic wave theory. Proc. R. Soc. Lond. A 320, 1541
(1970), 209 -- 226.
[12] PB Hunt, DI Robertson, RD Bretherton, and M Cr Royle. 1982. The SCOOT
on-line traffic signal optimisation technique. Traffic Engineering & Control 23, 4
(1982).
[13] PB Hunt, DI Robertson, RD Bretherton, and RI Winton. 1981. SCOOT-a traffic
responsive method of coordinating signals. Technical Report.
[14] Jiechuan Jiang, Chen Dun, and Zongqing Lu. 2018. Graph Convolutional Rein-
forcement Learning for Multi-Agent Cooperation. arXiv preprint arXiv:1810.09202
(2018).
[15] Peter Koonce et al. 2008. Traffic signal timing manual. Technical Report. United
States. Federal Highway Administration.
[16] Lior Kuyer, Shimon Whiteson, Bram Bakker, and Nikos Vlassis. 2008. Multia-
gent reinforcement learning for urban traffic control using coordination graphs.
In Joint European Conference on Machine Learning and Knowledge Discovery in
Databases. Springer, 656 -- 671.
[17] PR Lowrie. 1992. SCATS -- a traffic responsive method of controlling urban traffic.
Roads and traffic authority. NSW, Australia (1992).
[18] Tomoki Nishi, Keisuke Otaki, Keiichiro Hayakawa, and Takayoshi Yoshimura.
2018. Traffic Signal Control Based on Reinforcement Learning with Graph
Convolutional Neural Nets. In 2018 21st International Conference on Intelligent
Transportation Systems (ITSC). IEEE, 877 -- 883.
[19] Ann NowÃľ, Peter Vrancx, and Yann MichaÃńl De Hauwere. 2012. Game Theory
and Multi-agent Reinforcement Learning.
[20] LA Prashanth and Shalabh Bhatnagar. 2011. Reinforcement learning with func-
IEEE Transactions on Intelligent
tion approximation for traffic signal control.
Transportation Systems 12, 2 (2011), 412 -- 421.
[21] Roger P Roess, Elena S Prassas, and William R McShane. 2004. Traffic engineering.
Pearson/Prentice Hall.
[22] Bruno Castro Da Silva, Eduardo W. Basso, Filipo Studzinski Perotto, Ana L. C.
Bazzan, and Paulo Martins Engel. 2006. Improving reinforcement learning with
context detection. In International Joint Conference on Autonomous Agents &
Multiagent Systems.
[23] van der Pol et al. 2016. Coordinated Deep Reinforcement Learners for Traffic
Light Control. NIPS.
[24] Pravin Varaiya. 2013. The max-pressure controller for arbitrary networks of
signalized intersections. In Advances in Dynamic Network Modeling in Complex
Transportation Systems. Springer, 27 -- 66.
[25] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. In Advances in Neural Information Processing Systems. 5998 -- 6008.
[26] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro
arXiv preprint
Lio, and Yoshua Bengio. 2017. Graph attention networks.
arXiv:1710.10903 1, 2 (2017).
[27] Fo Vo Webster. 1958. Traffic signal settings. Technical Report.
[28] Hua Wei, Chacha Chen, Guanjie Zheng, Kan Wu, Vikash Gayah, Kai Xu, and
Zhenhui Li. 2019. PressLight: Learning Max Pressure Control to Coordinate
Traffic Signals in Arterial Network. In Proceedings of the 25th ACM SIGKDD
International Conference on Knowledge Discovery & Data Mining (KDD '19). ACM,
1290 -- 1298.
[29] Hua Wei, Guanjie Zheng, Vikash Gayah, and Zhenhui Li. 2019. A Survey on
Traffic Signal Control Methods. CoRR abs/1904.08117 (2019). arXiv:1904.08117
http://arxiv.org/abs/1904.08117
[30] Hua Wei, Guanjie Zheng, Huaxiu Yao, and Zhenhui Li. 2018.
Intellilight: A
reinforcement learning approach for intelligent traffic light control. In Proceedings
of the 24th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining. ACM, 2496 -- 2505.
[31] MA Wiering. 2000. Multi-agent reinforcement learning for traffic light con-
trol. In Machine Learning: Proceedings of the Seventeenth International Conference
(ICML'2000). 1151 -- 1158.
[32] 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).
[33] Huaxiu Yao, Xianfeng Tang, Hua Wei, Guanjie Zheng, Yanwei Yu, and Zhenhui Li.
2018. Modeling Spatial-Temporal Dynamics for Traffic Prediction. arXiv preprint
arXiv:1803.01254 (2018).
[34] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. 2016.
Image captioning with semantic attention. In Proceedings of the IEEE conference
on computer vision and pattern recognition. 4651 -- 4659.
[35] Huichu Zhang, Siyuan Feng, Chang Liu, Yaoyao Ding, Yichen Zhu, Zihan Zhou,
Weinan Zhang, Yong Yu, Haiming Jin, and Zhenhui Li. 2019. CityFlow: A Multi-
Agent Reinforcement Learning Environment for Large Scale City Traffic Scenario.
(2019).
|
1904.09630 | 5 | 1904 | 2019-11-21T15:11:55 | Genuine Personal Identifiers and Mutual Sureties for Sybil-Resilient Community Formation | [
"cs.MA",
"cs.SI"
] | While most of humanity is suddenly on the net, the value of this singularity is hampered by the lack of credible digital identities: Social networking, person-to-person transactions, democratic conduct, cooperation and philanthropy are all hampered by the profound presence of fake identities, as illustrated by Facebook's removal of 5.4Bn fake accounts since the beginning of 2019.
Here, we introduce the fundamental notion of a \emph{genuine personal identifier}---a globally unique and singular identifier of a person---and present a foundation for a decentralized, grassroots, bottom-up process in which every human being may create, own, and protect the privacy of a genuine personal identifier. The solution employs mutual sureties among owners of personal identifiers, resulting in a mutual-surety graph reminiscent of a web-of-trust. Importantly, this approach is designed for a distributed realization, possibly using distributed ledger technology, and does not depend on the use or storage of biometric properties. For the solution to be complete, additional components are needed, notably a mechanism that encourages honest behavior and a sybil-resilient governance system. | cs.MA | cs | Genuine Personal Identifiers and Mutual Sureties for
Sybil-Resilient Community Formation
Gal Shahaf1, Ehud Shapiro1 & Nimrod Talmon2
{gal.shahaf, ehud.shapiro}@weizmann.ac.il, [email protected]
1Weizmann Institute of Science
2Ben-Gurion University
9
1
0
2
v
o
N
1
2
]
A
M
.
s
c
[
5
v
0
3
6
9
0
.
4
0
9
1
:
v
i
X
r
a
Abstract. While most of humanity is suddenly on the net, the value
of this singularity is hampered by the lack of credible digital iden-
tities: Social networking, person-to-person transactions, democratic
conduct, cooperation and philanthropy are all hampered by the pro-
found presence of fake identities, as illustrated by Facebook's re-
moval of 5.4Bn fake accounts since the beginning of 2019.
Here, we introduce the fundamental notion of a genuine personal
identifier -- a globally unique and singular identifier of a person --
and present a foundation for a decentralized, grassroots, bottom-up
process in which every human being may create, own, and protect
the privacy of a genuine personal identifier. The solution employs
mutual sureties among owners of personal identifiers, resulting in a
mutual-surety graph reminiscent of a web-of-trust. Importantly, this
approach is designed for a distributed realization, possibly using dis-
tributed ledger technology, and does not depend on the use or storage
of biometric properties. For the solution to be complete, additional
components are needed, notably a mechanism that encourages hon-
est behavior [29] and a sybil-resilient governance system [30].
Introduction
Providing credible identities to all by 2030 is a UN Sustainable De-
velopment Goal. Yet, current top-down digital identity-granting solu-
tions are unlikely to close the 1Bn-people gap [6] in time, as they are
not working for citizens of failed states nor for people fleeing phys-
ical or political harshness [17, 5]. Concurrently, humanity is going
online at an astonishing rate, with more than half the world popu-
lation now being connected. Still, online accounts do not provide a
solution for credible digital identity either, as they may easily be fake,
resulting in lack of accountability and trust. For example, Facebook
reports the removal of 5.4Bn (!) fake accounts since the beginning of
2019 [16, 21].
The profound penetration of fake accounts on the net greatly ham-
pers its utility for credible human discourse and any ensuing delib-
erations and democratic decision making; it makes the net unsuit-
able for vulnerable populations, including children and the elderly;
it makes the use of the net for person-to-person transactions, notably
direct philanthropy, precarious; and in general it turns the net into an
inhuman, even dangerous, ecosystem. As an aside, we note that the
panacea of cryptocurrencies for the lack of credible personal iden-
tities on the net is the reckless employment of the environmentally-
harmful proof-of-work protocol.
Our aim is a conceptual and mathematical foundation for al-
lowing every person to create, own, and protect the privacy of a
globally-unique and singular identifier, henceforth referred to as
genuine personal identifier. We believe that successful deployment
and broad adoption of genuine personal identifiers will afford solu-
tions to these problems and more, providing for: egalitarian digital
democratic governance in local communities and in global move-
ments; digital cooperatives and digital credit unions; direct philan-
thropy; child-safe digital communities; preventing unwanted digi-
tal solicitation; banishing deep-fake (by marking as spam videos
not signed by a genuine personal identifier); credible and durable
digital identities for people fleeing political or economic harsh-
ness; accountability for criminal activities on the net; and egalitarian
cryptocurrencies employing an environmentally-friendly Byzantine-
agreement consensus protocol among owners of genuine personal
identifiers. Furthermore, genuine personal identifiers may provide
the necessary digital foundation for a notion of global citizenship
and, subsequently, for democratic global governance [31].
Granting an identity document by a state is a complex process as
it requires careful verification of the person's credentials. The pro-
cess culminates in granting the applicant a state-wide identifier that is
unique (no two people have the same identifier) and singular (no per-
son has two identifiers). Granting a genuine personal identifier might
seem even more daunting, as it needs to be globally-unique, not only
state-wide unique, except for following fundamental premise:
Every person deserves a genuine personal identifier. Thus, there are
no specific credentials to be checked, except for the existence of the
person. As a result, a solution for all people to create and own gen-
uine personal identifiers may focus solely on ensuring the one-to-one
correspondence between people and their personal identifiers.
A solution that is workable for all must be decentralized, dis-
tributed, grassroots, and bottom-up. Solid foundations are being laid
out by the notion of self-sovereign identities [22] and the W3C De-
centralized Identifiers [13] and Verifiable Claims [34] emerging stan-
dards, which aim to let people freely create and own identifiers and
associated credentials. We augment this freedom with the goal that
each person declares exactly one identifier as her genuine personal
identifier. We note that besides the genuine personal identifier, one
may create, own and use any number of identifiers of other types.
Becoming the owner of a genuine personal identifier is simple.
With a suitable app, this could be done with a click of a button:
1. Choose a new cryptographic key-pair (v, v−1), with v being the
public key and v−1 the private key.
2. Claim v to be your genuine personal identifier by publicly posting
a declaration that v is a genuine personal identifier, signed with
v−1.
Lo and behold! You have become the proud rightful owner of a gen-
uine personal identifier.1 Note that a declaration of a genuine per-
sonal identifier, by itself, does not reveal the person making the dec-
laration; it only reveals to all that someone who knows the secret key
for the public key v claims v as her genuine personal identifier. De-
pending on personality and habit, the person may or may not publicly
associate oneself with v. E.g., a person with truthful social media ac-
counts may wish to associate these accounts with its newly-minted
genuine personal identifier.
If becoming the rightful owner of a genuine personal identifier
is so simple, what could go wrong? In fact, so many things can go
wrong, that this paper is but an initial investigation into describing,
analyzing, and preventing them. Some of them are enumerated be-
low; h denotes an agent:
1. The key-pair (v, v−1) is not new, or else someone got hold of it
between Step 1 and Step 2 above. Either way, someone else has
declared v to be a genuine personal identifier prior to the declara-
tion by h. In which case h cannot declare v.
2. Agent h failed to keep v−1 secret so that other people, e.g. h(cid:48),
know v−1, in which case h(cid:48) is also an owner of v and, thus, v is
compromised. Figure 1 (left) illustrates a compromised personal
identifier.
3. The agent h intended to divulge his association with his public
key v only on a need-to-know basis, but the association of v and
h has become public knowledge, prompting agent h to replace his
genuine personal identifier v with a new one.
4. Agent h declared v as his genuine personal identifier, but later also
declared another personal identifier v(cid:48). Then, v and v(cid:48) are dupli-
cates, v(cid:48) is a sybil, and agent h is corrupt. An honest agent does
not declare sybils. Figure 1 illustrates honest and corrupt agents.
We aim to develop the foundation for genuine personal identifiers
utilizing basic concepts of public-key cryptography, graph theory, so-
cial choice theory and game theory. Here, we focus on utilizing the
first two disciplines; see [30] for sybil-resilient social choice, which
provides a foundation for democratic governance of digital commu-
nities with bounded sybil penetration; incorporating sybils in cooper-
ative game theory, with the goal of forming a sybil-free grand coali-
tion, is work in progress.
Related Work. Sybil-resilience has received considerable attention
in AI research [2, 3, 7, 8, 10, 11, 12, 30, 35, 36] as elaborated below.
Philosophically, genuine personal identifiers aim to bridge the gap
between agents and their corresponding identifiers. This distinction,
acknowledged in semiotics as the difference between signified and
signifier, strongly relates to the study of sense and reference initi-
ated by Frege [15] followed by vast literature in analytic philosophy
and the philosophy of language. Conceptually, the formal framework
suggested here may be viewed as an attempt to computationally re-
alize unique and singular signifiers of agents in a distributed setting.
Practically, digital identities are a subject of extensive study, with
many organizations aiming at providing solutions, notably Self-
Sovereign Identifiers [22]. Business initiatives include the Decen-
tralized Identity Foundation (identity.foundation), the Global Iden-
tity Foundation [14] and Sovrin [33]. None of these projects are con-
cerned with the uniqueness or singularity of personal identities.
1 As the public key may be quite long, one may also associate oneself with
a shorter "nickname", a hash of the public key, e.g. a 128-bit hash (as a
UUID) or a 256-bit hash (as common in the crypto world).
Other high-profile projects to provide nationwide digital identities
include India's Aadhaar system [23], Sierra Leone's Kiva identity
protocol [32] and the World Food Programmes cash aid distribution
program in refugee camps [19]. Here we are concerned with self-
sovereign and global personal identities, not bound to any national
boundaries or entities, and argue that top-down approaches fail to
provide such a solution. In this context, we mention the concept of
Proof of Personhood [7], aiming at providing unique and singular
identities by means of conducting face-to-face encounters, an ap-
proach suitable only for small communities.
Our solution is based upon the notion of trust, thus we men-
tion Andersen et al. [3], studying axiomatizations of trust systems.
They are not concerned, however, with sybils, but with quality of
recommendations. We mention work on sybil-resilient community
growth [26], describing algorithms for the growth of an online com-
munity that keep the fraction of sybils in it small; and work on sybil-
resilient social choice [30], describing aggregation methods to be ap-
plied in situations where sybils have infiltrated the electorate. In these
two papers, a notion of genuine identities and sybils is used without
specifying what they are; here, we define a concrete notion of gen-
uine personal identifiers, and derive from it a formal definition of
sybils and related notions of honest and corrupt agents and byzan-
tine identifiers. Finally, we mention the work of Conitzer
[12, 2]
regarding computerized tests for sybil fighting: A test that is hard for
a person to pass more than once [12] and a test that is hard to pass
simultaneously by one person [2].
Genuine Personal Identifiers
Ingredients. The ingredients needed for a realization of genuine per-
sonal identifiers are:
1. A set of agents. It is important to note that, mathematically, the
agents form a set (of unique entities) not a multiset (with dupli-
cates). Intuitively, it is best to think of agents as people (or other
physical beings with unique personal characteristics, unique per-
sonal history, and agency, such as intelligent aliens), which cannot
be duplicated, but not as software agents, which can be.
2. A way for agents to create cryptographic key-pairs. This can be
realized, e.g., using the RSA standard [27]. Our solution does not
require a global standard or a uniform implementation for pub-
lic key encryption: Different agents can use different technologies
for creating and using such key-pairs, as long as the signatures-
verification methods are declared.
3. A way for agents to sign strings using their key-pairs. As we as-
sume cryptographic hardness, an agent that does not know a cer-
tain key-pair cannot sign strings with this key-pair.
4. A bulletin board or public ledger, to which agents may post and
observe signed messages, where all agents observe the same order
of messages. The weaker requirement that the same order is ob-
served only eventually, as is standard with distributed ledger pro-
tocols, could also be accommodated. Considering partial orders is
a subject of future work.
Agents and their Personal Identifiers. We assume a set of agents
H that is fixed over time.2 Agents can create new key-pairs (v, v−1).
We assume that an agent that has a key-pair can sign a string, and de-
note by v(x) the string resulting from signing the string x with v−1.
Intuitively, each agent corresponds to a human being. Importantly,
2 Birth and death of agents will be addressed in future work.
Figure 1. A compromised identifier (left), honest agent (middle), and corrupt agent (right).
members of the set H of agents (e.g., containing all human beings)
cannot be referenced explicitly and, in particular, posted signed mes-
sages never refer directly to agents h ∈ H. A key motivation for
our work is providing people with digital genuine personal identi-
fiers without accessing any of their intrinsic (e.g. biometric) private
properties and without depending upon such properties as identifiers.
As we aim personal identifiers to be self-sovereign identities that
conform to the W3C Decentralized Identifiers emerging standards,
we let agents create and own their personal identifiers. An agent
h can publicly declare a personal identifier v for which it knows
the private key v−1. A personal identifier declaration has the form
v(gid(v)) and can be effected by agent h posting v(gid(v)) to a pub-
lic ledger. We denote this action by declareh(gid(v)) ∈ C. Recall
that all agents have the same view of the sequence of all declarations
made; subsequent work may relax this assumption.
Definition 1 (Personal Identifier) Let C be a sequence of personal
identifier declaration events and declareh(gid(v)) ∈ C the first dec-
laration event in which v occurs. Then v is a personal identifier and
h is the rightful owner of v, given C.
Definition 2 (Genuine Personal Identifier, Sybil, Honest, and
Corrupt Agents) Let C be a sequence of personal identifier dec-
laration events and h be the rightful owner of personal identifier v in
C. Then v is genuine if it is the first personal identifier declared in C
by h, else v is a sybil. An agent h is corrupt if it declares any sybils,
else h is honest. (All notions are relative to C.)
See Figure 1 and some remarks: (1) An agent is the rightful owner
of its genuine personal identifier as well as of any subsequent sybils
that it declares. (2) If h, the rightful owner of v, is corrupt, then its
first declared identifier is genuine and the rest of its declared iden-
tifiers are all sybils. (3) An honest agent may create and use many
key-pairs for various purposes, yet remain honest as long as it has
declared at most one public key as a personal identifier.
Mutual Sureties and Their Graphs
A key element of our approach is the pledging of mutual sureties by
agents. Intuitively, mutual surety pledges provide a notion of trust
between the owners of personal identifiers. They allow two agents
that know each other and know the personal identifiers declared by
each other to vouch for the other regarding the good standing of the
personal identifiers. This notion is key to help honest agents fend-off
sybils. We consider several types of mutual sureties, of increasing
strength, and illustrate the corresponding sybil-resilience each type
of mutual sureties can obtain.
For an agent, say h, to provide surety regarding the personal iden-
tifier of another agent, say h(cid:48), h first has to know h(cid:48). How this knowl-
edge is established is not specified in our formal framework, but this
is quite an onerous requirement that cannot be taken lightly or sat-
isfied casually. E.g., we may assume that one knows one's family,
friends and colleagues, and may diligently get to know new people
if one so chooses. We consider several types of sureties of increas-
ing strength, in which an agent h with personal identifier v makes
a pledge regarding the personal identifier v(cid:48) of another agent h(cid:48); all
assume that the agent h knows the agent h(cid:48). We describe four Surety
Types, which are cumulative as each includes all previous ones, ex-
plain on what basis one may choose to pledge each of them, and
present results that utilize them,
Surety of Type 1: Ownership of a personal identifier. Agent
pledges that agent h(cid:48) owns personal identifier v(cid:48).
Agent h(cid:48) can prove to agent h that she owns v(cid:48) without disclos-
ing v(cid:48)−1 to h. This can be done, for example, by h asking h(cid:48) to sign
a novel string x and verifying that v(cid:48)(x) is signed using v(cid:48)−1. This
surety type is the weakest of all four, it is the one given in "key sign-
ing parties", and is implicitly assumed by applications such as PGP
and web-of-trust [1]. For a given surety type, we say that the surety is
violated if its assertion does not hold; in particular, a surety of Type
1 is violated if h(cid:48) in fact does not know the secret key v(cid:48)−1.
h
In general, mutual surety between two agents with two per-
sonal identifiers is pledged by each of the two agents pledging
a surety to the personal identifier of the other agent.3 We de-
fine below three additional surety types, where the format of a
surety pledge of Type X by the owner of v to the owner of v(cid:48) is
v(suretyX(v(cid:48))), X ∈ {1, 2, 3, 4}. The corresponding surety event
is pledgehv(suretyX(v(cid:48))), and the surety enters into effect once
both parties have made the mutual pledges. We now take C to be a
record of both declaration events and pledge events.
3 We consider undirected graphs, as we require surety to be symmetric. In-
deed, one may consider directed sureties.
Definition 3 (Mutual Surety) The personal identifiers v, v(cid:48) have
there are
mutual surety of
h, h(cid:48) ∈ H for which pledgehv(suretyX(v(cid:48))) ∈ C
&
pledgeh(cid:48) v(cid:48)(suretyX(v)) ∈ C, in which case h and h(cid:48) are the wit-
nesses for the mutual surety between v and v(cid:48).
type X, X ∈ {1, 2, 3, 4},
if
A sequence of events induces a sequence of surety graphs in which
the vertices are personal identifiers that correspond to personal iden-
tifier declarations and the edges correspond to mutual surety pledges.
Definition 4 (Surety Graph) Let C = c1, c2, . . . be a sequence of
events and let Ck denote its first k ≥ 0 events. Then, for each k ≥ 0,
Ck induces a surety graph of type X, GXk = (Vk, EXk), X ∈
{1, 2, 3, 4}, as follows:4
Vk = {v declareh(gid(v)) ∈ Ck for some h ∈ H}
EXk = {(v, v
(cid:48)
) pledgehv(suretyX(v
pledgeh(cid:48) v
for some h, h
(cid:48)
(cid:48)
)) ∈ Ck,
(suretyX(v)) ∈ Ck,
(cid:48) ∈ Vk}
(cid:48) ∈ H, v, v
Remark 1 Observe that mutual sureties can be easily pledged by
agents, technically. However, we wish agents to be prudent and sin-
cere in their mutual surety pledges. Thus, we expect a mechanism
that, on one hand, rewards the pledging of sureties but, on the other
hand, punishes for surety violations, for example based on the ap-
proach of [29]. While the specifics of such a mechanism is beyond
the scope of the current paper, note that with such a mechanism in
place, the commissive illocutionary force [28] of a surety pledge will
come to bear.
Updating a Personal Identifier with Mutual Sureties
Once creating a genuine personal identifier is provided for, one must
also consider the many circumstances under which a person may
wish to update their personal identifier:
1. Identifier loss: The private key was lost.
2. Identifier theft: The private key was stolen, robbed, extorted, or
otherwise compromised.
3. Identifier breach of privacy: The association between the per-
sonal identifier and the person was accidentally or maliciously
disclosed with unwarranted consequences.
4. Identifier refresh: Proactive identifier update to protect against
all the above.
Update in Case of Loss or Theft. The personal identifier declara-
tion event declareh(gid(v)) establishes v as a personal identifier.
To support updating a personal identifier, we add the personal iden-
tifier update event declareh(gid(v, v(cid:48))), which declares that v(cid:48) is a
new personal identifier that replaces v. A public declaration of iden-
tifier update has the form v(gid(v, v(cid:48))), i.e., it is signed with the new
identifier. We refer to declarations of both types as personal iden-
tifier declarations, and extend the assumption that a new identifier
can be declared at most once to this broader definition of identifier
declaration. The validity of an identifier update declaration is defined
inductively, as follows.
4 We allow surety pledges to be made before the corresponding personal iden-
tifier declarations, as we do not see a reason to enforce order.
Definition 5 (Valid identifier Update declaration) Let C be a se-
quence of declarations, V the set of global identities declared in C,
and h ∈ H. A personal identifier update event over V has the form
declareh(gid(v, v(cid:48))), v, v(cid:48) ∈ V .
A personal identifier update event declareh(gid(v, v(cid:48))) ∈ C is
valid and h is the rightful owner of v if it is the first identifier decla-
(cid:3)
ration event of v and h is the rightful owner of v(cid:48).
Valid personal identifier declarations should form linear chains, one
for each agent, each starting from gid(v) and ending with the cur-
rently valid personal identifier of the agent:
Definition 6 (Identifier Provenance Chain) Let C be a sequence of
declarations and V the declared set of global identities. An identifier
provenance chain (provenance chain for short) is a subsequence of C
of the form (starting from the bottom):
declarehk (gid(vk, vk−1)),
declarehk−1 (gid(vk−1, vk−2)),
. . .
declareh1 (gid(v1)).
Such a provenance chain is valid if the declarations in it are valid.
Such a provenance chain is maximal if there is no declaration
declareh(gid(v, vk)) ∈ C
for any v ∈ V and h ∈ H. A personal identifier v is current in C if it
is the last identifier v = vk in a maximal provenance chain in C. (cid:3)
Note that it is very easy for an agent to make an update declaration
for its identifier. However, it is just as easy for an adversarial agent
wishing to steal the identifier to make such a declaration. Hence, this
ability must be coupled with a mechanism that protects the rightful
owner of an identifier from identifier theft through invalid identifier
update declarations. Here we propose to use a stronger type of mu-
tual sureties to support valid identifier update declarations and help
distinguish between them and invalid declarations.
Surety of Type 2: Rightful ownership of a personal identifier.
Agent h pledges that h(cid:48) is the rightful owner of personal iden-
tifier v(cid:48).
In addition to proving to h that it owns v(cid:48), h(cid:48) must provide evidence
that h(cid:48) itself, and not some other agent, has declared v(cid:48). A selfie
video of h(cid:48) pressing the declare button with v(cid:48), signed with a certified
timestamp promptly after the video was taken, and then signed by v(cid:48),
may constitute such evidence. A suitable app may record, timestamp,
and sign such a selfie video automatically during the creation of a
genuine personal identifier. In particular, this surety is violated if h(cid:48)
in fact did not declared v(cid:48) as a personal identifier.
Note that immediately following an identifier update declaration,
the new identifier may not have any surety edges incident to it. Thus,
as a crude measure, we may require that the identifier update would
come to bear only after all the Type 2 surety neighbors of the old
identifier, or a sufficiently large majority of them, would update their
mutual sureties to be with the new identifier. To achieve that, an agent
wishing to update its identifier would have to approach its neighbors
and to create such updated Type 2 mutual surety pledges.
Example 1 Consider two friends, agent h and agent h(cid:48) having a mu-
tual surety pledge between them. If h(cid:48) would lose her identifier, she
would create a new key-pair, make an identifier update declaration,
and ask h for a new mutual surety pledge between h's identifier and
h(cid:48)'s new identifier.
The following observation follows from: (1) a valid provenance
chain has a single owner; and (2) whether a Type 2 surety between
two identifiers is violated depends on their rightful owners.
Observation 1 Let C be a sequence of update declarations and
C1, C2 be two valid provenance chains in C. If a Type 2 surety pledge
between two global identities v1 ∈ C1, v2 ∈ C2 is valid, then any
Type 2 surety pledge between two personal identifiers in these prove-
nance chains, u1 ∈ C1, u2 ∈ C2 is valid.
The import of Observation 1 is that a Type 2 mutual surety can
be "moved along" valid provenance chains as they grow, without be-
ing violated, as it should be. Below we argue that invalid identifier
update declarations are quite easy to catch, thus the risk of stealing
identities can be managed. In effect, we show the value of Type 2
surety pledges in defending an identifier against theft via invalid up-
date declarations.
Let C be a sequence of declarations, C1, C2 be two provenance
chains in C, and assume that there is a valid Type 2 surety pledge be-
tween the two current global identifiers v1 ∈ C1, v2 ∈ C2, made by
h1 = Marry and h2 = John. Now assume that the identifier update
declaration c = declareh(gid(v, v1)) is made, namely, some agent
Sue (cid:54)= Marry has declared to replace v1 by v. Then, it will be hard for
Marry to secure surety from John and, if she attempts to do so, then
John will know that c is not valid and thus (if John is honest) a Type
2 mutual surety between v and v2 will not be established. Consider
the following case analysis:
• Assume Marry notices c. Then she would inform John that she did
• Assume that John notices c. He would approach Marry to update
the Type 2 mutual surety between them accordingly; Marry would
deny owning v, and thus John will know that c is invalid.
• Alternatively, Sue would approach John to update the Type 2 mu-
tual surety of v2 with v1 to be with v instead; John will see (or
suspect, if Sue did not reveal herself) that Sue is not Marry, will
double check with Marry and deem the declaration c invalid.
not declare c, and thus John will know that c is not valid.
Reset in Case of Breach of Privacy. Note that provenance chains
address identifier update in case of loss of theft, but do not address
breach of privacy, as the updated new identifier is publicly tied to the
previous one. To address breach of privacy, a person first has to inval-
idate his existing identifier, then create a new genuine identifier that
is not linked to the previous one; this may result in loss of any public
credit and goodwill associated with the old genuine personal identi-
fier, but there may be circumstances in which a person would need to
protect his privacy even at the expense of such loss. To facilitate that,
an identifier reset declaration uses the special value null and has the
form declareh(gid(v, null)), which nullifies v as a personal identi-
fier, and enters into effect if supported by those who initially provided
the surety to v. Agent h would then be free to declare a new personal
identifier v(cid:48), which is not tied to the now-defunct v, and at the same
time without v(cid:48) being a sybil and without h becoming corrupt as a
result of this declaration.
We note that using null, an initial genuine identifier declaration
could be of the form declareh(gid(null, v)), thus dispensing with
the unary declaration format declareh(gid(v)) altogether.
GDPR. In general, our approach does not imply any "data con-
trollers" or "data processors" [24] other than people and their trusted
friends, but if realized on a large scale it might require such, possibly
democratically-appointed by the large community that needs them.
Furthermore, our approach and does not record or store any "person-
ally identifiable information", and as such we believe is compliant
with GDPR [24]. One exception is perhaps the association of a gen-
uine personal identifier with the person that owns it, which could be
made public on purpose by the owner, inadvertently by the owner
or by another person, or maliciously by another person. Closely tied
to this exception is GDPR's "right to be forgotten" [24, Article 17],
which is notoriously difficult to realize in a distributed setting and
hence resulted in sweeping legal opinions regarding the inapplica-
bility of distributed ledger/blockchain technology for the storage of
personally identifiable information [25].
Our method of personal identifier reset first invalidates an existing
identifier (which could be coupled with a demand of erasure of every
reference to this identifier from any public data controller, e.g. Face-
book) and then creates an independent and unlinked new personal
identifier. This may be the first proposal on how to realize the "right
to be forgotten" in a decentralized setting.
Sybil- and Byzantine-Resilient Community Growth
Ideally, we would like to attain sybil-free communities, but acknowl-
edge that one cannot prevent sybils from being declared and, fur-
thermore, perfect detection and eradication of sybils is out of reach.
Thus, our aim is to provide the foundation for a digital community
of genuine personal identifiers to grow by admitting new identifiers
indefinitely, while retaining a bounded sybil penetration. As noted
above, democratic governance can be achieved even with bounded
sybils penetration [30].
Community history. For simplicity, we assume a single global com-
munity A and consider elementary transitions obtained by either
adding a single member to the community or removing a single com-
munity member:
Definition 7 (Elementary Community Transition) Let A, A(cid:48) de-
note two communities in V . We say that A(cid:48) is obtained from A by an
elementary community transition, and we denote it by A → A(cid:48), if:
• A(cid:48) = A, or
• A(cid:48) = A ∪ {v} for some v ∈ V \ A, or
• A(cid:48) = A \ {v} for some v ∈ A.
Definition 8 (Community History) Let C = c1, c2, ... be a se-
quence of events. A community history wrt. C is a sequence of com-
munities A = A1, A2, . . . such that Ai ⊆ Vi and Ai → Ai+1 holds
for every i ≥ 1.
We do not consider community governance in this paper, only the
effects of community decisions to add or remove members. Hence,
we assume that the sequence of events C includes the events addA(v)
and removeA(v). With this addition, C = c1, c2, . . . induces a com-
munity history A1, A2, ..., where Ai+1 = Ai∪{v} if ci = addA(v)
for v ∈ V \ Ai; Ai+1 = Ai \ {v} if ci = removeA(v) for v ∈ Ai;
else Ai+1 = Ai.
Definition 9 (Community, Sybil penetration rate) Let C be a se-
quence of events and let S ⊆ V denote the sybils in V wrt. C. A
community in V is a subset of identifiers A ⊆ V . The sybil penetra-
tion σ(A) of the community A is given by σ(A) =
.
A∩S
A
The following observation is immediate.
Observation 2 Let A1, A2, ... be the community history wrt. a se-
quence of declarations C. Assume that A1 ⊆ H, and that whenever
Ai+1 = Ai ∪ {v} for some v /∈ Ai, it holds that P r(v ∈ S) ≤ σ
for some fixed 0 ≤ σ ≤ 1. Then, the expected sybil penetration rate
for every Ai is at most σ.
That is, the observation above states that a sybil-free community can
keep its sybil penetration rate below σ, as long as the probability of
admitting a sybil to it is at most σ. While the simplicity of Observa-
tion 2 might seem promising, its premise is naively optimistic. Due
to the ease in which sybils can be created and to the benefits of own-
ing sybils in a democratic community, the realistic scenario is of a
hoard of sybils and a modest number of genuine personal identifiers
hoping to join the community. Furthermore, once a fraction of sybils
has already been admitted, it is reasonable to assume that all of them
(together with their perpetrators of course) would support the admis-
sion of further sybils. Thus, there is no reason to assume neither the
independence of candidates being sybils, nor a constant upper bound
on the probability of sybil admission to the community. Hence, in the
following we explore sybil-resilient community growth under more
realistic assumptions.
Sybil-Resilient Community Growth
A far more conservative assumption includes a process employed by
the community with the aim of detecting sybils. We shall use the ab-
stract notion of sybil detector in order to capture such process, that
may take the form of a query, a data-based comparison to other iden-
tifiers, or a personal investigation by some other agent. To leverage
this detector to sybil-resilient community growth regardless of the
sybil distribution among the candidates, we shall utilize a stronger
surety type, defined as follows:
Surety of Type 3: Rightful ownership of a genuine personal
identifier. Agent h pledges Surety Type 2 and that v(cid:48) is the gen-
uine personal identifier of h(cid:48).
Providing this surety requires a leap of faith. In addition to h ob-
taining from h(cid:48) a proof of rightful ownership of v(cid:48), h must also trust
h(cid:48) not to have declared any other personal identifier prior to declar-
ing v(cid:48). There is no reasonable way for h(cid:48) to prove this to h, hence the
leap of faith.
Since Type 3 sureties inherently aim to distinguish between gen-
uine identifiers and sybils, sybil-resilient community growth is es-
tablished upon the underlying G3 surety graph. Specifically, we con-
sider a setting where potential candidates to join the community are
identifiers with a surety obtained from current community members.
Conversely, we consider a violation of a surety in one direction as
a strong indication that the surety in the other direction is violated.
That is, if (v, v(cid:48)) ∈ E and v(cid:48) was shown to be sybil, (i.e., h(cid:48) has
declared some other v(cid:48)(cid:48) as a personal identifier before declaring v(cid:48) as
a personal identifier), then v should undergo a thorough investigation
in order to determine whether it is sybil as well.
Next, we formalize this intuition in a simple stochastic model
where admissions of new members are interleaved with random sybil
detection among community members:
(1) An identifier is admitted to the community via an elementary com-
munity transition A → A ∪ {v} only if there is some a ∈ A with
(a, v) ∈ E.
(2) Every admittance of a candidate is followed by a random sybil
detection within the community: An identifier a ∈ A is chosen
uniformly at random. If a is genuine it is declared as such. If a is
sybil, it is successfully detected with probability 0 < p ≤ 1.
(3) The detection of a sybil implies the successful detection of its en-
tire connected sybil component (with probability 1). That is, if a is
detected as sybil, then the entire connected component of a in the
sybil subgraph GS is detected and expelled from the community.
(4) The sybils are operated from at most j ≤ k disjoint sybil compo-
nents in A. Furthermore, we assume that sybils join sybil compo-
nents uniformly at random, i.e., a new sybil member has a surety
to a given sybil component with probability 1
k , else, it forms a new
sybil component with probability k−j
k .
Note that assumption (2) is far weaker than the premise in Obser-
vation 2 as it presumes nothing on the sybil penetration among the
candidates, but rather on the proactive ability to detect a sybil, once
examined. Assumption (3) exploits the natural cooperation among
sybils, especially if owned by the same agent, and assumes that if a
sybil is detected by a shallow random check with probability p, then
all its neighbours will be thoroughly investigated and will be detected
if sybil with probability 1, continuing the investigation iteratively un-
til the entire connected sybil component of the initially-detected sybil
is identified. In assumption (4), the parameter k and the locations of
the components are adversarial -- the attacker may choose how to op-
erate. While realistic attackers may also choose to which component
shall the new (sybil) member join, uniformity is assumed to simplify
the analysis. Possible relaxations of this model are future work.
For this setting we show an upper bound on the expected sybil pen-
etration, assuming bounded computational resources of the attacker.
Theorem 1 In the stochastic model described above, obtaining an
expected sybil penetration E[σ(A)] ≥ is NP-hard for every > 0.
Let Xi ⊆ A denote a sybil component within the commu-
Proof.
nity at time i. In the stochastic model described above, Xi is detected
and immediately expelled with probability p· Xi
Ai . The expected size
of the component in this model is obtained in a steady state, i.e., in a
state i in time where E[Xi+1] = Xi, that is:
(1 − px/n) · 1
k
· (x + 1) + (1 − px/n) · (1 − 1
k
where n := Ai and x = Xi. It follows that x2 + 1
pk = 0.
Solving this quadratic equation implies that the size of a single sybil
component in the steady state is x ≤ (cid:112)n/pk. It follows that the
number of sybils in the community in a steady state is xk ≤(cid:112)nk/p.
) · x = x,
k x − n
The crucial observation now is that operating from k nonempty
sybil components corresponds to obtaining an independent set of
size k (at least, choosing a single vertex in each component). The
theorem follows from the fact that approximating independent set
(cid:3)
within a constant factor is NP-hard (see, e.g., [4]).
The following corollary establishes an upper bound on the sybil
penetration rate regardless of the attacker's computational power.
The result is formulated in terms of the second eigenvalue of the
graph restricted to Ai. (λ(GAi ) is defined in the supplementary ma-
terials section.
Corollary 1 Let A = A1, A2, . . . be a community history wrt. a
sequence of events C. If every community Ai ∈ A with Ai = Ai−1 (cid:93)
{v} satisfies λ(GAi ) < λ, then the expected sybil penetration in
every Ai ∈ A under the stochastic model depicted above, is at most
(cid:112)λ/p.
utilize the notation A (cid:16) A(cid:48) to indicate that A(cid:48) was obtained from
A via a finite sequence of elementary community transitions of in-
cremental growth. Formally, A (cid:16) A(cid:48) if there exists k ∈ N with
A = A0 → A1 → ...Ak = A(cid:48), with Ai \ Ai−1 = 1 for all i ∈ [k].
Theorem 2 Let A = A1 (cid:16) A2 (cid:16) A3 (cid:16) . . . be a community his-
tory wrt. a sequence of events C. Set a sequence of degrees d1, d2, . . .
and parameters α, β, γ, δ ∈ [0, 1]. Assume:
1. deg(v) ≤ di for all v ∈ Ai, i ∈ N.
2. Every a ∈ Ai satisfies {x∈Ai (a,x)∈E}
A1 ≤ β.
A1∩B
3.
4. e(Ai∩H,Ai∩B)
5. Ai \ Ai−1 ≤ δAi−1, with β + δ ≤ 1
2 .
6. Φ(GAi ) > γ
α ·(cid:16) 1−β
(Ai∩H) ≤ γ.
≥ α.
volAi
di
(cid:17)
.
β
Then, every community Ai ∈ A has Byzantine penetration
β(Ai) ≤ β.
Roughly speaking, Theorem 2 suggests that whenever: (1) Each
graph GAi has a bounded degree di; (2) Sufficiently many edges are
within Ai; (3) Byzantine penetration to A1 is bounded; (4) Edges be-
tween harmless and byzantine identifiers are scarce; (5) Community
growth in each step is bounded; (6) The conductance within GAi is
sufficiently high; Then, the community may grow indefinitely with
bounded byzantine penetration.
As Theorem 2 allows byzantine resilient growth at a fixed rate δ,
it may be interpreted as an extension of a related result by Poupko
et al. [26], where new members were added one at a time. More-
over, [26] assumed a notion of honest (what we refer to as genuine)
and byzantine identities without defining what they are; here we for-
malize these notions and provide more sound definitions of harmless
and byzantine identifiers. A formal definition of graph conductance
Φ and the proof of Theorem 2 may be found in the supplementary
material.
Remark 2 A potential application of Theorem 2 is a byzantine-
resilient union of two communities. Let A, A(cid:48) ⊆ V denote two com-
munities that have some overlap (non-empty intersection) and wish
to unite into A2 := A ∪ A(cid:48). Then, if Theorem 2 holds for (A1, A2)
in case A1 := A and also in case A1 := A(cid:48), this would provide both
A and A(cid:48) the necessary guarantee that the union would not result in
an increase of the sybil penetration rate for either community.
Outlook
Digital identity systems face the "Decentralized Identity Trilemma",
of being (i) privacy-preserving, (ii) sybil-resilient and (iii) self-
sovereign, all at the same time [20]. It has been claimed that no ex-
isting identity system satisfies all three corners of the Trilemma [20];
our approach may be the first to do so.
While this paper provides a formal mathematical framework, we
aimed the constructions to be readily amenable to implementation.
Realizing the proposed solution entails developing additional com-
ponents, notably sybil-resilient governance mechanisms, e.g. along
the lines of [30]; a mechanism for encouraging honest behavior and
discouraging corrupt behavior, e.g. along the lines of [29]; and a
cryptocurrency to fuel such a mechanism and the system in general.
Once all components have been designed, we aim to implement, sim-
ulate, test, deploy, and evaluate the proposed framework, hopefully
realizing the potential of genuine personal identifiers.
Figure 2. Violations of Sureties of Type 3 and Type 4. The axis of time
goes from left to right, with the points in time in which v(cid:48) was declared by
h(cid:48) and the surety from h to h(cid:48) was declared. Then, the braces describe the
time regions in which, if another identifier v(cid:48)(cid:48) was to be declared by h(cid:48),
would correspond to a violation of a surety of Type 3 or 4.
Proof.
Recall that the size of the maximal independent set is
a trivial upper bound on k. The cardinality of an independent set
in a λ-expander is at most λn [18]; thus, k ≤ λn. It follows
that the number of sybils in the community in a steady state is
(cid:3)
xk ≤(cid:112)nk/p ≤ n(cid:112)λ/p.
Byzantine-Resilient Community Growth
Here we consider the challenge of byzantine-resilient community
growth. Intuitively, the term byzantines aims to capture identifiers
owned by agents that are acting maliciously, possibly in collabora-
tion with other malicious agents. Formally, we define byzantines as
follows.
Definition 10 (Byzantine and harmless identifiers, Byzantine
penetration) An identifier is said to be byzantine if it is either a
sybil or the personal identifier of a corrupt agent. Non-byzantine
identifiers are referred to as harmless. We denote the byzantine and
harmless identifiers in V by B, H ⊆ V , respectively. The byzantine
penetration β(A) of a community A ⊆ V is given by β(A) =
A∩B
.
Since S ⊆ B, it holds that σ(A) ≤ β(A) for every community A,
hence an upper bound on the byzantine penetration also provides an
upper bound on the sybil penetration.
A
As byzantine identifiers include genuine identifiers, they are inher-
ently harder to detect, and thus the detection-based model described
in Section is no longer applicable in this setting. Rather, to achieve
byzantine-resilient community growth, we rely on a stronger surety
type, defined as follows:
Surety of Type 4: Rightful ownership of a genuine personal
identifier by an honest agent. Agent h pledges Surety Type 3
and, furthermore, that v(cid:48) is a genuine personal identifier of an hon-
est agent h(cid:48).
Here h has to put even greater trust in h(cid:48): Not only does h has
to trust that the past actions of h(cid:48) resulted in v(cid:48) being her genuine
personal identifier, but she also has to take on faith that h(cid:48) has not
declared any sybils since and, furthermore, that h(cid:48) will not do so in
the future. Note that a Type 4 surety is violated if after h(cid:48) declares v(cid:48)
it ever declares some other v(cid:48)(cid:48) as a personal identifier. See Figure 2
for illustrations of violations of sureties of Types 3 and 4.
In the following, we provide sufficient conditions for Type 4
sureties to be used for byzantine-resilient community growth. We
References
[1] Alfarez Abdul-Rahman, 'The pgp trust model', in EDI-Forum: the
Journal of Electronic Commerce, volume 10 (3), pp. 27 -- 31, (1997).
[2] Garrett Andersen and Vincent Conitzer, 'Atucapts: Automated tests that
a user cannot pass twice simultaneously.', in Proceedings of IJCAI '16,
pp. 3662 -- 3669, (2016).
[3] R. Andersen, C. Borgs, J. Chayes, U. Feige, A. Flaxman, A. Kalai,
V. Mirrokni, and M. Tennenholtz, 'Trust-based recommendation sys-
tems: an axiomatic approach', in Proceedings of WWW '08, pp. 199 --
208, (2008).
[4] Sanjeev Arora and Boaz Barak, Computational complexity: a modern
approach, Cambridge University Press, 2009.
[5] R. Baird, K. Migiro, D. Nutt, A. Kwatra, S. Wilson, J. Melby,
A. Pendleton, M. Rodgers, and J. Davison. Human tide: the real mi-
gration crisis, 2007.
[6] World Bank.
Identification for development (ID4D) global dataset,
2018.
[7] M. Borge, E. Kokoris-Kogias, P. Jovanovic, L. Gasser, N. Gailly, and
B. Ford, 'Proof-of-personhood: Redemocratizing permissionless cryp-
tocurrencies', in Proceedings of EuroS&PW '17, pp. 23 -- 26, (2017).
[8] Maria Borge, Eleftherios Kokoris-Kogias, Philipp Jovanovic, Linus
Gasser, Nicolas Gailly, and Bryan Ford, 'Proof-of-personhood: Re-
democratizing permissionless cryptocurrencies', Proceedings of Eu-
roS&PW '17, 23 -- 26, (2017).
Jeff Cheeger, 'A lower bound for the smallest eigenvalue of the lapla-
cian', in Proceedings of the Princeton conference in honor of Professor
S. Bochner, (1969).
[9]
[10] V. Conitzer, N. Immorlica, J. Letchford, K. Munagala, and L. Wag-
man, 'False-name-proofness in social networks', in Proceedings of the
6th International Workshop on Internet and Network Economics (WINE
'10), pp. 209 -- 221, (2010).
[11] V. Conitzer and M. Yokoo, 'Using mechanism design to prevent false-
name manipulations', AI magazine, 31(4), 65 -- 78, (2010).
[12] Vincent Conitzer, 'Using a memory test to limit a user to one account',
in Proceedings of AMEC '08, pp. 60 -- 72, (2008).
[13] D. Longley C. Allen R. Grant M. Sabadello D. Reed, M. Sporny, 'De-
centralized identifiers (DIDs) v0. 12 -- data model and syntaxes for de-
centralized identifiers (DIDs)', Draft Community Group Report, 29,
(2018).
[14] The Global Identity Foundation. Global identity challenges, pitfalls and
solutions, 2014.
[15] Gottlob Frege, 'On sense and reference', oversatt av Max Black, i J.
Guit´errez-Rexach (red.): Semantics: Crictical concepts in linguistics,
1, 7 -- 25, (2003).
[16] Brian Fung and Ahiza Garcia, 'Facebook has shut down 5.4 billion fake
accounts this year', CNN, (November 2019).
[17] Charles Geisler and Ben Currens, 'Impediments to inland resettlement
under conditions of accelerated sea level rise', Land Use Policy, 66,
322 -- 330, (2017).
[18] Shlomo Hoory, Nathan Linial, and Avi Wigderson, 'Expander graphs
and their applications', Bulletin of the American Mathematical Society,
43(4), 439 -- 561, (2006).
[19] Nir Kshetri and Jeffrey Voas, 'Blockchain in developing countries', It
Professional, 20(2), 11 -- 14, (2018).
[20] Maciek Laskus. Decentralized identity trilemma, 2018. Available at
http://maciek.blog/dit/?cookie-state-change=1574327093444.
[21] Niall McCarthy, 'Facebook deleted more than 2 billion fake accounts
in the first quarter of the year', Forbes, (May 2019).
[22] A. Muhle, A. Gruner, T. Gayvoronskaya, and C. Meinel, 'A survey on
essential components of a self-sovereign identity', Computer Science
Review, 30, 80 -- 86, (2018).
[23] Government of India. Home - unique identification authority of india,
2018. Available at https://uidai.gov.in.
[24] European Parliament. General data protection regulation (gdpr).
[25] European Parliament. Blockchain and the general data protection regu-
lation: Can distributed ledgers be squared with european data protection
law?, 2019.
[26] Ouri Poupko, Gal Shahaf, Ehud Shapiro, and Nimrod Talmon, 'Sybil-
resilient conductance-based community growth', in Proceedings of
CSR '19, (2019). A preliminary version appears in https://arxiv.org/
abs/1901.00752.
[27] R. L. Rivest, A. Shamir, and L. Adleman, 'A method for obtaining dig-
ital signatures and public-key cryptosystems', Communications of the
ACM, 21(2), 120 -- 126, (1978).
[28]
J. R. Searle, S. Willis, and D. Vanderveken, Foundations of illocution-
ary logic, CUP Archive, 1985.
[29] Sven Seuken and David C Parkes, 'Sybil-proof accounting mechanisms
with transitive trust', in Proceedings of the 2014 international confer-
ence on Autonomous agents and multi-agent systems, pp. 205 -- 212. In-
ternational Foundation for Autonomous Agents and Multiagent Sys-
tems, (2014).
[30] G. Shahaf, E. Shapiro, and N. Talmon, 'Sybil-resilient reality-aware
social choice', in IJCAI '19, pp. 572 -- 579, (2019).
[31] E. Shapiro, 'Point: foundations of e-democracy', Communications of
the ACM, 61(8), 31 -- 34, (2018).
[32] Susan Staats, Alfonso Sintjago, and Renata Fitzpatrick, 'Kiva mi-
croloans in a learning community: An assignment for interdisciplinary
synthesis', Innovative Higher Education, 38(3), 173 -- 187, (2013).
[33] A. Tobin and D. Reed, 'The inevitable rise of self-sovereign identity',
The Sovrin Foundation, 29, (2016).
[34] W3C. Verifiable credentials data model 1.0, 2019.
[35] B. Waggoner, L. Xia, and V. Conitzer, 'Evaluating resistance to false-
name manipulations in elections.', in Proceedings of the 26rd AAAI
Conference on Artificial Intelligence (AAAI '12), (2012).
[36] L. Wagman and V. Conitzer, 'Optimal false-name-proof voting rules
with costly voting.', in Proceedings of the 22st AAAI Conference on
Artificial Intelligence (AAAI '08), pp. 190 -- 195, (2008).
Supplementary Material
Graph notations and terminology
We provide some more detailed notations regarding graphs and con-
ductance.
Let G = (V, E) be an undirected graph. The degree of a vertex
x ∈ V is deg(x) := {y ∈ V (x, y) ∈ E}. The volume of a
(cid:80)
given subset A ⊆ V is the sum of degrees of its vertices, vol(A) :=
in GA, and by volA(B) := (cid:80)
x∈A deg(x). Additionally, we denote the subgraph induced on the
set of vertices A as GA, by degA(x) the degree of vertex x ∈ A
x∈B degA(x) the volume of a set
B ⊆ A in GA. Given two disjoint subsets A, B ⊆ V , the size of the
cut between A and B is denoted by
e(A, B) = {(x, y) ∈ E x ∈ A, y ∈ B} .
Connectivity measures. In the following, we define two fundamen-
tal notions of graph connectivity that play a substantial role in safe
community growth.
Definition 11 (Combinatorial Conductance) Let G = (V, E) be a
graph. The conductance of G is defined by:
Φ(G) = min
∅(cid:54)=A⊂V
e(A, Ac)
min{vol(A), vol(Ac)} .
(Ac := V \ A is the complement of A.)
The following definition is an algebraic measure for connectivity.
Definition 12 (Algebraic conductance) Let G be a graph, and let
λn ≤ λn−1 ≤ ... ≤ λ2 ≤ λ1 be the eigenvalues of its random walk
matrix. Then, G is a said to be a λ-expander if its generalized second
eigenvalue λ(G) := maxi(cid:54)=1 λi satisfies λ(G) ≤ λ.
We note that the notions of conductance and algebraic conduc-
tance are tightly related via the celebrated Cheeger inequality [9].
We refer the reader to the text of Hoory et al. [18] for through ex-
position and elaborate discussion regarding conductance and graph
expansion.
Proof of Theorem 2
Theorem 2 follows by induction from the following Lemma:
Lemma 1 Let G = (V, E) be a surety graph with A ⊆ A(cid:48) ⊆ V ,
and set α, β, γ, δ ∈ [0, 1] and d > 0.
or equivalently
e(A(cid:48) ∩ H, A(cid:48) ∩ B)
dγA(cid:48) ∩ B
Assumptions (1) and (4) imply
≥ 1 − β
β
.
(5)
e(A(cid:48) ∩ H, A(cid:48) ∩ B)
dA(cid:48) ∩ H
≤ e(A(cid:48) ∩ H, A(cid:48) ∩ B)
volA(cid:48) (A(cid:48) ∩ H)
≤ γ ,
or equivalently
(cid:48) ∩ H ≥ e(A(cid:48) ∩ H, A(cid:48) ∩ B)
A
dγ
.
(6)
Combining Inequalities 5, 6 we get:
A(cid:48)
A(cid:48) ∩ B =
A(cid:48) ∩ B
A(cid:48) ∩ H + A(cid:48) ∩ B
≥ e(A(cid:48) ∩ H, A(cid:48) ∩ B)
(cid:18) 1 − β
dγA(cid:48) ∩ B
(cid:19)
+ 1 =
>
1
β
,
β
+ 1
where the first equality holds as A = (A ∩ H) (cid:93) (A ∩ B), the
second inequality stems from Equation 6 and the third inequality
stems from Equation 5. Flipping the nominator and the denominator
(cid:3)
then gives β(A(cid:48)) :=
A(cid:48)∩B
A(cid:48) < β.
Assume:
1. deg(v) ≤ d for all v ∈ A(cid:48).
2. Every a ∈ A(cid:48) satisfies {x∈A(cid:48) (a,x)∈E}
[The graph has a bounded degree].
≥ α.
d
[Sufficiently many edges are within members of A(cid:48)]
A ≤ β.
A∩B
3.
[Byzantine penetration to the initial community is bounded]
4. e(A(cid:48)∩H,A(cid:48)∩B)
volA(cid:48) (A(cid:48)∩H) ≤ γ.
[the edges between harmless and byzantine identifiers are rela-
tively scarce]
5. A(cid:48) \ A ≤ δA, with β + δ ≤ 1
2 .
[Community growth is bounded]
α ·(cid:16) 1−β
β
(cid:17)
.
6. Φ(GA(cid:48) ) > γ
[the conductance within A(cid:48) is sufficiently high]
Then, A(cid:48)∩B
A(cid:48) ≤ β.
Proof. We first note that due to A ⊆ A(cid:48), and assumptions (3), (5),
we have
(cid:48) \ A
A
(cid:48) ∩ B ≤ A ∩ B + A
≤ βA + δA
≤ A
A(cid:48)
2
2
As V = B (cid:93) H, it follows that
<
.
We now utilize assumption (1):
volA(cid:48) (A
(cid:48) ∩ B) :=
{x ∈ A
(cid:48) (a, x) ∈ E}
αd = αdA
(cid:48) ∩ B.
A
(cid:48) ∩ B < A
(cid:48) ∩ H.
(cid:88)
≥ (cid:88)
a∈A(cid:48)∩B
a∈A(cid:48)∩B
Similarly, we have
volA(cid:48) (A
(cid:48) ∩ H) ≥ αdA
(cid:48) ∩ H.
Inequalities 3 and 1 imply that
(cid:48) ∩ H) ≥ αdA
and together with Inequality 2, we have:
volA(cid:48) (A
(cid:48) ∩ B,
min{vol(A
(cid:48) ∩ B)} ≥ αdA
Now, Inequality 4 and assumption (6) imply that:
(cid:48) ∩ H), vol(A
(cid:48) ∩ B.
e(A(cid:48) ∩ H, A(cid:48) ∩ B)
αdA(cid:48) ∩ B
≥
>
e(A(cid:48) ∩ H, A(cid:48) ∩ B)
(cid:18) 1 − β
min{vol(A(cid:48) ∩ H), vol(A(cid:48) ∩ B)}
γ
α
(cid:19)
β
·
,
(1)
(2)
(3)
(4)
|
1609.05264 | 2 | 1609 | 2017-02-15T00:46:48 | Asynchronous and Dynamic Coverage Control Scheme for Persistent Surveillance Missions | [
"cs.MA",
"cs.RO",
"eess.SY"
] | A decomposition-based coverage control scheme is proposed for multi-agent, persistent surveillance missions operating in a communication-constrained, dynamic environment. The proposed approach decouples high-level task assignment from low-level motion planning in a modular framework. Coverage assignments and surveillance parameters are managed by a central base station, and transmitted to mobile agents via unplanned and asynchronous exchanges. Coverage updates promote load balancing, while maintaining geometric and temporal characteristics that allow effective pairing with generic path planners. Namely, the proposed scheme guarantees that (i) coverage regions are connected and collectively cover the environment, (ii) subregions may only go uncovered for bounded periods of time, (iii) collisions (or sensing overlaps) are inherently avoided, and (iv) under static event likelihoods, the collective coverage regions converge to a Pareto-optimal configuration. This management scheme is then paired with a generic path planner satisfying loose assumptions. The scheme is illustrated through simulated surveillance missions. | cs.MA | cs | Asynchronous and Dynamic Coverage Control
Scheme for Persistent Surveillance Missions
Jeffrey R. Peters, Sean J. Wang, Amit Surana, and Francesco Bullo
1
7
1
0
2
b
e
F
5
1
]
A
M
.
s
c
[
2
v
4
6
2
5
0
.
9
0
6
1
:
v
i
X
r
a
Abstract-A decomposition-based coverage control scheme is
proposed for multi-agent, persistent surveillance missions operat-
ing in a communication-constrained, dynamic environment. The
proposed approach decouples high-level task assignment from
low-level motion planning in a modular framework. Coverage as-
signments and surveillance parameters are managed by a central
base station, and transmitted to mobile agents via unplanned and
asynchronous exchanges. Coverage updates promote load balanc-
ing, while maintaining geometric and temporal characteristics
that allow effective pairing with generic path planners. Namely,
the proposed scheme guarantees that (i) coverage regions are
connected and collectively cover the environment, (ii) subregions
may only go uncovered for bounded periods of time, (iii) collisions
(or sensing overlaps) are inherently avoided, and (iv) under static
event likelihoods, the collective coverage regions converge to a
Pareto-optimal configuration. This management scheme is then
paired with a generic path planner satisfying loose assumptions.
The scheme is illustrated through simulated surveillance missions.
I. INTRODUCTION
A. Decomposition-Based Multi-Agent Surveillance
Modern exploratory and surveillance missions often utilize
autonomous vehicles or sensors to observe and monitor large
geographic areas. Example application domains include search
and rescue [14], environmental monitoring [32], warehouse
logistics [37], and military reconnaissance [7]. Such scenarios
require robust and flexible tools for autonomous coordination.
A number of planning strategies have been studied for single
agents, ranging from simple a priori tour construction [23]
to more complex methods involving Markov chains [34],
optimization [17], or Fourier analysis [16]. Unfortunately, it is
not straightforward to generalize single-agent strategies for use
in multi-agent missions: Naive approaches where each agent
follows an independent policy can result in poor performance
and introduce collision risks, while sophisticated generaliza-
tions often require joint optimizations that are intractable for
even modestly sized problems. Scaling issues are sometimes
This work has been sponsored by the U.S. Army Research Office and the
Regents of the University of California, through Contract Number W911NF-
09-D-0001 for the Institute for Collaborative Biotechnologies, and that the
content of the information does not necessarily reflect the position or the
policy of the Government or the Regents of the University of California, and
no official endorsement should be inferred.
Jeffrey R. Peters and Sean J. Wang are with the Mechanical Engineering
Department and the Center for Control, Dynamical Systems and Computa-
tion, University of California, Santa Barbara [email protected],
[email protected]
Amit Surana is with the Systems Department at United Technologies
Research Center, East Hartford, CT, [email protected]
Francesco Bullo is with the Mechanical Engineering Department and the
Center for Control, Dynamical Systems and Computation, University of
California, Santa Barbara [email protected]
Fig. 1. Difficulties arise when complete or pairwise updates are impossible.
With only single-region updates, two steps are required to move from the
left-most to the right-most configuration. In the top path, blue is updated first,
leaving an uncovered block at the intermediate step. In the bottom path, red
is updated first, leaving a block that belongs to both regions simultaneously.
alleviated through distributed control; however, such setups are
application specific and may require extensive efforts to pose
a mathematical problem that is suitable for use with formal
techniques [19]. In contrast, decomposition-based approaches,
which decouple the assignment and routing problem by first
dividing the workspace among agents, offer a straightforward,
modular framework to reasonably accomplish the desired
goals, despite sacrificing optimality in general.
the result
Communication constraints can, however, make it difficult
to effectively divide a dynamic workspace in real time, while
still ensuring that
is amenable for pairing with
single-agent motion planners. For example, many unmanned
missions require agents to transfer sensor data to a central
base station for analysis. With difficult terrain or hardware
limitations, these sporadic exchanges with the base station
may provide the only means of sharing real-time information
across agents. Applications that operate under this constraint
include underwater gliders that must surface to communicate
with a tower [28], data mules that periodically visit ground
robots [31], and supervisory missions where human operators
analyze data collected by unmanned vehicles [29]. Here,
updated mission information can only be relayed to one agent
at a time, rendering traditional dynamic partitioning schemes,
which rely on complete or pairwise coverage updates, impossi-
ble. As such, designers are forced to either allow for overlap in
coverage assignments, or sacrifice complete coverage (Fig. 1).
This work presents a decomposition-based surveillance
framework that operates under asynchronous one-to-base sta-
tion communication, a protocol in which data is transferred
solely via sporadic exchanges between a central base station
and autonomous agents [26]. The complete scheme consists
of two components: a dynamic partitioning component and
a single-agent routing component. The work herein focuses
primarily on developing the partitioning component, which,
in addition to dividing the workload among the agents, ma-
nipulates local variables in order to allow for effective pairing
with single-agent route planners. This development naturally
leads to a complete multi-agent framework in which high-level
coverage is coordinated by the base station, while individual
trajectories are generated independently via on-board planners.
B. Related Literature
Research relating to multi-agent coverage control is vast.
Typical strategies involve optimization [27], auctions [10],
biological meta-heuristics [4], potential fields [30], or space
decomposition [20]. Of particular relevance are multi-agent
persistent surveillance (persistent monitoring) problems, in
which a mobile sensor team is tasked with continual surveil-
lance of a region of interest, requiring each subregion to be
visited multiple (or infinitely many) times with the goal of
minimizing a cost, e.g., the time between visits or the likeli-
hood of detecting stochastic events [19]. Persistent surveillance
is a generalization of patrolling, where closed tours are sought
for the purpose of protecting or supervising an environment.
Patrolling has been studied extensively, with most current
solutions being based on operations research algorithms, non-
learning multi-agent systems, and multi-agent learning [1]. Pa-
trolling formulations are often one-dimensional and solutions
usually reduce to "back and forth" motions that do not readily
extend to general scenarios (see, e.g. [24]).
The framework proposed herein is based on workspace
decomposition, and uses an allocation strategy to reduce the
multi-agent problem into a set of single-agent problems.
This approach is common in multi-agent systems due to its
simplicity and scalability [19]. For persistent surveillance of
planar regions, decomposition-based approaches consist of two
primary components: partitioning and single-agent routing.
The most common approaches to optimal partitioning are
based on Voronoi partitions [21]. Effective schemes exist for
constructing centroidal Voronoi, equitable, or other types of
optimal partitions under various communication, sensing, and
workload constraints [5], [6], [25]. These strategies usually re-
quire convex environments. More general workspaces are usu-
ally addressed by discretizing the environment and considering
the resulting graph, on which any number of graph partitioning
schemes can be used [9]. In robotic contexts, the discrete
partitioning problem is often considered under communication
constraints, e.g., pairwise gossip [8] or asynchronous one-to-
base station communication [26]. The proposed partitioning
scheme most closely mirrors [26]; however, our approach em-
ploys additional logic to ensure the resultant coverage regions
are amenable for use in decomposition-based surveillance.
for persistent
Single-agent path planners
surveillance
most commonly operate over discrete environments,
i.e.,
graphs [13], [35], and solutions to classical problems, e.g., the
Traveling Salesperson Problem [11], may suffice. Stochasticity
can be introduced using tools such as Markov chains [34].
Persistent surveillance strategies for non-discrete regions (in
particular, open subsets of Euclidean space), are less common.
Here, strategies include the a priori construction of motion
routines [22], the adaptation of static coverage strategies [33],
the use of random fields [12], and spectral decomposition [16].
The modular framework herein does not require any particular
2
single-agent route planner, but rather can incorporate any
planner satisfying a mild assumption set (see Section V).
Despite the vast research on both topics individually, re-
markably few papers explicitly consider the implications of
combining geometric partitioning with continuous routing in
the context of multi-agent persistent surveillance. Research
that does exist is mostly preliminary, considering ideal com-
munication and employing simplistic methods. For example,
the authors of [2] employ a sweeping algorithm for partitioning
and guide vehicle motion via lawn-mower pattens. The authors
of [20] use rectangular partitions, while employing a reactive
routing policy which, in ideal cases, reduces to spiral search
patterns. The work in [18] uses slightly more sophisticated par-
titioning in tandem with lawn-mower motion trajectories. Also
relevant is [36], where partitions are based on the statistical
expectation of target presence; however, ideal communication
is assumed. Other works, e.g. [3], employ decomposition-
based structures, but focus on task-assignment with no detailed
treatment of the combined assignment/routing protocol.
C. Contributions
This work presents a decomposition-based, multi-agent cov-
erage control framework for persistent surveillance, which
requires only sporadic, unscheduled exchanges between agents
and a central base station. In particular, we focus on develop-
ing a sophisticated partitioning and coordination scheme that
is designed for pairing with generic single-agent trajectory
planners within a modular framework. Our setup encompasses
realistic constraints including restrictive communication, dy-
namic environments, and non-parametric event likelihoods.
Specifically, we develop a dynamic partitioning scheme
that assumes only asynchronous, one-to-base station commu-
nication. This algorithm governs region assignments, while
also introducing timing variables and manipulating the agents'
high-level surveillance parameters. We prove that our parti-
tioning strategy has properties that make it amenable for use
in a decomposition-based surveillance scheme: the produced
coverage regions collectively form a connected m-covering,
local likelihood functions have disjoint support, no subregion
remains uncovered indefinitely, coverage regions evolve at a
time-scale that allows for appropriate agent reactions, among
others. For static likelihoods under mild assumptions, we show
that the set of coverage regions and associated generators
converges to a Pareto optimal partition in finite time. We
combine our partitioning scheme with a generic single-agent
trajectory, and show that this combination guarantees collision
avoidance (no sensing overlap) when the trajectory planner
obeys natural restrictions. We illustrate our framework through
numerical examples.
For clarity and readability in what follows, we postpone all
Theorem proofs until the Appendix.
II. MISSION OBJECTIVES AND SOLUTION APPROACH
A team of m mobile agents, each equipped with an on-
board sensor,
is tasked with endlessly monitoring a non-
trivial, planar, large region of interest. The primary goal of
the mission is to collect sensor data about some key dynamic
3
load among the agents. During the exchange, the base station
also updates timing variables and governs the local likelihood
functions maintained by each agent so that, if no agent were
ever located outside of the support of its local likelihood at
any time, then the agents (sensors) would never be colocated.
Timing considerations also guarantee that 1) no region remains
uncovered, i.e., outside of the allowable area to which the
agents can travel, for an unbounded amount of time, and 2)
ample time is allotted for agents to vacate regions that are re-
assigned before the newly assigned agent takes over. Once the
update is complete, the link with the base is terminated and
the agent follows the trajectory found via its onboard planner.
III. PROBLEM SETUP
The base station, as well as each agent, has its own local
processor. "Global" information is stored by the base station,
while each agent only stores information pertinent to itself.
Convention 1. The subscripts i, j, or (cid:96) denote an entity
relevant to agent i, j, or (cid:96), resp. When attached to a set, the
subscripts i, j, (cid:96) indicate the component relevant to agent i, j
or (cid:96), resp. The superscript 'A' indicates an entity that is stored
on the agent's local processor.
A summary of the variables stored by relevant entities is
shown in Table I. We expand on these, and define other
relevant mathematical constructs here.
A. Agent Dynamics
Each agent (sensor) i is modeled as a point mass that moves
with maximum possible speed si > 0. Define s := {si}m
i=1.
B. Communication Protocol
There is a central base station with which the agents
communicate. We assume the following:
1) each agent can identify itself to the base station and
transmit sensor data,
2) there is a lower bound ∆ > 0 on the time between any
two successive exchanges involving the base station, and
3) there is an upper bound ∆ > 0 on the time between any
single agent's successive exchanges with the base.
Note that 2) implies that no two agents can communicate with
the base station simultaneously1. These communication spec-
ifications collectively define the one-to-base station protocol.
C. Surveillance Region
Consider the surveillance region as a finite grid of disjoint,
non-empty, simply-connected subregions. We represent
the
environment as a weighted graph G(Q) := (Q,E), where Q
is the set of vertices (each representative of a unique grid
element), and E is the edge set comprised of undirected,
weighted edges {k1, k2} spanning vertices representing adja-
cent regions2. The weight assigned to {k1, k2} is the distance
loss
between the appropriate regions. We assume without
1This constraint also prevents zeno behavior in the communication times
2Travel between the regions without entering other subregions is possible
Fig. 2. An example illustrating the proposed strategy. The dynamic partition-
ing component (executed by the base station) manages coverage regions and
introduces logic to prevent collisions, while the routing component (executed
onboard each agent) governs agent motion.
event or characteristic, e.g., an intruder. Data collected by
the agents is periodically relayed to a central base station
for analysis. Agents must move within the region to obtain
complete sensory information. Ideally, the agent motion/data
collection strategy should be coordinated so that:
1) the workload is balanced across agents,
2) no subregion goes unobserved indefinitely,
3) agents never collide (have sensor overlap), and
4) the search is biased toward regions of greater interest.
To achieve these goals, we adopt a decomposition-based
approach in which each agent's motion is restricted to lie
within a dynamically assigned coverage region.
The analysis herein focuses primarily on developing an
algorithmic framework for partitioning that defines, maintains,
and updates the agents' coverage assignments in a manner that
can be effectively paired with a typical single-agent route plan-
ner. More specifically, we seek a scheme to assign coverage
responsibilities, as well as provide high-level restrictions on
agent motion, so that lower-level, on-board trajectory planners
can operate independently within the proposed boundaries
to achieve global surveillance goals. In addition, we seek
algorithms that do not require peer-to-peer communication or
require the agents to be in constant contact with the base
station. As such, we assume only asynchronous, one-to-base
station communication (data transfer) [26], i.e., agents are
only assumed to sporadically communicate with a central base
station, subject to an upper bound on inter-communication
times, which are not specified a priori.
Broadly, our proposed solution operates as follows (see
Fig. 2). During each exchange with an agent,
the base-
station calculates the new coverage region solely for the
communicating agent, and transmits the update. Updates are
performed so as to guarantee connected coverage regions,
and, in the limit for static likelihoods, distribute the coverage
•Communication link is established•Sensor data sent to base station.•Base station calculates update•Agent receives updated region/variables•Agents follow continuous time surveillance within assigned regions.•Once waiting period is up, agents can enter any part of their assigned region•Agent waits a pre-specified time to enter newly assigned regions•During waiting period, other agent communicates with base, vacates region it no-longer owns. of generality that Q ⊂ N. Environmental information, e.g.,
terrain, is largely unknown at the mission onset; however, the
location of obstacles and prohibited areas, is known a priori
and is not included in the graphical representation G(Q).
Consider Q(cid:48) ⊆ Q. A vertex k1 ∈ Q is adjacent to Q(cid:48)
if k1 /∈ Q(cid:48) and there exists {k1, k2} ∈ E with k2 ∈ Q(cid:48).
Define G(Q(cid:48)) := (Q(cid:48),E(cid:48)), where E(cid:48) := {{k1, k2} ∈ Ek1, k2 ∈
Q(cid:48)}. A path on G(Q(cid:48)) between k1, kn ∈ Q(cid:48) is a sequence
(k1, k2, . . . , kn), where k1, k2, . . . , kn ∈ Q(cid:48) and {kr, kr+1} ∈
E(cid:48) for r ∈ {1, . . . , n − 1}. We say Q(cid:48) is connected if a path
exists in G(Q(cid:48)) between any k1, k2 ∈ Q(cid:48). Let dQ(cid:48) : Q(cid:48)×Q(cid:48) →
R≥0∪{∞} be the standard distance on G(Q(cid:48)), i.e., the length
of a shortest weighted path in G(Q(cid:48)) (if none exists, dQ(cid:48) takes
value ∞). Notice dQ(cid:48)(k1, k2) ≤ dQ(k1, k2) for any k1, k2 ∈
Q(cid:48). Also let dQ(cid:48) denote the map dQ(cid:48) : Q(cid:48)×2Q(cid:48) → R≥0∪{∞},
where dQ(cid:48)(k, Q(cid:48)(cid:48)) is the length of a shortest weighted path in
G(Q(cid:48)) between k and any vertex in Q(cid:48)(cid:48).
D. Coverage Regions
ing 1)(cid:83)m
(cid:84) Pj = ∅, if
An m-covering of Q is a family P = {Pi ⊆ Q}m
i=1 satisfy-
i=1 Pi = Q, and 2) Pi (cid:54)= ∅ for all i. Define Covm(Q)
as the set of all possible m-coverings of Q. An m-partition
of Q is an m-covering that also satisfies 3) Pi
i (cid:54)= j. An m-covering or m-partition P is connected if each
Pi is connected. In what follows, the base station maintains
an m-covering P of Q, and surveillance responsibilities are
assigned by pairing each agent i with Pi ∈ P (called agent
i's coverage region). Each agent maintains a copy P A
i of Pi.
The base station also stores a set c := {ci ∈ Q}m
i=1 (ci is the
generator of Pi), and each agent i maintains a copy cA
i of ci.
E. Identifiers, Timers, and Auxiliary Variables
i
i
}m
i=1, where P ID
k=1, and let P ID := {P ID
The proposed algorithms introduce logic and timing con-
siderations to ensure an effective overall framework. To each
k ∈ Q, assign an identifier IDk ∈ {1, . . . , m}. Define
ID := {IDk}Q
:=
{k ∈ Q IDk = i}. Notice P ID is an m-partition of Q. For
each agent i, define a timer Ti having dynamics Ti = −1
(cid:54)= 0, and Ti = 0 otherwise. The set T := {Ti}m
if Ti
i=1
is maintained by the base station. Similarly, each agent i
maintains a local timing variable τ A
i plays
is constant unless
a similar role to the timer Ti, note that τ A
i
explicitly changed by the algorithm, while Ti has autonomous
dynamics. Next, the base station maintains a set ω := {ωi}m
i=1,
where ωi is the time of agent i's last exchange with the base
i of ωi. Finally, each
station. Each agent maintains a copy ωA
⊆ P A
agent stores locally a subset P A,pd
i which, loosely,
collects vertices that have recently been added to P A
i .
i . Even though τ A
i
F. Likelihood Functions
A real-time estimate of the likelihood of events of interest
occurring within any subset of the surveillance region is
maintained by the base station in the form of a time-varying
probability mass function3 Φ : Q × R≥0 → R≥0. Since it is
3For any t, we have(cid:80)
k∈Q Φ(k, t) = 1
TABLE I
STORAGE SUMMARY
Stored by Base Station
Variable
G(Q)
P
c
ID
T
ω
Φ
Variable
G(Q)
P A
i
cA
i
P A,pd
i
τ A
i
ωA
i
ΦA
i
(cid:17)
(cid:17)
(cid:16)
T (t) ∈ Rm≥0
Description
Graphical representation of the environment
m-covering of Q (P ∈ Covm(Q))
Set of generators (c ∈ Qm)
Set of identifiers
Set of Timers
Set of most recent communication times
(cid:16)ID ∈ {1, . . . , m}Q(cid:17)
(cid:16)
(cid:1)
ω ∈ Rm≥0
Description
Graphical representation of the environment
Global likelihood function(cid:0)Φ : Q × R≥0 → R≥0
Coverage region(cid:0)P A
i ⊂ Q(cid:1)
i ∈ Q(cid:1)
Coverage region generator(cid:0)cA
(cid:17)
(cid:16)
i ∈ R(cid:1)
Local timing parameter(cid:0)τ A
Agent i's most recent communication time(cid:0)ωA
(cid:1)
Local likelihood function(cid:0)ΦA
i ∈ R≥0
Set of "recently added" vertices
: Q × R≥0 → R≥0
Stored by Agent i
⊆ P A
P A,pd
i
i
i
4
(cid:1)
difficult to precisely define load-balancing requirements over
arbitrarily time-varying environments, our analysis focuses on
achieving optimal performance in the case of a static environ-
ment, i.e., when Φ is time-invariant. We then show through
simulation that our proposed strategy still can be effectively
employed in the case of a quasi-static environment, i.e., when
Φ(k,·) is piecewise constant for any k. This situation arises in
many common scenarios, e.g., when estimates of Φ are only
updated during exchanges between the base station and the
mobile agents, which occur at discrete time-points.
: Q × R≥0 → R≥0
as the function that, loosely, represents the agent's local belief
regarding events of interest occurring in Q. The partitioning
algorithm will manipulate the support of each ΦA
to ensure
i
effective surveillance while also inherently preventing agent
collisions (sensor redundancy). Specifically, define
Define each agent's local likelihood ΦA
i
(cid:16)
if k ∈ P A
i ≥ τ A
i or k /∈ P A,pd
i
i and
t − ωA
(cid:17)
,
(1)
Φ(k, t),
ΦA
i (k, t) =
0,
otherwise.
i will be different4 from Φ.
In general, each function ΦA
Remark 1 (Global Data). If global knowledge of Φ is not
i can alternatively be
available instantaneously to agent i, ΦA
defined by replacing Φ(k, t) in (1) by Φ(k, ωA
i ). All subsequent
theorems hold under this alternative definition.
Remark 2 (Data Storage). The cost of storing a graph as
an adjacency list is O(Q + E). The generator set c, each
element of P , and the identifier set ID are stored as a integral
4Note that ΦA
i need not be normalized and thus may not be a time-varying
probability mass function in a strict sense
vectors. The timer set T and the set ω are are stored as real
vectors. Typically, only the values of Φ at vertices in Q are
needed, so Φ is stored as a time-varying real vector. Thus, the
cost of storage at the base station is O(mQ +E). Similarly,
each agent's local storage cost is O(Q + E).
IV. DYNAMIC COVERAGE UPDATE SCHEME
We adopt following convention for the remaining analysis.
Convention 2. Suppose that:
1) min∅ := max∅ := 0, and
2) when referring to a specific time instant, e.g., the instant
when an update is executed, the notation e+ and e−
refers to the value of the entity e immediately before
and after the instant in question, resp.
A. Additive Subset
We start with a definition.
i
, the additive
(k) ⊆ Q is the largest connected subset satisfying:
(cid:106) P add
Definition 1 (Additive Subset). Given k ∈ P ID
subset P add
1) P ID
2) for any h ∈ P add
a) Tj = 0, and
b)
(k) ∩ Pj, where j (cid:54)= i:
(k), and
dPj (h, cj).
i (k)(h, k) < 1
dP add
1
si
sj
i
i
i
i
The following characterizes well-posedness of Definition 1.
is connected and
(k) exists and is unique for
i
disjoint from (cid:83)
i
i
.
j(cid:54)=i Pj, then P add
Proposition 1 (Well-Posedness). If P ID
any k ∈ P ID
Proof. With the specified conditions, P ID
satisfies conditions 1-3 in Definition 1. Thus, P add
unique, maximally connected superset of P ID
same conditions.
i
i
is connected and
(k) is the
i
satisfying the
Under the conditions of Proposition 1, if h ∈ P add
(k), then
max{Tjj (cid:54)= i, h ∈ Pj} = 0 and there is a path from k to h
in G(P add
(k)) that is shorter than the optimal path spanning
cj and h within G(Pj), for any j (cid:54)= i with h ∈ Pj. Also note
that, by definition, additive subsets are connected.
i
i
B. Primary Update Algorithm
Define H : Qm × Covm(Q) × R≥0 → R≥0 ∪ {∞} by
H(c, P, t) =
dPi(k, ci)k ∈ Pi
(cid:88)
min
Φ(k, t),
(cid:27)
(cid:26) 1
si
k∈Q
where Φ is the global likelihood. If (i) each agent is solely
responsible for events within its own coverage region and (ii)
events occur proportionally to Φ, then H(c, P, t) is understood
as the expected time required for an appropriate agent to reach
a randomly occurring event from its region generator at time
t. Algorithm 1 defines the operations performed by the base
station when agent i communicates at time t0. Here, the input
∆H > 0 is a constant mission-specific parameter.
Consider the following initialization assumptions.
5
Algorithm 1: Timed One-To-Base Station Update
Data: t0, P , c, Φ, ω, ∆, ∆H, T , ID, s
Result: P , c, P A
i , ΦA
begin
i , P A,pd
i , ω, ωA
, T , τ A
i , cA
i
i , ID
i
i = P test
i = P ID
Initialize P ∗ = P test = P , c∗ = ctest = c
Set P ∗
if Ti > 0 and P ∗
Set τ A
i = τ A
for k ∈ P ID
i
Set P test
if H(ctest, P test, t) < H(c∗, P ∗, t) then
i = Pi then
i − t0 + ωi and ωA
i = P add
(k) and ctest
Set P ∗ = P test, c∗ = ctest
i = ωi = t0
i = k
else
do
i
i
i \P ID
Set P A,pd
Call Alg. 2 and obtain output ΦA
i = c∗
Set Pi = P A
i , ci = cA
for k ∈ Pi do Set IDk = i
= P ∗
i = P ∗
i
i , ω, T, τ A
i
i , ωA
i = ωi
return P , c, P A
i , cA
i , P A,pd
i
, T , τ A
i , ΦA
i , ω, ωA
i , ID
, Φ, ω, ∆, ∆H, T , s
(cid:111)
Algorithm 2: Timer Update
Data: t0, P , P ∗, c∗, P A,pd
Result: ΦA
begin
i , ω, T, τ A
i
i
(cid:110) 1
si
:= max
dPi(k, P ∗
i \P A,pd
∆Bf
i
for Each j (cid:54)= i satisfying Pj ∩ P ∗
dPj (k, Pj\P ∗
(cid:110) 1
∆Bf
j
Set Tj = ωj + ∆ − t0
:= max
sj
i
i
)k ∈ Pi\P ∗
(cid:54)= ∅ do
i )k ∈ Pj ∩ P ∗
i
i
(cid:111)
j − t0}
{ωj + ∆ + ∆Bf
i }
max, ∆Bf
i = ∆Bf
max, ωi = t0
max
j(cid:54)=i,Pj∩P ∗
i (cid:54)=∅
max = max{∆Bf
max + ∆H, τ A
i , according to (1) with updated variables
Find ∆Bf
max =
Redefine ∆Bf
Set Ti = ∆Bf
Construct ΦA
return ΦA
i , ω, T, τ A
i
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
Assumption 1 (Initialization). The following properties are
satisfied when t = 0:
1) P is a connected m-partition of Q,
2) P = P A = P ID, and
3) for all i ∈ {1, . . . , m},
i ∈ P A
i ,
= ∅,
i
a) ci = cA
b) P A,pd
c) Ti = ωi = ωA
d) τ A
e) ΦA
i = 0,
i = −∆H, and
i (·, 0) is defined according to Eq. 1.
Notice 1) and 3a) together imply that ci (cid:54)= cj for any j (cid:54)= i.
Our first result guarantees that the use of Algorithm 1 for
coverage region updates is well-posed.
Theorem 1 (Well-Posedness). Under Assumption 1, an update
scheme in which, upon each exchange, the base station and
the communicating agent update their respective variables via
Algorithm 1 is well-posed. That is, the operations required by
Algorithm 1 are well-posed at the time of execution.
i }m
Performing of updates with Algorithm 1 does not guarantee
that individual coverage regions (elements of P ), are disjoint
from one another. It does, however, guarantee that the m-
i }m
covering P , the local coverage region set P A := {P A
i=1,
and the local likelihoods {ΦA
i=1 retain properties that are
consistent with the goals of the decomposition-based coverage
framework. Namely, the coverings P and P A maintain connec-
tivity, and each function ΦA
i has support that is disjoint from
that of all other local likelihoods, yet still evolves to provide
reasonable global coverage. The manipulations in Algorithm 2
also ensure that agents are able to "safely" vacate areas that
are re-assigned before newly assigned agents enter. We detail
these and other properties for the remainder of this section.
C. Set Properties
The next result formalizes key set properties.
Theorem 2 (Set Properties). Suppose Assumption 1 holds,
and that, upon each exchange,
the base station and the
communicating agent update their respective variables via
Algorithm 1. Then, the following5 hold at any time t ≥ 0:
1) P ID is a connected m-partition of Q,
2) P is a connected m-covering of Q,
3) ci ∈ Pi and ci (cid:54)= cj for any i (cid:54)= j,
i (·, t)) ⊆ Pi for any i, and
4) supp(ΦA
i=1 supp(ΦA
i (·, t)) = ∅
5) (cid:84)m
Whenever additions are made to an agent's coverage region
during a call to Algorithm 1, the newly added vertices are
not immediately included in the instantaneous support6 of the
agent's local likelihood. As such, if each agent's movement is
restricted to lie within the aforementioned instantaneous sup-
port, then exploration of newly added regions is temporarily
prohibited following an update. This delay allows other agents
to vacate before the newly assigned agent enters. Conversely,
when regions are removed from an agent's coverage region,
Algorithm 1 guarantees a "safe" path, i.e., a path with no
collision risk, exists and persists long enough for the agent to
vacate. Let d := maxi
the prohibited region of agent i at time t, Prohi(t), as the
subset consisting of any newly added vertices that do not yet
i (·, t)), i.e., Prohi(t) := {k ∈ P A
i , t−ωA
belong to supp(ΦA
i <
}, we formalize this discussion as follows.
i and k ∈ P A,pd
τ A
Theorem 3 (Coverage Quality). Suppose Assumption 1 holds,
and that, upon each exchange,
the base station and the
communicating agent updates their respective variables via
Algorithm 1. Then, for any k ∈ Q and any t ≥ 0:
(cid:80){k1,k2}∈E dQ(k1, k2), and define
1
si
i
1) k belongs to at least one agent's coverage region Pi,
2) if k ∈ Prohi(t) for some i, then there exists t0 satisfying
t < t0 < t + ∆ + d such that, for all ¯t ∈ [t0, t0 + ∆H],
the vertex k belongs to the set Pi\Prohi(¯t), and
5supp(f ) denotes the support of a function f.
6The instantaneous support of ΦA
i at time t is defined as supp(ΦA
i (·, t)).
(cid:18)
¯t ∈
t, t +
(cid:16)
1
si
dP
−
i
k, P
ID,−
i
6
(cid:17)(cid:21)
,
3) if k is removed from Pi at time t, then, for all
ID,−
i
we have
a) P
b) there exists a length-minimizing path on G(P −
i )
from k into P
, and all of the vertices along
any such path (except the terminal vertex) belong
to the set ProhID+
⊆ Pi, and
ID,−
i
(¯t)\(cid:83)
j(cid:54)=ID+
k
Pj.
k
Theorems 2 and 3 allow Algorithm 1 to operate within a
decomposition-based framework to provide reasonable cover-
age with inherent collision avoidance. Indeed, when each agent
moves within its coverage region and avoids its prohibited
region, the theorems imply that each agent 1) can visit its
entire coverage region (connectedness), 2) allows adequate
time for other agents to vacate newly assigned regions before
entering, and 3) has a "safe" route into the remaining coverage
region if its current location is removed during an update.
Remark 3 (Local Variables). Theorems 2 and 3 also hold if
we replace P with P A and c with cA in the theorem statement.
Remark 4 (Bounds). Theorem 3 also holds when d is re-
defined as any other upper bound on the subgraph distance
between two arbitrarily chosen vertices of an arbitrarily
chosen connected subgraph of G.
D. Convergence Properties
Our proposed strategy differs from that of [26] mainly due
to the presence of logic (e.g., timer manipulations) to ensure
effective pairing with single-agent trajectory planners. Note
also that H differs from typical partitioning costs, since it
uses subgraph, rather than global graph, distances. As such,
convergence properties of the algorithms herein do not follow
readily from existing results. Consider the following definition.
Definition 2 (Pareto Optimality). The pair (c, P ) is Pareto
optimal at time t if (i) H(c, P, t) ≤ H(¯c, P, t) for any ¯c ∈ Qm,
and (ii) H(c, P, t) ≤ H(c, ¯P , t) for any ¯P ∈ Covm(Q).
The following result characterizes the dynamic evolution of
coverage regions with respect to Pareto optimality.
Theorem 4 (Convergence). Suppose Assumption 1 holds
and that, upon each exchange,
the base station and the
communicating agent updates their respective variables via
Algorithm 1. If Φ is static, i.e., Φ(·, t1) = Φ(·, t2) for all t1, t2,
then the m-covering P and the generators c converge in finite
time to an m-partition P ∗ of Q and a set c∗, resp. The pair
(c∗, P ∗) is Pareto optimal at any time following convergence.
Whenever the event likelihood is static (and Assumption 1
holds), Algorithm 1 causes coverage regions and generators to
collectively converge in finite time to a Pareto optimal config-
uration. That is, the agent's limiting coverage assignments are
"optimal" in that they balance the coverage load in a manner
that directly considers the likelihood Φ. Further, the entire
operation only relies on sporadic and unplanned information
exchange between agents and the base station.
Remark 5 (Voronoi Partitions). It can be shown that Pareto
optimality of (c∗, P ∗) in Theorem 4 implies that, following
convergence, P ∗ is a multiplicatively weighted Voronoi parti-
tion (generated by c∗, weighted by s, subject to density Φ(·, t))
by standard definitions (e.g., [26]). If the centroid set of each
dPi(k, h)Φ(k, t), then P ∗
Pi is defined as arg minh∈Pi
is also centroidal.
(cid:80)
k∈Pi
V. DECOMPOSITION-BASED SURVEILLANCE.
This section pairs the proposed partitioning framework
with a generic, single-vehicle trajectory planner, forming the
complete, decomposition-based coverage control framework.
A. Complete Routing Algorithm
lies entirely within the coverage region P A
Theorems 2 and 3 provide a number of useful insights when
using the dynamic partitioning updates of the previous section.
Namely, Theorem 2 ensures that (i) the instantaneous support
of each ΦA
i , and
i
(ii) the support of two distinct agents' local likelihoods do not
intersect. Theorem 3 states that (i) length of any interval on
which a given vertex is uncovered, i.e., belongs solely to agent
prohibited regions, cannot exceed a finite upper bound, and (ii)
the parameter ∆H is a lower bound on the length of time that
a recently uncovered vertex must remain covered before it can
become uncovered again. Since G(Q) is a discrete represen-
tation of the surveillance region, these facts together suggest
that an intelligent routing scheme that carefully restricts agent
motion according to the instantaneous support of the local
likelihood functions could achieve adequate coverage while
also accomplishing the ancillary goal of collision avoidance.
This motivates the following assumption.
Assumption 2 (Agent Motion). Each agent i has knowledge of
its position at any time t, and its on-board trajectory planner
operates under the following guidelines:
1) generated trajectories obey agent motion constraints,
2) trajectories are constructed incrementally and can be
altered in real-time, and
3) the agent is never directed to leave regions associated
with Pi or enter regions associated with Prohi(t).
Each agent precisely traverses generated trajectories.
Algorithm 3 presents the local protocol for Agent i.
B. Collision Avoidance
Although Assumption 2 locally prevents agents from leav-
ing assigned coverage regions or entering prohibited regions,
dynamic coverage updates can still result in undesirable agent
configurations. Indeed, an agent can be placed outside of
its own coverage region if the vertex corresponding to its
location is abruptly removed during an update. If this happens,
Algorithm 3 constructs a route from the agent's location back
into a region where there is no collision risk. With mild
assumptions, Theorem 3 guarantees that this construction 1) is
7
Algorithm 3: Motion Protocol for Agent i
Data: G(Q), ΦA
i , ωA
i
begin
i , P A,pd
i , P A
i , cA
, τ A
i
while True do
1
2
3
4
5
6
Increment trajectory via on-board planner
Follow trajectory
if Communication with the base station then
i = P A
i
Set P test
Obtain updated variables from base station
if Location lies within the set P A
i \P test
i
then
)
Find a minimum-length path in G(P test
from the currently occupied node into
P A
i
while Agent i is outside P A
i do
i
Follow the aforementioned route
well-defined, and 2) does not present the agent with a transient
collision risk. We formalize this result here.
Theorem 5 (Collision Avoidance). Suppose Assumptions 1
and 2 hold, and that each agent's initial position lies within its
initial coverage region Pi. Suppose further that each agent's
motion is locally governed according to Algorithm 3, where
the update in line 4 is calculated by the base station via
Algorithm 1. If the weight assigned to each edge in G(Q)
is an upper bound on the distance between the associated
regions, then no two agents will ever collide.
By running the schemes of Section IV in conjunction with
a motion planning scheme obeying Assumption 2, we obtain
a complete strategy that 1) only requires weak communication
assumptions, 2) provides dynamic load-balancing, and 3) has
inherent collision avoidance/efficiency properties.
VI. NUMERICAL EXAMPLES
This section presents numerical examples to illustrate the
functionality of the decomposition-based routing scheme. In
all examples, high-level coverage assignment updates are
performed by the base station via Algorithm 1 during each
exchange with an agent, while each agent's local processor
runs the motion protocol in Algorithm 3. For incremental
trajectory construction (Algorithm 3, line 1), we implement a
modified version of the Spectral Multiscale Coverage (SMC)
scheme in [15], which creates agent trajectories that mimic
ergodic dynamics while also locally constraining agent motion
to lie within the appropriate sets. This planner satisfies As-
sumption 2. Initial region generators were selected randomly
(enforcing the non-coincidence constraint), and each agent was
initially placed at its region generator. The initial covering P
was created from these generators by calculating a weighted
Voronoi partition. The remaining parameters were chosen
according to Assumption 1. During the simulation, randomly
chosen agents (chosen via random number generator) sporad-
ically communicated with the base station to receive coverage
8
assignment updates. Communication times were randomly
chosen, subject to a maximum inter-communication time ∆.
A. Static Likelihood
The first example is a 4 agent mission, which is executed
to a static,
over a 100 x 100 surveillance region subject
Gaussian likelihood centered at the bottom left corner. For
discretization, the region is divided into 400, 5 x 5 subregions,
and regions are considered adjacent if they share a horizontal
or vertical edge. Each agent has a maximum speed of 1 unit
distance per unit time, and the maximum inter-communication
time is ∆ = 10 time units. Figure 3 shows the evolution of
the coverage regions at various time points for an example
simulation run. Note that Figure 3 only shows each agent i's
active coverage region, i.e., the subset of Pi that does not
intersect its prohibited region Prohi(t). The family of active
coverage regions does not generally form an m-covering of
Q; however, elements of this family are connected and never
intersect as a result of inherent collision avoidance properties.
Figure 4 shows the mean values of the cost function H as
a function of time, calculated over the same 50 simulations
runs. Here, error bars represent
the range of cost values
achieved at select time points. The variance between runs
is due to the stochastic nature of the data-exchange patterns
between the agents and the base station. Notice that
the
cost
is a non-increasing function of time, as predicted in
the proof of Theorem 4, eventually settling as the coverage
regions/generators reach their limiting configuration, e.g., see
Figure 3. These configurations are guaranteed to be Pareto
optimal and, by Remark 5, to form a multiplicatively weighted
Voronoi partition. Since vehicles are assumed to have identical
maximum speeds, we see from the limiting configuration in
Figure 3 that the resultant coverage assignments provide load-
balancing that takes into account the event likelihood. If the
low-level trajectory planner biases trajectories according to the
event likelihood, this results in desirable coverage properties.
Under the modified SMC planner used here, the temporal
distribution of agent locations closely resembles the spatial
likelihood distribution in the limit, as shown in Figure 5.
Fig. 3. Snapshots of a 4 agent surveillance mission, assuming a static
Gaussian likelihood. Each agent's position, past trajectory, and active coverage
region are indicated by the colored triangle, line, and squares, resp..
The left plot in Figure 4 depicts the maximum amount
of time that any individual subregion went uncovered, i.e.
the subregion did not belong to any agent's active covering
Pi\Prohi(t), during each of 50 simulation runs. Here, the
Fig. 4. The maximum amount of time that any subregion went uncovered in
each of 50 simulation runs (left), and the value of the cost H as a function
of time, averaged over the same 50 runs (right).
maximum amount of time that any region went uncovered
was 186 units, though most trials had maximums of less than
75 units. This is well-below the loose bound ∆ + d = 770
predicted by Theorem 3 (see Remark 4). The right plot in
Fig. 5. Comparison between the (static) event likelihood Φ (left), and the
proportion of time that some agent occupied each subregion after significant
time has passed (10000 units) (right).
Further, during the simulation, no two agents ever occupied
the same space due to the careful parameter manipulations em-
ployed by Algorithm 1. Figure 6 illustrates the logic governing
these manipulations through a simplistic example. During the
Fig. 6. Simplified example illustrating how Algorithm 1 manipulates timing
parameters to prevent agent collisions.
first update, the blue agent acquires some of the red agent's
coverage region. Rather than immediately adding these regions
to its active covering, the blue region waits until sufficient
time has passed to guarantee that the red agent has updated
and moved out of the reassigned regions. Under Algorithm 3,
once the red agent communicates with the base, it immediately
vacates the re-assigned regions, after which the blue agent can
add the region to its active covering. This procedure guarantees
that no two agents will never have overlapping active coverings
and thus never collide (Theorem 5). This same logic is results
in inherent collision prevention over more complex scenarios.
B. Quasi-static Likelihood
We now illustrate how the proposed coverage framework
reacts to abrupt changes in the underlying likelihood, i.e.,
t = 0t = 50t = 150t = 250t = 350t = 10000501001502000510152025TimeFrequency01002003004002025303540455055Number of UpdatesCostBlue communicates with baseRed communicates, vacates regionBlue adds new regions when the likelihood is quasi-static. This type of scenario is
common in realistic missions, e.g., when the base-station's
estimate of the underlying likelihood is only re-formulated
if some agent's sensor data indicates a drastic change in the
underlying landscape. For this purpose, we adopt identical
parameters as in the first example, with the exception of the
likelihood Φ, whose spatial distribution abruptly switches at
select time-points. If the switches are sufficiently spaced in
comparison to the rate of convergence,
then the coverage
regions dynamically adjust to an optimal configuration that
is reflective of the current state. For example, Figure 8 shows
the coverage region evolution after the underlying likelihood
undergoes a single switch from the initial to the final density
shown in Figure 7 at time t = 2000.
In contrast, when
Fig. 7. The initial and final likelihood Φ(·, t).
Fig. 8. Coverage regions after the likelihood switches (see Fig. 7)
the underlying likelihood changes faster than the rate of
convergence, coverage regions are constantly in a transient
state. Despite this, the proposed framework still results in
some degree of load-balancing. To illustrate, Figure 9 shows
the value of the cost H during a simulation in which the
underlying likelihood switches at 12 randomly chosen time-
points over a 1000 unit horizon. Each switch re-defined the
spatial
likelihood as a Gaussian distribution centered at a
randomly selected location. Notice that the value of the cost
function monotonically decreases between the abrupt spikes
caused by changes in the underlying likelihood. Despite the
fact that convergence is not reached, coverage regions quickly
shift away from high-cost configurations, as indicated by the
sharp decreases in the cost shortly following each switch.
Fig. 9. Evolution of the cost H using a quasi-static likelihood with 12 random
switches (switch times are indicated by the stars).
9
VII. CONCLUSION
This work presents a modular, decomposition-based, cover-
age control framework for communication-constrained multi-
agent surveillance missions. In particular, our approach uses a
dynamic partitioning strategy to balance the surveillance load
across available agents, requiring only sporadic and unplanned
exchanges between individual agents and a base station. The
partitioning update algorithm also manages high-level timing
and logic parameters to guarantee that the resulting coverage
assignments have geometric and temporal properties that are
amenable for combination with generic single vehicle trajec-
tory planners. Under appropriate assumptions, the proposed
algorithms will produce a set of coverage regions forming a
Pareto optimal partition, while also ensuring collision avoid-
ance and quality of coverage guarantees.
Future work should further relax communication assump-
tions to reflect additional realistic hardware limitations, e.g.,
use of directional antennae for wireless transmission. Other
areas of future research include the addition of peer-to-peer, in
addition to central, communication, performance comparisons
between specific trajectory planners when used within our
framework, e.g., those involving ergodic Markov chains, and
further theoretical characterizations of performance.
APPENDIX: PROOFS
Proposition 2 (Set Membership). Suppose Assumption 1
holds, and that, at the time of each exchange occurring prior
to the fixed time ¯t ≥ 0, required algorithmic constructions are
well-posed so that the base station and the communicating
agent are able to update their respective variables via Algo-
rithm 1. Then, for any k ∈ Q at any time t ≤ ¯t:
1) k ∈ PIDk,
2) k belongs to at most 2 elements of P ,
3) if TIDk = 0, then k /∈ Pj for any j (cid:54)= IDk, and
4) if k ∈ Pj, j (cid:54)= IDk, then Pj ∩ P ID
k = i (cid:54)= j. Algorithm 1 defines P
(cid:96) = ∅ for (cid:96) /∈ {j, IDk}
Proof. Fix ¯t ≥ 0, k ∈ Q. When t = 0, P = P ID is an
m-partition of Q, implying the proposition. Since k is not
removed from PIDk or added to any Pi with i (cid:54)= IDk until
its first reassignment, i.e., when IDk is changed. Thus, the
proposition is true for all t prior to the first reassignment.
Suppose the proposition holds for all t prior to the pth
reassignment, which occurs at t = t0. Suppose that ID−
k = j,
i = P +
ID+
.
ID+
Thus, k ∈ P +
k
i and remains in these sets until another
ID+
k
reassignment. Statement 1 holds, therefore, for all t prior to the
p + 1st reassignment. Now note that, by Alg. 2, reassignment
cannot occur at t0 unless T −
j = 0. By inductive assumption,
statement 3 of the proposition holds when t = t−
0 , implying
for any (cid:96) (cid:54)= j. Upon reassignment, the timers
that k /∈ P −
j + ∆ − t0. Since
Tj, Ti are modified such that T +
1) IDk is unchanged as long as Tj > 0, and 2) agent j
communicates with the base prior to time ω+
j + ∆ (when k
is removed from Pj), we deduce that k solely belongs to Pj
and Pi until the p + 1st reassignment. By the same logic, at
any time t ≥ t+
0 at which Ti = 0 and the p + 1st reassignment
has not yet occurred, k ∈ Pi exclusively (addition to other
ID,+
i
= P +
= P +
(cid:96)
j , T +
i > ω+
Initial LikelihoodFinal Likelihoodt = 2000t = 2100t = 2200t = 400002004006008001000102030405060TimeCost 0 , implying P −
sets in P without reassignment is impossible). We deduce
that statements 2 and 3 hold for any t prior to the p + 1st
reassignment. Now note that, 1) statement 3 holds prior at
time t+
, 2) no agent claims additional
points vertices from P +
j unless Tj = 0, and 3) k cannot be
added to a coverage region without reassignment. As such,
(cid:96) = ∅ for any (cid:96) /∈ {j, IDk = i} prior to the p + 1st
Pj ∩ P ID
reassignment. From persistent communication and the bound
∆, the proposition follows by induction.
ID,−
i
j = P
i
i
i
(c+
(cid:17)
j(cid:54)=i Pj
(cid:83)
i = P add
= P +
ID,+
i
occurs, then P ID
i ∩(cid:16)(cid:83)
Proof of Theorem 1. It suffices to show that Def. 1 is well-
posed (Prop. 1) whenever additive sets are required. We
proceed by induction. When t = 0, P ID = P is a connected
m-partition of Q and thus, for any i, P ID
is disjoint from
j(cid:54)=i Pj. The same holds prior to the first exchange between
the base station and some agent. Thus, by Prop. 1, the first
call to Alg. 1 is well-posed. Now assume that, for all times
t prior to the pth call to Alg. 1, 1) P ID is a connected m-
partition of Q, and 2) if an exchange that requires construction
= ∅. This implies that
of P add
Prop. 2 holds at any time t prior to the p + 1st exchange.
Assume the pth communication occurs when t = t0 and
involves agent i. By the uniqueness of identifiers, P ID,+ is
an m-partition of Q. To show that P ID,+ is connected, first
ID,+
i . Since P +
note that P
i ) (connected
ID,−
i
(connected by inductive
by Def. 1), or P
= P
ID,+
i
follows. Now consider
assumption) connectivity of P
ID,−
(cid:54)= 0,
i
P ID
then P
and
j
ID,+
follows. Suppose T −
connectivity of P
j
is not connected. Then, P +
if it were,
j
k /∈ {i, j}, contradicting
there would exist k ∈ P +
Prop. 2, satement. Thus, there exists k1 ∈ P
such that 1)
k1 /∈ P add
i ), and 2) any length-minimizing path in G(P −
(c+
j )
j contains some k2 ∈ P add
spanning k1 and c+
i (c+
i ).
Select one such path and vertex k2. Assume without loss of
generality that there exists an edge {k1, k2} ∈ E. Def. 1
(cid:96) )(cid:96) (cid:54)= i, k2 ∈
implies 1
si
i } and thus 1
(k1, c+
P +
j ). Since
T −
j = 0 and ID−
i ) <
(cid:96) }, con-
dP +
tradicting k1 /∈ P add
is connected. Invoking
Prop. 2 statement 3, the inductive assumption therefore holds
for all times prior to the p + 1st exchange, thereby implying
well-posedness of the first p + 1 exchanges.
(k2, c+
(k2, c+
s(cid:96)
i ) < 1
dP +
dP +
sj
i ∪{k1}(k1, c+
= j, Prop. 2 implies 1
dP +
si
(cid:96) )(cid:96) (cid:54)= i, k1 ∈ P +
(k1, c+
ID,+
j
, j (cid:54)= i. If T −
j
ID,+
j
i ) < min{ 1
i ∪{k1}(k1, c+
j ) = min{ 1
(c+
is not connected:
j = 0 and P
j with ID+
i ). Thus, P
i ) = P +
ID,+
j
ID,+
j
(k1, c+
= P
dP +
dP +
dP +
(c+
1
sj
k1
s(cid:96)
si
j
i
i
i
(cid:96)
(cid:96)
j
j
i
Proof of Theorem 2.
Statement 1: The proof of Thm. 1 implies the statement.
Statement 2: P is an m-covering of Q since P ID is always
i ⊆ Pi for any
an m-partition of Q (statement 1), and P ID
i (Prop. 2, statement 1). P is connected, since Pi = P ID
(connected by statement 1) immediately following any update,
and Pi is unchanged in between updates.
Statement 3: It suffices to show that IDci = i for any t and
any i: this would imply ci (cid:54)= cj for any i (cid:54)= j, and ci ∈ Pi
(Prop. 2). By assumption, IDci = i for all i at t = 0. The same
holds for any t prior to the first exchange between any agent
and the base station. Suppose IDci = i for all i (thus ci (cid:54)= cj
i
10
i
i
i
i
j
j = ω+
. Since supp(ΦA
j + ∆ < t0 + T +
j (·, t)) = Pj = P A
j ) = 0 for any j, we have c+
j (·, t)). Thus, for all t > t0 +τ +
(c−
j , c−
= j, and induction proves the statement.
for any i (cid:54)= j) prior to the pth exchange. If agent i is the pth
communicating agent, lines 2 and 9 of Alg. 1 imply ID+
= i.
c+
j /∈ P add
i
(c+
Since dP
i ).
−
j
Thus, ID+
c+
j
Statements 4 and 5: Statement 4 follows from (1), noting that
i = Pi. Statement 5 holds by assumption when t = 0.
P A
Let k ∈ Q, and consider times when IDk changes (k is
re-assigned). Since supp(ΦA
for any j
at t = 0, statement 4 implies that, for any t prior to the
(·, t)),
first reassignment, k belongs exclusively to supp(ΦA
IDk
implying statement 5. Suppose statement 5 holds for all t prior
to the pth reassignment (occurring at time t0), and ID−
k = j,
k = i (cid:54)= j. Then, T −
ID+
j = 0 and k belongs exclusively to
0 (Prop. 2). By Alg. 1 and 2, k ∈ P A,pd,+
j when t = t−
P −
i (·, t))
j + ∆ − t0 ≥ τ A,+
and T +
i > ω+
i ≥ τ A,+
least T +
is not redefined for a duration of at
,
i (·, t)) when t ∈ [t+
(1) implies k /∈ supp(ΦA
0 + τ A,+
0 , t+
].
ID,−
Since k is re-assigned when t = t0, k ∈ P +
i \P
i
and
j + ∆ − t0. Agent j will communicate with the base
i
T +
j = ω+
at some time t1 < t0 + T +
i . Thus,
Ti > 0 when t = t1, and k is removed from both Pj and
i and before the p+1st
supp(ΦA
reassignment, k belongs exclusively to supp(φi(·, t)).
Proof of Theorem 3. Thm. 2 implies statement 1.
Statement 2: For any i, 1) Ti = 0 when t = 0, and 2)
dQ(cid:48)(k1, k2) ≤ d for any Q(cid:48) ⊆ Q, k1, k2 ∈ Q(cid:48). Thus, it
1
si
is straightforward to show that, for any i and any T , we have
the bound Ti ≤ ∆ + ∆H + d. We show that, for any i, the
i ≤ Ti− ∆H holds by induction: Ti = 0 and
bound τ A
i = −∆H when t = 0 , so τ A
i ≤ Ti − ∆H.
τ A
i = τ A
The bound similarly holds prior to the first exchange involving
i − t = τ A
i ≤ −∆H ≤ Ti − ∆H
any agent, since τ A
at any such time. Assume the bound holds prior to the pth
update (occuring at t = t0). Consider 2 cases: if agent i is the
i − ∆H;
communicating agent, then τ A,+
if not, then τ A,+
i = T +
implying
− t0 + ωA,+
the desired bound, or 2) T −
=
τ A,−
− t + ωA,−
i + ∆ − t0) −
i − ∆H. This logic extends to all times prior to the
i
∆H = T +
p + 1st exchange and the desired bound follows by induction.
Using the bounds from the previous paragraph, we have
i ≤ t + ∆ + d. Fix t and k ∈ Prohi(t). Then, k ∈
τ A
i + ωA
P A,+
= P +
('+' is
i
with respect to the fixed time t). Further, over the interval
[t, ωA,+
], the vertex k is not re-assigned, Pi is not
is unchanged . Therefore, k /∈ Prohi(t) at
augmented, and τ A
, we have t < t0 ≤ t+
i
time ωA,+
. If t0 := ωA,+
∆ + d. Since Ti ≥ τ A
, k is not re-assigned
, t0 + ∆H] ⊇
during the interval [ωA,+
[t0, t0 + ∆H]. Thus k ∈ Pi\Prohi(·) over the same interval.
Statement 3: Fix t and suppose k ∈ P −
(in this proof,
'+,−' are with respect to t). Then, 1) IDk changed (k was
reassigned) at time t0 < t, 2) agent i communicates with the
base at time t, and 3) no exchanges involving agent i occurred
during the interval [t0, t). Upon reassignment at time t0, Alg. 2
specifies that 1) Ti is reset to value ωA,−
i + ∆− t0, thus P ID
i − t + ωA
and either 1) T −
i = 0 and τ A,+
i − t + ωA
i − t + ωA
i +τ A,+
i + ∆H at time ωA,+
= τ A,−
i ≤ T −
i ] ⊇ [ωA,+
i \P +
i − ∆H = −∆H ≤ (ωA,+
i , k ∈ P A,pd,+
, and t − ωA,+
i − t + ωA
i + τ A,+
i +τ A,+
i + T +
i = τ +
i
< τ A,+
, ωA,+
:= T +
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
is unchanged over the interval [t0, t), 2) k is added to P A,pd
IDk
and 3) τ A
IDk
, TIDk are given values of at least
,
(cid:16)k, P
ID,−
i
(cid:27)
(cid:17) − t0
(cid:26)
ωA,−
i + ∆ +
1
si
i
k
k
k
i
(cid:96)
si
si
−
i
−
i
−
i
−
i
dP
dP
dP
dP
k∈P
−
i \P +
ID,−
i
ID,−
i
ID,−
i
⊆ P ID
ID,−
i
)].
)] ⊇ (t0, t].
) from k into P
i \P +
(k, P
i \P +
when TID+
i ⊆ ProhID+
ID,−
i
(k, P
i \P +
i , and 2) P −
ω := max
,
implying that PIDk, ProhIDk (·) remain unchanged over the
interval (t0, ω] ⊇ (t0, t + 1
(k, P
dP
Since coverage regions are connected and non-empty
= ∅ for any (cid:96) /∈ {i, ID+
i ∩ P ID
k }
(Thm. 2) and P −
on the interval (t0, t] (Prop. 2), 1) there exists a path of
ID,−
length dP
and every ver-
−
i
i
tex along any such path (except the terminal vertex) lyies
within P −
over the interval
(t0, ω] ⊇ [t, t + 1
)]. Since 1) each vertex
belongs to no more than two sets simultaneously (Prop. 2),
2) k ∈ P −
i , and 3) no agent claims any vertex in the set
ProhID+
> 0, vertices along the path (excluding
the terminal vertex) do not belong to Pj with j (cid:54)= IDk over
)]. To complete the proof,
the interval [t, t + 1
(k, P
si
note that Alg. 2 implies T +
), and thus
i > 1
(k, P
dP
−
si
ID,−
i
over [t, t + 1
P
(k, P
−
i
si
i
ID,−
i
Proposition 3 (Cost). Suppose Assumption 1 holds and that,
upon each exchange, the base station and the communicating
agent update their respective variables via Alg. 1. If Φ(·, t1) =
Φ(·, t2) for all t1, t2, then H(c, P ID,·) = H(c, P,·).
Proof. Since Φ is static, H(·,·, t1) = H(·,·, t2) for any t1, t2.
When t = 0, P = P ID and thus H(c, P ID, 0) = H(c, P, 0).
The same is true prior to the first exchange between any
agent and the base. Suppose that immediately prior to the pth
exchange (occurring at t = t0, involving agent i), we have
H(c−, P ID,−, t0) = H(c−, P −, t0). Recall that, for any agent
j, Pj and P ID
coincide immediately following any exchange
involving agent j and, if agent j claims vertices from Pi,
then Alg. 2 guarantees that agent i will communicate with
the base station before any additional vertices are claimed
by other agents. At the time of the pth update, this logic,
ID,−
j = ∅, for
along with Prop. 2,
implies that P
i
ID,−
all j (cid:54)= i. Noting that c+
i ∈ P
, we deduce that any
single k ∈ P
contributes equally to H(c+, P ID,+, t0)
i
and H(c+, P +, t0). If k ∈ P add
, then for any
j
i ) <
dP +
j ) (Def. 1). Therefore, k contributes equivalently
1
sj
to the values of both H(c+, P ID,+, t0) and H(c+, P +, t0).
(cid:54)= ∅. We show
j \P +
Now suppose k ∈ P +
j ∩ P +
(c+
if a length-minimizing
j , k):
that dP
j , k) = dP +
path in G(P +
j ) between c+
and k is also contained in
), then the result is trivial. Suppose that every such
G(P
minimum length path does leave G(P
). From Prop. 2
statement 4, every ¯k ∈ P +
¯k ∈ {i, j}.
Therefore, we assume without loss of generality that k is
to P +
adjacent
is adja-
cent to k and lies along a minimum-length path in G(P +
j )
j and k. Since k ∈ P +
spanning c+
, we must have
k ∈ P add
(c+
i ) as constructed during the update, which implies
(cid:54)= i such that k ∈ P +
dP +
be a vertex that
i \P
i )\P
(c+
j , we have
j must satisfy either ID+
i . Let k ∈ P +
i , where P +
ID,−
i
1
si
∩ P −
ID,−
i
ID,−
i
ID,+
j
ID,+
j
(k, c+
(k, c+
ID,+
j
(c+
j
j
i
i
i
i
j
j
i
11
i
j
j
j
(cid:96)
(cid:96)
i
s(cid:96)
(c+
dP +
dP +
dP +
dP +
j , k) = dP +
(k, c+
(k, c+
(k, c+
ID,+
j
(k, c+
i ∪{k}(k, c+
i ∪{k}(k, c+
(cid:96) }, contradicting k /∈ P add
(cid:96) )(cid:96) (cid:54)= i, ¯k ∈ P +
j ). Since T −
i ) < 1
sj
i ) < min{ 1
dP +
s(cid:96)
i ) < 1
dP +
sj
1
dP +
si
(cid:96) )(cid:96) (cid:54)= i, P +
(c+
(cid:96) } and thus
1
si
j = 0 and ID−
1
k =
si
(k, c+
j, Prop. 2 implies
j ) =
min{ 1
i ) ⊂
(c+
P +
j , k), which, by inductive
i . Thus, dP
assumption, implies that k contributes equally to the value of
both H(c+, P ID,+, t0) and H(c+, P +, t0). We conclude that
H(c+, P ID,+, t0) = H(c+, P +, t0). Since P , P ID, and c are
static between updates, the statement follows by induction.
Proof of Theorem 4. The value of H(c, P, t) is static in
between base station exchanges since P and c do not
change in between updates. Consider an update occurring
at t = t0 involving agent i. Noting Prop. 3, we have
H(c+, P +, t0) ≤ H(c−, P ID,−, t0) = H(c−, P −, t0). Thus,
the value H(c, P, t) is non-increasing as t → ∞. Since
Covm(Q) is finite, there must exist some time t0 after which
the value of H is static. Consider fixed t > t0 at which some
agent i communicates with the base. Since the value of H
does not change during the update, Alg. 1 implies that P ID
and c will be unchanged by the update. It follows that c and
i ⊆ Pi for any
P ID converge in finite time. Further, since P ID
ID,−
i (Prop. 2), we have P
i ) =
i
P +
i . Noting the persistence of communication imposed by ∆,
the same logic implies that after some finite time, P and P ID
are concurrent.
∪ P add
ID,−
i
ID,+
i
(c+
= P
= P
i
i
sj
h∈Pi
h∈Pi
k∈Q(cid:48) 1
si
that (cid:80)
dPi∪{k}(ci, k) ≥ 1
such that (cid:80)
dPi∪Q(cid:48)(ci, k) < (cid:80)
dPi(h, k)Φ(h, t) <(cid:80)
We show that the limiting configuration is Pareto optimal.
Consider t0, such that for all t > t0, c and P are static and P
is an m-partition of Q. Since timers Ti are only reset when P
is altered, we assume without loss of generality that Ti = 0
for all i ∈ {1, . . . , m} at any t > t0. Suppose that agent
i communicates with the base station at time t > t0. With
these assumptions, Alg. 1 implies that there exists no k ∈ Pi
dPi(h, ci)Φ(h, t)
(otherwise the value of H would be lowered by moving ci).
Similarly, for any k ∈ Pj with j (cid:54)= i that is adjacent to
Pi, we must have 1
dPj (cj, k). Indeed,
si
if this were not so, there would exist k ∈ P add
i )\P −
(c+
i ,
contradicting the convergence assumption, since P add
(c+
i ) =
there is no Q(cid:48) ⊂ Q\Pi such
P +
i . As such, for any i,
dPj (cj, k)k ∈
Pj, j (cid:54)= i}, which implies statement (ii) of Def. 2.
Proof of Theorem 5. By Assumption 2, no agent ever leaves
its assigned coverage region or enters its prohibited region,
provided no abrupt changes to these regions occur during
an update. Therefore, if no update ever occurs in which the
vertex corresponding to the communicating agent's location
is removed from the relevant agent's coverage region, then
the statement is immediate. Suppose now that, at some time
t = t0, agent i, whose location is associated with some
k ∈ P A,−
, communicates with the base station and k is
removed, i.e., k /∈ P A,+
. At time t+, agent i executes lines
5 and 6 of Alg. 3. Thm. 3, however, guarantees that 1) there
will exist a path in G(P A,−
ID,−
, 2)
i
P +
that all vertices along this path belong to ProhID+
j(cid:54)=ID+
j
)], and 3)
during the time period (t0, t0 + 1
(k, P
si
) between k and the set P
k∈Q(cid:48) min{ 1
\(cid:83)
ID,−
i
dP A,−
sj
i
i
i
i
k
k
i
ID,−
i
⊆ Pi := P A
k
i
.
until it reaches P A,+
i over the same interval. Therefore, if agent
P
i immediately starts moving along the path, its location will
lie exclusively within ProhID+
It remains to show that no agent i ever enters Prohi(t).
Whenever agent i is executing lines 1− 2 of Alg. 3, it follows
readily that it will not enter Prohi(t). We show that the same
holds when agent i is forced to execute line 5 and 6 of Alg. 3.
Without loss of generality, consider the update at time t0
previously described. Since k is re-assigned prior to the update
i = ∅ (since vertices in Pi cannot be
at time t0, we have Proh−
claimed unless Ti = 0, implying t0 − ωA,−
). Using
Prop. 2, we deduce that T +
> 0 and thus no vertices in
ID+
P A,−
k
, and no vertex on the
belongs to Prohi(t+
constructed path between k and P
0 ).
ID,−
), Prohi(·) remains
Since T +
(k, P
i
unchanged over the interval (t0, t0 + 1
si
can belong to P A,pd,+
ID,−
i
∩ P +
ID+
k
i > τ A,+
> τ A,−
ID,−
i
dP A,−
dP A,−
> 1
si
(k, P
)].
i
i
i
i
i
i
i
REFERENCES
[1] A. Almeida, G. Ramalho, H. Santana, P. Tedesco, T. Menezes, V. Cor-
ruble, and Y. Chevaleyre. Recent advances on multi-agent patrolling.
In Advances in Artificial Intelligence, volume 3171 of Lecture Notes in
Computer Science, pages 474–483. Springer, 2004.
[2] J.F. Araujo, P.B. Sujit, and J.B. Sousa. Multiple UAV area decom-
position and coverage. In Computational Intelligence for Security and
Defense Applications (CISDA), 2013 IEEE Symposium on, pages 30–37,
Singapore, April 2013. IEEE.
[3] B. Bethke, M. Valenti, and J. P. How. UAV task assignment.
IEEE
Robotics & Automation Magazine, 15(1):39–44, 2008.
[4] E. Bonabeau, M. Dorigo, and G. Theraulaz. Swarm Intelligence: From
Natural to Artificial Systems. Number 1. Oxford university press, 1999.
[5] J. Cort´es. Coverage optimization and spatial load balancing by robotic
sensor networks. IEEE Transactions on Automatic Control, 55(3):749–
754, 2010.
[6] J. Cort´es, S. Mart´ınez, T. Karatas, and F. Bullo. Coverage control
IEEE Transactions on Robotics and
for mobile sensing networks.
Automation, 20(2):243–255, 2004.
[7] S. R. Dixon and C. D. Wickens. Automation reliability in unmanned
aerial vehicle control: A reliance-compliance model of automation
dependence in high workload. Human Factors: The Journal of the
Human Factors and Ergonomics Society, 48(3):474–486, 2006.
[8] J. W. Durham, R. Carli, P. Frasca, and F. Bullo. Discrete partitioning and
coverage control for gossiping robots. IEEE Transactions on Robotics,
28(2):364–378, 2012.
[9] P. O. Fjallstrom. Algorithms for graph partitioning: A survey. Linkoping
Electronic Articles in Computer and Information Science, 3(10), 1998.
[10] B. P. Gerkey and M. J. Matari´c. Sold!: Auction methods for multi-
IEEE Transactions on Robotics and Automation,
robot coordination.
18(5):758–768, 2002.
[11] G. Gutin and A. P. Punnen. The Traveling Salesman Problem and Its
Variations. Springer, 2007.
[12] X. Lan and M. Schwager. Planning periodic persistent monitoring
In IEEE
trajectories for sensing robots in Gaussian random fields.
Int. Conf. on Robotics and Automation, pages 2415–2420, Karlsruhe,
Germany, May 2013.
[13] Gilbert Laporte. The vehicle routing problem: An overview of exact and
approximate algorithms. European Journal of Operational Research,
59(3):345–358, 1992.
[14] A. Macwan, G. Nejat, and B. Benhabib. Optimal deployment of robotic
teams for autonomous wilderness search and rescue. In IEEE/RSJ Int.
Conf. on Intelligent Robots & Systems, pages 4544–4549, San Francisco,
CA, USA, September 2011.
[15] G. Mathew and I. Mezic. Spectral multiscale coverage: A uniform
In IEEE Conf. on
coverage algorithm for mobile sensor networks.
Decision and Control, pages 7872–7877, Shanghai, China, December
2009.
[16] George Mathew, Amit Surana, and Igor Mezic. Uniform coverage
control of mobile sensor networks for dynamic target detection.
In
Decision and Control (CDC), 2010 49th IEEE Conference on, pages
7292–7299. IEEE, 2010.
12
[17] N. Mathew, S. L. Smith, and S. L. Waslander. Multirobot rendezvous
IEEE Transactions on
tasks.
planning for recharging in persistent
Robotics, 31(1):128–142, 2015.
[18] I. Maza and A. Ollero. Multiple UAV cooperative searching operation
using polygon area decomposition and efficient coverage algorithms. In
R. Alami, R. Chatila, and H. Asama, editors, Distributed Autonomous
Robotic Systems 6, pages 221–230. Springer, 2007.
[19] N. Nigam. The multiple unmanned air vehicle persistent surveillance
problem: A review. Machines, 2(1):13–72, 2014.
[20] N. Nigam and I. Kroo. Persistent surveillance using multiple unmanned
In Aerospace Conference, 2008 IEEE, pages 1–14, Big
air vehicles.
Sky, MT, USA, March 2008. IEEE.
[21] A. Okabe, B. Boots, K. Sugihara, and S. N. Chiu. Spatial Tessellations:
Concepts and Applications of Voronoi Diagrams. Wiley Series in
Probability and Statistics. John Wiley & Sons, 2 edition, 2000.
[22] J. Ousingsawat and M.G. Earl. Modified lawn-mower search pattern for
areas comprised of weighted regions. In American Control Conference,
pages 918–923. IEEE, 2007.
[23] F. Pasqualetti, J. W. Durham, and F. Bullo. Cooperative patrolling via
weighted tours: Performance analysis and distributed algorithms. IEEE
Transactions on Robotics, 28(5):1181–1188, 2012.
[24] F. Pasqualetti, F. Zanella, J. R. Peters, M. Spindler, R. Carli, and F. Bullo.
Camera network coordination for intruder detection. IEEE Transactions
on Control Systems Technology, 22(5):1669–1683, 2014.
[25] R. Patel, P. Frasca, and F. Bullo. Centroidal area-constrained partitioning
for robotic networks. ASME Journal of Dynamic Systems, Measurement,
and Control, 136(3):031024, 2014.
[26] R. Patel, P. Frasca, J. W. Durham, R. Carli, and F. Bullo. Dynamic
partitioning and coverage control with asynchronous one-to-base-station
IEEE Transactions on Control of Network Systems,
communication.
3(1):24–33, 2016.
[27] J. Peng and S. Akella. Coordinating multiple robots with kinodynamic
constraints along specified paths. The International Journal of Robotics
Research, 24(4):295–310, 2005.
[28] A. Pereira, H. Heidarsson, C. Oberg, D. Caron, B. Jones, and
G. Sukhatme. A communication framework for cost-effective operation
of AUVs in coastal regions. In A. Howard, K. Iagnemma, and A. Kelly,
editors, Field and Service Robotics, volume 62 of Tracts in Advanced
Robotics, pages 433–442. Springer, 2010.
[29] J. Peters, V. Srivastava, G. Taylor, A. Surana, M. P. Eckstein, and
F. Bullo. Mixed human-robot team surveillance: Integrating cognitive
IEEE Control Systems Magazine,
modeling with engineering design.
35(6):57–80, 2015.
[30] J. H. Reif and H. Wang. Social potential fields: a distributed behavioral
Robotics & Autonomous Systems,
control for autonomous robots.
27(3):171–194, 1999.
[31] R. C. Shah, S. Roy, S. Jain, and W. Brunette. Data MULEs: modeling
and analysis of a three-tier architecture for sparse sensor networks. Ad
Hoc Networks, 1(2-3):215–233, 2003.
[32] R. N. Smith, Y. Chao, P. P. Li, D. A. Caron, B. H. Jones, and G. S.
Sukhatme.
Planning and implementing trajectories for autonomous
underwater vehicles to track evolving ocean processes based on predic-
tions from a regional ocean model. International Journal of Robotics
Research, 29(12):1475–1497, 2010.
[33] D. E. Soltero, M. Schwager, and D. Rus. Generating informative paths
for persistent sensing in unknown environments. In IEEE/RSJ Int. Conf.
on Intelligent Robots & Systems, pages 2172–2179, Vilamoura, Portugal,
October 2012.
[34] V. Srivastava, F. Pasqualetti, and F. Bullo. Stochastic surveillance strate-
International Journal of Robotics
gies for spatial quickest detection.
Research, 32(12):1438–1458, 2013.
[35] P. Toth and D. Vigo, editors. The Vehicle Routing Problem. Monographs
on Discrete Mathematics and Applications. SIAM, 2001.
[36] J. Wood and J.k. Hedrick. Space partitioning and classification for multi-
target search and tracking by heterogeneous unmanned aerial system
In Infotech@Aerospace, St. Louis, MO, USA, March 2011.
teams.
AIAA 2011-1443.
[37] P. R. Wurman, R. D'Andrea, and M. Mountz. Coordinating hundreds of
cooperative, autonomous vehicles in warehouses. AI Magazine, 29(1):9–
20, 2008.
|
1802.01208 | 3 | 1802 | 2019-03-26T23:54:14 | Toward a Theory of Markov Influence Systems and their Renormalization | [
"cs.MA",
"math.PR",
"nlin.AO"
] | We introduce the concept of a Markov influence system (MIS) and analyze its dynamics. An MIS models a random walk in a graph whose edges and transition probabilities change endogenously as a function of the current distribution. This article consists of two independent parts: in the first one, we generalize the standard classification of Markov chain states to the time-varying case by showing how to "parse" graph sequences; in the second part, we use this framework to carry out the bifurcation analysis of a few important MIS families. We show that, in general, these systems can be chaotic but that irreducible MIS are almost always asymptotically periodic. We give an example of "hyper-torpid" mixing, where a stationary distribution is reached in super-exponential time, a timescale beyond the reach of any Markov chain. | cs.MA | cs |
Toward a Theory of Markov Influence Systems and their
Renormalization∗
Bernard Chazelle †
Abstract
We introduce the concept of a Markov influence system (MIS) and analyze its dynamics.
An MIS models a random walk in a graph whose edges and transition probabilities change
endogenously as a function of the current distribution. This article consists of two inde-
pendent parts: in the first one, we generalize the standard classification of Markov chain
states to the time-varying case by showing how to "parse" graph sequences; in the second
part, we use this framework to carry out the bifurcation analysis of a few important MIS
families. We show that, in general, these systems can be chaotic but that irreducible MIS
are almost always asymptotically periodic. We give an example of "hyper-torpid" mixing,
where a stationary distribution is reached in super-exponential time, a timescale beyond the
reach of any Markov chain.
Keywords: Random walks on time-varying graphs; Markov influence systems; Graph sequence parsing;
Renormalization; Hyper-torpid mixing; Chaos
1 Introduction
Nonlinear Markov chains are popular probabilistic models in the natural and social sciences.
They are commonly used in interacting particle systems, epidemic models, replicator dynamics,
mean-field games, etc. [5, 11, 15, 17, 20, 23, 31]. They differ from the linear kind by allowing
transition probabilities to vary as a function of the current state distribution. For example, a
traffic network might update its topology and edge transition rates adaptively to alleviate con-
gestion. The systems are Markovian in that the future depends only on the present: in this
work, the present will refer to the current state distribution rather than the single state presently
∗A preliminary version of this work appeared in the Proceedings of the 9th Innovations in Theoretical Com-
puter Science (ITCS), 2018. The Research was sponsored by the Army Research Office and the Defense Advanced
Research Projects Agency and was accomplished under Grant Number W911NF-17-1-0078. The views and conclu-
sions contained in this document are those of the authors and should not be interpreted as representing the official
policies, either expressed or implied, of the Army Research Office, the Defense Advanced Research Projects Agency,
or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government
purposes notwithstanding any copyright notation herein.
†Department of Computer Science, Princeton University, [email protected]
1
visited. The traditional formulation of these models comes from physics and relies on the clas-
sic tools of the trade: stochastic differential calculus, McKean interpretations, Feynman-Kac
models, Fokker-Planck PDEs, etc. [6, 8, 17, 23]. These techniques assume symmetries that are
typically absent from the "mesoscopic" scales of natural algorithms; they also often operate at
the thermodynamic limit, which rules out genuine agent-based modeling. Our goal is to initi-
ate a theory of discrete-time Markov chains whose topologies vary as a function of the current
probability distribution. Thus the entire theory of finite Markov chains should be recoverable as
a special case. Our contribution comes in two independent parts: the first one is a generalization
of the classification of Markov chains to the time-varying case; the second part is the bifurcation
analysis of Markov influence systems. The work highlights the signature trait of time-varying
random walks, which is the possibility of super-exponential ("hyper-torpid") mixing.
Renormalization. The term refers to a wide-reaching approach to complex systems that orig-
inated in quantum field theory and later expanded into statistical mechanics and dynamics.
Whether in its exact or coarse-grained form, the basic idea is to break down a complex sys-
tem into a hierarchy of simpler parts. When we define a dynamics on the original system (think
of interacting particles moving about) then the hierarchy itself creates its own dynamics between
the layers. This new "renormalized" dynamics can be entirely different from the original one.
Crucially, it can be both easier to analyze and more readily expressive of global properties. For
example, second-order phase transitions in the Ising model might correspond to fixed points of
the renormalized dynamics.
What is the relation to Markov chains? The standard classification of the states of a Markov
chain is an example of exact renormalization. Recall that the main idea behind the classification
is to express the chain as an acyclic directed graph, its condensation, whose vertices correspond
to its strongly connected components. This creates a two-level hierarchy (fig.1): a tree with
a root (the condensation) and its children (the strongly connected components). In a random
walk, the probability mass will flow entirely into the k (= 2) sinks of the condensation. The
stationary distribution lies on an attracting manifold of dimension k − 1 (= 1). The hierarchy
has only two levels. The situation is different with time-varying Markov chains, which can have
deep hierarchies.
Figure 1: The condensation of a graph.
Consider an infinite sequence (gk)k>0 of digraphs over a fixed set of labelled vertices. A
2
temporal random walk is defined by picking a starting vertex, moving to a random neighbor
in g1, then a random neighbor in g2, and so on, forever [9, 10, 19, 24, 32]. Note that a temporal
walk might not match a path in any of the graphs. How would one classify the states of this
"dynamic" Markov chain? Repeating the condensation decomposition at each step makes little
sense, as it carries no information about the temporal walks. Instead, we want to monitor when
and where temporal walks are extended. The cumulant graph compiles all such extensions and,
when the process stalls at time t(cid:48), reboots it by restarting from scratch at t = t(cid:48). We define
a grammar with which we parse the sequence (gk)k>0 accordingly. The method, explained in
detail in the next section, is very general and likely to be useful elsewhere.
Markov influence systems. All finite Markov chains oscillate periodically or mix to a station-
ary distribution. One key fact about their dynamics is that the timescales never exceed a single
exponential in the number of states. Allowing the transition probabilities to fluctuate over time
at random does not change that basic fact [2, 12, 13]. Markov influence systems are quite differ-
ent in that regard. Postponing formal definitions, let us think of an MIS for now as a dynamical
system defined by iterating the map f : x(cid:62) (cid:55)→ x(cid:62)S (x), where x is a probability distribution rep-
resented as a column vector in Rn and S (x) is a stochastic matrix that is piecewise-constant as a
function of x. We assume that the discontinuities are flats (ie, affine subspaces). The assumption
is not nearly as restrictive as it appears, as we explain with a simple example.
(cid:80)
Consider a random variable ξ over the distribution x and fix two n-by-n stochastic matrices
A and B. Define S (x) = A (resp. B) if varx ξ > 1 (resp. else); in other words, the random walk
picks one of two stochastic matrices at each step depending on the variance of ξ with respect
to the current state distribution x. Note that, in violation of our assumption, the discontinuity
is quadratic in x. This is not an issue because we can always linearize the variance: we begin
i, j (ξi − ξ j)2xix j and the fact that y := (xix j)i, j is a probability
with the identity varx ξ = 1
distribution. We form the Kronecker square T(y) = S (x) ⊗ S (x) and lift the system into the
2
(n2 − 1)-dimensional unit simplex to get a brand-new MIS defined by the map y(cid:62) (cid:55)→ y(cid:62)T(y).
We now have linear discontinuities. This same type of tensor lift can be used to linearize any
algebraic constraints; this requires making the polynomials over xi homogeneous, which we can
i xi = 1. Using ideas from [7], one can go even further than that
and base the stepwise edge selection on the outcome of any first-order logical formula we may
fancy (with the xi's acting as free variables). The key fact behind this result is that the first-
order theory of the reals is decidable by quantifier elimination. This allows us to pick the next
stochastic matrix at each time step on the basis of the truth value of a Boolean logic formula
with arbitrarily many quantifiers (see [7] for details). This discussion is meant to highlight the
fact that assuming linear discontinuities is not truly restrictive.
do by using the identity(cid:80)
We prove in this article that an irreducible Markov influence system is almost always asymp-
totically periodic. (An MIS is irreducible if S (x) forms an irreducible chain for each x.) We
extend this result to larger families of Markov influence systems. We also give an example of
"hyper-torpid" mixing: an MIS that converges to a stationary distribution in time equal to a
tower-of-twos in the size of the chain. This bound also applies to the period of certain peri-
odic MIS. The emergence of timescales far beyond the reach of standard Markov chains is a
distinctive feature of Markov influence systems.
Unlike in a standard random walk, the noncontractive eigenspace of an MIS may vary over
3
time. It is this spectral incoherence that renormalization attempts to "tame." To see why this has
a strong graph-theoretic flavor, observe that at each time step the support of the stationary distri-
bution can be read off the topology of the current graph: for example, the number of sinks in the
condensation is equal to the dimension of the principal eigenspace plus one. Renormalization
can thus be seen as an attempt to restore coherence to an ever-changing spectral landscape via a
dynamic hierarchy of graphs, subgraphs, and homomorphs.
The bifurcation analysis at the heart of the analysis depends on a notion of "general position"
aimed at bounding the growth rate of the induced symbolic dynamics [4,21,33]. The root of the
problem is a clash between order and randomness similar to the conflict between entropy and
energy encountered in the Ising model. The tension between these two "forces" is mediated by
the critical values of a perturbation parameter, which are shown to form a Cantor set of Hausdorff
dimension strictly less than 1. Final remark: There is a growing body of literature on dynamic
graphs [1, 3, 6, 16, 19, 22, 24, 25, 28, 29] and their random walks [2, 9 -- 13, 15, 20, 26, 27, 30, 32].
What distinguishes this work from its predecessors is that the changes in topology are induced
endogenously by the system itself (via a feedback loop).
2 How to Parse a Graph Sequence
Throughout this work, a digraph refers to a directed graph with vertices in [n] := {1, . . . , n}
and a self-loop at each vertex. Graphs and digraphs (words we use interchangeably) have no
multiple edges. We denote digraphs by lower-case letters (g, h, etc) and use boldface symbols
for sequences. A digraph sequence g = (gk)k>0 is an ordered, finite or infinite, list of digraphs
over the vertex set [n]. The digraph gi × g j consists of all the edges (x, y) such that there exist
an edge (x, z) in gi and another one (z, y) in g j for at least one vertex z. The operation × is
associative but not commutative; it corresponds roughly to matrix multiplication. We define the
cumulant(cid:81)≤k g = g1×···×gk and write(cid:81) g = g1×g2×··· for finite g. The cumulant indicates
all the pairs of vertices that can be joined by a temporal walk of a given length. The mixing time
of a random walk on a (fixed) graph depends on the speed at which information propagates and,
in particular, how quickly the cumulant becomes transitive. In the time-varying case, mixing is
a more complicated proposition, but the emergence of transitive cumulants is still what guides
the parsing process.
An edge (x, y) of a digraph g is leading if there is u such that (u, x) is an edge of g but (u, y)
is not. The non-leading edges form a subgraph of g, which is denoted by tf (g) and called the
transitive front of g. For example, tf (x → y → z) is the graph over x, y, z with the single edge
x → y (and the three self-loops); on the other hand, the transitive front of a directed cycle over
three or more vertices has no edges besides the self-loops. We denote by cl(g) the transitive
closure of g: it is the graph that includes an edge (x, y) for any two vertices x, y with a path from
x to y. Note that tf (g) (cid:22) g (cid:22) cl(g).
• An equivalent definition of the transitive front is that the edges of tf (g) are precisely the
pairs (i, j) such that Ci ⊆ C j, where Ck denotes the set of vertices l such that (l, k) is an
edge of g. Because each vertex has a self-loop, the inclusion Ci ⊆ C j implies that (i, j) is
an edge of g. If g is transitive, then tf (g) = g. The set-inclusion definition of the transitive
front shows that it is indeed transitive: ie, if (x, y) and (y, z) are edges, then so is (x, z).
4
Given two graphs g, h over the same vertex set, we write g (cid:22) h if all the edges of g are in
h (with strict inclusion denoted by the symbol ≺). Because of the self-loops, g, h (cid:22) g × h.
• A third characterization of tf (g) is as the unique densest graph h over [n] such that g× h =
g: we call this the maximally-dense property of the transitive front, and it is the motivation
behind our use of the concept.
Indeed, the failure of subsequent graphs to grow the
cumulant implies a structural constraint on them. This is the sort of structure that parsing
attempts to tease out.
observation that, because of the self-loops, the cumulant(cid:81)≤k g is monotonically nondecreasing
The parser. The parse tree of a (finite or infinite) graph sequence g = (gk)k>0 is a rooted tree
whose leaves are associated with g1, g2, . . . from left to right. The purpose of the parse tree
is to monitor the formation of new temporal walks as time progresses. This is based on the
with k (with all references to graph ordering being relative to (cid:22)). If the increase were strict at
each step then the parse tree would look like a fishbone. But, of course, the increase cannot go
on forever. To continue to extract structure even when the cumulant is "stuck" is what parsing is
all about. The underlying grammar consists of three productions: (1a) and (1b) renormalize the
graph sequence along the time axis, while (2) creates the hierarchical clustering of the graphs in
the sequence g.
1. Temporal renormalization We express the sequence g in terms of minimal subse-
quences with cumulants equal to(cid:81) g. There is a unique decomposition
g = g1, gm1, . . . , gk, gmk , gk+1
(cid:1) × gmi
The two productions below create the temporal parse tree.
such that
(i) g1 = g1, . . . , gm1−1; gi = gmi−1+1, . . . , gmi−1 (1 < i ≤ k); and gk+1 = gmk+1, . . . .
=(cid:81) g, for any i ≤ k; and(cid:81) gi ≺(cid:81) g, for any i ≤ k + 1.
(ii) (cid:0)(cid:81) gi
• Transitivization. Assume that(cid:81) g is not transitive. We define h = tf ((cid:81) g) and note
that h ≺(cid:81) g. It follows from the maximally-dense property of the transitive front
that k = 1. Indeed, k > 1 implies that(cid:81) g = (cid:0)(cid:81) g2
(cid:1) × gm1
(cid:9) =
tf ((cid:81) g), which contradicts the non-transitivity of(cid:81) g. We have the production
(cid:0) (g2) (cid:52) h(cid:1),
(cid:1) × gm2 (cid:22) tf(cid:8)(cid:0)(cid:81) g1
g −→ (cid:0) g1
(cid:1) gm1
(1a)
In the parse tree, the node for g has three children: the first one serves as the root
of the temporal parse subtree for g1; the second one is the leaf associated with the
graph gm1; the third one is a special node annotated with the label h, which serves
as the parent v of the node w rooting the parse subtree for g2. The node w is la-
beled (implicitly) by the transitive graph cl((cid:81) g2), which serves as a coarse-grained
approximation of(cid:81) g2. The purpose of annotating a special node with the label h
is to provide an intermediate approximation of(cid:81) g2 that is strictly finer than the
"obvious" cl((cid:81) g). These coarse-grained approximations are called sketches.
5
• Cumulant completion. Assume that(cid:81) g is transitive. We have the production
(cid:0) gk+1
(cid:1).
(1b)
g −→ (cid:0) g1
(cid:1) gm1
(cid:0) g2
(cid:1) gm2 ···(cid:0) gk
(cid:1) gmk
Note that the index k may be infinite and any of the subsequences gi might be empty
(for example, gk+1 if k = ∞).
2. Topological renormalization Network renormalization exploits the fact that the infor-
mation flowing across the system might get stuck in portions of the graph for some period
of time: when this happens, we cluster the graphs using topological renormalization. As
we just discussed, each node v of the temporal parse tree that is not annotated by (1a)
is labeled by the sketch cl((cid:81) g), where g is the graph sequence formed by the leaves of
the subtree rooted at v. In this way, every path from the root of the temporal parse tree
comes with a nested sequence of sketches h1 (cid:23) ··· (cid:23) hl. Pick two consecutive ones,
hi, hi+1: these are two transitive graphs whose strongly connected components, therefore,
are cliques. Let V1, . . . , Va and W1, . . . , Wb be the vertex sets of the cliques corresponding
to hi and hi+1, respectively. Since hi+1 is a subgraph of hi, it follows that each Vi is a
disjoint union of the form Wi1 ∪ ··· ∪ Wisi .
• Decoupling. We decorate the temporal parse tree with an additional tree connecting
the sketches present along each one of its paths. These topological parse trees are
formed by all the productions of the type:
Vi −→ Wi1 ··· Wisi
.
(1)
A sketch at a node v of the temporal tree can be viewed as an acyclic digraph over
cliques: its purpose is to place limits on the movement of the probability mass in
any temporal random walk corresponding to the leaves of the subtree rooted at v. In
particular, it indicates how decoupling might arise in the system.
The maximum depth of the temporal parse tree is O(n2) because each child's cumulant loses
at least one edge from its parent's (or grandparent's) cumulant. To see why the quadratic bound
is tight, consider a bipartite graph V = L ∪ R, where L = R and each pair from L × R is added
one at a time as a bipartite graph with a single nonloop edge; the leftmost path of the parse tree
is of quadratic length.
Left-to-right parsing. The temporal tree can be built on-line by scanning the graph sequence
g with no need to back up. Let g(cid:48) denote the sequence formed by appending the graph g to the
end of the finite graph sequence g. If g is empty, then the tree T (g(cid:48)) consists of a root with one
child labeled g. If g is not empty and(cid:81) g ≺(cid:81) g(cid:48), the root of T (g(cid:48)) has one child formed by the
root of T (g) as well as a child labeled g. Assume now that g is not empty and that(cid:81) g =(cid:81) g(cid:48).
Let v be the lowest internal node on the rightmost path of T (g) such that cv × g = cv, where cu
denotes the product of the graphs associated with the leaves of the subtree rooted at node u of
T (g). Let w be the rightmost child of v; note that v and w always exist. We explain how to form
T (g(cid:48)) by editing T (g).
6
1. cv is transitive and w is a leaf: We add a leaf labeled g as the new rightmost child z of v if
g = cv; otherwise, we do the same but, instead of attaching g directly to v, we make it the
unique child of z.
2. cv is transitive and w is not a leaf: We add a leaf labeled g as the new rightmost child of v
if cw × g = cv; otherwise, we add a new rightmost child z to v and we give z two children:
its left child is w and its subtree below; its right child is a leaf labeled g.
3. cv is not transitive and w is a leaf: We attach a new rightmost child z to v, and we make it
a special node annotated with the label tf (cv). We give z a single child z(cid:48), which is itself
the parent of a new leaf labeled g.
4. cv is not transitive and w is not a leaf: We note that cw × g (cid:22) tf (cv) ≺ cv. We attach a new
rightmost child z to v, and we make it a special node annotated with the label tf (cv). We
give z a single child z(cid:48), to which we attach two children: one of them is w (and its subtree)
and the other one is the leaf labeled g.
Undirected graphs. A graph is called undirected if any edge (x, y) with x (cid:44) y comes with
its companion (y, x). Consider a sequence of undirected graphs over [n]. We begin with the
observation that the cumulant of a subsequence might itself be directed; for example, the product
g1×g2 = (x ↔ y z)×(x y ↔ z) has a directed edge from x to z but not from z to x. We can use
undirectedness to strengthen the definition of the transitive front. Recall that tf (g) is the unique
densest graph h such that g × h = g. Its purpose is the following: if g is the current cumulant,
the transitive front of g is intended to include any edge that might appear in subsequent graphs
in the sequence without extending any path in g. Since, in the present case, the only edges
considered for extension will be undirected, we might as well require that h itself (unlike g)
should be undirected. In this way, we redefine the transitive front, now denoted by utf (g), as
the unique densest undirected graph h such that g × h = g. Its edge set includes all the pairs
(i, j) such that Ci = C j. Because of self-loops, the condition implies that (i, j) is an undirected
edge of g. This forms an equivalence relation among the vertices, so that utf (g) actually consists
of disconnected, undirected cliques. To see the difference with the directed case, we take our
previous example and note that tf (g1 × g2) has the edges (x, y), (x, z), (y, z), (z, y) in addition to
the self-loops, whereas utf (g1 × g2) has the single undirected edge (y, z) plus self-loops.
The depth of the parse tree can still be as high as quadratic in n. To see why, consider the
following recursive construction. Given a clique Ck over k vertices x1, . . . , xk at time t, attach
to it, at time t + 1, the undirected edge (x1, y). The cumulant gains the undirected edge (x1, y)
and the directed edges (xi, y) for i = 2, . . . , k. At time t + 2, . . . , t + k, visit each one of the k − 1
undirected edges (x1, xi) for i > 1, using single-edge undirected graphs. Each such step will see
the addition of a new directed edge (y, xi) to the cumulant, until it becomes the undirected clique
Ck+1. The quadratic lower bound on the tree depth follows immediately.
Backward parsing. The sequence of graphs leads to products where each new graph is mul-
tiplied to the right, as would happen in a time-varying Markov chain. Algebraically, the matri-
ces are multiplied from left to right. In diffusive systems (eg, multiagent agreement systems,
Hegselmann-Krause models, Deffuant systems), however, matrices are multiplied from right to
7
left. Although the dynamics can be quite different, the same parsing algorithm can be used.
Given a sequence g = (gk)k>0, its backward parsing is formed by applying the parser to the
sequence ←−g = (hk)k>0, where hk is derived from gk by reversing the direction of every edge, ie,
(x, y) becomes (y, x). Once the parse tree for ←−g has been built, we simply restore each edge to
its proper direction to produce the backward parse tree of g.
3 The Markov Influence Model
Let Sn−1 (or S when the dimension is understood) be the standard simplex (cid:8) x ∈ Rn x ≥
0 , (cid:107)x(cid:107)1 = 1(cid:9) and let S denote set of all n-by-n rational stochastic matrices. A Markov in-
fluence system (MIS ) is a discrete-time dynamical system with phase space S, which is defined
by the map f : x(cid:62) (cid:55)→ f (x) := x(cid:62)S(x), where x ∈ S and S is a function S (cid:55)→ S that is constant
over the pieces of a finite polyhedral partition P ={Pk} of S; we define f as the identity on the
discontinuities of the partition (fig.2).
(i) We define the digraph g(x) (and its corresponding Markov chain) formed by the positive
entries of S(x). To avoid inessential technicalities, we assume that the diagonal of each S(x)
is strictly positive (ie, g(x) has self-loops). In this way, any orbit of an MIS corresponds to a
lazy, time-varying random walk with transitions defined endogenously.1 We recall some basic
terminology. The orbit of x ∈ S is the infinite sequence ( f t(x))t≥0 and its itinerary is the corre-
sponding sequence of cells Pk visited in the process. The orbit is periodic if f t(x) = f s(x) for
any s = t modulo a fixed integer. It is asymptotically periodic if it gets arbitrarily close to a
periodic orbit over time.
Figure 2: A Markov influence system: each region Pk is associated with a single stochastic matrix. The
first five steps of an orbit are shown to visit regions P1, P2, P4, P4, P5 in this order.
(ii) The discontinuities in P are formed by hyperplanes in Rn of the form a(cid:62)
i x = 1 + δ,
where δ is confined to an interval Ω := [−ω, ω]. Assuming general position, we can pick a
1As discussed in the introduction, to access the full power of first-order logic in the stepwise choice of digraphs
requires nonlinear partitions, but these can be linearized by a suitable tensor construction [7].
8
small positive ω < 1/2 so that P remains (topologically) the same over Ω.2 Thus, the MIS
remains well-defined for all δ ∈ Ω. As we show in Section 5, the parameter δ is necessary to
keep chaos at bay.
(iii) The coefficient of ergodicity τ(M) of a matrix M is defined as half the maximum (cid:96)1-
distance between any two of its rows [31]. It is submultiplicative for stochastic matrices, a direct
consequence of the identity τ(M) = max
(cid:110)(cid:107)x(cid:62)M(cid:107)1 : x(cid:62)1 = 0 and (cid:107)x(cid:107)1 = 1
(cid:111)
.
(iv) Given ∆ ⊆ Ω, let Lt
∆ denote the set of t-long prefixes of any itinerary for any starting
position x ∈ S and any δ ∈ ∆. We define the ergodic renormalizer η∆ as the smallest integer
such that, for any t ≥ η∆ and any matrix sequence S1, . . . , St associated with an element of
∆, the product S1 ··· St is primitive (ie, some high enough power is a positive matrix) and its
Lt
coefficient of ergodicity is less than 1/2. We assume in this section that η := ηΩ < ∞ and
discuss in Section 4 how to relax this assumption via renormalization. Let D be the union of
0≤k≤t f −k(D) and
t≥0 Zt. Note that Z ⊆ S since the latter is the domain of f . Remarkably, for almost all
the hyperplanes from P in Rn (where δ is understood). We define Zt = (cid:83)
Z = (cid:83)
δ ∈ Ω, Zt becomes strictly equal to Z in a finite number of steps.3
Lemma 3.1. There is a constant c > 0 such that, for any ε > 0, there exists an integer ν ≤
2ηc log(1/ε) and a finite union K of intervals of total length less than ε such that Zν = Zν−1, for
any δ ∈ Ω \ K.
Note that Zν = Zν−1 implies that Z = Zν. Indeed, suppose that Zt+1 ⊃ Zt for t ≥ ν; then,
f t+1(y) ∈ D but f t(y) (cid:60) D for some y ∈ S; in other words, f ν(x) ∈ D but f ν−1(x) (cid:60) D for
x = f t−ν+1(y), which contradicts the equality Zν = Zν−1.
Corollary 3.2. For δ almost everywhere in Ω,4 every orbit is asymptotically periodic.
Proof. The equality Z = Zν implies the eventual periodicity of the symbolic dynamics. The
period cannot exceed the number of connected components in the complement of Z. Once an
itinerary becomes periodic at time to with period σ, the map f t can be expressed locally by
matrix powers. Indeed, divide t − to by σ and let q be the quotient and r the remainder; then,
locally, f t = gq ◦ f to+r, where g is specified by a stochastic matrix with a positive diagonal,
bidden" intervals. Define Kl = (cid:83)
which implies convergence to a periodic point at an exponential rate. Finally, apply Lemma 3.1
repeatedly, with ε = 2−l for l = 1, 2, . . . and denote by Kl be the corresponding union of "for-
l>0 Kl; then Leb(Kl) ≤ 21−l and hence
Leb(K∞) = 0. The lemma follows from the fact that any δ ∈ Ω outside of K∞ lies outside of Kl
(cid:3)
for some l > 0.
2Let Hδ be the set consisting of the hyperplane(cid:80)
on the set of vectors {ai}. Note that this problem arises only because of the constraint(cid:80)
i xi = 1 together with those used to define P. We assume that
H0 is in general position; hence so is Hδ for any δ ∈ Ω, where 0 < ω < 1/2 is smaller than a value that depends only
i xi = 1, since otherwise the
hyperplane arrangement is central around (0, . . . , 0,−1).
3Recall that both Z and Zt depend on δ ∈ Ω. All the constants used in this work may depend on the system's
parameters such as n, P (but not on δ). Dependency on other parameters is indicated by a subscript.
j≥l K j and K∞ = (cid:84)
4Meaning outside a subset of Ω of Lebesgue measure zero.
9
The corollary states that the set of "nonperiodic" values of δ has measure zero in parameter
space. Our result is actually stronger than that. We prove that the nonperiodic set can be covered
by a Cantor set of Hausdorff dimension strictly less than 1. The remainder of this section is
devoted to a proof of Lemma 3.1.
3.1 Shift spaces and growth rates
The growth exponent of a language is defined as limn→∞ 1
n maxk≤n log N(k), where N(k) is the
number of words of length k; for example, the growth exponent of {0, 1}∗ is 1 (all logarithms
taken to the base 2). The language consisting of all the itineraries of a Markov influence sys-
tem forms a shift space and its growth exponent is the topological entropy of its symbolic
dynamics [33] -- not to be confused with the topological entropy of the MIS itself. It can be
strictly positive, which is a sign of chaos. We show that, for a typical system, it is zero, the
key fact underlying periodicity. Let M1, . . . , MT be n-by-n matrices from a finite set M of
primitive stochastic rational matrices with positive diagonals, and assume that τ(M) < 1/2
for M ∈ M; hence τ(M1 ··· Mk) < 2−k. Because each product M1 ··· Mk is a primitive ma-
trix, it can be expressed as 1π(cid:62)
+ Qk (by Perron-Frobenius), where πk is its (unique) sta-
tionary distribution.5 If π is a stationary distribution for a stochastic matrix S , then its j-th
i πi(s j − si); hence, by the triangular inequality,
row s j satisfies s j − π(cid:62) = s j − π(cid:62)S = (cid:80)
(cid:107)s j − π(cid:62)(cid:107)1 ≤(cid:80)
M1 ··· Mk = 1π(cid:62)
i πi(cid:107)s j − si(cid:107)1 ≤ 2τ(S ). This implies that
(cid:107)Qk(cid:107)∞ ≤ 2τ(M1 ··· Mk) < 21−k.
(2)
+ Qk
k
k
Property U. Fix a vector a ∈ Qn, and denote by M (θ) the n-by-m matrix with the m column
vectors M1 ··· Mki a, where θ = (k1, . . . , km) is an increasing sequence of integers in [T]. We
say that property U holds if there exists a rational vector u = u(θ) such that 1(cid:62)u = 1 and
x(cid:62)M (θ)u does not depend on the variable x ∈ S.6 Property U is a quantifier elimination device
for expressing a notion of "general position" for an MIS. To see why, consider a simple statement
such as "the three points(cid:0)x, x2(cid:1),(cid:0)x + 1, (x + 1)2(cid:1), and(cid:0)x + 2, (x + 2)2(cid:1) cannot be collinear for
any value of x." This can be expressed by saying that a certain determinant polynomial in x
is constant. Likewise, the vector u manufactures a quantity, x(cid:62)M (θ)u, that "eliminates" the
variable x. Some condition on u is needed since otherwise we could pick u = 0. Note that
property U would be obvious if all the matrices Qk in (2) were null: indeed, we would have
x(cid:62)M (θ) = x(cid:62)1(b1, . . . , bm) = (b1, . . . , bm), where bi = π(cid:62)
a. This suggests that property U rests
ki
on the decaying properties of Qk.
To see the relevance of general position to the dynamics of an MIS, consider the iterates of a
small ball through the map f . To avoid chaos, it is intuitively obvious that these iterated images
should not fall across discontinuities too often. Fix such a discontinuity: if we think of the ball
(cid:1) are both primitive but their product is not.
(cid:1) and(cid:0) 0 1
(cid:0) 1 1
x(cid:62)(cid:0)11(cid:62)(cid:1) u = 1 for u = 1
1 1
1 0
n 1.
5Positive diagonals play a key role here because primitiveness is not closed under multiplication: for example,
6Because x is a probability distribution, property U does not imply that M (θ)u = 0; for example, we have
10
as being so small it looks like a point, then the case we are trying to avoid consists of many
points (the ball's iterates) lying on (or near) a given hyperplane. This is similar to the definition
of general position, which requires that a set of point should not lie on the same hyperplane.
Lemma 3.3. There exists a constant b > 0 (linear in n) such that, given any integer T > 0 and
any increasing sequence θ in [T] of length at least T 1−α/α, property U holds, where α := µ−b
and µ is the maximum number of bits needed to encode any entry of Mk for any k ∈ [T].
Proof. By choosing b large enough, we can ensure that T is as big as we want. The proof is
a mixture of algebraic and combinatorial arguments. We begin with a Ramsey-like statement
about stochastic matrices.
Fact 3.4. There is a constant d > 0 such that, if the sequence θ contains j0, . . . , jn with ji ≥
dµ ji−1 for each i ∈ [n], then property U holds.
Proof. By (2), (cid:107)Qka(cid:107)∞ < c02−k for constant c0 > 0. Note that Qk has rational entries over O(µk)
bits: the bound follows from the fact that the stationary distribution πk has rational coordinates
over O(µk) bits; as noted earlier, the constant factors may depend on n. We write M (θ) =
1a(cid:62)Π (θ) + Q (θ), where Π (θ) and Q (θ) are the n-by-m matrices formed by the m column vectors
πki and Qkia, respectively, for i ∈ [m]; recall that θ = (k1, . . . , km). The key fact is that the
dependency on x ∈ S is confined to the term Q (θ): indeed,
x(cid:62)M (θ)u = a(cid:62)Π (θ)u + x(cid:62)Q (θ)u.
(3)
This shows that, in order to satisfy property U, it is enough to ensure that Q (θ)u = 0 has a
solution such that 1(cid:62)u = 1. Let σ = ( j0, . . . , jn−1). If Q (σ) is nonsingular then, because each
one of its entries is a rational over O(µ jn−1) bits, we have det Q (σ) ≥ cµ jn−1
, for constant c1 > 0.
Let R be the (n + 1)-by-(n + 1) matrix derived from Q (σ) by adding the column Q jna to its right
and then adding a row of ones at the bottom. If R is nonsingular, then R u = (0, . . . , 0, 1)(cid:62) has
a (unique) solution in u and property U holds (after padding u with zeroes). Otherwise, we
expand the determinant of R along the last column. Suppose that det Q (σ) (cid:44) 0. By Hadamard's
inequality, all the cofactors are at most a constant c2 > 0 in absolute value; hence, for d large
enough,
1
0 = det R ≥ det Q (σ) − nc2(cid:107)Q jna(cid:107)∞ ≥ cµ jn−1
− nc2c02− jn > 0.
1
This contradiction implies that Q (σ) is singular, so (at least) one of its rows can be expressed as
a linear combination of the others. We form the n-by-n matrix R(cid:48) by removing that row from R,
= 0 to rewrite Q (θ)u = 0 as R(cid:48)u(cid:48) = (0, . . . , 0, 1)(cid:62),
together with the last column, and setting u jn
where u(cid:48) is the restriction of u(cid:48) to the columns indexed by R(cid:48). Having reduced the dimension
of the system by one variable, we can proceed inductively in the same way; either we terminate
with the discovery of a solution or the induction runs its course until n = 1 and the corresponding
1-by-1 matrix is null, so that the solution 1 works. Note that u has rational coordinates over
(cid:3)
O(µT) bits.
Let N(T) be the largest sequence θ in [T] such that property U does not hold. Divide [T]
into bins [(dµ)k, (dµ)k+1 − 1] for k ≥ 0. By Fact 3.4, the sequence θ can intersect at most
11
2n of them; thus, if T > t0, for some large enough t0 = (dµ)O(n), there is at least one empty
interval in T of length T/(dµ)2n+3. This gives us the recurrence N(T) ≤ T for T ≤ t0 and
N(T) ≤ N(T1) + N(T2), where T1 + T2 ≤ βT, for a positive constant β = 1 − (dµ)−2n−3. The
recursion to the right of the empty interval, say, N(T2), warrants a brief discussion. The issue is
that the proof of Fact 3.4 relies crucially on the property that Qk has rational entries over O(µk)
bits -- this is needed to lower-bound det Q (σ) when it is not 0. But this is not true any more,
because, after the recursion, the columns of the matrix M (θ) are of the form M1 ··· Mk a, for
T1 + L < k ≤ T, where L is the length of the empty interval and T = T1 + L + T2. Left as such,
the matrices use too many bits for the recursion to go through. To overcome this obstacle, we
observe that the recursively transformed M (θ) can be factored as AB, where A = M1 ··· MT1+L
and B consists of the column vectors MT1+L+1 ··· Mk a. The key observation now is that, if
x(cid:62)B u does not depend on x, then neither does x(cid:62)M (θ) u, since it can be written as y(cid:62)B u where
y = A(cid:62)x ∈ S. In this way, we can enforce property U while having restored the proper encoding
length for the entries of M (θ).
Plugging in the ansatz N(T) ≤ t0T γ, for some unknown positive γ < 1, we find by Jensen's
2 ) ≤ t021−γβγT γ. For the ansatz to hold true, we
inequality that, for all T > 0, N(T) ≤ t0(T γ
need to ensure that 21−γβγ ≤ 1. Setting γ = 1/(1−log β) < 1 completes the proof of Lemma 3.3.
(cid:3)
hyperplane in Rn. We consider a set of canonical intervals of length ρ (or less): Dρ =(cid:8) [kρ, (k +
1)ρ] ∩ Ω k ∈ Z(cid:9), where ρ > 0 (specified below), Ω = ωI, I := [−1, 1], and 0 < ω < 1/2.
Define φk(x) = x(cid:62)M1 ··· Mk for x ∈ Rn and k ≤ T; and let hδ : a(cid:62)x = 1 + δ be some
Roughly, the "general position" lemma below says that, for most δ, the φk-images of any ρ-wide
cube centered in the simplex S cannot come very near the hyperplane hδ for most values of
k ≤ T. This may be counterintuitive. After all, if the stochastic matrices Mi are the identity, the
images stay put, so if the initial cube collides then all of the images will! The point is that Mi
is primitive so it cannot be the identity. The low coefficients of ergodicity will also play a key
role. The crux of the lemma is that the exclusion set U does not depend on the choice of x ∈ S.
A point of notation: α refers to its use in Lemma 3.3.
Lemma 3.5. For any real ρ > 0 and any integer T > 0, there exists U ⊆ Dρ of size cT = 2O(µT),
where cT is independent of ρ, such that, for any ∆ ∈ Dρ\U and x ∈ S, there are at most T 1−α/α
integers k ≤ T such that φk(X) ∩ h∆ (cid:44) ∅, where X = x + ρIn and h∆ :=(cid:83)
+ T γ
1
δ∈∆ hδ.
Proof. In what follows, b0, b1, . . . refer to suitably large positive constants (which, we shall
recall, may depend on n, a, etc). We assume the existence of more than T 1−α/α integers k ≤ T
such that φk(X) ∩ h∆ (cid:44) ∅ for some ∆ ∈ Dρ and draw the consequences: in particular, we infer
certain linear constraints on δ; by negating them, we define the forbidden set U and ensure the
conclusion of the lemma. Let k1 < ··· < km be the integers in question, where m > T 1−α/α. For
each i ∈ [m], there exists x(i) ∈ X and δi ∈ ∆ such that x(i)(cid:62)M1 ··· Mkia − 1 − δi ≤ ρ. Note
that δi − δ ≤ ρ for some δ ∈ ∆ common to all i ∈ [m]. By the stochasticity of the matrices,
By Lemma 3.3, there is a rational vector u such that 1(cid:62)u = 1 and x(cid:62)M (θ)u = ψ(M (θ), a) does
not depend on the variable x ∈ S; on the other hand, x(cid:62)M (θ)u − (1 + δ) ≤ b1ρ. Two quick
remarks: (i) the term 1 + δ is derived from (1 + δ)1(cid:62)u = 1 + δ; (ii) b1 ≤ (b0 + 2)(cid:107)u(cid:107)1, where
it follows that (cid:0)x(i) − x(cid:1)(cid:62)M1 ··· Mkia ≤ b0 ρ; hence x(cid:62)M1 ··· Mkia − 1 − δ ≤ (b0 + 2)ρ.
12
u is a rational over O(µT) bits. We invalidate the condition on k1, . . . , km by keeping δ outside
the interval ψ(M (θ), a) − 1 + b1ρI, which rules out at most 2(b1 + 1) = 2O(µT) intervals from Dρ.
Repeating this for all sequences (k1, . . . , km) raises the number of forbidden intervals by a factor
(cid:3)
of at most 2T .
Topological entropy. We identify the family M with the set of all matrices of the form
S1 ··· Sk, for η ≤ k ≤ 3η (η = ηΩ), where the matrix sequence S1, . . . , Sk matches some ele-
Ω. By definition of the ergodic renormalizer, any M ∈ M is primitive and τ(M) < 1/2;
ment of Lk
furthermore, both µ and logM are in O(η). Our next result shows that the topological entropy
of the shift space of itineraries vanishes.
Lemma 3.6. For any real ρ > 0 and any integer T > 0, there exist tρ = O(η log ρ) and an
exclusion set V ⊆ Dρ of size dT = 2O(T) such that, for any ∆ ∈ Dρ\V, any integer t ≥ tρ, and
any σ ∈ Lt
(cid:9)(cid:12)(cid:12)(cid:12) ≤ ηbT 1−η−b, for constant b > 0.
∆, log
(cid:12)(cid:12)(cid:12)(cid:8) σ(cid:48) σ · σ(cid:48) ∈ Lt+T
∆
Proof.
In the lemma, tρ (resp. dT ) is independent of T (resp. ρ). The main point is that
the exponent of T is bounded away from 1. We define V as the union of the sets U formed
by applying Lemma 3.5 to each one of the hyperplanes hδ involved in P and every possible
phase space S× ∆ for the dynamical system induced by the map f↑ :(cid:0)x(cid:62), δ) (cid:55)→ ( x(cid:62)S(x), δ(cid:1). The
sequence of T matrices in M. This increases cT to 2O(ηT). Fix ∆ ∈ Dρ\V and consider the (lifted)
system is piecewise-linear with respect to the polyhedral partition P↑ of Rn+1 formed by treating
δ as a variable in hδ. Let Υt be a continuity piece for f t↑, ie, a maximal region of S × ∆ over
which the t-th iterate of f↑ is linear. Reprising the argument leading to (2), any matrix sequence
S1, . . . , St matching an element of Lt
∆ is such that S1 ··· St = 1π(cid:62) + Q, where
(cid:107)Q(cid:107)∞ < 22−t/η.
(4)
Thus there exists tρ = O(η log ρ) such that, for any t ≥ tρ, f t↑(Υt) ⊆ (x + ρIn) × ∆, for some
x = x(t, Υt) ∈ S. Consider a nested sequence Υ1 ⊇ Υ2 ⊇ ··· . Note that Υ1 is a cell of P↑,
f k↑ (Υk+1) ⊆ f k↑ (Υk), and Sl is the stochastic matrix used to map f l−1↑
(Υl) to f l↑(Υl) (ignoring the
δ-axis). We say there is a split at k if Υk+1 ⊂ Υk, and we show that, given any t ≥ tρ, there
are only O(ηT 1−α/α) splits between t and t + ηT, where α = η−b, for constant b.7 We may
confine our attention to splits caused by the same hyperplane hδ since P features only a constant
number of them. Arguing by contradiction, we assume the presence of at least 6ηT 1−α/α splits,
which implies that at least N := 2T 1−α/α of those splits occur for values of k at least 2η apart.
This is best seen by binning [t + 1, t + ηT] into T intervals of length η and observing that at
least 3N intervals must feature splits. In fact, this proves the existence of N splits at positions
separated by a least two consecutive bins. Next, we use the same binning to produce the matrices
M1, . . . , MT , where M j = St+1+( j−1)η ··· St+ jη.
Suppose that all of the N splits occur for values k of the form t + jη. In this case, a straight-
forward application of Lemma 3.5 is possible: we set X × ∆ = f t↑(Υt) and note that the functions
7We may have to scale b up by a constant factor since µ = O(η) and, by Lemma 3.3, α = µ−b.
13
φk are all products of matrices from the family M, which happen to be η-long products. The
number of splits, 2T 1−α/α, exceeds the number allowed by the lemma and we have a contra-
diction. If the splits do not fall neatly at endpoints of the bins, we use the fact that M includes
matrix products of any length between η and 3η. This allows us to reconfigure the bins so as to
form a sequence M1, . . . , MT with the splits occurring at the endpoints: for each split, merge its
own bin with the one to its left and the one to its right (neither of which contains a split) and use
the split's position to subdivide the resulting interval into two new bins; we leave all the other
bins alone.8 This leads to the same contradiction, which implies the existence of fewer than
O(ηT 1−α/α) splits at k ∈ [t, t + ηT]; hence the same bound on the number of strict inclusions
in the nested sequence Υt ⊇ ··· ⊇ Υt+ηT . The set of all such sequences forms a tree of depth
ηT, where each node has at most a constant number of children and any path from the root has
O(ηT 1−α/α) nodes with more than one child. Rescaling T to ηT and raising b completes the
(cid:3)
proof.
3.2 Proof of Lemma 3.1
We show that the nonperiodic δ-intervals can be covered by a Cantor set of Hausdorff dimension
less than one. All the parameters below refer to Lemma 3.6. We fix ε > 0 and ∆ ∈ Dρ\V and
assume that δ ∈ ∆. Set T = 2η2b, ρ = ε/(2dT ), and ν = tρ + kT, where k = cη log(1/ρ) for a large
enough constant c > 0. Since tρ = O(η log ρ) and dT = 2O(T), we have
ν = 2ηO(1) log(1/ε).
(5)
Let M be the matrix S1 ··· Sν, where S1, . . . , Sν the matrix sequence matching an element of Lν
∆.
By (4), diam(cid:96)∞(S M) ≤ 22−ν/η. There exists a point xM such that, given any point y ∈ S whose ν-
th iterate f ν(y) = zT is specified by the matrix M, that is, zT = yT M, we have (cid:107)xM−y(cid:107)∞ ≤ 22−ν/η.
Consider a discontinuity hδ : a(cid:62)
i x = 1 + δ of the system. Testing which side of it the point z lies
is equivalent to checking the point xM instead with respect to hδ(cid:48) for some δ(cid:48) that differs from δ
by O(2−ν/η). It follows that adding an interval of length O(2−ν/η) to the exclusion set V ensures
that all the ν-th iterates f ν(y) (specified by M) lie strictly on the same side of hδ for all δ ∈ ∆.
∆ and every ∆ ∈ Dρ\V increases the length covered by V from
Repeating this for every string Lν
∆2−ν/η/ρ) < ε. This last bound follows from the
its original dT ρ = ε/2 to at most dT ρ + O(Lν
consequence of Lemma 3.6 that logLν
∆ ≤ kηbT 1−η−b + O(tρ). Thus, for any δ ∈ Ω outside a set
of intervals covering a length less than ε, no f ν(x) lies on a discontinuity. It follows that, for any
(cid:3)
such δ, we have Zν = Zν−1 and, by (5), the proof is complete.
4 Applications
We can use the renormalization and bifurcation analysis techniques developed above to resolve
several important families of MIS. We discuss two simple cases here.
8We note the possibility of an inconsequential decrease in T caused by the merges. Also, we can now see clearly
why Lemma 3.5 is stated in terms of the slab h∆ and not the hyperplane hδ. This allows us to express splitting caused
by the hyperplane a(cid:62)x = 1 + δ in lifted space Rn+1.
14
Irreducible systems
4.1
A Markov influence system is called irreducible if the Markov chain g(x) is irreducible for all
x ∈ S; given the assumed presence of self-loops, each chain is ergodic. All the digraphs g(x)
of an irreducible MIS are strongly connected. Every step from the beginning sees growth in the
cumulant until it is a clique. To see why, assume by contradiction that the cumulant fails to
<k g =(cid:81)≤k g for k ≤ m1, where m1 is the smallest index for which
grow at an earlier step, ie,(cid:81)
(cid:81)≤m1 g is a clique. If so, then gk is a subgraph of tf ((cid:81)
(cid:81)
<k g). Because the latter is transitive and it
has in gk a strongly connected subgraph that spans all the vertices, it must be a clique; therefore
<m1 g is a clique, which contradicts our assumption. Because m1 ≤ n, each product of at most
n graphs is a clique, so its coefficient of ergodicity is at most 1 − γ, for γ > 0. Since the number
of distinct matrices is finite, their positive entries are uniformly bounded away from zero; hence
ηΩ is bounded from below by a positive constant (which may depend on n). By Lemma 3.1 and
Corollary 3.2, we conclude with:
Theorem 4.1. Typically, every orbit of an irreducible Markov influence system is asymptotically
periodic.
4.2 Weakly irreducible systems
We strengthen the previous result by assuming a fixed partition of the vertices such that each
digraph g(x) consists of disjoint strongly connected graphs defined over the subsets V1, . . . , Vl
of the partition. This means that no edges ever join two distinct Vi, V j and, within each Vi,
the graphs are always strongly connected with self-loops. Irreducible systems correspond to
the case l = 1. What makes weak irreducibility interesting is that the systems are not simply
the union of independent irreducible systems. Indeed, communication flows among states in
two ways: (i) directly, vertices collect information from incoming neighbors to update their
states; and (ii) indirectly, via the polyhedral partition P, the sequence of graphs for Vi may be
determined by the current states within other groups V j. In the extreme case, we can have the
co-evolution of two systems V1 and V2, each one depending entirely on the other one yet with
no links between the two of them. If the two subsystems were independent, their joint dynamics
could be expressed as a direct sum and resolved separately. This cannot be done, in general, and
the bifurcation analysis requires some modifications.
By using topological renormalization, we can partition the vertex set [n] into l subsets and
deal with each one separately. For the bifurcation analysis, Fact 3.4 relies on the rank-l expan-
sion
l(cid:88)
(cid:16)(cid:88)
(cid:17)
x(cid:62)M (θ) =
a(cid:62)Π (θ)
i
+ x(cid:62)Q (θ) ,
i=1
x j
j∈Vi
i=1 1Viπ(cid:62)
where (i) M1 ··· Mk =(cid:80)l
dependency is confined to the sums si :=(cid:80)
+ Qk; (ii) all vectors 1Vi and πi,k have support in Vi; (iii) Qk
is block-diagonal and (cid:107)Qk(cid:107)∞ < 21−k; (iv) Π (θ)
and Q (θ) are formed, respectively, by the column
vectors πi,k j and Qk ja for j ∈ [m], with θ = (k1, . . . , km). Property U no longer holds, however:
if l > 1, indeed, it is no longer true that x(cid:62)M (θ)u is independent of the variable x ∈ S. The
j∈Vi x j for i ∈ [l]. The key observation is that these
sums are time-invariant. We assume them to be rational and we redefine the phase space as
i,k
i
15
the invariant manifold(cid:81)l
(cid:0)si SVi−1(cid:1). The rest of the proof mimics the irreducible case, whose
conclusion therefore still applies.
i=1
Theorem 4.2. Typically, every orbit of a weakly irreducible Markov influence system is asymp-
totically periodic.
5 Hyper-Torpid Mixing and Chaos
Among the MIS that converge to a single stationary distribution, we show that the mixing time
the MIS
can be super-exponential. Very slow clocks can be designed in the same manner:
is periodic with a period of length equal to a tower-of-twos. The creation of new timescales
is what most distinguishes MIS from standard Markov chains. As we mentioned earlier, the
system can be chaotic. We prove all of these claims below.
5.1 A super-exponential mixer
How can reaching a fixed point distribution take so long? Before we answer this question for-
mally, we provide a bit of intuition. Imagine having three unit-volume water reservoirs A, B, C
alongside a clock that rings at noon every day. Initially, the clock is at 1pm and A is full while
B and C are empty. Reservoir A transfers half of its content to B and repeats this each hour until
the clock rings noon. At this point, reservoir A empties into C the little water that it has left
and B empties its content into A. At 1pm, we resume what we did the day before at the same
hour, ie, A transfers half of its water content to B, etc. This goes on until some day, the noon
hour rings and reservoir C finds its more than half full. (Note that the water level of C rises by
about 10−3 every day.) At this point, both B and C transfer all their water back to A, so that at
1pm on that same day, we are back to square one. The original 12-step clock has been extended
into a new clock of period roughly 1,000. The proof below shows how to simulate this iterative
process with an MIS.
Theorem 5.1. There exist Markov influence systems that mix to a stationary distribution in time
equal to a tower-of-twos of height linear in the number of states.
Proof. We construct an MIS with a periodic orbit of length equal to a power-of-twos of height
proportional to n; this is the function f (n) = 2 f (n−1), with f (1) = 1. It is easy to turn it into
one with an orbit that is attracted to a stationary distribution (a fixed point) with an equivalent
mixing rate, and we omit this part of the discussion. Assume, by induction, that we have a
Markov influence system M cycling through states 1, . . . , p, for p ≥ 4. We build another one
with period at least 2p by adding a "gadget" to it consisting of a graph over the vertices 1, 2, 3
with probability distribution (x, y, z) ∈ S. We initialize the system by placing M in state 1 (ie,
1pm in our clock example) and setting x = 1. The dynamic graph is specified by these rules:
1. Suppose that M is in state 1, . . . , p − 1. The graph has the edge (1, 2), which is assigned
probability 1/2, as is the self-loop at 1. There are self-loops at 2 and 3.
16
2. Suppose that M is in state p. If z ≤ 1/2, then the graph has the edge (2, 1) and (1, 3), both
of them assigned probability 1, with one self-loop at 3. If z > 1/2, then the graph has the
edge (2, 1) and (3, 1), both of them assigned probability 1, with one self-loop at 3.
Suppose that M is in state 1 and that y = 0 and z ≤ 1/2. When M reaches state p − 1, then
x = (1 − z)21−p and y = (1 − z)(1 − 21−p). Since z ≤ 1/2, the system cycles back to state 1 with
the updates z ← x + z and x ← y. Note that z increases by a number between 2−p and 21−p.
Since z begins at 0, such increases will occur consecutively at least p2p/4 ≥ 2p times, before x
is reset to 1. The construction on top of M adds three new vertices so we can push this recursion
roughly n/3 times to produce a Markov influence system that is periodic with a period of length
equal to a tower-of-twos of height roughly n/3.
We need to tie up a few loose ends. The construction needs to recognize state p − 1 by
a polyhedral cell; in fact, any state will do. The easiest choice is state 1, which corresponds
to x ≥ 1 (to express it as an inequality). The base case of our inductive construction consists
of a two-vertex system of period p = 4 with initial distribution (1, 0). If x > 21−p, the graph
has an edge from 1 to 2 and a self-loop at 1, both of them assigned probability 1/2; else an
edge from 2 to 1 given probability 1 to reset the system. Finally, the construction assumes
probabilities summing up to 1 within each of the (cid:98)(n−2)/3(cid:99) +1 gadgets, which is clearly wrong:
we fix this by dividing the probability weights equally among the gadget and adjusting the linear
(cid:3)
discontinuities appropriately.
5.2 Chaos
We give a simple 5-state construction with chaotic symbolic dynamics. The idea is to build an
MIS that simulates the classic baker's map.
1
y ←
2(y + 1)
2y
y+1
if y < −1
if −1 ≤ y ≤ 0.
Writing z = (y + 1)/(y − 1), we note that −1 ≤ z < 1 and it evolves according to z (cid:55)→ 2z + 1 if
z ≤ 0, and z (cid:55)→ 2z− 1 otherwise, a map that conjugates with the baker's map and is known to be
chaotic [14].
17
2
0
0
0
0
A =
1
3
if x1 + x2 > x4
1 0 0
1 2 0
0 3 0
0 0 2
0 0 0
0
0
0
1
3
and
B =
for x ∈ S4. We focus our attention on Σ =(cid:8) (x1, x2, x4) 0 < x1 ≤ x4/2 ≤ x2 < x4
check that it is an invariant manifold. At time 0, we fix x4 = 1/4 and x5 = 0; at all times, of
course, x3 = 1 − x1 − x2 − x4 − x5. The variable y:= (2x2 − x4)/(2x1 − x4) is always nonpositive
over Σ. It evolves as follows:
1
1
0
0
0
1
3
0
2
0
0
0
2
0
3
0
0
0
0
0
2
0
else,
(cid:9), and easily
0
0
0
1
3
Acknowledgments
I wish to thank Maria Chudnovsky and Ramon van Handel for helpful comments.
References
[1] Anagnostopoulos, A., Kumar, R., Mahdian, M., Upfal, E., Vandin, F. Algorithms on evolving
graphs, Proc. 3rd Innovations in Theoretical Computer Science ITCS (2012) 149 -- 160.
[2] Avin, C., Kouck´y, M., Lotker, Z. How to explore a fast-changing world (cover time of a simple
random walk on evolving graphs), Proc. 35th ICALP (2008), 121 -- 132.
[3] Barrat, A., Barth´elemy, M., Vespignani, A. Dynamical Processes on Complex Networks, Cam-
bridge University Press, 2008.
[4] Bruin, H., Deane, J.H.B. Piecewise contractions are asymptotically periodic, Proc. American
Mathematical Society 137, 4 (2009), 1389 -- 1395.
[5] Cao, M., Spielman, D.A. Morse, A.S. A lower bound on convergence of a distributed network
consensus algorithm, 44th IEEE Conference on Decision and Control, and the European Control
Conference, Seville, Spain (2005), 2356 -- 2361.
[6] Castellano, C., Fortunato, S., Loreto, V. Statistical physics of social dynamics, Rev. Mod. Phys. 81
(2009), 591 -- 646.
[7] Chazelle, B. Diffusive influence systems, SIAM J. Comput. 44 (2015), 1403 -- 1442.
[8] Chazelle, B. Jiu, Q., Li, Q., Wang, C. Well-posedness of the limiting equation of a noisy consensus
model in opinion dynamics, J. Differential Equations 263 (2017), 365 -- 397.
[9] Condon, A., Hernek, D. Random walks on colored graphs, Random Structures and Algorithms 5
(1994), 285 -- 303.
[10] Condon, A., Lipton, R.J. On the complexity of space bounded interactive proofs, Proc. 30th IEEE
Symp. on Foundations of Computer Science (1989), 462 -- 267.
[11] Coppersmith, D., Diaconis, P. Random walk with reinforcement (1986), Unpublished manuscript.
[12] Denysyuk, O., Rodrigues, L. Random walks on directed dynamic graphs, Proc. 2nd International
Workshop on Dynamic Networks: Algorithms and Security (DYNAS10), Bordeaux, France, July,
2010. Also arXiv:1101.5944 (2011).
[13] Denysyuk, O., Rodrigues, L. Random walks on evolving graphs with recurring topologies, Proc.
28th International Symposium on Distributed Computing (DISC), Austin, Texas, USA, October
2014.
[14] Devaney, R.L. An Introduction to Chaotic Dynamical Systems, 2nd edition, Westview Press, 2003.
[15] Diaconis, P. Recent progress on de Finetti's notions of exchangeability, Bayesian Statistics 3
(1988).
[16] Fagnani, F., Frasca, P. Introduction to Averaging Dynamics over Networks, Lecture Notes in Con-
trol and Information Sciences, 472, Springer, 2018.
18
[17] Frank, T.D. Strongly nonlinear stochastic processes in physics and the life sciences, ISRN Mathe-
matical Physics 2013, Article ID 149169 (2013).
[18] Holme, P., Modern temporal networks theory: a colloquium, European Physical Journal B, 88, 234
(2015).
[19] Holme, P., Saramaki, J. Temporal networks, Physics Reports 519 (2012), 97 -- 125.
[20] Iacobelli, G., Figueiredo, D.R. Edge-attractor random walks on dynamic networks, J. Complex
Networks 5 (2017), 84 -- 110.
[21] Katok, A., Hasselblatt, B. Introduction to the Modern Theory of Dynamical Systems, Cambridge
University Press; revised ed. edition (December 28, 1996).
[22] Kempe, D., Kleinberg, J. Kumar, A. Connectivity and inference problems for temporal networks,
J. Computer and System Sciences 64 (2002), 820 -- 842.
[23] Kolokoltsov, V.N. Nonlinear Markov Processes and Kinetic Equations, Cambridge Tracks in Math-
ematics 182, Cambridge Univ. Press, 2010.
[24] Michail, O. An introduction to temporal graphs: an algorithmic perspective, Internet Mathematics
12 (2016).
[25] Michail, O., Spirakis, P.G. Elements of the theory of dynamic networks, Commun. ACM 61, 2
(2018), 72 -- 72.
[26] Nguyen, G.H., Lee, J.B., Rossi, R.A., Ahmed, N., Koh, E., Kim, S. Dynamic network embeddings:
from random walks to temporal random walks, 2018 IEEE International Conference on Big Data
(2018), 1085 -- 1092.
[27] Perra, N., Baronchelli, A., Mocanu, D., Gonc¸alves, B., Pastor-Satorras, R., Vespignani, A. Random
walks and search in time-varying networks, Phys. Rev. Lett. 109, 238701, 2012.
[28] Proskurnikov, A.V., Tempo, R. A tutorial on modeling and analysis of dynamic social networks, I.
Annual Reviews in Control 43 (2017), 65 -- 79.
[29] Proskurnikov, A.V., Tempo, R. A tutorial on modeling and analysis of dynamic social networks,
Part II. Annual Reviews in Control 45 (2018), 166 -- 190.
[30] Ramiro, V., Lochin, E., S´enac, P., Rakotoarivelo, T. Temporal random walk as a lightweight com-
munication infrastructure for opportunistic networks, 2014 IEEE 15th International Symposium
(WoWMoM), June 2014.
[31] Seneta, E. Non-Negative Matrices and Markov Chains, Springer, 2nd ed., 2006.
[32] Starnini, M., Baronchelli, A., Barrat, A., Pastor-Satorras, R. Random walks on temporal networks,
Phys. Rev. E 85, 5 (2012), 056115 -- 26.
[33] Sternberg, S. Dynamical Systems, Dover Books on Mathematics, 2010.
19
|
1708.02938 | 1 | 1708 | 2017-08-09T07:15:47 | A Novel Formal Agent-based Simulation Modeling Framework of an AIDS Complex Adaptive System | [
"cs.MA",
"nlin.PS",
"q-bio.PE"
] | HIV/AIDS spread depends upon complex patterns of interaction among various sub-sets emerging at population level. This added complexity makes it difficult to study and model AIDS and its dynamics. AIDS is therefore a natural candidate to be modeled using agent-based modeling, a paradigm well-known for modeling Complex Adaptive Systems (CAS). While agent-based models are also well-known to effectively model CAS, often times models can tend to be ambiguous and the use of purely text-based specifications (such as ODD) can make models difficult to be replicated. Previous work has shown how formal specification may be used in conjunction with agent-based modeling to develop models of various CAS. However, to the best of our knowledge, no such model has been developed in conjunction with AIDS. In this paper, we present a Formal Agent-Based Simulation modeling framework (FABS-AIDS) for an AIDS-based CAS. FABS-AIDS employs the use of a formal specification model in conjunction with an agent-based model to reduce ambiguity as well as improve clarity in the model definition. The proposed model demonstrates the effectiveness of using formal specification in conjunction with agent-based simulation for developing models of CAS in general and, social network-based agent-based models, in particular. | cs.MA | cs | A Novel Formal Agent-based
Simulation Modeling
Framework of an AIDS Complex
Adaptive System
Amnah Siddiqa1, Muaz Niazi2,*
1Research Centre for Modeling and Simulation, National University of Sciences and Technology,
Islamabad, Pakistan
2Office of Research, Innovation & Commercialization, Bahria University, Islamabad, Pakistan
*Corresponding author [email protected]
Cite as:
Siddiqa, A., & Niazi, M. (2013). A Novel Formal Agent-Based Simulation Modeling Framework of an AIDS
Complex Adaptive System. International Journal of Agent Technologies and Systems (IJATS), 5(3), 33-53.
doi:10.4018/ijats.2013070103
http://www.igi-global.com/article/a-novel-formal-agent-based-simulation-modeling-framework-of-
an-aids-complex-adaptive-system/97688
Abstract
HIV/AIDS spread depends upon complex patterns of interaction among various
sub-sets emerging at population level. This added complexity makes it difficult to
study and model AIDS and its dynamics. AIDS is therefore a natural candidate to be
modeled using agent-based modeling, a paradigm well-known for modeling Complex
Adaptive Systems (CAS). While agent-based models are also well-known to effectively
model CAS, often times models can tend to be ambiguous and the use of purely text-
based specifications (such as ODD) can make models difficult to be replicated.
Previous work has shown how formal specification may be used in conjunction with
agent-based modeling to develop models of various CAS. However, to the best of our
knowledge, no such model has been developed in conjunction with AIDS. In this
paper, we present a Formal Agent-Based Simulation modeling framework (FABS-AIDS)
for an AIDS-based CAS. FABS-AIDS employs the use of a formal specification model in
conjunction with an agent-based model to reduce ambiguity as well as improve clarity
in the model definition. The proposed model demonstrates the effectiveness of using
formal specification in conjunction with agent-based simulation for developing models
of CAS in general and, social network-based agent-based models, in particular.
Keywords: Agent based modeling, AIDS, Complex Adaptive System, sub-
populations, complex networks, formal specification
1. Introduction
As noted in reports by the World Health Organization (WHO) (WHO, 2011), the Acquired
Immunodeficiency Syndrome (AIDS) contributes 0.8% (i.e., approximately 34 million people) to the
global burden of diseases. While the mode of spread of AIDS and its prevention strategies are quite
well-known, a clear indication of the complexity of HIV/AIDS spread can be noted by its continual
prevalence in regions globally. The causative agent of AIDS is a retrovirus termed as the "Human
Immunodeficiency Virus (HIV)". WHO reports note that areas such as Sub-Saharan Africa have nearly
1 in every 20 adults (4.9%) living with HIV thereby accounting for 69% of the people living with HIV
worldwide.
The variation in the emergent pattern of the AIDS/HIV pandemic in global populations is
primarily due to the numerous and complex interactions of a large number of heterogeneous
entities or agents1 (Dean & Fenton, 2010) and populations. One of the pioneers of the theory of
"Complex Adaptive Systems" (CAS), John Holland (JH Holland, 1992) forecasted that AIDS spread can
be modeled as a CAS; systems which are traditionally associated with the observance of unusual
1 "agent" being used here as a technical term, which can refer to a person, a group, a cell or even
a particular country or sub-population (M. Niazi & A. Hussain, 2011).
patterns, termed as "emergent behaviors" (Boccara, 2010; Muaz A Niazi & Amir Hussain, 2011).
Considering AIDS as a CAS would imply that we can expect HIV/AIDS to exhibit emergent
phenomena as a result of non-linear interactions of its constituent agents. However, to the best of
our knowledge, although such patterns are expected and predicted by CAS researchers (such as
Holland), such patterns have not been modeled or highlighted much in existing literature which has
traditionally focused on studying AIDS primarily from the healthcare perspective.
In this paper, the research question addressed is as to how a formal specification model can
be coupled with an agent-based model of AIDS/HIV spread for examining HIV allowing for the
discovery of emergent patterns. We use existing techniques such as studied in previous literature
(M. A. Niazi & A. Hussain, 2011a) to develop a Formal Agent-based Simulation Framework of
HIV/AIDS spread (FABS-HIV). The structure of the rest of the paper is as follows: Section 2 presents
background. In section 3, first a formal specification framework of the HIV/AIDS spread "FABS-HIV" is
described followed by the description of an agent-based model developed in line with this formal
specification. Section 4 presents implementation details of the agent-based model as well as
simulation experiments, results and discussion. In the final section, the paper is concluded and
future research directions are outlined.
2. Background
This section first provides background and related work about modeling HIV/AIDS spread,
formal specification and Agent-based.
2.1. HIV Disease Spread
The HIV/AIDS epidemic has received considerable interest in the last two decades. With specific
journals dedicated to different aspects of studying HIV/AIDS, there is considerable information
available on both the actual virology of the epidemic as well as on infection spread. In addition, it is
well-known how protection from HIV/AIDS can be achieved. However, in spite of this considerable
amount of literature and wide-spread knowledge, the AIDS epidemic still poses a serious problem
with worldwide prevalence.
Complex dynamics in AIDS were suggested by Jones and Handcock for disease spread associated
within sexual social networks (Jones & Handcock, 2003). A key problem in investigative studies of
HIV/AIDS is in the collection of real-time data of disease spread due to challenges such as the taboo
nature of the way it spreads in populations and subsequently how HIV transforms into AIDS and the
hitherto unknown exact factors which causes the HIV virus to behave in a seemingly unpredictable
manner (Chin & Kroesen, 1999). Thus, while not all cases of HIV spread can be monitored easily,
post-infection data can be considerably scarce (Allen et al., 2003). This can result in developing
results based on speculations and are limited by the extent of the sampled data. Thus, even the
guidelines for determining non-probabilistic sample sizes are virtually nonexistent in most affected
countries, mentioned by Guest et al. (Guest, Bunce, & Johnson, 2006). Whereas Pollack et al.
(Pollack, Osmond, Paul, & Catania, 2005) note that risk behavior can be overestimated in the case it
is defined broadly. In other words, while the existing schemes of data collection have given
interesting results, there is actually a possibility of considerable room for further exploration and
investigation in how different studies are practically structured and conducted.
2.2. Formal Specification
Formal specification is a type of formal methods employed for mathematical modeling of real-
world systems (Hall, 1990). While modeling a system, it is necessary to ensure that all the system
components have been incorporated at the desired level of abstraction. Thus it is usually demanding
and very effective strategy to use formal methods for the correct implementation of the system
under investigation. The use of formal specification in the design phase of system modeling is
suggested by Bowen (Bowen, 1995) for earlier debugging of the system. "Z" is an ISO standard
formal specification language developed in 1970s at Oxford University (Woodcock & Davies, 1996).
It has been used to model large scale real systems including IBM CICS (Houston & King, 1991) and
cardiac pacemaker (Verdasca et al., 2005). Formal semantics of Z specification
includes
mathematical notations based on set like and first order predicate calculus. These notations are
represented using formal structure known as "schemas" in Z specification.
In this paper, we have employed a Z-based formal specification model for HIV/AIDS modeling as
part of FABS, an established modeling framework which has previously been shown to effectively
model and analyze different types of CAS such as sensing emergence such as flocking near Wireless
Sensor Networks (M. A. Niazi & A. Hussain, 2011a, 2011b) and emergent behavior of citations of
scholars as they progress in their research careers (Hussain & Niazi, 2013). Further to the
justification of using formal specification is the fact that in spite of various modeling methods
applied to investigate the HIV/AIDS spread dynamics there is lack of effective monitoring of the AIDS
disease perhaps because of missing specification related to the modeling of CAS in context of
HIV/AIDS spread. The reason behind this dilemma needs to be investigated through a modeling
paradigm which takes into account a formal specification model of the system as offered by Z formal
specification language.
2.3. Agent-based Modeling
Agent based modeling (ABM) is a simulation paradigm which has close ties with actual scientific
experiments thus making it a good technique for evaluation different paradigms (Li, Brimicombe, &
Li, 2008). What makes ABM unique for modeling complex adaptive systems is the way an ABM is
typically designed. ABMs have been used in domains as diverse as Biological Sciences (Bailey,
Lawrence, Shang, Katz, & Peirce, 2009; Siddiqa et al., 2009), disease spread models such as for
Dengue (Jacintho, Batista, Ruas, Marietto, & Silva, 2010), Social Sciences (Dennard, Richardson, &
Morçöl, 2008), Economics (Cartier, 2004) and Computer Sciences (Muaz Niazi & Hussain, 2010; M. A.
Niazi & A. Hussain, 2011a). ABM has been termed as a revolution in the prestigious journal
"Proceedings of the National Academy of Sciences" (Bankes, 2002). Holland also notes that CAS can
be better studied by the use of agent-based models (John Holland, 2006). A recent review by Niazi
(M. A. Niazi, 2013) presents literature about modeling CAS using agent-based and complex-network
based models.
3. Model Development
This section includes description of informal model concepts needed for modeling FABS-HIV. It
further includes description of formal specification model developed using Z and subsequently
transformed working agent-based model.
3.1.
Informal Description of Model Concepts
Modeling of HIV/AIDS spread involves dealing with heterogeneous population subsets with complex
intra and inter communication patterns. The infection starts in a closed population sub-group2
sharing common disease spreading behavior usually termed as high risk category. These high risk
categories are further classified according to the mode of infection transmission used by them for
example Injecting Drug Users (IDU) share syringes for intravenous drug intake and sex workers
communicate through physical interaction as a mode of infection transmission etc. These sub-groups
become a core source to transmit the infection to other population sub-groups which might be
intermediate or secondary contractors of the infection. The intermediate individuals directly
contract the disease from high risk individuals and in turn become a source to spread it to the
population strata which is not directly exposed to any of the sources of HIV infection (thus termed
secondary). Thus a complete picture of HIV/AIDS spread system demand the modeling of both intra
and inter communication pattern involved in these population sub-groups instead of individually
modeling each category as is the case with previous studies.
We in this study will limit our model to a single high risk category i.e., Female Sex Workers (FSWs) for
simplicity's sake. The FSWs interact directly with their clients. Their clients in turn interact with their
social partners belonging to low risk population sub-group2 which are further divided into two
categories. These two categories include both the socially committed and non-committed partners.
The interaction between individuals belonging to any of these sub-groups might or might not be
using preventive measures, another important factor in determining the shape of infection
pandemic. This informal description enlists the key concepts to explain how infection actually travels
in population. The same pattern is replicable for other high risk categories which is the reason for
limiting our model to single high risk category for conduction of exploratory analysis of interaction
pattern.
3.2.
The FABS-HIV Framework
The FABS modeling framework involve two phases of model construction. First a formal model
of HIV/AIDS spread using a mathematical framework originally based on a formal specification
language Z is constructed. This formal specification model aims to validate all the informal needs of
the system under investigation at the selected level of abstraction besides providing the
unambiguous representation of concepts. Next an agent based model based on the formal
specification framework is developed as a proof of concept. Subsequent simulation experiments
performed by means of setting a 100% condom usage in populations (as control setting) were used
for validation. Our results based on 95% confidence interval demonstrate the existence of complex
2 "Sub-group" refers to a subset of people in the population which share commonalities in terms of
disease spreading behavior
emergent interaction pattern observed as backflow of HIV infection in FSWs population based on
coupling habits of their clients. Thus, in general, our experiments demonstrate why there exists a
need for a realistic division of sub-populations in HIV studies in line with the local population sexual
habits and behavior implying AIDS/HIV spread to be modeled as CAS. Our results suggest that
instead of simply dividing populations between high risk and low-risk categories, studies should
examine populations based on topological structure exhibited within and among the sub-
populations. Following a more logic division in HIV studies could thus assist in a better evaluation
and hopefully development of better mechanisms for controlling the spread of the HIV/AIDS spread
worldwide.
As such the formal agent-based simulation framework consists of three types of parameters i.e.,
hard-coded, programmable and measured. The hard-coded parameters allows
for
the
representation of key concepts based on the facts associated with the system, programmable
parameters allow for state initializing of the system and measurable parameters allow for the
subsequent monitoring of behavior dynamics.
The model is described using some key concepts required for modeling of FABS-HIV. The model dealt
with a population consisting of various sub-populations that were categorized on the basis of risk
behavior. These included one HIV/AIDS high risk population; i.e., Female Sex Workers (FSWs) in this
case and three seemingly low risk populations: 1) heterosexual men (termed "primaries") being the
people who have occasional interaction with the FSWs as their clients), 2) heterosexual women
(termed "secondaries") since they do not have any direct interaction with the FSWs and acquire
infection via primaries being their socially committed partner, and 3) heterosexual women who are
not committed partners of the primaries, termed exsecondaries. The next key concept is to model
the interactions between all these sub-population categories. The interactions required for modeling
of FABS-HIV can be concluded as follows:
a) The interaction between primary and secondary
b) The interaction between primary and exsecondary
c) The interaction between primary and FSWs
All these interactions are based on the moral behavior of primary population which can be
described using two more concepts i.e., commitment level and adopting preventive measures such
as condom usage. A high commitment level of a primary with its socially committed partner is
indicative of keeping away from all other types of partnerships and in turn the infection itself and
vice versa. Similarly, using the preventive measures like condom usage is also an indicator of partially
staying away from the infection. We assume some basic knowledge of Z notation for simplification.
3.2.1. Sets
Sets have been used in the model to declare the user defined types in the model. We started by
defining the given sets.
The set "PERSON" is the set of all persons in a given population:
[PERSON] == {x:ℕx>=0 }
Each person in the population carries different attributes regarding population sub-group
category, infection status, gender, and their social status for coupling with their partners. Each of
these concepts was adequately defined by associating different attributes with each PERSON entity.
The free type definition "PERSONTYPE" depicts the different types of individuals in a given
population; i.e., fsw, primary, secondary and exsecondary. A PERSON entity can belong to any one
of these types. Formal definition of PERSONTYPE is as follows:
[PERSONTYPE] == {fsw, primary, secondary, exsecondary}
The free type "PERSONSTATE" was used to declare the health state of the individual which could
either be infected or uninfected:
[PERSONSTATE]=={infected, unifected}
The next free type "PERSONGENDER" was defined as a means of declaring the gender of the
individual which could either be male or female:
[PERSONGENDER]=={male,female}
The next set "INFECTED" defined all the infected individuals. It belongs to both programmable
and measurable component of the model. It is used to monitor the system state encompassing
infected population only, at a given time point:
[INFECTED]=={n:INFECTED0<n<#persons∧ personstate=infected }
where #persons is a count of total no of persons in the population.
Likewise the set "PARTNER" is the set of all partners of a person. The persontype of partner was
restricted to either primary or secondary since the partner is used to model the relationship of
socially recognized commitment between these two entities:
[PARTNER] = = {x: PARTNER persontype= primary ∨ persontype= secondary}
The set "COUPLE" was used to declare the pairs of sexually interacting individuals in the
population. Each couple models the interaction of primaries with any of the other female sub-
populations. Therefore, the predicate of the couple was restricted to at least one primary:
[COUPLE]== {x,y:COUPLE x∈primary ∨y∈primary}
The set "𝔽" was defined as the set of natural numbers greater than or equal to 0 and less than or
equal to 100. The natural numbers between 1-100 provided a scale for adjusting user-desired
variables like commitment and condom usage as a property of the primary population. It was also a
programmable parameter:
[𝔽] = = { x: ℕ 0<x≤100}
Several user-defined global variables were declared with axiomatic declaration as programmable
parameters. The variables "maxprimary, maxsecondary, maxfsw, maxinfectedfsw, maxexsecondary,
and tobecoupled" were used for declaration of the number of primaries, secondaries, FSWs, the
infected FSWs, the exsecondaries, and the number of pairs to be formed between primaries and
secondaries, respectively. These global variables were provided for customized modeling of the
initial population:
maxprimary:ℕ
maxsecondary:ℕ
maxfsw:ℕ
maxinfectedfsw:ℕ
maxessecondary:ℕ
tobecoupled:ℕ
3.2.2. State Schemas
A state schema is a special structure used for the declaration of the system state. Each system
state requires a specific set of objects, variables, entities and functions to ensure that all system
requirements are being met. The states of FABS-HIV might be summed in three concise
declarations:
a) State declaration for each population sub-group
b) State declaration of socially committed partnerships
c) State declaration for coupling action based on heterogeneous male population behavior
As a first step, the state schemas for all the sub-populations were declared to accommodate for
their heterogeneity. The first population subtype object declared was Fsw. Every Fsw was declared
with a set of attributes; i.e., persontype, gender and state. The predicate function restricts the
person type to Fsw, and the gender to female:
Fsw
type : PERSONTYPE
state : PERSONSTATE
gender : GENDER
persontype = fsw
gender = female
The next population subtype defined was Primary. The predicate function restricts the person
type to Primary, gender to male and declares it with a socially committed partner which must belong
to Secondary subtype:
Primary
type : PERSONTYPE
state : PERSONSTATE
gender : GENDER
apartner : PARTNER
persontype = Primary
gender =male
apartner= Secondary
The next sub-population declared was Secondary. The predicate function restricts the person
type to secondary, gender to female and declares it with a partner who must belong to primary
subtype:
Secondary
type : PERSONTYPE
state : PERSONSTATE
gender : GENDER
apartner : PARTNER
persontype = Secondary
gender = female
apartner = Primary
Exsecondary is an object of type Secondary. This category is used to depict the females who
belong to low risk population with no socially committed partner at given time point. However they
might be involved in occasional sexual encounters. The predicate function restricts the exsecondary
to a null partner:
ExSecondary
exsecondary : Secondary
partner= Null
After defining all the population sub-groups adequately, we formulated the state schema
Partners to store the pointers of the partners as a couple. This schema tends to create a pair of
socially selected partners necessarily should be a primary (a male) and a secondary (a female). The
aim of this object is to only declare the socially committed partners in a population. It has nothing to
do with the coupling (sexual interaction) between these partners yet. The term coupling refers to
the sexual act between heterosexual individuals.
The set 𝔽PARTNER depicts a set of finite elements. Each element of this set consists of two
entities, thus forming a pair. Each pair consists of two elements which holds pointer to each other.
Each created pair can be called by using notation 𝔽PARTNER. However, the elements of each pair are
accessible using notation PARTNER. Three conditions were checked in the predicate. First condition
checked for both elements of a pair to either belong with Primary or Secondary. Second condition
monitored that no input entity is reusable i.e., an individual who has been used to form a pair is not
used again in another partnership. Third condition implied that the pair formed uses the user-
defined range for input values. The numbers maxprimary and maxsecondary depict the global
variables used to declare the user defined values for the maximum number of primaries and
seconadries, respectively. The total number of pairs to be formed was declared using a global
variable tobecoupled.:
Partners
x,y: PERSON
apair: 𝔽PARTNER
(x∨y)∧(Primary∨Secondary)
x∧y∉PARTNERS
#primary ≤ maxprimary
#primary ≤ tobecoupled
#seconadry ≤ maxsecondary
#secondary ≤ tobecoupled
The next state schema Link modeled the heterosexual sexual interaction between two
individuals. Two variables are needed to model such an interaction. In the interaction, to model a
male population one would always need a primary and to model a female population, a person
could belong to any of the female sub-populations. Secondly the primary would always go for its
own secondary partner from secondary sub-set. The purpose was achieved by calling three variables.
Two of these were the elements of a pair created through Makepartner schema. It was depicted by
notation x:PARTNER and y:PARTNER. Another person variable was used to model rest of two female
subpopulations and restricting it to belong to either of two female populations i.e., fsw and
exsexondary in the predicate. The variable acouple held the pointers to the coupling individuals in
the pair. Two variables commitment and threshold-commitment declares the user-defined value of
commitment level associated with primary population and a threshold value for implementation of
commitment preference rule:
Link
p1 : PARTNER
p2:PARTNER
p3:PERSON
acouple : 𝔽COUPLE
commitment, threshold-commitment : 𝔽
P3 ∧ ( fsw ∨ exsecondary)
commitment< threshold-commitment<=commitment
The next step is initialization of the declared states to describe the system state when it is first
started. The initializing elucidates the working of concepts as hard-coded parameters i.e., population
sub-groups representation and programmable parameters i.e., partnership declaration and coupling
frequency etc.
Each fsw was a heterosexual female who was initially uninfected. The state schema fsw was
initialized as InitFsw with type fsw, state uninfected, and gender female:
InitFsw
Fsw
type = fsw
state =uninfected
gender = female
Each primary was a heterosexual male who was initially uninfected with no socially committed
partner. The state schema primary was initialized as InitPrimary with type primary, state uninfected,
gender male, and a null partner:
InitPrimary
Primary
type = primary
state= uninfected
gender =male
apartner = NULL
Each secondary was a heterosexual female who was initially uninfected with no socially
committed partner. The state schema secondary was initialized as InitSecondary with type
secondary, state uninfected, gender female and a null partner:
InitSecondary
Secondary
type =secondary
state = uninfected
gender = female
apartner = NULL
Each exsecondary was the element of secondary population with no socially committed partner.
The state schema exsecondary was initialized as InitExsecondary with type secondary, state
uninfected, gender female and a null partner:
InitExsecondary
Exsecondary
type =secondary
state = uninfected
gender = female
apartner = NULL
The partner schema was initialized as InitPartner. The partnership between the primary
(declared as variable x) and secondary (declared as variable y) sub-populations as socially committed
partners was initialized in this schema. The number of primaries and secondaries to be declared as
socially committed partners was chosen through the user-defined value of global variables
maxprimary and maxsecondary. The maxprimary and maxsecondary in this schema were set to
#maxprimary and #maxsecondary respectively. There was no value stored in the couple yet;
therefore, the value for acouple was set to an empty set(∅).
InitPartner
Partner
x=Primary
y=Secondary
maxprimary = #maxprimary
maxsecondary =#maxsecondary
tobecoupled=#tobecoupled
acouple=∅
The Link schema was initialized as InitLink schema. Initially, each variable was set to null as no
coupling was performed.
InitLink
Link
p1 = Null
p2= Null
p3=Null
acouple =∅
commitment=Null
threshold-commitment=Null
Next all the global variables were declared using axiomatic declaration. The user-defined
variables maxprimary, maxsecondary, maxfsw and maxinfectedfsw were declared to be positive
integers. The variables coupledprimary and coupledsecondary were declared to be zero initially as
these variables will be updated by their respective usage in further operations.The threshold-
condomusage and threshold-commitment were declared to be any value between 0 and 100
inclusive.
maxprimary:ℕ
maxsecondary:ℕ
maxfsw:ℕ
maxinfectedfsw:ℕ
maxessecondary:ℕ
tobecoupled:ℕ
maxprimary => 0
maxsecondary=>0
maxfsw=>0
maxinfectedfsw=>0
maxexsecondary=>0
tobecoupled=0
3.2.3. Operation Schemas
An operation schema is the special structure in which the change in the states of the system is
illustrated. It performs all the functions on all components using predicate specified conditions. The
results of operation schemas produce output of measured parameters needed for system analysis
and monitoring.
The first operation of FABS-HIV was to initialize the population. For this purpose, a schema
SetupInitialPopulation was defined. The schema used several global variables to deploy user-desired
inputs. The global variables used in this schema were: maxprimary, maxsecondary, maxexsecondary,
maxfsws and maxinfectedfsws taking input for the maximum no of primaries, secondaries, fsws, and
the infected fsws in the initial population:
SetupInitialPopulation
primaries: PRIMARY
secondaries : SECONDARIES
exsecondaries:SECONDARIES
fsws: FSWS
infectedfsws: FSWS
primaries′ = #maxprimaries
secondaries′= #maxsecondaries
exsecondaries′= #maxexsecondaries
fsws′= #maxfsws
infectedfsws′= #maxinfected fsws
The aim of the next operation was to declare the primaries and secondaries as social partners of
each other. The operation schema was declared as "MakePartners". The operation on the Partner
schema is performed which has been included with a delta(Δ) sign to show the change in it. Two
objects of type Primary and Secondary were taken as inputs. Both of these individuals were checked
for the condition that the count of coupled individuals in the simulation was still required or not
according to the initialized values. They were additionally checked for not being the elements of
PARTNER set. The total number of pairs to be formed was checked using several global variables i.e.,
maxprimary, maxsecondary and tobecoupled through predicate conditions. The fulfillment of all the
conditions would result in making the primary and secondary partners of each other:
MakePartners
ΔPartner
x? : Primary
y? : Secondary
x=Primary
y=Secondary
x∉PARTNER
y∉PARTNER
#primary < #maxprimary
#primary<#tobecoupled
#secondary< #maxsecondary
#secondary<#tobecoupled
apair′=𝔽PARTNER ∪{x?, y?}
The next operation schema Coupling performed operation of coupling on the Link schema. A
primary will always go for coupling with his partner if the commitment is greater or equal to the
threshold value. Otherwise, it will go for the coupling with FSW. This interaction was modeled as a
binary relation defined as acouple between the selected objects through commitment preference
rules:
Coupling
ΔLink
p1?, p2?:PARTNER
p3?:Fsw
commitment?, threshold-commitment? : 𝔽
p3?=Fsw
commitment?>=threshold-commitment?
acouple′ = (p1? , p2?) ∪ acouple
commitment?<threshold-commitment?
acouple′ = (p1?, p3?) ∪ acouple
The next schema ApplyCondomUsage performed an operation on selected coupling pair from
Coupling schema as an input to apply the condom usage rule. This operation modeled the infection
spread in a given population based on the frequency of preventive measure deployed by the
primarys. If the condomusage was less than the threshold value, the infection would spread and vice
versa. The predicate function would check the state of both the persons in the pair and infect the
other if one was infected; otherwise, the state would remain the same:
Applyingcondomusage
ΔCoupling
x?: COUPLE
y?:COUPLE
condomusage?, threshold-condomusage?:𝔽
condomusage? >=threshold-condomusage?
x′=x?
y′=y?
condomusage?< threshold-condomusage?
x?∉INFECTED
y?∉INFECTED
x′={x?} ∪ INFECTED
y′= {y?} ∪ INFECTED
3.3.
Agent-based Model of FABS-HIV
This section formally describes the translated agent-based model developed as a proof-of-
concept based on the formal specification presented for FABS-HIV in order to study AIDS/HIV as a
CAS. The model was developed using NetLogo (Wilensky, 1999), a freely available agent-based
modeling tool that has previously been used extensively to model complex phenomena (M. A. K.
Niazi, 2011).
3.3.1. Agents
Based on the formal specification model, we modeled three sub-populations: a high risk
category i.e., FSWs, their clients termed primaries, and the partners of the primaries termed
secondaries (Fig. 1). To differentiate between them, a color code was used where yellow color
indicated FSWs, blue indicated primaries, pink indicated secondaries and red indicated HIV/AIDS
infected individuals.
3.3.2. Interactions
The model used "links" to represent two people engaged in a sexual relationship. These links
created couples. There were essentially two kinds of couplings envisioned in this model:
1) an interaction between FSWs and primaries
2) an interaction between primaries and secondaries. These again were of two types:
between primaries and their socially selected secondary partners (e.g. married partner)
between primaries and their other secondary partners (which are perhaps not coupled with
any other primary)
To perform verification of the model, each time there was a coupling, a link was shown on the
screen between the two agents as a connecting line as (Fig 1).
3.3.3. Model Input Parameters (Agent Behavior)
Several input parameters were provided as controls in the model to adjust the behavior of the
agents. The number of sub-populations could be adjusted through the sliders on the graphical user
interphase (GUI) of the model (Fig 2). The differentiation between secondaries of the two types
could be adjusted with the help of another slider named "ex-sec", where ex-sec were the low risk
heterosexual women other than the long-term selected partners of the primaries. The coupling
between primaries and secondaries was controlled through a coupling variable based on per-month
frequency of interactions (coupling) and the coupling between the primaries and FSWs was
controlled through a variable adjusted according to the average client coupling recorded for the
FSWs (Fig. 2). Two key confounding parameters; i.e. the commitment level of the primaries and the
condom-usage among the primaries was used to study their impact on the spread of HIV/AIDS. The
values of these parameters could also be adjusted according to the sliding scales provided in the
model (Fig. 2). All the user-desired adjustable parameters were chosen on the basis of reported
statistical parameters used for empirical data collection of HIV/AIDS spread(UNAIDS, 2005)
(UNAIDS, 2010).
3.3.4. Model Output Parameters (Counters)
All changes occurring in the model are time-dependant. Several output monitors including plots
and counters (display screens with relevant numbers) were used for displaying the updated variable
state of the agents at a given time. The estimates of the number of infected FSWs, their direct
partners and secondaries were displayed as output counters. Furthermore, the model also provided
estimates of the number of non-committed-secondaries (the number of secondaries used in the
simulation which were not committed with any of the primaries), non-committed-infected-
secondaries (the number of infected non-committed secondaries), and the total-infected population,
a number that gave the overall picture of the infection spreading in the population. More
importantly, the model also displayed the numbers of FSWs which were back-infected from infected
primaries, FSW-back-infected. In other words, these infections resulted due to the interaction of
infected primaries with non-infected FSWs, a value that demonstrates the effects of the existence of
a complex hidden network for HIV back-flow. Finally, the number of back-infections of primaries
from secondaries; i.e., primaries-back-infected was also displayed individually. The infection curve
for all sub-populations was displayed through a line plot on ABM screen.
4. Results and Discussion
In this section, we first give an overview of the implementation details followed by model
calibration. Finally, we present experiments and discussion of results.
4.1.
Implementation details
The desired input parameters were incorporated using input scales provided on the model GUI.
Next the setup button was used to initialize the model based on adjusted parameter settings. Time
control provided to run the simulation allows performing it either indefinitely or for a specific time
period. These controls were provided in the form of two "go" buttons, which allowed the control of
the simulation runs in respective manners.
4.2.
Model Calibration and Methods
Our model allows for calibration of certain parameters using the real data which include the
number of sub-groups of agent population, the coupling frequency of primaries and FSWs, condom
usage, infection rate, and time period. We performed verification of our model using two
confounding parameters i.e., commitment level of primaries and variation in condom usage among
the primaries.
Variation in commitment level of the primaries: we experimented with a 0%
commitment level since it would mean that the primary agent had no regard for his
partner regarding sexual practice in regular life, while a 100% commitment level
meant that the agent was totally committed to the partner. The value of percent
commitment level was adjusted from 0 to a commitment level of 100 percent in
increments of 20.
Variation in condom-usage among the primaries: We experimented with a 0%
condom-usage level since it would mean that the infection would not propagate at
all while a 100% condom-usage level meant that the infection would propagate
exponentially. The value of percent commitment level was adjusted from 0 to a
commitment level of 100 percent in increments of 20.
Each point in all experiments was tested in 50 individual simulations. Every time tick represents
one month as the smallest collective time unit. The simulation results were rendered in a graphical
format for easy understanding. An error line function was used to get a clear picture of the
minimum, maximum and average values of the data (to a 95% confidence interval) to observe
various trends (Figs. 3 and 4).
To ensure realism in the simulations, the input data was based on real sources (UNAIDS, 2005)
(UNAIDS, 2010) for Pakistani population except the coupling parameter which was analyzed by
incrementing it through equally spaced variation intervals. The results obtained from the various
experiments conducted below thus reflect the trends found in the Pakistani population.
4.3.
Experiment no 1: Test of Variation in the Commitment Level
The first simulation experiment tested the effect of commitment level of the primary partners of
FSWs on the spread of infection among the various agents being tested. The commitment level of
the primary population was studied since we felt that their actions could seriously affect the form
and shape that an HIV epidemic could take within a population.
Simulation results, while varying the commitment level among the primaries, revealed that the
average infection among FSWs increased if the commitment of primaries with the secondaries went
down (Fig. 3A). On the other hand, when the commitment level of the primaries increased with
their partners, the infection among the primaries went down drastically by nearly 50% (Fig. 3B). The
results in Fig. 3C show that the infection in FSWs could actually be significantly increased due to a
lack of commitment level among the primaries with their partners, revealing that the infection was
actually due to a back-flow from the primaries to the FSWs. In Fig. 3D, the same trend was observed
for back-infected primaries, revealing the possibility that it was not the infected FSWs passing the
disease to the primaries as they were quite few in number; rather, the primaries could actually be
mostly "back-infected" from the secondaries since the infections went significantly down with an
increase in the commitment level. The results from Fig. 3, panels A and C confirm the existence of
back-flow emergence patterns involved in the HIV/AIDS spread which could not be calculated
through traditional methods and thus prove FABS-HIV to be presentable as an adequate modeling
framework for analyzing a CAS like HIV/AIDS spread. Interestingly, the rate of infected secondaries
were seen to be very low when primaries had no regard of commitment perhaps because of high
coupling rate with FSWs instead of their partners which gradually increased with higher commitment
levels(Fig. 3E). The infection rate within the secondaries decreased gradually after 50 percent
commitment level, as expected. This could perhaps be owing to the lesser infection flow in the
secondaries because of low infection rates among primaries due to higher commitment and lesser
couplings with the FSWs (Fig 3E). Fig. 3F reveals the overall trend of varying commitment level
which shows up as normal distribution in the form of a bell-shaped curve.
4.4.
Experiment no 2: Test of Variation in Condom Usage
Next, the variation in condom usage on infection rates was studied among the various actors
using simulation modeling with the assumption that any interaction using condoms would not
propagate infection. Thus, condom usage here assumed proper usage without any tears or
exchange of fluids. The infection rate obtained by varying commitment level was plotted against the
different types of agents to reveal trends in each population as before.
It was observed that infections in FSWs could be significantly reduced by nearly 50% if the
condom usage of primaries was increased incrementally (Fig. 4A). With a 100% condom usage, the
simulation experiment could not stop, verifying the program. The same effect was observed on the
back-flow of infection to the FSWs (Fig. 4C), suggesting the possibility of infection from back-flow as
mentioned above.
Effect of increase in condom usage among the primaries revealed a similar effect on infection as
in FSWs. However, the rate of decrease in infection rates initially among the primaries was observed
to be slower than that observed among the FSWs (Fig. 4B) and the same was observed with the
back-flow of infection to the primaries from the secondaries (Fig. 4D)
Fig. 4E reveals the effect of variation of condom usage on the overall number of infected
secondaries. As expected, with a 100% condom use, the infections went to zero, demonstrating the
verification of the simulation program. Finally, the effect of increase in condom use on total number
of infections revealed a similar trend (Fig.4F). In all these calculations, the number avg-client-
pmonth (Average Transactional Clients per month for FSWs) was also taken from the self-reported
data.
4.5. Correlation with Related Work
While, to the best of our knowledge, there is no other work which presents a FABS in the area of
HIV/AIDS, in this section we present a comparison of our presented work with related work.
A review of AIDS propogation modeling using agent-based model has previously been presented
in (Tirado-Ramos & Kelley, 2013). One of the earliest agent-based model of HIV spread among
injecting drug users (IDUs) was conducted by Atkins et al. in 1996 to simulate the spread of HIV from
an index case to other IDUs using contaminated needles (Atkinson). Newer agent-based models
consider network topology such as geographic isolation, social norms and also sexual taboos such as
by Osgood et al. (Osgood, Moavenzadeh, & Rhee, 2006). Such studies have shown that depending
upon the mode of transmission of the disease agents, the epidemic can take on different profiles. As
such, a variety of intervention and policy decisions may have to be used to counter the effects of
epidemic transmission such as those observed in Taiwan, Botswana, India, and South Africa
(Nagelkerke et al.). Teweldemedhin et al. (Teweldemedhin, Marwala, & Mueller, 2004) use an agent-
based modeling approach to model HIV spread in populations. Alam et al. (Alam, Meyer, & Norling,
2006) use an agent-based modeling approach to study the impact of HIV/AIDS in the context of
socio-economic stressors. Mei et al. (Mei, Sloot, Quax, Zhu, & Wang, 2010) use agent-based
networks to explain HIV epidemic in homosexual males in Amsterdam. Nagoski (Nagoski, 2006)
presents an agent-based disease diffusion model to study heterogeneous sexual motivation. All
these models are encompassing small set of agent/entities and the interactions as well, thus not
employing the full potential of ABM to model a CAS. A previous paper on modeling HIV/AIDS as a
CAS has been presented by Niazi et al. in (MuazA Niazi, Siddiqa, & Fortino, 2013).
Other previous studies based on formal methods include a fuzzy mathematical model developed
by Morio et al. (Morio et al.) to analyze the heterosexual HIV transmission in Japan. A mathematical
framework of concurrent partnerships on the transmission of HIV has been presented by Morris and.
Kretzschmar in (Morris & Kretzschmar, 1997). Other mathematical models developed for the
understanding of the HIV/AIDS epidemic and its transmission have been reviewed by Boily et al. in
(Boily, Lowndes, & Alary, 2002). These models however lack a formal specification or modeling
HIV/AIDS spread as a CAS.
In short, to the best of our knowledge, previous work either did not expand the use of agent-
based modeling to the extent of modeling AIDS as a CAS or else if models were presented, they did
not employ the use of formal specification. Likewise, previously presented models using formal
methods or mathematical models have not connected these models to agent-based models. The
current paper can be considered as extension of previous work on AIDS by means of developing a
formal specification model in addition to expanding agent-based modeling to study emergent
patterns in AIDS spread.
5. Conclusions & Future Work
In this paper, our goal was to present an agent-based model of AIDS spread, modeled as a CAS,
coupled and correlated with a formal specification model. In particular, however, as an outcome of
extensive simulation experiments, we have noted how a more natural sub-division of populations
might lead to discovery of complex emergent behaviors in HIV epidemic spread. The results
demonstrate how a more natural division of population preferably in the form of networks might
allow for the detection of complex, emergent patterns in HIV spread.
We have demonstrated the use of a formal framework in conjunction with agent-based
simulation model as an effective method in order to study complex diseases like HIV/AIDS. This
formal framework allows conducting an investigative study for analyzing a complex adaptive
environment like HIV/AIDS spread effectively. First a formal specification model of HIV/AIDS spread
was developed using Z which allowed for a clear and unambiguous representation of the underlying
complex social network. Next the formal specification was translated into an agent-based model as a
proof of concept.
Extensive simulation experiments presented in a 95% confidence interval, demonstrated that in
the case of FSWs, there can be back-infection patterns originating based on sexual habits and
partnerships of their clients with a different section of population. We have also shown how the
model can be verified and validated by evaluating the use of protection in physical relations as a
protective measure from disease spread. Sensitivity analysis provided a proof of the utility of
dividing populations
into further sub-populations. The discovered emergent patterns are
demonstrated since HIV spread is closely tied with the realities associated with the taboos and
norms of the human society. The traditional modeling techniques could not predict the existence of
complex hidden patterns and thus FABS-HIV proves to be presentable as an adequate modeling
framework for analyzing a CAS like HIV/AIDS spread.
We demonstrate theoretically that infection emergence can change considerably depending
upon the variations in the high risk behavior of certain populations. Our results also confirm the
significantly positive role of both condom usage and commitment levels on stopping this possible
backflow of infection in FSWs. Positive awareness of both these factors can play a significant role in
better understanding of the disease diffusion rates specific to a population and also in planning
effective intervention strategies.
Although the formal framework presented here has been described in the context of FSWs, it
provides a complete description of the complex patterns involved in the flow of HIV/AIDS. In the
future, further research can be conducted on the use of complex network methods to investigate
trends in AIDS spread. In addition, FABS-based models can be developed for other sub-populations
based on the behavior of high risk individuals such as injecting drug users (IDUs).
Fig.1. Screen shot of agents used during simulation modeling showing female sex
workers, FSWs (yellow), infected FSWs (red), their clients, termed primaries (blue) and
partners of primaries, termed secondaries (red). Sexual coupling between agents of different
sub-populations during simulation is shown by vertical lines between agents. For each time
tick (equally-spaced time interval), the interactions appear as lines between the relevant
agents and then disappear in the next time tick for the sake of clarity.
Fig.2. Screen shot of the graphical user interface (GUI) of the agent-based model of the
FABS-HIV. The image shows sliders that allowed modulation of various sub-populations
within the model as well as the level of confounding parameters like commitment levels and
condom usage among the various agents.
Fig.3. Results of simulation modeling by varying the commitment level among primaries.
The graphs represent the estimated number of infected agents with increase in commitment
levels: A) female sex workers (FSWs) infected, B) primaries infected, C) FSWs back-
infected, D) primaries back-infected E) secondaries infected, F) total infected. The bars show
the minimum, maximum, and average numbers estimated by simulation modeling using 50
rounds of simulations conducted per point.
Fig.4. Results of simulation modeling by varying the condom use among primaries and
secondaries. The graphs represent the estimated number of infected agents with increase in
condom usage: A) female sex workers (FSWs) infected, B) primaries infected, C) FSWs
back-infected, D) primaries back-infected, E) secondaries infected, and F) total infected. The
bars show the minimum, maximum, and average numbers estimated by simulation modeling
using 50 rounds of simulations conducted per point.
References
Alam, S., Meyer, R., & Norling, E. (2006). Using Agent-Based Modeling to Understand the Impact of HIV/AIDS in the Context of Socio-
Economic Stressors Centre for Policy Modelling Report CPM-06-167.
Allen, S., Meinzen-Derr, J., Kautzman, M., Zulu, I., Trask, S., Fideli, U., . . . Haworth, A. (2003). Sexual behavior of HIV discordant
couples after HIV counseling and testing. Aids, 17(5), 733.
Atkinson, J. (1996). A simulation model of the dynamics of HIV transmission in intravenous drug users. Computers and Biomedical
Research, 29(4), 338-349.
Bailey, A. M., Lawrence, M. B., Shang, H., Katz, A. J., & Peirce, S. M. (2009). Agent-based model of therapeutic adipose-derived stromal
cell trafficking during ischemia predicts ability to roll on P-selectin. PLoS Comput Biol, 5(2), e1000294. doi:
10.1371/journal.pcbi.1000294
Bankes, S. C. (2002). Agent-based modeling: A revolution? Proceedings of the National Academy of Sciences, 99(90003), 7199-7200.
Boccara, N. (2010). Modeling complex systems (Second ed.): Springer.
Boily, M., Lowndes, C., & Alary, M. (2002). The impact of HIV epidemic phases on the effectiveness of core group interventions: insights
from mathematical models. Sexually transmitted infections, 78(suppl 1), i78.
Bowen, J. (1995). Formal specification and documentation using Z: a case study approach: Citeseer.
Cartier, M. (2004). An Agent-Based Model of Innovation Emergence in Organizations: Renault and Ford Through the Lens of
Evolutionism. Comput. Math. Organ. Theory, 10(2), 147-153. doi: http://dx.doi.org/10.1023/B:CMOT.0000039167.91320.df
Chin, D., & Kroesen, K. (1999). Disclosure of HIV Infection Among Asian/Pacific Islander American Women: Cultural Stigma and
Support* 1. Cultural diversity and ethnic minority psychology, 5(3), 222-235.
Dean, H. D., & Fenton, K. A. (2010). Addressing social determinants of health in the prevention and control of HIV/AIDS, viral hepatitis,
sexually transmitted infections, and tuberculosis. Public Health Reports, 125(Suppl 4), 1.
Dennard, L. F., Richardson, K. A., & Morçöl, G. g. (2008). Complexity and policy analysis : tools and concepts for designing robust
policies in a complex world. Goodyear, AZ: ISCE Pub.
Guest, G., Bunce, A., & Johnson, L. (2006). How many interviews are enough? Field methods, 18(1), 59.
Hall, A. (1990). Seven myths of formal methods. IEEE software, 7(5), 11-19.
Holland, J. (1992). Complex adaptive systems. Daedalus, 121(1), 17-30.
Holland, J. (2006). Studying Complex Adaptive Systems. Journal of Systems Science and Complexity, 19(1), 1-8. doi: 10.1007/s11424-006-
0001-z
Houston, I., & King, S. (1991). CICS project report experiences and results from the use of Z in IBM. Paper presented at the VDM'91
Formal Software Development Methods.
Hussain, A., & Niazi, M. (2013). Toward a Formal, Visual Framework of Emergent Cognitive Development of Scholars. Cognitive
Computation, 1-12. doi: 10.1007/s12559-013-9219-y
Jacintho, L. F. O., Batista, A. F. M., Ruas, T. L., Marietto, M. G. B., & Silva, F. A. (2010). An agent-based model for the spread of the
Dengue fever: a swarm platform simulation approach.
Jones, J., & Handcock, M. (2003). Social networks (communication arising): Sexual contacts and epidemic thresholds. Nature, 423(6940),
605-606.
Li, Y., Brimicombe, A. J., & Li, C. (2008). Agent-based services for the validation and calibration of multi-agent models. Computers,
Environment and Urban Systems, 32(6), 464-473. doi: DOI: 10.1016/j.compenvurbsys.2008.09.002
Mei, S., Sloot, P., Quax, R., Zhu, Y., & Wang, W. (2010). Complex agent networks explaining the HIV epidemic among homosexual men
in Amsterdam. Mathematics and Computers in Simulation, 80(5), 1018-1030.
Mitchell, M. (2009). Complexity: a guided tour: Oxford University Press, USA.
Morio, S., Soda, K., HashimotoY, S., Fukutomiy, K., Ichikawa, S., Kamakura, M., & Nakayama, H. (1996). Simulation of the heterosexual
HIV/AIDS epidemic in Japan by a fuzzy mathematical model. Yonago Acta medica, 39(2), 83-98.
Morris, M., & Kretzschmar, M. (1997). Concurrent partnerships and the spread of HIV. Aids, 11(5), 641.
Nagelkerke, N. J. D., Jha, P., Vlas, S. J., Korenromp, E. L., Moses, S., Blanchard, J. F., & Plummer, F. A. (2002). Modelling HIV/AIDS
epidemics in Botswana and India: impact of interventions to prevent transmission. Bulletin of the World Health Organization, 80,
89-96.
Nagoski, E. (2006). An Agent Based Model of Disease Diffusion in the Context of Heterogeneous Sexual Motivation. PhD, Indiana
University. Retrieved from http://hdl.handle.net/2022/7351
Niazi, M., & Hussain, A. (2010). Agent based tools for modeling and simulation of self-organization in peer-to-peer, ad-hoc and other
complex networks. IEEE Communications Magazine, 47(3), 163 – 173.
Niazi, M., & Hussain, A. (2011). Agent-based computing from multi-agent systems to agent-based models: a visual survey. Scientometrics,
89(2), 479-499. doi: 10.1007/s11192-011-0468-9
Niazi, M., Siddiqa, A., & Fortino, G. (2013). Modeling AIDS Spread in Social Networks. In M. Klusch, M. Thimm & M. Paprzycki (Eds.),
Multiagent System Technologies (Vol. 8076, pp. 361-371): Springer Berlin Heidelberg.
Niazi, M. A. (2013). Complex Adaptive Systems Modeling: A multidisciplinary Roadmap. Complex Adaptive Systems Modeling, 1(1), 1.
Niazi, M. A., & Hussain, A. (2011a). A Novel Agent-Based Simulation Framework for Sensing in Complex Adaptive Environments.
Sensors Journal, IEEE, 11(2), 404-412.
Niazi, M. A., & Hussain, A. (2011b). Sensing Emergence in Complex Systems. Sensors Journal, IEEE, 11(10), 2479-2480.
Niazi, M. A., & Hussain, A. (2011). Sensing Emergence in Complex Systems. IEEE Sensors Journal. doi: 10.1109/JSEN.2011.2142303
Niazi, M. A. K. (2011). Towards A Novel Unified Framework for Developing Formal, Network and Validated Agent-Based Simulation
from
PhD, University
Retrieved
Systems.
Stirling,
Models
https://dspace.stir.ac.uk/handle/1893/3365
Adaptive
Complex
of
of
Stirling.
Osgood, N., Moavenzadeh, F., & Rhee, A. J. (2006). An agent-based approach to HIV/AIDS epidemic modeling: a case study of Papua New
Guinea.
Pollack, L., Osmond, D., Paul, J., & Catania, J. (2005). Evaluation of the Center for Disease Control and Prevention's HIV behavioral
surveillance of men who have sex with men: sampling issues. Sexually transmitted diseases, 32(9), 581.
Siddiqa, A., Niazi, M. A., Mustafa, F., Bokhari, H., Hussain, A., Akram, N., . . . Iqbal, S. (2009, 15-16 Aug. 2009). A new hybrid agent-
based modeling & simulation decision support system for breast cancer data analysis. Paper presented at the Information and
Communication Technologies, 2009. ICICT '09. International Conference on.
Teweldemedhin, E., Marwala, T., & Mueller, C. (2004, 5-8 Dec. 2004). Agent-based modelling: a case study in HIV epidemic. Paper
presented at the Hybrid Intelligent Systems, 2004. HIS '04. Fourth International Conference on.
Tirado-Ramos, A., & Kelley, C. (2013). Simulation of HIV Infection Propagation Networks: A Review of the State of the Art in Agent-
Based Approaches. International Journal of Agent Technologies and Systems (IJATS), 5(1), 53-63. doi: 10.4018/jats.2013010104
UNAIDS, W. (2005). HIV/AIDS Surveillance Project Report (HASP) 2003 and 2005
UNAIDS, W. (2010). Second generation surveillance for HIV: The next decade. Islamabad: HASP, Pakistan.
Verdasca, J., Telo da Gama, M. M., Nunes, A., Bernardino, N. R., Pacheco, J. M., & Gomes, M. C. (2005). Recurrent epidemics in small
world networks. Journal of Theoretical Biology, 233(4), 553-561. doi: DOI: 10.1016/j.jtbi.2004.10.031
WHO. (2011). Adult HIV Prevalence (15-49 years) by WHO Region (G. H. Observatory, Trans.): World Health Organization.
Wilensky, U. (1999). NetLogo. Evanston, IL, 1999: Center for Connected Learning Comp.-Based Modeling, Northwestern University.
Retrieved from http://ccl.northwestern.edu/netlogo
Woodcock, J., & Davies, J. (1996). Using Z: specification, refinement, and proof: Prentice Hall Upper Saddle River, NJ.
|
1810.05587 | 3 | 1810 | 2019-08-30T19:33:23 | A Survey and Critique of Multiagent Deep Reinforcement Learning | [
"cs.MA",
"cs.AI",
"cs.LG"
] | Deep reinforcement learning (RL) has achieved outstanding results in recent years. This has led to a dramatic increase in the number of applications and methods. Recent works have explored learning beyond single-agent scenarios and have considered multiagent learning (MAL) scenarios. Initial results report successes in complex multiagent domains, although there are several challenges to be addressed. The primary goal of this article is to provide a clear overview of current multiagent deep reinforcement learning (MDRL) literature. Additionally, we complement the overview with a broader analysis: (i) we revisit previous key components, originally presented in MAL and RL, and highlight how they have been adapted to multiagent deep reinforcement learning settings. (ii) We provide general guidelines to new practitioners in the area: describing lessons learned from MDRL works, pointing to recent benchmarks, and outlining open avenues of research. (iii) We take a more critical tone raising practical challenges of MDRL (e.g., implementation and computational demands). We expect this article will help unify and motivate future research to take advantage of the abundant literature that exists (e.g., RL and MAL) in a joint effort to promote fruitful research in the multiagent community. | cs.MA | cs | A Survey and Critique of Multiagent Deep Reinforcement Learning$
Pablo Hernandez-Leal, Bilal Kartal and Matthew E. Taylor
{pablo.hernandez,bilal.kartal,matthew.taylor}@borealisai.com
Borealis AI
Edmonton, Canada
9
1
0
2
g
u
A
0
3
]
A
M
.
s
c
[
3
v
7
8
5
5
0
.
0
1
8
1
:
v
i
X
r
a
Abstract
Deep reinforcement learning (RL) has achieved outstanding results in recent years. This has led
to a dramatic increase in the number of applications and methods. Recent works have explored
learning beyond single-agent scenarios and have considered multiagent learning (MAL) scenarios.
Initial results report successes in complex multiagent domains, although there are several challenges
to be addressed. The primary goal of this article is to provide a clear overview of current multiagent
deep reinforcement learning (MDRL) literature. Additionally, we complement the overview with a
broader analysis: (i) we revisit previous key components, originally presented in MAL and RL, and
highlight how they have been adapted to multiagent deep reinforcement learning settings. (ii) We
provide general guidelines to new practitioners in the area: describing lessons learned from MDRL
works, pointing to recent benchmarks, and outlining open avenues of research. (iii) We take a
more critical tone raising practical challenges of MDRL (e.g., implementation and computational
demands). We expect this article will help unify and motivate future research to take advantage of
the abundant literature that exists (e.g., RL and MAL) in a joint effort to promote fruitful research
in the multiagent community.
1. Introduction
Almost 20 years ago Stone and Veloso's seminal survey [1] laid the groundwork for defining the
area of multiagent systems (MAS) and its open problems in the context of AI. About ten years
ago, Shoham, Powers, and Grenager [2] noted that the literature on multiagent learning (MAL)
was growing and it was not possible to enumerate all relevant articles. Since then, the number of
published MAL works continues to steadily rise, which led to different surveys on the area, ranging
from analyzing the basics of MAL and their challenges [3, 4, 5], to addressing specific subareas:
game theory and MAL [2, 6], cooperative scenarios [7, 8], and evolutionary dynamics of MAL [9].
In just the last couple of years, three surveys related to MAL have been published:
learning in
non-stationary environments [10], agents modeling agents [11], and transfer learning in multiagent
RL [12].
The research interest in MAL has been accompanied by successes in artificial intelligence,
first, in single-agent video games [13]; more recently, in two-player games, for example, playing
$Earlier versions of this work had the title: "Is multiagent deep reinforcement learning the answer or the question?
A brief survey"
1
Go [14, 15], poker [16, 17], and games of two competing teams, e.g., DOTA 2 [18] and StarCraft
II [19].
While different techniques and algorithms were used in the above scenarios, in general, they are
all a combination of techniques from two main areas: reinforcement learning (RL) [20] and deep
learning [21, 22].
RL is an area of machine learning where an agent learns by interacting (i.e., taking actions)
within a dynamic environment. However, one of the main challenges to RL, and traditional ma-
chine learning in general, is the need for manually designing quality features on which to learn.
Deep learning enables efficient representation learning, thus allowing the automatic discovery of
features [21, 22]. In recent years, deep learning has had successes in different areas such as com-
puter vision and natural language processing [21, 22]. One of the key aspects of deep learning is the
use of neural networks (NNs) that can find compact representations in high-dimensional data [23].
In deep reinforcement learning (DRL) [23, 24] deep neural networks are trained to approximate
the optimal policy and/or the value function.
In this way the deep NN, serving as function
approximator, enables powerful generalization. One of the key advantages of DRL is that it enables
RL to scale to problems with high-dimensional state and action spaces. However, most existing
successful DRL applications so far have been on visual domains (e.g., Atari games), and there is
still a lot of work to be done for more realistic applications [25, 26] with complex dynamics, which
are not necessarily vision-based.
DRL has been regarded as an important component in constructing general AI systems [27]
and has been successfully integrated with other techniques, e.g., search [14], planning [28], and
more recently with multiagent systems, with an emerging area of multiagent deep reinforcement
learning (MDRL)[29, 30].1
Learning in multiagent settings is fundamentally more difficult than the single-agent case due to
the presence of multiagent pathologies, e.g., the moving target problem (non-stationarity) [2, 5, 10],
curse of dimensionality [2, 5], multiagent credit assignment [31, 32], global exploration [8], and
relative overgeneralization [33, 34, 35]. Despite this complexity, top AI conferences like AAAI,
ICML, ICLR, IJCAI and NeurIPS, and specialized conferences such as AAMAS, have published
works reporting successes in MDRL. In light of these works, we believe it is pertinent to first, have
an overview of the recent MDRL works, and second, understand how these recent works relate to
the existing literature.
This article contributes to the state of the art with a brief survey of the current works in
MDRL in an effort to complement existing surveys on multiagent learning [36, 10], cooperative
learning [7, 8], agents modeling agents [11], knowledge reuse in multiagent RL [12], and (single-
agent) deep reinforcement learning [23, 37].
First, we provide a short review of key algorithms in RL such as Q-learning and REINFORCE
(see Section 2.1). Second, we review DRL highlighting the challenges in this setting and reviewing
recent works (see Section 2.2). Third, we present the multiagent setting and give an overview of
key challenges and results (see Section 3.1). Then, we present the identified four categories to
group recent MDRL works (see Figure 1):
• Analysis of emergent behaviors: evaluate single-agent DRL algorithms in multiagent scenarios
(e.g., Atari games, social dilemmas, 3D competitive games).
• Learning communication: agents learn communication protocols to solve cooperative tasks.
1We have noted inconsistency in abbreviations such as: D-MARL, MADRL, deep-multiagent RL and MA-DRL.
2
(a) Analysis of emergent behaviors
(b) Learning communication
(c) Learning cooperation
(d) Agents modeling agents
Figure 1: Categories of different MDRL works.
(a) Analysis of emergent behaviors: evaluate single-agent DRL
algorithms in multiagent scenarios. (b) Learning communication: agents learn with actions and through messages.
(c) Learning cooperation: agents learn to cooperate using only actions and (local) observations. (d) Agents modeling
agents: agents reason about others to fulfill a task (e.g., cooperative or competitive). For a more detailed description
see Sections 3.3 -- 3.6 and Tables 1 -- 4.
• Learning cooperation: agents learn to cooperate using only actions and (local) observations.
• Agents modeling agents: agents reason about others to fulfill a task (e.g., best response
learners).
For each category we provide a description as well as outline the recent works (see Section 3.2
and Tables 1 -- 4). Then, we take a step back and reflect on how these new works relate to the existing
literature. In that context, first, we present examples on how methods and algorithms originally
introduced in RL and MAL were successfully been scaled to MDRL (see Section 4.1). Second,
we provide some pointers for new practitioners in the area by describing general lessons learned
from the existing MDRL works (see Section 4.2) and point to recent multiagent benchmarks (see
Section 4.3). Third, we take a more critical view and describe practical challenges in MDRL, such
as reproducibility, hyperparameter tunning, and computational demands (see Section 4.4). Then,
we outline some open research questions (see Section 4.5). Lastly, we present our conclusions from
this work (see Section 5).
Our goal is to outline a recent and active area (i.e., MDRL), as well as to motivate future
research to take advantage of the ample and existing literature in multiagent learning. We aim to
enable researchers with experience in either DRL or MAL to gain a common understanding about
recent works, and open problems in MDRL, and to avoid having scattered sub-communities with
little interaction [2, 10, 11, 38].
2. Single-agent learning
This section presents the formalism of reinforcement learning and its main components before
outlining deep reinforcement learning along with its particular challenges and recent algorithms.
3
For a more detailed description we refer the reader to excellent books and surveys on the area [39,
20, 23, 40, 24].
2.1. Reinforcement learning
RL formalizes the interaction of an agent with an environment using a Markov decision process
(MDP) [41]. An MDP is defined by the tuple (cid:104)S,A, R, T, γ(cid:105) where S represents a finite set of states.
A represents a finite set of actions. The transition function T : S × A × S → [0, 1] determines the
probability of a transition from any state s ∈ S to any state s(cid:48) ∈ S given any possible action a ∈ A.
The reward function R : S ×A×S → R defines the immediate and possibly stochastic reward that
an agent would receive given that the agent executes action a while in state s and it is transitioned
to state s(cid:48), γ ∈ [0, 1] represents the discount factor that balances the trade-off between immediate
rewards and future rewards.
MDPs are adequate models to obtain optimal decisions in single agent fully observable environ-
ments.2 Solving an MDP will yield a policy π : S → A, which is a mapping from states to actions.
An optimal policy π∗ is the one that maximizes the expected discounted sum of rewards. There are
different techniques for solving MDPs assuming a complete description of all its elements. One of
the most common techniques is the value iteration algorithm [44], which requires a complete and
accurate representation of states, actions, rewards, and transitions. However, this may be difficult
to obtain in many domains. For this reason, RL algorithms often learn from experience interacting
with the environment in discrete time steps.
Q-learning. One of the most well known algorithms for RL is Q-learning [45]. It has been devised
for stationary, single-agent, fully observable environments with discrete actions. A Q-learning agent
keeps the estimate of its expected payoff starting in state s, taking action a as Q(s, a). Each tabular
entry Q(s, a) is an estimate of the corresponding optimal Q∗ function that maps state-action pairs
to the discounted sum of future rewards starting with action a at state s and following the optimal
policy thereafter. Each time the agent transitions from a state s to a state s(cid:48) via action a receiving
payoff r, the Q table is updated as follows:
Q(s, a) ← Q(s, a) + α[(r + γ max
a(cid:48)
Q(s(cid:48), a(cid:48))) − Q(s, a)]
(1)
with the learning rate α ∈ [0, 1]. Q-learning is proven to converge to Q∗ if state and action spaces
are discrete and finite, the sum of the learning rates goes to infinity (so that each state-action pair
is visited infinitely often) and that the sum of the squares of the learning rates is finite (which
is required to show that the convergence is with probability one) [46, 45, 47, 48, 49, 50, 51].
The convergence of single-step on-policy RL algorithms, i.e, SARSA (λ = 0), for both decaying
exploration (greedy in the limit with infinite exploration) and persistent exploration (selecting
actions probabilistically according to the ranks of the Q values) was demonstrated by Singh et
al. [52]. Furthermore, Van Seijen [53] has proven convergence for Expected SARSA (see Section 3.1
for convergence results in multiagent domains).
2A Partially Observable Markov Decision Process (POMDP) [42, 43] explicitly models environments where the
agent no longer sees the true system state and instead receives an observation (generated from the underlying system
state).
4
REINFORCE (Monte Carlo policy gradient). In contrast to value-based methods, which do not
try to optimize directly over a policy space [54], policy gradient methods can learn parameterized
policies without using intermediate value estimates.
Policy parameters are learned by following the gradient of some performance measure with
gradient descent [55]. For example, REINFORCE [56] uses estimated return by Monte Carlo (MC)
methods with full episode trajectories to learn policy parameters θ, with π(a; s, θ) ≈ π(a; s), as
follows
∇π(At; St, θt)
π(At; St, θt)
(2)
where Gt represents the return, α is the learning rate, and At ∼ π. A main limitation is that policy
gradient methods can have high variance [54].
θt+1 = θt + αGt
The policy gradient update can be generalized to include a comparison to an arbitrary baseline
of the state [56]. The baseline, b(s), can be any function, as long as it does not vary with the
action; the baseline leaves the expected value of the update unchanged, but it can have an effect
on its variance [20]. A natural choice for the baseline is a learned state-value function, this reduces
the variance, and it is bias-free if learned by MC.3 Moreover, when using the state-value function
for bootstrapping (updating the value estimate for a state from the estimated values of subsequent
states) it assigns credit (reducing the variance but introducing bias), i.e., criticizes the policy's
action selections. Thus, in actor-critic methods [54], the actor represents the policy, i.e., action-
selection mechanism, whereas a critic is used for the value function learning. In the case when
the critic learns a state-action function (Q function) and a state value function (V function), an
advantage function can be computed by subtracting state values from the state-action values [20,
60]. The advantage function indicates the relative quality of an action compared to other available
actions computed from the baseline, i.e., state value function. An example of an actor-critic
algorithm is Deterministic Policy Gradient (DPG) [61]. In DPG [61] the critic follows the standard
Q-learning and the actor is updated following the gradient of the policy's performance [62], DPG
was later extended to DRL (see Section 2.2) and MDRL (see Section 3.5). For multiagent learning
settings the variance is further increased as all the agents' rewards depend on the rest of the agents,
and it is formally shown that as the number of agents increase, the probability of taking a correct
gradient direction decreases exponentially [63]. Recent MDRL works addressed this high variance
issue, e.g., COMA [64] and MADDPG [63] (see Section 3.5).
Policy gradient methods have a clear connection with deep reinforcement learning since the
policy might be represented by a neural network whose input is a representation of the state, whose
output are action selection probabilities or values for continuous control [65], and whose weights
are the policy parameters.
2.2. Deep reinforcement learning
While tabular RL methods such as Q-learning are successful in domains that do not suffer
from the curse of dimensionality, there are many limitations: learning in large state spaces can be
prohibitively slow, methods do not generalize (across the state space), and state representations
need to be hand-specified [20]. Function approximators tried to address those limitations, using
3Action-dependant baselines had been proposed [57, 58], however, a recent study by Tucker et al. [59] found that
in many works the reason of good performance was because of bugs or errors in the code, rather than the proposed
method itself.
5
for example, decision trees [66], tile coding [67], radial basis functions [68], and locally weighted
regression [69] to approximate the value function.
Similarly, these challenges can be addressed by using deep learning, i.e., neural networks [69, 66]
as function approximators. For example, Q(s, a; θ) can be used to approximate the state-action
values with θ representing the neural network weights. This has two advantages, first, deep learning
helps to generalize across states improving the sample efficiency for large state-space RL problems.
Second, deep learning can be used to reduce (or eliminate) the need for manually designing features
to represent state information [21, 22].
However, extending deep learning to RL problems comes with additional challenges including
non-i.i.d. (not independently and identically distributed) data. Many supervised learning methods
assume that training data is from an i.i.d. stationary distribution [70, 22, 71]. However, in RL,
training data consists of highly correlated sequential agent-environment interactions, which violates
the independence condition. Moreover, RL training data distribution is non-stationary as the agent
actively learns while exploring different parts of the state space, violating the condition of sampled
data being identically distributed [72].
In practice, using function approximators in RL requires making crucial representational de-
cisions and poor design choices can result in estimates that diverge from the optimal value func-
tion [73, 69, 74, 75, 76, 77]. In particular, function approximation, bootstrapping, and off-policy
learning are considered the three main properties that when combined, can make the learning to di-
verge and are known as the deadly triad [77, 20]. Recently, some works have shown that non-linear
(i.e., deep) function approximators poorly estimate the value function [78, 59, 79] and another
work found problems with Q-learning using function approximation (over/under-estimation, in-
stability and even divergence) due to the delusional bias: "delusional bias occurs whenever a
backed-up value estimate is derived from action choices that are not realizable in the underlying
policy class"[80]. Additionally, convergence results for reinforcement learning using function ap-
proximation are still scarce [74, 81, 82, 83, 80]; in general, stronger convergence guarantees are
available for policy-gradient methods [55] than for value-based methods [20].
Below we mention how the existing DRL methods aim to address these challenges when briefly
reviewing value-based methods, such as DQN [13]; policy gradient methods, like Proximal Policy
Optimization (PPO) [60]; and actor-critic methods like Asynchronous Advantage Actor-Critic
(A3C) [84]. We refer the reader to recent surveys on single-agent DRL [23, 37, 24] for a more
detailed discussion of the literature.
Value-based methods. The major breakthrough work combining deep learning with Q-learning was
the Deep Q-Network (DQN) [13]. DQN uses a deep neural network for function approxima-
tion [87]4 (see Figure 2) and maintains an experience replay (ER) buffer [89, 90] to store interactions
(cid:104)s, a, r, s(cid:48)(cid:105). DQN keeps an additional copy of neural network parameters, θ−, for the target net-
work in addition to the θ parameters to stabilize the learning, i.e., to alleviate the non-stationary
data distribution.5 For each training iteration i, DQN minimizes the mean-squared error (MSE)
between the Q-network and its target network using the loss function:
4Before DQN, many approaches used neural networks for representing the Q-value function [88], such as Neural
Fitted Q-learning [87] and NEAT+Q [75].
5Double Q-learning [91] originally proposed keeping two Q functions (estimators) to reduce the overestimation
bias in RL, while still keeping the convergence guarantees, later it was extended to DRL in Double DQN [92] (see
Section 4.1).
6
Figure 2: Deep Q-Network (DQN) [13]: Inputs are four stacked frames; the network is composed of several layers:
Convolutional layers employ filters to learn features from high-dimensional data with a much smaller number of
neurons and Dense layers are fully-connected layers. The last layer represents the actions the agent can take (in
this case, 10 possible actions). Deep Recurrent Q-Network (DRQN) [85], which extends DQN to partially observable
domains [42], is identical to this setup except the penultimate layer (1× 256 Dense layer) is replaced with a recurrent
LSTM layer [86].
Li(θi) = Es,a,r,s(cid:48)[(r + γmaxa(cid:48)Q(s(cid:48), a(cid:48); θ−
(3)
where target network parameters θ− are set to Q-network parameters θ periodically and mini-
batches of (cid:104)s, a, r, s(cid:48)(cid:105) tuples are sampled from the ER buffer, as depicted in Figure 3.
i ) − Q(s, a; θi))2]
The ER buffer provides stability for learning as random batches sampled from the buffer helps
alleviating the problems caused by the non-i.i.d. data. However, it comes with disadvantages,
such as higher memory requirements and computation per real interaction [93]. The ER buffer is
mainly used for off-policy RL methods as it can cause a mismatch between buffer content from
earlier policy and from the current policy for on-policy methods [93]. Extending the ER buffer
for the multiagent case is not trivial, see Sections 3.5, 4.1 and 4.2. Recent works were designed
to reduce the problem of catastrophic forgetting (this occurs when the trained neural network
performs poorly on previously learned tasks due to a non-stationary training distribution [94, 95])
and the ER buffer, in DRL [96] and MDRL [97].
DQN has been extended in many ways, for example, by using double estimators [91] to reduce
the overestimation bias with Double DQN [92] (see Section 4.1) and by decomposing the Q-function
with a dueling-DQN architecture [98], where two streams are learned, one estimates state values
and another one advantages, those are combined in the final layer to form Q values (this method
improved over Double DQN).
In practice, DQN is trained using an input of four stacked frames (last four frames the agent has
encountered). If a game requires a memory of more than four frames it will appear non-Markovian
to DQN because the future game states (and rewards) do not depend only on the input (four
frames) but rather on the history [99]. Thus, DQN's performance declines when given incomplete
state observations (e.g., one input frame) since DQN assumes full state observability.
Real-world tasks often feature incomplete and noisy state information resulting from partial
observability (see Section 2.1). Deep Recurrent Q-Networks (DRQN) [85] proposed using recurrent
neural networks, in particular, Long Short-Term Memory (LSTMs) cells [86] in DQN, for this
setting. Consider the architecture in Figure 2 with the first dense layer after convolution replaced
7
Figure 3: Representation of a DQN agent that uses an experience replay buffer [89, 90] to keep (cid:104)s, a, r, s(cid:48)(cid:105) tuples for
minibatch updates. The Q-values are parameterized with a NN and a policy is obtained by selecting (greedily) over
those at every timestep.
by a layer of LSTM cells. With this addition, DRQN has memory capacity so that it can even
work with only one input frame rather than a stacked input of consecutive frames. This idea has
been extended to MDRL, see Figure 6 and Section 4.2. There are also other approaches to deal
with partial observability such as finite state controllers [100] (where action selection is performed
according to the complete observation history) and using an initiation set of options conditioned
on the previously employed option [101].
Policy gradient methods. For many tasks, particularly for physical control, the action space is
continuous and high dimensional where DQN is not suitable. Deep Deterministic Policy Gradient
(DDPG) [65] is a model-free off-policy actor-critic algorithm for such domains, based on the DPG
algorithm [61] (see Section 2.1). Additionally, it proposes a new method for updating the networks,
i.e., the target network parameters slowly change (this could also be applicable to DQN), in contrast
to the hard reset (direct weight copy) used in DQN. Given the off-policy nature, DDPG generates
exploratory behavior by adding sampled noise from some noise processes to its actor policy. The
authors also used batch normalization [102] to ensure generalization across many different tasks
without performing manual normalizations. However, note that other works have shown batch
normalization can cause divergence in DRL [103, 104].
Asynchronous Advantage Actor-Critic (A3C) [93] is an algorithm that employs a parallelized
asynchronous training scheme (using multiple CPU threads) for efficiency.
It is an on-policy
RL method that does not use an experience replay buffer. A3C allows multiple workers to
simultaneously interact with the environment and compute gradients locally. All the workers
pass their computed local gradients to a global NN which performs the optimization and syn-
chronizes with the workers asynchronously (see Figure 4). There is also the Advantage Actor-
Critic (A2C) method [105] that combines all the gradients from all the workers to update the
global NN synchronously. The loss function for A3C is composed of two terms: policy loss (ac-
tor), Lπ, and value loss (critic), Lv. A3C parameters are updated using the advantage function
A(st, at; θv) = Q(s, a)− V (s), commonly used to reduce variance (see Section 2.1). An entropy loss
for the policy, H(π), is also commonly added, which helps to improve exploration by discouraging
premature convergence to suboptimal deterministic policies [93]. Thus, the loss function is given
by: LA3C = λvLv + λπLπ − λHEs∼π[H(π(s,·, θ)] with λv, λπ, and λH , being weighting terms on the
individual loss components. Wang et al. [106] took A3C's framework but used off-policy learning
8
Figure 4: Asynchronous Advantage Actor-Critic (A3C) employs multiple (CPUs) workers without needing an ER
buffer. Each worker has its own NN and independently interacts with the environment to compute the loss and
gradients. Workers then pass computed gradients to the global NN that optimizes the parameters and synchronizes
with the worker asynchronously. This distributed system is designed for single-agent deep RL. Compared to different
DQN variants, A3C obtains better performance on a variety of Atari games using substantially less training time
with multiple CPU cores of standard laptops without a GPU [93]. However, we note that more recent approaches
use both multiple CPU cores for more efficient training data generation and GPUs for more efficient learning.
to create the Actor-critic with experience replay (ACER) algorithm. Gu et al. [107] introduced
the Interpolated Policy Gradient (IPG) algorithm and showed a connection between ACER and
DDPG: they are a pair of reparametrization terms (they are special cases of IPG) when they are
put under the same stochastic policy setting, and when the policy is deterministic they collapse
into DDPG.
Jaderberg et al. [84] built the Unsupervised Reinforcement and Auxiliary Learning (UNREAL)
framework on top of A3C and introduced unsupervised auxiliary tasks (e.g., reward prediction)
to speed up the learning process. Auxiliary tasks in general are not used for anything other
than shaping the features of the agent, i.e., facilitating and regularizing the representation learn-
(cid:80)
ing process [108, 109]; their formalization in RL is related to the concept of general value func-
tions [20, 110]. The UNREAL framework optimizes a combined loss function LUNREAL ≈ LA3C +
i λATiLATi, that combines the A3C loss, LA3C, together with auxiliary task losses LATi, where
λATi are weight terms (see Section 4.1 for use of auxiliary tasks in MDRL). In contrast to A3C, UN-
REAL uses a prioritized ER buffer, in which transitions with positive reward are given higher prob-
ability of being sampled. This approach can be viewed as a simple form of prioritized replay [111],
which was in turn inspired by model-based RL algorithms like prioritized sweeping [112, 113].
Another distributed architecture is the Importance Weighted Actor-Learner Architecture (IM-
9
PALA) [114]. Unlike A3C or UNREAL, IMPALA actors communicate trajectories of experience
(sequences of states, actions, and rewards) to a centralized learner, thus IMPALA decouples acting
from learning.
Trust Region Policy Optimization (TRPO) [60] and Proximal Policy Optimization (PPO) [115]
are recently proposed policy gradient algorithms where the latter represents the state-of-the art
with advantages such as being simpler to implement and having better empirical sample complex-
ity. Interestingly, a recent work [79] studying PPO and TRPO arrived at the surprising conclusion
that these methods often deviate from what the theoretical framework would predict: gradient
estimates are poorly correlated with the true gradient and value networks tend to produce inaccu-
rate predictions for the true value function. Compared to vanilla policy gradient algorithms, PPO
prevents abrupt changes in policies during training through the loss function, similar to early work
by Kakade [116]. Another advantage of PPO is that it can be used in a distributed fashion, i.e,
Distributed PPO (DPPO) [117]. Note that distributed approaches like DPPO or A3C use paral-
lelization only to improve the learning by more efficient training data generation through multiple
CPU cores for single agent DRL and they should not be considered multiagent approaches (except
for recent work which tries to exploit this parallelization in a multiagent environment [118]).
Lastly, there's a connection between policy gradient algorithms and Q-learning [119] within the
framework of entropy-regularized reinforcement learning [120] where the value and Q functions are
slightly altered to consider the entropy of the policy. In this vein, Soft Actor-Critic (SAC) [121] is
a recent algorithm that concurrently learns a stochastic policy, two Q-functions (taking inspiration
from Double Q-learning) and a value function. SAC alternates between collecting experience with
the current policy and updating from batches sampled from the ER buffer.
We have reviewed recent algorithms in DRL, while the list is not exhaustive, it provides an
overview of the different state-of-art techniques and algorithms which will become useful while
describing the MDRL techniques in the next section.
3. Multiagent Deep Reinforcement Learning (MDRL)
First, we briefly introduce the general framework on multiagent learning and then we dive into
the categories and the research on MDRL.
3.1. Multiagent Learning
Learning in a multiagent environment is inherently more complex than in the single-agent
case, as agents interact at the same time with environment and potentially with each other [5].
The independent learners, a.k.a. decentralized learners approach [122] directly uses single-agent
algorithms in the multi-agent setting despite the underlying assumptions of these algorithms being
violated (each agent independently learns its own policy, treating other agents as part of the
environment). In particular the Markov property (the future dynamics, transitions, and rewards
depend only on the current state) becomes invalid since the environment is no longer stationary [4,
6, 123]. This approach ignores the multiagent nature of the setting entirely and it can fail when
an opponent adapts or learns, for example, based on the past history of interactions [2]. Despite
the lack of guarantees, independent learners have been used in practice, providing advantages with
regards to scalability while often achieving good results [8].
To understand why multiagent domains are non-stationary from agents' local perspectives,
consider a simple stochastic (also known as Markov) game (S,N ,A,T ,R), which can be seen as
an extension of an MDP to multiple agents [124, 125]. One key distinction is that the transition,
10
T , and reward function, R, depend on the actions A = A1 × ...× AN of all, N , agents, this means,
R = R1 × ... × RN and T = S × A1 × ... × AN .
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):6
V π
i (s) =
π(s, a)
T (s, ai, a−i, s(cid:48))[Ri(s, ai, a−i, s(cid:48)) + γVi(s(cid:48))].
(cid:88)
a∈A
(cid:88)
s(cid:48)∈S
(4)
(5)
Consequently, the optimal policy is dependent on the other agents' policies,
π∗
i (s, ai, π−i) = arg max
(πi,π−i)
i
V
πi
(cid:88)
a∈A
arg max
πi
(s) =
(cid:88)
s(cid:48)∈S
πi(s, ai)π−i(s, a−i)
T (s, ai, a−i, s(cid:48))[Ri(s, ai, a−i, s(cid:48)) + γV
(πi,π−i)
i
(s(cid:48))].
Specifically, the opponents' joint policy π−i(s, a−i) can be non-stationary, i.e., changes as the
opponents' policies change over time, for example with learning opponents.
Convergence results. Littman [125] studied convergence properties of reinforcement learning joint
action agents [126] in Markov games with the following conclusions: in adversarial environments
(zero-sum games) an optimal play can be guaranteed against an arbitrary opponent, i.e., Minimax
Q-learning [124]. In coordination environments (e.g., in cooperative games all agents share the same
reward function), strong assumptions need be made about other agents to guarantee convergence
to optimal behavior [125], e.g., Nash Q-learning [127] and Friend-or-Foe Q-learning [128]. In other
types of environments no value-based RL algorithms with guaranteed convergence properties are
known [125].
Recent work on MDRL have addressed scalability and have focused significantly less on con-
vergence guarantees, with few exceptions [129, 130, 131, 132]. One notable work has shown a
connection between update rules for actor-critic algorithms for multiagent partially observable set-
tings and (counterfactual) regret minimization:7 the advantage values are scaled counterfactual
regrets. This lead to new convergence properties of independent RL algorithms in zero-sum games
with imperfect information [136]. The result is also used to support policy gradient optimization
against worst-case opponents, in a new algorithm called Exploitability Descent [137].8
We refer the interested reader to seminal works about convergence in multiagent domains [139,
140, 141, 142, 143, 144, 145, 146, 147, 148, 149]. Note that instead of convergence, some MAL
algorithms have proved learning a best response against classes of opponents [150, 151, 152].
There are other common problems in MAL, including action shadowing [34, 33], the curse of
dimensionality [5], and multiagent credit assignment [32]. Describing each problem is out of the
6In this setting each agent independently executes a policy, however, there are other cases where this does not
hold, for example when agents have a coordinated exploration strategy.
7 Counterfactual regret minimization is a technique for solving large games based on regret minimization [133, 134]
due to a well-known connection between regret and Nash equilibria [135]. It has been one of the reasons of successes
in Poker [16, 17].
8This algorithm is similar to CFR-BR [138] and has the main advantage that the current policy convergences
rather than the average policy, so there is no need to learn the average strategy, which requires large reservoir buffers
or many past networks.
11
scope of this survey. However, we refer the interested reader to excellent resources on general
MAL [4, 153, 154], as well as surveys in specific areas: game theory and multiagent reinforce-
ment learning [5, 6], cooperative scenarios [7, 8], evolutionary dynamics of multiagent learning [9],
learning in non-stationary environments [10], agents modeling agents [11], and transfer learning in
multiagent RL [12].
3.2. MDRL categorization
In Section 2.2 we outlined some recent works in single-agent DRL since an exhaustive list is out
of the scope of this article. This explosion of works has led DRL to be extended and combined with
other techniques [23, 37, 29]. One natural extension to DRL is to test whether these approaches
could be applied in a multiagent environment.
We analyzed the most recent works (that are not covered by previous MAL surveys [10, 11]
and we do not consider genetic algorithms or swarm intelligence in this survey) that have a clear
connection with MDRL. We propose 4 categories which take inspiration from previous surveys [1,
5, 7, 11] and that conveniently describe and represent current works. Note that some of these works
fit into more than one category (they are not mutually exclusive), therefore their summaries are
presented in all applicable Tables 1-4, however, for the ease of exposition when describing them in
the text we only do so in one category. Additionally, for each work we present its learning type,
either a value-based method (e.g., DQN) or a policy gradient method (e.g., actor-critic); also, we
mention if the setting is evaluated in a fully cooperative, fully competitive or mixed environment
(both cooperative and competitive).
• Analysis of emergent behaviors. These works, in general, do not propose learning algorithms
-- their main focus is to analyze and evaluate DRL algorithms, e.g., DQN [155, 156, 157],
PPO [158, 157] and others [159, 157, 160], in a multiagent environment. In this category we
found works which analyze behaviors in the three major settings: cooperative, competitive
and mixed scenarios; see Section 3.3 and Table 1.
• Learning communication [161, 160, 162, 163, 164, 165]. These works explore a sub-area
in which agents can share information with communication protocols, for example through
direct messages [162] or via a shared memory [165]. This area is attracting attention and it
had not been explored much in the MAL literature. See Section 3.4 and Table 2.
• Learning cooperation. While learning to communicate is an emerging area, fostering coop-
eration in learning agents has a long history of research in MAL [7, 8].
In this category
the analyzed works are evaluated in either cooperative or mixed settings. Some works in
this category take inspiration from MAL (e.g., leniency, hysteresis, and difference rewards
concepts) and extend them to the MDRL setting [35, 166, 167]. A notable exception [168]
takes a key component from RL (i.e., experience replay buffer) and adapts it for MDRL. See
Section 3.5 and Table 3.
• Agents modeling agents. Albrecht and Stone [11] presented a thorough survey in this topic
and we have found many works that fit into this category in the MDRL setting, some taking
inspiration from DRL [169, 170, 171], and others from MAL [172, 173, 64, 174, 175]. Modeling
agents is helpful not only to cooperate, but also for modeling opponents [172, 169, 171, 173],
inferring goals [170], and accounting for the learning behavior of other agents [64]. In this
category the analyzed algorithms present their results in either a competitive setting or a
mixed one (cooperative and competitive). See Section 3.6 and Table 4.
12
In the rest of this section we describe each category along with the summaries of related works.
3.3. Emergent behaviors
Some recent works have analyzed the previously mentioned independent DRL agents (see Sec-
tion 3.1) from the perspective of types of emerging behaviors (e.g., cooperative or competitive).
One of the earliest MDRL works is by Tampuu et al. [155], which had two independent DQN
learning agents to play the Atari Pong game. Their focus was to adapt the reward function for the
learning agents, which resulted in either cooperative or competitive emergent behaviors.
Leibo et al. [156] meanwhile studied independent DQNs in the context of sequential social
dilemmas: a Markov game that satisfies certain inequalities [156]. The focus of this work was
to highlight that cooperative or competitive behaviors exist not only as discrete (atomic) actions,
but they are temporally extended (over policies). In the related setting of one shot Markov social
dilemmas, Lerer and Peysakhovich [176] extended the famous Tit-for-Tat (TFT)9 strategy [187]
for DRL (using function approximators) and showed (theoretically and experimentally) that such
agents can maintain cooperation. To construct the agents they used self-play and two reward
schemes: selfish and cooperative. Previously, different MAL algorithms were designed to foster
cooperation in social dilemmas with Q-learning agents [188, 189].
Self-play is a useful concept for learning algorithms (e.g., fictitious play [190]) since under certain
classes of games it can guarantee convergence10 and it has been used as a standard technique in
previous RL and MAL works [192, 14, 193]. Despite its common usage self-play can be brittle
to forgetting past knowledge [194, 172, 195] (see Section 4.5 for a note on the role of self-play
as an open question in MDRL). To overcome this issue, Leibo et al. [159] proposed Malthusian
reinforcement learning as an extension of self-play to population dynamics. The approach can be
thought of as community coevolution and has been shown to produce better results (avoiding local
optima) than independent agents with intrinsic motivation [196]. A limitation of this work is that it
does not place itself within the state of the art in evolutionary and genetic algorithms. Evolutionary
strategies have been employed for solving reinforcement learning problems [197] and for evolving
function approximators [75]. Similarly, they have been used multiagent scenarios to compute
approximate Nash equilibria [198] and as metaheuristic optimization algorithms [199, 200, 7, 201].
Bansal et al. [158] explored the emergent behaviors in competitive scenarios using the Mu-
JoCo simulator [202]. They trained independent learning agents with PPO and incorporated two
main modifications to deal with the MAL nature of the problem. First, they used exploration
rewards [203] which are dense rewards that allow agents to learn basic (non-competitive) behaviors
-- this type of reward is annealed through time giving more weight to the environmental (com-
petitive) reward. Exploration rewards come from early work in robotics [204] and single-agent
RL [205], and their goal is to provide dense feedback for the learning algorithm to improve sam-
ple efficiency (Ng et al. [206] studied the theoretical conditions under which modifications of the
reward function of an MDP preserve the optimal policy). For multiagent scenarios, these dense
rewards help agents in the beginning phase of the training to learn basic non-competitive skills,
increasing the probability of random actions from the agent yielding a positive reward. The second
9TFT originated in an iterated prisoner's dilemma tournament and later inspired different strategies in MAL [185],
its generalization, Godfather, is a representative of leader strategies [186].
10The average strategy profile of fictitious players converges to a Nash equilibrium in certain classes of games, e.g.,
two-player zero-sum and potential games [191].
13
Table 1: These papers analyze emergent behaviors in MDRL. Learning type is either value-based (VB) or policy
gradient (PG). Setting where experiments were performed: cooperative (CO), competitive (CMP) or mixed. A
detailed description is given in Section 3.3.
Work
Tampuu et al. [155]
Leibo et al. [156]
Lerer
Peysakhovich [176]
Leibo et al. [159]
and
Bansal et al. [158]
Raghu et al. [157]
Lazaridou et al. [161]
Mordatch
Abbeel [160]
and
Summary
Train DQN agents to play Pong.
Train DQN agents to play sequential social dilemmas.
Propose DRL agents able to cooperate in social dilem-
mas.
Propose Malthusian reinforcement learning which ex-
tends self-play to population dynamics.
Train PPO agents in competitive MuJoCo scenarios.
Train PPO, A3C, and DQN agents in attacker-defender
games.
Train agents represented with NN to learn a communi-
cation language.
Learn communication with an end-to-end differentiable
model to train with backpropagation.
Learning
Setting
VB
VB
VB
VB
PG
VB, PG
PG
PG
CO&CMP
Mixed
Mixed
Mixed
CMP
CMP
CO
CO
Table 2: These papers propose algorithms for learning communication. Learning type is either value-based (VB)
or policy gradient (PG). Setting were experiments were performed: cooperative (CO) or mixed. A more detailed
description is given in Section 3.4.
Algorithm
Lazaridou et al. [161]
and
Mordatch
Abbeel [160]
RIAL [162]
DIAL [162]
CommNet [163]
BiCNet [164]
MD-MADDPG [165]
MADDPG-MD [177]
Summary
Train agents represented with NN to learn a communica-
tion language.
Learn communication with an end-to-end differentiable
model to train with backpropagation.
Use a single network (parameter sharing) to train agents
that take environmental and communication actions.
Use gradient sharing during learning and communication
actions during execution.
Use a continuous vector channel for communication on a
single network.
Use the actor-critic paradigm where communication occurs
in the latent space.
Use of a shared memory as a means to multiagent commu-
nication.
Extend dropout technique to robustify communication
when applied in multiagent scenarios with direct commu-
nication.
Learning
Setting
PG
PG
VB
VB
PG
PG
PG
PG
CO
CO
CO
CO
CO
Mixed
CO
CO
14
Table 3: These papers aim to learn cooperation. Learning type is either value-based (VB) or policy gradient (PG).
Setting where experiments were performed: cooperative (CO), competitive (CMP) or mixed. A more detailed
description is given in Section 3.5.
Algorithm
Lerer
Peysakhovich [176]
MD-MADDPG [165]
and
MADDPG-MD [177]
RIAL [162]
DIAL [162]
DCH/PSRO [172]
Fingerprints [168]
Lenient-DQN [35]
Hysteretic-
DRQN [166]
WDDQN [178]
FTW [179]
VDN [180]
QMIX [181]
COMA [167]
PS-DQN, PS-TRPO,
PS-A3C [182]
MADDPG [63]
Summary
Propose DRL agents able to cooperate in social dilem-
mas.
Use of a shared memory as a means to multiagent com-
munication.
Extend dropout technique to robustify communication
when applied in multiagent scenarios with direct com-
munication.
Use a single network (parameter sharing) to train agents
that take environmental and communication actions.
Use gradient sharing during learning and communication
actions during execution.
Policies can overfit to opponents: better compute approx-
imate best responses to a mixture of policies.
Deal with ER problems in MDRL by conditioning the
value function on a fingerprint that disambiguates the
age of the sampled data.
Achieve cooperation by leniency, optimism in the value
function by forgiving suboptimal (low-rewards) actions.
Achieve cooperation by using two learning rates, depend-
ing on the updated values together with multitask learn-
ing via policy distillation.
Achieve cooperation by leniency, weighted double estima-
tors, and a modified prioritized experience replay buffer.
Agents act in a mixed environment (composed of team-
mates and opponents), it proposes a two-level architec-
ture and population-based learning.
Decompose the team action-value function into pieces
across agents, where the pieces can be easily added.
Decompose the team action-value function together with
a mixing network that can recombine them.
Use a centralized critic and a counter-factual advantage
function based on solving the multiagent credit assign-
ment.
Propose parameter sharing for learning cooperative
tasks.
Use an actor-critic approach where the critic is aug-
mented with information from other agents, the actions
of all agents.
Learning
VB
PG
PG
VB
VB
VB
VB
VB
VB
VB
PG
VB
VB
PG
Setting
Mixed
CO
CO
CO
CO
CO & CMP
CO
CO
CO
CO
Mixed
Mixed
Mixed
Mixed
VB, PG
CO
PG
Mixed
15
Table 4: These papers consider agents modeling agents. Learning type is either value-based (VB) or policy gradient
(PG). Setting where experiments were performed: cooperative (CO), competitive (CMP) or mixed. A more detailed
description is given in Section 3.6.
Algorithm
MADDPG [63]
DRON [169]
DPIQN,
DPIRQN [171]
SOM [170]
NFSP [173]
PSRO/DCH [172]
M3DDPG [183]
LOLA [64]
ToMnet [174]
Bayes-
Deep
ToMoP [175]
Deep BPR+[184]
Summary
Use an actor-critic approach where the critic is augmented
with information from other agents, the actions of all
agents.
Have a network to infer the opponent behavior together
with the standard DQN architecture.
Learn policy features from raw observations that represent
high-level opponent behaviors via auxiliary tasks.
Assume the reward function depends on a hidden goal of
both agents and then use an agent's own policy to infer the
goal of the other agent.
Compute approximate Nash equilibria via self-play and two
neural networks.
Policies can overfit to opponents: better compute approxi-
mate best responses to a mixture of policies.
Extend MADDPG with minimax objective to robustify the
learned policy.
Use a learning rule where the agent accounts for the param-
eter update of other agents to maximize its own reward.
Use an architecture for end-to-end learning and inference
of diverse opponent types.
Best respond to opponents using Bayesian policy reuse, the-
ory of mind, and deep networks.
Bayesian policy reuse and policy distillation to quickly best
respond to opponents.
Learning
PG
Setting
Mixed
VB
VB
PG
VB
PG
PG
PG
PG
VB
VB
Mixed
Mixed
Mixed
CMP
CO & CMP
Mixed
Mixed
Mixed
CMP
CO & CMP
16
contribution was opponent sampling which maintains a pool of older versions of the opponent to
sample from, in contrast to using the most recent version.
Raghu et al. [157] investigated how DRL algorithms (DQN, A2C, and PPO) performed in
a family of two-player zero-sum games with tunable complexity, called Erdos-Selfridge-Spencer
games [207, 208]. Their reasoning is threefold: (i) these games provide a parameterized family
of environments where (ii) optimal behavior can be completely characterized, and (iii) support
multiagent play. Their work showed that algorithms can exhibit wide variation in performance as
the algorithms are tuned to the game's difficulty.
Lazaridou et al. [161] proposed a framework for language learning that relies on multiagent
communication. The agents, represented by (feed-forward) neural networks, need to develop an
emergent language to solve a task. The task is formalized as a signaling game [209] in which two
agents, a sender and a receiver, obtain a pair of images. The sender is told one of them is the target
and is allowed to send a message (from a fixed vocabulary) to the receiver. Only when the receiver
identifies the target image do both agents receive a positive reward. The results show that agents
can coordinate for the experimented visual-based domain. To analyze the semantic properties11 of
the learned communication protocol they looked whether symbol usage reflects the semantics of
the visual space, and that despite some variation, many high level objects groups correspond to the
same learned symbols using a t-SNE [210] based analysis (t-SNE is a visualization technique for
high-dimensional data and it has also been used to better understand the behavior of trained DRL
agents [211, 212]). A key objective of this work was to determine if the agent's language could be
human-interpretable. To achieve this, learned symbols were grounded with natural language by
extending the signaling game with a supervised image labelling task (the sender will be encouraged
to use conventional names, making communication more transparent to humans). To measure the
interpretability of the extended game, a crowdsourced survey was performed, and in essence, the
trained agent receiver was replaced with a human. The results showed that 68% of the cases,
human participants picked the correct image.
Similarly, Mordatch and Abbeel [160] investigated the emergence of language with the difference
that in their setting there were no explicit roles for the agents (i.e., sender or receiver). To learn,
they proposed an end-to-end differentiable model of all agent and environment state dynamics over
time to calculate the gradient of the return with backpropagation.
3.4. Learning communication
As we discussed in the previous section, one of the desired emergent behaviors of multiagent
interaction is the emergence of communication [161, 160]. This setting usually considers a set of
cooperative agents in a partially observable environment (see Section 2.2) where agents need to
maximize their shared utility by means of communicating information.
Reinforced Inter-Agent Learning (RIAL) and Differentiable Inter-Agent Learning (DIAL) are
two methods using deep networks to learn to communicate [162]. Both methods use a neural
net that outputs the agent's Q values (as done in standard DRL algorithms) and a message to
communicate to other agents in the next timestep. RIAL is based on DRQN and also uses the
concept of parameter sharing, i.e., using a single network whose parameters are shared among all
agents. In contrast, DIAL directly passes gradients via the communication channel during learning,
and messages are discretized and mapped to the set of communication actions during execution.
11The vocabulary that agents use was arbitrary and had no initial meaning. To understand its emerging semantics
they looked at the relationship between symbols and the sets of images they referred to [161].
17
Memory-driven (MD) communication was proposed on top of the Multi-Agent Deep Determin-
istic Policy Gradient (MADDPG) [63] method. In MD-MADDPG [165], the agents use a shared
memory as a communication channel: before taking an action, the agent first reads the memory,
then writes a response. In this case the agent's policy becomes dependent on its private observation
and its interpretation of the collective memory. Experiments were performed with two agents in
cooperative scenarios. The results highlighted the fact that the communication channel was used
differently in each environment, e.g., in simpler tasks agents significantly decrease their memory
activity near the end of the task as there are no more changes in the environment; in more complex
environments, the changes in memory usage appear at a much higher frequency due to the presence
of many sub-tasks.
Dropout [213] is a technique to prevent overfitting (in supervised learning this happens when the
learning algorithm achieves good performance only on a specific data set and fails to generalize) in
neural networks which is based on randomly dropping units and their connections during training
time. Inspired by dropout, Kim et al. [177] proposed a similar approach in multiagent environments
where direct communication through messages is allowed. In this case, the messages of other agents
are dropped out at training time, thus the authors proposed the Message-Dropout MADDPG
algorithm [177]. This method is expected to work in fully or limited communication environments.
The empirical results show that with properly chosen message dropout rate, the proposed method
both significantly improves the training speed and the robustness of learned policies (by introducing
communication errors) during execution time. This capability is important as MDRL agents trained
in simulated or controlled environments will be less fragile when transferred to more realistic
environments.
While RIAL and DIAL used a discrete communication channel, CommNet [163] used a con-
tinuous vector channel. Through this channel agents receive the summed transmissions of other
agents. The authors assume full cooperation and train a single network for all the agents. There
are two distinctive characteristics of CommNet from previous works: it allows multiple communi-
cation cycles at each timestep and a dynamic variation of agents at run time, i.e., agents come and
go in the environment.
In contrast to previous approaches, in Multiagent Bidirectionally Coordinated Network (BiC-
Net) [164], communication takes place in the latent space (i.e., in the hidden layers). It also uses
parameter sharing, however, it proposes bidirectional recurrent neural networks [214] to model the
actor and critic networks of their model. Note that in BiCNet agents do not explicitly share a
message and thus it can be considered a method for learning cooperation.
Learning communication is an active area in MDRL with many open questions, in this context,
we refer the interested reader to a recent work by Lowe et al. [215] where it discusses common pitfalls
(and recommendations to avoid those) while measuring communication in multiagent environments.
3.5. Learning cooperation
Although explicit communication is a new emerging trend in MDRL, there has already been a
large amount of work in MAL for cooperative settings12 that do not involve communication [7, 8].
Therefore, it was a natural starting point for many recent MDRL works.
Foerster et al. [168] studied the simple scenario of cooperation with independent Q-learning
agents (see Section 3.1), where the agents use the standard DQN architecture of neural networks
12There is a large body of research on coordinating multiagent teams by specifying communication protocols [216,
217]: these expect agents to know the team's goal as well as the tasks required to accomplish the goal.
18
and an experience replay buffer (see Figure 3). However, for the ER to work, the data distribution
needs to follow certain assumptions (see Section 2.2) which are no loger valid due to the multiagent
nature of the world: the dynamics that generated the data in the ER no longer reflect the current
dynamics, making the experience obsolete [168, 90]. Their solution is to add information to the
experience tuple that can help to disambiguate the age of the sampled data from the replay memory.
Two approaches were proposed. The first is Multiagent Importance Sampling which adds the
probability of the joint action so an importance sampling correction [70, 218] can computed when
the tuple is later sampled for training. This was similar to previous works in adaptive importance
sampling [219, 220] and off-environment RL [221]. The second approach is Multiagent Fingerprints
which adds the estimate (i.e., fingerprint) of other agents' policies (loosely inspired by Hyper-
Q [150], see Section 4.1). For the practical implementation, good results were obtained by using
the training iteration number and exploration rate as the fingerprint.
Gupta et al. [182] tackled cooperative environments in partially observable domains without
explicit communication. They proposed parameter sharing (PS) as a way to improve learning in
homogeneous multiagent environments (where agents have the same set of actions). The idea is to
have one globally shared learning network that can still behave differently in execution time, i.e.,
because its inputs (individual agent observation and agent index) will be different. They tested
three variations of this approach with parameter sharing: PS-DQN, PS-DDPG and PS-TRPO,
which extended single-agent DQN, DDPG and TRPO algorithms, respectively. The results showed
that PS-TRPO outperformed the other two. Note that Foerster et al. [162] concurrently proposed
a similar concept, see Section 3.4.
Lenient-DQN (LDQN) [35] took the leniency concept [222] (originally presented in MAL) and
extended their use to MDRL. The purpose of leniency is to overcome a pathology called rela-
tive overgeneralization [34, 223, 224]. Similar to other approaches designed to overcome relative
overgeneralization (e.g., distributed Q-learning [225] and hysteretic Q-learning [8]) lenient learn-
ers initially maintain an optimistic disposition to mitigate the noise from transitions resulting in
miscoordination, preventing agents from being drawn towards sub-optimal but wide peaks in the
reward search space [97]. However, similar to other MDRL works [168], the LDQN authors expe-
rienced problems with the ER buffer and arrived at a similar solution: adding information to the
experience tuple, in their case, the leniency value. When sampling from the ER buffer, this value
is used to determine a leniency condition; if the condition is not met then the sample is ignored.
In a similar vein, Decentralized-Hysteretic Deep Recurrent Q-Networks (DEC-HDRQNs) [166]
were proposed for fostering cooperation among independent learners. The motivation is similar
to LDQN, making an optimistic value update, however, their solution is different. Here, the
authors took inspiration from Hysteretic Q-learning [8], originally presented in MAL, where two
learning rates were used. A difference between lenient agents and hysteretic Q-learning is that
lenient agents are only initially forgiving towards teammates. Lenient learners over time apply
less leniency towards updates that would lower utility values, taking into account how frequently
observation-action pairs have been encountered. The idea being that the transition from optimistic
to average reward learner will help make lenient learners more robust towards misleading stochastic
rewards [222]. Additionally, in DEC-HDRQNs the ER buffer is also extended into concurrent
experience replay trajectories, which are composed of three dimensions: agent index, the episode,
and the timestep; when training, the sampled traces have the same starting timesteps. Moreover,
to improve on generalization over different tasks, i.e., multi-task learning[226], DEC-HDRQNs
make use of policy distillation [227, 228] (see Section 4.1). In contrast to other approaches, DEC-
19
Figure 5: A schematic view of the architecture used in FTW (For the Win) [179]: two unrolled recurrent neural
networks (RNNs) operate at different time-scales, the idea is that the Slow RNN helps with long term temporal
correlations. Observations are latent space output of some convolutional neural network to learn non-linear features.
Feudal Networks [229] is another work in single-agent DRL that also maintains a multi-time scale hierarchy where
the slower network sets the goal, and the faster network tries to achieve them. Fedual Networks were in turn, inspired
by early work in RL which proposed a hierarchy of Q-learners [230, 231].
HDRQNS are fully decentralized during learning and execution.
Weighted Double Deep Q-Network (WDDQN) [178] is based on having double estimators. This
idea was originally introduced in Double Q-learning [91] and aims to remove the existing overes-
timation bias caused by using the maximum action value as an approximation for the maximum
expected action value (see Section 4.1). It also uses a lenient reward [222] to be optimistic during
initial phase of coordination and proposes a scheduled replay strategy in which samples closer to
the terminal states are heuristically given higher priority; this strategy might not be applicable
for any domain. For other works extending the ER to multiagent settings see MADDPG [63],
Sections 4.1 and 4.2.
While previous approaches were mostly inspired by how MAL algorithms could be extended
to MDRL, other works take as base the results by single-agent DRL. One example is the For The
Win (FTW) [179] agent which is based on the actor-learner structure of IMPALA [114] (see Sec-
tion 2.2). The authors test FTW in a game where two opposing teams compete to capture each
other's flags [232]. To deal with the MAL problem they propose two main additions: a hierarchical
two-level representation with recurrent neural networks operating at different timescales, as de-
picted in Figure 5, and a population based training [233, 234, 235] where 30 agents were trained in
parallel together with a stochastic matchmaking scheme that biases agents to be of similar skills.
20
The Elo rating system [236] was originally devised to rate chess player skills,13 TrueSkill [237]
extended Elo by tracking uncertainty in skill rating, supporting draws, and matches beyond 1 vs
1; α−Rank is a more recent alternative to ELO [238]. FTW did not use TrueSkill but a sim-
pler extension of Elo for n vs n games (by adding individual agent ratings to compute the team
skill). Hierarchical approaches were previously proposed in RL, e.g., Feudal RL [230, 231], and
were later extended to DRL in Feudal networks [229]; population based training can be considered
analogous to evolutionary strategies that employ self-adaptive hyperparameter tuning to modify
how the genetic algorithm itself operates [234, 239, 240]. An interesting result from FTW is that
the population-based training obtained better results than training via self-play [192], which was
a standard concept in previous works [14, 193]. FTW used heavy compute resources, it used 30
agents (processes) in parallel where every training game lasted 4500 agent steps (≈ five minutes)
and agents were trained for two billion steps (≈ 450K games).
Lowe et al. [63] noted that using standard policy gradient methods (see Section 2.1) on mul-
tiagent environments yields high variance and performs poorly. This occurs because the variance
is further increased as all the agents' rewards depend on the rest of the agents, and it is for-
mally shown that as the number of agents increase, the probability of taking a correct gradient
direction decreases exponentially [63]. Therefore, to overcome this issue Lowe et al. proposed the
Multi-Agent Deep Deterministic Policy Gradient (MADDPG) [63], building on DDPG [65] (see
Section 2.2), to train a centralized critic per agent that is given all agents' policies during train-
ing to reduce the variance by removing the non-stationarity caused by the concurrently learning
agents. Here, the actor only has local information (turning the method into a centralized training
with decentralized execution) and the ER buffer records experiences of all agents. MADDPG was
tested in both cooperative and competitive scenarios, experimental results show that it performs
better than several decentralized methods (such as DQN, DDPG, and TRPO). The authors men-
tion that traditional RL methods do not produce consistent gradient signals. This is exemplified
in a challenging competitive scenarios where agents continuously adapt to each other causing the
learned best-response policies oscillate -- for such a domain, MADDPG is shown to learn more
robustly than DDPG.
Another approach based on policy gradients is the Counterfactual Multi-Agent Policy Gradients
(COMA) [167]. COMA was designed for the fully centralized setting and the multiagent credit
assignment problem [241], i.e., how the agents should deduce their contributions when learning
in a cooperative setting in the presence of only global rewards. Their proposal is to compute a
counterfactual baseline, that is, marginalize out the action of the agent while keeping the rest of
the other agents' actions fixed. Then, an advantage function can be computed comparing the
current Q value to the counterfactual. This counterfactual baseline has its roots in difference
rewards, which is a method for obtaining the individual contribution of an agent in a cooperative
multiagent team [241]. In particular, the aristocrat utility aims to measure the difference between
an agent's actual action and the average action [31]. The intention would be equivalent to sideline
the agent by having the agent perform an action where the reward does not depend on the agent's
actions, i.e., to consider the reward that would have arisen assuming a world without that agent
having ever existed (see Section 4.2).
On the one hand, fully centralized approaches (e.g., COMA) do not suffer from non-stationarity
13Elo uses a normal distribution for each player skill, and after each match, both players' distributions are updated
based on measure of surprise, i.e., if a user with previously lower (predicted) skill beats a high skilled one, the
low-skilled player is significantly increased.
21
Figure 6: (a) Deep Policy Inference Q-Network: receives four stacked frames as input (similar to DQN, see Figure 2).
(b) Deep Policy Inference Recurrent Q-Network: receives one frame as input and has an LSTM layer instead of a
fully connected layer (FC). Both approaches [171] condition the QM value outputs on the policy features, hP I , which
are also used to learn the opponent policy πo.
but have constrained scalability. On the other hand, independent learning agents are better suited
to scale but suffer from non-stationarity issues. There are some hybrid approaches that learn a
centralized but factored Q value function [242, 243]. Value Decomposition Networks (VDNs) [180]
decompose a team value function into an additive decomposition of the individual value functions.
Similarly, QMIX [181] relies on the idea of factorizing, however, instead of sum, QMIX assumes a
mixing network that combines the local values in a non-linear way, which can represent monotonic
action-value functions. While the mentioned approaches have obtained good empirical results, the
factorization of value-functions in multiagent scenarios using function approximators (MDRL) is
an ongoing research topic, with open questions such as how well factorizations capture complex
coordination problems and how to learn those factorizations [244] (see Section 4.4).
3.6. Agents modeling agents
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 [11]. An early work for
modeling agents while using deep neural networks was the Deep Reinforcement Opponent Network
(DRON) [169]. The idea is to have two networks: one which evaluates Q-values and a second
one that learns a representation of the opponent's policy. Moreover, the authors proposed to have
several expert networks to combine their predictions to get the estimated Q value, the idea being
that each expert network captures one type of opponent strategy [245]. This is related to previous
22
works in type-based reasoning from game theory [246, 139] later applied in AI [245, 11, 247]. The
mixture of experts idea was presented in supervised learning where each expert handled a subset of
the data (a subtask), and then a gating network decided which of the experts should be used [248].
DRON uses hand-crafted features to define the opponent network. In contrast, Deep Policy
Inference Q-Network (DPIQN) and its recurrent version, DPIRQN [171] learn policy features di-
rectly from raw observations of the other agents. The way to learn these policy features is by
means of auxiliary tasks [84, 110] (see Sections 2.2 and 4.1) that provide additional learning goals,
in this case, the auxiliary task is to learn the opponents' policies. This auxiliary task modifies the
loss function by computing an auxiliary loss: the cross entropy loss between the inferred opponent
policy and the ground truth (one-hot action vector) of the opponent. Then, the Q value function
of the learning agent is conditioned on the opponent's policy features (see Figure 6), which aims to
reduce the non-stationarity of the environment. The authors used an adaptive training procedure
to adjust the attention (a weight on the loss function) to either emphasize learning the policy fea-
tures (of the opponent) or the respective Q values of the agent. An advantage of these approaches
is that modeling the agents can work for both opponents and teammates [171].
In many previous works an opponent model is learned from observations. Self Other Modeling
(SOM) [170] proposed a different approach, this is, using the agent's own policy as a means to
predict the opponent's actions. SOM can be used in cooperative and competitive settings (with
an arbitrary number of agents) and infers other agents' goals. This is important because in the
evaluated domains, the reward function depends on the goal of the agents. SOM uses two networks,
one used for computing the agents' own policy, and a second one used to infer the opponent's goal.
The idea is that these networks have the same input parameters but with different values (the
agent's or the opponent's). In contrast to previous approaches, SOM is not focused on learning
the opponent policy, i.e., a probability distribution over next actions, but rather on estimating the
opponent's goal. SOM is expected to work best when agents share a set of goals from which each
agent gets assigned one at the beginning of the episode and the reward structure depends on both
of their assigned goals. Despite its simplicity, training takes longer as an additional optimization
step is performed given the other agent's observed actions.
There is a long-standing history of combining game theory and MAL [2, 6, 193]. From that
context, some approaches were inspired by influential game theory approaches. Neural Fictitious
Self-Play (NFSP) [173] builds on fictitious (self-) play [190, 249], together with two deep networks to
find approximate Nash equilibria 14 in two-player imperfect information games [251] (for example,
consider Poker: when it is an agent's turn to move it does not have access to all information
about the world). One network learns an approximate best response (−greedy over Q values) to
the historical behavior of other agents and the second one (called the average network) learns to
imitate its own past best response behaviour using supervised classification. The agent behaves
using a mixture of the average and the best response networks depending on the probability of an
anticipatory parameter [252]. Comparisons with DQN in Leduc Holdem Poker revealed that DQN's
deterministic strategy is highly exploitable. Such strategies are sufficient to behave optimally in
single-agent domains, i.e., MDPs for which DQN was designed. However, imperfect-information
games generally require stochastic strategies to achieve optimal behaviour [173]. DQN learning
experiences are both highly correlated over time, and highly focused on a narrow state distribution.
14Nash equilibrium [250] is a solution concept in game theory in which no agent would choose to deviate from its
strategy (they are a best response to others' strategies). This concept has been explored in seminal MAL algorithms
like Nash-Q learning [127] and Minimax-Q learning [124, 128].
23
In contrast to NFSP agents whose experience varies more smoothly, resulting in a more stable data
distribution, more stable neural networks and better performance.
The (N)FSP concept was further generalized in Policy-Space Response Oracles (PSRO) [172],
where it was shown that fictitious play is one specific meta-strategy distribution over a set of
previous (approximate) best responses (summarized by a meta-game obtained by empirical game
theoretic analysis [253]), but there are a wide variety to choose from. One reason to use mixed meta-
strategies is that it prevents overfitting15 the responses to one specific policy, and hence provides a
form of opponent/teammate regularization. An approximate scalable version of the algorithm leads
to a graph of agents best-responding independently called Deep Cognitive Hierarchies (DCHs) [172]
due to its similarity to behavioral game-theoretic models [255, 256].
Minimax is a paramount concept in game theory that is roughly described as minimizing the
worst case scenario (maximum loss) [251]. Li et al. [183] took the minimax idea as an approach
to robustify learning in multiagent environments so that the learned robust policy should be able
to behave well even with strategies not seen during training. They extended the MADDPG algo-
rithm [63] to Minimax Multiagent Deep Deterministic Policy Gradients (M3DDPG), which updates
policies considering a worst-case scenario: assuming that all other agents act adversarially. This
yields a minimax learning objective which is computationally intractable to directly optimize. They
address this issue by taking ideas from robust reinforcement learning [257] which implicitly adopts
the minimax idea by using the worst noise concept [258]. In MAL different approaches were pro-
posed to assess the robustness of an algorithm, e.g., guarantees of safety [152, 259], security [260]
or exploitability [261, 262, 263].
Previous approaches usually learned a model of the other agents as a way to predict their
behavior. However, they do not explicitly account for anticipated learning of the other agents, which
is the objective of Learning with Opponent-Learning Awareness (LOLA) [64]. LOLA optimizes the
expected return after the opponent updates its policy one step. Therefore, a LOLA agent directly
shapes the policy updates of other agents to maximize its own reward. One of LOLA's assumptions
is having access to opponents' policy parameters. LOLA builds on previous ideas by Zhang and
Lesser [264] where the learning agent predicts the opponent's policy parameter update but only
uses it to learn a best response (to the anticipated updated parameters).
Theory of mind is part of a group of recursive reasoning approaches[265, 245, 266, 267] in
which agents have explicit beliefs about the mental states of other agents. The mental states
of other agents may, in turn, also contain beliefs and mental states of other agents, leading to
a nesting of beliefs [11]. Theory of Mind Network (ToMnet) [174] starts with a simple premise:
when encountering a novel opponent, the agent should already have a strong and rich prior about
how the opponent should behave. ToMnet has an architecture composed of three networks: (i) a
character network that learns from historical information, (ii) a mental state network that takes the
character output and the recent trajectory, and (iii) the prediction network that takes the current
state as well as the outputs of the other networks as its input. The output of the architecture
is open for different problems but in general its goal is to predict the opponent's next action. A
main advantage of ToMnet is that it can predict general behavior, for all agents; or specific, for a
particular agent.
Deep Bayesian Theory of Mind Policy (Bayes-ToMoP) [175] is another algorithm that takes
inspiration from theory of mind [268]. The algorithm assumes the opponent has different stationary
15Johanson et al. [254] also found "overfitting" when solving large extensive games (e.g., poker) -- the performance
in an abstract game improved but it was worse in the full game.
24
strategies to act and changes among them over time [269]. Earlier work in MAL dealt with
this setting, e.g., BPR+ [270] extends the Bayesian policy reuse16 framework [271] to multiagent
settings (BPR assumes a single-agent environment; BPR+ aims to best respond to the opponent
in a multiagent game). A limitation of BPR+ is that it behaves poorly against itself (self-play),
thus, Deep Bayes-ToMoP uses theory of mind to provide a higher-level reasoning strategy which
provides an optimal behavior against BPR+ agents.
Deep BPR+ [184] is another work inspired by BPR+ which uses neural networks as value-
function approximators. It not only uses the environment reward but also uses the online learned
opponent model [272, 273] to construct a rectified belief over the opponent strategy. Additionally, it
leverages ideas from policy distillation [227, 228] and extends them to the multiagent case to create a
distilled policy network. In this case, whenever a new acting policy is learned, distillation is applied
to consolidate the new updated library which improves in terms of storage and generalization (over
opponents).
4. Bridging RL, MAL and MDRL
This section aims to provide directions to promote fruitful cooperations between sub-communities.
First, we address the pitfall of deep learning amnesia, roughly described as missing citations to
the original works and not exploiting the advancements that have been made in the past. We
present examples on how ideas originated earlier, for example in RL and MAL, were successfully
extended to MDRL (see Section 4.1). Second, we outline lessons learned from the works analyzed
in this survey (see Section 4.2). Then we point the readers to recent benchmarks for MDRL (see
Section 4.3) and we discuss the practical challenges that arise in MDRL like high computational
demands and reproducibility (see Section 4.4). Lastly, we pose some open research challenges and
reflect on their relation with previous open questions in MAL [11] (see Section 4.5).
4.1. Avoiding deep learning amnesia: examples in MDRL
This survey focuses on recent deep works, however, in previous sections, when describing recent
algorithms, we also point to original works that inspired them. Schmidhuber said "Machine learning
is the science of credit assignment. The machine learning community itself profits from proper
credit assignment to its members" [274]. In this context, we want to avoid committing the pitfall
of not giving credit to original ideas that were proposed earlier, a.k.a. deep learning amnesia.
Here, we provide some specific examples of research milestones that were studied earlier, e.g., RL
or MAL, and that now became highly relevant for MDRL. Our purpose is to highlight that existent
literature contains pertinent ideas and algorithms that should not be ignored. On the contrary,
they should be examined and cited [275, 276] to understand recent developments [277].
Dealing with non-stationarity in independent learners. It is well known that using independent
learners makes the environment non-stationary from the agent's point of view [4, 123]. Many MAL
algorithms tried to solve this problem in different ways [10]. One example is Hyper-Q [150] which
accounts for the (values of mixed) strategies of other agents and includes that information in the
state representation, which effectively turns the learning problem into a stationary one. Note that in
16Bayesian policy reuse assumes an agent with prior experience in the form of a library of policies. When a novel
task instance occurs, the objective is to reuse a policy from its library based on observed signals which correlate to
policy performance [271].
25
this way it is possible to even consider adaptive agents. Foerster et al. [162] make use of this insight
to propose their fingerprint algorithm in an MDRL problem (see Section 3.5). Other examples
include the leniency concept [222] and Hysteretic Q-learning [8] originally presented in MAL, which
now have their "deep" counterparts, LDQNs [35] and DEC-HDRQNs[166], see Section 3.5.
Multiagent credit assignment. In cooperative multiagent scenarios, it is common to use either
local rewards, unique for each agent, or global rewards, which represent the entire group's perfor-
mance [278]. However, local rewards are usually harder to obtain, therefore, it is common to rely
only on the global ones. This raises the problem of credit assignment: how does a single agent's ac-
tions contribute to a system that involves the actions of many agents [32]. A solution that came from
MAL research that has proven successful in many scenarios is difference rewards [241, 278, 279],
which aims to capture an agent's contribution to the system's global performance. In particular
the aristocrat utility aims to measure the difference between an agents actual action and the av-
erage action [31], however, it has a self-consistency problem and in practice it is more common
to compute the wonderful life utility [280, 31], which proposes to use a clamping operation that
would be equivalent to removing that player from the team. COMA [167] builds on these concepts
to propose an advantage function based on the contribution of the agent, which can be efficiently
computed with deep neural networks (see Section 3.5).
Multitask learning. In the context of RL, multitask learning [226] is an area that develops agents
that can act in several related tasks rather than just in a single one [281]. Distillation, roughly
defined as transferring the knowledge from a large model to a small model, was a concept originally
introduced for supervised learning and model compression [282, 228].
Inspired by those works,
Policy distillation [227] was extended to the DRL realm. Policy distillation was used to train a much
smaller network and to merge several task-specific policies into a single policy, i.e., for multitask
learning.
In the MDRL setting, Omidshafiei et al. [166] successfully adapted policy distillation
within Dec-HDRQNs to obtain a more general multitask multiagent network (see Section 3.5).
Another example is Deep BPR+ [184] which uses distillation to generalize over multiple opponents
(see Section 3.6).
Auxiliary tasks. Jaderberg et al. [84] introduced the term auxiliary task with the insight that
(single-agent) environments contain a variety of possible training signals (e.g., pixel changes). These
tasks are naturally implemented in DRL in which the last layer is split into multiple parts (heads),
each working on a different task. All heads propagate errors into the same shared preceding part
of the network, which would then try to form representations, in its next-to-last layer, to support
all the heads [20]. However, the idea of multiple predictions about arbitrary signals was originally
suggested for RL, in the context of general value functions [110, 20] and there still open problems,
for example, better theoretical understanding [109, 283]. In the context of neural networks, early
work proposed hints that improved the network performance and learning time. Suddarth and
Kergosien [284] presented a minimal example of a small neural network where it was shown that
adding an auxiliary task effectively removed local minima. One could think of extending these
auxiliary tasks to modeling other agents' behaviors [285, 160], which is one of the key ideas that
DPIQN and DRPIQN [171] proposed in MDRL settings (see Section 3.6).
Experience replay. Lin [90, 89] proposed the concept of experience replay to speed up the credit
assignment propagation process in single agent RL. This concept became central to many DRL
works [72] (see Section 2.2). However, Lin stated that a condition for the ER to be useful is
26
that "the environment should not change over time because this makes past experiences irrelevant
or even harmful" [90]. This is a problem in domains where many agents are learning since the
environment becomes non-stationary from the point of view of each agent. Since DRL relies
heavily on experience replay, this is an issue in MDRL: the non-stationarity introduced means that
the dynamics that generated the data in the agent's replay memory no longer reflect the current
dynamics in which it is learning [162]. To overcome this problem different methods have been
proposed [168, 35, 166, 178], see Section 4.2.
Double estimators. Double Q-learning [91] proposed to reduce the overestimation of action values
in Q-learning, this is caused by using the maximum action value as an approximation for the
maximum expected action value. Double Q-learning works by keeping two Q functions and was
proven to convergence to the optimal policy [91]. Later this idea was applied to arbitrary function
approximators, including deep neural networks, i.e., Double DQN [92], which were naturally applied
since two networks were already used in DQN (see Section 2.2). These ideas have also been recently
applied to MDRL [178].
4.2. Lessons learned
We have exemplified how RL and MAL can be extended for MDRL settings. Now, we outline
general best practices learned from the works analyzed throughout this paper.
• Experience replay buffer in MDRL. While some works removed the ER buffer in MDRL [162]
it is an important component in many DRL and MDRL algorithms. However, using the
standard buffer (i.e., keeping (cid:104)s, a, r, s(cid:48)(cid:105)) will probably fail due to a lack of theoretical guar-
antees under this setting, see Sections 2.2 and 4.1. Adding information in the experience
tuple that can help disambiguate the sample is the solution adopted in many works, whether
a value based method [168, 35, 166, 178] or a policy gradient method [63]. In this regard,
it is an open question to consider how new DRL ideas could be best integrated into the
ER [286, 111, 287, 288, 96] and how those ideas would fare in a MDRL setting.
• Centralized learning with decentralized execution. Many MAL works were either fully cen-
tralized or fully decentralized approaches. However, inspired by decentralized partially ob-
servable Markov decison processes (DEC-POMDPs) [289, 290],17 in MDRL this new mixed
paradigm has been commonly used [168, 35, 181, 172, 167, 63] (a notable exception are
DEC-HDRQNs [166] which perform learning and execution in a decentralized manner, see
Section 3.5). Note that not all real-world problems fit into this paradigm and it is more
common for robotics or games where a simulator is generally available [162]. The main ben-
efit is that during learning additional information can be used (e.g., global state, action, or
rewards) and during execution this information is removed.
• Parameter sharing. Another frequent component in many MDRL works is the idea of sharing
parameters, i.e., training a single network in which agents share their weights. Note that, since
agents could receive different observations (e.g., in partially observable scenarios), they can
still behave differently. This method was proposed concurrently in different works [292, 162]
and later it has been successfully applied in many others [163, 164, 168, 180, 181].
17Centralized planning and decentralized execution is also a standard paradigm for multiagent planning [291].
27
(a) Multiagent object transportation
(b) Pommerman
Figure 7: (a) A fully cooperative benchmark with two agents, Multiagent Object Trasportation.
cooperative-competitive domain with four agents, Pommerman. For more MDRL benchmarks see Section 4.3.
(b) A mixed
• Recurrent networks. Recurrent neural networks (RNNs) enhanced neural networks with a
memory capability, however, they suffer from the vanishing gradient problem, which ren-
ders them inefficient for long-term dependencies [293]. However, RNN variants such as
LSTMs [86, 294] and GRUs (Gated Recurrent Unit) [295] addressed this challenge. In single-
agent DRL, DRQN [85] initially proposed idea of using recurrent networks in single-agent
partially observable environments. Then, Feudal Networks [229] proposed a hierarchical ap-
proach [230], multiple LSTM networks with different time-scales, i.e., the observation input
schedule is different for each LSTM network, to create a temporal hierarchy so that it can
better address the long-term credit assignment challenge for RL problems. Recently, the
use of recurrent networks has been extended to MDRL to address the challenge of partially
observability [158, 162, 164, 166, 180, 181, 170, 171, 174] for example, in FTW [179], depicted
in Figure 5 and DRPIRQN [171] depicted in Figure 6. See Section 4.4 for practical challenges
(e.g., training issues) of recurrent networks in MDRL.
• Overfitting in MAL. In single-agent RL, agents can overfit to the environment [296]. A similar
problem can occur in multiagent settings [254], agents can overfit, i.e., an agent's policy can
easily get stuck in a local optima and the learned policy may be only locally optimal to other
agents' current policies [183]. This has the effect of limiting the generalization of the learned
policies [172]. To reduce this problem, a solution is to have a set of policies (an ensemble)
and learn from them or best respond to the mixture of them [172, 63, 169]. Another solution
has been to robustify algorithms -- a robust policy should be able to behave well even with
strategies different from its training (better generalization) [183].
4.3. Benchmarks for MDRL
Standardized environments such as the Arcade Learning Environment (ALE) [297, 298] and
OpenAI Gym [299] have allowed single-agent RL to move beyond toy domains. For DRL there are
open-source frameworks that provide compact and reliable implementations of some state-of-the-
art DRL algorithms [300]. Even though MDRL is a recent area, there are now a number of open
sourced simulators and benchmarks to use with different characteristics, which we describe below.
28
• Fully Cooperative Multiagent Object Transporation Problems (CMOTPs)18 were originally
presented by Busoniu et al. [36] as a simple two-agent coordination problem in MAL. Palmer
et al. [35] proposed two pixel-based extensions to the original setting which include narrow
passages that test the agents' ability to master fully-cooperative sub-tasks, stochastic rewards
and noisy observations, see Figure 7a.
• The Apprentice Firemen Game19 (inspired by the classic climb game [126]) is another two-
agent pixel-based environment that simultaneously confronts learners with four pathologies
in MAL: relative overgeneralization, stochasticity, the moving target problem, and alter ex-
ploration problem [97].
• Pommerman [301] is a multiagent benchmark useful for testing cooperative, competitive
and mixed (cooperative and competitive) scenarios.
It supports partial observability and
communication among agents, see Figure 7b. Pommerman is a very challenging domain
from the exploration perspective as the rewards are very sparse and delayed [302]. A recent
competition was held during NeurIPS-201820 and the top agents from that competition are
available for training purposes.
• Starcraft Multiagent Challenge [303] is based on the real-time strategy game StarCraft II
and focuses on micromanagement challenges,21 that is, fine-grained control of individual
units, where each unit is controlled by an independent agent that must act based on local
observations. It is accompanied by a MDRL framework including state-of-the-art algorithms
(e.g., QMIX and COMA).22
• The Multi-Agent Reinforcement Learning in Malmo (MARL O) competition [304] is another
multiagent challenge with multiple cooperative 3D games23 within Minecraft. The scenarios
were created with the open source Malmo platform [305], providing examples of how a wider
range of multiagent cooperative, competitive and mixed scenarios can be experimented on
within Minecraft.
• Hanabi is a cooperative multiplayer card game (two to five players). The main characteristic
of the game is that players do not observe their own cards but other players can reveal
information about them. This makes an interesting challenge for learning algorithms in
particular in the context of self-play learning and ad-hoc teams [306, 307, 308]. The Hanabi
Learning Environment [309] was recently released24 and it is accompanied with a baseline
(deep RL) agent [310].
• Arena [311] is platform for multiagent research25 based on the Unity engine [312]. It has 35
multiagent games (e.g., social dilemmas) and supports communication among agents. It has
basseline implementations of recent DRL algorithms such as independent PPO learners.
18https://github.com/gjp1203/nui_in_madrl
19https://github.com/gjp1203/nui_in_madrl
20https://www.pommerman.com/
21https://github.com/oxwhirl/smac
22https://github.com/oxwhirl/pymarl
23https://github.com/crowdAI/marlo-single-agent-starter-kit/
24https://github.com/deepmind/hanabi-learning-environment
25https://github.com/YuhangSong/Arena-BuildingToolkit
29
• MuJoCo Multiagent Soccer [313] uses the MuJoCo physics engine [202]. The environment
simulates a 2 vs. 2 soccer game with agents having a 3-dimensional action space.26
• Neural MMO [314] is a research platform27 inspired by the human game genre of Massively
Multiplayer Online (MMO) Role-Playing Games. These games involve a large, variable num-
ber of players competing to survive.
4.4. Practical challenges in MDRL
In this section we take a more critical view with respect to MDRL and highlight different
practical challenges that already happen in DRL and that are likely to occur in MDRL such
as reproducibility, hyperparameter tuning, the need of computational resources and conflation of
results. We provide pointers on how we think those challenges could be (partially) addressed.
Reproducibility, troubling trends and negative results. Reproducibility is a challenge in RL which
is only aggravated in DRL due to different sources of stochasticity: baselines, hyperparameters,
architectures [315, 316] and random seeds [317]. Moreover, DRL does not have common practices
for statistical testing [318] which has led to bad practices such as only reporting the results when
algorithms perform well, sometimes referred as cherry picking [319] (Azizzadenesheli also describes
cherry planting as adapting an environment to a specific algorithm [319]). We believe that together
with following the advice on how to design experiments and report results [320], the community
would also benefit from reporting negative results [321, 322, 318, 323] for carefully designed hypoth-
esis and experiments.28 However, we found very few papers with this characteristic[324, 325, 326]
-- we note that this is not encouraged in the ML community; moreover, negative results reduce
the chance of paper acceptance [320]. In this regard, we ask the community to reflect on these
practices and find ways to remove these obstacles.
Implementation challenges and hyperparameter tuning. One problem is that canonical implementa-
tions of DRL algorithms often contain additional non-trivial optimizations -- these are sometimes
necessary for the algorithms to achieve good performance [79]. A recent study by Tucker et al. [59]
found that several published works on action-dependant baselines contained bugs and errors --
those were the real reason of the high performance in the experimental results, not the proposed
method. Melis et al. [327] compared a series of works with increasing innovations in network ar-
chitectures and the vanilla LSTMs [86] (originally proposed in 1997). The results showed that,
when properly tuned, LSTMs outperformed the more recent models. In this context, Lipton and
Steinhardt noted that the community may have benefited more by learning the details of the hy-
perparameter tuning [320]. A partial reason for this surprising result might be that this type of
networks are known for being difficult to train [293] and there are recent works in DRL that re-
port problems when using recurrent networks [182, 328, 329, 330]. Another known complication is
catastrophic forgetting (see Section 2.2) with recent examples in DRL [157, 92] -- we expect that
26https://github.com/deepmind/dm_control/tree/master/dm_control/locomotion/soccer
27https://github.com/openai/neural-mmo
28This idea was initially inspired by the Workshop "Critiquing and Correcting Trends in Machine Learning" at
NeurIPS 2018 where it was possible to submit Negative results papers: "Papers which show failure modes of existing
algorithms or suggest new approaches which one might expect to perform well but which do not. The aim is to
provide a venue for work which might otherwise go unpublished but which is still of interest to the community."
https://ml-critique-correct.github.io/
30
these issues would likely occur in MDRL. The effects of hyperparameter tuning were analyzed in
more detail in DRL by Henderson et al. [315], who arrived at the conclusion that hyperparameters
can have significantly different effects across algorithms (they tested TRPO, DDPG, PPO and
ACKTR) and environments since there is an intricate interplay among them [315]. The authors
urge the community to report all parameters used in the experimental evaluations for accurate
comparison -- we encourage a similar behavior for MDRL. Note that hyperparameter tuning is re-
lated to the troubling trend of cherry picking in that it can show a carefully picked set of parameters
that make an algorithm work (see previous challenge). Lastly, note that hyperparameter tuning
is computationally very expensive, which relates to the connection with the following challenge of
computational demands.
Computational resources. Deep RL usually requires millions of interactions for an agent to learn [331],
i.e., low sample efficiency [332], which highlights the need for large computational infrastructure
in general. The original A3C implementation [93] uses 16 CPU workers for 4 days to learn to play
an Atari game with a total of 200M training frames29 (results are reported for 57 Atari games).
Distributed PPO used 64 workers (presumably one CPU per worker, although this is not clearly
stated in the paper) for 100 hours (more than 4 days) to learn locomotion tasks [117]. In MDRL,
for example, the Atari Pong game, agents were trained for 50 epochs, 250k time steps each, for a
total of 1.25M training frames [155]. The FTW agent [179] uses 30 agents (processes) in parallel
and every training game lasts for five minues; FTW agents were trained for approximately 450K
games ≈4.2 years. These examples highlight the computational demands sometimes needed within
DRL and MDRL.
Recent works have reduced the learning of an Atari game to minutes (Stooke and Abbeel [334]
trained DRL agents in less than one hour with hardware consisting of 8 GPUs and 40 cores).
However, this is (for now) the exception and computational infrastructure is a major bottleneck
for doing DRL and MDRL, especially for those who do not have such large compute power (e.g.,
most companies and most academic research groups) [212, 322].30 Within this context we propose
two ways to address this problem. (1) Raising awareness: For DRL we found few works that study
the computational demands of recent algorithms [335, 331]. For MDRL most published works do
not provide information regarding computational resources used such as CPU/GPU usage, memory
demands, and wall-clock computation. Therefore, the first way to tackle this issue is by raising
awareness and encouraging authors to report metrics about computational demands for accurately
comparison and evaluation.
(2) Delve into algorithmic contributions. Another way to address
these issues is to prioritize the algorithmic contribution for the new MDRL algorithms rather than
the computational resources spent.
Indeed, for this to work, it needs to be accompanied with
high-quality reviewers.
We have argued to raise awareness on the computational demands and report results, however,
there is still the open question on how and what to measure/report. There are several dimensions
to measure efficiency: sample efficiency is commonly measured by counting state-action pairs used
for training; computational efficiency could be measured by number of CPUs/GPUs and days used
for training. How do we measure the impact of other resources, such as external data sources or
29It is sometimes unclear in the literature what is the meaning of frame due to the "frame skip" technique. It is
therefore suggested to refer to "game frames" and "training frames" [333].
30One recent effort by Beeching et al. [212] proposes to use only "mid-range hardware" (8 CPUs and 1 GPU) to
train deep RL agents.
31
annotations?31 Similarly, do we need to differentiate the computational needs of the algorithm
itself versus the environment it is run in? We do not have the answers, however, we point out that
current standard metrics might not be entirely comprehensive.
In the end, we believe that high compute based methods act as a frontier to showcase bench-
marks [19, 18], i.e., they show what results are possible as data and compute is scaled up (e.g.,
OpenAI Five generates 180 years of gameplay data each day using 128,000 CPU cores and 256
GPUs [18]; AlphaStar uses 200 years of Starcraft II gameplay [19]); however, lighter compute based
algorithmic methods can also yield significant contributions to better tackle real-world problems.
Occam's razor and ablative analysis. Finding the simplest context that exposes the innovative
research idea remains challenging, and if ignored leads to a conflation of fundamental research
(working principles in the most abstract setting) and applied research (working systems as complete
as possible). In particular, some deep learning papers are presented as learning from pixels without
further explanation, while object-level representations would have already exposed the algorithmic
contribution. This still makes sense to remain comparable with established benchmarks (e.g.,
OpenAI Gym [299]), but less so if custom simulations are written without open source access, as it
introduces unnecessary variance in pixel-level representations and artificially inflates computational
resources (see previous point about computational resources).32
In this context there are some
notable exceptions where the algorithmic contribution is presented in a minimal setting and then
results are scaled into complex settings: LOLA [64] first presented a minimalist setting with a two-
player two-action game and then with a more complex variant; similarly, QMIX [181] presented its
results in a two-step (matrix) game and then in the more involved Starcraft II micromanagement
domain [303].
4.5. Open questions
Finally, here we present some open questions for MDRL and point to suggestions on how to
approach them. We believe that there are solid ideas in earlier literature and we refer the reader
to Section 4.1 to avoid deep learning amnesia.
• On the challenge of sparse and delayed rewards.
Recent MDRL competitions and environments have complex scenarios where many actions
are taken before a reward signal is available (see Section 4.3). This sparseness is already a
challenge for RL [20, 338] where approaches such as count-based exploration/intrinsic moti-
vation [196, 339, 340, 341, 342] and hierarchical learning [343, 344, 111] have been proposed
to address it -- in MDRL this is even more problematic since the agents not only need to
learn basic behaviors (like in DRL), but also to learn the strategic element (e.g., competi-
tive/collaborative) embedded in the multiagent setting. To address this issue, recent MDRL
approaches applied dense rewards [206, 205, 204] (a concept originated in RL) at each step
to allow the agents to learn basic motor skills and then decrease these dense rewards over
31NeurIPS 2019 hosts the "MineRL Competition on Sample Efficient Reinforcement Learning using Human Priors"
where the primary goal of the competition is to foster the development of algorithms which can efficiently leverage
human demonstrations to drastically reduce the number of samples needed to solve complex, hierarchical, and sparse
environments [336].
32Cuccu, Togelius and Cudr´e-Mauroux achieved state-of-the-art policy learning in Atari games with only 6 to 18
neurons [337]. The main idea was to decouple image processing from decision-making.
32
time in favor of the environmental reward [158], see Section 3.3. Recent works like OpenAI
Five [18] uses hand-crafted intermediate rewards to accelerate the learning and FTW [179]
lets the agents learn their internal rewards by a hierarchical two-tier optimization. In single
agent domains, RUDDER [345] has been recently proposed for such delayed sparse reward
problems. RUDDER generates a new MDP with more intermediate rewards whose optimal
solution is still an optimal solution to the original MDP. This is achieved by using LSTM
networks to redistribute the original sparse reward to earlier state-action pairs and automat-
ically provide reward shaping. How to best extend RUDDER to multiagent domains is an
open avenue of research.
• On the role of self-play.
Self-play is a cornerstone in MAL with impressive results [147, 127, 145, 346, 143]. While
notable results had also been shown in MDRL [173, 193], recent works have also shown that
plain self-play does not yield the best results. However, adding diversity, i.e., evolutionary
methods [239, 240, 233, 234] or sampling-based methods, have shown good results [158, 179,
159]. A drawback of these solutions is the additional computational requirements since they
need either parallel training (more CPU computation) or memory requirements. Then, it is
still an open problem to improve the computational efficiency of these previously proposed
successful methods, i.e., achieving similar training stability with smaller population sizes that
uses fewer CPU workers in MAL and MDRL (see Section 4.4 and Albrecht et al. [11, Section
5.5]).
• On the challenge of the combinatorial nature of MDRL.
Monte Carlo tree search (MCTS) [347] has been the backbone of the major breakthroughs be-
hind AlphaGo [14] and AlphaGo Zero [15] that combined search and DRL. A recent work [348]
has outlined how search and RL can be better combined for potentially new methods. How-
ever, for multiagent scenarios, there is an additional challenge of the exponential growth of
all the agents' action spaces for centralized methods [349]. One way to tackle this challenge
within multiagent scenarios is the use of search parallelization [350, 351]. Given more scalable
planners, there is room for research in combining these techniques in MDRL settings.
To learn complex multiagent interactions some type of abstraction [352] is often needed, for
example, factored value functions [353, 354, 242, 243, 355, 356] (see QMIX and VDN in
Section 3.5 for recent work in MDRL) try to exploit independence among agents through
(factored) structure; however, in MDRL there are still open questions such as understanding
their representational power [244] (e.g., the accuracy of the learned Q-function approxima-
tions) and how to learn those factorizations, where ideas from transfer planning techniques
could be useful [357, 103]. In transfer planning the idea is to define a simpler "source prob-
lem" (e.g., with fewer agents), in which the agent(s) can plan [357] or learn [103]; since it
is less complex than the real multiagent problem, issues such as the non-stationarity of the
environment can be reduced/removed. Lastly, another related idea are influence abstrac-
tions [358, 359, 10], where instead of learning a complex multiagent model, these methods
try to build smaller models based on the influence agents can exert on one another. While
this has not been sufficiently explored in actual multiagent settings, there is some evidence
that these ideas can lead to effective inductive biases, improving effectiveness of DRL in such
local abstractions [360].
33
5. Conclusions
Deep reinforcement learning has shown recent success on many fronts [13, 14, 16] and a natural
next step is to test multiagent scenarios. However, learning in multiagent environments is fun-
damentally more difficult due to non-stationarity, the increase of dimensionality, and the credit-
assignment problem, among other factors [1, 5, 10, 147, 241, 361, 97].
This survey provides broad overview of recent works in the emerging area of Multiagent Deep
Reinforcement Learning (MDRL). First, we categorized recent works into four different topics:
emergent behaviors, learning communication, learning cooperation, and agents modeling agents.
Then, we exemplified how key components (e.g., experience replay and difference rewards) origi-
nated in RL and MAL need to be adapted to work in MDRL. We provided general lessons learned
applicable to MDRL, pointed to recent multiagent benchmarks and highlighted some open research
problems. Finally, we also reflected on the practical challenges such as computational demands
and reproducibility in MDRL.
Our conclusions of this work are that while the number of works in DRL and MDRL are
notable and represent important milestones for AI, at the same time we acknowledge there are also
open questions in both (deep) single-agent learning [38, 298, 362, 79] and multiagent learning [363,
364, 365, 366, 367, 368]. Our view is that there are practical issues within MDRL that hinder
its scientific progress: the necessity of high compute power, complicated reproducibility (e.g.,
hyperparameter tuning), and the lack of sufficient encouragement for publishing negative results.
However, we remain highly optimistic of the multiagent community and hope this work serves to
raise those issues, encounter good solutions, and ultimately take advantage of the existing literature
and resources available to move the area in the right direction.
Acknowledgements
We would like to thank Chao Gao, Nidhi Hegde, Gregory Palmer, Felipe Leno Da Silva and
Craig Sherstan for reading earlier versions of this work and providing feedback, to April Cooper for
her visual designs for the figures in the article, to Frans Oliehoek, Sam Devlin, Marc Lanctot, Nolan
Bard, Roberta Raileanu, Angeliki Lazaridou, and Yuhang Song for clarifications in their areas of
expertise, to Baoxiang Wang for his suggestions on recent deep RL works, to Michael Kaisers, Daan
Bloembergen, and Katja Hofmann for their comments about the practical challenges of MDRL,
and to the editor and three anonymous reviewers whose comments and suggestions increased the
quality of this work.
References
[1] P. Stone, M. M. Veloso, Multiagent Systems - A Survey from a Machine Learning Perspective., Autonomous
Robots 8 (3) (2000) 345 -- 383.
[2] Y. Shoham, R. Powers, T. Grenager, If multi-agent learning is the answer, what is the question?, Artificial
Intelligence 171 (7) (2007) 365 -- 377.
[3] E. Alonso, M. D'inverno, D. Kudenko, M. Luck, J. Noble, Learning in multi-agent systems, Knowledge Engi-
neering Review 16 (03) (2002) 1 -- 8.
[4] K. Tuyls, G. Weiss, Multiagent learning: Basics, challenges, and prospects, AI Magazine 33 (3) (2012) 41 -- 52.
[5] L. Busoniu, R. Babuska, B. De Schutter, A Comprehensive Survey of Multiagent Reinforcement Learning, IEEE
Transactions on Systems, Man and Cybernetics, Part C (Applications and Reviews) 38 (2) (2008) 156 -- 172.
[6] A. Now´e, P. Vrancx, Y.-M. De Hauwere, Game theory and multi-agent reinforcement learning, in: Reinforce-
ment Learning, Springer, 2012, pp. 441 -- 470.
34
[7] L. Panait, S. Luke, Cooperative Multi-Agent Learning: The State of the Art, Autonomous Agents and Multi-
Agent Systems 11 (3).
[8] L. Matignon, G. J. Laurent, N. Le Fort-Piat, Independent reinforcement learners in cooperative Markov games:
a survey regarding coordination problems, Knowledge Engineering Review 27 (1) (2012) 1 -- 31.
[9] D. Bloembergen, K. Tuyls, D. Hennes, M. Kaisers, Evolutionary Dynamics of Multi-Agent Learning: A Survey.,
Journal of Artificial Intelligence Research 53 (2015) 659 -- 697.
[10] P. Hernandez-Leal, M. Kaisers, T. Baarslag, E. Munoz de Cote, A Survey of Learning in Multiagent Environ-
ments - Dealing with Non-Stationarity.
URL http://arxiv.org/abs/1707.09183
[11] S. V. Albrecht, P. Stone, Autonomous agents modelling other agents: A comprehensive survey and open
problems, Artificial Intelligence 258 (2018) 66 -- 95.
[12] F. L. Silva, A. H. R. Costa, A survey on transfer learning for multiagent reinforcement learning systems,
Journal of Artificial Intelligence Research 64 (2019) 645 -- 703.
[13] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K.
Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra,
S. Legg, D. Hassabis, Human-level control through deep reinforcement learning, Nature 518 (7540) (2015)
529 -- 533.
[14] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou,
V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap,
M. Leach, K. Kavukcuoglu, T. Graepel, D. Hassabis, Mastering the game of Go with deep neural networks
and tree search, Nature 529 (7587) (2016) 484 -- 489.
[15] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai,
A. Bolton, et al., Mastering the game of Go without human knowledge, Nature 550 (7676) (2017) 354.
[16] M. Moravc´ık, M. Schmid, N. Burch, V. Lis´y, D. Morrill, N. Bard, T. Davis, K. Waugh, M. Johanson, M. Bowl-
ing, DeepStack: Expert-level artificial intelligence in heads-up no-limit poker, Science 356 (6337) (2017) 508 --
513.
[17] N. Brown, T. Sandholm, Superhuman AI for heads-up no-limit poker: Libratus beats top professionals, Science
359 (6374) (2018) 418 -- 424.
[18] Open AI Five, https://blog.openai.com/openai-five, [Online; accessed 7-September-2018] (2018).
[19] O. Vinyals, I. Babuschkin, J. Chung, M. Mathieu, M. Jaderberg, W. M. Czarnecki, A. Dudzik, A. Huang,
P. Georgiev, R. Powell, T. Ewalds, D. Horgan, M. Kroiss, I. Danihelka, J. Agapiou, J. Oh, V. Dalibard,
D. Choi, L. Sifre, Y. Sulsky, S. Vezhnevets, J. Molloy, T. Cai, D. Budden, T. Paine, C. Gulcehre, Z. Wang,
T. Pfaff, T. Pohlen, Y. Wu, D. Yogatama, J. Cohen, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, C. Apps,
K. Kavukcuoglu, D. Hassabis, D. Silver, AlphaStar: Mastering the Real-Time Strategy Game StarCraft II,
https://deepmind.com/blog/alphastar-mastering-real-time-strategy-game-starcraft-ii/ (2019).
[20] R. S. Sutton, A. G. Barto, Reinforcement learning: An introduction, 2nd Edition, MIT Press, 2018.
[21] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (7553) (2015) 436.
[22] J. Schmidhuber, Deep learning in neural networks: An overview, Neural networks 61 (2015) 85 -- 117.
[23] K. Arulkumaran, M. P. Deisenroth, M. Brundage, A. A. Bharath, A Brief Survey of Deep Reinforcement
Learning .
URL http://arXiv.org/abs/1708.05866v2
reinforcement learning, Foundations and Trends R(cid:13) in Machine Learning 11 (3-4) (2018) 219 -- 354.
[24] V. Fran¸cois-Lavet, P. Henderson, R. Islam, M. G. Bellemare, J. Pineau, et al., An introduction to deep
[25] Y. Yang, J. Hao, M. Sun, Z. Wang, C. Fan, G. Strbac, Recurrent Deep Multiagent Q-Learning for Autonomous
Brokers in Smart Grid, in: Proceedings of the Twenty-Seventh International Joint Conference on Artificial
Intelligence, Stockholm, Sweden, 2018.
[26] J. Zhao, G. Qiu, Z. Guan, W. Zhao, X. He, Deep reinforcement learning for sponsored search real-time bidding,
in: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,
ACM, 2018, pp. 1021 -- 1030.
[27] B. M. Lake, T. D. Ullman, J. Tenenbaum, S. Gershman, Building machines that learn and think like people,
Behavioral and Brain Sciences 40.
[28] A. Tamar, S. Levine, P. Abbeel, Y. Wu, G. Thomas, Value Iteration Networks., NIPS (2016) 2154 -- 2162.
[29] G. Papoudakis, F. Christianos, A. Rahman, S. V. Albrecht, Dealing with non-stationarity in multi-agent deep
reinforcement learning, arXiv preprint arXiv:1906.04737.
[30] T. T. Nguyen, N. D. Nguyen, S. Nahavandi, Deep reinforcement learning for multi-agent systems: A review of
challenges, solutions and applications, arXiv preprint arXiv:1812.11794.
35
[31] D. H. Wolpert, K. Tumer, Optimal payoff functions for members of collectives, in: Modeling complexity in
economic and social systems, 2002, pp. 355 -- 369.
[32] A. K. Agogino, K. Tumer, Unifying Temporal and Structural Credit Assignment Problems., in: Proceedings
of 17th International Conference on Autonomous Agents and Multiagent Systems, 2004.
[33] N. Fulda, D. Ventura, Predicting and Preventing Coordination Problems in Cooperative Q-learning Systems,
in: Proceedings of the Twentieth International Joint Conference on Artificial Intelligence, Hyderabad, India,
2007, pp. 780 -- 785.
[34] E. Wei, S. Luke, Lenient Learning in Independent-Learner Stochastic Cooperative Games., Journal of Machine
Learning Research.
[35] G. Palmer, K. Tuyls, D. Bloembergen, R. Savani, Lenient Multi-Agent Deep Reinforcement Learning., in:
International Conference on Autonomous Agents and Multiagent Systems, 2018.
[36] L. Busoniu, R. Babuska, B. De Schutter, Multi-agent reinforcement learning: An overview, in: D. Srinivasan,
L. C. Jain (Eds.), Innovations in Multi-Agent Systems and Applications - 1, Springer Berlin Heidelberg, Berlin,
Heidelberg, 2010, pp. 183 -- 221.
[37] Y. Li, Deep reinforcement learning: An overview, CoRR abs/1701.07274. arXiv:1701.07274.
URL http://arxiv.org/abs/1701.07274
[38] A. Darwiche, Human-level intelligence or animal-like abilities?, Commun. ACM 61 (10) (2018) 56 -- 67.
[39] M. Wiering, M. Van Otterlo, Reinforcement learning, Adaptation, learning, and optimization 12.
[40] L. P. Kaelbling, M. L. Littman, A. W. Moore, Reinforcement learning: A survey, Journal of artificial intelligence
research 4 (1996) 237 -- 285.
[41] M. L. Puterman, Markov decision processes: Discrete stochastic dynamic programming, John Wiley & Sons,
Inc., 1994.
[42] A. R. Cassandra, Exact and approximate algorithms for partially observable Markov decision processes, Ph.D.
thesis, Computer Science Department, Brown University (May 1998).
[43] K. J. Astrom, Optimal control of Markov processes with incomplete state information, Journal of mathematical
analysis and applications 10 (1) (1965) 174 -- 205.
[44] R. Bellman, A Markovian decision process, Journal of Mathematics and Mechanics 6 (5) (1957) 679 -- 684.
[45] J. Watkins, Learning from delayed rewards, Ph.D. thesis, King's College, Cambridge, UK (Apr. 1989).
[46] T. Kamihigashi, C. Le Van, Necessary and Sufficient Conditions for a Solution of the Bellman Equation to be
the Value Function: A General Principle.
URL https://halshs.archives-ouvertes.fr/halshs-01159177
[47] J. Tsitsiklis, Asynchronous stochastic approximation and Q-learning, Machine Learning 16 (3) (1994) 185 -- 202.
[48] T. Jaakkola, M. I. Jordan, S. P. Singh, Convergence of stochastic iterative dynamic programming algorithms,
in: Advances in neural information processing systems, 1994, pp. 703 -- 710.
[49] C. Szepesv´ari, M. L. Littman, A unified analysis of value-function-based reinforcement-learning algorithms,
Neural computation 11 (8) (1999) 2017 -- 2060.
[50] E. Even-Dar, Y. Mansour, Learning rates for Q-learning, Journal of Machine Learning Research 5 (Dec) (2003)
1 -- 25.
[51] C. Szepesv´ari, Algorithms for reinforcement learning, Synthesis lectures on artificial intelligence and machine
learning 4 (1) (2010) 1 -- 103.
[52] S. Singh, T. Jaakkola, M. L. Littman, C. Szepesv´ari, Convergence results for single-step on-policy
reinforcement-learning algorithms, Machine learning 38 (3) (2000) 287 -- 308.
[53] H. Van Seijen, H. Van Hasselt, S. Whiteson, M. Wiering, A theoretical and empirical analysis of Expected
Sarsa, in: IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning, Nashville, TN,
USA, 2009, pp. 177 -- 184.
[54] V. R. Konda, J. Tsitsiklis, Actor-critic algorithms, in: Advances in Neural Information Processing Systems,
2000.
[55] R. S. Sutton, D. A. McAllester, S. P. Singh, Y. Mansour, Policy Gradient Methods for Reinforcement Learning
with Function Approximation., in: Advances in Neural Information Processing Systems, 2000.
[56] R. J. Williams, Simple statistical gradient-following algorithms for connectionist reinforcement learning, Ma-
chine learning 8 (3-4) (1992) 229 -- 256.
[57] H. Liu, Y. Feng, Y. Mao, D. Zhou, J. Peng, Q. Liu, Action-depedent control variates for policy optimization
via stein's identity, in: International Conference on Learning Representations, 2018.
[58] S. Gu, T. Lillicrap, Z. Ghahramani, R. E. Turner, S. Levine, Q-prop: Sample-efficient policy gradient with an
off-policy critic, in: International Conference on Learning Representations, 2017.
[59] G. Tucker, S. Bhupatiraju, S. Gu, R. E. Turner, Z. Ghahramani, S. Levine, The mirage of action-dependent
36
baselines in reinforcement learning, in: International Conference on Machine Learning, 2018.
[60] J. Schulman, S. Levine, P. Abbeel, M. I. Jordan, P. Moritz, Trust Region Policy Optimization., in: 31st
International Conference on Machine Learning, Lille, France, 2015.
[61] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, M. Riedmiller, Deterministic policy gradient algorithms,
in: ICML, 2014.
[62] R. Hafner, M. Riedmiller, Reinforcement learning in feedback control, Machine learning 84 (1-2) (2011) 137 --
169.
[63] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, I. Mordatch, Multi-Agent Actor-Critic for Mixed Cooperative-
Competitive Environments., in: Advances in Neural Information Processing Systems, 2017, pp. 6379 -- 6390.
[64] J. N. Foerster, R. Y. Chen, M. Al-Shedivat, S. Whiteson, P. Abbeel, I. Mordatch, Learning with Opponent-
Learning Awareness., in: Proceedings of 17th International Conference on Autonomous Agents and Multiagent
Systems, Stockholm, Sweden, 2018.
[65] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, D. Wierstra, Continuous control
with deep reinforcement learning, in: International Conference on Learning Representations, 2016.
[66] L. D. Pyeatt, A. E. Howe, et al., Decision tree function approximation in reinforcement learning, in: Proceed-
ings of the third international symposium on adaptive systems: evolutionary computation and probabilistic
graphical models, Vol. 2, Cuba, 2001, pp. 70 -- 77.
[67] R. S. Sutton, Generalization in reinforcement learning: Successful examples using sparse coarse coding, in:
Advances in neural information processing systems, 1996, pp. 1038 -- 1044.
[68] R. M. Kretchmar, C. W. Anderson, Comparison of CMACs and radial basis functions for local function
approximators in reinforcement learning, in: Proceedings of International Conference on Neural Networks
(ICNN'97), Vol. 2, IEEE, 1997, pp. 834 -- 837.
[69] J. A. Boyan, A. W. Moore, Generalization in reinforcement learning: Safely approximating the value function,
in: Advances in neural information processing systems, 1995, pp. 369 -- 376.
[70] C. M. Bishop, Pattern recognition and machine learning, Springer, 2006.
[71] M. Riemer, I. Cases, R. Ajemian, M. Liu, I. Rish, Y. Tu, G. Tesauro, Learning to learn without forgetting by
maximizing transfer and minimizing interference, CoRR abs/1810.11910. arXiv:1810.11910.
URL http://arxiv.org/abs/1810.11910
[72] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, M. Riedmiller, Playing Atari with
Deep Reinforcement Learning.
URL http://arxiv.org/abs/1312.5602v1
[73] G. J. Gordon, Approximate solutions to Markov decision processes, Tech. rep., Carnegie-Mellon University
(1999).
[74] L. Baird, Residual algorithms: Reinforcement learning with function approximation, in: Machine Learning
Proceedings 1995, 1995, pp. 30 -- 37.
[75] S. Whiteson, P. Stone, Evolutionary function approximation for reinforcement learning, Journal of Machine
Learning Research 7 (May) (2006) 877 -- 917.
[76] J. Achiam, E. Knight, P. Abbeel, Towards Characterizing Divergence in Deep Q-Learning, CoRR
abs/1903.08894. arXiv:1903.08894.
URL http://arxiv.org/abs/1903.08894
[77] H. van Hasselt, Y. Doron, F. Strub, M. Hessel, N. Sonnerat, J. Modayil, Deep reinforcement learning and the
deadly triad, CoRR abs/1812.02648. arXiv:1812.02648.
URL http://arxiv.org/abs/1812.02648
[78] S. Fujimoto, H. van Hoof, D. Meger, Addressing function approximation error in actor-critic methods, in:
International Conference on Machine Learning, 2018.
[79] A. Ilyas, L. Engstrom, S. Santurkar, D. Tsipras, F. Janoos, L. Rudolph, A. Madry, Are deep policy gradient
algorithms truly policy gradient algorithms?, CoRR abs/1811.02553. arXiv:1811.02553.
URL http://arxiv.org/abs/1811.02553
[80] T. Lu, D. Schuurmans, C. Boutilier, Non-delusional Q-learning and value-iteration, in: Advances in Neural
Information Processing Systems, 2018, pp. 9949 -- 9959.
[81] J. N. Tsitsiklis, B. Van Roy, Analysis of temporal-diffference learning with function approximation, in: Ad-
vances in neural information processing systems, 1997, pp. 1075 -- 1081.
[82] F. S. Melo, S. P. Meyn, M. I. Ribeiro, An analysis of reinforcement learning with function approximation, in:
Proceedings of the 25th international conference on Machine learning, ACM, 2008, pp. 664 -- 671.
[83] D. Ernst, P. Geurts, L. Wehenkel, Tree-based batch mode reinforcement learning, Journal of Machine Learning
Research 6 (Apr) (2005) 503 -- 556.
37
[84] M. Jaderberg, V. Mnih, W. M. Czarnecki, T. Schaul, J. Z. Leibo, D. Silver, K. Kavukcuoglu, Reinforcement
Learning with Unsupervised Auxiliary Tasks., in: International Conference on Learning Representations, 2017.
[85] M. Hausknecht, P. Stone, Deep Recurrent Q-Learning for Partially Observable MDPs, in: International Con-
ference on Learning Representations, 2015.
[86] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (8) (1997) 1735 -- 1780.
[87] M. Riedmiller, Neural fitted Q iteration -- first experiences with a data efficient neural reinforcement learning
method, in: European Conference on Machine Learning, Springer, 2005, pp. 317 -- 328.
[88] R. H. Crites, A. G. Barto, Elevator group control using multiple reinforcement learning agents, Machine
learning 33 (2-3) (1998) 235 -- 262.
[89] L. J. Lin, Programming robots using reinforcement learning and teaching., in: AAAI, 1991, pp. 781 -- 786.
[90] L.-J. Lin, Self-improving reactive agents based on reinforcement learning, planning and teaching, Machine
learning 8 (3-4) (1992) 293 -- 321.
[91] H. V. Hasselt, Double Q-learning, in: Advances in Neural Information Processing Systems, 2010, pp. 2613 --
2621.
[92] H. Van Hasselt, A. Guez, D. Silver, Deep reinforcement learning with double Q-learning, in: Thirtieth AAAI
Conference on Artificial Intelligence, 2016.
[93] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, K. Kavukcuoglu, Asynchronous
methods for deep reinforcement learning, in: International Conference on Machine Learning, 2016, pp. 1928 --
1937.
[94] M. McCloskey, N. J. Cohen, Catastrophic interference in connectionist networks: The sequential learning
problem, in: Psychology of learning and motivation, Vol. 24, Elsevier, 1989, pp. 109 -- 165.
[95] I. J. Goodfellow, M. Mirza, D. Xiao, A. Courville, Y. Bengio, An empirical investigation of catastrophic
forgetting in gradient-based neural networks.
URL https://arxiv.org/abs/1312.6211
[96] D. Isele, A. Cosgun, Selective experience replay for lifelong learning, in: Thirty-Second AAAI Conference on
Artificial Intelligence, 2018.
[97] G. Palmer, R. Savani, K. Tuyls, Negative update intervals in deep multi-agent reinforcement learning, in: 18th
International Conference on Autonomous Agents and Multiagent Systems, 2019.
[98] Z. Wang, T. Schaul, M. Hessel, H. Van Hasselt, M. Lanctot, N. De Freitas, Dueling network architectures for
deep reinforcement learning, in: International Conference on Machine Learning, 2016.
[99] M. Hauskrecht, Value-function approximations for partially observable Markov decision processes, Journal of
Artificial Intelligence Research 13 (1).
[100] N. Meuleau, L. Peshkin, K.-E. Kim, L. P. Kaelbling, Learning finite-state controllers for partially observable
environments, in: Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence, 1999, pp.
427 -- 436.
[101] D. Steckelmacher, D. M. Roijers, A. Harutyunyan, P. Vrancx, H. Plisnier, A. Now´e, Reinforcement learning in
pomdps with memoryless options and option-observation initiation sets, in: Thirty-Second AAAI Conference
on Artificial Intelligence, 2018.
[102] S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate
shift, Proceedings of the 32nd International Conference on Machine Learning (2015) 448 -- 456.
[103] E. Van der Pol, F. A. Oliehoek, Coordinated deep reinforcement learners for traffic light control, in: Proceedings
of Learning, Inference and Control of Multi-Agent Systems at NIPS, 2016.
[104] T. Salimans, D. P. Kingma, Weight normalization: A simple reparameterization to accelerate training of deep
neural networks, in: Advances in Neural Information Processing Systems, 2016, pp. 901 -- 909.
[105] OpenAI Baselines: ACKTR & A2C, https://openai.com/blog/baselines-acktr-a2c/, [Online; accessed
29-April-2019] (2017).
[106] Z. Wang, V. Bapst, N. Heess, V. Mnih, R. Munos, K. Kavukcuoglu, N. de Freitas, Sample efficient actor-critic
with experience replay, arXiv preprint arXiv:1611.01224.
[107] S. S. Gu, T. Lillicrap, R. E. Turner, Z. Ghahramani, B. Scholkopf, S. Levine, Interpolated policy gradient:
Merging on-policy and off-policy gradient estimation for deep reinforcement learning, in: Advances in Neural
Information Processing Systems, 2017, pp. 3846 -- 3855.
[108] E. Shelhamer, P. Mahmoudieh, M. Argus, T. Darrell, Loss is its own reward: Self-supervision for reinforcement
learning, ICLR workshops.
[109] M. G. Bellemare, W. Dabney, R. Dadashi, A. A. Taıga, P. S. Castro, N. L. Roux, D. Schuurmans, T. Lat-
timore, C. Lyle, A Geometric Perspective on Optimal Representations for Reinforcement Learning, CoRR
abs/1901.11530. arXiv:1901.11530.
38
URL http://arxiv.org/abs/1901.11530
[110] R. S. Sutton, J. Modayil, M. Delp, T. Degris, P. M. Pilarski, A. White, D. Precup, Horde: A scalable
real-time architecture for learning knowledge from unsupervised sensorimotor interaction, in: The 10th Inter-
national Conference on Autonomous Agents and Multiagent Systems-Volume 2, International Foundation for
Autonomous Agents and Multiagent Systems, 2011, pp. 761 -- 768.
[111] T. Schaul, J. Quan, I. Antonoglou, D. Silver, Prioritized Experience Replay, in: International Conference on
Learning Representations, 2016.
[112] A. W. Moore, C. G. Atkeson, Prioritized sweeping: Reinforcement learning with less data and less time,
Machine learning 13 (1) (1993) 103 -- 130.
[113] D. Andre, N. Friedman, R. Parr, Generalized prioritized sweeping, in: Advances in Neural Information Pro-
cessing Systems, 1998, pp. 1001 -- 1007.
[114] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu, T. Harley, I. Dun-
ning, et al., IMPALA: Scalable distributed Deep-RL with importance weighted actor-learner architectures, in:
International Conference on Machine Learning, 2018.
[115] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal Policy Optimization Algorithms.
URL http://arxiv.org/abs/1707.06347
[116] S. M. Kakade, A natural policy gradient, in: Advances in neural information processing systems, 2002, pp.
1531 -- 1538.
[117] N. Heess, D. TB, S. Sriram, J. Lemmon, J. Merel, G. Wayne, Y. Tassa, T. Erez, Z. Wang, S. M. A. Eslami,
M. A. Riedmiller, D. Silver, Emergence of Locomotion Behaviours in Rich Environments.
URL http://arXiv.org/abs/1707.02286v2
[118] G. Bacchiani, D. Molinari, M. Patander, Microscopic traffic simulation by cooperative multi-agent deep rein-
forcement learning, in: AAMAS, 2019.
[119] J. Schulman, P. Abbeel, X. Chen, Equivalence Between Policy Gradients and Soft Q-Learning, CoRR
abs/1704.06440. arXiv:1704.06440.
URL http://arxiv.org/abs/1704.06440
[120] T. Haarnoja, H. Tang, P. Abbeel, S. Levine, Reinforcement learning with deep energy-based policies, in:
Proceedings of the 34th International Conference on Machine Learning-Volume 70, 2017, pp. 1352 -- 1361.
[121] T. Haarnoja, A. Zhou, P. Abbeel, S. Levine, Soft actor-critic: Off-policy maximum entropy deep reinforcement
learning with a stochastic actor, in: International Conference on Machine Learning, 2018.
[122] M. Tan, Multi-Agent Reinforcement Learning: Independent vs. Cooperative Agents, in: Machine Learning
Proceedings 1993 Proceedings of the Tenth International Conference, University of Massachusetts, Amherst,
June 27 -- 29, 1993, 1993, pp. 330 -- 337.
[123] G. J. Laurent, L. Matignon, L. Fort-Piat, et al., The world of independent learners is not Markovian, Interna-
tional Journal of Knowledge-based and Intelligent Engineering Systems 15 (1) (2011) 55 -- 64.
[124] M. L. Littman, Markov games as a framework for multi-agent reinforcement learning, in: Proceedings of the
11th International Conference on Machine Learning, New Brunswick, NJ, USA, 1994, pp. 157 -- 163.
[125] M. L. Littman, Value-function reinforcement learning in Markov games, Cognitive Systems Research 2 (1)
(2001) 55 -- 66.
[126] C. Claus, C. Boutilier, The dynamics of reinforcement learning in cooperative multiagent systems, in: Proceed-
ings of the 15th National Conference on Artificial Intelligence, Madison, Wisconsin, USA, 1998, pp. 746 -- 752.
[127] J. Hu, M. P. Wellman, Nash Q-learning for general-sum stochastic games, The Journal of Machine Learning
Research 4 (2003) 1039 -- 1069.
[128] M. L. Littman, Friend-or-foe Q-learning in general-sum games, in: Proceedings of 17th International Conference
on Autonomous Agents and Multiagent Systems, Williamstown, MA, USA, 2001, pp. 322 -- 328.
[129] X. Song, T. Wang, C. Zhang, Convergence of multi-agent learning with a finite step size in general-sum games,
in: 18th International Conference on Autonomous Agents and Multiagent Systems, 2019.
[130] D. Balduzzi, S. Racaniere, J. Martens, J. Foerster, K. Tuyls, T. Graepel, The mechanics of n-player differ-
entiable games, in: Proceedings of the 35th International Conference on Machine Learning, Proceedings of
Machine Learning Research, Stockholm, Sweden, 2018, pp. 354 -- 363.
[131] J. P´erolat, B. Piot, O. Pietquin, Actor-critic fictitious play in simultaneous move multistage games, in: 21st
International Conference on Artificial Intelligence and Statistics, 2018.
[132] G. Bono, J. S. Dibangoye, L. Matignon, F. Pereyron, O. Simonin, Cooperative multi-agent policy gradient, in:
European Conference on Machine Learning, 2018.
[133] T. W. Neller, M. Lanctot, An introduction to counterfactual regret minimization, in: Proceedings of Model AI
Assignments, The Fourth Symposium on Educational Advances in Artificial Intelligence (EAAI-2013), 2013.
39
[134] M. Zinkevich, M. Johanson, M. Bowling, C. Piccione, Regret minimization in games with incomplete informa-
tion, in: Advances in neural information processing systems, 2008, pp. 1729 -- 1736.
[135] A. Blum, Y. Monsour, Learning, regret minimization, and equilibria, in: Algorithmic Game Theory, Cambridge
University Press, 2007, Ch. 4.
[136] S. Srinivasan, M. Lanctot, V. Zambaldi, J. P´erolat, K. Tuyls, R. Munos, M. Bowling, Actor-critic policy
optimization in partially observable multiagent environments, in: Advances in Neural Information Processing
Systems, 2018, pp. 3422 -- 3435.
[137] E. Lockhart, M. Lanctot, J. P´erolat, J. Lespiau, D. Morrill, F. Timbers, K. Tuyls, Computing approximate
equilibria in sequential adversarial games by exploitability descent, CoRR abs/1903.05614. arXiv:1903.05614.
URL http://arxiv.org/abs/1903.05614
[138] M. Johanson, N. Bard, N. Burch, M. Bowling, Finding optimal abstract strategies in extensive-form games,
in: Twenty-Sixth AAAI Conference on Artificial Intelligence, 2012.
[139] E. Kalai, E. Lehrer, Rational learning leads to Nash equilibrium, Econometrica: Journal of the Econometric
Society (1993) 1019 -- 1045.
[140] T. W. Sandholm, R. H. Crites, Multiagent reinforcement learning in the iterated prisoner's dilemma, Biosystems
37 (1-2) (1996) 147 -- 166.
[141] S. Singh, M. Kearns, Y. Mansour, Nash convergence of gradient dynamics in general-sum games, in: Proceed-
ings of the Sixteenth conference on Uncertainty in artificial intelligence, Morgan Kaufmann Publishers Inc.,
2000, pp. 541 -- 548.
[142] B. Banerjee, J. Peng, Adaptive policy gradient in multiagent learning, in: Proceedings of the second interna-
tional joint conference on Autonomous agents and multiagent systems, ACM, 2003, pp. 686 -- 692.
[143] A. Greenwald, K. Hall, Correlated Q-learning, in: Proceedings of 17th International Conference on Autonomous
Agents and Multiagent Systems, Washington, DC, USA, 2003, pp. 242 -- 249.
[144] M. Bowling, Convergence problems of general-sum multiagent reinforcement learning, in: International Con-
ference on Machine Learning, 2000, pp. 89 -- 94.
[145] M. Bowling, Convergence and no-regret in multiagent learning, in: Advances in Neural Information Processing
Systems, Vancouver, Canada, 2004, pp. 209 -- 216.
[146] M. Zinkevich, A. Greenwald, M. L. Littman, Cyclic equilibria in Markov games, in: Advances in Neural
Information Processing Systems, 2006, pp. 1641 -- 1648.
[147] M. Bowling, M. Veloso, Multiagent learning using a variable learning rate, Artificial Intelligence 136 (2) (2002)
215 -- 250.
[148] M. Kaisers, K. Tuyls, FAQ-learning in matrix games: demonstrating convergence near Nash equilibria, and
bifurcation of attractors in the battle of sexes, in: AAAI Workshop on Interactive Decision Theory and Game
Theory, San Francisco, CA, USA, 2011, pp. 309 -- 316.
[149] M. Wunder, M. L. Littman, M. Babes, Classes of Multiagent Q-learning Dynamics with epsilon-greedy Ex-
ploration, in: Proceedings of the 35th International Conference on Machine Learning, Haifa, Israel, 2010, pp.
1167 -- 1174.
[150] G. Tesauro, Extending Q-learning to general adaptive multi-agent systems, in: Advances in Neural Information
Processing Systems, Vancouver, Canada, 2003, pp. 871 -- 878.
[151] M. Weinberg, J. S. Rosenschein, Best-response multiagent learning in non-stationary environments, in: Pro-
ceedings of the 3rd International Conference on Autonomous Agents and Multiagent Systems, New York, NY,
USA, 2004, pp. 506 -- 513.
[152] D. Chakraborty, P. Stone, Multiagent learning in the presence of memory-bounded agents, Autonomous Agents
and Multi-Agent Systems 28 (2) (2013) 182 -- 213.
[153] G. Weiss (Ed.), Multiagent Systems, 2nd Edition, (Intelligent Robotics and Autonomous Agents series), MIT
Press, Cambridge, MA, USA, 2013.
[154] Multiagent Learning, Foundations and Recent Trends, https://www.cs.utexas.edu/~larg/ijcai17_
tutorial/multiagent_learning.pdf, [Online; accessed 7-September-2018] (2017).
[155] A. Tampuu, T. Matiisen, D. Kodelja, I. Kuzovkin, K. Korjus, J. Aru, J. Aru, R. Vicente, Multiagent cooper-
ation and competition with deep reinforcement learning, PLOS ONE 12 (4) (2017) e0172395.
[156] J. Z. Leibo, V. Zambaldi, M. Lanctot, J. Marecki, Multi-agent Reinforcement Learning in Sequential Social
Dilemmas, in: Proceedings of the 16th Conference on Autonomous Agents and Multiagent Systems, Sao Paulo,
2017.
[157] M. Raghu, A. Irpan, J. Andreas, R. Kleinberg, Q. Le, J. Kleinberg, Can Deep Reinforcement Learning solve
Erdos-Selfridge-Spencer Games?, in: Proceedings of the 35th International Conference on Machine Learning,
2018.
40
[158] T. Bansal, J. Pachocki, S. Sidor, I. Sutskever, I. Mordatch, Emergent Complexity via Multi-Agent Competition.,
in: International Conference on Machine Learning, 2018.
[159] J. Z. Leibo, J. Perolat, E. Hughes, S. Wheelwright, A. H. Marblestone, E. Du´enez-Guzm´an, P. Sunehag,
I. Dunning, T. Graepel, Malthusian reinforcement learning, in: 18th International Conference on Autonomous
Agents and Multiagent Systems, 2019.
[160] I. Mordatch, P. Abbeel, Emergence of grounded compositional language in multi-agent populations, in: Thirty-
Second AAAI Conference on Artificial Intelligence, 2018.
[161] A. Lazaridou, A. Peysakhovich, M. Baroni, Multi-Agent Cooperation and the Emergence of (Natural) Lan-
guage, in: International Conference on Learning Representations, 2017.
[162] J. N. Foerster, Y. M. Assael, N. De Freitas, S. Whiteson, Learning to communicate with deep multi-agent
reinforcement learning, in: Advances in Neural Information Processing Systems, 2016, pp. 2145 -- 2153.
[163] S. Sukhbaatar, A. Szlam, R. Fergus, Learning Multiagent Communication with Backpropagation, in: Advances
in Neural Information Processing Systems, 2016, pp. 2244 -- 2252.
[164] P. Peng, Q. Yuan, Y. Wen, Y. Yang, Z. Tang, H. Long, J. Wang, Multiagent Bidirectionally-Coordinated Nets
for Learning to Play StarCraft Combat Games.
URL http://arxiv.org/abs/1703.10069
[165] E. Pesce, G. Montana, Improving coordination in multi-agent deep reinforcement learning through memory-
driven communication, CoRR abs/1901.03887. arXiv:1901.03887.
URL http://arxiv.org/abs/1901.03887
[166] S. Omidshafiei, J. Pazis, C. Amato, J. P. How, J. Vian, Deep Decentralized Multi-task Multi-Agent Reinforce-
ment Learning under Partial Observability, in: Proceedings of the 34th International Conference on Machine
Learning, Sydney, 2017.
[167] J. N. Foerster, G. Farquhar, T. Afouras, N. Nardelli, S. Whiteson, Counterfactual Multi-Agent Policy Gradi-
ents., in: 32nd AAAI Conference on Artificial Intelligence, 2017.
[168] J. N. Foerster, N. Nardelli, G. Farquhar, T. Afouras, P. H. S. Torr, P. Kohli, S. Whiteson, Stabilising Experience
Replay for Deep Multi-Agent Reinforcement Learning., in: International Conference on Machine Learning,
2017.
[169] H. He, J. Boyd-Graber, K. Kwok, H. Daume, Opponent modeling in deep reinforcement learning, in: 33rd
International Conference on Machine Learning, 2016, pp. 2675 -- 2684.
[170] R. Raileanu, E. Denton, A. Szlam, R. Fergus, Modeling Others using Oneself in Multi-Agent Reinforcement
Learning., in: International Conference on Machine Learning, 2018.
[171] Z.-W. Hong, S.-Y. Su, T.-Y. Shann, Y.-H. Chang, C.-Y. Lee, A Deep Policy Inference Q-Network for Multi-
Agent Systems, in: International Conference on Autonomous Agents and Multiagent Systems, 2018.
[172] M. Lanctot, V. F. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. P´erolat, D. Silver, T. Graepel, A Uni-
fied Game-Theoretic Approach to Multiagent Reinforcement Learning., in: Advances in Neural Information
Processing Systems, 2017.
[173] J. Heinrich, D. Silver, Deep Reinforcement Learning from Self-Play in Imperfect-Information Games.
URL http://arxiv.org/abs/1603.01121
[174] N. C. Rabinowitz, F. Perbet, H. F. Song, C. Zhang, S. M. A. Eslami, M. Botvinick, Machine Theory of Mind.,
in: International Conference on Machine Learning, Stockholm, Sweden, 2018.
[175] T. Yang, J. Hao, Z. Meng, C. Zhang, Y. Z. Z. Zheng, Towards Efficient Detection and Optimal Response
against Sophisticated Opponents, in: IJCAI, 2019.
[176] A. Lerer, A. Peysakhovich, Maintaining cooperation in complex social dilemmas using deep reinforcement
learning, CoRR abs/1707.01068. arXiv:1707.01068.
URL http://arxiv.org/abs/1707.01068
[177] W. Kim, M. Cho, Y. Sung, Message-Dropout: An Efficient Training Method for Multi-Agent Deep Reinforce-
ment Learning, in: 33rd AAAI Conference on Artificial Intelligence, 2019.
[178] Y. Zheng, J. Hao, Z. Zhang, Weighted double deep multiagent reinforcement learning in stochastic cooperative
environments.
URL http://arXiv.org/abs/1802.08534
[179] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie, N. C. Rabi-
nowitz, A. S. Morcos, A. Ruderman, N. Sonnerat, T. Green, L. Deason, J. Z. Leibo, D. Silver, D. Hassabis,
K. Kavukcuoglu, T. Graepel, Human-level performance in 3d multiplayer games with population-based rein-
forcement learning, Science 364 (6443) (2019) 859 -- 865. arXiv:https://science.sciencemag.org/content/
364/6443/859.full.pdf, doi:10.1126/science.aau6249.
URL https://science.sciencemag.org/content/364/6443/859
41
[180] P. Sunehag, G. Lever, A. Gruslys, W. M. Czarnecki, V. F. Zambaldi, M. Jaderberg, M. Lanctot, N. Sonnerat,
J. Z. Leibo, K. Tuyls, T. Graepel, Value-Decomposition Networks For Cooperative Multi-Agent Learning Based
On Team Reward., in: Proceedings of 17th International Conference on Autonomous Agents and Multiagent
Systems, Stockholm, Sweden, 2018.
[181] T. Rashid, M. Samvelyan, C. S. de Witt, G. Farquhar, J. N. Foerster, S. Whiteson, QMIX - Monotonic
Value Function Factorisation for Deep Multi-Agent Reinforcement Learning., in: International Conference on
Machine Learning, 2018.
[182] J. K. Gupta, M. Egorov, M. Kochenderfer, Cooperative multi-agent control using deep reinforcement learning,
in: G. Sukthankar, J. A. Rodriguez-Aguilar (Eds.), Autonomous Agents and Multiagent Systems, Springer
International Publishing, Cham, 2017, pp. 66 -- 83.
[183] S. Li, Y. Wu, X. Cui, H. Dong, F. Fang, S. Russell, Robust multi-agent reinforcement learning via minimax
deep deterministic policy gradient, in: AAAI Conference on Artificial Intelligence, 2019.
[184] Y. Zheng, Z. Meng, J. Hao, Z. Zhang, T. Yang, C. Fan, A Deep Bayesian Policy Reuse Approach Against
Non-Stationary Agents, in: Advances in Neural Information Processing Systems, 2018, pp. 962 -- 972.
[185] R. Powers, Y. Shoham, Learning against opponents with bounded memory, in: Proceedings of the 19th Inter-
national Joint Conference on Artificial Intelligence, Edinburg, Scotland, UK, 2005, pp. 817 -- 822.
[186] M. L. Littman, P. Stone, Implicit Negotiation in Repeated Games, ATAL '01: Revised Papers from the 8th
International Workshop on Intelligent Agents VIII.
[187] R. Axelrod, W. D. Hamilton, The evolution of cooperation, Science 211 (27) (1981) 1390 -- 1396.
[188] E. Munoz de Cote, A. Lazaric, M. Restelli, Learning to cooperate in multi-agent social dilemmas, in: Proceed-
ings of the 5th International Conference on Autonomous Agents and Multiagent Systems, Hakodate, Hokkaido,
Japan, 2006, pp. 783 -- 785.
[189] J. L. Stimpson, M. A. Goodrich, Learning to cooperate in a social dilemma: A satisficing approach to bar-
gaining, in: Proceedings of the 20th International Conference on Machine Learning (ICML-03), 2003, pp.
728 -- 735.
[190] G. W. Brown, Iterative solution of games by fictitious play, Activity analysis of production and allocation
13 (1) (1951) 374 -- 376.
[191] D. Monderer, L. S. Shapley, Fictitious play property for games with identical interests, Journal of economic
theory 68 (1) (1996) 258 -- 265.
[192] G. Tesauro, Temporal difference learning and TD-Gammon, Communications of the ACM 38 (3) (1995) 58 -- 68.
[193] M. Bowling, N. Burch, M. Johanson, O. Tammelin, Heads-up limit hold'em poker is solved, Science 347 (6218)
(2015) 145 -- 149.
[194] J. Z. Leibo, E. Hughes, M. Lanctot, T. Graepel, Autocurricula and the emergence of innovation from social
interaction: A manifesto for multi-agent intelligence research, CoRR abs/1903.00742. arXiv:1903.00742.
URL http://arxiv.org/abs/1903.00742
[195] S. Samothrakis, S. Lucas, T. Runarsson, D. Robles, Coevolving game-playing agents: Measuring performance
and intransitivities, IEEE Transactions on Evolutionary Computation 17 (2) (2013) 213 -- 226.
[196] M. Bellemare, S. Srinivasan, G. Ostrovski, T. Schaul, D. Saxton, R. Munos, Unifying count-based exploration
and intrinsic motivation, in: Advances in Neural Information Processing Systems, 2016, pp. 1471 -- 1479.
[197] D. E. Moriarty, A. C. Schultz, J. J. Grefenstette, Evolutionary algorithms for reinforcement learning, Journal
of Artificial Intelligence Research 11 (1999) 241 -- 276.
[198] F. A. Oliehoek, E. D. De Jong, N. Vlassis, The parallel Nash memory for asymmetric games, in: Proceedings
of the 8th annual conference on Genetic and evolutionary computation, ACM, 2006, pp. 337 -- 344.
[199] L. Bull, T. C. Fogarty, M. Snaith, Evolution in multi-agent systems: Evolving communicating classifier systems
for gait in a quadrupedal robot, in: Proceedings of the 6th International Conference on Genetic Algorithms,
Morgan Kaufmann Publishers Inc., 1995, pp. 382 -- 388.
[200] L. Bull, Evolutionary computing in multi-agent environments: Operators, in: International Conference on
Evolutionary Programming, Springer, 1998, pp. 43 -- 52.
[201] H. Iba, Emergent cooperation for multiple agents using genetic programming, in: International Conference on
Parallel Problem Solving from Nature, Springer, 1996, pp. 32 -- 41.
[202] E. Todorov, T. Erez, Y. Tassa, MuJoCo - A physics engine for model-based control, Intelligent Robots and
Systems (2012) 5026 -- 5033.
[203] V. Gullapalli, A. G. Barto, Shaping as a method for accelerating reinforcement learning, in: Proceedings of
the 1992 IEEE international symposium on intelligent control, IEEE, 1992, pp. 554 -- 559.
[204] S. Mahadevan, J. Connell, Automatic programming of behavior-based robots using reinforcement learning,
Artificial intelligence 55 (2-3) (1992) 311 -- 365.
42
[205] G. Konidaris, A. Barto, Autonomous shaping: Knowledge transfer in reinforcement learning, in: Proceedings
of the 23rd international conference on Machine learning, ACM, 2006, pp. 489 -- 496.
[206] A. Y. Ng, D. Harada, S. J. Russell, Policy invariance under reward transformations: Theory and application
to reward shaping, in: Proceedings of the Sixteenth International Conference on Machine Learning, 1999, pp.
278 -- 287.
[207] P. Erdos, J. L. Selfridge, On a combinatorial game, Journal of Combinatorial Theory, Series A 14 (3) (1973)
298 -- 301.
[208] J. Spencer, Randomization, derandomization and antirandomization: three games, Theoretical Computer
Science 131 (2) (1994) 415 -- 429.
[209] D. Fudenberg, J. Tirole, Game Theory, The MIT Press, 1991.
[210] L. v. d. Maaten, G. Hinton, Visualizing data using t-SNE, Journal of machine learning research 9 (Nov) (2008)
2579 -- 2605.
[211] T. Zahavy, N. Ben-Zrihem, S. Mannor, Graying the black box: Understanding DQNs, in: International Con-
ference on Machine Learning, 2016, pp. 1899 -- 1908.
[212] E. Beeching, C. Wolf, J. Dibangoye, O. Simonin, Deep Reinforcement Learning on a Budget: 3D Control and
Reasoning Without a Supercomputer, CoRR abs/1904.01806. arXiv:1904.01806.
URL http://arxiv.org/abs/1904.01806
[213] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, R. Salakhutdinov, Dropout: a simple way to prevent
neural networks from overfitting, The Journal of Machine Learning Research 15 (1) (2014) 1929 -- 1958.
[214] M. Schuster, K. K. Paliwal, Bidirectional recurrent neural networks, IEEE Transactions on Signal Processing
45 (11) (1997) 2673 -- 2681.
[215] R. Lowe, J. Foerster, Y.-L. Boureau, J. Pineau, Y. Dauphin, On the pitfalls of measuring emergent communi-
cation, in: 18th International Conference on Autonomous Agents and Multiagent Systems, 2019.
[216] M. Tambe, Towards flexible teamwork, Journal of artificial intelligence research 7 (1997) 83 -- 124.
[217] B. J. Grosz, S. Kraus, Collaborative plans for complex group action, Artificial Intelligence 86 (2) (1996)
269 -- 357.
[218] D. Precup, R. S. Sutton, S. Singh, Eligibility traces for off-policy policy evaluation, in: Proceedings of the
Seventeenth International Conference on Machine Learning., 2000.
[219] J. Frank, S. Mannor, D. Precup, Reinforcement learning in the presence of rare events, in: Proceedings of the
25th international conference on Machine learning, ACM, 2008, pp. 336 -- 343.
[220] T. I. Ahamed, V. S. Borkar, S. Juneja, Adaptive importance sampling technique for markov chains using
stochastic approximation, Operations Research 54 (3) (2006) 489 -- 504.
[221] K. A. Ciosek, S. Whiteson, Offer: Off-environment reinforcement learning, in: Thirty-First AAAI Conference
on Artificial Intelligence, 2017.
[222] D. Bloembergen, M. Kaisers, K. Tuyls, Lenient frequency adjusted Q-learning, in: Proceedings of the 22nd
Belgian/Netherlands Artificial Intelligence Conference, 2010.
[223] L. Panait, K. Sullivan, S. Luke, Lenience towards teammates helps in cooperative multiagent learning, in:
Proceedings of the 5th International Conference on Autonomous Agents and Multiagent Systems, Hakodate,
Japan, 2006.
[224] L. Panait, K. Tuyls, S. Luke, Theoretical advantages of lenient learners: An evolutionary game theoretic
perspective, JMLR 9 (Mar) (2008) 423 -- 457.
[225] M. Lauer, M. Riedmiller, An algorithm for distributed reinforcement learning in cooperative multi-agent sys-
tems, in: In Proceedings of the Seventeenth International Conference on Machine Learning, 2000.
[226] R. Caruana, Multitask learning, Machine learning 28 (1) (1997) 41 -- 75.
[227] A. A. Rusu, S. G. Colmenarejo, C. Gulcehre, G. Desjardins, J. Kirkpatrick, R. Pascanu, V. Mnih,
K. Kavukcuoglu, R. Hadsell, Policy Distillation, in: International Conference on Learning Representations,
2016.
[228] G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, in: NIPS Deep Learning Work-
shop, 2014.
[229] A. S. Vezhnevets, S. Osindero, T. Schaul, N. Heess, M. Jaderberg, D. Silver, K. Kavukcuoglu, FeUdal Networks
for Hierarchical Reinforcement Learning., International Conference On Machine Learning.
[230] P. Dayan, G. E. Hinton, Feudal reinforcement learning, in: Advances in neural information processing systems,
1993, pp. 271 -- 278.
[231] S. P. Singh, Transfer of learning by composing solutions of elemental sequential tasks, Machine Learning 8 (3-4)
(1992) 323 -- 339.
[232] Capture
the Flag:
the
emergence of
complex cooperative agents,
https://deepmind.com/blog/
43
capture-the-flag/, [Online; accessed 7-September-2018] (2018).
[233] C. D. Rosin, R. K. Belew, New methods for competitive coevolution, Evolutionary computation 5 (1) (1997)
1 -- 29.
[234] J. Lehman, K. O. Stanley, Exploiting open-endedness to solve problems through the search for novelty., in:
ALIFE, 2008, pp. 329 -- 336.
[235] M. Jaderberg, V. Dalibard, S. Osindero, W. M. Czarnecki, J. Donahue, A. Razavi, O. Vinyals, T. Green,
I. Dunning, K. Simonyan, et al., Population based training of neural networks.
URL http://arxiv.org/abs/1711.09846
[236] A. E. Elo, The rating of chessplayers, past and present, Arco Pub., 1978.
[237] R. Herbrich, T. Minka, T. Graepel, TrueSkill: a Bayesian skill rating system, in: Advances in neural information
processing systems, 2007, pp. 569 -- 576.
[238] S. Omidshafiei, C. Papadimitriou, G. Piliouras, K. Tuyls, M. Rowland, J.-B. Lespiau, W. M. Czarnecki,
M. Lanctot, J. Perolat, R. Munos, α-Rank: Multi-Agent Evaluation by Evolution, Scientific Reports 9.
[239] T. Back, Evolutionary algorithms in theory and practice: evolution strategies, evolutionary programming,
genetic algorithms, Oxford university press, 1996.
[240] K. A. De Jong, Evolutionary computation: a unified approach, MIT press, 2006.
[241] K. Tumer, A. Agogino, Distributed agent-based air traffic flow management, in: Proceedings of the 6th Inter-
national Conference on Autonomous Agents and Multiagent Systems, Honolulu, Hawaii, 2007.
[242] C. Guestrin, D. Koller, R. Parr, Multiagent planning with factored MDPs, in: Advances in neural information
processing systems, 2002, pp. 1523 -- 1530.
[243] J. R. Kok, N. Vlassis, Sparse cooperative Q-learning, in: Proceedings of the twenty-first international conference
on Machine learning, ACM, 2004, p. 61.
[244] J. Castellini, F. A. Oliehoek, R. Savani, S. Whiteson, The Representational Capacity of Action-Value Net-
works for Multi-Agent Reinforcement Learning, in: 18th International Conference on Autonomous Agents and
Multiagent Systems, 2019.
[245] P. J. Gmytrasiewicz, P. Doshi, A framework for sequential planning in multiagent settings, Journal of Artificial
Intelligence Research 24 (1) (2005) 49 -- 79.
[246] J. C. Harsanyi, Games with incomplete information played by Bayesian players, I -- III Part I. The basic model,
Management science 14 (3) (1967) 159 -- 182.
[247] S. Barrett, P. Stone, S. Kraus, A. Rosenfeld, Teamwork with Limited Knowledge of Teammates., in: Pro-
ceedings of the Twenty-Seventh AAAI Conference on Artificial Intelligence, Bellevue, WS, USA, 2013, pp.
102 -- 108.
[248] R. A. Jacobs, M. I. Jordan, S. J. Nowlan, G. E. Hinton, et al., Adaptive mixtures of local experts., Neural
computation 3 (1) (1991) 79 -- 87.
[249] J. Heinrich, M. Lanctot, D. Silver, Fictitious self-play in extensive-form games, in: International Conference
on Machine Learning, 2015, pp. 805 -- 813.
[250] J. F. Nash, Equilibrium points in n-person games, Proceedings of the National Academy of Sciences 36 (1)
(1950) 48 -- 49.
[251] J. Von Neumann, O. Morgenstern, Theory of games and economic behavior, Vol. 51, Bull. Amer. Math. Soc,
1945.
[252] J. S. Shamma, G. Arslan, Dynamic fictitious play, dynamic gradient play, and distributed convergence to Nash
equilibria, IEEE Transactions on Automatic Control 50 (3) (2005) 312 -- 327.
[253] W. E. Walsh, R. Das, G. Tesauro, J. O. Kephart, Analyzing complex strategic interactions in multi-agent
systems, AAAI-02 Workshop on Game-Theoretic and Decision-Theoretic Agents (2002) 109 -- 118.
[254] M. Johanson, K. Waugh, M. Bowling, M. Zinkevich, Accelerating best response calculation in large extensive
games, in: Twenty-Second International Joint Conference on Artificial Intelligence, 2011.
[255] C. F. Camerer, T.-H. Ho, J.-K. Chong, A cognitive hierarchy model of games, The Quarterly Journal of
Economics 119 (3) (2004) 861.
[256] M. Costa Gomes, V. P. Crawford, B. Broseta, Cognition and Behavior in Normal -- Form Games: An Experi-
mental Study, Econometrica 69 (5) (2001) 1193 -- 1235.
[257] J. Morimoto, K. Doya, Robust reinforcement learning, Neural computation 17 (2) (2005) 335 -- 359.
[258] L. Pinto, J. Davidson, R. Sukthankar, A. Gupta, Robust adversarial reinforcement learning, in: Proceedings
of the 34th International Conference on Machine Learning-Volume 70, JMLR. org, 2017, pp. 2817 -- 2826.
[259] R. Powers, Y. Shoham, T. Vu, A general criterion and an algorithmic framework for learning in multi-agent
systems, Machine Learning 67 (1-2) (2007) 45 -- 76.
[260] J. W. Crandall, M. A. Goodrich, Learning to compete, coordinate, and cooperate in repeated games using
44
reinforcement learning, Machine Learning 82 (3) (2011) 281 -- 314.
[261] M. Johanson, M. A. Zinkevich, M. Bowling, Computing Robust Counter-Strategies., in: Advances in Neural
Information Processing Systems, Vancouver, BC, Canada, 2007, pp. 721 -- 728.
[262] P. McCracken, M. Bowling, Safe strategies for agent modelling in games, in: AAAI Fall Symposium, 2004, pp.
103 -- 110.
[263] S. Damer, M. Gini, Safely using predictions in general-sum normal form games, in: Proceedings of the 16th
Conference on Autonomous Agents and Multiagent Systems, Sao Paulo, 2017.
[264] C. Zhang, V. Lesser, Multi-agent learning with policy prediction, in: Twenty-Fourth AAAI Conference on
Artificial Intelligence, 2010.
[265] P. J. Gmytrasiewicz, E. H. Durfee, Rational Coordination in Multi-Agent Environments, Autonomous Agents
and Multi-Agent Systems 3 (4) (2000) 319 -- 350.
[266] C. F. Camerer, T.-H. Ho, J.-K. Chong, Behavioural Game Theory: Thinking, Learning and Teaching, in:
Advances in Understanding Strategic Behavior, New York, 2004, pp. 120 -- 180.
[267] D. Carmel, S. Markovitch, Incorporating opponent models into adversary search, in: AAAI/IAAI, Vol. 1, 1996,
pp. 120 -- 125.
[268] H. de Weerd, R. Verbrugge, B. Verheij, How much does it help to know what she knows you know? An
agent-based simulation study, Artificial Intelligence 199-200 (C) (2013) 67 -- 92.
[269] P. Hernandez-Leal, M. Kaisers, Towards a Fast Detection of Opponents in Repeated Stochastic Games, in:
G. Sukthankar, J. A. Rodriguez-Aguilar (Eds.), Autonomous Agents and Multiagent Systems: AAMAS 2017
Workshops, Best Papers, Sao Paulo, Brazil, May 8-12, 2017, Revised Selected Papers, 2017, pp. 239 -- 257.
[270] P. Hernandez-Leal, M. E. Taylor, B. Rosman, L. E. Sucar, E. Munoz de Cote, Identifying and Tracking Switch-
ing, Non-stationary Opponents: a Bayesian Approach, in: Multiagent Interaction without Prior Coordination
Workshop at AAAI, Phoenix, AZ, USA, 2016.
[271] B. Rosman, M. Hawasly, S. Ramamoorthy, Bayesian Policy Reuse, Machine Learning 104 (1) (2016) 99 -- 127.
[272] P. Hernandez-Leal, Y. Zhan, M. E. Taylor, L. E. Sucar, E. Munoz de Cote, Efficiently detecting switches
against non-stationary opponents, Autonomous Agents and Multi-Agent Systems 31 (4) (2017) 767 -- 789.
[273] P. Hernandez-Leal, M. Kaisers, Learning against sequential opponents in repeated stochastic games, in: The
3rd Multi-disciplinary Conference on Reinforcement Learning and Decision Making, Ann Arbor, 2017.
[274] J. Schmidhuber, Critique of Paper by "Deep Learning Conspiracy" (Nature 521 p 436), http://people.idsia.
ch/~juergen/deep-learning-conspiracy.html (2015).
paper?,
arXiv
really
cite
an
have
to
[275] Do
I
http://approximatelycorrect.com/2017/08/01/
do-i-have-to-cite-arxiv-paper/, [Online; accessed 21-May-2019] (2017).
[276] Collaboration & Credit Principles, How can we be good stewards of collaborative trust?, http://colah.
github.io/posts/2019-05-Collaboration/index.html, [Online; accessed 31-May-2019] (2019).
[277] H. Wang, B. Raj, E. P. Xing, On the origin of deep learning, CoRR abs/1702.07800. arXiv:1702.07800.
URL http://arxiv.org/abs/1702.07800
[278] A. K. Agogino, K. Tumer, Analyzing and visualizing multiagent rewards in dynamic and stochastic domains,
Autonomous Agents and Multi-Agent Systems 17 (2) (2008) 320 -- 338.
[279] S. Devlin, L. M. Yliniemi, D. Kudenko, K. Tumer, Potential-based difference rewards for multiagent reinforce-
ment learning., in: 13th International Conference on Autonomous Agents and Multiagent Systems, AAMAS
2014, Paris, France, 2014.
[280] D. H. Wolpert, K. R. Wheeler, K. Tumer, General principles of learning-based multi-agent systems, in: Pro-
ceedings of the Third International Conference on Autonomous Agents, 1999.
[281] M. E. Taylor, P. Stone, Transfer learning for reinforcement learning domains: A survey, The Journal of Machine
Learning Research 10 (2009) 1633 -- 1685.
[282] C. Bucilua, R. Caruana, A. Niculescu-Mizil, Model compression, in: Proceedings of the 12th ACM SIGKDD
international conference on Knowledge discovery and data mining, ACM, 2006, pp. 535 -- 541.
[283] Y. Du, W. M. Czarnecki, S. M. Jayakumar, R. Pascanu, B. Lakshminarayanan, Adapting auxiliary losses using
gradient similarity, arXiv preprint arXiv:1812.02224.
[284] S. C. Suddarth, Y. Kergosien, Rule-injection hints as a means of improving network performance and learning
time, in: Neural Networks, Springer, 1990, pp. 120 -- 129.
[285] P. Hernandez-Leal, B. Kartal, M. E. Taylor, Agent Modeling as Auxiliary Task for Deep Reinforcement Learn-
ing, in: AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 2019.
[286] M. Andrychowicz, F. Wolski, A. Ray, J. Schneider, R. Fong, P. Welinder, B. McGrew, J. Tobin, P. Abbeel,
W. Zaremba, Hindsight experience replay, in: Advances in Neural Information Processing Systems, 2017.
[287] Z. C. Lipton, K. Azizzadenesheli, A. Kumar, L. Li, J. Gao, L. Deng, Combating Reinforcement Learning's
45
Sisyphean Curse with Intrinsic Fear.
URL http://arxiv.org/abs/1611.01211v8
[288] T. De Bruin, J. Kober, K. Tuyls, R. Babuska, Experience selection in deep reinforcement learning for control,
The Journal of Machine Learning Research 19 (1) (2018) 347 -- 402.
[289] D. S. Bernstein, R. Givan, N. Immerman, S. Zilberstein, The complexity of decentralized control of Markov
decision processes, Mathematics of operations research 27 (4) (2002) 819 -- 840.
[290] F. A. Oliehoek, C. Amato, et al., A concise introduction to decentralized POMDPs, Springer, 2016.
[291] F. A. Oliehoek, M. T. Spaan, N. Vlassis, Optimal and approximate Q-value functions for decentralized
POMDPs, Journal of Artificial Intelligence Research 32 (2008) 289 -- 353.
[292] J. K. Gupta, M. Egorov, M. J. Kochenderfer, Cooperative Multi-agent Control using deep reinforcement
learning, in: Adaptive Learning Agents at AAMAS, Sao Paulo, 2017.
[293] R. Pascanu, T. Mikolov, Y. Bengio, On the difficulty of training recurrent neural networks, in: International
conference on machine learning, 2013, pp. 1310 -- 1318.
[294] K. Greff, R. K. Srivastava, J. Koutnik, B. R. Steunebrink, J. Schmidhuber, LSTM: A Search Space Odyssey,
IEEE Transactions on Neural Networks and Learning Systems 28 (10) (2017) 2222 -- 2232.
[295] J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical evaluation of gated recurrent neural networks on sequence
modeling, in: Deep Learning and Representation Learning Workshop, 2014.
[296] S. Whiteson, B. Tanner, M. E. Taylor, P. Stone, Protecting against evaluation overfitting in empirical reinforce-
ment learning, in: 2011 IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning
(ADPRL), IEEE, 2011, pp. 120 -- 127.
[297] M. G. Bellemare, Y. Naddaf, J. Veness, M. Bowling, The arcade learning environment: An evaluation platform
for general agents, Journal of Artificial Intelligence Research 47 (2013) 253 -- 279.
[298] M. C. Machado, M. G. Bellemare, E. Talvitie, J. Veness, M. Hausknecht, M. Bowling, Revisiting the ar-
cade learning environment: Evaluation protocols and open problems for general agents, Journal of Artificial
Intelligence Research 61 (2018) 523 -- 562.
[299] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, W. Zaremba, OpenAI Gym,
arXiv preprint arXiv:1606.01540.
[300] P. S. Castro, S. Moitra, C. Gelada, S. Kumar, M. G. Bellemare, Dopamine: A Research Framework for Deep
Reinforcement Learning.
URL http://arxiv.org/abs/1812.06110
[301] C. Resnick, W. Eldridge, D. Ha, D. Britz, J. Foerster, J. Togelius, K. Cho, J. Bruna, Pommerman: A Multi-
Agent Playground.
URL http://arxiv.org/abs/1809.07124
[302] C. Gao, B. Kartal, P. Hernandez-Leal, M. E. Taylor, On Hard Exploration for Reinforcement Learning: a Case
Study in Pommerman, in: AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment,
2019.
[303] M. Samvelyan, T. Rashid, C. S. de Witt, G. Farquhar, N. Nardelli, T. G. J. Rudner, C. Hung, P. H. S. Torr,
J. N. Foerster, S. Whiteson, The StarCraft Multi-Agent Challenge, CoRR abs/1902.04043. arXiv:1902.04043.
URL http://arxiv.org/abs/1902.04043
[304] D. P´erez-Li´ebana, K. Hofmann, S. P. Mohanty, N. Kuno, A. Kramer, S. Devlin, R. D. Gaina, D. Ionita,
The multi-agent reinforcement learning in Malmo (MARL O) competition, CoRR abs/1901.08129. arXiv:
1901.08129.
URL http://arxiv.org/abs/1901.08129
[305] M. Johnson, K. Hofmann, T. Hutton, D. Bignell, The Malmo platform for artificial intelligence experimenta-
tion., in: IJCAI, 2016, pp. 4246 -- 4247.
[306] P. Stone, G. Kaminka, S. Kraus, J. S. Rosenschein, Ad Hoc Autonomous Agent Teams: Collaboration without
Pre-Coordination., in: 32nd AAAI Conference on Artificial Intelligence, Atlanta, Georgia, USA, 2010, pp.
1504 -- 1509.
[307] M. Bowling, P. McCracken, Coordination and adaptation in impromptu teams, in: Proceedings of the Nine-
teenth Conference on Artificial Intelligence, Vol. 5, 2005, pp. 53 -- 58.
[308] S. V. Albrecht, S. Ramamoorthy, A game-theoretic model and best-response learning method for ad hoc
coordination in multiagent systems, in: Proceedings of the 12th International Conference on Autonomous
Agents and Multi-agent Systems, Saint Paul, MN, USA, 2013.
[309] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin, S. Moitra,
E. Hughes, I. Dunning, S. Mourad, H. Larochelle, M. G. Bellemare, M. Bowling, The Hanabi Challenge: A
New Frontier for AI Research.
46
URL https://arxiv.org/abs/1902.00506
[310] M. Hessel, J. Modayil, H. Van Hasselt, T. Schaul, G. Ostrovski, W. Dabney, D. Horgan, B. Piot, M. Azar,
D. Silver, Rainbow: Combining improvements in deep reinforcement learning, in: Thirty-Second AAAI Con-
ference on Artificial Intelligence, 2018.
[311] Y. Song, J. Wang, T. Lukasiewicz, Z. Xu, M. Xu, Z. Ding, L. Wu, Arena: A general evaluation platform and
building toolkit for multi-agent intelligence, CoRR abs/1905.08085. arXiv:1905.08085.
URL http://arxiv.org/abs/1905.08085
[312] A. Juliani, V. Berges, E. Vckay, Y. Gao, H. Henry, M. Mattar, D. Lange, Unity: A general platform for
intelligent agents, CoRR abs/1809.02627. arXiv:1809.02627.
URL http://arxiv.org/abs/1809.02627
[313] S. Liu, G. Lever, J. Merel, S. Tunyasuvunakool, N. Heess, T. Graepel, Emergent coordination through compe-
tition, in: International Conference on Learning Representations, 2019.
[314] J. Suarez, Y. Du, P. Isola, I. Mordatch, Neural MMO: A massively multiagent game environment for training
and evaluating intelligent agents, CoRR abs/1903.00784. arXiv:1903.00784.
URL http://arxiv.org/abs/1903.00784
[315] P. Henderson, R. Islam, P. Bachman, J. Pineau, D. Precup, D. Meger, Deep Reinforcement Learning That
Matters., in: 32nd AAAI Conference on Artificial Intelligence, 2018.
[316] P. Nagarajan, G. Warnell, P. Stone, Deterministic implementations for reproducibility in deep reinforcement
learning.
URL http://arxiv.org/abs/1809.05676
[317] K. Clary, E. Tosch, J. Foley, D. Jensen, Let's play again: Variability of deep reinforcement learning agents in
Atari environments, in: NeurIPS Critiquing and Correcting Trends Workshop, 2018.
[318] J. Z. Forde, M. Paganini, The scientific method in the science of machine learning, in: ICLR Debugging
Machine Learning Models workshop, 2019.
[319] K. Azizzadenesheli, Maybe a few considerations in reinforcement learning research?, in: Reinforcement Learn-
ing for Real Life Workshop, 2019.
[320] Z. C. Lipton, J. Steinhardt, Troubling trends in machine learning scholarship, in: ICML Machine Learning
Debates workshop, 2018.
[321] R. Rosenthal, The file drawer problem and tolerance for null results., Psychological bulletin 86 (3) (1979) 638.
[322] D. Sculley, J. Snoek, A. Wiltschko, A. Rahimi, Winner's curse? on pace, progress, and empirical rigor, in:
ICLR Workshop, 2018.
[323] O. Gencoglu, M. van Gils, E. Guldogan, C. Morikawa, M. Suzen, M. Gruber, J. Leinonen, H. Hut-
tunen, Hark side of deep learning -- from grad student descent to automated machine learning, arXiv preprint
arXiv:1904.07633.
[324] K. Azizzadenesheli, B. Yang, W. Liu, E. Brunskill, Z. Lipton, A. Anandkumar, Surprising negative results for
generative adversarial tree search, in: Critiquing and Correcting Trends in Machine Learning Workshop, 2018.
[325] C. Lyle, P. S. Castro, M. G. Bellemare, A comparative analysis of expected and distributional reinforcement
learning, in: Thirty-Third AAAI Conference on Artificial Intelligence, 2019.
[326] B. Kartal, P. Hernandez-Leal, M. E. Taylor, Using Monte Carlo tree search as a demonstrator within asyn-
chronous deep RL, in: AAAI Workshop on Reinforcement Learning in Games, 2019.
[327] G. Melis, C. Dyer, P. Blunsom, On the state of the art of evaluation in neural language models, in: International
Conference on Learning Representations, 2018.
[328] Deep Reinforcement Learning: Pong from Pixels, https://karpathy.github.io/2016/05/31/rl/, [Online;
accessed 7-May-2019] (2016).
[329] V. Firoiu, W. F. Whitney, J. B. Tenenbaum, Beating the World's Best at Super Smash Bros. with Deep
Reinforcement Learning, CoRR abs/1702.06230. arXiv:1702.06230.
URL http://arxiv.org/abs/1702.06230
[330] C. Gao, P. Hernandez-Leal, B. Kartal, M. E. Taylor, Skynet: A Top Deep RL Agent in the Inaugural Pom-
merman Team Competition, in: 4th Multidisciplinary Conference on Reinforcement Learning and Decision
Making, 2019.
[331] D. Amodei, D. Hernandez, AI and Compute (2018).
URL https://blog.openai.com/ai-and-compute
[332] Y. Yu, Towards sample efficient reinforcement learning., in: IJCAI, 2018, pp. 5739 -- 5743.
[333] F. P. Such, V. Madhavan, E. Conti, J. Lehman, K. O. Stanley, J. Clune, Deep neuroevolution: Genetic
algorithms are a competitive alternative for training deep neural networks for reinforcement learning, CoRR
abs/1712.06567.
47
URL http://arxiv.org/abs/1712.06567
[334] A. Stooke, P. Abbeel, Accelerated methods for deep reinforcement learning, CoRR abs/1803.02811. arXiv:
1803.02811.
URL http://arxiv.org/abs/1803.02811
[335] M. Babaeizadeh, I. Frosio, S. Tyree, J. Clemons, J. Kautz, Reinforcement learning through asynchronous
advantage actor-critic on a GPU, in: International Conference on Learning Representations, 2017.
[336] W. H. Guss, C. Codel, K. Hofmann, B. Houghton, N. Kuno, S. Milani, S. P. Mohanty, D. P. Liebana,
R. Salakhutdinov, N. Topin, M. Veloso, P. Wang, The MineRL Competition on Sample Efficient Reinforcement
Learning using Human Priors, CoRR abs/1904.10079. arXiv:1904.10079.
URL http://arxiv.org/abs/1904.10079
[337] G. Cuccu, J. Togelius, P. Cudr´e-Mauroux, Playing Atari with six neurons, in: Proceedings of the 18th Interna-
tional Conference on Autonomous Agents and MultiAgent Systems, International Foundation for Autonomous
Agents and Multiagent Systems, 2019, pp. 998 -- 1006.
[338] A. Ecoffet, J. Huizinga, J. Lehman, K. O. Stanley, J. Clune, Go-explore: a new approach for hard-exploration
problems, arXiv preprint arXiv:1901.10995.
[339] R. I. Brafman, M. Tennenholtz, R-max-a general polynomial time algorithm for near-optimal reinforcement
learning, Journal of Machine Learning Research 3 (Oct) (2002) 213 -- 231.
[340] A. L. Strehl, M. L. Littman, An analysis of model-based interval estimation for Markov decision processes,
Journal of Computer and System Sciences 74 (8) (2008) 1309 -- 1331.
[341] J. Schmidhuber, A possibility for implementing curiosity and boredom in model-building neural controllers,
in: Proc. of the international conference on simulation of adaptive behavior: From animals to animats, 1991,
pp. 222 -- 227.
[342] A. G. Barto, Intrinsic motivation and reinforcement learning, in: Intrinsically motivated learning in natural
and artificial systems, Springer, 2013, pp. 17 -- 47.
[343] T. D. Kulkarni, K. Narasimhan, A. Saeedi, J. Tenenbaum, Hierarchical deep reinforcement learning: Integrating
temporal abstraction and intrinsic motivation, in: Advances in neural information processing systems, 2016,
pp. 3675 -- 3683.
[344] T. G. Dietterich, Ensemble Methods in Machine Learning, in: MCS Proceedings of the First International
Workshop on Multiple Classifier Systems, Springer Berlin Heidelberg, Cagliari, Italy, 2000, pp. 1 -- 15.
[345] J. A. Arjona-Medina, M. Gillhofer, M. Widrich, T. Unterthiner, S. Hochreiter, RUDDER: Return Decomposi-
tion for Delayed Rewards.
URL http://arxiv.org/abs/1806.07857
[346] V. Conitzer, T. Sandholm, AWESOME: A general multiagent learning algorithm that converges in self-play
and learns a best response against stationary opponents, Machine Learning 67 (1-2) (2006) 23 -- 43.
[347] C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P. I. Cowling, P. Rohlfshagen, S. Tavener, D. Perez,
S. Samothrakis, S. Colton, A survey of Monte Carlo tree search methods, IEEE Transactions on Computational
Intelligence and AI in games 4 (1) (2012) 1 -- 43.
[348] T. Vodopivec, S. Samothrakis, B. Ster, On Monte Carlo tree search and reinforcement learning, Journal of
Artificial Intelligence Research 60 (2017) 881 -- 936.
[349] B. Kartal, J. Godoy, I. Karamouzas, S. J. Guy, Stochastic tree search with useful cycles for patrolling problems,
in: Robotics and Automation (ICRA), 2015 IEEE International Conference on, IEEE, 2015, pp. 1289 -- 1294.
[350] B. Kartal, E. Nunes, J. Godoy, M. Gini, Monte Carlo tree search with branch and bound for multi-robot task
allocation, in: The IJCAI-16 Workshop on Autonomous Mobile Service Robots, 2016.
[351] G. Best, O. M. Cliff, T. Patten, R. R. Mettu, R. Fitch, Dec-MCTS: Decentralized planning for multi-robot
active perception, The International Journal of Robotics Research 38 (2-3) (2019) 316 -- 337.
[352] Y.-M. De Hauwere, P. Vrancx, A. Nowe, Learning multi-agent state space representations, in: Proceedings of
the 9th International Conference on Autonomous Agents and Multiagent Systems, Toronto, Canada, 2010, pp.
715 -- 722.
[353] C. Guestrin, M. Lagoudakis, R. Parr, Coordinated reinforcement learning, in: ICML, Vol. 2, 2002, pp. 227 -- 234.
[354] C. Guestrin, D. Koller, R. Parr, S. Venkataraman, Efficient solution algorithms for factored MDPs, Journal of
Artificial Intelligence Research 19 (2003) 399 -- 468.
[355] C. Amato, F. A. Oliehoek, Scalable Planning and Learning for Multiagent POMDPs, in: AAAI, 2015, pp.
1995 -- 2002.
[356] F. A. Oliehoek, Interactive Learning and Decision Making - Foundations, Insights & Challenges., International
Joint Conference on Artificial Intelligence.
[357] F. A. Oliehoek, S. Whiteson, M. T. Spaan, Approximate solutions for factored Dec-POMDPs with many
48
agents, in: Proceedings of the 2013 international conference on Autonomous agents and multi-agent systems,
International Foundation for Autonomous Agents and Multiagent Systems, 2013, pp. 563 -- 570.
[358] R. Becker, S. Zilberstein, V. Lesser, C. V. Goldman, Solving transition independent decentralized Markov
decision processes, Journal of Artificial Intelligence Research 22 (2004) 423 -- 455.
[359] F. A. Oliehoek, S. J. Witwicki, L. P. Kaelbling, Influence-based abstraction for multiagent systems, in: Twenty-
Sixth AAAI Conference on Artificial Intelligence, 2012.
[360] M. Suau de Castro, E. Congeduti, R. A. Starre, A. Czechowski, F. A. Oliehoek, Influence-based abstraction in
deep reinforcement learning, in: Adaptive, Learning Agents workshop, 2019.
[361] E. Wei, D. Wicke, D. Freelan, S. Luke, Multiagent Soft Q-Learning.
URL http://arXiv.org/abs/1804.09817
[362] R. R. Torrado, P. Bontrager, J. Togelius, J. Liu, D. Perez-Liebana, Deep Reinforcement Learning for General
Video Game AI.
URL http://arxiv.org/abs/1806.02448
[363] P. A. Ortega, S. Legg, Modeling friends and foes.
URL http://arxiv.org/abs/1807.00196
[364] Y. Yang, R. Luo, M. Li, M. Zhou, W. Zhang, J. Wang, Mean field multi-agent reinforcement learning, in:
Proceedings of the 35th International Conference on Machine Learning, Stockholm Sweden, 2018.
[365] A. Grover, M. Al-Shedivat, J. K. Gupta, Y. Burda, H. Edwards, Learning Policy Representations in Multiagent
Systems., in: International Conference on Machine Learning, 2018.
[366] C. K. Ling, F. Fang, J. Z. Kolter, What game are we playing? end-to-end learning in normal and extensive
form games, in: Twenty-Seventh International Joint Conference on Artificial Intelligence, 2018.
[367] S. Omidshafiei, D. Hennes, D. Morrill, R. Munos, J. Perolat, M. Lanctot, A. Gruslys, J.-B. Lespiau, K. Tuyls,
Neural Replicator Dynamics, arXiv e-prints (2019) arXiv:1906.00190arXiv:1906.00190.
[368] S. Khadka, S. Majumdar, K. Tumer, Evolutionary Reinforcement Learning for Sample-Efficient Multiagent
Coordination, arXiv e-prints (2019) arXiv:1906.07315arXiv:1906.07315.
49
|
1902.01642 | 1 | 1902 | 2019-02-05T11:15:55 | Agent-Based Simulation Modelling for Reflecting on Consequences of Digital Mental Health | [
"cs.MA",
"cs.CY"
] | The premise of this working paper is based around agent-based simulation models and how to go about creating them from given incomplete information. Agent-based simulations are stochastic simulations that revolve around groups of agents that each have their own characteristics and can make decisions. Such simulations can be used to emulate real life situations and to create hypothetical situations without the need for real-world testing prior. Here we describe the development of an agent-based simulation model for studying future digital mental health scenarios. An incomplete conceptual model has been used as the basis for this development. To define differences in responses to stimuli we employed fuzzy decision making logic. The model has been implemented but not been used for structured experimentation yet. This is planned as our next step. | cs.MA | cs | Agent-Based Simulation Modelling for Reflecting on
Consequences of Digital Mental Health
A working paper based on the 2018 Summer Internship Report by Daniel Stroud, who was
supervised by Peer-Olaf Siebers and Christian Wagner
Finalised by Peer-Olaf Siebers (University of Nottingham, School of Computer Science, Nottingham,
NG8 1BB, UK, [email protected])
Published on 05/02/2019
Abstract
The premise of this working paper is based around agent-based simulation models and how to go
about creating them from given incomplete information. Agent-based simulations are stochastic
simulations that revolve around groups of agents that each have their own characteristics and can
make decisions. Such simulations can be used to emulate real life situations and to create hypothetical
situations without the need for real-world testing prior. Here we describe the development of an
agent-based simulation model for studying future digital mental health scenarios. An incomplete
conceptual model has been used as the basis for this development. To define differences in responses
to stimuli we employed fuzzy decision making logic. The model has been implemented but not been
used for structured experimentation yet. This is planned as our next step.
Keywords
Digital Mental Health; Agent Based Modelling; Simulation; Fuzzy Logic Decision Making
1 Introduction
Agent-Based Simulation (ABS) is a powerful paradigm that can be used for conducting what-if analysis
of human centric systems (Siebers and Aickelin 2008). Agents are comparable to non-player characters
in games. In principle, ABS enables exploring the interaction of different groups of stakeholders, where
the latter are often people in a specific context, from land owners in conservation, to shoppers in
marketing.
A key challenge in ABS is the appropriate specification of agents in order for them to behave similarly
to humans. Commonly, so-called archetypes are established, which capture groups within a
population sharing similar behaviour (Zhang et al 2012). This project is designed to combine fuzzy set
theory with ABS in order to enable simulations where agents can be efficiently designed to replicate
core behavioural aspects of human stakeholders, thus providing a pathway for human-inspired ABS.
My internship was based on creating a simulation from an incomplete conceptual model that was
previously created by my supervisor Peer-Olaf Siebers and his colleagues Elvira Perez Vallejos and
Tommy Nilsson (Siebers et al unpublished) using the Engineering Agent Based Social Simulation
framework (or EABSS for short). The EABSS supports a structured way of co-creation by embedding
design philosophies and tools from Software Engineering and using these for driving its focus group
communications. The process is guided by a set of predefined table templates in combination with the
use of UML diagrams - a graphical notation used in Software Engineering to conduct system analysis
and design (Fowler 2004) - for capturing the thoughts of academic and non-academic participants.
1
I started by learning to use AnyLogic (XJTEK 2018) and by reading up on the subject of technology with
mental health. After this, I created a very simple simulation of a hospital scenario including doctors,
patient and robot doctors (the technology). This simulation gave me something that I could add to and
build upon using the conceptual model given to me by Peer. I continued adding to the conceptual
model and the simulation model until I had a detailed enough simulation that I could start catering it
towards the hypotheses defined in the conceptual model I was given. As well as this, I added a small
bit of fuzzy logic to the simulation to familiarize myself with the area before the second half of my
internship. The fuzzy logic was used as the decision making for doctors and robot doctors to decide
how long they should treat patients for, and was used for visitors to decide whether they should visit
or not. After catering towards the hypotheses I had a fully working simulation based off the conceptual
model given to me.
2 Proof-of-Principle Agent-Based Simulation Model
For the first half I created an ABS based on a conceptual model given to me by Peer Olaf Siebers that
was created in a focus group (Siebers et al, unpublished). In Agent-Based Modelling (ABM), a system
is modelled as a collection of autonomous decision-making entities called agents. Each agent
individually assesses its situation and makes decisions on the basis of a set of rules (Bonabeau 2002).
He wanted me to create a simple model that would visualize and give valuable outputs to the problems
posed in the conceptual model. Some scans of the original conceptual model are shown in Figure 1.
The conceptual model had the aim "reflect on consequences of digital mental health solutions" and
used the EABSS framework approach (Siebers and Klügl 2017) for driving the model development
process. The information within it allowed me to develop a simulation model and run some
experiments. My simulation was a hospital scenario with doctors, patients, visitors and robot doctors
(modelled as proactive machines). The idea of the hospital scenario was to have doctors and proactive
machines treating patients with occasional visits from visitors and then to track the mental state of
the patients. Depending on the different ratios of human doctors to proactive machines and how
much the patients communicated, their mental states would vary.
2
EABSS driven system analysis
EABSS driven initial model scope table
3
EABSS model scope table after iteration
Figure 1: Scans of parts of the original conceptual model that I worked off
The initial steps of the internship were to create a very simple hospital scenario that wasn't very
specific to certain hypotheses, this would allow me to add more detail later depending on which
hypotheses from the conceptual model I chose to focus on. My initial model included doctors,
proactive machines and patients. The doctors and proactive machines would both treat the patients
and there was no mental state for the patients at this point. They treated the patients based off a
linked list containing patients that needed a check-up or requested to be seen. For the implementation
I used the free version of AnyLogic PLE (XJTEK 2018). Figure 2 shows how the simulation looked like
during runtime at this point (stage 1).
Figure 2: Example screenshot, showing what the simulation looked like at stage 1
4
The next step after this was adding visitors and differences between the doctors and proactive
machines. I initially added visitors that would be created at a certain time every day and then after an
hour they would all be removed from the simulation. This method caused some problems due to the
50,000 agent limit on the free version of AnyLogic, so I had to create a certain number of visitors at
the start of the simulation and then have them visit and leave the hospital every day for the duration
of the simulation. In order to create differences between the doctors and proactive machines, firstly I
made them do slightly different jobs and secondly, I added a mental state (satisfaction) variable to the
patients so that they are more satisfied if they have been treated by a doctor than a proactive
machine. The colour behind each patient represents their mental state. Figure 3 shows how the
simulation looked like during runtime at this point (stage 2).
Figure 3: Example screenshot, showing what the simulation looked like at stage 2
3 Proof-of-Principle Fuzzy Logic Extension
At this point I had a simulation that was fully functional and allowed me to track the patients' mental
states over time, however I had not focussed on certain hypotheses from the original conceptual
model.
Before fully focussing on specific hypotheses I decided to create and add some fuzzy logic systems into
the simulation to give me a better idea of the area for the second half of the internship and because
some areas of the simulation required decision making that could be made using fuzzy logic systems.
Fuzzy logic is an approach to computing based on "degrees of truth" rather than the usual "true or
false" (1 or 0) Boolean logic on which the modern computer is based. The idea of fuzzy logic was first
advanced by Dr Lotfi Zadeh of the University of California at Berkeley in the 1960s (Rouse n.d.). I added
three FLSs overall and their purposes were to decide the amount of time doctors and proactive
machines should treat patients for and also to decide whether visitors should visit patients and if so,
how long for. I created these FLSs initially using JuzzyOnline (Wagner and Pierfitte n.d.) and then used
those to help write and implement the Java into the AnyLogic simulation model I created earlier.
Following are the fuzzy logic graphs showing the inputs, outputs and membership functions of the FLS'
for a doctor (Figure 4), a robot (Figure 5), and a visitor (Figure 6).
5
Figure 4: Inputs, outputs and membership functions of the FLS' for a doctor
Figure 5: Inputs, outputs and membership functions of the FLS' for a robot
6
Figure 6: Inputs, outputs and membership functions of the FLS' for a visitor
4 Applying EABSS for Model Re-Development and Documentation
After adding fuzzy logic into the simulation I decided to focus on 2 of the hypotheses posed by the
original conceptual model. The two hypotheses I focused on were 'Trust between machines and
humans will affect human-human relationships in a negative way' and 'Machine-look influences our
decisions'.
To aim the simulation towards the first hypothesis I had to add trust variables to each patient. These
variables changed over time depending on the treatment of robots or doctors on that patient. As well
as this I created a network between the patients so that a patient is connected to the patients in the
beds next to them and the beds next to those beds. If patients are connected in a network, then their
current trust can affect other patient's trust as if they are talking to each other and sharing opinions.
Then to be able to view the relationships/trust between patients they are connected with lines that
change colour depending on the patients' difference in trust values. If there is a large difference, then
the line is shown in red but if they have similar trust values then they are like each other and trust
each other so the line is shown in green. Anywhere in between is shown in yellow.
To tackle the second hypothesis, I created two different groups of robots, human-like robots and
robot-like robots. If a human-like robot treats a patient, then they are more satisfied and trust the
robot more than if is robot-like. From this I can track the trust and opinions of robots and see it spread
throughout the patients using the network explained before. As well as this I added two groups of
7
doctors, a senior group and a junior group. The senior doctors treat patients quicker than the junior
doctors meaning that they can treat more patients overall.
In the following I present the conceptual model for final version of hospital simulation using the EABSS
framework:
Title
Exploratory Study of Proactive Machines in a Hospital Scenario
Context
Proactive machines are an idea that in the future hospitals will have automated machines as well as
Doctors and Nurses that can help and treat patients. This project uses the idea of proactive machines
to test the impact of technology on mental state and opinions within patients. The project also
incorporates the use of fuzzy logic for some of the decision making.
Gathering Knowledge
All information for this project has been gathered from a conceptual model created by a focus group
lead by Peer-Olaf Siebers which had the aim of reflecting on the consequences of digital mental health
solutions considering the relationship between machines and humans.
Step 1: Defining Objectives
Aim and objectives
Reflect on the consequences of digital mental health solutions
Hypotheses
Trust between machines and humans will affect human-human relationships in a negative way
Machine-look influences our decisions
Experimental factors
Number of doctors, proactive machines and patients
Number of beds
Responses
Mental state of patients, e.g. happiness
Opinions of each patient on the doctors and robots
The trust of the patients on the robots
Step 2: Defining the Scope
Level of abstraction
For this model the high level scope decision was to only consider the treatment of patients by doctors
and proactive machines and to consider the visitors visiting patients. Any movement of agents or
geographical location were not considered as there is no need for this and it would add more
complexity which is not needed. I have chosen to use a hypothetical hospital setting to keep it realistic
without having the problems of gathering data which would be required if a real-world hospital was
used.
8
Scope table
The next step was to define the scope of the simulation model, considering which elements should be
included/excluded and stating why I have made this decision. The result can be found in Table 1.
Table 1: Scope table (part 1/2)
Table 1: Scope table (part 2/2)
9
Step 3: Defining Key Activities
For this stage, use case diagrams were made to show the specific use cases that each actor must do.
Figure 7 shows the use case diagram from the original conceptual model and includes a lot of actors
that are not in the final simulation. Figure 8 shows the use case diagram from the final simulation that
includes just the actors and use cases that are included in the final simulation.
Figure 7: Use case diagram of original conceptual model
Figure 8: Use case diagram of final simulation
Step 4: Defining Stereotypes
I decided to use a habit table approach to create the stereotypes. Stereotypes were only needed for
the doctors and robots because these were the only agents where stereotyping would have been
relevant to the hypotheses. Table 2 shows that a doctor can be one of two stereotypes, a senior or
junior. This difference in level means that the senior doctors treat patients quicker than the junior
doctors. Table 3 shows that robots can also be one of two stereotypes, they can either be more
humanlike in appearance or more robot-like. Whichever one they are, affects the patients opinions
on them when they are treated by them.
10
Table 2: Doctor stereotypes
Table 3: Robot stereotypes
Step 5: Defining Agent and Object Templates
Class diagram
Figure 9 shows a class diagram for the whole simulation. It shows each agent/class and what is
contained inside them as well as how each agent/class is connected to each other. The solid black
diamonds represent strong aggregation also known as composition.
11
Figure 9: Class diagram for final simulation
State machine diagrams
The figures below (Figure 10-14) show the state machine diagrams for each of the agents. Most agents
have very simple state machines with usually less than 3 main states. They are mostly self-explanatory
with names on each of the states that describe what they do. I have made use of three different types
of transitions which are timeout transitions, condition transitions and message transitions. Timeout
transitions trigger after a certain amount of time, condition transitions trigger when a condition is true
and message transitions trigger when a message is received.
Figure 10: Patient agent
Figure 11: Doctor agent
12
Figure 12: Robot agent
Figure 13: Visitor agent
Figure 14: Bed agent
Step 6: Defining Interactions
The figures below (Figure 15-17) show the most significant tasks and interactions of the simulation
represented in sequence diagrams. I haven't transferred all interactions into sequence diagrams
because there would be too many and it is not important to represent every interaction with a
sequence diagram.
13
Figure 15: Patient getting treated by robot
Figure 16: Patient getting treated by doctor
Figure 17: Visitor visits patient
5 The Final Version of the Hospital Simulation Model
Once the conceptual model was completed I have updated the simulation model to reflect all design
features. Figure 18 shows a screenshot of the final hospital simulation model while running. The
benefit of this final model in comparison to the previous versions created along the way is that this
version is catered towards two of the hypotheses on the original conceptual model so, these being
"Trust between machines and humans will affect human-human relationships in a negative way" and
14
"Machine-look influences our decisions (trust)". This means that we can see valuable outputs that are
related to the original ideas of the conceptual model e.g. the lines connecting each patient to other
patients show how their trust values differ, if the line is red it means that those two patients' trust
values differ a lot meaning their relationship is not good but if it is green it means their relationship is
good. These colours are affected depending on lots of different aspects in this model that were not
included in the previous, simpler models.
Figure 18: Screenshot of the final version of the simulation in progress
6 Conclusions
During my internship I finalised the conceptualisation of a simulation model, based on incomplete
information previously gathered by Peer-Olaf Siebers and his colleagues, for exploratory studies of the
impact of proactive machines in a hospital scenario. I have then implemented the simulation model,
and documented my implementation using the EABSS framework.
The next step will be a more in-depth validation of the model. In particular we need to establish face
validity by showing it to the people who provided us with the incomplete conceptual model. Once this
has been done and the colleagues are happy with our interpretation of the information provided, we
would need to run some experiments to test the hypotheses define in the conceptual model.
NB: The final version of the hospital simulation model is available upon request. Please get in touch
with Peer-Olaf Siebers ([email protected]).
Acknowledgement
The internship this working paper is based upon was run jointly by Peer-Olaf Siebers and Christian
Wagner. The original (incomplete) conceptual model this work is based upon was donated by Peer-
Olaf Siebers, Tommy Nilsson, and Elvira Perez Vallejos. Many thanks to all supporters!
15
References
Bonabeau E (2002) Agent-based modeling: Methods and techniques for simulating human systems.
Proceedings of the National Academy of Sciences, 99(suppl 3), pp. 7280-7287.
Fowler M (2004) UML Distilled: A Brief Guide to the Standard Object Modeling Language. Addison-
Wesley Professional
Rouse M (n.d.) Definition of fuzzy logic. https://searchenterpriseai.techtarget.com/definition/fuzzy-
logic (last accessed 30/08/2018)
Siebers PO, Perez Vallejos E, and Nilsson T (unpublished) Conceptual model draft "Reflecting on
consequences of digital mental health solutions"
Siebers PO and Aickelin U (2008) Introduction to multi-agent simulation. In: Adam F and Humphreys
P (Eds). Encyclopaedia of Decision Making and Decision Support Technologies, Pennsylvania:
Idea Group Publishing, pp. 554-564.
Siebers PO and Klügl F (2017) What software engineering has to offer to agent-based social
simulation. In: Edmonds B and Meyer R (eds.) 'Simulating Social Complexity: A Handbook' - 2e,
pp. 81-117. New York: Springer.
Wagner C and Pierfitte M (n.d.) JuzzyOnline.
https://ieeexplore.ieee.org/abstract/document/6891548/ (last accessed 30/08/2018)
XJTEK (2018) AnyLogic software. https://www.anylogic.com/ (last accessed 30/08/2018)
Zhang T, Siebers PO and Aickelin U (2012) A Three-Dimensional Model of Residential Energy
Consumer Archetypes for Local Energy Policy Design in the UK. Energy Policy, pp. 102-110.
16
|
0809.0009 | 2 | 0809 | 2012-05-18T14:36:33 | Distributed Parameter Estimation in Sensor Networks: Nonlinear Observation Models and Imperfect Communication | [
"cs.MA",
"cs.IT",
"cs.IT"
] | The paper studies distributed static parameter (vector) estimation in sensor networks with nonlinear observation models and noisy inter-sensor communication. It introduces \emph{separably estimable} observation models that generalize the observability condition in linear centralized estimation to nonlinear distributed estimation. It studies two distributed estimation algorithms in separably estimable models, the $\mathcal{NU}$ (with its linear counterpart $\mathcal{LU}$) and the $\mathcal{NLU}$. Their update rule combines a \emph{consensus} step (where each sensor updates the state by weight averaging it with its neighbors' states) and an \emph{innovation} step (where each sensor processes its local current observation.) This makes the three algorithms of the \textit{consensus + innovations} type, very different from traditional consensus. The paper proves consistency (all sensors reach consensus almost surely and converge to the true parameter value,) efficiency, and asymptotic unbiasedness. For $\mathcal{LU}$ and $\mathcal{NU}$, it proves asymptotic normality and provides convergence rate guarantees. The three algorithms are characterized by appropriately chosen decaying weight sequences. Algorithms $\mathcal{LU}$ and $\mathcal{NU}$ are analyzed in the framework of stochastic approximation theory; algorithm $\mathcal{NLU}$ exhibits mixed time-scale behavior and biased perturbations, and its analysis requires a different approach that is developed in the paper. | cs.MA | cs | IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
1
Distributed Parameter Estimation in Sensor
Networks: Nonlinear Observation Models and
Imperfect Communication
Soummya Kar, Jos´e M. F. Moura and Kavita Ramanan
Abstract
The paper studies distributed static parameter (vector) estimation in sensor networks with nonlinear observation
models and noisy inter-sensor communication. It introduces separably estimable observation models that generalize
the observability condition in linear centralized estimation to nonlinear distributed estimation. It studies two distributed
estimation algorithms in separably estimable models, the NU (with its linear counterpart LU) and the NLU. Their
update rule combines a consensus step (where each sensor updates the state by weight averaging it with its neighbors'
states) and an innovation step (where each sensor processes its local current observation.) This makes the three
algorithms of the consensus + innovations type, very different from traditional consensus. The paper proves consistency
(all sensors reach consensus almost surely and converge to the true parameter value,) efficiency, and asymptotic
unbiasedness. For LU and NU, it proves asymptotic normality and provides convergence rate guarantees. The three
algorithms are characterized by appropriately chosen decaying weight sequences. Algorithms LU and NU are analyzed
in the framework of stochastic approximation theory; algorithm NLU exhibits mixed time-scale behavior and biased
perturbations, and its analysis requires a different approach that is developed in the paper.
Keywords: Asymptotic normality, consensus, consensus + innovations, consistency, distributed parameter estimation,
Laplacian, separable estimable, spectral graph theory, stochastic approximation, unbiasedness.
A. Background and Motivation
I. INTRODUCTION
The paper studies distributed inference, in particular, distributed estimation, as consensus+innovations algorithms
that generalize distributed consensus by combining, at each time step, cooperation among agents (consensus) with
assimilation of their observations (innovations). Our consensus + innovations algorithms contrast with: i) standard
consensus, see the extensive literature, e.g., [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14],
where in each time step only local averaging of the neighbors' states occurs, and no observations are processed;
and ii) distributed estimation algorithms, see recent literature,e.g., [15], [16], [17], [18], [19], where between
measurement updates a large number of consensus steps (theoretically, an infinite number of steps) is taken.
Combined consensus+innovations algorithms are natural when a distributed network estimates a spatially varying
random field defined at M spatial locations, say, for simplicity, a temperature field. The goal is to reconstruct at
each and every sensor an accurate image of the entire spatial distribution of the M-dimensional field, assuming
that at each time step each sensor makes a noisy measurement of the temperature at its single location. Without
cooperation (no consensus step,) the processing of the successive temperature readings at each sensor (successive
innovation steps) leads to a reliable estimate of the temperature at the sensor location -- but provides no clue about the
temperature distribution at the other M − 1 locations. On the other hand, if sensors cooperate (consensus iterates,)
but only process the initial measurement, as in traditional consensus, they converge to the average temperature across
2
1
0
2
y
a
M
8
1
]
A
M
.
s
c
[
2
v
9
0
0
0
.
9
0
8
0
:
v
i
X
r
a
Manuscript received August 29, 2008; revised February 05, 2011; accepted November 15, 2011. The work of Soummya Kar and Jos´e
M. F. Moura was supported by NSF under grants CCF-1011903 and CCF-1018509, and by AFOSR grant FA95501010291. The work of Kavita
Ramanan was supported by the NSF under grants DMS 0405343 and CMMI 0728064.
Soummya Kar and Jos´e M. F. Moura are with the Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh,
PA, USA 15213 (e-mail: [email protected], [email protected], ph: (412) 268-6341, fax: (412) 268-3890.)
Kavita Ramanan is with the Division of Applied Mathematics, Brown University, Providence, RI 02912 (e-mail: Kavita [email protected].)
Communicated by M. Lops, Associate Editor for Detection and Estimation.
Digital Object Identifier 10.1109/TIT.2012.219450
c(cid:13)2012 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or
future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale
or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
2
the field, not to an estimate of the M-dimensional temperature distribution. The distributed consensus+innovations
algorithms that we introduce achieve both; each sensor converges to an estimate of the entire M-dimensional field
by combining consensus and processing of the sensors measurements. Subsequent to this paper, analysis of detection
consensus+innovations type algorithms is, e.g., in [20], [21].
Important questions that arise with consensus+innovations algorithms include: i) convergence: do the algorithms
converge and if so in what sense; ii) consensus: do the agents reach a consensus on their field estimates; iii) dis-
tributed versus centralized: how good is the distributed field estimate at each sensor when compared with the
centralized estimate obtained by a fusion center, in other words are the distributed estimate sequences consistent,
and asymptotically unbiased, efficient, or normal; and iv) rate of convergence: what is the rate at which the distributed
estimators converge. These questions are very distinct from the convergence issues considered in the "consensus
only" literature.
We present three distributed consensus+innovations inference algorithms: LU for linear observation models (as
when each sensor makes a noisy reading of the temperature at its location, see Section II-E;) and two algorithms,
NU and NLU, for nonlinear observation models (like in power grids when each sensor measures a phase differential
through a sinusoidal modulation, see Section IV-D.) The paper introduces the conditions on the sensor observations
model (separable estimability that we define) and on the communication network (connectedness on average) for
the distributed estimates to converge. The separable estimability, akin to global observability, and connectedness, is
an intuitively pleasing condition and in a sense minimal -- distributed estimation cannot do better than (the optimal)
centralized estimator, hence, the model better be (globally) observable (but not necessarily locally observable;) and
if the sensors need to cooperate to assimilate the data collected by the distributed network, the network should be
connected (on average,) or the sensors will work in isolation.
In contrast with other settings, e.g., linear distributed least-mean-square (LMS) approaches to parameter esti-
mation, e.g, [22], [23], [24], [25], we study distributed estimation in the usual framework of linear or nonlinear
observations of a (vector) parameter in noise1, when the dimension, Mn, of the observation at each sensor n
is Mn (cid:28) M, and the parameter estimation model is locally unobservable, i.e., each individual sensor cannot
recover the entire M-dimensional parameter from its Mn-dimensional observation, even if noiseless. Through
cooperation (consensus) a (local) sensor estimator may converge to an estimate of the entire M-dimensional field,
by simultaneously combining at each time i, its estimate, its observation (innovation), and the estimates received
from the sensors with which it communicates. We show in the paper conditions under which this holds.
We extend this distributed estimation model to include sensor and link or communication channel failures, random
communication protocols, and quantized communication. These conditions make the problem more realistic when
a large number of agents are involved since inexpensive sensors are bounded to fail at random times, packet loss
in wireless digital communications cause links to fail intermittently, agents can communicate asynchronously via a
random protocol like gossip or one of its variants, and the agents may be resource constrained and have a limited
bit budget for communication. We make no distributional assumptions on the sensors and link failures, they can
be spatially correlated, [28], but are temporally uncorrelated2. We show that, under these broad conditions, the
three distributed estimation algorithms, LU, NU, and NLU, are consistent if the observation model is separable
estimable (see Section III-A) and the network is connected on average.
Algorithms LU, NU, and NLU: LU applies when the noisy observations are linear on the parameter. For the
linear model, the separably estimable condition reduces to a rank condition on the global observability Grammian.
LU combines at each time iteration the consensus term with the innovations associated with the new observation.
Note that, in this algorithm, as well as with the other two nonlinear algorithms, the dimension of the local observation
for sensor n, Mn, is much smaller than the dimension M of the field, i.e., Mn (cid:28) M. The algorithm NU generalizes
LU to nonlinear separably estimable models. It is very important to note that, in both algorithms, LU and NU, the
same asymptotically decaying to zero time-varying weight sequence is associated with the consensus and innovation
updates; in other words, both the consensus and innovation terms of the algorithm exhibit the same decay rate.
Because of this, it is enough to resort to stochastic approximation techniques to prove consistency, asymptotic
unbiasedness, and asymptotic normality for both algorithms, LU and NU. For a treatment of general distributed
stochastic algorithms see [29], [30], [31], [32]. Beyond consistency, we characterize explicitly for the LU algorithm
1In this paper, we restrict attention to static parameter (fields). Time-varying parameters/signals are considered elsewhere, see, for example,
[26],[27] for estimation/filtering of fading (non-stationary) parameters and general time-varying linear dynamical systems, respectively.
2This dynamic network is more general and subsumes the erasure network model, where the link failures are independent over space and
time.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
3
the asymptotic variance and compare it with the asymptotic variance of the centralized optimal scheme. For the
NU algorithm, and general models, it is difficult to find explicitly a Lyapounov function (as needed by stochastic
approximation). However, with weaker assumptions on the nonlinear observation model (Lipschitz continuity and
certain growth properties,) we guarantee the existence of a Lyapounov function; hence, demonstrate asymptotic
normality of the NU estimates, see Theorems 18 and 19 in Section III-C. These conditions are much easier to
verify than guessing the form of a Lyapounov function. Also, in the proof of Theorem 18, we actually show how to
use these conditions to determine a Lyapounov function explicitly, which can then be used to analyze convergence
rates.
The third algorithm, NLU, applies when the observation models are only continuous and not Lipschitz continuous.
NLU is however a mixed time-scale algorithm, where the consensus time-scale dominates the innovations time-
scale, and consists of unbiased perturbations (detailed explanation is provided in the paper.) Because of this mixed
time-scales, the NLU algorithm does not fall under the purview of standard stochastic approximation theory, and
to show its consistency requires an altogether different framework as developed in the paper, see Theorems 21
and 22, in Section IV.
The two algorithms NLU and NU represent different tradeoffs. We show consistency for NLU under weaker
assumptions (observation model continuity) than for NU (Lipschitz continuity plus growth conditions.) On the
other hand, when these more stringent conditions hold, NU provides convergence rate guarantees and asymptotic
normality; these follow from standard stochastic approximation theory that apply to NU but not to NLU.
Brief comment on the literature. We contrast our work with relevant recent literature on distributed estimation.
Papers [15], [17], [33], [18] study estimation in static networks, where either the sensors take a single snapshot of
the field at the start and then initiate distributed consensus protocols (or more generally distributed optimization,
as in [17]) to fuse the initial estimates, or the observation rate of the sensors is assumed to be much slower
than the inter-sensor communicate rate, thus permitting a separation of the two time-scales. On the contrary, our
consensus+innovations algorithm combines fusion (consensus) and observation (innovation) updates in the same
iteration. The network is dynamic with channel failures, the protocols are random, and the sensors fail. Unlike [15],
[17], [33], [18], our approach does not require distributional assumptions on the observation noise, and we make
explicit the structural assumptions on the observation model (separable estimability) and network connectivity
needed to guarantee consistent parameter estimates at every sensor. These structural assumptions are quite weak
and are necessary for centralized estimators to obtain consistency.
There is considerable work in linear distributed least-mean-square (LMS) approaches to parameter estimation in
static networks, e.g., [22], [23], [24], [25]. While LMS is also a consensus+innovation type algorithm, we show
how our linear algorithm LU and LMS are quite distinct, with a very different setup and goal. In LMS, for example,
for channel estimation or channel equalization, [34], in adaptive filtering, [35], or in system identification, see [36],
the observations zn(i) are the output of a noisy finite impulse response channel (or a linear system to be identified)
excited by a random input sequence u(i) (these random input sequences are the regressors.) The unknown channel
impulse response θ is to be estimated by a stochastic gradient type algorithm that has available (in the channel
estimation or training phase) both the random inputs and the regressors. In contrast, in the distributed estimation
problem we study, for example, for the LU, the observations at sensor n and time i are
For example, in (1), the observation matrices Hn(i) could be of the form,
zn(i) = Hn(i)θ∗ + ζn(i)
Hn(i) =
1
p
δn(i)H n
(1)
(2)
where δn(i) is a zero-one Bernoulli variable to account for sensor failures, p > 0 denotes the sensing probability,
and the mean value H n models the normal operation of the sensor, e.g., measuring the local temperature, or
an average of local temperatures. Equation (1) is the usual model in parameter estimation or waveform filtering,
while (2) extends this model in a significant way to random intermittent measurements. In our distributed estimation
algorithms, we do not know the random observation matrix Hn(i) (only its first and second moment), while in the
LMS where they play the role of the regressors, they are usually known to the LMS algorithm.
We contrast further our linear distributed LU with linear distributed LMS. References [22], [24], [25] use non-
decaying combining weights that lead to a residual tracking error; under appropriate assumptions, these algorithms
can be adapted to certain time-varying tracking scenarios; we consider time varying processes in other work, [26],
[27]. Reference [23] considers decaying weight sequences as we do in LU, thereby establishing also L2 convergence
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
4
to the desired parameter value. All these works deal with distributed linear problems, while our work emphasizes
distributed estimators for linear and nonlinear sensor observation models and establishes their convergence properties.
We present the necessary (minimal) structural conditions that the distributed sensing model (given) and the inter-
sensor communication network should satisfy to guarantee the existence of successful distributed estimators. Also,
apart from treating generic separably estimable nonlinear observation models, in the linear case, our algorithms NU
and LU lead to asymptotic normality in addition to consistency and asymptotic unbiasedness in random time-varying
networks with quantized inter-sensor communication and sensor failures.
Remark. We noted that the NLU algorithm is mixed time scale; this means a stochastic algorithm where two
potentials act in the same update step with different weight or gain sequences. This should not be confused with
(centralized) stochastic algorithms with coupling (see [37]), where a quickly switching parameter influences the
relatively slower dynamics of another state, leading to averaged dynamics. We note further in this context that [38]
(and references therein) develops methods to analyze mixed time scale (centralized) algorithms in the context of
simulated annealing. In [38], the role of our innovation (new observation) potential is played by a martingale
difference term. However, in our study, the innovation is not a martingale difference process, and a key step in the
analysis is to derive pathwise strong approximation results to characterize the rate at which the innovation process
converges to a martingale difference process.
We briefly comment on the organization of the remaining of the paper. The rest of this section introduces notation
and preliminaries to be adopted throughout the paper. To motivate the generic nonlinear problem, we study the linear
case (algorithm LU) in Section II. Section III studies the generic separably estimable models and the algorithm
NU, whereas algorithm NLU is presented in Section IV. Finally, Section V concludes the paper.
B. Notation
(cid:16)
(cid:17)
k 1k1T
√
1/
k
and dithered quantization to be used in the sequel.
For completeness, this subsection sets notation and presents preliminaries on algebraic graph theory, matrices,
Preliminaries: We adopt the following notation. Rk: the k-dimensional Euclidean space; Ik: k×k identity matrix;
k : the rank one k × k projector, whose
1k, 0k: column vector of ones and zeros in Rk, respectively; Pk = 1
1k; (cid:107)·(cid:107): the standard
only non-zero eigenvalue is one, and the corresponding normalized eigenvector is
Euclidean 2-norm when applied to a vector and the induced 2-norm when applied to matrices, which is equivalent
to the matrix spectral radius for symmetric matrices; θ ∈ U ⊂ RM : the parameter to be estimated; θ∗: the true (but
unknown) value of the parameter θ; xn(i) ∈ RM : the estimate of θ∗ at time i at sensor n -- without loss of generality
(wlog), the initial estimate, xn(0), at time 0 at sensor n is a non-random quantity; (Ω,F): common measurable
space where all the random objects are defined; Pθ∗ [·] and Eθ∗ [·]: the probability and expectation operators when
the true (but unknown) parameter value is θ∗ -- when the context is clear, we abuse notation by dropping the subscript.
All inequalities involving random variables are to be interpreted a.s. (almost surely.)
Spectral graph theory: For the undirected graph G = (V, E), V = [1··· N ] is the set of nodes or vertices,
V = N, and E is the set of edges. The unordered pair (n, l) ∈ E if there exists an edge between nodes n and l.
The graph G is simple if devoid of self-loops and multiple edges and connected if there exists a path3 between each
pair of nodes. The neighborhood of node n is Ωn = {l ∈ V (n, l) ∈ E}. The degree dn = Ωn of node n is the
number of edges with n as one end point, and D = diag (d1 ··· dN ) is the degree matrix, the diagonal matrix with
diagonal entries the degrees dn. The structure of the graph can be described by the symmetric N × N adjacency
matrix, A = [Anl], Anl = 1, if (n, l) ∈ E, Anl = 0, otherwise. The graph Laplacian matrix, L, is L = D − A;
it is a a positive semidefinite matrix whose eigenvalues can be ordered as 0 = λ1(L) ≤ λ2(L) ≤ ··· ≤ λN (L).
1N being the corresponding normalized eigenvector. The
The smallest eigenvalue λ1(l) is zero, with
multiplicity of the zero eigenvalue equals the number of connected components of the network; for a connected
graph, λ2(L) > 0. This second eigenvalue is the algebraic connectivity or the Fiedler value of the network; see
[39], [40], [41] for detailed treatment of graphs and their spectral theory.
Kronecker product: Since we are dealing with vector parameters, most of the matrix manipulations will involve
Kronecker products. For example, the Kronecker product of the N ×N matrix L and IM will be an N M ×N M ma-
trix, denoted by L⊗IM . We will deal often with matrices of the form C = [INM − bL ⊗ IM − aINM − PN ⊗ IM ].
It follows from the properties of Kronecker products and of the matrices L and PN that the eigenvalues of the
matrix C are −a and 1 − bλi(L) − a, 2 ≤ i ≤ N, each being repeated M times.
√
1/
(cid:16)
(cid:17)
N
3A path between nodes n and l of length m is a sequence (n = i0, i1, · · · , im = l) of vertices, such that (ik, ik+1) ∈ E ∀ 0 ≤ k ≤ m− 1.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
5
[k1∆,··· , kM ∆]T(cid:12)(cid:12)(cid:12) ki ∈ Z, ∀i
(cid:111)
(cid:110)
We now review results from statistical quantization theory.
Dithered quantization: We assume that all sensors are equipped with identical uniform, dithered quantizers
q(·) : RM → QM applied componentwise, with countable alphabet QM =
. We
assume the dither satisfies the Schuchman conditions (see [42], [43], [44], [45],) so that the error sequence for
subtractively dithered systems ([43]) {ε(i)}i≥0
(3)
is an i.i.d. sequence of uniformly distributed random variables on [−∆/2, ∆/2), which is independent of the
input sequence {y(i)}i≥0. In (3), the dither sequence {ν(i)}i≥0 is i.i.d. uniformly distributed random variables on
[−∆/2, ∆/2), independent of the input sequence {y(i)}i≥0; we refer to [46] where we use this model and make
Consistency and asymptotic unbiasedness: We recall standard definitions from sequential estimation theory
ε(i) = q(y(i) + ν(i)) − (y(i) + ν(i))
further relevant comments.
(see, for example, [47]).
Definition 1 (Consistency) : A sequence of estimates {x•(i)}i≥0 is called consistent if
(cid:104)
Pθ∗
lim
i→∞
(cid:105)
x•(i) = θ∗
= 1, ∀θ∗ ∈ U
or, in other words, if the estimate sequence converges a.s. to the true parameter value. The above definition of
consistency is also called strong consistency. When the convergence is in probability, we get weak consistency. In
this paper, we use the term consistency to mean strong consistency, which implies weak consistency.
Definition 2 (Asymptotic Unbiasedness) :
A sequence of estimates {x•(i)}i≥0 is called asymptotically unbiased if
Eθ∗ [x•(i)] = θ∗, ∀θ∗ ∈ U
lim
i→∞
II. DISTRIBUTED LINEAR PARAMETER ESTIMATION: ALGORITHM LU
In this section, we consider the algorithm LU for distributed parameter estimation when the observation model
is linear. This problem motivates the generic separably estimable nonlinear observation models considered in
Sections III and IV. Section II-A sets up the distributed linear estimation problem and presents the algorithm LU.
Section II-B establishes the consistency and asymptotic unbiasedness of the LU algorithm, where we show that,
under the LU algorithm, all sensors converge a.s. to the true parameter value, θ∗. Convergence rate analysis
(asymptotic normality) is carried out in Section II-C, while Section II-E illustrates LU with an example.
A. Problem Formulation: Algorithm LU
Let θ∗ ∈ RM be an M-dimensional parameter to be estimated by a network of N sensors. Sensor n makes the
observations:
zn(i) = Hn(i)θ∗ + ζn(i) ∈ RMn
(6)
Each sensor observes only a subset of Mn of the components of θ∗, or Mn linear combinations of a few components
of θ∗, with Mn (cid:28) M. We make the following assumptions.
(A.1)Observation Noise: The observation noise process,
with finite second moment. In particular, the observation noise covariance is bounded and independent of i
1 (i),··· , ζ T
is i.i.d. zero mean,
i≥0
(cid:110)
ζ(i) =(cid:2)ζ T
E(cid:2)ζ(i)ζ T (j)(cid:3) = Sζδij, ∀i, j ≥ 0
N (i)(cid:3)T(cid:111)
(7)
where the Kronecker symbol δij = 1 if i = j and zero otherwise. Note that the observation noises at different
sensors may be correlated during a particular iteration. Eqn. (7) states only temporal independence. The spatial
correlation of the observation noise makes our model applicable to practical sensor network problems, for instance,
for distributed target localization, where the observation noise is generally correlated across sensors.
(cid:2)H 1,··· , H N
(A.2)Sensor Failures: The observation matrices, {[H1(i),··· , HN (i)]}i≥0, form an i.i.d. sequence with mean
(cid:3) and finite second moment. In particular, we have
Hn(i) = H n + (cid:101)Hn(i), ∀i, n
(4)
(5)
(8)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
6
(cid:110)(cid:104)(cid:101)H1(i),··· , (cid:101)HN (i)
(cid:105)(cid:111)
where, H n = E [Hn(i)] , ∀i, n and the sequence
is zero mean i.i.d. with finite second
moment. Here, also, we require only temporal independence of the observation matrices, but allow them to be
spatially correlated. For example, Hn(i) = δn(i)H n, with δn(i) an iid sequence of Bernoulli variables modeling
intermittent sensor failures.
Remark 3 The LU update does not use the instantaneous observation matrices, Hn(i), only their ensemble
averages. This is a distinction between LU and LMS. LMS (for example, in adaptive filtering) assumes the
random matrices Hn(i) are, together with the observations, also available (see Chapter 4 of [48]). In parameter
estimation, the Hn(i) model sensor failures and LU has no control over their instantiations. Hence, while in LMS
it may be reasonable to use the instantaneous values of the random regressors Hn(i), in the setting we consider,
the instantaneous realizations of the observation matrices are not available.
i≥0
L(i) = L +(cid:101)L(i), ∀i ≥ 0
(A.3)Mean Connectedness, Link Failures, and Random Protocols: The graph Laplacians
(9)
are a sequence of i.i.d. matrices with mean L = E [L(i)]. We make no distributional assumptions on the {L(i)}.
Although independent at different times, during the same iteration, the link failures can be spatially dependent,
i.e., correlated. This is more general and subsumes the erasure network model, where the link failures are
independent over space and time. Wireless sensor networks motivate this model since interference among the
wireless communication channels correlates the link failures over space, while, over time, it is still reasonable to
assume that the channels are memoryless or independent. Connectedness of the graph is an important issue. The
random instantiations G(i) of the graph need not be connected; in fact, all these instantiations may be disconnected.
We only require the graph to be connected on average. This is captured by requiring that λ2
to capture a broad class of asynchronous communication models; for example, the random asynchronous gossip
protocol analyzed in [49] satisfies λ2
(A.4) Independence: The sequences {L(i)}i≥0, {ζn(i)}1≤n≤N, i≥0, {Hn(i)}1≤n≤N,i≥0, {νm
nl(i)} are mutually
independent.
We introduce the distributed observability condition required for convergence of the LU linear estimation algorithm.
(cid:0)L(cid:1) > 0 and hence falls under this framework.
(cid:0)L(cid:1) > 0, enabling us
Definition 4 (Distributed observability) The observation system (6) is distributedly observable if the matrix G is
full rank
G =
T
n H n
H
(10)
N(cid:88)
n=1
This distributed observability extends the observability condition for a centralized estimator that is needed to get
a consistent estimate of the parameter θ∗. We note that the information available to the n-th sensor at any time i
about the corresponding observation matrix is just the mean H n, and not the random Hn(i). Hence, the state
update equation uses only the H n's, as given in (11) below.
(A.5) Observability: The distributed observation system (6) is distributedly observable in the sense of definition 4.
Algorithm LU: We consider now the algorithm LU for distributed parameter estimation in the linear observation
model (6). Starting from some initial deterministic estimate of the parameters4, xn(0) ∈ RM , each sensor n
generates a sequence of estimates, {xn(i)}i≥0 by the following distributed iterative algorithm:
xn(i + 1) = xn(i) − α(i)
(xn(i) − q (xl(i)
(11)
b
(cid:88)
(cid:0)zn(i) − H nxn(i)(cid:1)
l∈Ωn(i)
+νnl(i))) − H
T
n
4The initial states may be random, we assume deterministic for notational simplicity.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
7
where {q(xl(i) + νnl(i))}l∈Ωn(i) is the dithered quantized exchanged data. In (11), the sequence of weights {α(i)}
satisfies the persistence condition B5 given in the Appendix A; b > 0 is a constant and {α(i)}i≥0 is a sequence
of weights with properties to be defined below. The algorithm (11) is distributed because for sensor n it involves
only the data from the sensors in its neighborhood Ωn(i). Using (3), the state update can be written as
We rewrite (12) in compact form. Define the random vectors, Υ(i) and Ψ(i) ∈ RNM with vector components
It follows from the Schuchman conditions on the dither, see Section I-B and [46], that
b
(cid:88)
εnl(i)
(xn(i) − xl(i))
T
n
− H
l∈Ωn(i)
l∈Ωn(i)
l∈Ωn(i)
xn(i + 1) = xn(i) − α(i)
Υn(i) = − (cid:88)
(cid:0)zn(i) − H nxn(i)(cid:1) − bνnl(i) − bεnl(i)
νnl(i) and Ψn(i) = − (cid:88)
E(cid:104)(cid:107)Ψ(i)(cid:107)2(cid:105)
≤ N (N − 1)M ∆2
E(cid:104)(cid:107)Υ(i)(cid:107)2(cid:105)
E(cid:104)(cid:107)Ψ(i)(cid:107)2(cid:105) ≤ N (N − 1)M ∆2
N (i)(cid:3)T and define the matrices
(cid:104)
E(cid:104)(cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:105) ≤ 2 sup
E(cid:104)(cid:107)Υ(i)(cid:107)2(cid:105)
E [Υ(i)] = E [Ψ(i)] = 0, ∀i
+ 2 sup
= sup
(cid:105)
sup
i
sup
i
12
3
i
i
i
+
= ηq
(cid:105)
T
N H N
DH = diag
DH = DH D
T
N
(cid:104)
H
H
T
1 ··· H
T
H = diag
(cid:104)
T
1 H 1 ··· H
(cid:17)
T
H
z(i) − D
b(L(i) ⊗ IM )x(i)D
(cid:16)
(Υ(i) + Ψ(i)) (Υ(i) + Ψ(i))T(cid:105)
T
H x(i)
+ bΥ(i) + bΨ(i)
(cid:105)
−DH
Sq = E(cid:104)
(cid:16)
Then, the compact vector form of the LU algorithm is
x(i + 1) = x(i) − α(i)
In the LU algorithm (19), the covariance matrix of the noise is defined as
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
from which we then have
x(i) =(cid:2)xT
1 (i)··· xT
The iterations in (11) can be written in compact form. Stack all sensors state estimates in a long state vector estimate
Markov: We characterize the state vector estimate x(i). Consider the filtration, {F x
From (A1) -- (A4), L(i), z(i), Υ(i), Ψ(i) are independent of F x
B. Consistency of LU
F x
i = σ
x(0),{L(j), z(j), Υ(j), Ψ(j)}0≤j<i
i ; so, {x(i),F x
We consider consistency and asymptotic unbiasedness.
Lemma 5 Consider LU under Assumptions (A.1)-(A.5). Then,(cid:2)bL ⊗ IM + DH
i }i≥0, given by
(cid:17)
i }i≥0 is a Markov process.
(cid:3) is symmetric positive definite.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
8
Proof: Symmetricity is obvious. It also follows from the properties of Laplacian matrices and the structure
of DH that these matrices are positive semidefinite. Then the matrix (cid:2)bL ⊗ IM + DH
the matrix(cid:2)bL ⊗ IM + DH
(cid:3) is not positive definite. Then, there exists, x ∈ RNM , such that x (cid:54)= 0 and
being the sum of two positive semidefinite matrices. To prove positive definiteness, assume, on the contrary, that
(cid:3) is positive semidefinite,
xT(cid:2)bL ⊗ IM + DH
xT(cid:2)L ⊗ IM
(cid:3) x = 0
(cid:3) x = 0, xT DH x = 0
From the positive semidefiniteness of L ⊗ IM and DH, and the fact that b > 0, it follows
Partition x as x =(cid:2)xT
1 ··· xT
N
(cid:3)T
the fact that λ2(L) > 0, that (23) holds iff
where a ∈ RM , and a (cid:54)= 0. Also, (23) implies
N(cid:88)
n=1
Let G be as in (10). Equations (25) and (24) imply
(23)
, xn ∈ RM ,∀1 ≤ n ≤ N. It follows from the properties of Laplacian matrices and
xn = a, ∀n
xT
n H
T
n H nxn = 0
(24)
(25)
a contradiction, since G > 0 by Assumption (A.5) and a (cid:54)= 0. Thus,(cid:2)bL ⊗ IM + DH
Theorem 6 (LU: Asymptotic unbiasedness) Let the LU algorithm under (A.1)-(A.5). Then, {xn(i)}i≥0, at sensor n
(cid:3) > 0.
aT Ga = 0,
(26)
is asymptotically unbiased
(22)
(27)
(28)
(29)
(30)
Proof: Taking expectations on both sides of (19) and by the independence assumption (A.4),
E [xn(i)] = θ∗, 1 ≤ n ≤ N
lim
i→∞
E [x(i + 1)] = E [x(i)] − α(i)(cid:2)b(cid:0)L ⊗ IM
E [z(i)](cid:3)
E [x(i)] − DH
+DH
(cid:1) E [x(i)] +
Subtracting 1N ⊗ θ∗ from both sides of (28), noting that
(cid:0)L ⊗ IM
(cid:1) (1N ⊗ θ∗) = 0,
we have
(cid:0)bL ⊗ IM + DH
matrix(cid:2)bL ⊗ IM + DH
Let λmin
E [z(i)] = DH (1N ⊗ θ∗)
DH
E [x(i + 1)] − 1N ⊗ θ∗ =(cid:2)INM − α(i)(cid:0)bL ⊗ IM +
(cid:1), λmax
(cid:0)bL ⊗ IM + DH
(cid:3) (Lemma 5.) Since α(i) → 0 (Assumption (B.5), Appendix A),
+ DH )] [E [x(i)] − 1N ⊗ θ∗]
λmax
Continuing the recursion in (31), we have, for i > i0,
∃i0 (cid:51): α(i0) ≤
1
(31)
(cid:1) be the smallest and largest eigenvalues of the positive definite
(cid:1) , ∀i ≥ i0
(cid:0)bL ⊗ IM + DH
i−1(cid:89)
(cid:2)INM − α(j)(cid:0)bL ⊗ IM +
[E [x(i0)] − 1N ⊗ θ∗]
(32)
(33)
j=i0
E [x(i)] − 1N ⊗ θ∗ =
+ DH )]
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Eqn. (33) implies
It follows from (32)
Eqns. (34,35) now give for i > i0
i−1(cid:89)
j=i0
i > i0
(cid:107)INM − α(j)
(cid:107)E [x(i)] − 1N ⊗ θ∗(cid:107) ≤
(cid:1)(cid:13)(cid:13)(cid:107)E [x(i0)] − 1N ⊗ θ∗(cid:107) ,
(cid:0)bL ⊗ IM + DH
(cid:13)(cid:13)INM − α(j)(cid:0)bL ⊗ IM + DH
(cid:1)(cid:13)(cid:13) =
(cid:1) , j ≥ i0
(cid:0)bL ⊗ IM + DH
i−1(cid:89)
(cid:1)(cid:1)(cid:107)E [x(i0)] − 1N ⊗ θ∗(cid:107) ,
(cid:0)bL ⊗ IM + DH
1 − α(j)λmin
(1 − α(j)
λmin
j=i0
(cid:107)E [x(i)] − 1N ⊗ θ∗(cid:107) ≤
Finally, from the inequality 1 − a ≤ e−a, 0 ≤ a ≤ 1, get
(cid:0)bL ⊗ IM + DH
Since, λmin
(cid:107)E [x(i)] − 1N ⊗ θ∗(cid:107) ≤ e−λmin(bL⊗IM +DH)(cid:80)i−1
(cid:1) > 0 and the weight sequence sums to infinity, the theorem follows since
(cid:107)E [x (i0)] − 1N ⊗ θ∗(cid:107) ,
i > i0
α(j)
j=i0
(cid:107)E [x(i)] − 1N ⊗ θ∗(cid:107) = 0
lim
i→∞
9
(34)
(35)
(36)
(37)
(38)
Before proceeding to Theorems 7 and 10 establishing the consistency and asymptotic normality of the LU, the
reader may refer to Appendix A, where useful results on stochastic approximation are discussed.
= 1
(cid:105)
lim
i→∞
n is consistent
xn(i) = θ∗, ∀n
(Appendix A). Recall the filtration, {F x
and noting that
Theorem 7 (LU: Consistency) Consider LU under (A.1) -- (A.5). Then, the estimate sequence {xn(i)}i≥0 at sensor
P(cid:104)
Proof: The proof follows by showing that {x(i)}i≥0 satisfies the Assumptions (B.1)-(B.5) of Theorem 29
i }i≥0, in (21). Rewrite (19) by adding and subtracting the vector 1N ⊗ θ∗
(cid:1) (1N ⊗ θ∗) = 0
(cid:20)
(cid:1) (x(i)−
b(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:16)
x(i)+
(cid:21)
H 1N ⊗ θ∗
−1N ⊗ θ∗) + b
+DH (x(i)−1N ⊗ θ∗)−DH
x(i + 1) = x(i) − α(i)
(cid:0)L ⊗ IM
z(i)−D
(cid:17)
(40)
(41)
(39)
T
+
+ bΥ(i) + bΨ(i)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
10
In the notation of Theorem 29, Appendix A, let R (x) and Γ (i + 1, x, ω) as in (42) and (43) below and rewrite
(41)
x(i + 1) = x(i) + α(i) [R(x(i)) + Γ (i + 1, x(i), ω)]
R (x) = −(cid:2)bL ⊗ IM + DH
Γ (i + 1, x, ω) = −(cid:104)
b
(cid:3) (x − 1N ⊗ θ∗)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:17)
H 1N ⊗ θ∗
T
x−
−DH
z(i) − D
(cid:16)
+ bΥ(i) + bΨ(i)
(cid:105)
Under the Assumptions (A.1)-(A.5), for fixed i+1, the random family, {Γ (i + 1, x, ω)}x∈RNM , is F x
i+1 measurable,
zero-mean and independent of F x
i . Hence, the assumptions (B.1)-(B.2) of Theorem 29 are satisfied.
We now show the existence of a stochastic potential function V (·) satisfying the remaining Assumptions (B.3)-
(B.4) of Theorem 29. To this end, define
V (x) = (x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
(cid:3)
(x − 1N ⊗ θ∗)
Clearly, V (x) ∈ C2 with bounded second order partial derivatives. It follows from the positive definiteness of
(cid:3) (Lemma 5), that
(cid:2)bL ⊗ IM + DH
Since the matrix(cid:2)bL ⊗ IM + DH
hence, there exists a constant c1 > 0, such that
V (1N ⊗ θ∗) = 0, V (x) > 0, x (cid:54)= 1N ⊗ θ∗
(cid:3) is positive definite, the matrix(cid:2)bL ⊗ IM + DH
(x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
(x − 1N ⊗ θ∗) ≥
(cid:3)2
≥ c1(cid:107)x − 1N ⊗ θ∗(cid:107)2, ∀x ∈ RNM
(cid:3)2 is also positive definite and
(45)
(42)
(43)
(44)
(46)
(47)
(48)
(49)
(50)
It then follows that
sup
(cid:107)x−1N⊗θ∗(cid:107)>
− 2
inf
(cid:107)x−1N⊗θ∗(cid:107)>
(R (x) , Vx (x)) =
(cid:110)
(x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
(cid:3)2
(cid:111)
(x − 1N ⊗ θ∗)
c1 (cid:107)x − 1N ⊗ θ∗(cid:107)2
≤ −2
inf
≤ −2c12 < 0
(cid:107)x−1N⊗θ∗(cid:107)>
Thus, Assumption (B.3) is satisfied. From (42)
From (43) and the independence Assumption (A.4)
= − 1
2
(R (x) , Vx (x))
(cid:107)R (x)(cid:107)2 =
= (x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
E(cid:104)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:105)
(cid:20)
(x − 1N ⊗ θ∗)T(cid:16)
b(cid:101)L(i) ⊗ IM
(cid:20)(cid:13)(cid:13)(cid:13)DH
(cid:16)
+ b2E(cid:104)(cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:105)
z(i) − D
= E
+ E
=
(cid:3)2
(x − 1N ⊗ θ∗)
(cid:21)
(cid:17)2
H 1N ⊗ θ∗
T
(x − 1N ⊗ θ∗)
(cid:17)(cid:13)(cid:13)(cid:13)2(cid:21)
+
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Since the random matrix (cid:101)L(i) takes values in a finite set, there exists a constant c2 > 0, such that, ∀x ∈ RNM ,
Again, since(cid:0)bL ⊗ IM + DH
≤ c2(cid:107)x − 1N ⊗ θ∗(cid:107)2
(x − 1N ⊗ θ∗) ≤
(cid:17)2
(x − 1N ⊗ θ∗)T(cid:16)
b(cid:101)L(i) ⊗ IM
(cid:1) is positive definite, there exists a constant c3 > 0, such that, ∀x ∈ RNM ,
(x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
(cid:3) (x − 1N ⊗ θ∗) ≥
(cid:21)
(cid:20)
(cid:17)2
(x − 1N ⊗ θ∗)T(cid:16)
b(cid:101)L(i) ⊗ IM
(cid:3)(x − 1N ⊗ θ∗)
(x − 1N ⊗ θ∗)T(cid:2)bL ⊗ IM + DH
+ b2E(cid:104)(cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:105)
E(cid:104)(cid:13)(cid:13)DH z(i) − DH 1N ⊗ θ∗(cid:13)(cid:13)2(cid:105)
≥ c3(cid:107)x − 1N ⊗ θ∗(cid:107)2.
(x − 1N ⊗ θ∗)
≤ c2
c3
= c4V (x)
> 0. The term
for some constant c4 = c2
c3
We then have from (51)-(52)
≤
E
11
(51)
(52)
(53)
is bounded by a finite constant c5 > 0, as it follows from Assumptions (A.1)-(A.5). We then have from (49)-(50)
(cid:107)R (x)(cid:107)2 + E(cid:104)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:105) ≤
(R (x) , Vx (x)) + c4V (x) + c5 ≤
≤ − 1
2
≤ c6 (1 + V (x)) − 1
2
(R (x) , Vx (x))
xn(i) = θ∗, ∀n] = 1
P[ lim
i→∞
where c6 = max (c4, c5) > 0. This verifies Assumption (B.4) of Theorem 29. Assumption (B.5) is satisfied by the
choice of {α(i)}i≥0. It then follows that the process {x(i)}i≥0 converges a.s. to 1N ⊗ θ∗. In other words,
which establishes consistency of LU.
The proof above can be modified to show L2 convergence of the sensor estimates to θ∗. Due to the fact that the
LU update rule is linear, the driving noise terms are L2 bounded, and the stable (as shown in the proof) Lyapunov
function V (·) assumes a positive definite quadratic form. Hence, by studying the recursion of the deterministic
sequence {E[V (x(i))]} and by similar arguments5 as in [46] (Lemma 4), we conclude the following:
Lemma 8 (Mean square convergence) Let the hypotheses of Theorem 7 hold and, in addition, the weight sequence
{α(i)} satisfy the following:
(54)
(55)
(56)
(57)
where a > 0 and .5 < τ ≤ 1. Then, the a.s. convergence in Theorem 7 holds in L2 also, i.e., for all n,
a
α(i) =
(i + 1)τ
E(cid:104)(cid:107)xn(i) − θ∗(cid:107)2(cid:105)
= 0
lim
i→∞
C. Asymptotic Variance: LU
In this subsection, we carry out a convergence rate analysis of the LU algorithm by studying its moderate
deviation characteristics. We summarize here some definitions and terminology from the statistical literature, used
to characterize the performance of sequential estimation procedures (see [47]).
5Note, that Lemma 4 in [46] does not assume the additional term due to new observations at each iteration. However, this does not pose
difficulties as the observation weights are the same as the consensus weights.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
12
Definition 9 (Asymptotic Normality) A sequence of estimates {x•(i)}i≥0 is asymptotically normal if for every
θ∗ ∈ U, there exists a positive semidefinite matrix S(θ∗) ∈ RM×M , such that,
i (x•(i) − θ∗) =⇒ N (0M , S(θ∗))
(58)
√
The matrix S(θ∗) is called the asymptotic variance of the estimate sequence {x•(i)}i≥0.
In the following we prove the asymptotic normality of the LU algorithm and explicitly characterize the resulting
lim
i→∞
asymptotic variance. To this end, define
...
...
SH = E
(cid:101)H1(i)
DH
DH
(cid:101)H1(i)
1N θ∗
...(cid:101)HN (i)
T
1N θ∗
(cid:3) and recall Sζ, Sq in (7) and (20).
(cid:1) be the smallest eigenvalue of(cid:2)bL ⊗ IM + DH
...(cid:101)HN (i)
...
...
(cid:0)bL ⊗ IM + DH
Let λmin
We now state the main result of this subsection, establishing the asymptotic normality of the LU algorithm.
Theorem 10 (LU: Asymptotic efficiency/ normality) Let the LU algorithm under (A.1)-(A.5) with link weight se-
quence, {α(i)}i≥0 that is given by:
a
, ∀i
α(i) =
i + 1
for some constant a > 0. Let {x(i)}i≥0 be the state sequence generated. Then, if a >
1
2λmin(bL⊗IM +DH)
, we have
(i) (x(i) − 1N ⊗ θ∗) =⇒ N (0, S(θ∗))
(59)
(60)
(61)
(62)
(63)
(64)
(65)
where
and
(cid:112)
(cid:112)
(cid:90) ∞
S(θ∗) = a2
eΣvS0eΣvdv,
Σ = −a(cid:2)bL ⊗ IM + DH
(cid:3) +
0
1
2
I,
S0 = SH + DH SζD
T
H + b2Sq
In particular, at any sensor n, the estimate sequence, {xn(i)}i≥0 is asymptotically normal:
(i) (xn(i) − θ∗) =⇒ N (0, Snn(θ∗))
where, Snn(θ∗) ∈ RM×M denotes the n-th principal block of S(θ∗).
Proof: The proof involves a step-by-step verification of Assumptions (C.1)-(C.5) of Theorem 29 (Appendix A),
since the Assumptions (B.1)-(B.5) are already shown to be satisfied (see, Theorem 7.) Recall R (x) and Γ (i + 1, x, ω)
from Theorem 7 ((42)-(43)). From (42), Assumption (C.1) of Theorem 29 is satisfied with
B = −(cid:2)bL ⊗ IM + DH
(cid:3)
and δ (x) ≡ 0. Assumption (C.2) is satisfied by hypothesis, while the condition a >
Σ = −a(cid:2)bL ⊗ IM + DH
(cid:3) +
1
2
INM = aB +
1
2
INM
(66)
1
2λmin(bL⊗IM +DH)
implies
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
is stable, and hence Assumption (C.3). To verify Assumption (C.4), we have from Assumption (A.4)
From the i.i.d. assumptions, we note that all the three terms on the R.H.S. of (67) are independent of i, and, in
particular, the last two terms are constants. For the first term, we note that
from the bounded convergence theorem, as the entries of
are bounded and
For the second term on the R.H.S. of (67), we have
lim
(cid:17)
= b2E
E
x→1N⊗θ∗
xxT(cid:16)(cid:101)L(i) ⊗ IM
(cid:110)(cid:101)L(i)
A (i, x) = E(cid:2)Γ (i + 1, x, ω) ΓT (i + 1, x, ω)(cid:3)
(cid:20)(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)T(cid:21)
xxT(cid:16)(cid:101)L(i) ⊗ IM
+ E(cid:2)(cid:0)DH z(i) − DH 1N ⊗ θ∗(cid:1)
(cid:0)DH z(i) − DH 1N ⊗ θ∗(cid:1)T(cid:105)
(Υ(i) + Ψ(i)) (Υ(i) + Ψ(i))T(cid:105)
+ b2E(cid:104)
(cid:17)T(cid:21)
(cid:20)(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
E(cid:104)(cid:0)DH z(i)−DH 1N ⊗ θ∗(cid:1)(cid:0)DH z(i)−DH 1N ⊗ θ∗(cid:1)T(cid:105)
1N θ∗
DH
(cid:101)H1(i)
...(cid:101)HN (i)
T +
1N θ∗
DH
(cid:101)H1(i)
...(cid:101)HN (i)
+ E(cid:104)
(cid:105)
(1N ⊗ θ∗) = 0
= E
(cid:111)
...
...
...
...
DH ζζ T D
T
H
= 0
i≥0
= SH + DH SζD
T
H
=
x−
(x − θ∗)−
(cid:107)Γ (i + 1, x, ω)(cid:107)2 =
−(cid:0)DH z(i) − DH 1N ⊗ θ∗(cid:1) + bΥ(i) + bΨ(i)(cid:13)(cid:13)2
(cid:13)(cid:13)(cid:13)b
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:13)(cid:13)2
−(cid:0)DH z(i) − DH 1N ⊗ θ∗(cid:1) + bΥ(i) + bΨ(i)
(cid:20)(cid:13)(cid:13)(cid:13)(cid:16)
(cid:13)(cid:13)(cid:13)2
b(cid:101)L(i) ⊗ IM
+ b2 (cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:105)
+(cid:13)(cid:13)DH z(i) − DH 1N ⊗ θ∗(cid:13)(cid:13)2
(x − θ∗)
≤ 9
(cid:17)
+
13
(67)
(68)
(69)
(70)
(71)
(72)
where the last step follows from (59),(7). Finally, we note the third term on the R.H.S. of (67) is b2Sq, see (20).
We thus have from (67)-(70)
We now verify Assumption (C.5). Consider a fixed > 0. We note that (157) is a restatement of the uniform
lim
integrability of the random family,(cid:8)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:9)
(cid:13)(cid:13)(cid:13)b
i→∞, x→x∗ A (i, x) = SH + DH SζD
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
i≥0, (cid:107)x−θ∗(cid:107)<. From (43), we have
T
H + b2Sq = S0
(cid:104)(cid:107)y1(cid:107)2 + (cid:107)y2(cid:107)2 + (cid:107)y3(cid:107)2(cid:105)
(cid:13)(cid:13)(cid:13)2 ≤ c22
14
, for vectors y1, y2, y3.
(73)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
where we used (69) and the inequality, (cid:107)y1 + y2 + y3(cid:107)2 ≤ 9
From (51) we note that, if (cid:107)x − θ∗(cid:107) < ,(cid:13)(cid:13)(cid:13)(cid:16)
(cid:17)
b(cid:101)L(i) ⊗ IM
(cid:111)
From (72), the family [defined in (74) below](cid:110)(cid:101)Γ (i + 1, x, ω)
(cid:8)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:9)
dominates the family
(x − θ∗)
where
(cid:101)Γ (i + 1, x, ω) = 9
i≥0,(cid:107)x−θ∗(cid:107)<
(cid:104)
i≥0, (cid:107)x−θ∗(cid:107)< ,
c22 +(cid:13)(cid:13)DH z(i) − DH 1N ⊗ θ∗(cid:13)(cid:13)2
+ b2 (cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:105)
The family
(cid:111)
(cid:110)(cid:101)Γ (i + 1, x, ω)
(cid:8)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:9)
is i.i.d. and hence uniformly integrable (see [50]). Then the family
(cid:110)(cid:101)Γ (i + 1, x, ω)
(cid:111)
i≥0, (cid:107)x−θ∗(cid:107)< is also uniformly integrable since it is dominated by the uniformly integrable
(see [50]). Thus (C.1)-(C.5) are verified and the theorem follows.
i≥0, (cid:107)x−θ∗(cid:107)<
family
(74)
i≥0, (cid:107)x−θ∗(cid:107)<
N(cid:88)
D. A Simulation Example
Fig. 1 (b) shows the performance of LU for the network of N = 45 sensors in Fig. 1 (a), where the sensors
are deployed randomly on a 25 × 25 grid. The sensors communicate in a fixed radius and are further constrained
to have a maximum of 6 neighbors per node. The true parameter θ∗ ∈ R45. Each node is associated with a single
n , the unit vector of zeros, except entry n that is 1. For the experiment, each
component of θ∗, i.e., H n = eT
component of θ∗ is generated by an instantiation of a zero mean Gaussian random variable of variance 25. The
parameter θ∗ represents the state of the field to be estimated. In this example, the field is white, stationary, and
hence each sample of the field has the same Gaussian distribution and is independent of the others. More generally,
the components of θ∗ may correspond to random field samples, as dictated by the sensor deployment, that can
possibly arise from the discretization of a field governed by a PDE. Each sensor observes the corresponding field
component in additive Gaussian noise. For example, sensor 1 observes z1(t) = θ∗1 + ζ1(t), where ζ1(t) ∼ N (0, 1).
Clearly, such a model satisfies the distributed observability condition
G =
T
n H n = I45 = G−1
H
(75)
n=1
Fig. 1(b) shows the normalized error at every sensor plotted against the iteration index i for an instantiation of the
algorithm. The normalized error for the n-th sensor at time i is given by the quantity (cid:107)xn(i) − θ∗(cid:107) /45, i.e., the
estimation error normalized by the dimension of θ∗. We note that the errors converge to zero as established by the
theoretical findings. The decrease is rapid at the beginning and slows down at i increases. This is a standard property
of stochastic approximation based algorithms, consequence of the decreasing weight sequence α(i) required for
convergence. From the plots, although the individual sensors are low rank observations of the true parameter, by
collaborating, each sensor reconstructs the true parameter value, as desired.
E. An Example
From Theorem 10 and (59), we note that the asymptotic variance is independent of θ∗, if the observation matrices
are non-random. In that case, it is possible to optimize (minimize) the asymptotic variance over the weights a and
b. In the following, we study a special case permitting explicit computations and that leads to interesting results.
Consider a scalar parameter (M = 1) and let each sensor n have the same i.i.d. observation model,
zn(i) = hθ∗ + ζn(i)
(76)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
15
(a)
(b)
Fig. 1.
estimation error at each sensor.
Illustration of distributed linear parameter estimation. (a) Example network deployment of 45 nodes. (b) Convergence of normalized
where h (cid:54)= 0 and {ζn(i)}i≥0, 1≤n≤N is a family of independent zero mean Gaussian random variables with variance
σ2. In addition, assume unquantized inter-sensor exchanges. We define the average asymptotic variance per sensor
attained by the algorithm LU as
where S is given by (62) in Theorem 10. From Theorem 10, we have S0 = σ2h2IN and, hence, from (62)
From (63) the eigenvalues of 2Σv are(cid:2)−2abλn(L) −(cid:0)2ah2 − 1(cid:1)(cid:3) v for 1 ≤ n ≤ N and we have
N
=
0
a2σ2h2
SLU =
1
N
Tr (S)
(cid:18)(cid:90) ∞
(cid:19)
(cid:90) ∞
Tr(cid:0)e2Σv(cid:1) dv
e2Σvdv
0
SLU =
a2σ2h2
N
Tr
(cid:90) ∞
N(cid:88)
N(cid:88)
n=1
a2σ2h2
N
a2σ2h2
SLU =
=
=
e[−2abλn(L)−(2ah2−1)]vdv
0
1
2abλn(L) + (2ah2 − 1)
N
n=1
a2σ2h2
N (2ah2 − 1)
+
a2σ2h2
N
N(cid:88)
n=2
1
2abλn(L) + (2ah2 − 1)
(77)
(78)
(79)
In this case, the constraint a >
2λmin(bL⊗IM +DH )
optimum a, b design to minimize SLU is given by
S∗
LU =
1
in Theorem 10 reduces to a > 1
2h2 , and hence the problem of
inf
2h2 , b>0
a> 1
SLU
(80)
It is to be noted, that the first term on the last step of (79) is minimized at a = 1
non-negative under the constraint) goes to zero as b → ∞ for any fixed a > 0. Hence, we have
h2 and the second term (always
The above shows that, by setting a = 1
h2 and b sufficiently large in LU, SLU is arbitrarily close to S∗
LU
S∗
LU =
σ2
N h2
(81)
.
051015202505101520255010015020025030000.511.522.53IterationsErrorpercomponentofθ∗ IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
16
We compare this optimum achievable asymptotic variance per sensor, S∗
LU
, attained by LU to that attained by a
centralized scheme. In the centralized scheme, there is a central estimator, which receives measurements from all
the sensors and computes an estimate based on all measurements. In this case, the sample mean estimator is an
efficient estimator (in the sense of Cram´er-Rao) and the estimate sequence {xc(i)}i≥0 is given by
(cid:88)
n,i
xc(i) =
1
N ih
zn(i)
(82)
(83)
and we have
√
i (xc(i) − θ∗) ∼ (0,Sc)
where, Sc is the variance (which is also the one-step Fisher information in this case, see, [47]) and is given by
Sc =
σ2
N h2
(84)
From (81) we note that,
S∗
LU = Sc
(85)
Thus the average asymptotic variance attainable by the distributed algorithm LU is the same as that of the optimum
(in the sense of Cram´er-Rao) centralized estimator having access to all information simultaneously. This is an
interesting result, as it holds irrespective of the network topology. In particular, however sparse the inter-sensor
communication graph is, the optimum achievable asymptotic variance is the same as that of the centralized efficient
estimator. Note that weak convergence itself is a limiting result, and, hence, the rate of convergence in (61) in
Theorem 10 will, in general, depend on the network topology.
F. Some generalizations
We discuss some generalizations of the basic LU scheme before proceeding to the nonlinear observation models
addressed in the subsequent sections. We start by revisiting the scalar example in Section II-E for which the
distributed LU is shown to achieve the performance of the optimal centralized estimator. Interestingly, the above
example is not an isolated special case and has several important implications. The observation that by increasing
b > 0 we can achieve asymptotic variance as close as desired to the centralized estimator hints to a more general
time-scale separation in the case of unquantized transmissions. Intuitively, for a fixed b > 0, the weight associated to
the consensus potential is bα(i), which goes to zero at the same rate as that of the innovation potential. Hence, in the
long run, a non-negligible (in the scale {α(i)}) amount of time is required to disseminate new information acquired
by a sensor. In other words, the rate of uncertainty reduction in LU depends on both the rate of new information
acquisition at the sensors and the rate of information dissemination in the network. On the contrary, in a centralized
scenario, no additional time is incurred for information dissemination and the rate of uncertainty reduction is the
same as the rate of information acquisition. This is manifested, in general, in the asymptotic variance of LU,
which is larger than its centralized counterpart due to the additional overhead of the mixing terms (Theorem 10).
This suggests that, if the mixing can be carried out at a faster scale, the additional overhead due to the mixing
time will not be observed at the time scale of observation acquisition and, in effect, the distributed scheme will
lead to similar asymptotic variance as in the centralized setting. This is noted in Section II-E, where increasing
the relative weight b of the consensus or mixing potential leads to a time scale separation between information
dissemination and acquisition. Increasing b beyond bounds suggests that we replace the decreasing weight sequence
{α(i)} from the consensus term and retain it with a constant weight, or more generally, a weight sequence {β(i)},
that asymptotically dominates {α(i)}. Such a mixed time scale extension of the LU is introduced and analyzed
in [26]. The results in [26] show that the conclusion in Section II-E for the scalar example (the distributed achieves
the centralized performance in terms of asymptotic variance) holds in more general vector parameter settings by
appropriately tuning the consensus and innovation weights. This is significant, as it justifies the applicability of
distributed estimation schemes over centralized approaches.
The development in this paper assumes stationarity of the sensor observations over time. While this is applicable
and is a commonly used assumption in many statistical models, some scenarios inherently lead to non-stationary
observation time series. For example, consider a distributed sensor network monitoring a target that fades over
time. In this example, the sensor observation models are no longer stationary as the SNR (signal to noise ratio)
decays over time, the decay rate being a function of the fading characteristics. Treating nonstationarity requires
modification of the algorithm (intuitively, the update rules are no longer stationary) and is pursued in [26].
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
17
The above did not exploit the physical significance of the parameter θ. That θ may itself come from a spa-
tially distributed random field was only implicit in the distributed observation model. Typical examples include
instrumenting a spatially distributed random field (say a temperature surface) with a sensor network. Another
example is of cyberphysical systems, where a network of physical entities equipped with sensors are deployed over
a large geographical region. A well known example in this setting is the power grid, a large distributed network
of generators and loads. Our results imply that, under appropriate observability conditions, the physical field θ
may be reconstructed completely at each node (sensor).6 However, for such systems, the parameter θ representing
the physical field is quite large dimensional, may be of the order of 103 or more, as exemplified by the power
grid.7 It is then impractical and unnecessary to reconstruct the high dimensional parameter in its entirety at each
node. On the other hand, the node may be interested only in its state, or those of its close neighbors. In general,
the observation at each sensor reflects the coupling of a few local physical states and hence, acting alone, a node
may not be able to recover its state uniquely. In [51], we develop approaches to address this problem, where each
node wants to reconstruct a few components8 of the large state vector. The estimation approach would lead to
low dimensional data exchanges between neighboring sensors (nodes) and local estimate updates would involve
only those components, the node wants to reconstruct. Due to the partial information exchange between sensors
and the fact that sensors may have different goals, the distributed observability no longer culminates to the sum
of network connectivity and global connectivity, but requires more subtle relations between the observation model
and the network topology. In general, the scope of such problems of distributed estimation with partial inter-sensor
information exchange is quite broad and challenging, and we refer the reader to [51] (Chapter 5) for an exposition.
III. NONLINEAR OBSERVATION MODELS: AGORITHM NU
The previous section developed the algorithm LU for distributed parameter estimation when the observation model
is linear. In this section, we extend the previous development to accommodate more general classes of nonlinear
observation models. We comment briefly on the organization of this section. In Section III-B, we introduce notation
and setup the problem, and in Section III-C we present the NU algorithm for distributed parameter estimation for
nonlinear observation models and establish conditions for its consistency.
A. Nonlinear Observation Models
Similar to Section II, let θ∗ ∈ U ⊂ RM be the true but unknown parameter value. We assume that the domain
U is an open set in RM . In the general case, the observation model at each sensor n consists of an i.i.d. sequence
{zn(i)}i≥0 in RMN with
(cid:90)
Pθ∗ [zn(i) ∈ D] =
dFn,θ∗ , ∀ D ∈ BMN
(86)
where Fn,θ∗ denotes the distribution function of the random vector zn(i). For consistent parameter estimates, even
in centralized settings, some form of observability needs to be imposed on the nonlinear model. In the following,
we assume that the distributed observation model is separably estimable, a notion which we introduce now.
D
Definition 11 (Separably Estimable) Let {zn(i)}i≥0 be the i.i.d. observation sequence at sensor n, where 1 ≤ n ≤
N. We call the parameter estimation problem to be separably estimable, if there exist functions gn(·) : RMN (cid:55)−→
RM , ∀1 ≤ n ≤ N, such that the function h(·) : U (cid:55)−→ RM is continuous and invertible on U
N(cid:88)
n=1
h(θ) =
1
N
Eθ [gn(zn(i))]
(87)
Remark 12 Before providing examples of separably estimable observation models and demonstrating the applica-
bility of the notion, we comment on the definition.
6A node, in this context, refers to the physical entity at a geographical location, for example, a generator in a power grid. The sensing or
measurement unit associated to a node is referred to as a sensor. The state of a node represents the field intensity at that point, for example,
the phase of a generator.
7For problems involving infinite dimensional systems, such as the temperature distribution over a domain in the Euclidean space, any reasonable
discretization would lead to a large dimensional θ.
8These components may vary from node to node.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
18
T
(i) We note that the factor 1
N in (87) is just for notational convenience, as will be seen later. In fact, the 1
N can be
absorbed by redefining the functions gn(·). Also, it is implicitly assumed that the random vectors gn(zn(i))
are integrable w.r.t. the measures Pθ for θ ∈ U.
(ii) Let h(U) ⊂ RM denote the range of h(·). The continuity of h(·) implies that h(U) is open. Let h−1 : h(U) (cid:55)−→
RM denote the inverse of h(·) (which is necessarily continuous on h(U).) It then follows that h−1(·) has a
measurable extension defined over all of RM . In the following we will assume that h−1(·) has been measurably
extended and, by abusing notation, denote this extension by h−1.
(iii) We will show that the notion of separably estimable models introduced above is, in fact, necessary and
sufficient to guarantee the existence of consistent distributed estimation procedures for a wide range of practical
scenarios. This condition may also be viewed as a natural generalization of the observability constraint
of Assumption (A.5) in the linear model. Indeed, if, assuming the linear model, we define gn(zn(i)) =
n zn(i), ∀1 ≤ n ≤ N in (87), we have h(θ) = Gθ, where G is defined in (10). Then, invertibility of (87)
H
is equivalent to Assumption (A.5), i.e., to invertibility of G; hence, the linear model is an example of a
separably estimable problem. Note that, if an observation model is separably estimable, then the choice of
functions gn(·) is not unique. Indeed, given a separably estimable model, it is important to figure out an
appropriate decomposition, as in (87), because the convergence properties of the algorithms (Algorithm NU,
Section III-C) to be studied are intimately related to the behavior of these functions. Finally, we note that, in
general, M (cid:54)= M, and the dimension M of the range space of h(·) is very much linked to the memory and
transmission requirements of the distributed algorithm NLU to be studied in Section IV. In this sense, the
function h(·) plays the role of a complete sufficient statistic as used in classical (centralized) estimation, the
major difference being the distributed computability (to be made precise later) of h(·) in the current setting.
In Sections III-C and IV, respectively, we will present algorithms NU and NLU for distributed parameter
estimation in separably estimable models. While the NLU provides consistent parameter estimates for all separably
estimable models, the NU requires further (mainly of the Lipschitz type) conditions on the functions gn(·) and h(·).
However, in cases where the NU is applicable, it automatically leads to convergence rate guarantees in the context
of asymptotic normality. These differences are further clarified in Section V. Before discussing these algorithms in
detail, we provide examples of separably estimably models in the following.
Examples: Signal in additive noise models
We now demonstrate an important and large class of distributed observation models possessing the separably
estimable property, thus justifying the generality and applicability of the notion.
A wide range of observation models are of the signal in additive noise type. In particular, for each n, denote
by {ζn(i)} the zero mean i.i.d. observation noise at the n-th sensor of arbitrary distribution (the distribution may
vary from sensor to sensor.) The sensor observation model is said to be of signal in additive noise type, if the
observation sequence {zn(i)} at the n-th sensor is of the form:
zn(i) = fn(θ∗) + ζn(i)
(88)
Here fn : U (cid:55)−→ RMn denotes the transformed (nonlinearly) signal (or parameter) observed at sensor n, further
corrupted by additive noise. The following simple proposition characterizes the subclass of signal in additive noise
observable models with the separably estimable property:
Proposition 13 Let f : U (cid:55)−→ R(cid:80)N
additive noise observation model (see (88)) is separably estimable if f (·) is continuous and invertible on U.
pothetical centralized estimator having access to all the sensor observations at all times. Clearly, the (cid:80)N
Before providing the rather straightforward proof, we note the consequences of Proposition 13. Consider a hy-
n=1 Mn
dimensional i.i.d. observation sequence {z(i)} at such a center is given by:
N (θ)]T . Then, the above signal in
n=1 Mn be defined by, f (θ) = [f T
1 (θ)··· f T
z(i) = f (θ∗) + ζ(i)
(89)
In general, for arbitrary statistics of the noise sequence {ζ(i)}, it is necessary that the function f be invertible, for the
center to yield a consistent estimate of the parameter. In fact, for consistent centralized estimates, the invertibility of
f (·) is required, even when the observation noise is identically zero. On the other hand, Proposition 13 asserts that
the invertibility of f (·) (and its continuity) is sufficient to guarantee that the model is separably estimable and hence
the existence of consistent distributed estimation schemes. Hence, at least in the class of widely adopted signal in
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
19
additive noise models, centralized observability is equivalent to distributed observability (formulated here in terms
of separable estimability.) This further justifies the notion of separable estimability as a reasonable generalization
of the concept of centralized observability to distributed nonlinear settings. In Section IV-D, we will show that the
NLU algorithm provides a completely distributed approach to the static phase estimation problem in power grids
of generators and loads based on line flow measurements, an important practical example of a distributed nonlinear
signal in additive noise model.
gn : RMn (cid:55)−→ R(cid:80)N
Proof: The proof follows in a straightforward manner from the definition. For each n, define the function
n=1 Mn by
(90)
Recall 0M1 ∈ RM1 denotes the column vector of M1 zeros and so on. By the independence of the noise sequence
{ζ(i)}, it the follows that
gn(y) = [0T
M1
0T
M2
··· yT ··· 0T
MN ]T , ∀y ∈ RMn
Eθ∗ [gn(zn(i))] = f (θ∗)
(91)
N(cid:88)
n=1
The continuity and invertibility of f (·) then establishes the separable estimability of the model (Definition 11) by
the correspondence h(·) = 1
By using the same arguments we demonstrate a larger class of separably estimable models as follows:
N f (·).
Proposition 14 Let the observation sequence {zn(i)} at the n-th sensor be of the form:
zn(i) = fn(θ∗, ζ 1
n(i)} ∈ RM 1
n(i)) + ζ 2
n(i)
n(i), ζ 2
n (cid:55)−→ RMn, {ζ 1
where fn : U × RM 1
mean. Assuming that the moments exist, define the function f n : U (cid:55)−→ RMn, for each n, by
Further, let f : U (cid:55)−→ R(cid:80)N
is separably estimable if f (·) is continuous and invertible on U.
(cid:2)fn(θ∗, ζ 1
n(i))(cid:3)
n=1 Mn be defined by, f (θ) = [f
f n(θ) = Eθ
n × RMn is a temporally i.i.d. sequence and {ζ 2
T
1 (θ)··· f
T
N (θ)]T . Then, the observation model in (92)
(92)
n(i)} is zero
(93)
Remark 15 The generic model considered in Proposition 14 subsumes the class of signals with multiplicative noise
n(i) to zero.
models, by suitably defining the functions fn(θ∗, ζ 1
We also note that a general guideline for choosing the functions gn(·) for the signal in additive noise type models
based on problem data is given in (90). From a similar line of reasoning, it follows that the same choice of gn(·)
works for the larger class of separably estimable models considered in Proposition 14.
n(i)) and setting the additive noise component ζ 2
In the following subsection, we present the algorithm NU for distributed parameter estimation in nonlinear
separably estimable observation models.
B. Algorithm NU and Assumptions
Before introducing the algorithm, we formally state the generic observation and communication assumptions
required by the NU.
(D.1)Separably Estimable Model: The nonlinear observation model (86) is separably estimable (Definition 11).
In particular, at iteration i, the observations across different sensors need not be independent. In other words,
we allow spatial correlation, but require temporal independence. Also, other than the structural assumption of
separable estimability, no assumptions are required on the noise statistics, in particular, its distribution.
(D.2)Random Link Failure, Quantized Communication: The random link failure model is the model given in
sequence, as in (II-A)) are mutually independent. Let M =(cid:80)N
Section I-B; similarly, we assume quantized inter-sensor communication with subtractive dithering.
(D.3)Independence and Moment Assumptions: The sequences {L(i)}i≥0,{zn(i)}1≤n≤N, i≥0,{νm
nl(i)} (dither
n=1 Mn and define hn : RM (cid:55)−→ RM , by
hn(θ) = Eθ [gn(zn(i))] , ∀1 ≤ n ≤ N
(94)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
We make the assumption ∀θ ∈ U:
Eθ
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1
N
N(cid:88)
n=1
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2=η(θ) < ∞,
gn(zn(i)) − h(θ)
20
(95)
We thus assume the existence of quadratic moments of the (transformed) random variables gn(zn(i)). For example,
under the reasonable hypotheses of Propositions 13-14, the functions gn(·) may be taken to be linear, and
Assumption (D.3) then coincides with the existence of quadratic moment of the observations zn(i). In general,
since the choice of the functions gn(·) for a separably estimable model is not unique, the moment Assumption (D.3)
may enter as a selection criterion of the transformations gn(·).
In Section III-C and Section IV, we give two algorithms, NU and NLU, respectively, for the distributed estimation
problem (D.1)-(D.3) and provide conditions for consistency and other properties of the estimates.
C. Algorithm NU
In this subsection, we present the algorithm NU for distributed parameter estimation in separably estimable
Algorithm NU: Each sensor n performs the following estimate update:
models under Assumptions (D.1)-(D.3).
xn(i + 1) = xn(i) − α(i)
β (xn(i)−
(96)
(cid:88)
l∈Ωn(i)
q(xl(i) + νnl(i))) + Kn (hn(xn(i)) − gn(zn(i)))
(cid:8)xn(i) ∈ RM(cid:9)
based on xn(i), {q(xl(i) + νnl(i))}l∈Ωn(i), and zn(i), which are all available to it at time i. The sequence,
i≥0, is the estimate (state) sequence generated at sensor n. The weight sequence {α(i)}i≥0 satisfies
the persistence condition of Assumption (B.5) and β > 0 is chosen to be an appropriate constant. Finally,
Kn ∈ RM×M is an appropriately chosen matrix gain, possibly varying from sensor to sensor. Similar to (12)
the above update can be written in compact form as
x(i + 1) = x(i) − α(i) [β(L(i) ⊗ IM )x(i)+
+ K (M (x(i)) − J(z(i))) + Υ(i) + Ψ(i)]
(97)
1 (i)··· xT
N (i)]T is the vector of sensor states (estimates). The
where Υ(i), Ψ(i) are as in (13)-(15) and x(i) = [xT
functions M (x(i)) and J(z(i)) are given by
M (x(i)) =(cid:2)hT
J(z(i)) =(cid:2)gT
N (xN (i))(cid:3)T
N (zN (i))(cid:3)T
1 (x1(i))··· hT
1 (z1(i))··· gT
and K = diag(K1,··· ,KN ) is the block diagonal matrix of gains.
gains Kn to be IM ), the NU reduces to the LU updates (19).
As an example, for the linear observation model, by defining gn(zn(i)) to be H
,
(98)
(99)
T
n zn(i) (and choosing the matrix
We note that the update scheme in (97) is nonlinear and hence convergence properties can, in general, be
characterized through the existence of appropriate stochastic Lyapunov functions. In particular, if we can show that
the iterative scheme in (97) falls under the purview of a general result like Theorem 29, we can establish properties
like consistency, normality etc. To this end, we note, that (97) can be written as
x(i + 1) = x(i) − α(i)(cid:2)β(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
x(i) + K (M (x(i)) − M (1N ⊗ θ∗))
+ β
−K (J(z(i)) − M (1N ⊗ θ∗)) + Υ(i) + Ψ(i)]
(cid:1) (x(i) − 1N ⊗ θ∗)
(100)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
which becomes in the notation of Theorem 29
x(i + 1)=x(i)+α(i)[R(x(i)) + Γ (i + 1, x(i), ω)]
R (x) = −(cid:2)β(cid:0)L ⊗ IM
Γ (i + 1, x, ω) = −(cid:104)
(cid:1) (x − 1N ⊗ θ∗)
+ K (M (x) − M (1N ⊗ θ∗))(cid:3)
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
x−
β
− K (J(z(i)) − M (1N ⊗ θ∗)) + Υ(i) + Ψ(i)
(cid:18)
Fi=σ
x(0),
(cid:110)
(cid:111)
L(j),{zn(j)}1≤N ,Υ(j), Ψ(j)
0≤j<i
(cid:105)
(cid:19)
Consider the filtration, {Fi}i≥0,
21
(101)
(102)
(103)
(104)
(105)
(106)
(107)
Clearly, under (D.1)-(D.3), the {x(i)}i≥0 generated by NU is Markov w.r.t. {Fi}i≥0, and the definition in (103)
renders the random family, {Γ (i + 1, x, ω)}x∈RNM , Fi+1 measurable, zero-mean, and independent of Fi for fixed
i + 1. Thus (B.1)-(B.2) of Theorem 29 are satisfied, and we have the following.
Proposition 16 (NU:Consistency/ asymp. normality) Let the sequence {x(i)}i≥0 be generated by NU. Let R (x) , Γ (i + 1, x, ω) ,Fi
be as in (102)-(103). Then, if there exists a function V (x) satisfying (B.3)-(B.4) at x∗ = 1N ⊗ θ∗, the estimate
sequence {xn(i)}i≥0 at any sensor n is consistent. In other words,
Pθ∗ [ lim
i→∞
xn(i) = θ∗, ∀n] = 1
If, in addition, (C.1)-(C.4) are satisfied, the sequence {xn(i)}i≥0 at any sensor n is asymptotically normal.
Proposition 16 states that, a.s. asymptotically, the network reaches consensus, and the estimates at each sensor
converge to the true value of the parameter vector θ(cid:63). The Proposition relates these convergence properties of NU to
the existence of suitable Lyapunov functions. For a particular observation model characterized by the corresponding
functions hn(·), gn(·), if one can come up with an appropriate Lyapunov function satisfying the assumptions
of Proposition 16, then consistency and asymptotic normality are guaranteed. Existence of a suitable Lyapunov
condition is sufficient for consistency, but may not be necessary. In particular, there may be observation models
for which the NU algorithm is consistent, but there exists no Lyapunov function satisfying the assumptions of
Proposition 16.9 Also, even if a suitable Lyapunov function exists, it may be difficult to guess its form, because
there is no systematic (constructive) way of coming up with Lyapunov functions for generic models.
However, for our problem of interest, some additional weak assumptions on the observation model, for example,
Lipschitz continuity of the functions hn(·), will guarantee the existence of suitable Lyapunov functions, thus
establishing convergence properties of the NU algorithm. The rest of this subsection studies this issue and presents
different sufficient conditions on the observation model, which guarantee that the assumptions of Proposition 16 are
satisfied, leading to the a.s. convergence of the NU algorithm. For the development in the rest of the subsection,
we assume that M = M in the decomposition (87) and Kn = IM for all n. The extensions of Theorems 18-19 to
M (cid:54)= M and arbitrary gains Kn are immediate. We start with the following definition:
Definition 17 (Consensus Subspace) We define the consensus subspace, C ⊂ RM N as
y ∈ RNM(cid:12)(cid:12)(cid:12) y = 1N ⊗(cid:101)y,(cid:101)y ∈ RM(cid:111)
(cid:110)
C =
For y ∈ RNM , we denote its component in C by yC and its orthogonal component by y⊥
C
algorithm (Assumptions (D.1)-(D.3)). Further, ∀θ (cid:54)=(cid:101)θ ∈ RM , 1 ≤ n ≤ N, let the functions hn(·) be:
Theorem 18 (NU:Consistency under Lipschitz on hn) Let {x(i)}i≥0 be the state sequence generated by the NU
1) Lipschitz continuous with constants kn > 0, i.e.,
.
(cid:107)hn(θ) − hn((cid:101)θ)(cid:107) ≤ kn(cid:107)θ −(cid:101)θ(cid:107)
9This is because converse theorems in stability theory do not hold in general, see, [52].
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
2)
Define K as
(cid:16)
(cid:17)T(cid:16)
θ −(cid:101)θ
hn(θ) − hn((cid:101)θ)
(cid:17) ≥ 0.
K = max(k1,··· , kN )
(cid:104)
Pθ∗
lim
i→∞
(cid:105)
xn(i) = θ∗, ∀n
= 1
Then, for every β > 0, the estimate sequence is consistent. In other words,
The proof is in Appendix B. The conditions in (107)-(108) are much easier to verify than guessing a Lyapunov
function. Also, as will be shown in the proof, the conditions in Theorem 18 determine a Lyapunov function explicitly,
which may be used to analyze properties like convergence rate. The Lipschitz assumption is quite common in the
stochastic approximation literature, while the assumption in (108) holds for a large class of functions. As a matter
of fact, in the one-dimensional case (M = 1), it is satisfied if the functions hn(·) are non-decreasing. Also, in
general, it can be shown from the proof (Appendix B) that the Lipschitz continuity assumption in Theorem 18
may be replaced by continuity of the functions hn(·), 1 ≤ n ≤ N, and linear growth conditions, i.e., for constants
cn,1, cn,2 > 0,
(cid:107)hn(θ)(cid:107)2≤cn,1+cn,2(cid:107)θ(cid:107)2,∀θ ∈ RM , 1 ≤ n ≤ N.
(111)
We now present another set of sufficient conditions that guarantee consistency of NU. If the observation model
is separably estimable, in some cases even if the underlying model is nonlinear, it may be possible to choose the
functions, gn(·), such that the function h(·) possesses nice properties. This is the next result.
Theorem 19 (NU:Consistency -- h strict monotonicity) Consider the NU algorithm (Assumptions (D.1)-(D.3)). Sup-
pose that the functions gn(·) can be chosen, such that the functions hn(·) are Lipschitz continuous with constants
kn > 0 and the function h(·) satisfies
(cid:16)
(cid:17)T(cid:16)
θ −(cid:101)θ
h(θ) − h((cid:101)θ)
(cid:17)≥γ(cid:107)θ −(cid:101)θ(cid:107)2, ∀θ,(cid:101)θ ∈ RM
22
(108)
(109)
(110)
(112)
(113)
for some constant γ > 0. Then, for
K = max(k1,··· , kN ),
if
the algorithm NU is consistent, i.e.,
Pθ∗
(cid:104)
β >
K 2 + Kγ
γλ2L
,
xn(i) = θ∗, ∀n
lim
i→∞
(cid:105)
= 1
The proof is provided in Appendix B. We comment that, in comparison to Theorem 18, strengthening the assumptions
on h(·), see (112), considerably weakens the assumptions on the functions hn(·). Eqn. (112) is an analog of strict
monotonicity. For example, if h(·) is linear, the left hand side of (112) becomes a quadratic and the condition says
that this quadratic is strictly away from zero, i.e., monotonically increasing with rate γ.
IV. NONLINEAR OBSERVATION MODELS: ALGORITHM NLU
In this Section, we present the algorithm NLU for distributed estimation in separably estimable observation
models. As explained later, this is a mixed time-scale algorithm, where the consensus time-scale dominates the
observation update time-scale as time progresses. The NLU algorithm is based on the fact that, for separably
estimable models, it suffices to know h(θ∗), because θ∗ can be unambiguously determined from the invertible
function h(θ∗). To be precise, if the function h(·) has a continuous inverse, then any iterative scheme converging to
h(θ∗) will lead to consistent estimates, obtained by inverting the sequence of iterates. The algorithm NLU is shown
to yield consistent and unbiased estimators at each sensor for any separably observable model, under the assumption
that the function h(·) has a continuous inverse. Thus, the algorithm NLU presents a more reliable alternative than
the algorithm NU, because, as shown in Section III-C, the convergence properties of the latter can be guaranteed
only under certain assumptions on the observation model. We briefly comment on the organization of this section.
The NLU algorithm for separably estimable observation models is presented in Section IV-A. Section IV-B offers
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
23
interpretations of the NLU algorithm and presents the main results regarding consistency, mean-square convergence,
asymptotic unbiasedness proved in the paper. In Section IV-C we prove the main results about the NLU algorithm
and provide insights behind the analysis (in particular, why standard stochastic approximation results cannot be
used directly to give its convergence properties.) Finally, Section V presents discussions on the NLU algorithm
and suggests future research directions.
A. Algorithm NLU
Algorithm NLU: Let x(0) = [xT
1 ··· xT
N ]T be the initial set of states (estimates) at the sensors. The NLU
generates the sequence {xn(i)} ∈ RM at the n-th sensor according to the distributed recursive scheme:
h(xn(i)) −
xn(i + 1) = h−1
(cid:88)
− β(i)
[h(xn(i)) − q (h(xl(i)) + νnl(i))]
l∈Ωn(i)
− α(i) [h(xn(i)) − gn(zn(i))]
based on the information,
xn(i),{q (h(xl(i)) + νnl(i))}l∈Ωn(i) , zn(i),
available to it at time i (we assume that at time i sensor l sends a quantized version of h(xl(i)) + νnl(i) to sensor
n.) Here h−1(·) denotes the inverse of the function h(·) and {β(i)}i≥0 ,{α(i)}i≥0 are appropriately chosen weight
sequences. In the sequel, we analyze the NLU algorithm under the model Assumptions (D.1)-(D.3), and in addition
we assume:
(D.4): There exists 1 > 0, such that, ∀θ ∈ U, the following moment exists:
(cid:104)(cid:107)J(z(i)) − P J(z(i))(cid:107)2+1(cid:105)
Eθ
= κ(θ) < ∞,
where J(z(i)) is defined in (99) and the matrix P is given by
(114)
(115)
(116)
(117)
(118)
(119)
(120)
P =
(1N ⊗ IM ) (1N ⊗ IM )T
(cid:0)1N 1T
N
(cid:1) ⊗ IM
1
N
1
N
=
= PN ⊗ IM
The above moment condition is slightly stronger than the moment assumption required by the NU algorithm in
(95), where only existence of the quadratic moment of the random variables gn(zn(i)) was assumed. For example,
for the models considered in Propositions 13-14, the above condition coincides with the existence of slightly higher
than quadratic moments of the observations zn(i). The latter is clearly justified for any reasonable observation
noise distribution.
We also define, ∀θ ∈ U:
(cid:104)(cid:107)J(z(i)) − P J(z(i))(cid:107)2(cid:105)
(D.5): The weight sequences {α(i)}i≥0, and {β(i)}i≥0 are given by
Eθ [(cid:107)J(z(i)) − P J(z(i))(cid:107)] = κ1(θ) < ∞
= κ2(θ) < ∞.
Eθ
(i + 1)τ1
where a, b > 0 are constants. We assume the following:
.5 < τ1, τ2 ≤ 1, τ1 >
α(i) =
, β(i) =
a
b
(i + 1)τ2
1
2 + 1
+ τ2, 2τ2 > τ1
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
24
We note that, under Assumption (D.4), 1 > 0, such weight sequences always exist. As an example, if
then the choice τ1 = 1 and τ2 = .505 satisfies the inequalities in (120).
To write the NLU in a more compact form, introduce the transformed state sequence, {(cid:101)x(i)}i≥0, where (cid:101)x(i) =
[(cid:101)xT
1 (i)···(cid:101)xT
N (i)]T ∈ RNM and the iterations are
= .49,
2+1
1
(cid:101)x(i + 1) =(cid:101)x(i) − β(i) (L(i) ⊗ IM )(cid:101)x(i)−
−α(i) [(cid:101)x(i) − J(z(i))] − β(i) (Υ(i) + Ψ(i))
(cid:104)(cid:0)h−1((cid:101)x1(i))(cid:1)T ···(cid:0)h−1((cid:101)xN (i))(cid:1)T(cid:105)T
(121)
x(i) =
(122)
Here Υ(i), Ψ(i) ∈ RNM model the dithered quantization error effects as in algorithm NU resulting from the
quantized transmissions in (114). The update model in (121) is a mixed time-scale procedure, where the consensus
time-scale is determined by the weight sequence {β(i)}i≥0. On the other hand, the observation update time-scale
is governed by the weight sequence {α(i)}i≥0. It follows from Assumption (D.5) that τ1 > τ2, which in turn
α(i) → ∞ as i → ∞. Thus, the consensus time-scale dominates the observation update time-scale as
implies, β(i)
the algorithm progresses making it a mixed time-scale algorithm that does not directly fall under the purview of
stochastic approximation results like Theorem 29. Also, the presence of the random link failures and quantization
noise (which operate at the same time-scale as the consensus update) precludes standard approaches like time-scale
separation for the limiting system.
Remark 20 We comment on the distributed implementation of the NLU. Based on (114) and (121)-(122), the
NLU may be implemented either in the estimate domain with {x(i)} as the algorithm state sequence or in the
require the sensors to store and transmit the instantaneous estimate xn(i), however, implementation of the update
involves computation of the functions h(xn(i)) followed by an inverse (h−1(·)) at every step. On the other hand,
the implementation in the transformed domain, (121)-(122), requires the sensors to store and transmit the states
transformed domain with {(cid:101)x(i)} as the state sequence. The implementation in the estimate domain, (114), would
(cid:101)xn(i). The advantage in the latter implementation form is that the transformed state update rule, (121), is linear in
the state(cid:101)x(i) and, in particular, does not require function computations and inverses at each step. (We note that the
computations, the NLU may be implemented in the transformed domain with(cid:101)xn(i) being the state at a sensor n.
inverse in (122) may not be implemented at all iterations and does not affect the propagation of the transformed
state sequence. In fact, (122) may be implemented only once to obtain the actual estimates from the transformed
state sequence when the latter converge based on a suitable stopping criterion.) Hence, in practice, to simplify
B. Algorithm NLU: Discussions and Main Results
We comment on the NLU algorithm. As is clear from (121)-(122), the NLU algorithm operates in a transformed
domain. As a matter of fact, the function h(·) (c.f. definition 11) can be viewed as an invertible transformation on
the parameter space U. The transformed state sequence, {(cid:101)x(i)}i≥0, is then a transformation of the estimate sequence
{x(i)}i≥0, and, as seen from (121), the evolution of the sequence {(cid:101)x(i)}i≥0 is linear. This is an important feature
of the NLU algorithm, which is linear in the transformed domain, although the underlying observation model is
nonlinear. Intuitively, this approach can be thought of as a distributed stochastic version of homomorphic filtering
(see [53]), where, by suitably transforming the state space, linear filtering is performed on a certain non-linear
problem of filtering. In our case, for models of the separably estimable type, the function h(·) then plays the
role of the analogous transformation in homomorphic filtering, and, in this transformed space, one can design
linear estimation algorithms with desirable properties. This makes the NLU algorithm significantly different from
algorithm NU, with the latter operating on the untransformed space and is non-linear. This linear property of the
NLU algorithm in the transformed domain leads to nice statistical properties (for example, consistency asymptotic
unbiasedness) under much weaker assumptions on the observation model than required by the nonlinear NU
algorithm, but not asymptotic normality.
We now state the main results about the NLU algorithm developed in the paper. We show that, if the observation
model is separably estimable, then, in the transformed domain, the NLU algorithm is consistent. More specifically,
mean-squared sense to h(θ∗). We note that, unlike the NU algorithm, this only requires the observation model to
be separably estimable and no other conditions on the functions hn(·), h(·). We summarize these in the following
theorem.
if θ∗ is the true (but unknown) parameter value, then the transformed sequence {(cid:101)x(i)}i≥0 converges a.s. and in
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Theorem 21 Consider the NLU algorithm under the Assumptions (D.1)-(D.5), and the sequence {(cid:101)x(i)}i≥0 gener-
ated according to (121). We then have
25
(cid:104)
Pθ∗
lim
i→∞
lim
i→∞
Eθ∗
(cid:105)
(cid:101)xn(i) = h(θ∗), ∀1 ≤ n ≤ N
(cid:104)(cid:107)(cid:101)xn(i) − h(θ∗)(cid:107)2(cid:105)
Eθ∗ [(cid:101)xn(i)] = h(θ∗), ∀1 ≤ n ≤ N
= 0, ∀1 ≤ n ≤ N
= 1
(123)
(124)
(cid:104)
Pθ∗
lim
i→∞
lim
i→∞
In particular,
In other words, in the transformed domain, the estimate sequence {(cid:101)xn(i)}i≥0 at sensor n, is consistent, asymptot-
lim
i→∞
(125)
ically unbiased and converges in mean-squared sense to h(θ∗).
As an immediate consequence of Theorem 21, we have the following result, which characterizes the statistical
properties of the untransformed state sequence {x(i)}i≥0.
Theorem 22 Consider the NLU algorithm under the Assumptions (D.1)-(D.5). Let {x(i)}i≥0 be the state sequence
generated, as given by (121)-(122). We then have
(cid:105)
xn(i) = θ∗, ∀ 1 ≤ n ≤ N
= 1
(126)
In other words, the NLU algorithm is consistent.
If in addition, the function h−1(·) is Lipschitz continuous, the NLU algorithm is asymptotically unbiased, i.e.,
(127)
Eθ∗ [xn(i)] = θ∗, ∀ 1 ≤ n ≤ N
The next subsection is concerned with the proofs of Theorems 21, 22.
C. Consistency and Asymptotic Unbiasedness of NLU: Proofs of Theorems 21,22
The present subsection is devoted to proving the consistency and unbiasedness of the NLU algorithm under
the stated Assumptions. The proof is lengthy and we start by explaining why standard stochastic approximation
results like Theorem 29 do not apply directly. A careful inspection shows that there are essentially two different
time-scales embedded in (121). The consensus time-scale is determined by the weight sequence {β(i)}i≥0, whereas
the observation update time-scale is governed by the weight sequence {α(i)}i≥0. It follows from Assumption (D.5)
α(i) → ∞ as i → ∞. Thus, the consensus time-scale dominates the observation
that τ1 > τ2, which, in turn, implies β(i)
update time-scale as the algorithm progresses making it a mixed time-scale algorithm that does not directly fall
under the purview of stochastic approximation results like Theorem 29. Also, the presence of the random link
failures and quantization noise (which operate at the same time-scale as the consensus update) precludes standard
approaches like time-scale separation for the limiting system.
Finally, we note that standard stochastic approximation assumes that the state evolution follows a stable deter-
ministic system perturbed by zero-mean stochastic noise. More specifically, if {y(i)}i≥0 is the sequence of interest,
Theorem 29 assumes that {y(i)}i≥0 evolves as
(128)
where {γ(i)}i≥0 is the weight sequence, Γ(i + 1, ω, y(i)) is the zero-mean noise. If the sequence {y(i)}i≥0 is
supposed to converge to y0, it further assumes that R(y0) = 0 and y0 is a stable equilibrium of the deterministic
system
y(i + 1)=y(i) + γ(i)[R(y(i))+Γ(i + 1, ω, y(i))]
yd(i + 1) = yd(i) + γ(i)R(yd(i))
(129)
The NU algorithm (and its linear version, LU) falls under the purview of this, and we can establish convergence
properties using standard stochastic approximation (see Sections II,III-B.) However, the NLU algorithm cannot
be represented in the form of (128), even ignoring the presence of multiple time-scales. Indeed, as established by
stochastically perturbed system around 1N ⊗ h(θ∗) we have
Theorem 21, the sequence {(cid:101)x(i)}i≥0 is supposed to converge to 1N ⊗ h(θ∗) a.s. and hence writing (121) as a
(cid:101)x(i + 1) =(cid:101)x(i) + γ(i) [R((cid:101)x(i)) + Γ(i + 1, ω,(cid:101)x(i))]
(130)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
26
where,
R((cid:101)x(i)) = −β(i)(cid:0)L ⊗ IM
Γ(i + 1, ω,(cid:101)x(i)) = −β(i)
(cid:1) ((cid:101)x(i) − 1N ⊗ h(θ∗))−
− α(i) ((cid:101)x(i) − 1N ⊗ h(θ∗))
(cid:16)(cid:101)L(i) ⊗ IM
((cid:101)x(i)−
(cid:17)
−1N ⊗ h(θ∗)) − β(i) (Υ(i) + Ψ(i)) +
+ α(i) (J(z(i)) − 1N ⊗ h(θ∗))
Although, R(1N ⊗ h(θ∗)) = 0 in the above decomposition, the noise Γ(i + 1, ω,(cid:101)x(i)) is not unbiased as the term
(J(z(i)) − 1N ⊗ h(θ∗)) is not zero-mean.
With the above discussion in mind, we proceed to the proof of Theorems 21,22, which we develop in stages.
(131)
(132)
(cid:1)(cid:98)x(i)−
(cid:1)(cid:101)x◦(i)−
The detailed proofs of the intermediate results are provided in Appendix F.
− α(i) [(cid:101)x◦(i) − J(z(i))]
quantization noises Υ(i) and Ψ(i) are not included. In other words, in the absence of link failures and quantization,
(cid:101)x◦(i + 1) =(cid:101)x◦(i) − β(i)(cid:0)L ⊗ IM
sequence, {(cid:101)x◦(i)}i≥0:
In parallel to the evolution of the state sequence {x(i)}i≥0, we consider the following update of the auxiliary
with (cid:101)x◦(0) = (cid:101)x(0). Note that in (131) the random Laplacian L is replaced by the average Laplacian L and the
the recursion (121) reduces to (131), i.e., the sequences {(cid:101)x(i)}i≥0 and {(cid:101)x◦(i)}i≥0 are the same.
sequence {(cid:98)x(i)}i≥0 given by (cid:98)x(i + 1) =(cid:98)x(i) − β(i)(cid:0)L ⊗ IM
with (cid:98)x(0) = (cid:101)x(0). Like before, the recursions (121,122) will reduce to (132) when there are no link failures.
quantizing (cid:101)x(i) in (121) and not from quantizing (cid:98)x(i) in (132).
Now consider the sequence whose recursion adds as input to the recursion in (131) the quantization noises Υ(i)
and Ψ(i). In other words, in the absence of link failures, but with quantization included, define similarly the
− α(i) [(cid:98)x(i) − J(z(i))] − β(i) (Υ(i) + Ψ(i))
However, notice that in (132) the quantization noise sequences Υ(i) and Ψ(i) are the sequences resulting from
(1N ⊗ IM )T(cid:101)x(i)
Define the instantaneous averages over the network as
N(cid:88)
N(cid:88)
N(cid:88)
N(cid:88)
we first show that the average sequence, (cid:8)(cid:101)x◦avg(i)(cid:9)
standard stochastic approximation arguments. Then we show that the sequence {(cid:101)x◦(i)}i≥0 reaches consensus a.s.,
that the sequence {(cid:101)x(i)}i≥0 also reaches consensus a.s. with h(θ∗) as the limiting consensus value.
(I): The first step consists of studying the convergence properties of the sequence (cid:8)(cid:101)x◦avg(i)(cid:9)
and clearly the limiting consensus value must be h(θ∗). Intuitively, the a.s. consensus comes from the fact that,
after a sufficiently large number of iterations, the consensus effect dominates over the observation update effect,
thus asymptotically leading to consensus. The final step in the proof uses a series of comparison arguments to show
We sketch the main steps of the proof here. While proving consistency and mean-squared sense convergence,
, converges a.s. to h(θ∗). This can be done by invoking
We now detail the proofs of Theorems 21,22 in the following steps.
(1N ⊗ IM )T(cid:101)x◦(i)
x◦avg(i) =
(cid:101)x◦avg(i) =
x◦n(i) =
(cid:101)x◦n(i) =
xavg(i) =
(cid:101)xavg(i) =
xn(i) =
(cid:101)xn(i) =
(1N ⊗ IM )T x◦(i)
(1N ⊗ IM )T x(i)
1
N
1
N
1
N
1
N
, see (131), for
1
N
1
N
1
N
1
N
(133)
n=1
n=1
n=1
n=1
i≥0
i≥0
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
which we establish the following result.
Lemma 23 Consider the sequence, {(cid:101)x◦(i)}i≥0, given by (131), under the Assumptions (D.1)-(D.5). Then,
(cid:104)
(cid:105)
(cid:101)x◦(i) = 1N ⊗ h(θ∗)
(cid:104)(cid:107)(cid:101)x◦(i) − 1N ⊗ h(θ∗)(cid:107)2(cid:105)
lim
i→∞
Pθ∗
Eθ∗
= 1
lim
i→∞
Lemma 23 says that the sequence {(cid:101)x◦(i)}i≥0 converges a.s. and in L2 to 1N ⊗ h(θ∗). For proving Lemma 23
we first consider the corresponding average sequence {(cid:101)x◦avg(i)}i≥0, see (133). For the sequence {(cid:101)x◦avg(i)}i≥0, we
Lemma 24 Consider the sequence,(cid:8)(cid:101)x◦avg(i)(cid:9)
(cid:104)
can invoke stochastic approximation algorithms to prove that it converges a.s. and in L2 to h(θ∗). This is carried
out in Lemma 24, which we state now.
, given by (133), under the Assumptions (D.1)-(D.5). Then,
= 0
i≥0
(135)
= 1
(136)
(cid:105)
(cid:101)x◦avg(i) = h(θ∗)
(cid:104)(cid:13)(cid:13)(cid:101)x◦avg(i) − h(θ∗)(cid:13)(cid:13)2(cid:105)
lim
i→∞
Pθ∗
Eθ∗
lim
i→∞
= 0
(137)
The arguments in Lemmas 24,23 and subsequent results require the following property of real number sequences,
which we state here (see Appendix C for proof.)
Lemma 25 Let the sequences {r1(i)}i≥0 and {r2(i)}i≥0 be given by
(138)
where a1, a2, δ2 ≥ 0 and 0 ≤ δ1 ≤ 1. Then, if δ1 = δ2, there exists B > 0, such that, for sufficiently large
non-negative integers, j < i,
, r2(i) =
(i + 1)δ1
(i + 1)δ2
r1(i) =
a1
a2
27
(134)
(139)
(140)
(cid:34)(cid:32) i−1(cid:89)
0 ≤ i−1(cid:88)
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
k=j
l=k+1
(1 − r1(l))
r2(k)
≤ B
(cid:33)
(cid:33)
(cid:35)
(cid:35)
Moreover, the constant B can be chosen independently of i, j. Also, if δ1 < δ2, then, for arbitrary fixed j,
(We use the convention that,(cid:81)i−1
lim
i→∞
k=j
l=k+1
l=k+1 (1 − rl) = 1, for k = i − 1.)
(1 − r1(l))
r2(k)
= 0
We note that Lemma 25 essentially studies stability of time-varying deterministic scalar recursions of the form:
y(i + 1) = r1(i)y(i) + r2(i)
(141)
where {y(i)}i≥0 is a scalar sequence evolving according to (141) with y(0) = 0, and the sequences {r1(i)}i≥0
and {r2(i)}i≥0 are given by (138).
(II): In this step, we study the convergence properties of the sequence {(cid:98)x(i)}i≥0, see (132), for which we establish
Lemma 26 Consider the sequence {(cid:98)x(i)}i≥0 given by (132) under the Assumptions (D.1)-(D.5). We have
the following result.
(cid:104)
(cid:105)
(cid:98)x(i) = 1N ⊗ h(θ∗)
(cid:104)(cid:107)(cid:98)x(i) − 1N ⊗ h(θ∗)(cid:107)2(cid:105)
lim
i→∞
Pθ∗
Eθ∗
= 1
= 0
(142)
(143)
The proof of Lemma 26 is given in Appendix E, and mainly consists of a comparison argument involving the
sequences(cid:8)(cid:101)x◦avg(i)(cid:9)
lim
i→∞
and {(cid:98)x(i)}i≥0.
i≥0
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
28
(III): This is the final step in the proofs of Theorems 21,22. The proof of Theorem 21 consists of a comparison
argument between the sequences {(cid:98)x(i)}i≥0 and {(cid:101)x(i)}i≥0, which is detailed in Appendix F. The proof of
Theorem 22, also detailed in Appendix F, is a consequence of Theorem 21 and the Assumptions.
D. Application: Distributed Static Phase Estimation in Smart Grids
In this subsection we show that our development of the NLU for separably estimable observation models
leads to a completely distributed solution of the static phase estimation problem in smart grids. We briefly review
the application scenario in the following, for a more complete treatment of the classical problem of static phase
estimation in power grids the reader is referred to one of the many existing excellent textbooks, for example, [54].
For our purpose, we may assume the power grid to be a physical network of N generators and loads (hereafter
called nodes), interconnected through transmission lines. The physical grid may then be modeled as a network
Gp = (V, Ep), where10 Ep denotes the set of transmission lines or interconnections. The physical state of a node
n consists of the pair (Vn, θn), denoting the voltage magnitude and the phase angle respectively. The real power
flowing through the transmission line connecting nodes n and l is then given by (see [55])
Pnl =V 2
nanl−VnVlanl cos(θnl) + VnVlbnl sin(θnl)
(144)
where anl + jbnl is the complex line admittance and θnl = θn − θl. In view of the physical network structure, the
following assumptions on the physical grid are supposed to hold:
(P.1) The physical grid, represented by the graph (V, Ep), is connected. This is reasonable, as the physical power
grid is often studied with aggregated models that have dense interconnections, see, for example, the benchmark
IEEE 30 bus and IEEE 118 bus systems ([55]).
(P.2) The real and imaginary parts, anl, bnl, of the line admittance are taken to be zero, if no direct physical
connection (link) exists between the nodes n and l. Similarly, if nodes n and l are connected by a transmission
line, we assume both the components anl, bnl, of the line admittance to be non-zero. In particular, from (144),
the real power flow Pnl between nodes n and l is non-zero iff
there exists a physical transmission line
connecting the nodes. Also, Vn (cid:54)= 0 for all n.
In the following, we will assume that the node voltages are known constants and the unknown parameter of interest
is the vector of node phases θ = [θ1,··· , θN ]T . This is justified by the common assumption of phase-voltage
decoupling in power grids, where the voltage magnitude is generally seen to fluctuate at a much slower time-scale
than the phase (see [54]). Also, to keep the exposition simple, we assume that the node phase differences are small,
i.e., cos(θnl) ≈ 1, commonly used in the steady state grid operating regime ([54]). With these simplifications, the
real power flow in a transmission line connecting nodes n and l is approximately given by
Pnl(θ) = VnVlbnl sin(θnl)
(145)
The goal of centralized static phase estimation is to estimate the unknown vector θ of phases by using line flow data.
Since, only relative quantities (phase differences) are involved in this problem, it is customary to assume (see [55])
that one of the nodes is a slack (or reference) bus, whose phase is a known constant. W.l.o.g. we assume that node
N is the slack bus in our system, whose phase angle θN is a known constant. Hence, the effective parameter vector
is θ = [θ1,··· , θN−1]T ∈ RN−1. We now provide conditions for the distributed observation model (148) to be
separably estimable. We show that our NLU algorithm can be used to obtain a distributed solution to this problem,
leading to a consistent estimate of θ at each sensor. To setup the distributed observation model, let Em ⊂ Ep denote
the set of physical transmission lines equipped with power flow measuring devices (usually some form of relays,
see [55].) The successive power flow measurements, {znl(i)} at the physical line (n, l) (assuming (n, l) ∈ Em) are
then noisy versions of the power flow Pnl, i.e.,
znl(i) = Pnl(θ) + ζnl(i)
= VnVlbnl sin(θnl) + ζnl(i)
(146)
10Note that the physical connections Ep and the inter-node communication links E (to be used by the distributed information processing
algorithms) are, in general, different. This is a common feature of cyberphysical architectures, where a sensor network is instrumented on
top of an existing physical infrastructure. In the following, we will assume that each physical node is equipped with a sensor for information
processing, although the inter-sensor communication topology may be different from the physical inter-node connections. Also, the terms nodes
and sensors will be used synonymously in the sequel.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
29
where, {ζnl(i)} is the zero mean i.i.d. measurement noise. For distributed information processing, we assume that
the measurement sequence {znl(i)} is forwarded to one of the adjacent nodes n or l. As will be seen, the particular
choice of n or l is not important, as long as it stays constant for all i. In general, denoting by Ωm
n to be the set of
physical neighbors of node n w.r.t. the physical graph (V, Em), let Mn ⊂ Ωm
n denote the set such that l ∈ Mn if
the line flow data {znl(i)} is forwarded to node n. The observation sequence, {zn(i)}, is then
zn(i) = {znl(i), ∀ l ∈ Mn}
(147)
By (146)-(147) and the development in Section III-A, the above distributed observation model corresponds to the
signal in additive noise type. Following the notation in Section III-A, the observation process {zn(i)} may be
written as
(148)
where fn : U (cid:55)−→ RMn and ζn(i) are given by
zn(i) = fn(θ) + ζn(i)
fn(θ) = [VnVlbnl sin(θnl), l ∈ Mn]T
ζn(i) = [ζnl, l ∈ Mn]T .
(149)
f (θ)=[f T
n=1 Mn by
1 (θ)··· f T
N (θ)]T=[Pnl(θ), (n, l)∈ Em]T.
n=1 [0, π/2) or ×N−1
We now provide conditions for the distributed observation model (148) to be separably estimable.
Proposition 27 Depending on the phase θN of the reference bus N (as to whether θN ∈ [0, π/2) or θN ∈
f : U (cid:55)−→ R(cid:80)N
[−π/4, π/4)), let the parameter domain U ⊂ RN−1 be either ×N−1
n=1 [−π/4, π/4). Define, ∀θ ∈ U,
(150)
Then, if the graph (V, Em) (with Em as the edge set) is connected and assumption (P.2) holds, f (·) is invertible
on U and the observation model (148) is separably estimable.
Before proceeding to the proof we comment on Proposition 27. The observation model in (148) is of the signal
in additive noise type and hence, by the development in Section III-A, the invertibility of f (·) is equivalent to
separable estimability and necessary for the consistency or observability of the centralized estimator also (see the
text following Proposition 13.) Proposition 27 shows that the invertibility of f (·) holds if the graph formed by the
physical transmission links equipped with power flow measuring devices is connected.
Proof: The proof is based on a simple inductive argument. First, we note that the continuity of f (·) on U holds
trivially. To establish the invertibility of f (·) on U, it suffices to show that we can uniquely recover the value of
θ ∈ U given the value f (θ). To this end, assume that f (θ) is given. Recall the form of f (θ), as in (150). Since, θN
is known, given f (θ), the components θn, n ∈ Ωm
N may be uniquely determined. Indeed, knowing f (θ) amounts
N . Hence, under assumption (P.2), and the fact that θ ∈ U, we
to knowing the values of the quantities Pn,N , n ∈ Ωm
can uniquely determine θn, n ∈ Ωm
N . Once the components
n for each n ∈ J1 may be uniquely
θn, n ∈ J1 are known, by using similar reasoning, the components θl, l ∈ Ωm
determined. Hence, in the second step, the set of known components J2 ⊂ V is
N . To continue the induction, define J1 ⊂ V by J1 = Ωm
Continuing the same recursion, the set of components known at the k-th step, Jk is given by
J2 = {l ∈ Ωm
n n ∈ J1}
Jk = {l ∈ Ωm
n n ∈ Jk−1}
Note that J1 ⊂ J2 ⊂ ··· ⊂ Jk ⊂ ··· . However, the number of nodes is finite; hence, the sets Jk cannot increase
forever. Due to the connectivity of the graph (V, Em) in a finite number of steps k0 (at most equal to the diameter
of the graph), the process will converge with Jk0 = V . Hence, all components of θ will be uniquely determined,
establishing the invertibility of f (·).
The following result demonstrates the applicability of the NLU to distributed consistent phase estimation in
power grids. It follows from Theorem 22.
Theorem 28 Consider the power grid described above and let the observation process {zn(i)} at the n-th node
(sensor) be given by (148). Let the measurement noise process {ζ(i)} satisfy the moment conditions (D.4) and
the physical grid conditions in the hypothesis of Proposition 27 hold. Suppose the nodes are instrumented with a
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
30
communication architecture satisfying assumptions (D.2)-(D.3). Let {xn(i)} be the estimate sequence of the vector
θ∗ of phase angles generated at node n by an instantiation of the NLU distributed parameter estimation algorithm
under assumption (D.5). Then,
Pθ∗ (xn(i) = θ∗, ∀n) = 1, ∀ θ∗ ∈ U
(151)
V. CONCLUSION
This paper studies linear and nonlinear distributed (vector) parameter estimation problems as may arise in
constrained sensor networks. Our problem statement is quite general, including communication among sensors that
is quantized, noisy, and with channels that fail at random times. These are characteristic of packet communication
in wireless sensor networks. We introduce a generic observability condition, the separable estimability condition,
that generalizes to distributed estimation the general observability condition of centralized parameter estimation.
We study three recursive distributed estimators, ALU, NU, and NLU. We study their asymptotic properties,
namely: consistency, asymptotic unbiasedness, and for the ALU and NU algorithms their asymptotic normality. The
NLU works in a transformed domain where the recursion is actually linear, and a final nonlinear transformation,
justified by the separable estimability condition, recovers the parameter estimate (a stochastic generalization of
homeomorphic filtering.) For example, Theorem 21 shows that, in the transformed domain, the NLU leads to
consistent and asymptotically unbiased estimators at every sensor for all separably estimable observation models
satisfying (D.4)11. Since, the function h(·) is invertible, for practical purposes, a knowledge of h(θ∗) is sufficient for
knowing θ∗. In that respect, the algorithm NLU is much more applicable than the algorithm NU, which requires
further assumptions on the observation model for the existence of consistent and asymptotically unbiased estimators.
However, in case, the algorithm NU is applicable, it provides convergence rate guarantees (for example, asymptotic
normality) that follow from standard stochastic approximation theory. On the other hand, the algorithm NLU does
not fall under the purview of standard stochastic approximation theory (see Section IV-C) and hence does not inherit
these convergence rate properties. In this paper, we presented a convergence theory of the three algorithms under
broad conditions. An interesting future research direction is to establish a convergence rate theory for the NLU
algorithm (and in general, distributed stochastic algorithms of this form, which involve mixed time-scale behavior
and biased perturbations.)
APPENDIX A
SOME RESULTS ON STOCHASTIC APPROXIMATION
We present some classical results on stochastic approximation from [56] regarding the convergence properties of generic
stochastic recursive procedures, which will be used to characterize the convergence properties (consistency, convergence rate)
of the LU algorithm.
Theorem 29 Let(cid:8)x(i) ∈ Rl(cid:9)
i≥0 be a random sequence:
x(i + 1) = x(i) + α(i) [R(x(i)) + Γ (i + 1, x(i), ω)]
(152)
where, R(·) : Rl (cid:55)−→ Rl is Borel measurable and {Γ(i, x, ω)}i≥0, x∈Rl is a family of random vectors in Rl, defined on a
probability space (Ω,F,P), and ω ∈ Ω is a canonical element. Let the following sets of assumptions hold:
(B.1): The function Γ(i,·,·) : Rl × Ω −→ Rl is Bl ⊗ F measurable for every i; Bl is the Borel algebra of Rl.
(B.2): There exists a filtration {Fi}i≥0 of F, such that, for each i, the family of random vectors {Γ (i, x, ω)}x∈Rl is Fi
measurable, zero-mean and independent of Fi−1.
(If Assumptions (B.1)-(B.2) hold, {x(i)}i≥0, is Markov.)
(B.3): There exists a function V (x) ∈ C2 with bounded second order partial derivatives and a point x∗ ∈ Rl satisfying:
∗
) = 0, V (x) > 0, x (cid:54)= x
∗
,
V (x
lim(cid:107)x(cid:107)→∞ V (x) = ∞,
(R (x) , Vx (x)) < 0, ∀ > 0
sup
<(cid:107)x−x∗(cid:107)< 1
11The NLU requires a slightly stronger moment condition, Assumption (D.4). However, for reasonable observation noise statistics arising in
practice, this assumption is justified.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
31
where Vx (x) denotes the gradient (vector) of V (·) at x.
(B.4): There exist constants k1, k2 > 0, such that,
(cid:107)R (x)(cid:107)2 + E(cid:2)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:3) ≤ k1 (1 + V (x))−
(B.5): The weight sequence {α(i)}i≥0 satisfies
α(i) > 0,
(C.1): The function R (x) admits the representation
− k2 (R (x) , Vx (x))
(cid:88)
i≥0
αi = ∞,
(cid:88)
i≥0
α2(i) < ∞
R (x) = B (x − x
∗
) + δ (x)
where
lim
x→x∗
(Note, in particular, if δ (x) ≡ 0, then (155) is satisfied.)
(C.2): The weight sequence, {α(i)}i≥0 is of the form,
(cid:107)δ (x)(cid:107)
(cid:107)x − x∗(cid:107) = 0
α(i) =
, ∀i ≥ 0
a
i + 1
(153)
(154)
(155)
(156)
where a > 0 is a constant (note that (C.2) implies (B.5)).
(C.3): Let I be the l × l identity matrix and a, B as in (156) and (154), respectively. Then, the matrix Σ = aB + 1
(C.4): The entries of the matrices, ∀i ≥ 0, x ∈ Rl,
2 I is stable.
A (i, x) = E(cid:104)
Γ (i + 1, x, ω) ΓT (i + 1, x, ω)
,
(cid:105)
R→∞ sup
lim
(cid:107)x−x∗(cid:107)<
sup
i≥0
(cid:107)Γ (i + 1, x, ω)(cid:107)2 dP = 0
(cid:107)Γ(i+1,x,ω)(cid:107)>R
(157)
Let Assumptions (B.1)-(B.5) hold for {x(i)}i≥0 in (152). Then, starting from an arbitrary initial state, the Markov process,
are finite, and the following limit exists:
(C.5): There exists > 0, such that
Then we have the following:
{x(i)}i≥0, converges a.s. to x∗. In other words,
(cid:110)√
(cid:111)
The normalized process,
i≥0
(C.5) are also satisfied. In particular, as i → ∞
i (x(i) − x∗)
i→∞, x→x∗ A (i, x) = S0
lim
(cid:90)
P(cid:104)
∗(cid:105)
√
i (x(i) − x
∗
) =⇒ N (0, S)
lim
i→∞ x(i) = x
= 1
(158)
, is asymptotically normal if, besides Assumptions (B.1)-(B.5), Assumptions (C.1)-
where =⇒ denotes convergence in distribution (weak convergence.) Also, the asymptotic variance, S, in (159) is
(cid:90) ∞
S = a2
eΣvS0eΣT vdv
Proof: For a proof see [56] (c.f. Theorems 4.4.4, 6.6.1).APPENDIX B
0
PROOFS OF THEOREMS 18,19
Proof of Theorem 18
(159)
(160)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
32
Proof: We noted the recursive scheme (97) satisfies Assumptions (B.1)-(B.2) of Theorem 29. To prove consistency, we
verify Assumptions (B.3)-(B.4). Let the Lyapunov function
V (x) = (cid:107)x − 1N ⊗ θ
∗(cid:107)2
Clearly,
V (1N ⊗ θ
∗
) = 0, V (x) > 0, x (cid:54)= 1N ⊗ θ
∗
, lim(cid:107)x(cid:107)→∞ V (x) = ∞
By Assumptions (107)-(108), h(·) is Lipschitz continuous and
(cid:16)
θ −(cid:101)θ
(cid:17)T(cid:16)
(cid:17)
> 0, ∀ θ (cid:54)=(cid:101)θ ∈ RM
where (162) follows from the invertibility of h(·) and
h(θ) − h((cid:101)θ)
N(cid:88)
1
N
n=1
h (θ) =
hn (θ) , ∀ θ ∈ RM
Recall R (x) , Γ (i + 1, x, ω) in (102)-(103). Then
(R (x) , Vx (x)) =
− 2β (x − 1N ⊗ θ
− 2 (x − 1N ⊗ θ
∗
)T(cid:0)L ⊗ IM
= −2β (x − 1N ⊗ θ
∗
N(cid:88)
(cid:104)
− 2
(xn − θ
∗
∗
)
)T(cid:0)L ⊗ IM
(cid:1) (x − 1N ⊗ θ
∗
)T [M (x) − M (1N ⊗ θ
)−
(cid:1) (x − 1N ⊗ θ
(cid:105) ≤ 0
)T (hn(xn) − hn(θ
))
)]
∗
∗
∗
where the last step follows from the positive-semidefiniteness of L ⊗ IM and (108). To verify Assumption (B.3), show
n=1
Assume, on the contrary, (165) not satisfied. Then from (164)
sup
<(cid:107)x−1N θ∗(cid:107)< 1
(R (x) , Vx (x)) < 0, ∀ > 0
, such that
(cid:17)
k≥0 has a limit point,(cid:98)x, such that ≤ (cid:107)(cid:101)x−1N θ∗(cid:107) ≤ 1
= 0
(167)
,
(161)
(162)
(163)
(164)
(165)
(166)
(168)
(169)
(170)
(171)
(172)
Then, there exists a sequence,(cid:8)xk(cid:9)
Since(cid:8)x ∈ RNM < (cid:107)x − 1N θ∗(cid:107) < 1
k≥0 in
and, by continuity of (R (x) , Vx (x)):
From (108) and (164), we then have
(R (x) , Vx (x)) = 0, ∀ > 0
(cid:111)
<(cid:107)x−1N θ∗(cid:107)< 1
sup
(cid:16)
lim
k→∞
R(xk), Vx(xk)
x ∈ RNM(cid:12)(cid:12)(cid:12) < (cid:107)x − 1N θ∗(cid:107) < 1
(cid:110)
(cid:9) is relatively compact,(cid:8)xk(cid:9)
(R((cid:101)x), Vx((cid:101)x)) = 0
)T(cid:0)L ⊗ IM
(cid:1) ((cid:101)x − 1N ⊗ θ
)T (hn((cid:101)xn) − hn(θ
((cid:101)xn − θ
(cid:101)xn = a, ∀n
∗
∗
∗
((cid:101)x − 1N ⊗ θ
∗
) = 0,
)) = 0, ∀n
The inequalities in (170) then imply
(a − θ
∗
)T (h(a) − h(θ
∗
)) = 0
The equality in (169) and the properties of the Laplacian imply that (cid:101)x ∈ C and hence there exists a ∈ RM , such that,
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
33
which is a contradiction by (162) since a (cid:54)= θ∗. Thus, we have (165) that verifies Assumption (B.3). Finally, we note that,
(cid:107)R (x)(cid:107)2 =(cid:13)(cid:13)β(cid:0)L ⊗ IM
≤ 4β2(cid:13)(cid:13)(cid:0)L ⊗ IM
(cid:1) (x − 1N ⊗ θ
(cid:1) (x − 1N ⊗ θ
)(cid:13)(cid:13)2
+
∗
∗
) +
))(cid:13)(cid:13)2
∗
(173)
+ (M (x) − M (1N ⊗ θ
≤ 4β2λN (L)(cid:107)x − 1N ⊗ θ
+ 4(cid:107)M (x) − M (1N ⊗ θ
)(cid:107)2
∗
∗(cid:107)2 + 4K 2(cid:107)x − 1N ⊗ θ
∗(cid:107)2
where the second step follows from the Lipschitz continuity of hn(·) and K is defined in (109). To verify Assumption (B.4),
we have then along similar lines as in Theorem 7
(cid:107)R (x)(cid:107)2 + E(cid:2)(cid:107)Γ (i + 1, x, ω)(cid:107)2(cid:3) ≤ k1(1 + V (x))
(174)
≤ k1(1 + V (x)) − (R (x) , Vx (x))
for a constant k1 > 0 (the last step follows from (164).) Hence, the assumptions are satisfied and the claim follows.
Proof of Theorem 19
Proof: The recursive scheme in (97) satisfies Assumptions (B.1)-(B.2) of Theorem 29. To prove consistency, we verify
Assumptions (B.3)-(B.4). Consider the Lyapunov function
V (x) = (cid:107)x − 1N ⊗ θ
∗(cid:107)2
(175)
Clearly,
V (1N ⊗ θ
∗
) = 0, V (x) > 0, x (cid:54)= 1N ⊗ θ
∗
, lim(cid:107)x(cid:107)→∞V (x) = ∞
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Recall the definitions of R (x) , Γ (i + 1, x, ω) in (102)-(103), and the consensus subspace in (106). We then have
(R (x) , Vx (x)) =
)T(cid:0)L ⊗ IM
(cid:1) (x − 1N ⊗ θ
− 2β (x − 1N ⊗ θ
∗
∗
)−
− 2 (x − 1N ⊗ θ
∗
)T [M (x) − M (1N ⊗ θ
∗
)]
34
(176)
≤ −2βλ2(L)(cid:107)xC⊥(cid:107)2−
− 2 (x − 1N ⊗ θ
∗
≤ −2βλ2(L)(cid:107)xC⊥(cid:107)2+
− 2 (x − 1N ⊗ θ
∗
(cid:13)(cid:13)(cid:13)(x − 1N ⊗ θ
+ 2
− 2 (x − 1N ⊗ θ
∗
)T [M (x) − M (xC)]
∗
)T [M (xC) − M (1N ⊗ θ
)]
∗
)T [M (x) − M (xC)]
)T [M (xC) − M (1N ⊗ θ
∗
)]
≤ −2βλ2(L)(cid:107)xC⊥(cid:107)2 + 2K(cid:107)xC⊥(cid:107)(cid:107)x − 1N ⊗ θ
∗(cid:107)
− 2 (x − 1N ⊗ θ
∗
)T [M (xC) − M (1N ⊗ θ
∗(cid:107)−
∗
)]
= −2βλ2(L)(cid:107)xC⊥(cid:107)2 + 2K(cid:107)xC⊥(cid:107)(cid:107)x − 1N ⊗ θ
− 2xTC⊥ [M (xC) − M (1N ⊗ θ
)]
∗
− 2 (xC − 1N ⊗ θ
≤ −2βλ2(L)(cid:107)xC⊥(cid:107)2 + 2K(cid:107)xC⊥(cid:107)(cid:107)x − 1N ⊗ θ
∗
)T [M (xC) − M (1N ⊗ θ
∗(cid:107)+
∗
(cid:13)(cid:13)(cid:13)xTC⊥ [M (xC) − M (1N ⊗ θ
2
− 2 (xC − 1N ⊗ θ
∗
)T [M (xC) − M (1N ⊗ θ
∗(cid:107)+
∗(cid:107) − 2γ (cid:107)xC − 1N ⊗ θ
∗(cid:107)2
≤ −2βλ2(L)(cid:107)xC⊥(cid:107)2 + 2K(cid:107)xC⊥(cid:107)(cid:107)x − 1N ⊗ θ
+ 2K(cid:107)xC⊥(cid:107)(cid:107)xC − 1N ⊗ θ
=(cid:0)−2βλ2(L) + 2K(cid:1)(cid:107)xC⊥(cid:107)2+
(cid:13)(cid:13)(cid:13)
∗
)]
(cid:13)(cid:13)(cid:13)
)]
∗
)]
+ 4K(cid:107)xC⊥(cid:107)(cid:107)xC − 1N ⊗ θ
∗(cid:107) − 2γ (cid:107)xC − 1N ⊗ θ
∗(cid:107)2
where the second to last step is justified because xC = 1N ⊗(cid:101)y for some (cid:101)y ∈ RM and
)T [hn((cid:101)y) − hn(θ
)T [M (xC) − M (1N ⊗ θ
(xC − 1N ⊗ θ
((cid:101)y − θ
)] =
∗
∗
∗
N(cid:88)
N(cid:88)
[hn((cid:101)y) − hn(θ
)T [h((cid:101)y) − h(θ
n=1
n=1
∗
∗
= ((cid:101)y − θ
= N ((cid:101)y − θ
)T
∗
= γ (cid:107)xC − 1N ⊗ θ
∗(cid:107)2
∗
)]
)] ≥ N γ (cid:107)(cid:101)y − θ
∗(cid:107)2
(177)
∗
)]
It can be shown that, if β > K2+Kγ
γλ2L
, the term on the R.H.S. of (176) is always non-positive. We thus have
By the continuity of (R (x) , Vx (x)) and the relative compactness of
similar lines as in Theorem 18 that
(R (x) , Vx (x)) ≤ 0, ∀x ∈ RM N
x ∈ RNM(cid:12)(cid:12)(cid:12) < (cid:107)x − 1N θ∗(cid:107) < 1
(cid:110)
sup
<(cid:107)x−1N θ∗(cid:107)< 1
(R (x) , Vx (x)) < 0, ∀ > 0
(cid:111)
(178)
, we can show along
(179)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
35
verifying Assumption (B.3). Assumption (B.4) is verified in similar manner to Theorem 18 and the result follows.
(180)
(181)
(182)
(183)
(184)
APPENDIX C
PROOF OF LEMMA 25
Proof of Lemma 25: We prove for the case δ1 < 1 first. Consider j sufficiently large, such that,
Then, for k ≥ j, using 1 − a ≤ e−a, 0 ≤ a ≤ 1:
i−1(cid:89)
l=k+1
r1(i) ≤ 1, ∀i ≥ j
(1 − r1(l)) ≤ e
−(cid:80)i−1
l=k+1 r1(l)
It follows from the properties of the Riemann integral that
i−1(cid:88)
l=k+1
r1(l) =
l=k+1
a1
(l + 1)δ1
i−1(cid:88)
(cid:90) i+1
(cid:104)
(i + 1)1−δ1 − (k + 2)1−δ1(cid:105)
(cid:33)
1
tδ1
a1
1 − δ1
k+2
dt
≥ a1
=
(cid:35)
(k+2)1−δ1(cid:105)
≤
r2(l)
a2
(k + 1)δ2
(1 − r1(l))
(i+1)1−δ1
a1
1−δ1
e
i−1(cid:88)
(cid:20)
k=j
− a1
1−δ1
a2e
(i+1)1−δ1
a1
1−δ1
e
(k+2)1−δ1
1
(k + 1)δ2
We thus have from (181)-(182)
i−1(cid:88)
i−1(cid:88)
k=j
k=j
(cid:34)(cid:32) i−1(cid:89)
(cid:104)
− a1
1−δ1
e
l=k+1
=
(cid:21)
(cid:35)
(cid:21)
From properties of Riemann integration, for j large enough:
(cid:21)
≤
(cid:20)
i−1(cid:88)
(k+2)1−δ1
1
a1
1−δ1
e
(k + 1)δ2
(cid:34)
i−1(cid:88)
k=j
(k+2)1−δ1
a1
1−δ1
e
(cid:20)
a1
1−δ1
k1−δ1 1
kδ2
e
(i+1)1−δ1
1
k=j
(cid:20)
i−1(cid:88)
i+1(cid:88)
k=j
= 2δ2
= 2δ2
k=j+2
= 2δ2 e
a1
1−δ1
(k+2)1−δ1
a1
1−δ1
e
1
( k
2 + 1)δ2
(cid:21)
(k + 2)δ2
1
(cid:21)
(i + 1)δ2
+ 2δ2
+
i(cid:88)
(cid:20)
a1
1−δ1
e
k1−δ1 1
kδ2
k=j+2
≤ 2δ2 e
a1
1−δ1
(i+1)1−δ1
1
(i + 1)δ2
+ 2δ2
+
(cid:90) i+1
j+2
(cid:20)
a1
1−δ1
e
t1−δ1 1
tδ2
(cid:21)
dt
36
(185)
(186)
(cid:90) i+1
(cid:20)
a1
1−δ1
e
j+2
a1
1−δ1
e
t1−δ1 1
tδ2
(cid:21)
(cid:21)
t1−δ1 1
tδ1
tδ2−δ1
(i+1)1−δ1
a1
1−δ1
e
= a1
(cid:90) i+1
(cid:20)
j+2
= a1
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
(1 − r1(l))
r2(i)
=
(cid:35)
(cid:33)
i−1(cid:88)
(cid:20)
k=j
l=k+1
− a1
1−δ1
a2e
(i+1)1−δ1
k=j
(i+1)1−δ1 1
a1
1−δ1
≤ 2δ2 a2e
dt + C1
dt + C1
(cid:21)
1
a1
1−δ1
e
(k+2)1−δ1
(cid:104)
(cid:82) i+1
j+2
e
(k + 1)δ2
t1−δ1 1
tδ2
a1
1−δ1
≤
(cid:105)
dt
2δ2 a2
(i+1)δ2
a1
1−δ1
(cid:82) i+1
e
j+2
(cid:104)
(cid:82) i+1
e
j+2
a1
1−δ1
a1
1−δ1
e
2δ2 a2
(cid:82) i+1
(cid:104)
j+2
e
(cid:104)
+2δ2 a2
(i+1)1−δ1
t1−δ1 1
tδ2
a1
1−δ1
(i+1)1−δ1
(cid:105)
dt
(cid:105)
t1−δ1 1
tδ2
tδ2−δ1
e
a1
1−δ1
t1−δ1 1
tδ2
(cid:105)
dt
dt + C1
=
2δ2 a2
(i + 1)δ2
≤ 2δ2 a2
(i + 1)δ2
+
+
a1
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Again by the fundamental theorem of calculus,
where C1 = C1(j) > 0 for sufficiently large j. From (184)-(185) we have
The second term stays bounded if δ1 = δ2 and goes to zero as i → ∞ if δ1 < δ2, thus establishing the Lemma for the case
δ1 < 1. Also, in the case δ1 = δ2, we have from (186):
(1 − r1(l))
r2(i)
≤ 2δ2 a2
(i + 1)δ2
+
(187)
(cid:35)
(cid:33)
(cid:104)(cid:82) i+1
j+2
2δ2 a2
a1
(cid:104)
k=j
l=k+1
+
a1 + C1
≤ 2δ2 a2 +
2δ2 a2
a1
1−δ1
e
t1−δ1 1
tδ2
(cid:105)
(cid:105)−1
dt
thus making the choice of B in (139) independent of i, j.
Now consider the case δ1 = 1. Consider j sufficiently large, such that,
r1(i) ≤ 1, ∀i ≥ j
(188)
Using a similar set of manipulations for k ≥ j, we have
i−1(cid:89)
l=k+1
(1 − r1(l)) ≤ e
1
l+1
l=k+1
−a1
−a1
(cid:80)i−1
(cid:82) i+1
−a1ln( i+1
k+2 )
(k + 2)a1
(i + 1)a1
k+2
1
t dt
≤ e
= e
=
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
37
We thus have
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
k=j
l=k+1
(cid:33)
(1 − r1(l))
(cid:35)
r2(i)
≤
a2
(i + 1)a1
≤ 2δ2 a2
(i + 1)a1
=
2δ2 a2
(i + 1)a1
k=j
i−1(cid:88)
i−1(cid:88)
i+1(cid:88)
k=j
(k + 2)a1
(k + 1)δ2
(k + 2)a1
(k + 2)δ2
ka1
kδ2
k=j+2
Now, if a1 ≥ δ2, then
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
(cid:33)
(cid:35)
≤
(189)
(1 − r1(l))
r2(i)
k=j+2
ka1−δ2
i+1(cid:88)
(i + 1)a1−δ2 +
(cid:20)
(i + 1)a1−δ2 +
k=j
l=k+1
≤ 2δ2 a2
(i + 1)a1
=
2δ2 a2
(i + 1)a1
≤ 2δ2 a2
(i + 1)a1
i(cid:88)
(cid:90) i+1
k=j+2
(cid:21)
ka1−δ2
ta1−δ2 dt
=
2δ2 a2
(i + 1)δ2
+
2δ2 a2
a − δ2 + 1
j+2
(i + 1)a−δ2+1 − (j + 2)a−δ2+1
(i + 1)a1
The second term is bounded if δ2 = 1 and vanishes if δ2 > 1. If a1 < δ2 is resolved similarly.
APPENDIX D
PROOFS OF LEMMAS 24,23
Proof of Lemma 24
Proof: It follows from (131) and (133), and the fact that
that the evolution of the sequence,(cid:8)(cid:101)x◦
avg(i)(cid:9)
(cid:101)x
avg(i + 1) =(cid:101)x
We note that (191) can be written as(cid:101)x
avg(i+1) =(cid:101)x
◦
◦
i≥0 is given by
(cid:1) = 0
N(cid:88)
(1N ⊗ IM )T(cid:0)L ⊗ IM
(cid:34)(cid:101)x
avg(i)− 1
◦
N
avg(i)+α(i)(cid:2)R((cid:101)x
avg(i))+Γ(i+1,(cid:101)x
avg(i)−α(i)
◦
n=1
◦
◦
(cid:35)
gn(zn(i))
avg(i), ω)(cid:3)
◦
where
R(y) = − (y − h(θ
∗
)) ,
N(cid:88)
n=1
Γ(i + 1, y, ω) =
1
N
gn(zn(i)) − h(θ
∗
), y ∈ RM
Such a definition of R(·), Γ(·) clearly satisfies Assumptions (B.1)-(B.2) of Theorem 29. Now, defining
V (y) = (cid:107)y − h(θ
∗
)(cid:107)2
we have
V (h(θ
∗
)) = 0,V (y) > 0,y (cid:54)= h(θ
∗
), lim(cid:107)y(cid:107)→∞V (y) = ∞
(190)
(191)
(192)
(193)
(194)
(195)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Also, we have for > 0
sup
<(cid:107)y−h(θ∗)(cid:107)< 1
(R(y), Vy(y)) =
=
sup
<(cid:107)y−h(θ∗)(cid:107)< 1
≤ −22
< 0
(cid:0)−2(cid:107)y − h(θ
∗
)(cid:107)2(cid:1)
thus verifying Assumption (B.3). Finally from (95) and (192)-(193), we have
(cid:107)R(y)(cid:107)2 + Eθ∗(cid:2)(cid:107)Γ(i + 1, y, ω)(cid:107)2(cid:3) =
∗
∗
)(cid:107)2 + η(θ
= (cid:107)y − h(θ
≤ k1(1 + V (y))
≤ k1(1 + V (y)) − (R(y), Vy(y))
)
for k1 = max(1, η(θ∗)). Thus the Assumptions (B.1)-(B.4) are satisfied, and we have the claim in (136).
To establish (137), we note that, for sufficiently large i,
(cid:104)(cid:13)(cid:13)(cid:101)x
Eθ∗
)(cid:13)(cid:13)2(cid:105)
∗
avg(i) − h(θ
◦
=
= (1 − α(i − 1))2Eθ∗
avg(i − 1) − h(θ
◦
(cid:104)(cid:13)(cid:13)(cid:101)x
(cid:104)(cid:13)(cid:13)(cid:101)x
+ α2(i − 1)η(θ
∗
)
+ α2(i − 1)η(θ
∗
)
+
∗
)(cid:13)(cid:13)2(cid:105)
)(cid:13)(cid:13)2(cid:105)
+
≤ (1 − α(i − 1))Eθ∗
avg(i − 1) − h(θ
◦
∗
38
(196)
(197)
(198)
where the last step follows from the fact that 0 ≤ (1 − α(i)) ≤ 1 for sufficiently large i. Continuing the recursion in (198), we
have for sufficiently large j ≤ i
k=j α(k) → ∞ as i → ∞ because 0.5 < τ1 ≤ 1. Thus, the first term in (199) goes
to zero as i → ∞. The second term in (199) falls under the purview of Lemma 25 with δ1 = τ1 and δ2 = 2τ1 and hence goes
to zero as i → ∞. We thus have
Eθ∗
avg(i) − h(θ
◦
∗
= 0
(200)
)(cid:13)(cid:13)2(cid:105)
(cid:104)(cid:13)(cid:13)(cid:101)x
lim
i→∞
Proof of Lemma 23
≤
Eθ∗
i−1(cid:89)
≤(cid:16)
From Assumption (D.5), we note that(cid:80)i−1
k=j
e
(cid:104)(cid:13)(cid:13)(cid:101)x
avg(i) − h(θ
◦
∗
(1 − α(k))
)(cid:13)(cid:13)2(cid:105) ≤
(cid:13)(cid:13)(cid:101)x
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
k=j α(k)(cid:17)(cid:13)(cid:13)(cid:101)x
−(cid:80)i−1
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
+ η(θ
k=j
∗
∗
)
+ η(θ
)
k=0
l=k+1
avg(0) − h(θ
◦
∗
)(cid:13)(cid:13)2 +
(cid:33)
)(cid:13)(cid:13)2 +
(cid:33)
(199)
(cid:35)
(cid:35)
(1 − α(l))
α2(k)
(1 − α(l))
α2(k)
l=k+1
avg(0) − h(θ
◦
∗
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Proof: Recall from (131) and (191) that the evolution of the sequences {(cid:101)x◦(i)}i≥0 and(cid:8)(cid:101)x◦
avg(i)(cid:9)
i≥0 are given by
◦
(i + 1) =(cid:101)x
(cid:101)x
avg(i + 1) =(cid:101)x
◦
(cid:101)x
(cid:104)
Pθ∗
lim
i→∞
◦
◦
(i)−
(cid:1)(cid:101)x
(i) − β(i)(cid:0)L ⊗ IM
− α(i) [(cid:101)x(i) − J(z(i))]
avg(i)−
◦
N(cid:88)
avg(i)(cid:1)(cid:13)(cid:13) = 0
(cid:34)(cid:101)x
(i) −(cid:0)1N ⊗(cid:101)x
avg(i) − 1
◦
N
− α(i)
n=1
◦
◦
(cid:13)(cid:13)(cid:101)x
(cid:35)
gn(zn(i))
(cid:105)
= 1
(1N ⊗ IM ) (1N ⊗ IM )T
1
N
◦
◦
◦
(i) = 1N ⊗(cid:101)x
(i + 1) −(cid:0)1N ⊗(cid:101)x
=(cid:2)INM − β(i)(cid:0)L ⊗ IM
avg(i), P 1N ⊗(cid:101)x
avg(i) = 1N ⊗(cid:101)x
avg(i + 1)(cid:1) =
(cid:1) − α(i)INM − P(cid:3) [(cid:101)x
−(cid:0)1N ⊗(cid:101)x
◦
◦
◦
avg(i),∀i
◦
(i)−
◦
avg(i)(cid:1)(cid:3)
+ α(i) [J(z(i)) − P J(z(i))]
To establish the claim (134), Lemma 23, we prove
Recall the matrix
P =
and note that
P(cid:101)x
From (201)-(202), we then have (cid:101)x
Choose δ satisfying
0 < δ < τ1 − 1
2 + 1
− τ2.
Such a choice exists by Assumption (D.5). Now claim:
lim
i→∞
1
(i + 1)
1
2+1
+δ
(cid:107)J(z(i)) − P J(z(i))(cid:107) = 0
Indeed, consider any > 0. We then have from Assumption (D.4) and Chebyshev's inequality
(cid:107)J(z(i)) − P J(z(i))(cid:107) >
≤
1
Pθ∗
1
+δ
2+1
(i + 1)
≤(cid:88)
(cid:2)(cid:107)J(z(i)) − P J(z(i))(cid:107)2+1(cid:3)=
1
(i + 1)1+δ(2+1)2+1
1
(i + 1)1+δ(2+1)
i≥0
κ(θ∗)
2+1
(cid:88)
i≥0
< ∞
39
(201)
(202)
(203)
(204)
(205)
(206)
(207)
(208)
(cid:34)
Pθ∗
(cid:88)
i≥0
Eθ
(cid:34)
(cid:34)
(cid:35)
= 1
(cid:35)
(cid:35)
= 0
It then follows from the Borel-Cantelli Lemma (see [50]) that for arbitrary > 0
Pθ∗
1
(i + 1)
1
2+1
+δ
(cid:107)J(z(i)) − P J(z(i))(cid:107) > i.o.
Pθ∗ [Ω2] = 1, where the sequence(cid:8)(cid:101)x◦
where i.o. stands for infinitely often. Since the above holds for arbitrarily small, we have (see [50]) the a.s. claim in (207).
Consider the set Ω1 ⊂ Ω with Pθ∗ [Ω1] = 1, where the a.s. property in (207) holds. Also, consider the set Ω2 ⊂ Ω with
now show that, on Ω3, the sample paths of the sequence {(cid:101)x◦(i)}i≥0 converge to (1N ⊗ h(θ∗)), thus proving the Lemma. In
i≥0 converges to h(θ∗). Let Ω3 = Ω1 ∩ Ω2. It is clear that Pθ∗ [Ω3] = 1. We will
avg(i)(cid:9)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
40
the following we index the sample paths by ω to emphasize the fact that we are establishing properties pathwise.
From (205), we have on ω ∈ Ω3(cid:13)(cid:13)(cid:101)x
◦
(i + 1, ω) −(cid:0)1N ⊗(cid:101)x
≤(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
◦
avg(i + 1, ω)(cid:1)(cid:13)(cid:13) ≤
(cid:1) − α(i)INM − P(cid:13)(cid:13)
(cid:13)(cid:13)(cid:101)x
(i, ω) −(cid:0)1N ⊗(cid:101)x
◦
For sufficiently large i, we have
From (208) for ω ∈ Ω3 we can choose > 0 and j(ω) such that ∀i ≥ j(ω)
Let j(ω) be sufficiently large such that (209) is also satisfied in addition to (210). We then have for ω ∈ Ω3, i ≥ j(ω)
◦
avg(i, ω)(cid:1)(cid:13)(cid:13)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
−δ
(i + 1)τ1− 1
1
2+1
(i + 1)
1
2+1
+δ
[J(z(i, ω)) − P J(z(i, ω))]
(cid:1)−α(i)INM − P(cid:13)(cid:13) ≤ 1 − β(i)λ2(L)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ .
[J(z(i, ω)) − P J(z(i, ω))]
◦
◦
avg(i, ω)(cid:1)(cid:13)(cid:13) ≤
(cid:0)1 − β(k)λ2(L)(cid:1)(cid:107)(cid:101)x
−(cid:0)1N ⊗(cid:101)x
(cid:0)1 − β(l)λ2(L)(cid:1)(cid:33)
(cid:34)(cid:32) i−1(cid:89)
(cid:0)1 − β(k)λ2(L)(cid:1) ≤ e
l=k+1
−λ2(L)(cid:80)i−1
−bλ2(L)(cid:80)i−1
(j(ω), ω)−
avg(j(ω), ω)(cid:1)(cid:13)(cid:13) +
(cid:35)
◦
1
(k + 1)τ1− 1
2+1
−δ
k=j(ω)
= e
k=j(ω)
1
(k+1)τ2
a
+
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(cid:13)(cid:13)I−β(i)(cid:0)L ⊗ IM
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)
(i, ω) −(cid:0)1N ⊗(cid:101)x
i−1(cid:89)
i−1(cid:88)
(i + 1)
k=j(ω)
2+1
+δ
1
◦
1
(cid:13)(cid:13)(cid:101)x
≤
+ a
k=j(ω)
i−1(cid:89)
(209)
(210)
(211)
For the first term on the R.H.S. of (211) we note that
k=j(ω)
β(k)
(212)
which goes to zero as i → ∞ since τ2 < 1 by Assumption (D.5). Hence the first term on the R.H.S. of (211) goes to zero
as i → ∞. The summation in the second term on the R.H.S. of (211) falls under the purview of Lemma 25 with δ1 = τ2
− δ. It follows from the choice of δ in (206) and Assumption (D.5) that δ1 < δ2 and hence the term
and δ2 = τ1 − 1
(cid:80)i−1
(cid:34)(cid:16)(cid:81)i−1
2+1
(cid:0)1 − β(l)λ2(L)(cid:1)(cid:17)
k=j(ω)
l=k+1
→ 0 as i → ∞. We then conclude from (211) that, for ω ∈ Ω3
(cid:35)
(i, ω) −(cid:0)1N ⊗(cid:101)x
−δ
1
(cid:13)(cid:13)(cid:101)x
◦
τ1− 1
2+1
(k+1)
lim
i→∞
avg(i, ω)(cid:1)(cid:13)(cid:13) = 0
◦
(213)
The Lemma then follows from the fact that Pθ∗ [Ω3] = 1.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
41
To establish (135), we have from (205)
Taking expectations on both sides and from (117)
◦
◦
◦
◦
◦
◦
(cid:13)(cid:13)(cid:101)x
+ α2(i)(cid:107)J(z(i)) − P J(z(i))(cid:107)2
avg(i + 1)(cid:1)(cid:13)(cid:13)2 ≤
(i + 1) −(cid:0)1N ⊗(cid:101)x
≤(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:1) − α(i)INM − P(cid:13)(cid:13)2
avg(i)(cid:1)(cid:13)(cid:13)2 +
(cid:13)(cid:13)(cid:101)x
(i) −(cid:0)1N ⊗(cid:101)x
+ 2α(i)(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:1) − α(i)INM − P(cid:13)(cid:13)
avg(i)(cid:1)(cid:13)(cid:13)(cid:107)J(z(i)) − P J(z(i))(cid:107) +
(cid:13)(cid:13)(cid:101)x
(i) −(cid:0)1N ⊗(cid:101)x
avg(i + 1)(cid:1)(cid:13)(cid:13)2(cid:105) ≤
(cid:104)(cid:13)(cid:13)(cid:101)x
(i + 1) −(cid:0)1N ⊗(cid:101)x
≤(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:1) − α(i)INM − P(cid:13)(cid:13)2
(cid:104)(cid:13)(cid:13)(cid:101)x
avg(i)(cid:1)(cid:13)(cid:13)2(cid:105)
(i) −(cid:0)1N ⊗(cid:101)x
(cid:1) − α(i)INM − P(cid:13)(cid:13) κ1 (θ
+ 2α(i)(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:104)(cid:13)(cid:13)(cid:101)x
avg(i)(cid:1)(cid:13)(cid:13)2(cid:105)
(i) −(cid:0)1N ⊗(cid:101)x
(cid:1) − α(i)INM − P(cid:13)(cid:13) κ1 (θ
+ 2α(i)(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
Eθ∗
Eθ∗
∗
◦
◦
◦
◦
◦
◦
∗
Eθ∗
+ α2(i)κ2(θ
+
)
+
) +
∗
)
where we used the inequality that ∀i(cid:13)(cid:13)(cid:101)x
Choose j sufficiently large such that ∀i ≥ j
For i ≥ j, it can be shown that, for c1 > 0 a constant:
◦
◦
(i)−(cid:0)1N ⊗(cid:101)x
avg(i)(cid:1)(cid:13)(cid:13)≤(cid:13)(cid:13)(cid:101)x
avg(i)(cid:1)(cid:13)(cid:13)2+1.
(cid:1) − α(i)INM − P(cid:13)(cid:13) 1 − β(i)λ2(L).
◦
◦
(i)−(cid:0)1N ⊗(cid:101)x
(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:104)(cid:13)(cid:13)(cid:101)x
(i + 1) −(cid:0)1N ⊗(cid:101)x
≤(cid:2)1 − β(i)λ2(L) + 2α(i)κ1(θ
(cid:104)(cid:13)(cid:13)(cid:101)x
(i) −(cid:0)1N ⊗(cid:101)x
Eθ∗
Eθ∗
◦
◦
◦
avg(i + 1)(cid:1)(cid:13)(cid:13)2(cid:105) ≤
)(cid:3)
avg(i)(cid:1)(cid:13)(cid:13)2(cid:105)
∗
◦
+ α(i)c1.
Now choose j1 ≥ j and 0 < c2 < λ2(L)12 such that,
1 − β(i)λ2(L) + 2α(i)κ1(θ
∗
) ≤ 1 − β(i)c2, ∀i ≥ j1
Then the claim in (135) follows because for i ≥ j1
Eθ∗
(cid:104)(cid:13)(cid:13)(cid:101)x
i−1(cid:89)
k=j1
◦
◦
avg(i)(cid:1)(cid:13)(cid:13)2(cid:105) ≤
(i) −(cid:0)1N ⊗(cid:101)x
Eθ∗
(cid:104)(cid:13)(cid:13)(cid:101)x
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
(1 − β(k)c2)
◦
+ c1
(j1) −(cid:0)1N ⊗(cid:101)x
◦
avg(j)(cid:1)(cid:13)(cid:13)2(cid:105)
(cid:33)
(cid:35)
+
(1 − β(l)c2)
α(k)
12Such a choice exists because τ1 > τ2.
k=j1
l=k+1
(214)
(215)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
42
and the first and second terms on the R.H.S. of (214) vanish as i → ∞ by the argument in (212) and Lemma 25, respectively.
APPENDIX E
PROOF OF LEMMA 26
Proof of Lemma 26: From (131) and (132) we have
(cid:98)x(i + 1) −(cid:101)x
◦
(i + 1) =(cid:2)INM − β(i)(cid:0)L ⊗ IM
(cid:1)−
− α(i)INM
(i)] − β(i) (Υ(i) + Ψ(i))
For sufficiently large j, we have
We then have from (216), for i ≥ j,
◦
◦
(cid:13)(cid:13) ≤ 1 − α(i),∀i ≥ j
(cid:3) [(cid:98)x(i) −(cid:101)x
(cid:1) − α(i)INM
(i + 1)(cid:107)2(cid:105) ≤
(i)(cid:107)2(cid:105)
(cid:104)(cid:107)(cid:98)x(i) −(cid:101)x
+ β2(i)Eθ∗(cid:2)(cid:107)Υ(i) + Ψ(i)(cid:107)2(cid:3)
(cid:104)(cid:107)(cid:98)x(i) −(cid:101)x
(i)(cid:107)2(cid:105)
(cid:107)(cid:98)x(j) −(cid:101)x
i−1(cid:89)
(cid:35)
(cid:33)
(cid:34)(cid:32) i−1(cid:89)
(1 − α(k))
+ ηqβ2(i)
k=j
+
◦
◦
◦
(1 − α(l))
β2(k)
(j)(cid:107)2+
(cid:13)(cid:13)I − β(i)(cid:0)L ⊗ IM
(cid:104)(cid:107)(cid:98)x(i + 1) −(cid:101)x
Eθ∗
≤ (1 − α(i))2 Eθ∗
≤ (1 − α(i)) Eθ∗
(cid:104)(cid:107)(cid:98)x(i) −(cid:101)x
◦
Eθ∗
(i)(cid:107)2(cid:105)≤
i−1(cid:88)
+ ηq
(216)
(217)
(218)
(219)
(220)
where the last step follows from the fact that 0 ≤ (1 − α(i)) ≤ 1 for i ≥ j and (16). Continuing the recursion, we have
The first and second terms on the R.H.S. of (220) vanish as i → ∞, respectively because 1) of an argument similar to the proof
of Lemma 24, and 2) by Lemma 25, with δ1 = τ1, δ2 = 2τ2, since by Assumption (D.5), 2τ2 > τ1. Thus:
which shows that the sequence {(cid:107)(cid:98)x(i) −(cid:101)x◦(i)(cid:107)}i≥0 converges to 0 in L2 (mean-squared sense). We then have from Lemma 23
= 0
(221)
k=j
l=k+1
◦
Eθ∗
lim
i→∞
(cid:104)(cid:107)(cid:98)x(i) −(cid:101)x
(i)(cid:107)2(cid:105)
(cid:104)(cid:107)(cid:98)x(i) − 1N ⊗ h(θ
)(cid:107)2(cid:105) ≤
(cid:104)(cid:107)(cid:98)x(i) −(cid:101)x
(i)(cid:107)2(cid:105)
(cid:104)(cid:107)(cid:101)x
+ 2 lim
i→∞
Eθ∗
Eθ∗
∗
◦
◦
≤ 2 lim
i→∞
+
Eθ∗
lim
i→∞
(222)
)(cid:107)2(cid:105)
∗
= 0
(i) − 1N ⊗ h(θ
thus establishing the claim in (143).
We now show that the sequence {(cid:107)(cid:98)x(i) −(cid:101)x◦(i)(cid:107)}i≥0 also converges a.s. to a finite random variable. Choose j sufficiently
43
(223)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
large as in (217). We then have from (216)
◦
(i) =
(cid:98)x(i) −(cid:101)x
i−1(cid:89)
(cid:0)INM − β(k)(cid:0)L ⊗ IM
(cid:34)(cid:32) i−1(cid:89)
− i−1(cid:88)
k=j
(cid:0)INM − β(l)(cid:0)L ⊗ IM
k=j
l=k+1
(cid:34)(cid:32) i−1(cid:89)
− i−1(cid:88)
k=j
l=k+1
(cid:0)INM − β(l)(cid:0)L ⊗ IM
(j))−
◦
(cid:1) − α(k)I(cid:1)((cid:98)x(j) −(cid:101)x
(cid:1) − α(l)I(cid:1)(cid:33)
(cid:1) − α(l)I(cid:1)(cid:33)
β(k)Υ(k)
−
(cid:35)
(cid:35)
β(k)Ψ(k)
The first term on the R.H.S. of (223) converges a.s. to zero as i → ∞ by a similar argument as in the proof of Lemma 24.
Since the sequence {Υ(i)}i≥0 is i.i.d., the second term is a weighted summation of independent random vectors. Define the
triangular array of weight matrices, {Ai,k, j ≤ k ≤ i − 1}i>j, by
l=k+1
i−1(cid:89)
(cid:0)INM − β(l)(cid:0)L ⊗ IM
(cid:0)INM − β(l)(cid:0)L ⊗ IM
Ai,k =
(cid:34)(cid:32) i−1(cid:89)
k=j
l=k+1
(cid:1) − α(l)I(cid:1) β(k)
(cid:1) − α(l)I(cid:1)(cid:33)
(cid:35)
i−1(cid:88)
β(k)Υ(k)
=
Ai,kΥ(k)
k=j
i−1(cid:88)
k=j
(cid:107)Ai,k(cid:107)2 ≤
≤ lim sup
i→∞
(cid:34)(cid:32) i−1(cid:89)
i−1(cid:88)
k=j
l=k+1
(cid:33)
(1 − α(l))
(cid:35)
β2(k)
= 0
sup
i>j
i−1(cid:88)
k=j
(cid:107)Ai,k(cid:107)2 = C3 < ∞
(224)
(225)
(226)
(227)
We then have
i−1(cid:88)
By Lemma 25 and Assumption (D.5) we note that
lim sup
i→∞
It then follows that
(cid:110)(cid:80)i−1
(cid:111)
The sequence
weighted summations of independent random vectors (see [57], [58], [59]).
k=j Ai,kΥ(k)
i>j
then converges a.s. to a finite random vector by standard results from the limit theory of
In a similar way, the last term on the R.H.S of (223) converges a.s. to a finite random vector since by the properties of
dither the sequence {Ψ(i)}i≥0 is i.i.d. It then follows from (223) that the sequence {(cid:98)x(i) −(cid:101)x◦(i)}i≥0 converges a.s. to a
finite random vector, which in turn implies that the sequence {(cid:107)(cid:98)x(i) −(cid:101)x◦(i)(cid:107)}i≥0 converges a.s. to a finite random variable.
However, we have already shown that the sequence {(cid:107)(cid:98)x(i) −(cid:101)x◦(i)(cid:107)}i≥0 converges in mean-squared sense to 0. It then follows
from the uniqueness of the mean-squared and a.s. limit, that the sequence {(cid:107)(cid:98)x(i) −(cid:101)x◦(i)(cid:107)}i≥0 converges a.s. to 0. In other
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
words,
(cid:104)
i→∞(cid:107)(cid:98)x(i) −(cid:101)x
lim
◦
Pθ∗
(cid:105)
(i)(cid:107) = 0
= 1
The claim in (142) then follows from (228) and Lemma 23.
APPENDIX F
PROOFS OF THEOREMS 21,22
Proof of Theorem 21
Proof: Recall the evolution of the sequences {(cid:101)x(i)}i≥0, {(cid:98)x(i)}i≥0 in (121) and (132).
Then writing L(i) = L +(cid:101)L(i) and using the fact that
(cid:17)(cid:98)xC⊥ (i), ∀i
(cid:17)(cid:98)x(i) =
(cid:16)(cid:101)L(i) ⊗ IM
we have from (121) and (132) (cid:101)x(i + 1) −(cid:98)x(i + 1) = [INM − β(i) (L(i) ⊗ IM )−
(cid:16)(cid:101)L(i) ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)(cid:98)xC⊥ (i)
44
(228)
(229)
(230)
− β(i)
(231)
(232)
lim
i→∞ y(i) = 0
= 1
To prove (123), it clearly suffices (from Lemma 26) to prove
For ease of notation, introduce the sequence {y(i)}i≥0, given by
y(i + 1) =(cid:2)INM − β(i)(cid:0)L ⊗ IM
−α(i)INM ] ((cid:101)x(i) −(cid:98)x(i)) − β(i)
y(i) =(cid:101)x(i) −(cid:98)x(i)
(cid:105)
(cid:104)
Pθ∗
From (230), the evolution of the sequence {y(i)}i≥0 is:
(cid:1) − α(i)INM
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)(cid:98)xC⊥ (i). However, from Lemma 26:
The sequence {y(i)}i≥0 is not uniformly bounded, in general, because of β(i)
(cid:105)
(cid:104)
i→∞(cid:98)xC⊥ (i) = 0
and, hence, asymptotically, its effect diminishes. However, {(cid:98)xC⊥ (i)}i≥0 is not uniformly bounded over sample paths and, hence,
(cid:40) aa min(a, R)
yR(i + 1) =(cid:2)INM − β(i)(cid:0)L ⊗ IM
(cid:1) − α(i)INM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
((cid:98)xC⊥ (i))R
(cid:105)
For a vector, the truncation operation applies componentwise. For R > 0, we also consider the sequences, {yR(i)}i≥0:
we use truncation arguments (see, e.g., [56]). For a scalar a, define its truncation (a)R at level R > 0 by
(cid:3) y(i)−
(cid:17)(cid:98)xC⊥ (i)
(cid:16)(cid:101)L(i) ⊗ IM
We will show that for every R > 0
(cid:3) yR(i)−
if a (cid:54)= 0
if a = 0
y(i) − β(i)
(a)R =
0
Pθ∗
lim
yR(i)−
(233)
(234)
(235)
(236)
(237)
= 1
− β(i)
− β(i)
(cid:104)
Pθ∗
lim
i→∞ yR(i) = 0
= 1
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
45
Now, the sequence {(cid:98)xC⊥ (i)}i≥0 converges a.s. to zero, and, hence, for every > 0, there exists R() > 0 (see [50]), such that
and, hence, from (233)-(236)
Pθ∗
This, together with (237), will then imply
(cid:20)
Pθ∗
sup
i≥0
(cid:20)
(cid:21)
(cid:13)(cid:13)(cid:13)(cid:98)xC⊥ (i) − ((cid:98)xC⊥ (i))R()(cid:13)(cid:13)(cid:13) = 0
(cid:21)
(cid:13)(cid:13)(cid:13) = 0
(cid:13)(cid:13)(cid:13)y(i) − yR()(i)
(cid:105)
(cid:104)
sup
i≥0
> 1 −
lim
i→∞ y(i) = 0
Pθ∗
> 1 −
> 1 −
Since > 0 is arbitrary in (240), we will be able to conclude (123). Thus, the proof reduces to establishing (237) for every
R > 0, which is carried out in the following.
For a given R > 0 consider the recursion given in (236). Choose ε1 > 0 and ε2 < 0 such that
(241)
Because τ2 > .5 in Assumption (D.5) permits such choice of ε1, ε2. Let ρ > 0 be constant and define V : N × RNM (cid:55)−→ R+
1 − ε2 < 2τ2 − ε1.
(cid:1) x + ρiε2 .
V (i, x) = iε1 xT(cid:0)L ⊗ IM
(cid:110)
x(0),
L(j),{zn(j)}1≤N , Υ(j), Ψ(j)
(242)
(cid:19)
(cid:111)
0≤j<i
Recall the filtration {Fi}i≥0 in (104)
Fi = σ
(cid:18)
(238)
(239)
(240)
(243)
to which all the processes of interest are adapted. We now show that there exists an integer iR > 0 sufficiently large, such that
the process {V (i, yR(i))}i≥iR
. To this end, we note that, using
the recursion (236):
yT
(i + 1)ε1 yT
+2β(i)α(i)yT
−2β(i)yT
−2α(i)yT
yR,C⊥ (i)−
= (i + 1)ε1(cid:104)
Eθ∗ [V (i + 1, yR(i + 1))Fi] − V (i, yR(i)) =
R(i + 1)(cid:0)L ⊗ IM
R,C⊥ (i)(cid:0)L ⊗ IM
− iε1 yT
R,C⊥ (i)(cid:0)L ⊗ IM
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:104)
is a non-negative supermartingale w.r.t. the filtration {Fi}i≥iR
(cid:1) yR(i + 1) + ρ(i + 1)ε2
R(i)(cid:0)L ⊗ IM
(cid:1) yR(i) − ρiε2
(cid:1) yR,C⊥ (i)−
(cid:1)2
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:1) yR,C⊥ (i)+
(cid:1)2
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:1)3
(cid:1) yR,C⊥ (i)+
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:17)(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:1)
(cid:12)(cid:12)(cid:12) Fi
(cid:105)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:17)(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:1)
((cid:98)xC⊥ (i))R(cid:12)(cid:12)(cid:12) Fi
(cid:105)
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)(cid:0)L ⊗ IM
(cid:17)R(cid:16)(cid:101)L(i) ⊗ IM
(cid:1)
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:105)(cid:105)
((cid:98)xC⊥ (i))R Fi
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:1) yR,C⊥ (i) − ρiε2
(cid:104)
(cid:20)(cid:16)(cid:98)xTC⊥ (i)
+ (i + 1)ε2 − iε1 yT
+2β2(i)Eθ∗
yR,C⊥ (i)
+
+β2(i)Eθ∗
yT
R,C⊥ (i)
+β2(i)Eθ∗
+
+
yR,C⊥ (i)+
+β2(i)yT
yR,C⊥ (i)+
yT
R,C⊥ (i)
+α2(i)yT
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
46
where we repeatedly used the fact that
and (cid:101)L(i) is independent of Fi.
In going to the next step we use the following inequalities, where c1 > 0 is a constant:
(cid:1) yR,C⊥ (i)
(cid:17)
yR,C⊥ (i)
(cid:1) yR(i) =(cid:0)L ⊗ IM
(cid:0)L ⊗ IM
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
R,C⊥ (i)(cid:0)L ⊗ IM
yR,C⊥ (i) ≥ λ2
yR(i) =
yT
2(L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
(cid:1) yR,C⊥ (i)
N (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
yR,C⊥ (i) ≤ λ2
=
λ2
2(L)
λN (L)
yT
≥ λ2
2(L)
λN (L)
R,C⊥ (i)(cid:0)L ⊗ IM
yT
=
λ2
N (L)
λ2(L)
≤ λ2
yT
N (L)
λ2(L)
R,C⊥ (i)(cid:0)L ⊗ IM
yT
(cid:1)2
(cid:1)2
λN (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
R,C⊥ (i)(cid:0)L ⊗ IM
λ2(L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
R,C⊥ (i)(cid:0)L ⊗ IM
λ2(L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:1)3
yT
=
λ3
N (L)
λ2(L)
≤ λ3
N (L)
λ2(L)
(cid:1) yR,C⊥ (i)
N (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
yR,C⊥ (i) ≤ λ3
(cid:1) yR,C⊥ (i)
(244)
(245)
(246)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
Eθ∗
Eθ∗
yT
yT
(cid:1)
(cid:104)
R,C⊥ (i)
R,C⊥ (i)
λ2
yT
yR,C⊥ (i)
yR,C⊥ (i)
≤ c1λN (L)
≤ λN (L)Eθ∗
≤ c1λN (L)Eθ∗
(cid:12)(cid:12)(cid:12) Fi
(cid:105) ≤
(cid:21)
(cid:13)(cid:13)(cid:13)2 Fi
(cid:17)(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
(cid:20)(cid:13)(cid:13)(cid:13)(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
(cid:104)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2(cid:12)(cid:12)(cid:12) Fi
(cid:105)
= c1λN (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:1) yR,C⊥ (i)
(cid:17)(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:104)
(cid:1)
((cid:98)xC⊥ (i))R(cid:12)(cid:12)(cid:12) Fi
(cid:16)(cid:101)L(i) ⊗ IM
(cid:105)
(cid:17)
(cid:20)(cid:13)(cid:13)(cid:13)yT
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)(cid:13)(cid:13)(cid:13)
(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)((cid:98)xC⊥ (i))R(cid:13)(cid:13)(cid:13)(cid:12)(cid:12)(cid:12) Fi
(cid:1)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)(cid:101)L(i) ⊗ IM
(cid:13)(cid:13)(cid:0)L ⊗ IM
≤ Rc1λN (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)
≤ Rc1λN (L) + Rc1λN (L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:20)(cid:16)(cid:98)xTC⊥ (i)
(cid:17)(cid:0)L ⊗ IM
(cid:1)
(cid:21)
((cid:98)xC⊥ (i))R (cid:12)(cid:12)(cid:12) Fi
(cid:17)R(cid:16)(cid:101)L(i) ⊗ IM
(cid:16)(cid:101)L(i) ⊗ IM
(cid:17)
≤ Rc1λN (L) +
Rc1λN (L)
≤ Eθ∗
R,C⊥ (i)
λ2(L)
yT
Eθ∗
(cid:21)
(cid:1) yR,C⊥ (i)
We go from (249) to (250) because
Eθ∗
47
(247)
(248)
(249)
(250)
(251)
(252)
(253)
(254)
(255)
≤ R2c1λN (L)
(i + 1)ε1 − iε1 ≤ ε1(i + 1)ε1−1
ρ(i + 1)ε2 − ρiε2 ≤ ρε2iε2−1.
(cid:13)(cid:13)(cid:13)((cid:98)xC⊥ (i))R(cid:13)(cid:13)(cid:13) ≤ R. Using inequalities (244)-(254), we have from (243)
(cid:104)
(cid:20)
(cid:105) − V (i, yR(i)) ≤
V (i + 1, yR(i + 1))
(cid:12)(cid:12)(cid:12) Fi
(i + 1)ε1
+2β(i)α(i)
ε1
(i + 1)1 − 2β(i)
+β2(i)
λ2
N (L)
λ2(L)
λ2
2(L)
λN (L)
λ3
N (L)
λ2(L)
− 2α(i)+
+2β2(i)
Rc1λN (L)
λ2(L)
(cid:20)
+
1
2τ2 − ε1
c1λN (L)
+
λ2
+α2(i)+β2(i)
(cid:1) yR,C⊥ (i)+
(cid:0)2Rc1λN (L) + R2c1λN (L)(cid:1) + ρε2iε2−1
(cid:1) yR,C⊥ (i), the coefficient −2β(i)(i + 1)ε1 dominates
(cid:21)
(cid:21)
yT
R,C⊥ (i)(cid:0)L ⊗ IM
R,C⊥ (i)(cid:0)L ⊗ IM
(cid:12)(cid:12)(cid:12) Fi
For the first term on the R.H.S. of (255) involving yT
all other coefficients eventually (τ2 < 1 by Assumption (D.5)); hence, the first term on the R.H.S. of (255) becomes negative
eventually (for sufficiently large i). The second term on the R.H.S. of (255) becomes negative eventually because ρε2 < 0 and
1 − ε2 < 2τ2 − ε1 by assumption. Hence there exists sufficiently large i, say iR, such that,
(cid:104)
Eθ∗
is a non-negative supermartingale w.r.t. the filtration {Fi}i≥iR
(cid:105) − V (i, yR(i)) ≤ 0,∀i ≥ iR.
V (i + 1, yR(i + 1))
This shows {V (i, yR(i))}i≥iR
. Thus, {V (i, yR(i))}i≥iR
where
Since(cid:80)
we have
and similarly
P (L(i) ⊗ IM ) = P(cid:0)L ⊗ IM
(cid:1) = 0, ∀i
P(cid:101)x(i + 1) = P(cid:101)x(i) − α(i) [P(cid:101)x(i) − P J(z(i))]
− β(i)P (Υ(i) + Ψ(i))
P(cid:98)x(i + 1) = P(cid:98)x(i) − α(i) [P(cid:98)x(i) − P J(z(i))]
48
(256)
(257)
(258)
(259)
(260)
(261)
(262)
(266)
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
converges a.s. to a finite random variable (see [50]). Clearly, the sequence ρiε2 goes to zero as ε2 < 0. Then:
(cid:104)
R(i)(cid:0)L ⊗ IM
(cid:105)
(cid:1) yR(i) exists and is finite
= 1
Pθ∗
Since iε1 → ∞ as i → ∞, it follows
i→∞ iε1 yT
lim
R(i)(cid:0)L ⊗ IM
Since yT
(cid:104)
Pθ∗
i→∞ yT
lim
(cid:1) yR(i) ≥ λ2(L)(cid:13)(cid:13)yR,C⊥ (i)(cid:13)(cid:13)2, from (256) we have
(cid:105)
R(i)(cid:0)L ⊗ IM
(cid:104)
(cid:105)
(cid:1) yR(i) = 0
Pθ∗
i→∞ yR,C⊥ (i) = 0
lim
= 1
= 1
To establish (237) we note that
yR,C(i) = 1N ⊗ yR,avg(i)
yR,avg(i + 1) = (1 − α(i)) yR,avg(i)
i≥0 α(i) = ∞, it follows from standard arguments that yR,avg(i) → 0 as i → ∞. We then have from (258)
(cid:104)
Pθ∗
(cid:105)
i→∞ yR,C(i) = 0
lim
= 1
which together with (257) establishes (237). The claim in (123) then follows from the arguments above.
We now prove the claim in (124). Recall the matrix P in (204). Using the fact,
Since the sequences {P(cid:101)x(i)}i≥0 and {P(cid:98)x(i)}i≥0 follow the same recursion and start with the same initial state P(cid:101)x(0), they
(263)
− β(i)P (Υ(i) + Ψ(i))
are equal, and we have ∀i
P y(i) = P ((cid:101)x(i) −(cid:98)x(i))
= 0
(264)
From (233) we then have
y(i + 1) =(cid:2)INM − β(i)(cid:0)L ⊗ IM
(cid:17)
(cid:16)(cid:101)L(i) ⊗ IM
− β(i)
(cid:1) − α(i)INM − P(cid:3) y(i)−
(cid:17)(cid:98)x(i)
(cid:16)(cid:101)L(i) ⊗ IM
y(i) − β(i)
By Lemma 26, to prove the claim in (123), it suffices to prove
From Lemma 26, we note that the sequence {(cid:98)x(i)}i≥0 converges in L2 to 1N ⊗ h(θ∗) and hence L2 bounded, i.e., there
lim
i→∞
(265)
exists constant c3 > 0, such that,
Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) = 0
Eθ∗(cid:2)(cid:107)(cid:98)x(i)(cid:107)2(cid:3) ≤ c3 < ∞
sup
i≥0
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
49
Choose j large enough, such that, for i ≥ j
(cid:1) − α(i)INM − P(cid:13)(cid:13) ≤ 1 − β(i)λ2(L)
(cid:13)(cid:13)INM − β(i)(cid:0)L ⊗ IM
(cid:13)(cid:13)(cid:13) ≤ c2 for some constant c2 > 0, we have for i ≥ j,
(cid:13)(cid:13)(cid:13)(cid:101)L(i)
Noting that (cid:101)L(i) is independent of Fi and
Eθ∗(cid:2)(cid:107)y(i + 1)(cid:107)2(cid:3) =
(cid:104)
yT (i)(cid:0)INM − β(i)(cid:0)L ⊗ IM
(cid:16)(cid:101)L(i)
(cid:17)2
y(i) + β2(i)(cid:98)xT (i)
(cid:21)
(cid:17)2(cid:98)x(i)
(cid:16)(cid:101)L(i)
≤(cid:0)1 − β(i)λ2(L)(cid:1) Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) + c2
(cid:18)
(cid:18)
(cid:1) − α(i)INM − P(cid:1)2
(cid:16)(cid:101)L(i)
(cid:17)2(cid:98)x(i)
2β2(i)Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3)
Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3)
θ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3)
(cid:19)
1 − β(i)λ2(L) + c2
2β2(i) + 2β2(i)c2
2c
+β2(i)yT (i)
+β2(i)yT (i)
2c3β2(i) +
2β2(i)c2
2c
(cid:19)
Eθ∗
+ c2
y(i)
≤
1
2
3
E 1
2
1
2
3
(cid:18)
(cid:19)
+ β2(i)
2c3 + 2c2
c2
2c
1
2
3
where in the last step we used the inequality
θ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) ≤ Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) + 1
E 1
2
Now similar to Lemma 23, choose j1 ≥ j and 0 < c4 < λ2(L), such that,
1 − β(i)λ2(L) + c2
3 ≤ 1 − β(i)c4, ∀i ≥ j1
2β2(i) + 2β2(i)c2
2c
1
2
(267)
(268)
Then, for i ≥ j1, from (267)
Eθ∗(cid:2)(cid:107)y(i + 1)(cid:107)2(cid:3) ≤ (1 − β(i)c4) Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) +
(cid:19)
from which we conclude that limi→∞ Eθ∗(cid:2)(cid:107)y(i)(cid:107)2(cid:3) = 0 by Lemma 25 (see also Lemma 23.)
Proof: Consistency follows from the fact that by Theorem 21 the sequence {(cid:101)x(i)}i≥0 converges a.s. to 1N ⊗ h(θ∗), and
Proof of Theorem 22
1
2c3 + 2c2
c2
2c
2
3
+ β2(i)
(cid:18)
the function h−1(·) exists and is continuous on the open set U.
To establish the second claim, we note that, if h−1(·) is Lipschitz continuous, there exists constant k > 0, such that
(cid:13)(cid:13)h
−1((cid:101)y1) − h
−1((cid:101)y2)(cid:13)(cid:13) ≤ k (cid:107)(cid:101)y1 −(cid:101)y2(cid:107) , ∀ (cid:101)y1,(cid:101)y2 ∈ RM
Since L2 convergence implies L1, we then have from Theorem 21 for 1 ≤ n ≤ N
Eθ∗ [(cid:107)xn(i) − θ
lim
∗(cid:107)]
∗
i→∞(cid:107)Eθ∗ [xn(i) − θ
= lim
i→∞
≤ k lim
i→∞
](cid:107) ≤ lim
i→∞
Eθ∗(cid:2)(cid:13)(cid:13)h
−1 ((cid:101)xn(i)) − h
Eθ∗ [(cid:107)(cid:101)xn(i) − h(θ
∗
−1 (h(θ
)(cid:107)]
))(cid:13)(cid:13)(cid:3)
∗
which establishes the theorem.
= 0
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
50
REFERENCES
[1] R. Olfati-Saber and R. M. Murray, "Consensus problems in networks of agents with switching topology and time-delays," IEEE Trans.
Automat. Contr., vol. 49, no. 9, pp. 1520 -- 1533, Sept. 2004.
[2] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile autonomous agents using nearest neighbor rules," IEEE Trans.
Autom. Control, vol. AC-48, no. 6, pp. 988 -- 1001, June 2003.
[3] L. Xiao and S. Boyd, "Fast linear iterations for distributed averaging," Syst. Contr. Lett., vol. 53, pp. 65 -- 78, 2004.
[4] S. Kar and J. M. F. Moura, "Sensor networks with random links: Topology design for distributed consensus," IEEE Transactions on Signal
Processing, vol. 56, no. 7, pp. 3315 -- 3326, July 2008.
[5] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in sensor networks with communication channel noise and random link
failures," in 41st Asilomar Conference on Signals, Systems, and Computers, Pacific Grove, CA, Nov. 2007.
[6] S. Kar and J. M. F. Moura, "Distributed average consensus in sensor networks with quantized inter-sensor communication," in Proceedings
of the 33rd International Conference on Acoustics, Speech, and Signal Processing, Las Vegas, Nevada, USA, April 1-4 2008.
[7] Y. Hatano and M. Mesbahi, "Agreement over random networks," in 43rd IEEE Conference on Decision and Control, vol. 2, Dec. 2004,
pp. 2010 -- 2015.
[8] T. C. Aysal, M. Coates, and M. Rabbat, "Distributed average consensus using probabilistic quantization," in IEEE/SP 14th Workshop on
Statistical Signal Processing Workshop, Maddison, Wisconsin, USA, August 2007, pp. 640 -- 644.
[9] M. E. Yildiz and A. Scaglione, "Differential nested lattice encoding for consensus problems," in ACM/IEEE Information Processing in
Sensor Networks, Cambridge, MA, April 2007.
[10] A. Kashyap, T. Basar, and R. Srikant, "Quantized consensus," Automatica, vol. 43, pp. 1192 -- 1203, July 2007.
[11] P. Frasca, R. Carli, F. Fagnani, and S. Zampieri, "Average consensus on networks with quantized communication," Submitted to the Int.
J. Robust and Nonlinear Control, 2008.
[12] A. Nedic, A. Olshevsky, A. Ozdaglar, and J. N. Tsitsiklis, "On distributed averaging algorithms and quantization effects," Technical Report
2778, LIDS-MIT, Nov. 2007.
[13] M. Huang and J. Manton, "Stochastic approximation for consensus seeking: mean square and almost sure convergence," in Proceedings
of the 46th IEEE Conference on Decision and Control, New Orleans, LA, USA, Dec. 12-14 2007.
[14] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooperation in networked multi-agent systems," IEEE Proceedings, vol. 95,
no. 1, pp. 215 -- 233, January 2007.
[15] A. Das and M. Mesbahi, "Distributed linear parameter estimation in sensor networks based on Laplacian dynamics consensus algorithm,"
in 3rd Annual IEEE Communications Society on Sensor and Ad Hoc Communications and Networks, vol. 2, Reston, VA, USA, 28-28
Sept. 2006, pp. 440 -- 449.
[16] R. Olfati-Saber, "Distributed Kalman filter with embedded consensus filters," in CDC-ECC'05, 44th IEEE Conference on Decision and
Control and 2005 European Control Conference, 2005, pp. 8179 -- 8184.
[17] I. D. Schizas, A. Ribeiro, and G. B. Giannakis, "Consensus in Ad Hoc WSNs with noisy links - part I: Distributed estimation of deterministic
signals," IEEE Transactions on Signal Processing, vol. 56, no. 1, pp. 350 -- 364, January 2008.
[18] U. A. Khan and J. M. F. Moura, "Distributing the Kalman filter for large-scale systems," IEEE Transactions on Signal Processing, vol. 56,
no. 10, p. 49194935, October 2008.
[19] R. Carli, A. Chiuso, L. Schenato, and S. Zampieri, "Distributed Kalman filtering using consensus strategies," IEEE Journal on Selected
Areas of Communications, vol. 26, no. 4, pp. 622 -- 633, September 2008.
[20] D. Bajovic, D. Jakovetic, J. Xavier, B. Sinopoli, and J. M. F. Moura, "Distributed detection via Gaussian running consensus: Large
deviations asymptotic analysis," IEEE Transactions on Signal Processing, vol. 59, no. 5, May 2011.
[21] D. Jakovetic, J. M. F. Moura, and J. Xavier, "Distributed detection over noisy networks: Large deviations analysis," August 2011, accepted
for publication, IEEE Transactions on Signal Processing, vol. 60, 2012.
[22] C. G. Lopes and A. H. Sayed, "Diffusion least-mean squares over adaptive networks: Formulation and performance analysis," IEEE
Transactions on Signal Processing, vol. 56, no. 7, pp. 3122 -- 3136, July 2008.
[23] S. Stankovic, M. Stankovic, and D. Stipanovic, "Decentralized parameter estimation by consensus based stochastic approximation," in 46th
IEEE Conference on Decision and Control, New Orleans, LA, USA, 12-14 Dec. 2007, pp. 1535 -- 1540.
[24] I. Schizas, G. Mateos, and G. Giannakis, "Stability analysis of the consensus-based distributed LMS algorithm," in Proceedings of the
33rd International Conference on Acoustics, Speech, and Signal Processing, Las Vegas, Nevada, USA, April 1-4 2008, pp. 3289 -- 3292.
[25] S. Ram, V. Veeravalli, and A. Nedic, "Distributed and recursive parameter estimation in parametrized linear state-space models," to appear
in IEEE Transactions on Automatic Control, vol. 55, no. 2, pp. 488 -- 492, February 2010.
[26] S. Kar and J. M. F. Moura, "Convergence rate analysis of distributed gossip (linear parameter) estimation: Fundamental limits and tradeoffs,"
IEEE Journal on Selected Topics on Signal Processing, vol. 5, 2011.
[27] S. Kar and J. M. F. Moura, "Gossip and distributed Kalman filtering: Weak consensus under weak detectability," IEEE Transactions on
Signal Processing, vol. 59, no. 4, pp. 1766-1784, April 2011.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
51
[28] D. Jakovetic, J. ao Xavier, and J. M. F. Moura, "Weight optimization for consenus algorithms with correlated switching topology," IEEE
Transactions on Signal Processing, vol. 58, no. 7, pp. 3788 -- 3801, July 2010.
[29] J. N. Tsitsiklis, "Problems in decentralized decision making and computation," Ph.D., Massachusetts Institute of Technology, Cambridge,
MA, 1984.
[30] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization algorithms,"
IEEE Trans. Autom. Control, vol. AC-31, no. 9, pp. 803 -- 812, September 1986.
[31] D. Bertsekas, J. Tsitsiklis, and M. Athans, "Convergence theories of distributed iterative processes: A survey," Technical Report for
Information and Decision Systems, Massachusetts Inst. of Technology, Cambridge, MA, 1984.
[32] H. Kushner and G. Yin, "Asymptotic properties of distributed and communicating stochastic approximation algorithms," Siam J. Control
and Optimization, vol. 25, no. 5, pp. 1266 -- 1290, Sept. 1987.
[33] S. Kar, S. A. Aldosari, and J. M. F. Moura, "Topology for distributed inference on graphs," IEEE Transactions on Signal Processing,
vol. 56, no. 6, pp. 2609 -- 2613, June 2008.
[34] A. H. Sayed, Fundamentals of Adaptive Filtering, ser. Wiley-Interscience. Hoboken, NJ: John Wiley & Sons, 2003.
[35] V. Solo and X. Kong, Adaptive Signal Processing Algorithms: Stability and Performance. Prentice Hall, 1995.
[36] L. Ljung, System Identification: Theory for the User. Prentice Hall, 1999.
[37] V. S. Borkar, Stochastic Approximation: A Dynamical Systems Viewpoint. Cambridge, UK: Cambridge University Press, 2008.
[38] S. B. Gelfand and S. K. Mitter, "Recursive stochastic algorithms for global optimization in Rd," SIAM J. Control Optim., vol. 29, no. 5,
pp. 999 -- 1018, September 1991.
[39] F. R. K. Chung, Spectral Graph Theory. Providence, RI : American Mathematical Society, 1997.
[40] B. Mohar, "The Laplacian spectrum of graphs," in Graph Theory, Combinatorics, and Applications, Y. Alavi, G. Chartrand, O. R.
Oellermann, and A. J. Schwenk, Eds. New York: J. Wiley & Sons, 1991, vol. 2, pp. 871 -- 898.
[41] B. Bollobas, Modern Graph Theory. New York, NY: Springer Verlag, 1998.
[42] L. Schuchman, "Dither signals and their effect on quantization noise," IEEE Trans. Commun. Technol., vol. COMM-12, pp. 162 -- 165,
December 1964.
[43] S. P. Lipshitz, R. A. Wannamaker, and J. Vanderkooy, "Quantization and dither: A theoretical survey," J. Audio Eng. Soc., vol. 40, pp.
355 -- 375, May 1992.
[44] A. B. Sripad and D. L. Snyder, "A necessary and sufficient condition for quantization errors to be uniform and white," IEEE Trans. Acoust.,
Speech, Signal Processing, vol. ASSP-25, pp. 442 -- 448, October 1977.
[45] R. M. Gray and T. G. Stockham, "Dithered quantizers," IEEE Trans. Information Theory, vol. 39, pp. 805 -- 811, May 1993.
[46] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in sensor networks: Quantized data and random link failures,"
IEEE Transactions on Signal Processing, vol. 58, no. 3, pp. 1383 -- 1400, March 2010, initial post: Dec. 2007. [Online]. Available:
http://arxiv.org/abs/0712.1609
[47] E. Lehmann, Theory of Point Estimation.
[48] H.-F. Chen and L. Guo, Identification and Stochastic Adaptive Control. Boston: Birkhauser, 1991.
[49] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE/ACM Trans. Netw., vol. 14, no. SI, pp. 2508 -- 2530,
John Wiley and Sons, Inc., 1983.
2006.
[50] O. Kallenberg, Foundations of Modern Probability, 2nd ed. Springer Series in Statistics., 2002.
[51] S. Kar, "Large scale networked dynamical systems: Distributed inference," Ph.D., Carnegie Mellon University, Pittsburgh, PA, 2010.
[52] N. Krasovskii, Stability of Motion. Stanford University Press, 1963.
[53] A. V. Oppenheim and R. W. Schafer, Digital Signal Processing. Prentice-Hall, 1975.
[54] M. D. Ilic' and J. Zaborszky, Dynamics and Control of Large Electric Power Systems. Wiley, 2000.
[55] J. S. Thorp, A. G. Phadke, and K. J. Karimi, "Real time voltage-phasor measurements for static state estimation," IEEE Transactions on
Power Apparatus and Systems, vol. 104, no. 11, pp. 3098 -- 3106, November 1985.
[56] M. Nevel'son and R. Has'minskii, Stochastic Approximation and Recursive Estimation. Providence, Rhode Island: American Mathematical
Society, 1973.
[57] Y. Chow, "Some convergence theorems for independent random variables," Ann. Math. Statist., vol. 37, pp. 1482 -- 1493, 1966.
[58] Y. Chow and T. Lai, "Limiting behavior of weighted sums of independent random variables," Ann. Prob., vol. 1, pp. 810 -- 824, 1973.
[59] W. Stout, "Some results on the complete and almost sure convergence of linear combinations of independent random variables and
martingale differences," Ann. Math. Statist., vol. 39, pp. 1549 -- 1562, 1968.
IEEE TRANSACTIONS ON INFORMATION THEORY, VOL. 58, NO. 6, JUNE 2012
52
Soummya Kar (S'05 -- M'10) received the B.Tech. degree in Electronics and Electrical Communication Engineering
from the Indian Institute of Technology, Kharagpur, India, in May 2005 and the Ph.D. degree in electrical and computer
engineering from Carnegie Mellon University, Pittsburgh, PA, in 2010. From June 2010 to May 2011 he was with the
EE Department at Princeton University as a Postdoctoral Research Associate. He is currently an Assistant Research
Professor of ECE at Carnegie Mellon University. His research interests include performance analysis and inference in
large-scale networked systems, adaptive stochastic systems, stochastic approximation, and large deviations.
PLACE
PHOTO
HERE
PLACE
PHOTO
HERE
Jos´e M. F. Moura (S'71 -- M'75 -- SM'90 -- F'94) received degrees from Instituto Superior T´ecnico (IST), Portugal, and
from the Massachusetts Institute of Technology (MIT), Cambridge, MA. He is University Professor at Carnegie Mellon
University (CMU), was on the faculty at IST and has been visiting Professor at MIT. His interests include statistical
and algebraic signal and image processing.
Dr. Moura is Division IX Director of the IEEE, was President of the IEEE Signal Processing Society (SPS), and
Editor in Chief of the IEEE Transactions on Signal Processing. Dr. Moura is a Fellow from IEEE and the AAAS
and an Academy of Sciences of Portugal corresponding member. He has received several awards including the SPS
Technical Achievement Award. In 2010, he was elected University Professor at CMU.
PLACE
PHOTO
HERE
Kavita Ramanan received her Ph.D. from the Division of Applied Mathematics at Brown University in 1998. After a
post-doctoral position at the Technion in Haifa, Israel, Dr. Ramanan joined the Mathematical Sciences Center of Bell
Laboratories, Lucent, as a Member of Technical Staff. In 2003, Dr. Ramanan returned to academia as an Associate
Professor at the Department of Mathematical Sciences at Carnegie Mellon University (CMU), Pittsburgh. In 2010, she
moved to the Division of Applied Mathematics at Brown University where she is a Professor of Applied Mathematics.
Dr. Ramanan works on probability theory, stochastic processes and their applications. Since 2008, she has also been
an adjunct professor at the Chennai Mathematical Institute, India.
|
1804.10781 | 1 | 1804 | 2018-04-28T10:12:47 | The Sharer's Dilemma in Collective Adaptive Systems of Self-Interested Agents | [
"cs.MA"
] | In collective adaptive systems (CAS), adaptation can be implemented by optimization wrt. utility. Agents in a CAS may be self-interested, while their utilities may depend on other agents' choices. Independent optimization of agent utilities may yield poor individual and global reward due to locally interfering individual preferences. Joint optimization may scale poorly, and is impossible if agents cannot expose their preferences due to privacy or security issues. In this paper, we study utility sharing for mitigating this issue. Sharing utility with others may incentivize individuals to consider choices that are locally suboptimal but increase global reward. We illustrate our approach with a utility sharing variant of distributed cross entropy optimization. Empirical results show that utility sharing increases expected individual and global payoff in comparison to optimization without utility sharing. We also investigate the effect of greedy defectors in a CAS of sharing, self-interested agents. We observe that defection increases the mean expected individual payoff at the expense of sharing individuals' payoff. We empirically show that the choice between defection and sharing yields a fundamental dilemma for self-interested agents in a CAS. | cs.MA | cs |
The Sharer's Dilemma in Collective Adaptive
Systems of Self-Interested Agents
Lenz Belzner1, Kyrill Schmid2, Thomy Phan2, Thomas Gabor2, and Martin
Wirsing2
1 MaibornWolff
2 LMU Munich
Abstract. In collective adaptive systems (CAS), adaptation can be im-
plemented by optimization wrt. utility. Agents in a CAS may be self-
interested, while their utilities may depend on other agents' choices. In-
dependent optimization of agent utilities may yield poor individual and
global reward due to locally interfering individual preferences. Joint op-
timization may scale poorly, and is impossible if agents cannot expose
their preferences due to privacy or security issues.
In this paper, we study utility sharing for mitigating this issue. Sharing
utility with others may incentivize individuals to consider choices that are
locally suboptimal but increase global reward. We illustrate our approach
with a utility sharing variant of distributed cross entropy optimization.
Empirical results show that utility sharing increases expected individual
and global payoff in comparison to optimization without utility sharing.
We also investigate the effect of greedy defectors in a CAS of sharing,
self-interested agents. We observe that defection increases the mean ex-
pected individual payoff at the expense of sharing individuals' payoff. We
empirically show that the choice between defection and sharing yields a
fundamental dilemma for self-interested agents in a CAS.
1
Introduction
In collective adaptive systems (CAS), adaptation can be implemented by opti-
mization wrt. utility, e.g. using multi-agent reinforcement learning or distributed
statistical planning [1,2,3,4,5]. Agents in a CAS may be self-interested, while
their utilities may depend on other agents' choices. This kind of situation arises
frequently when agents are competing for scarce resources. Independent opti-
mization of each agent's utility may yield poor individual and global payoff due
to locally interfering individual preferences in the course of optimization [6,7].
Joint optimization may scale poorly, and is impossible if agents do not want to
expose their preferences due to privacy or security issues [8].
A minimal example of such a situation is the coin game [9] (cf. Figure 1.
Here, a yellow and a blue agent compete for coins. The coins are also colored in
yellow or blue. Both agents can decide whether to pick up the coin or not. If both
agents opt to pick up the coin, one of them receives it uniformly at random. If
an agent picks up a coin of its own color, it receives a reward of 2. If it picks up
a differently colored coin, it gets a reward of one. Each agent wants to maximize
its individual reward. If agents act purely self-interested, then each agent tries
to pick up each coin, resulting in suboptimal global reward. However, if rewards
can be shared among agents, then agents will only pick up coins of their own
color. They receive a share that is high enough to compensate for not picking
up differently colored coins. This increases individual and global reward alike.
There are many examples for this kind of situation. For example, energy
production in the smart grid can be modeled in terms of a CAS of self-interested
agents. Each participant has to decide locally how much energy to produce. Each
agent wants to maximize its individual payoff by selling energy to consumers in
the grid. However, the price is depending on global production. Also, global
overproduction is penalized. Routing of vehicles poses similar problems. Each
vehicle wants to reach its destination in a minimal amount of time. However,
roads are a constrained resource, and for a globally optimal solution, only a
fraction of vehicles should opt for the shortest route. In both scenarios, the
ability of agents to share payoff may increase individual and global reward alike.
Fig. 1. Two agents competing for a coin: if agent 1 (yellow) on the left side happens
to get the coin it will get a reward of +1 whereas agent 2 (blue) will get a reward of
+2 for it. If there is a fifty-fifty chance for an agent to get the coin when both agents
are trying to collect it, the expected values are 0.5 for agent 1 and 1 for agent 2 when
both agents independently optimize their utility. In contrast, if there is the possibility
to share reward then agents could learn to do the following: agent 1 (yellow) resists to
collect the coin. That increases the blue agent's probability for getting a reward to 1.
The blue agent transfers reward (e.g. 1) to the yellow agent. This leaves agents with
expected values of 1 each and therefore defines a strong Pareto improvement compared
to the former outcome.
In this paper, we study distributed optimization with utility sharing for miti-
gating the issue of contrasting individual goals at the cost of expected individual
and global reward. To illustrate our ideas, we propose a utility sharing variant of
distributed cross entropy optimization. Empirical results show that utility shar-
ing increases expected individual and global payoff in comparison to optimization
without utility sharing.
+1+2E = 0.5E = 1+1+2E = 1E = 1 We then investigate the effect of defectors participating in a CAS of sharing,
self-interested agents. We observe that defection increases the mean expected in-
dividual payoff at the expense of sharing individuals' payoff. We empirically show
that the choice between defection and sharing yields a fundamental dilemma for
self-interested agents in a CAS.
The paper makes the following contributions.
– We motivate utility sharing as a means to mitigate conflicts and increase
expected individual and global reward in CAS of self-interested agents.
– We propose distributed optimization with sharing (DOS) as an algorithm to
realize utility sharing in self-interested CAS.
– We evaluate DOS empirically, showing that it increases individual and global
reward in expectation.
– We investigate the effect of defecting, non-sharing individuals in a group of
self-interested sharing agents. We show that the choice between defection
and cooperation yields a fundamental dilemma for self-interested agents in
collective adaptive systems.
The remainder of the paper is structured as follows. In Section 2 we discuss
related work. We introduce DOS in Section 3. We discuss our empirical results
and the Sharer's Dilemma in Section 4. We conclude in Section 5.
2 Related Work
In general, we see our work in the context of collective adaptive systems (CAS)
[2,3] and multi-agent systems [10]. In particular, we are interested in CAS where
agents are adaptive through optimization of actions of policies wrt. a given in-
dividual or global utility function. These settings can for example be modeled
in terms of distributed constrained optimization problems [11], or as stochastic
games [12].
Searching for optimal actions or learning policies can be done by open- or
closed-loop planning, potentially enhanced with learned components such as
search exploration policies or value functions [5,13,14,15,16]. Another approach
for learning optimal policies in multi agent domains such as CAS is multi agent
reinforcement learning (MARL) [1,17] and its modern variants based on deep
learning for scaling up to more complex domains[4,18,19]. A recent example
of planning-based deep MARL combines open-loop search and learned value
functions in fully cooperative multi-agent domains [5].
In the case of self-interested agents, the Coco-Q algorithm was proposed [20].
Coco-Q has been evaluated for discrete two-player matrix games, and requires
explicit knowledge of other agents' utilities. In some sense, our study of sharing
in CAS extends the Coco-Q approach to continuous optimization with more
than two agents. Also, we model the amount sharing as a free parameter to be
learned in the course of optimization.
In the context of a research on emergent social effects in MARL [6,7,21,9], a
recent report investigated the effects of inequity aversion and utility sharing in
temporally extended dilemmas [22]. The authors state that "it remains to be seen
whether emergent inequity-aversion can be obtained by evolving reinforcement
learning agents" [22]. Our current work is a first step into this direction, and
shows that the question of whether to share or not poses a dilemma in and
for itself, at least in the case of stateless optimization (in contrast to learning
policies).
3 Distributed Optimization with Sharing
We model decision making in a CAS as a stochastic game (X, N, A, p, R) [12].
– X is a finite set of states.
– N = {0, ..., n} is a finite set of agents.
– A = ×i∈N Ai is a set of joint actions. Ai is a finite set of actions for agent i.
– p(x(cid:48)x, a) is a distribution modeling the probability that executing action
– R = {ri}i∈N , ri : X × A → R is a set of reward functions, one for each agent.
In the following, we assume X = {x} consists of a single state, and ∀a ∈ A :
a ∈ A in state x ∈ X yields state x(cid:48) ∈ X.
p(xx, a) = 1. As x is unique, we will not consider it in further notation.
jointly for all agents, that is maxa∈A : (cid:80)
We assume that ri is available to agent i in terms of a generative model that
may be queried for samples a, e.g. a simulation of the application domain. Each
agent only has access to its own reward function, but does not know the reward
functions of other agents.
The task of a self-interested agent i is to find an action that maximizes its
payoff. However, its payoff ri(a), a ∈ A in general depends on the choices of
other agents. One way to deal with this dependency is to perform optimization
i∈N ri(a). However, in a CAS with
self-interested agents, each participant tries to maximize its individual reward.
Also, in many situations participating agents would not want to expose their
individual reward functions to others due to privacy or security issues [8]. In
these situations, joint optimization wrt. global reward is not feasible. Note that
optimization of self-interested individuals is non-stationary due to changes in
others' choices as they optimize for themselves.
3.1 Reward Sharing
We define agents' utilities as ui. We consider the two different cases we are
interested in:
1. Individual, purely self-interested optimization
2. Self-interested optimization with the option to share individual rewards
Pure Self-Interest When optimizing independently and purely self-interested,
ui(a) = ri(a).
Sharing Sharing agents choose a share si ∈ R, si ≥ 0 additionally to ai. We
denote the joint shares by s = ×i∈N si. Given n agents, a joint action a ∈ A and
a joint share s ∈ Rn, si ≥ 0 for all i, we define individual agents' utility ui for
distributed optimization with sharing as follows. 3
(cid:80)
ui(a, s) = ri(a) − si +
j,j(cid:54)=i sj
n − 1
(1)
Shares are uniformly distributed among all other agents. There are no bilat-
eral shares. Note that this sharing mechanism is an arbitrary choice.
For example, sharing yields the following utilities for two agents.
u0(a, s) = r0(a) − s0 + s1
u1(a, s) = r1(a) − s1 + s0
3.2 Distributed Optimization with Sharing
We now give a general formulation of distributed optimization with sharing
(DOS). DOS is shown in Algorithm 1. Each agent maintains a policy πi(ai),
i.e. a distribution over actions and shares. It is initialized with an arbitrary
prior distribution. A rational agent wants to optimize its policy such that the
expectation of reward is maximized: max Eari(a), where a ∼ ×i∈N πi(ai). Note
that optimization of an individual's policy depends on the policies of all other
agents. Also note that policy optimization of self-interested individuals is non-
stationary due to changes in others' policies as they optimize for themselves.
After initialization, DOS performs the following steps for a predefined number
of iterations.
1. Each agent samples a multiset of nsample actions from its policy and com-
municates it to other agents.
2. A list of joint actions is constructed from the communicated action lists of
other agents.
3. The utility of each joint action is determined according to Equation 1.
4. The policy is updated in a way that increases the likelihood of sampling
high-utility actions and shares.
After niter iterations, each agent samples an action and a share from its
policy, executes the action, and shares reward accordingly. The resulting joint
action yields the global result of DOS.
3.3 Cross-Entropy DOS
In general, DOS is parametric w.r.t. modeling and updating of policies πi. As
an example, we instantiate DOS with cross entropy optimization [23]. We label
this instantiation CE-DOS.
3 We can account for the change of signature of ui by extending the action space Ai
of each agent accordingly: As,i = Ai × R, As = ×i∈N As,i.
each agent samples a list of nsample actions from πi
broadcast sampled actions
for each agent i do
Algorithm 1 Distributed Optimization with Sharing (DOS)
1: initialize πi for each agent i
2: for niter iterations do
3:
for each agent i do
4:
5:
6:
7:
8:
9:
10: for each agent i do
11:
build joint actions a
determine utility ui(a) according to Eq. 1
update πi to increase the likelihood of high-utility samples
execute ai and share si(a) sampled from πi
For CE-DOS, we model a policy π as isotropic normal distribution N (µ, σ).
I.e., each parameter of an action is sampled from a normal distribution that
is independent from other action parameter distributions. Note that it is also
possible to model policies in terms of normal distribution with full covariance,
but the simpler and computationally less expensive isotropic representation suf-
fices for our illustrative concerns. As prior CE-DOS requires initial mean µ0 and
standard deviation σ0 for a policy (cf. Algorithm 2, line 1). I.e. initial actions
before any optimization are sampled as follows.
ai ∼ N (µ0, σ0)
(2)
Updating a policy (cf. Algorithm 1, line 12 - 15) is done by recalculating
mean and variance of the normal distribution. We want the update to increase
the expected sample utility. For each of niter iterations, we sample nsample actions
and shares ai, si ∼ πi from each agent's policy, and build the corresponding joint
actions a = ×i∈N ai and shares s = ×i∈N si.
Each agent evaluates sampled actions and shares according to its utility
ui(a, s). From the set of evaluated samples of each agent, we drop a fraction
ψ ∈ (0, 1] of samples from the set wrt. their utilities. That is, we only keep high
utility samples in the set. We then compute mean and variance of the action pa-
rameters in the reduced set, and use them to update the policy. A learning rate
α ∈ (0, 1] determines the impact of the new mean and variance on the existing
distribution parameters: E.g. let µt and σt be the mean and standard deviation
of a normal distribution modeling a policy at iteration t, then
µt+1 = (1 − α)µt + αµnew
σt+1 = (1 − α)σt + ασnew
where µnew and σnew are mean and standard deviation of the elite samples. We
require a lower bound σmin on the standard deviation of policies in order to
maintain a minimum amount of exploration.
The hyperparameters of CE-DOS are thus as follows.
– A stochastic game (X, N, A, p, R)
– Number of iterations niter
– Number of samples nsample from the policy at each iteration
– Prior mean µ0 and standard deviation σ0 for policies
– Lower bound σmin on the policy standard deviations
– Fraction ψ ∈ (0, 1] of elite samples to keep
– Learning rate α ∈ (0, 1]
sample nsample actions and shares ai, si ∼ πi
clip si such that si ≥ 0
broadcast sampled actions and shares
for each agent i do
Algorithm 2 Cross Entropy DOS
1: Intitialize πi ← N (µ0, σ0) for each agent i
2: for niter iterations do
3:
for each agent i do
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16: for each agent i do
17:
18:
build joint actions a = ×i∈N ai and shares s = ×i∈N si
determine utility ui(a, s) according to Eq. 1
keep ψ · nsample elite samples a, s with highest utility
compute µnew and σnew from ai, si in the elite samples
µt+1 ← (1 − α)µt + αµnew
σt+1 ← (1 − α)σt + ασnew
σt+1 ← max(σt+1, σmin)
πi ← N (µt+1, σt+1)
ai, si ∼ πi
execute ai and share si
4 Experimental Results and the Sharer's Dilemma
We experimentally analyzed the effects of sharing in collective adaptive systems
of self-interested agents.
4.1 Domains
We evaluated the effect of sharing utilities with CE-DOS in two synthetic do-
mains. In these domains, a CAS of self-interested agents has to balance individual
and global resource consumption (or production, respectively).
For example, the energy market in the smart grid can be modeled as a CAS of
self-interested agents. Each participant has to decide locally how much energy to
produce. Each agent wants to maximize its individual payoff by selling energy to
consumers in the grid. Therefore, each agent would like to maximize its individual
energy production. However, the selling price per unit is typically non-linearly
depending on global production. For example, global overproduction is penalized.
There are a number of corresponding real world problems, for example en-
ergy production and consumption in the smart grid, traffic routing, passenger
distribution to individual ride hailing participants, cargo distribution on trans-
port as a service, routing of packets in networks, distribution of computational
load to computers in a cluster, and many more.
We now define two market models (simple and logistic) as domains for eval-
uating the effects of sharing in CAS of self-interested agents.
agents' action space, ai ∈ Ai models the production amount. The sum(cid:80)
Simple Market We model individual and global production, and use their
relation for calculating utilities in such a scenario. We set Ai = R1 as individual
i∈N ai
models the global production.
We define the reward of each agent as the relation of its own individual
resource consumption to the global resource consumption. I.e. the reward cor-
relates to an agents market share. We introduce a slope parameter ξ to control
the utility slope of individual and global consumption.
(cid:16)(cid:80)
ai
j∈N aj
(cid:17)ξ
ri(a) =
(3)
In this setup, a rational agent would like to increase its own consumption
until saturation. I.e. a monopoly is able to produce cheaper than two small
producers, and therefore an inequal production amount unlocks more global
reward. If all agents act rationally by maximizing their individual ai, in general
the corresponding equilibrium is not equal to the global optimum.
Logistic Market We modeled another market scenario for investigating the
effects of sharing in CAS of self-interested agents. As before, each agent has
to choose the amount of energy to use for production of a particular good. I.e.
Ai ∈ [.1, 4], as in the simple market domain. Note that this is an arbitrary choice.
Each agent has a logistic production curve pi : Ai → [0, 1] as a function
of its invested energy. For example, this models different production machine
properties. The logistic curve pi is given as follows.
1
pi(ai) =
1 + e−c(ai−o)
(4)
Here, c ∈ R defines the steepness of the logistic function, and o ∈ R determines
the offset on the x-axis.
Global production prod is the sum of individual production(cid:80)
i pi(ai). A price
function (i.e. an inverse logistic function) defines the price per produced unit,
given global production prod .
price(prod ) = 1 −
1
1 + e−c(prod−o)
(5)
the global price.
ri(a) = pi(ai) · price(prod ) where prod =
pj(aj)
(6)
The reward for an agent is defined as the product of its produced units and
(cid:88)
j∈N
Figure 2 shows an example of production and price functions in the logistic
market domain.
Fig. 2. Example production functions (left) and global price function (right) in the
logistic market domain.
4.2 Setup
For our experiments, we used the following setup of CE-DOS.4
– We consider a stochastic game with n agents, that is N = {1, ..., n}.
– We set n = 10, n = 50 and n = 100 in our experiments.
– Individual action spaces were set as Ai = [.1, 4].
– We define the individual reward functions as given by Equation 3.
– We set the number of iterations niter for CE-DOS to 100.
– We draw nsample = 100 samples from the policy per iteration for each agent.
– Prior mean µ0 and standard deviation σ0 were set to 0 and 1, respectively.
– We set the fraction of elite samples ψ = 0.25.
– We set the learning rate α = 0.5.
– We set the minimal policy standard deviation σmin = 0.2.
We sampled domain parameters uniformely from the following intervals.
– We sampled the slope parameter ξ from [2, 4] in the simple market domain.
– We sampled logistic steepness c and offset o from [1, 3] for all production
and cost functions in our experiments with the logistic market domain.
4 We plan to publish our code upon publication.
We varied the number of sharing agents to measure the effect of defecting
(i.e. non-sharing) agents that participate in the stochastic game together with
sharing individuals.
Note that for the results we report here, we clipped the sharing values such
that agents are only able to share up to their current reward, i.e. si ≤ ri(a) for a
given a ∈ A. In general, other setups with unbound sharing are possible as well.
4.3 Effect of Sharing on Global Reward
Figure 3 shows the mean global utility gathered for varying numbers of shar-
ing agents. We can observe that the fraction of sharing agents correlates with
global utility. We also see that the effect of sharing increases with the number
of participating agents.
Figure 4 shows the mean individual shared value for the corresponding ex-
perimental setups. We can see that the amount of shared value correlates with
global reward. I.e. the more value shared, the higher the global reward. We also
see that the number of participating agents correlates with the effect of sharing.
4.4 Sharer's Dilemma
Figure 5 shows the Schelling diagrams for the corresponding experiments. A
Schelling diagram compares the mean individual utility of sharers and defectors
based on the global number of sharing agents [24]. We can see that agents that
choose to defect gather more individual utility than the sharing ones.
The shape of the Schelling diagrams in Figure 5 shows that sharing in col-
lective adaptive systems with self-interested agents yields a dilemma in our ex-
perimental setups.
Should an individual agent share or defect?
There is no rational answer to this question for an individual self-interested
agent. If the agent chooses to share, it may be exploited by other agents that
are defecting. However, if the agent chooses to defect, it may hurt its individual
return by doing so in comparison to having chosen to share.
Note that the amount of sharing is a free parameter to be optimized by DOS.
This means that all behavior we observe in our experiments is emergent. The
combination of available resources, interdependency of agents' actions and the
ability to share lets agents decide to share with others based on their intrinsic
motivation.
Our results illustrate a potential reason for emergence of cooperation and
inequity aversion in CAS of only self-interested agents. They also give an ex-
planation to the existence of punishment of individuals that exploit societal
cooperation at the cost of sharing individuals' and global reward.
5 Conclusion
We summarize the ideas in this paper, discuss limitations and implications of
our results, and outline venues for further research.
Fig. 3. Global utility gathered for varying numbers of sharing agents in the simple
market (left column) and logistic market (right column) domains. 10 agents (top row),
50 agents (center row) and 100 agents (bottom row) in total. Solid line shows empirical
mean of 10 experimental runs, shaded areas show .95 confidence intervals. Best viewed
on screen in color.
Fig. 4. Mean individual shares for varying numbers of sharing agents in the simple
market (left column) and logistic market (right column) domains. 10 agents (top row),
50 agents (center row) and 100 agents (bottom row) in total. Solid line shows empirical
mean of 10 experimental runs, shaded areas show .95 confidence intervals. Best viewed
on screen in color.
Fig. 5. Schelling diagrams showing mean individual utility for defectors and sharers,
for varying numbers of sharing agents in the simple market (left column) and logistic
market (right column) domains. Note the log scale on the y-axis. 10 agents (top row),
50 agents (center row) and 100 agents (bottom row) in total. 10 experimental runs.
Best viewed on screen in color.
5.1 Summary
In collective adaptive systems (CAS), adaptation can be implemented by opti-
mization wrt. utility. Agents in a CAS may be self-interested, while their utilities
may depend on other agents' choices. Independent optimization of each agent's
utility may yield poor individual and global payoff due to locally interfering in-
dividual preferences in the course of optimization. Joint optimization may scale
poorly, and is impossible if agents do not want to expose their preferences due
to privacy or security issues.
In this paper, we studied distributed optimization with sharing for mitigat-
ing this issue. Sharing utility with others may incentivize individuals to consider
choices that are locally suboptimal but increase global reward. To illustrate our
ideas, we proposed a utility sharing variant of distributed cross entropy optimiza-
tion. Empirical results show that utility sharing increases expected individual
and global payoff in comparison to optimization without utility sharing.
We also investigated the effect of defectors participating in a CAS of shar-
ing, self-interested agents. We observed that defection increases the mean ex-
pected individual payoff at the expense of sharing individuals' payoff. We empir-
ically showed that the choice between defection and sharing yields a fundamental
dilemma for self-interested agents in a CAS.
5.2 Limitations
A central limitation of CE-DOS is its state- and memoryless optimization. In
our formulation of utility sharing self-interested agents optimize an individual
action and share that maximizes their utility. However, our formulation does not
account for learning decision policies based on a current state and other learning
agents. In this case, the utility of each agent would also depend on concrete
states, transition dynamics and potentially also on models agents learn about
other participants [25,26].
As there is no temporal component to the optimization problems that we
studied in this paper, it is also not possible to study the effect of gathering
wealth in our current setup. We think that the dynamics of sharing in temporally
extended decision problems may differ from the ones in stateless optimization.
For example, corresponding observations have been made for game theoretic
dillemas, where optimal strategies change when repeating a game (in contrast to
the optimal strategy when the game is only played once) [27]. Similar research
has been conducted in the field of reinforcement learning, however not accounting
for utility sharing so far [6].
We also want to point out that exposing shares eventually provides ground
for attack for malicious agents [8]. Albeit indirectly, exposed shares carry in-
formation about individual utility landscapes, allowing attackers to potentially
gather sensitive information about agents' internal motivations. Agents in criti-
cal application domains should consider this weakness when opting to share.
5.3 Future Work
In future work, we would like to transfer our approach to temporally extended
domains and model sharing in CAS with multi-agent reinforcement learning.
Hopefully, this would enable studying sharing and the Sharer's Dilemma in more
complex domains.
We also think that there are many interesting options for realizing sharing
besides equal distribution as formulated in Eq. 1. For example, our formulation
does not allow for bilateral shares or formation of coalitions. Also, we would be
interested to study the effect of wealth on emergent cooperation and defection.
Another interesting line would be to investigate the effects of punishment in
CAS of self-interested agents.
As an application domain, it would be interesting to exploit the duality of
planning and verification. For example, agents utility could model individual
goal satisfaction probability. Sharing could be used to increase individual and
global goal satisfaction probability in CAS.
References
1. Tan, M.: Multi-agent reinforcement learning: Independent vs. cooperative agents.
In: Proceedings of the tenth international conference on machine learning. (1993)
330–337
2. Hillston, J., Pitt, J., Wirsing, M., Zambonelli, F.: Collective adaptive systems:
qualitative and quantitative modelling and analysis (dagstuhl seminar 14512). In:
Dagstuhl Reports. Volume 4., Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik
(2015)
3. Belzner, L., Holzl, M., Koch, N., Wirsing, M.: Collective autonomic systems: To-
wards engineering principles and their foundations. In: Transactions on Founda-
tions for Mastering Change I. Springer (2016) 180–200
4. Foerster, J., Nardelli, N., Farquhar, G., Torr, P., Kohli, P., Whiteson, S., et al.:
Stabilising experience replay for deep multi-agent reinforcement learning. arXiv
preprint arXiv:1702.08887 (2017)
5. Phan, T., Belzner, L., Gabor, T., Schmid, K.: Leveraging statistical multi-agent
online planning with emergent value function approximation. In: Proceedings of the
17th Conference on Autonomous Agents and Multi Agent Systems, International
Foundation for Autonomous Agents and Multiagent Systems (2018)
6. Leibo, J.Z., Zambaldi, V., Lanctot, M., Marecki, J., Graepel, T.: Multi-agent re-
inforcement learning in sequential social dilemmas. In: Proceedings of the 16th
Conference on Autonomous Agents and Multi Agent Systems, International Foun-
dation for Autonomous Agents and Multiagent Systems (2017) 464–473
7. Perolat, J., Leibo, J.Z., Zambaldi, V., Beattie, C., Tuyls, K., Graepel, T.: A multi-
agent reinforcement learning model of common-pool resource appropriation. In:
Advances in Neural Information Processing Systems. (2017) 3646–3655
8. Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., Dafoe, A.,
Scharre, P., Zeitzoff, T., Filar, B., et al.: The malicious use of artificial intelligence:
Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228 (2018)
9. Lerer, A., Peysakhovich, A.: Maintaining cooperation in complex social dilemmas
using deep reinforcement learning. arXiv preprint arXiv:1707.01068 (2017)
10. Van der Hoek, W., Wooldridge, M.: Multi-agent systems. Foundations of Artificial
Intelligence 3 (2008) 887–928
11. Fioretto, F., Pontelli, E., Yeoh, W.: Distributed constraint optimization problems
and applications: A survey. arXiv preprint arXiv:1602.06347 (2016)
12. Shapley, L.S.: Stochastic games. Proceedings of the national academy of sciences
39(10) (1953) 1095–1100
13. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al.: Master-
ing the game of go with deep neural networks and tree search. nature 529(7587)
(2016) 484–489
14. Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanc-
tot, M., Sifre, L., Kumaran, D., Graepel, T., et al.: Mastering chess and shogi
by self-play with a general reinforcement learning algorithm.
arXiv preprint
arXiv:1712.01815 (2017)
15. Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A.,
Hubert, T., Baker, L., Lai, M., Bolton, A., et al.: Mastering the game of go without
human knowledge. Nature 550(7676) (2017) 354
16. Anthony, T., Tian, Z., Barber, D.: Thinking fast and slow with deep learning
and tree search. In: Advances in Neural Information Processing Systems. (2017)
5366–5376
17. Littman, M.L.: Markov games as a framework for multi-agent reinforcement learn-
ing. In: Machine Learning Proceedings 1994. Elsevier (1994) 157–163
18. 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. (2016) 2137–2145
19. Tampuu, A., Matiisen, T., Kodelja, D., Kuzovkin, I., Korjus, K., Aru, J., Aru,
J., Vicente, R.: Multiagent cooperation and competition with deep reinforcement
learning. PloS one 12(4) (2017) e0172395
20. Sodomka, E., Hilliard, E., Littman, M., Greenwald, A.: Coco-q: Learning in
In: International Conference on Machine
stochastic games with side payments.
Learning. (2013) 1471–1479
21. Peysakhovich, A., Lerer, A.: Prosocial learning agents solve generalized stag hunts
better than selfish ones. arXiv preprint arXiv:1709.02865 (2017)
22. Hughes, E., Leibo, J.Z., Philips, M.G., Tuyls, K., Du´enez-Guzm´an, E.A.,
Castaneda, A.G., Dunning, I., Zhu, T., McKee, K.R., Koster, R., et al.: Inequity
aversion resolves intertemporal social dilemmas. arXiv preprint arXiv:1803.08884
(2018)
23. Kroese, D.P., Rubinstein, R.Y., Cohen, I., Porotsky, S., Taimre, T.: Cross-entropy
In: Encyclopedia of Operations Research and Management Science.
method.
Springer (2013) 326–333
24. Schelling, T.C.: Hockey helmets, concealed weapons, and daylight saving: A study
of binary choices with externalities. Journal of Conflict resolution 17(3) (1973)
381–428
25. Foerster, J.N., Chen, R.Y., Al-Shedivat, M., Whiteson, S., Abbeel, P., Mordatch,
I.: Learning with opponent-learning awareness. arXiv preprint arXiv:1709.04326
(2017)
26. Rabinowitz, N.C., Perbet, F., Song, H.F., Zhang, C., Eslami, S., Botvinick, M.:
Machine theory of mind. arXiv preprint arXiv:1802.07740 (2018)
27. Sandholm, T.W., Crites, R.H.: Multiagent reinforcement learning in the iterated
prisoner's dilemma. Biosystems 37(1-2) (1996) 147–166
|
1810.11634 | 1 | 1810 | 2018-10-27T12:33:03 | Agent-based models of collective intelligence | [
"cs.MA",
"cs.SI",
"nlin.AO"
] | Collective or group intelligence is manifested in the fact that a team of cooperating agents can solve problems more efficiently than when those agents work in isolation. Although cooperation is, in general, a successful problem solving strategy, it is not clear whether it merely speeds up the time to find the solution, or whether it alters qualitatively the statistical signature of the search for the solution. Here we review and offer insights on two agent-based models of distributed cooperative problem-solving systems, whose task is to solve a cryptarithmetic puzzle. The first model is the imitative learning search in which the agents exchange information on the quality of their partial solutions to the puzzle and imitate the most successful agent in the group. This scenario predicts a very poor performance in the case imitation is too frequent or the group is too large, a phenomenon akin to Groupthink of social psychology. The second model is the blackboard organization in which agents read and post hints on a public blackboard. This brainstorming scenario performs the best when there is a stringent limit to the amount of information that is exhibited on the board. Both cooperative scenarios produce a substantial speed up of the time to solve the puzzle as compared with the situation where the agents work in isolation. The statistical signature of the search, however, is the same as that of the independent search. | cs.MA | cs | Agent-based models of collective intelligence
Sandro M. Reia,1 Andr´e C. Amado,2 and Jos´e F. Fontanari1
1Instituto de F´ısica de Sao Carlos, Universidade de Sao Paulo,
Caixa Postal 369, 13560-970 Sao Carlos, Sao Paulo, Brazil
2Departamento de F´ısica, Universidade Federal de Pernambuco, 50670-901, Recife, PE, Brazil
Collective or group intelligence is manifested in the fact that a team of cooperating agents can
solve problems more efficiently than when those agents work in isolation. Although cooperation is,
in general, a successful problem solving strategy, it is not clear whether it merely speeds up the time
to find the solution, or whether it alters qualitatively the statistical signature of the search for the
solution. Here we review and offer insights on two agent-based models of distributed cooperative
problem-solving systems, whose task is to solve a cryptarithmetic puzzle. The first model is the
imitative learning search in which the agents exchange information on the quality of their partial
solutions to the puzzle and imitate the most successful agent in the group. This scenario predicts a
very poor performance in the case imitation is too frequent or the group is too large, a phenomenon
akin to Groupthink of social psychology. The second model is the blackboard organization in which
agents read and post hints on a public blackboard. This brainstorming scenario performs the best
when there is a stringent limit to the amount of information that is exhibited on the board. Both
cooperative scenarios produce a substantial speed up of the time to solve the puzzle as compared with
the situation where the agents work in isolation. The statistical signature of the search, however, is
the same as that of the independent search.
I.
INTRODUCTION
The development of software packages to tackle almost
any technical problem we can think of, as well as of
friendly interfaces between people and computers (see,
e.g., [1] in this issue), has contributed to the tremendous
increase of the productivity of today's society, thus rati-
fying the 1960s vision of the computer pioneer Doug En-
gelbart that the human intellect could be augmented with
the aid of computers [2]. In fact, if human intelligence
is gauged by the capacity and speed to solve problems,
which seems a sensible perspective for the scientific and
technological milieus, then it is difficult to disagree with
Engelbart's rationale.
However, since the general intelligence of a person --
the g factor -- has strong genetic (about 50%) and devel-
opmental (the remaining 50%) components there is little
adults can do to increase their intelligence (see [3] for a
lucid discussion of these controversial issues), thus lim-
iting the growth of the most important element of the
human-computer partnership. Although the natural and
widely employed way to circumvent this limitation is to
consider group work, only recently psychologists have put
forward evidence supporting a general collective intelli-
gence factor -- the so-called c factor -- that explains the
group performance on a variety of tasks [4]. Surprisingly,
the c factor does not seem to be strongly correlated with
the average or maximum individual intelligence of group
members. It is correlated instead with the average social
sensitivity of the group members [4]. Therefore, we could,
in principle, augment the group intelligence by properly
selecting the group composition and internal organiza-
tion.
Actually, the study of the influence of the organiza-
tion and,
in particular, of intra-group communication
patterns (i.e., who can communicate with whom) on
the problem-solving performance of groups dates back at
least to the 1950s [5 -- 7] (see [8 -- 10] for more recent con-
tributions). Understanding this influence is, of course, of
immense value because problem solving (e.g., drug de-
sign, traffic engineering, software development) by task
forces represents a substantial portion of the economy of
developed countries nowadays [11].
As argued above, we see organizational design as a by-
product of the study of collective intelligence or, more
specifically, of distributed cooperative problem-solving
systems. The key feature of these systems is that their
members exchange information about their progress to-
wards the completion of a goal [12, 13]. There are many
common-sense assumptions in this field, e.g., that a group
of cooperating individuals is more efficient than those
same individuals working in isolation or that diversity
is always beneficial to the group performance, that were
not fully scrutinized through a powerful (in the explana-
tory sense) analytical tool of physics, namely, the math-
ematical and computational modeling of complex phe-
nomena using minimal models. A minimal model should
exhibit a good balance between simplicity and realism
and should be successful in reducing a complex collective
phenomenon to its functional essence (see [14, 15] for use
of this approach to elucidate the physics of mind).
In fact, despite the extensive use of optimization
heuristics inspired on cooperative systems, such as the
particle swarm optimization algorithm [16] and the adap-
tive culture heuristic [17, 18], to search for optimal or
near optimal solutions of combinatorial problems, we
know little about the factors that make cooperation ef-
fective, as well as about the universal character (if any) of
the quantitative improvements that results from it [13].
This is so because those heuristics and the problems they
are set to solve are too complex to yield to a first-principle
analysis. Here we review recent attempts to study dis-
8
1
0
2
t
c
O
7
2
]
A
M
.
s
c
[
1
v
4
3
6
1
1
.
0
1
8
1
:
v
i
X
r
a
tributed cooperative problem-solving systems using min-
imal models following the research strategy set forth by
Huberman in the late 1980s [12, 13, 19]. We consider
two scenarios of distributed cooperative systems where
the goal of the agents is to solve a particular cryptarith-
metic puzzle [see eq. 1)], which is an illustrative example
of the class of constraint satisfaction problems that bases
most studies on problem solving [20].
In the first scenario we endow the agents with the ca-
pacity to evaluate the goodness of the partial solutions of
their peers and imitate the more successful agent in the
group. In the context of collective intelligence or global
brains, imitative learning is probably the most important
factor as neatly expresses this quote by Bloom "Imitative
learning acts like a synapse, allowing information to leap
the gap from one creature to another" [21]. Since imi-
tation is central to the remarkable success of our species
[22, 23] (see [24] for a discussion of this issue), we ex-
pect that this scenario may be of relevance to the or-
ganization of real-world task groups [25].
In fact, we
found that if the agents are too propense to imitate their
more successful peers or if the group is too large then
the group performance is catastrophic when compared
with the baseline situation where the agents work inde-
pendently of each other [26, 27]. This is similar to the
classic Groupthink phenomenon of social psychology that
occurs when everyone in a group starts thinking alike [28].
Avoiding this sort of harmful effect is the task of orga-
nizational designers and we have verified that two rather
natural interventions are partially effective to circumvent
Groupthink, namely, decreasing the connectivity of the
agents so as to delay the propagation of misleading infor-
mation through the system [29], and allowing diversity in
the agents' propensities to imitate their peers [30]. How-
ever, these interventions have an unwelcome side effect:
the degradation of the optimal performance of the group,
which is achieved in the case of fully connected homoge-
neous agents.
The
second scenario of distributed cooperative
problem-solving systems that we consider here is the
blackboard organization that was introduced in the Arti-
ficial Intelligence domain in the 1980s and is now part of
the AI problem-solving toolkit [31]. In this organization,
the agents read and write hints to a central blackboard
that can be accessed by all members of the group. Hence
the blackboard scenario describes the common view of
task-forces as teams of specialists exchanging ideas on
possible approaches to solve a problem and displaying
the promising suggestions in a public blackboard. In this
scenario, there is no need to assume that the agents are
capable of quantifying the goodness of their partial so-
lutions to the cryptarithmetic puzzle as in the imitative
learning scenario (see, however, [32]). To our knowledge,
this was the first distributed cooperative problem-solving
system studied using an agent-based minimal model [13].
Contrary to the claims of that original study, however, we
found that the search using the blackboard organization
exhibits the same statistical signature of the independent
2
search [32]. Most unexpectedly, we found that limiting
the amount of information displayed on the board can
markedly boost the performance of the group. This is an
original result that we offer in this review paper, which
illustrates well the power of minimal models to reveal
relevant hidden features of complex systems.
The rest of this short review paper is organized as fol-
lows. In Section II we present the particular cryptarith-
metic problem that we use throughout the paper, de-
fine the costs associated to the digit-to-letter assignments
that are necessary to implement the imitative learning
search strategy, and introduce the definition of hint that
is necessary for implementing the blackboard organiza-
tion. In Section III we present our measure of the group
performance, which is proportional to the time for an
agent in the group to find the solution and hence to halt
the search. By dividing this time by the size of the state
space of the problem and multiplying it by the number
of agents in the group we obtain the computational cost
of the search, which is our performance measure. In Sec-
tion IV we describe the imitative learning search strategy
and discuss its performance on the cryptarithmetic puz-
zle for the simplest case where all agents interact with
each other, i.e., the pattern of communication is a fully
connected network. In Section V we present the minimal
model for the blackboard organization and show that the
probability distribution of the computational costs is an
exponential distribution as in the case of the independent
search. The mean computational cost of the search for
the standard blackboard organization, however, is about
ten times lower than for the independent search. Also in
that section, we introduce the limited space blackboard
scenario where the agents have to compete for space in
the board. In this case, the computational cost can be
reduced by another factor of ten using a judicious choice
of the blackboard size. Finally, Section VI is reserved to
our concluding remarks and to advance some avenues for
future research.
II. THE CRYPTARITHMETIC PROBLEM
Cryptarithmetic problems such as
DON ALD + GERALD = ROBERT
(1)
are constraint satisfaction problems in which the task is
to find unique digit-to-letter assignments so that the in-
teger numbers represented by the words add up correctly
[33]. In the cryptarithmetic problem (1), there are 10! dif-
ferent digit-to-letter assignments, of which only one is the
solution to the problem, namely, A = 4, B = 3, D = 5,
E = 9, G = 1, L = 8, N = 6, O = 2, R = 7, T = 0
so that DON ALD = 526485, GERALD = 197485 and
ROBERT = 723970. In this paper we will focus only on
the cryptarithmetic problem (1) because its state space is
the largest possible for this type of puzzle (a cryptarith-
metic puzzle has at most 10 different letters) and because
it facilitates the replication of our findings. Use of ran-
domly generated cryptarithmetic puzzles as well as dis-
tinct optimization problems, such as finding the global
maximum of NK-fitness landscapes [34], has yielded the
same (qualitative) results [26, 27].
The imitative learning search strategy that will be dis-
cussed in Sect. IV requires that we assign a cost to each
digit-to-letter assignment, which is viewed as a measure
of the goodness of the answer represented by that assign-
ment. A natural choice for the cost function is [35]
c = ROBERT − (DON ALD + GERALD) .
(2)
For example, the digit-to-letter assignment A = 0, B = 2,
D = 9, E = 4, G = 8, L = 1, N = 7, O = 6, R = 3, T =
5 yields ROBERT = 362435, DON ALD = 967019 and
GERALD = 843019 and the cost assigned to it is c =
1447603. We should note that the cost value (2) applies
to all digit-to-letter assignments except those for which
R = 0, D = 0 and G = 0, which are invalid assignments
since they violate the rule of the cryptarithmetic puzzles
that an integer number should not have the digit 0 at
its leftmost position. Hence for those assignments we fix
an arbitrary large cost value, namely, c = 108, so that
now they become valid assignments but have the highest
cost among all assignments.
If the cost of a digit-to-
letter assignment is c = 0 then it is the solution to the
cryptarithmetic problem.
Although we could easily think up clever alternatives
to the cost function (2), we recall that our aim is not to
design efficient algorithms to solve cryptarithmetic prob-
lems but to explore cooperative strategies that improve
the efficiency of group work [12]. In that sense, the cho-
sen problem (1) is quite challenging in that it offers many
misleading clues -- local minima of the cost (2) and wrong
hints -- which may lure the search away from the solution.
However, as already mentioned, we stress that the main
advantage of considering a specific problem is the easy to
replicate and verify our claims.
At this stage it is convenient to describe the minimal
or elementary move in the state space composed of the
10! possible digit-to-letter assignments. Starting from a
particular digit-to-letter assignment, say, A = 0, B = 2,
D = 9, E = 4, G = 8, L = 1, N = 7, O = 6, R = 3,
T = 5 we choose two different letters at random and in-
terchange the digits assigned to them. For example, if
we pick letters D and T then the assignment that results
from the application of the elementary move is A = 0,
B = 2, D = 5, E = 4, G = 8, L = 1, N = 7, O = 6,
R = 3, T = 9. Any two valid digit-to-letter assignments
that are connected by the elementary move are said to
be neighbor assignments. Hence each digit-to-letter as-
signment in the state space of problem (1) has exactly 45
neighbors. Clearly, the repeated application of our ele-
mentary move allows us to explore the entire state space
of the cryptarithmetic problem. We can check all assign-
ments and their neighbors to find the number of min-
ima, i.e., those assignments that have a cost (strictly)
lower than the cost of their neighbors. We find that the
3
cryptarithmetic problem (1) has 102 minima in total: a
single global minimum and 101 local minima. We recall,
however, that the existence and characteristics of the lo-
cal minima are strongly dependent on the choices of the
cost function and of the elementary move in the state
space.
An important feature of cryptarithmetic puzzles,
which makes them a testbed for cooperative strategies, is
the existence of hints that may hint on the suitability of
a particular digit-to-letter assignment. A hint is a set of
letters in a same column that add up correctly modulo
10. For example, considering the third column (from left
to right) of the problem (1) we have B = +N +R where
= 0, 1 and the sum is done modulo 10. The case = 1
accounts for the possibility that an 1 is carried from the
sum of the letters in the fourth column. Of course, for
the rightmost column (D + D = T ) the only possibility
is = 0. For this column there are 9 different hints:
(D = 1, T = 2), (D = 2, T = 4), (D = 3, T = 6), etc.
Each one of the columns + L + L = R, + A + A = E
and + O + E = O has 18 different hints (9 for = 0
and 9 for = 1), whereas columns + N + R = B and
+ D + G = R have 144 different hints (72 for = 0
and 72 for = 1) each. Hence there are a total of 351
distinct hints but only 6 of them yield the solution of
the puzzle (1). We note that we could further reduce
this number by eliminating hints in the leftmost column
such that + D + G > 9 with the sum now done modulo
1. However, we choose not to implement this rule since
what matters is that the total number of hints is much
smaller than the size of the state space and that the six
correct hints are contemplated in our definition of hint.
In the blackboard organization, the communication be-
tween the agents is achieved by posting and reading hints
in a public blackboard [13] and so there is no need to in-
troduce a cost for each digit-to-letter assignment (see [32]
for a scheme where the hints are displayed together with
the costs of the agents that posted them).
III. THE COMPUTATIONAL COST
The efficiency of the search for the solution of the
cryptarithmetic puzzle is measured by the computational
cost that is defined as follows. Let us consider a group
composed of M agents so that each agent is represented
by a digit-to-letter assignment. The agents explore the
state space following a search strategy that specifies the
rules for updating their digit-to-letter assignments. Each
time a randomly chosen agent updates its digit-to-letter
assignment we increment the time t by the quantity
∆t = 1/M , so that during the increment from t to t + 1
exactly M , not necessarily distinct, agents are updated.
The search ends when one of the agents finds the solu-
tion to the puzzle and we denote by t∗ the time when
this happens. Since we expect that t∗ will increase with
the size of the state space (10! for our cryptarithmetic
puzzle) and that, at least for the independent search, it
will decrease with the reciprocal of the number of agents,
we define the computational cost C of the search as
C = M t∗/10!
(3)
so that C is on the order of 1 for the independent search,
regardless of the group size. Next we study the statistical
properties of the computational cost for two cooperative
problem-solving scenarios, namely, the imitative learning
search and the blackboard organization.
IV. THE IMITATIVE LEARNING SEARCH
This search strategy is based on the presumption that
the cost (2) offers a clue on the goodness of the digit-to-
letter assignment so that it may be advantageous to copy
or imitate agents whose assignments have low cost. More
pointedly, at time t, a randomly chosen agent -- the target
agent -- can choose between two actions. The first action,
which happens with probability 1−p, is the elementary or
minimal move in the space space described in Section II.
The second action, which happens with probability p, is
the imitation of the model agent, which is the agent with
the lowest cost digit-to-letter assignment in the group at
time t. To illustrate the copying process let us assume
for the sake of concreteness that the target agent has
the assignment A = 0, B = 2, D = 9, E = 4, G = 8,
L = 1, N = 7, O = 6, R = 3, T = 5 whose cost is
c = 1447603 and the model agent has the assignment
A = 5, B = 3, D = 9, E = 4, G = 8, L = 1, N = 6,
O = 2, R = 7, T = 0 whose cost is c = 1050568. In the
copying process the target agent selects at random one
of the distinct digit-to-letter assignments in the model
agent and assimilates it.
In our example, the distinct
assignments occur for the letters A, B, N , O, R and
T . Say that letter B is chosen, so that the target agent
has to assimilate the assignment B = 3. To do that the
target agent simply interchanges the digits assigned to
the letters B and R, as in the elementary move, so that
the resulting assignment becomes A = 0, B = 3, D = 9,
E = 4, G = 8, L = 1, N = 7, O = 6, R = 2, T = 5
whose cost is c = 1545613. As expected, the result of
imitation is the increase of the similarity between the
target and the model agents, which may not necessarily
lead to a decrease of the cost of the target agent, as in
our example. The case p = 0 corresponds to the baseline
situation where the M agents explore the state space
independently.
It is important to note that in the case the target agent
is identical to the model agent, and this situation is not
uncommon since the imitation process reduces the diver-
sity of the group, the target agent executes the elemen-
tary move with probability one. This procedure is differ-
ent from that used in [26], in which agents identical to
the model agent are not updated in the imitation action.
Both implementations yield qualitatively similar results,
except in the regime where imitation is extremely fre-
quent, i.e., for p ≈ 1. In particular, for p = 1 the imple-
4
FIG. 1. Mean computational cost (cid:104)C(cid:105) as function of the imita-
tion probability p for groups of size (top to bottom at p = 0.2)
M = 2, 5, 8, 25 and 200. For the independent search (p = 0)
we find (cid:104)C(cid:105) ≈ 1.02 regardless of the group size.
mentation in which the model agent is unchanged results
in the search being permanently stuck in a local mini-
mum [26], whereas in the implementation in which the
model agent executes the elementary move actually leads
to the optimal performance for very small groups, as we
will show next. The procedure adopted here was used in
most studies of the imitative learning search [27, 36].
Figure 1 shows the mean computational cost (cid:104)C(cid:105) as
function of the imitation probability p obtained by av-
eraging over 105 independent runs. For groups of size
M < 5 the performance always improves with increas-
ing p and for those small groups the strategy of always
imitating the lowest cost agent (i.e., p = 1) is optimal.
For M = 5 there appears a minimum at p ≈ 0.96 with
computational cost (cid:104)C(cid:105) ≈ 0.032 that corresponds to the
best performance of the imitative learning search for the
entire space of the model parameters M and p. This
amounts to more than a thirtyfold improvement on the
group performance as compared with the independent
search. For large groups, increase of the imitation prob-
ability p can lead to catastrophic results due to the trap-
ping of the search around the local minima. This harmful
effect appears in large groups only and it is due to the
existence of several copies of the model agent carrying
a low cost digit-to-letter assignment (local minimum).
This makes it very hard to explore other regions of the
state space through the elementary move, since the ex-
tra copies attract the updated model agent back to the
local minimum. However, for small group sizes, the ele-
mentary move can easily carry the agents away from the
local minima as illustrated in Fig. 1. The optimal perfor-
mance for M > 5, which is determined by the minimum
of the curve (cid:104)C(cid:105) vs. p, degrades smoothly with increas-
ing M . We note that when the model parameters are
close to their optimal values, the distribution of proba-
bility of the computational costs is well-described by an
exponential distribution [26].
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1<C>p 5
imitative learning search is markedly different from the
well-known genetic algorithms [39]. In fact, the elemen-
tary move can be seen as the counterpart of mutations
with the caveat that in the genetic algorithm mutation
is an error of the reproduction process, whereas in the
imitative search the elementary move and the imitation
procedure are mutually exclusive processes. The analogy
between the imitation and the crossover processes is even
more far-fetched. The model agent is a mandatory par-
ent in all mates but it contributes a single gene (i.e., a
single digit to letter assignment) to the offspring which
then replaces the other parent, namely, the target agent.
Since the contributed gene is not random - it must be
absent in the target agent - the genetic analogy is clearly
inappropriate and so the imitative learning search stands
on its own as a search strategy.
V. THE BLACKBOARD ORGANIZATION
The popular view of working groups as teams of spe-
cialists that exchange ideas on possible approaches to
solve a problem and write the promising lines of investi-
gation in a public display is the inspiration for the black-
board organization [31]. The study of a minimal model
of these brainstorming groups, which considers M agents
and a central blackboard where the agents can read and
write hints, suggested that the blackboard organization
could produce a superlinear speedup of the solution time
t∗ (see Section III) with respect to the number of group
members M [13]. We recall that for the independent
search the speedup is linear, i.e., t∗ ∝ 1/M , provided
that M is not too large in order to avoid duplication of
work. In our problem, duplication of work will occur for
unrealistically large groups, M (cid:29) 10!, only. If the super-
linear speedup claim were correct, then it would offer a
nice qualitative evidence of the benefits of cooperation to
problem-solving systems. However, recent evidences in-
dicate that whereas the blackboard organization actually
produces a significant improvement on the performance
of the search, it does not change the nature of the search
which exhibits the same characteristics of the indepen-
dent search and, in particular, the same scaling of t∗
with M [32].
An advantage of blackboard systems is that they do
not need the introduction of arbitrary cost functions to
weight the quality of the digit-to-letter assignments. As it
will be clear in our analysis of limited space blackboards,
the number of hints exhibited by an assignment is an
effective, albeit indirect, measure of its quality. A limited
space blackboard can exhibit at most B hints so that,
when the blackboard is full, the agents must erase hints
to make room for their own hints on the board. Next we
describe the dynamics of the blackboard organization.
At the initial time, t = 1, all agents' digit-to-letter
assignments are selected with equal probability from the
pool of the 10! valid assignments. Each agent then checks
for all possible hints of its digit-to-letter assignment (see
FIG. 2. Mean computational cost (cid:104)C(cid:105) as function of the group
size M for imitation probability p = 0.5 (circles), 0.6 (trian-
gles), 0.7 (inverted triangles) and 1 (squares). The horizontal
line at (cid:104)C(cid:105) = 1.02 indicates the mean cost for the independent
search (p = 0).
Figure 2 illustrates more neatly the existence of a group
size that optimizes the performance of the group for a
fixed imitation probability. Hence the conjecture that
the efficacy of imitative learning could be a factor de-
terminant of the group size of social animals [26] (see
[37, 38] for a discussion of the standard selective pres-
sures on group size in nature). As pointed out before,
the best performance overall is achieved for small groups
with high imitation probability.
The relevant finding revealed in Figs. 1 and 2 is that
the group performance can be optimized by tuning the
two parameters of the model, namely, the group size M
and the imitation probability p. Hence collective intelli-
gence can be augmented through a judicious choice of the
behavioral characteristics of the group members, mod-
eled here by the imitation probability p, and of the group
size M . As already pointed out, the catastrophic per-
formance exhibited by large groups and high imitation
probabilities is akin to the Groupthink phenomenon [28],
when everyone in a group starts thinking alike, which
can occur when people put unlimited faith in a talented
leader (the model agent, in our case). In addition to us-
ing small group sizes, there are two other ways to avoid
the trapping in the local minima. The first way is de-
creasing or delaying the influence of the model agent by
reducing the connectivity of the network [29]. The second
is allowing some diversity in the imitation probabilities
of the agents since agents characterized by p ≈ 0 will
rarely be trapped in the local minima [30]. Although
these solutions are effective in avoiding the catastrophic
performance for large M and p they have the unwanted
side effect of degrading the optimal performance, which is
obtained using a fully connected influence network with
homogeneous imitation probabilities, as in the model de-
scribed before.
We note that, despite some superficial similarities, the
0 0.2 0.4 0.6 0.8 101020304050607080<C>M Section II), singles out the novel hints (i.e., the hints
that are not already displayed on the blackboard) and
chooses one of them at random to post on the board. In
this process, the agent also makes a list of the hints that
are displayed on the board and that do not appear in its
digit-to-letter assignment. Those are the different hints.
Let us assume that the limited space board of size B is
full, so the agent must make room to post its selected
hint. To do so the agent selects one of the different hints
on the board at random and replaces it by its hint. In
case the board is not full, the agent simply posts the
selected hint on the board. We have tested many variants
of this pick-and-replace procedure and found that they
produce only negligible quantitative changes on the group
performance and so do not affect our conclusions.
Once the initial states of the agents and of the black-
board are set up, the agents can update their digit-to-
letter assignments by performing two actions: the ele-
mentary move described in Section II and the assimila-
tion of one of the hints displayed on the blackboard. The
search procedure develops as follows.
It begins with a
randomly chosen agent -- the target agent -- picking a hint
at random from the blackboard. In the case that there
are no hints (i.e, the blackboard is empty), or that the
target agent is already using the chosen hint, the agent
performs the elementary move; otherwise it assimilates
the hint. The assimilation of a hint by the target agent
involves the relocation of at most six digits of its digit-to-
letter assignment. For example, consider the assimilation
of the hint (N = 1, R = 4, B = 5) by an agent that has
the assignment A = 0, B = 2, D = 9, E = 4, G = 8,
L = 1, N = 7, O = 6, R = 3, T = 5. This can be done
sequentially using the same assimilation procedure of the
imitative learning search described in Section IV. First,
the assignment N = 1 is assimilated, yielding A = 0,
B = 2, D = 9, E = 4, G = 8, L = 7, N = 1, O = 6,
R = 3, T = 5, then R = 4, yielding A = 0, B = 2, D = 9,
E = 3, G = 8, L = 7, N = 1, O = 6, R = 4, T = 5 and
finally B = 5 resulting in the digit-to-letter assignment
A = 0, B = 5, D = 9, E = 3, G = 8, L = 7, N = 1,
O = 6, R = 4, T = 2 that exhibits the desired hint. As
usual, after the target agent is updated, we increment
the time t by the quantity ∆t = 1/M .
After any of the events -- elementary move or assimi-
lation of a hint from the blackboard -- the target agent
checks for all possible hints from its new assignment and
executes the pick-and-replace procedure described be-
fore. In addition, if the solution of the puzzle is found
the search halts and the time t = t∗ is recorded.
Figure 3 shows the distribution of probability P (C) of
the computational cost for the independent search and
for the blackboard organization in the case there is no
space limitation on the blackboard so it can display all
351 distinct hints. Our results show that those distribu-
tions are exponential for both search strategies, contrary
to the suggestion of [13] that the exponential distribution,
which characterizes the independent search, would be re-
placed by a lognormal distribution for the blackboard
6
FIG. 3. Probability distribution of the computational cost
(3) for the independent search and the blackboard organiza-
tion with M = 100 agents. There is no space limitation on the
blackboard, i.e., B = 351. These distributions were generated
using 106 independent runs. The curve fitting the data of the
independent search is P (C) = 0.98 exp (−0.98C), whereas
the data of the blackboard system is fitted by P (C) =
10 exp (−10C).
organization. From a quantitative perspective, however,
the blackboard organization produces a tenfold decrease
of the computational cost as compared with the indepen-
dent search. In particular, (cid:104)C(cid:105) ≈ 0.10 for the unlimited
space blackboard and (cid:104)C(cid:105) ≈ 1.02 for the independent
search. We note that the elementary move is slightly less
efficient to explore the state space than the replacement
of the entire digit-to-letter assignment (global move) used
in Refs. [13, 32]. This is so because it is not too unlikely to
reverse a change made by the elementary move. For ex-
ample, the probability to reverse a change in a subsequent
trial is 2/90 for the the elementary move, whereas it is
1/10! for the global move. Interestingly, the replacement
of the global by the elementary move has no discernible
effect on the performance of the blackboard organization.
Figure 4, which shows the mean computational cost
as function of the system size, proves that t∗ scales with
1/M (and hence (cid:104)C(cid:105) is independent of the system size
M ) for both the blackboard and the independent search
strategies. For the blackboard organization, the increase
of the computational cost due to duplication of work oc-
curs for much smaller group sizes than for the indepen-
dent search since the blackboard reduces the effective size
of the state space to be explored by the agents. The main
point is that the unlimited blackboard system is not re-
ally a cooperative problem-solving system, since once the
blackboard is filled out, which happens in a very short
time [32], the agents will pick hints on the board and
explore the state space independently of each other.
This situation changes dramatically when the size B of
the blackboard is limited so the agents have to compete
for space to write their hints on the board. At first sight,
one would expect that limiting the information avail-
0.001 0.01 0.1 1 1000.20.40.60.81independentsearchblackboardP(C)C 7
from the board. This probability is defined as the ra-
tio between the number of correct hints selected and the
total number of hint selections, averaged over all agents
during a run. The result is then averaged over 105 inde-
pendent runs. Not surprisingly, Figs. 5 and 6 reveal the
strong correlation between (cid:104)C(cid:105) and φ so that the better
performance of limited space blackboards is consequence
of the higher odds of selecting a correct hint from the
board. But the reason these odds are higher for limited
boards is not obvious at all. For instance, consider a null
model in which the hints displayed on the board of size
B are selected randomly without replacement from the
pool of 351 hints. Since the cryptarithmetic puzzle (1)
has only six correct hints, the probability that the board
displays exactly k correct hints is given by an hyperge-
ometric distribution. Now, given that the blackboard
displays k correct hints, the probability that the agents
selects one of them is simply k/B. Hence the probability
that an agent selects a correct hint from a board of size
B in this null model is
k
B
=
6
351
≈ 0.017,
(4)
6(cid:88)
k=0
φ =
(cid:0)6
k
(cid:1)
(cid:1)(cid:0)351−6
(cid:0)351
(cid:1)
B−k
B
FIG. 4. Mean computational cost (cid:104)C(cid:105) as function of the sys-
tem size M for the independent search (triangles) and the
blackboard organization (circles). There is no space limita-
tion on the blackboard, i.e., B = 351. Each symbol represents
the average over 105 independent runs. The error bars are
smaller than the symbol sizes.
which does not depend on the board size. We note that
in this null model the limitation of the number of hints
displayed on the board does not affect the group perfor-
mance at all, contrary to the naive expectation that it
would harm that performance.
The explanation for the dependence on B shown in
Fig. 6 (and consequently in Fig. 5) is that the hints dis-
played on the board are not a random sample of the pool
of hints, as assumed in the null model. The somewhat
subtle reason for the bias towards the correct hints is that
whenever an agent assimilates a hint from the board it
must relocate up to six digits of its original digit-to-letter
assignment. In doing so, it is very likely to eliminate any
previous hints it carried, except if those hints, i.e., the
hint copied from the board and the hints that are already
part of the agent's assignment, are the correct hints. In
that sense, correct hints are insensitive to the radical re-
arrangement of digits resulting from the assimilation of
another correct hint from the board and this explains its
higher frequency in the blackboard.
As can be hinted from Fig. 5 the dependence of the
computational cost on the number of agents M is quite
complex and changes qualitatively for different values of
the board size B, so we will leave a detailed discussion of
this point to a future contribution. Here we mention only
that in a typical situation the computational cost initially
decreases with increasing M until it reaches a minimum
value, beyond which it begins to increase till it levels
off and becomes size-independent, provided M is not too
large, as in the case of the unlimited size blackboard (see
Fig. 4).
Finally, the fact that the blackboard organization
works so well for just a single agent (M = 1), which seems
to use the blackboard as an external memory to store the
FIG. 5. Mean computational cost (cid:104)C(cid:105) as function of the black-
board size B for systems of size M = 1 (triangles), 10 (in-
verted triangles) and 100 (circles). Each symbol represents
the average over 105 independent runs. The error bars are
smaller than the symbol sizes.
able to the agents would harm the group performance.
However, Fig. 5 shows that, except for small board sizes
(B < 20 for the data shown in the figure), increasing
the number of hints displayed on the board actually de-
grades the group performance, regardless of the group
size. In addition, for each group size there is a value of
B that minimizes the computational cost. In particular,
for M = 1 we find this optimum at B = 20, for M = 10
at B = 7 and for M = 100 at B = 15.
To understand the counterintuitive finding that lim-
iting the amount of information displayed on the black-
board improves the group performance, we present in Fig.
6 the probability φ that an agent selects a correct hint
0 0.2 0.4 0.6 0.8 1100101102103104105106107independent searchblackboard<C>M 0 0.02 0.04 0.06 0.08 0.1050100150200250300350<C>B 8
imal or not is the number of model parameters. The imi-
tative learning model has two parameters only: the copy
or imitation propensity p of the agents and the group size
M . The blackboard organization has two parameters too:
the group size M and the blackboard size B. The value
of a minimal model (or, for that matter, of any model)
should be gauged by the unexpectedness of its predic-
tions. In fact, the optimization of the group performance
for a specific group size and the Groupthink-like phe-
nomenon observed in the study of the imitative learning
model, as well as the improved performance of the black-
board organization in the case the amount of available
information is limited, bear witness to the importance of
those models.
From a quantitative perspective, we note that the best
performance of the imitative search strategy is achieved
for M = 5 and p = 0.96 and yields the mean compu-
tational cost (cid:104)C(cid:105) ≈ 0.03 (see Fig. 1), whereas the best
performance of the blackboard organization is achieved
for M = B = 10 with mean cost (cid:104)C(cid:105) ≈ 0.01 (see Fig. 5).
For the sake of comparison, we recall that the mean com-
putational cost of the independent search is (cid:104)C(cid:105) ≈ 1.02,
regardless of the group size, so our two cooperative work
scenarios produce a substantial boost on the performance
of the group of agents.
Although the two minimal models of distributed coop-
erative problem-solving systems presented here are easy
to formulate and simulate in a computer, they exhibit
some features that preclude any simple analytical ap-
proach as, for instance, the need to select the best digit-
to-letter assignment to serve as model for the agents
and the non-local effect of assimilating a hint from the
blackboard. Perhaps, these obstacles may yield to more
powerful and sophisticated mathematical tools, such as
the kinetic theory of active particles [43], so as to make
the study of collective intelligence more appealing to the
mathematics community [44].
To conclude, we note that the key issues that motivated
the proposal of the original minimal model of cooperative
problem-solving systems remain unanswered [13]. For in-
stance, the question whether cooperative work can alter
the statistical signature of the search on the state space of
the combinatorial problem is still open, since the compu-
tational costs of the models studied here are distributed
by exponential probability distributions as in the case of
the independent search (see Fig. 3). Moreover, for large
group sizes M the time t∗ to find the solution decreases
with 1/M for both the blackboard organization and the
independent search (see Fig. 4), whereas it actually in-
creases with increasing M for the imitative search due to
the Groupthink phenomenon (see Fig. 2). A qualitative
beneficial effect of cooperation should result in the scal-
ing t∗ ∝ 1/M α with the exponent α > 1, but producing
a model with this attribute has proved an elusive task so
far.
FIG. 6. Probability φ that an agent selects a correct hint
from the blackboard as function of the blackboard size B for
systems of size M = 1 (triangles), 10 (inverted triangles) and
100 (circles). Each symbol represents the average over 105
independent runs. The error bars are smaller than the symbol
sizes. The horizontal line at φ = 6/351 is the prediction of
the random blackboard null model.
hints discovered during its exploration of the state space,
indicates that the main role of the blackboard is not the
promotion of cooperation between the agents as initially
thought [13]: the blackboard serves as a collective mem-
ory storage device for the otherwise memoryless agents.
We note that for the reputation blackboard, where the
hints are posted together with the cost (2) of the agent,
the best performance is achieved in the case of a single
agent [32]. Interestingly, in the context of the wisdom-
of-crowds effect [40] (see also [41]), the performance of
a single individual is usually improved if its estimate is
taken as the average of its previous estimates -- the so-
called crowd within [42]. These findings in very distinct
contexts make evident the difficulty to disentangle the
memory from the cooperation effects without the aid of
the minimal model approach.
VI. CONCLUSION
Rather than advance new search heuristics to solve
combinatorial problems, the goal of our approach to
study collective intelligence is to assess quantitatively
and systematically the potential of cooperation to solve
problems in very simplified scenarios. Of course, once
the conditions that optimize the efficiency of cooperative
work are understood, this knowledge can be used to de-
vise cost-effective search heuristics, which is ultimately
the goal of the research on collective intelligence.
Here we have reviewed and offered original insights on
two minimal models of distributed cooperative problem-
solving systems, namely, the imitative learning search
strategy and the blackboard organization. A good crite-
rion to determine whether a mathematical model is min-
0.02 0.04 0.06 0.08 0.1050100150200250300350φB ACKNOWLEDGMENTS
The research of J.F.F. is supported in part by grants
2017/23288-0, Funda¸cao de Amparo `a Pesquisa do Es-
9
tado de Sao Paulo (FAPESP) and 305058/2017-7, Con-
selho Nacional de Desenvolvimento Cient´ıfico e Tec-
nol´ogico (CNPq).
is supported by FAPESP
scholarship 2015/17277-0 and A.C.A is supported by a
CAPES scholarship.
S.M.R.
[1] Fallani FDV, Bassett DS. Network neuroscience for op-
timizing brain-computer interfaces. Phys Life Rev [this
issue].
[2] Engelbart DC. Augmenting Human Intellect: A Con-
ceptual Framework. SRI Summary Report AFOSR-3223;
1962.
[3] Deary IJ. Intelligence: A Very Short Introduction. Ox-
ford, UK: Oxford University Press; 2001.
[4] Woolley AW, Chabris CF, Pentland A, Malone NHT.
Evidence for a Collective Intelligence Factor in the Per-
formance of Human Groups. Science 2010; 333: 686 -- 688.
[5] Bavelas A. Communication Patterns in Task-Oriented
Groups. J Acoustical Soc Amer 1950; 22: 725 -- 730.
[6] Guetzkow H, Simon H A. The Impact of Certain Com-
munication Nets upon Organization and Performance in
Task-oriented Groups. Manage Sci 1955; 1: 233 -- 250.
[7] Leavitt HJ. Some Effects of Certain Communication Pat-
terns on Group Performance. J Abnorm Soc Psych 1951;
46: 38 -- 50.
2001.
[22] Blackmore S. The Meme Machine. Oxford: Oxford Uni-
versity Press; 2000.
[23] Rendell L, Boyd R, Cownden D, Enquist M, Eriksson
K, Feldman MW, Fogarty L, Ghirlanda S, Lillicrap T,
Laland KN. Why Copy Others? Insights from the Social
Learning Strategies Tournament. Science 2010; 328: 208 --
213.
[24] Kivinen O, Piiroinen T. Evolutionary Understanding of
Consciousness -- a Transactional Naturalist Approach.
Phys Life Rev [this issue].
[25] Lazer D, Friedman A. The Network Structure of Explo-
ration and Exploitation. Admin Sci Quart 2007; 52: 667 --
694.
[26] Fontanari JF. Imitative Learning as a Connector of Col-
lective Brains. PLoS ONE 2014; 9: e110517.
[27] Fontanari JF. Exploring NK fitness landscapes using im-
itative learning. Eur Phys J B 2015; 88: 251.
[28] Janis IL. Groupthink: psychological studies of policy de-
[8] Mason W, Watts DJ. Collaborative Learning in Net-
cisions and fiascoes. Boston: Houghton Mifflin; 1982.
works. Proc Natl Acad Sci 2012; 109: 764 -- 769.
[9] Reia SM, Herrmann S, Fontanari JF. Impact of centrality
on cooperative processes. Phys Rev E 2017; 95: 022305.
[10] Reia SM, Fontanari JF. Effect of group organization on
the performance of cooperative processes. Ecol Complex
2017; 30: 47 -- 56.
[11] Page SE. The Difference: How the Power of Diversity
Creates Better Groups, Firms, Schools, and Societies.
Princeton, NJ: Princeton University Press; 2007.
[29] Fontanari JF, Rodrigues FA. Influence of network topol-
ogy on cooperative problem-solving systems. Theor
Biosci 2016; 135: 101 -- 110.
[30] Fontanari JF. When more of the same is better. EPL
2016; 113: 28009.
[31] Corkill DD. Blackboard Systems. AI Expert 1991; 6: 40 --
47.
[32] Fontanari JF. Reputation blackboard systems. Cogn Syst
Res 2018; 50: 29 -- 35.
[12] Huberman BA. The performance of cooperative pro-
[33] Averbach B, Chein O. Problem Solving Through Recre-
cesses. Physica D 1990; 42: 38 -- 47.
[13] Clearwater SH, Huberman BA, Hogg T. Cooperative So-
lution of Constraint Satisfaction Problems. Science 1991;
254: 1181 -- 1183.
[14] Perlovsky L. Toward physics of the mind: Concepts, emo-
tions, consciousness, and symbols. Phys Life Rev 2006;
3: 23 -- 55.
[15] Schoeller F, Perlovsky L, Arseniev D. Physics of mind:
Experimental confirmations of theoretical predictions.
Phys Life Rev 2018; doi: 10.1016/j.plrev.2017.11.021
[16] Bonabeau E, Dorigo M, Theraulaz G. Swarm Intelli-
gence: From Natural to Artificial Systems. Oxford, UK:
Oxford University Press; 1999.
[17] Kennedy J. Thinking is social: Experiments with the
adaptive culture model. J Conflict Res 1998; 42: 56 -- 76.
[18] Fontanari JF. Social interaction as a heuristic for com-
binatorial optimization problems. Phys Rev E 2010; 82:
056118.
[19] Huberman BA, Hogg T. Phase transitions in artificial
intelligence systems. Artif Intell 1987; 33: 155 -- 171.
[20] Newell A, Simon HA. Human problem solving. Engle-
wood Cliffs, NJ: Prentice-Hall; 1972.
[21] Bloom H. Global Brain: The Evolution of Mass Mind
from the Big Bang to the 21st Century. New York: Wiley;
ational Mathematics. San Francisco: Freeman; 1980.
[34] Kauffman S, Levin S. Towards a general theory of adap-
tive walks on rugged landscapes. J Theor Biol 1987; 128:
11 -- 45.
[35] Abbasian R, Mazloom M. Solving Cryptarithmetic Prob-
lems Using Parallel Genetic Algorithm. In: Proceedings
of the Second International Conference on Computer
and Electrical Engineering. Washington, DC: IEEE Com-
puter Society; 2009. p. 308 -- 312.
[36] Fontanari JF. Awareness improves problem-solving per-
formance. Cogn Syst Res 2017; 45: 52 -- 58.
[37] Wilson E. Sociobiology. Cambridge, MA: Harvard Uni-
versity Press; 1975.
[38] Dunbar RIM. Neocortex size as a constraint on group
size in primates. J Human Evol 1992; 22: 469 -- 493.
[39] Goldberg DE. Genetic Algorithms in Search, Optimiza-
tion, and Machine Learning. Reading: Addison-Wesley;
1989.
[40] Galton F. Vox populi. Nature 1907; 75: 450 -- 451.
[41] Surowiecki J. The wisdom of crowds: Why the Many
Are Smarter than the Few and How Collective Wisdom
Shapes Business, Economies, Societies, and Nations. New
York: Random House; 2004.
[42] Vul E, Pashler H. Measuring the crowd within: prob-
abilistic representations within individuals. Psychol Sci
2008; 19: 645 -- 647.
10
[43] Bellomo N, Bianca C, Delitala M. Complexity analysis
and mathematical tools towards the modelling of living
systems. Phys Life Rev 2009; 6: 144 -- 175.
[44] Burini D, De Lillo S, Gibelli L. Collective learning dy-
namics modeling based on the kinetic theory of active
particles. Phys Life Rev 2016; 16: 123 -- 139.
|
1809.04918 | 1 | 1809 | 2018-09-13T12:44:48 | Coordination-driven learning in multi-agent problem spaces | [
"cs.MA",
"cs.AI"
] | We discuss the role of coordination as a direct learning objective in multi-agent reinforcement learning (MARL) domains. To this end, we present a novel means of quantifying coordination in multi-agent systems, and discuss the implications of using such a measure to optimize coordinated agent policies. This concept has important implications for adversary-aware RL, which we take to be a sub-domain of multi-agent learning. | cs.MA | cs | Coordination-driven learning in multi-agent problem spaces
Sean L. Barton, Nicholas R. Waytowich, and Derrik E. Asher
U.S. Army Research Laboratory, Aberdeen Proving Ground
Aberdeen, Maryland 21005
8
1
0
2
p
e
S
3
1
]
A
M
.
s
c
[
1
v
8
1
9
4
0
.
9
0
8
1
:
v
i
X
r
a
Abstract
We discuss the role of coordination as a direct learning objec-
tive in multi-agent reinforcement learning (MARL) domains.
To this end, we present a novel means of quantifying coordi-
nation in multi-agent systems, and discuss the implications of
using such a measure to optimize coordinated agent policies.
This concept has important implications for adversary-aware
RL, which we take to be a sub-domain of multi-agent learn-
ing.
1 Introduction
Modern reinforcement learning (RL) has demonstrated a
number of striking achievements in the realm of intelligent
behavior by leveraging the power of deep neural networks
(Mnih et al. 2015). However, like any deep-learning system,
RL agents are vulnerable to adversarial attacks that seek to
undermine their learned behaviors (Huang et al. 2017). In
order for RL agents to function effectively along side hu-
mans in real-world problems, their behaviors must be re-
silient against such adversarial assaults.
Promisingly, there is recent evidence showing deep RL
agents learn policies robust to adversary attacks at test
time when they train with adversaries during learning
(Behzadan and Munir 2017). This has important implica-
tions for robust deep RL, as it suggests that security against
attacks can be derived from learning. Here, we build on this
idea and suggest that deriving adversary-aware agents from
learning is a subset of the multi-agent reinforcement learn-
ing (MARL) problem.
At the heart of this problem is the need for an individ-
ual agent to coordinate its actions with those taken by other
agents (Fulda and Ventura 2007). Given the role of inter-
agent coordination in MARL, we suggest that operationaliz-
ing coordination between agent actions as a direct learning
objective may lead to better policies for multi-agent tasks.
Here, we present a quantitative metric that can be used to
measure the degree of coordination between agents over the
Copyright c(cid:13) by the papers authors. Copying permitted for private
and academic purposes. In: Joseph Collins, Prithviraj Dasgupta,
Ranjeev Mittu (eds.): Proceedings of the AAAI Fall 2018 Sympo-
sium on Adversary-Aware Learning Techniques and Trends in Cy-
bersecurity, Arlington, VA, USA, 18-19 October, 2018, published
at http://ceur-ws.org
course of learning. Further, we present a research concept
for using this metric to shape agent learning towards coordi-
nated behavior, as well as the impact that different degrees
of coordination can have on multi-agent task performance.
1.1 Adversary-aware RL as MARL
Understanding adversary-aware RL agents in terms of
MARL is straightforward when we consider that train-
ing in the presence of adversarial attacks is similar
to training in the presence of agents pursuing compet-
ing goals. In competitive RL, outcomes are often con-
sidered zero-sum, when agents reward/loss are in di-
rect opposition (Busoniu, Babuska, and De Schutter 2008;
Crandall and Goodrich 2011). In the case of attacks on RL
agents, the adversary's goal is typically to learn a cost func-
tion that, when optimized, minimizes the returns of the at-
tacked agent (Pattanaik et al. 2017). Thus, the adversary's
reward is the opposite of the attacked agent's.
If we take seriously these comparisons, the problem of
creating adversary-aware agents is largely one of develop-
ing agents that can learn to coordinate their behaviors effec-
tively with the actions of an adversary so as to minimize the
impact of its attacks. Thus, adversary-aware RL is an inher-
ently multi-agent problem.
1.2 Coordination in MARL
In MARL problems, the simultaneous actions of multi-
ple actors obfuscate the ground truth from any individual
agent. This uncertainty about the state of the world is pri-
marily studied in terms of 1) partial-observability wherein
the information about a given state is only probabilistic
(Omidshafiei et al. 2017), and 2) non-stationarity where the
goal of the task is "moving" with respect to any individual
agent's perspective (Hernandez-Leal et al. 2017).
on
the
degree
tasks
depends
are
able
To the
extent
critically
that uncertainty from an agent's
perspective can be resolved, performance in multi-
agent
to
which agents
efforts
(Matignon, Laurent, and Le Fort-Piat 2012). With MARL
collaborative goals, individual agents must learn policies
that increase their own reward without diminishing the re-
ward received by other agents. Simple tasks, such as matrix
or climbing games, present straightforward constraints that
promote the emergence of coordination between agents, as
to coordinate
their
these small state-space problems make the pareto-optimal
solution readily discoverable.
dination in more complex cases, and does not provide any
new information during learning.
Matignon et al. (2012) enumerate the challenges for co-
operative MARL, and show that no single algorithm is
successful at achieving better performance. Instead, ex-
isting algorithms tend to address specific challenges at
the expense of others. Further, in more complex state-
spaces pareto-optimal solutions can be "shadowed" by in-
dividually optimal solutions that constrain learned behav-
ior to selfish policies (Fulda and Ventura 2007). This un-
dermines the performance gains achievable through co-
ordinated actions in MARL problems. For
these rea-
sons, coordination between agents can only be guaran-
teed in limited cases where the challenges of MARL can
be reasonably constrained (Lauer and Riedmiller 2000). As
such, partial-observability and non-stationarity are prob-
lems that must be overcome for coordination to emerge
(Matignon, Laurent, and Le Fort-Piat 2012). For complex
tasks, modern advances with DNNs have leveraged joint
action learning to overcome the inherent uncertainty of
MARL (Foerster et al. 2017). Indeed these algorithms show
improved performance over decentralized and independent
learning alternatives.
Though this work is promising, we recently showed
that when coordination is directly measured, it cannot ex-
plain the improved performance of these algorithms in all
cases (Barton et al. In Press). Coordination between agents,
as measured by the causal influence between agent actions
(method described below), was found to be almost indis-
tinguishable from hard-coded agents forced to act indepen-
dently. This leads to an interesting question about how to
achieve coordinated actions between learning agents in real-
world tasks where there is strong interest for the deployment
of RL-equipped agents.
2 Approach
A possible solution to overcome the challenges of MARL is
to address coordination directly. This concept was recently
put to the test in several simple competitive tasks being per-
formed by two deep RL agents (Foerster et al. 2018). The
study explicitly took into account an opponent's change in
learning parameters during its own learning step. Account-
ing for opponent behavior during learning in this manner
was shown to yield human-like cooperative behaviors previ-
ously unobserved in MARL agents.
In a similar thrust, we propose here that coordination
should not be left to emerge from the constraints on the
multi-agent task, but instead be a direct objective of learn-
ing. This may be accomplished by providing a coordination
measure in the loss of a MARL agent's optimization step.
2.1 A novel measure for coordination in MARL
The first step towards optimizing coordinated behavior in
MARL is to define an adequate measure of coordination.
Historically, coordinated behavior has been evaluated by
agent performance in tasks where cooperation is explicitly
required (Lauer and Riedmiller 2000). As we showed previ-
ously, performance alone is insufficient for evaluating coor-
Fortunately a metric borrowed from ecological research
has shown promise as a quantitative measure of inter-agent
coordination, independent of performance. Convergent cross
mapping (CCM) quantifies the unique causal influence one
time-series has on another (Sugihara et al. 2012). This is
accomplished by embedding each time-series in its own
high dimensional attractor space, and then using the embed-
ded data of one time-series as a model for the other. Each
model's accuracy is taken as a measure of the causal influ-
ence between the two time-series.
In multi-agent tasks, we can define collaboration to be the
amount of causal influence between time-series of agent ac-
tions, as measured by CCM. The advantage of this metric
is that it provides a measure of coordination between agents
that is independent of performance, and thus can be used
as a novel training signal to optimize coordinated behav-
ior. Thus, coordination is no longer exclusively an emergent
property of the task, but rather a signal for driving agents'
learned behavior.
2.2 Coordination in an example MARL task
We propose an experimental paradigm that is designed to
measure the role of coordination in a continuous coopera-
tive/competitive task: online learning of coordination during
multi-agent predator-prey pursuit. In this exemplary experi-
ment, CCM is used as a direct learning signal that influences
how agents learn to complete a cooperative task.
The task is an adaptation of discrete predator-prey pursuit
(Benda 1985) into a continuous bounded 2D particle envi-
ronment with three identical predator agents and a single
prey agent. Predators score points each time they make con-
tact with the prey, while the prey's points are decremented if
contacted by any predator.
Typically, agent learning would be driven solely by the en-
vironmental reward (in this case, agent score). With this typ-
ical framework, coordination may emerge, but is not guaran-
teed (see (Barton et al. In Press)). In contrast, CCM provides
a direct measure of inter-agent coordination, which can be
used to modify agent learning through the incorporation of
CCM as a term in learning loss. This can be done either in-
directly as a secondary reward or directly as a term applied
during back-propagation. Thus, learned behavior is shaped
by both, task success and inter-agent coordination.
This paradigm provides an opportunity for coordination
to be manipulated experimentally by setting a desired coor-
dination threshold. As agents learn, they should coordinate
their behaviors with their partners and/or adversaries up to
this threshold. Minimizing this threshold should yield agents
that optimize the task at the expense of a partner, while
maximizing this threshold would likely produce high dimen-
sional oscillations between agent actions that ignore task de-
mands. Effective coordination likely lies between these ex-
tremes. Thus, we can directly observe the impact of coor-
dinated behaviors in a MARL environment by varying this
coordination threshold. To our knowledge, this has not been
previously attempted.
3 Implications and Discussion
Explicit coordination between agents can lead to greater
success in multi-agent systems. Our concept provides a
paradigm shift towards making coordination between agents
an intended goal of learning. In contrast, many previous
MARL approaches assume that coordination will emerge as
performance is optimized. In summary, we suggest that co-
ordination is better thought of as a necessary driver of learn-
ing, as important as (or possibly more important than) per-
formance measures alone.
Our proposed use of CCM as a signal for inter-agent co-
ordination provides a new source of information for learning
agents that can be integrated into a compound loss function
during learning. This would allow agents to learn coordi-
nated behaviors explicitly, rather than gambling on agents
discovering coordinated policies during exploration.
With the addition of coordination driven learning, the
policies an agent learns will not take into account adversary
behavior by chance, but rather by design. Such an algorithm
would actively seek out policies that account for the actions
of partners and competitors, limiting the policy search space
to those that reason over the behavior of other agents in the
system. We believe this is a reasonable avenue for more ef-
ficiently training mulit-agent policies.
Driving learning with coordination creates an opportunity
for the development of agents that are inherently determined
to coordinate their actions with a human partner. This is im-
portant, as without such a drive it is not clear how to guaran-
tee that humans and agents will work well together. In partic-
ular, if modeling of human policies is too difficult for agents,
they may settle on policies that try to minimize the degree
of coordination in an attempt to recover some selfishly op-
timal behavior. Forcing coordination to be optimized during
learning ensures that agents only seek out policies that are
well integrated with the actions of their partners.
Our concept, as presented here, is to promote coordinated
behaviors in intelligent learning agents by providing a quan-
titative measure of coordination that can be optimized dur-
ing learning. The importance of implementing coordination
to overcome adversarial attacks in the MARL problem can-
not be understated. Furthermore, an explicit drive towards
coordinated behavior between intelligent agents constitutes
a significant advancement within the fields of artificial intel-
ligence and computational learning.
Acknowledgements This research was sponsored by the
Army Research Laboratory and was accomplished under
Cooperative Agreement Number W911NF-18-2-0058. The
views and conclusions contained in this document are those
of the authors and should not be interpreted as represent-
ing the official policies, either expressed or implied, of
the Army Research Laboratory or the U.S. Government.
The U.S. Government is authorized to reproduce and dis-
tribute reprints for Government purposes notwithstanding
any copyright notation herein.
References
laborative emergent behavior in multi-agent reinforcement
learning.
In 1st International Conference on Human Sys-
tems Engineering and Design. IHSED.
[Behzadan and Munir 2017] Behzadan, V., and Munir, A.
2017. Whatever does not kill deep reinforcement learning,
makes it stronger. arXiv preprint arXiv:1712.09344.
[Benda 1985] Benda, M. 1985. On optimal cooperation of
knolwedge sources. Technical Report BCS-G2010-28.
[Busoniu, Babuska, and De Schutter 2008] Busoniu,
L.;
Babuska, R.; and De Schutter, B. 2008. A comprehen-
sive survey of multiagent reinforcement learning.
IEEE
Transactions on Systems, Man, And Cybernetics-Part C:
Applications and Reviews, 38 (2), 2008.
[Crandall and Goodrich 2011] Crandall,
and
Goodrich, M. A.
2011. Learning to compete, coordi-
nate, and cooperate in repeated games using reinforcement
learning. Machine Learning 82(3):281 -- 314.
J. W.,
[Foerster et al. 2017] Foerster, J.; Farquhar, G.; Afouras, T.;
Nardelli, N.; and Whiteson, S. 2017. Counterfactual Multi-
Agent Policy Gradients. arXiv:1705.08926 [cs].
arXiv:
1705.08926.
[Foerster et al. 2018] Foerster, J.; Chen, R. Y.; Al-Shedivat,
M.; Whiteson, S.; Abbeel, P.; and Mordatch, I. 2018. Learn-
ing with opponent-learning awareness. In Proceedings of the
17th International Conference on Autonomous Agents and
MultiAgent Systems, 122 -- 130. International Foundation for
Autonomous Agents and Multiagent Systems.
[Fulda and Ventura 2007] Fulda, N., and Ventura, D. 2007.
Predicting and preventing coordination problems in cooper-
ative q-learning systems. In IJCAI, volume 2007, 780 -- 785.
[Hernandez-Leal et al. 2017] Hernandez-Leal, P.; Kaisers,
M.; Baarslag, T.; and de Cote, E. M. 2017. A survey
of learning in multiagent environments: Dealing with non-
stationarity. arXiv preprint arXiv:1707.09183.
[Huang et al. 2017] Huang, S.; Papernot, N.; Goodfellow, I.;
Duan, Y.; and Abbeel, P. 2017. Adversarial attacks on neural
network policies. arXiv preprint arXiv:1702.02284.
[Lauer and Riedmiller 2000] Lauer, M., and Riedmiller, M.
2000. An algorithm for distributed reinforcement learning
in cooperative multi-agent systems. In In Proceedings of the
Seventeenth International Conference on Machine Learn-
ing. Citeseer.
[Matignon, Laurent, and Le Fort-Piat 2012] Matignon, L.;
Laurent, G. J.; and Le Fort-Piat, N. 2012.
Independent
reinforcement learners in cooperative markov games: a
survey regarding coordination problems. The Knowledge
Engineering Review 27(1):1 -- 31.
[Mnih et al. 2015] Mnih, V.; Kavukcuoglu, K.; Silver, D.;
Rusu, A. A.; Veness, J.; Bellemare, M. G.; Graves, A.; Ried-
miller, M.; Fidjeland, A. K.; Ostrovski, G.; Petersen, S.;
Beattie, C.; Sadik, A.; Antonoglou, I.; King, H.; Kumaran,
D.; Wierstra, D.; Legg, S.; and Hassabis, D. 2015. Human-
level control through deep reinforcement learning. Nature
518(7540):529 -- 533.
[Barton et al. In Press] Barton, S. L.; Waytowich, N. R.;
Zaroukian, E.; and Asher, D. E. In Press. Measuring col-
[Omidshafiei et al. 2017] Omidshafiei, S.; Pazis, J.; Amato,
C.; How, J. P.; and Vian, J. 2017. Deep Decentralized Multi-
task Multi-Agent Reinforcement Learning under Partial Ob-
servability. arXiv:1703.06182 [cs]. arXiv: 1703.06182.
[Pattanaik et al. 2017] Pattanaik, A.; Tang, Z.; Liu, S.; Bom-
mannan, G.; and Chowdhary, G. 2017. Robust deep rein-
forcement learning with adversarial attacks. arXiv preprint
arXiv:1712.03632.
[Sugihara et al. 2012] Sugihara, G.; May, R.; Ye, H.; Hsieh,
C.-h.; Deyle, E.; Fogarty, M.; and Munch, S. 2012. Detect-
ing causality in complex ecosystems. science 1227079.
|
1711.09229 | 1 | 1711 | 2017-11-25T11:28:00 | A-Evac: the evacuation simulator for stochastic environment | [
"cs.MA"
] | We introduce an open-source software Aamks for fire risk assessment. This article focuses on a component of Aamks - an evacuation simulator named a-evac. A-evac models evacuation of humans in the fire environment produced by CFAST fire simulator. In the article we discuss the probabilistic evacuation approach, automatic planning of exit routes, the interactions amongst the moving evacuees and the impact of smoke on the humans. The results consist of risk values based on FED, F-N curves and evacuation animations. | cs.MA | cs | Noname manuscript No.
(will be inserted by the editor)
A-Evac: the evacuation simulator for stochastic
environment
PRE-PRINT
Adam Krasuski · Karol Krenski
Received: 24.11.2017 / Accepted: date
Abstract We introduce an open-source software Aamks for fire risk assess-
ment. This article focuses on a component of Aamks -- an evacuation simulator
named a-evac. A-evac models evacuation of humans in the fire environment
produced by CFAST fire simulator. In the article we discuss the probabilis-
tic evacuation approach, automatic planning of exit routes, the interactions
amongst the moving evacuees and the impact of smoke on the humans. The
results consist of risk values based on FED, F-N curves and evacuation ani-
mations.
Keywords Evacuation Modeling · Stochastic Simulations · Risk Manage-
ment · Monte Carlo
1 Introduction
These days, there's not much you can understand about what is going
on around you if you do not understand the uncertainty attached to
pretty much every phenomenon.
-- J. N. Tsitsiklis
There is a continuous progress in understanding the fire phenomenon, its
impact on the structure and on the reaction of humans and safety systems.
A. Krasuski
The Main School of Fire Service
Slowackiego 52/54
01-629 Warsaw, Poland
E-mail: [email protected]
K. Krenski
The Main School of Fire Service
Slowackiego 52/54
01-629 Warsaw, Poland
E-mail: [email protected]
7
1
0
2
v
o
N
5
2
]
A
M
.
s
c
[
1
v
9
2
2
9
0
.
1
1
7
1
:
v
i
X
r
a
2
Adam Krasuski, Karol Krenski
There are scientific methods and models of the fire and the emergency scene,
there are computer implementations, but the complexity of the domain im-
pedes the more widespread use of these tools.
Currently, the most typical approach for assessing the safety of the building
is a precise choosing of the input parameters for a small number of lengthy,
detailed simulations. This procedure is managed by a practitioner, based on
his experience. However, based on heuristics and biases [31,22] we have con-
cerns that human judgement surpasses statistical calculations. The alternative
is to let computer randomly choose the parameters and run thousands of sim-
ulations. The resulting collection allows us, after further processing, to judge
on the safety of the building.
2 Aamks, the multisimulations platform
We created Aamks -- the platform for running simulations of fires and then
running the evacuation simulations, but thousands of them for a single project.
This is the Monte-Carlo approach. We use CFAST, which is a rough, but a
fast fire simulator. This allows us to explore the space of possible scenarios and
assess the probability of them. The second component of risk -- consequences --
is taken from an evacuation simulator capable to model evacuation in the fire
environment. We use a-evac as the evacuation simulator which we have built
from scratch. The multisimulation is a handy name for what we are doing.
Aamks tries to assess the risk of failure of humans evacuation from a building
under fire. We applied methodology proposed in [16, 8, 1] -- stochastic simula-
tions based on the Simple Monte-Carlo approach [7]. Our primary goal was
to develop an easy to use engineering tool rather than a scientific tool, which
resulted in: AutoCAD plugin for creating geometries, web based interface,
predefined setups of materials and distributions of various aspects of buildings
features, etc. The workflow is as follows: The user draws a building layout
or exports an existing one. Next, the user defines a few parameters including
the type of the building, the safety systems in the building, etc. Finally, they
launch a defined number of stochastic simulations. As a result they obtain
the distributions of the safety parameters, namely: available safe egress time
(ASET), required safe egress time (RSET), fractional effective dose (FED),
hot layer height and temperature and F-N curves as well as the event tree and
risk matrix.
Fortran is a popular language for coding simulations of physical systems.
CFAST and FDS+Evac are coded in fortran. Since we don't create a fire sim-
ulator we code Aamks in python which is more comfortable due to extremely
rich collection of libraries. We decided that borrowing Evac from FDS and inte-
grating it with Aamks would be harder for us than to code our own evacuation
simulator, hence a-evac was born. There's also a higher chance of attracting
new python developers than fortran developers for our project.
Aamks consists of the following modules:
A-Evac: the evacuation simulator for stochastic environment
3
-- a-geom, geometry processing: AutoCAD plugin, importing geometry, ex-
tracting topology of the building, navigating in the building, etc.
-- a-evac, directing evacuees across the building and altering their states
-- a-fire, CFAST and FDS binaries and processing of their outputs
-- a-gui, web application for user's input and for the results visualisation
-- a-montecarlo, stochastic producer of thousands of input files for CFAST
and a-evac
-- a-results, post-processing the results and creating the content for reports,
-- a-manager, managing computations on the grid/cluster of computers
-- a-installer
3 A-evac, the evacuation simulator
In the following subsections we describe the internals of a-evac, sometimes
with the necessary Aamks context.
3.1 Geometry of the environment
The Aamks workflow starts with a 3D geometry where fires and evacuations
will be simulated. We need to represent the building, which contains one or
more floors. Each floor can consist of compartments and openings in them,
named respectively COMPAS and VENTS in CFAST. Our considerations are
narrowed to rectangular geometries. There are two basic ways for representing
architecture geometries: a) cuboids can define the insides of the rooms (type-
a-geometry) or b) cuboids can define the walls / obstacles (type-b-geometry).
CFAST uses the type-a-geometry. We create CFAST geometries from the input
files of the following format (there are more entities than presented here):
{
"FLOOR 1":
{
"ROOM": [
[ [ 3.0 , 4.8 , 0.0 ] , [ 4.8 , 6.5 , 3.0 ] ] ,
[ [ 3.0 , 6.5 , 0.0 ] , [ 6.8 , 7.4 , 3.0 ] ]
] ,
"COR": [
[ [ 6.2 , 0.2 , 0.0 ] , [ 7.6 , 4.8 , 3.0 ] ]
] ,
"D": [
[ [ 3.9 , 3.4 , 0.0 ] , [ 4.8 , 3.4 , 2.0 ] ]
] ,
"W": [
[ [ 1.2 , 3.4 , 1.0 ] , [ 2.2 , 3.4 , 2.0 ] ]
] ,
"HOLE": [
[ [ 3.0 , 6.5 , 0.0 ] , [ 4.8 , 6.5 , 3.0 ] ]
Adam Krasuski, Karol Krenski
4
}
]
}
ROOM and COR(RIDOR) belong to COMPAS. D(OOR), W(INDOW)
and HOLE belong to VENTS. HOLE is a result of CFAST restrictions -- it
is an artificial entity which serves to merge two compartments into a single
compartment as shown on Figure 1.
Fig. 1 The concept of a HOLE: a) the room in reality, b) the room representation in
CFAST: two rectangles for separated calculations, but open to each other via HOLE.
All the entities in the example belong to the same FLOOR 1. The triplets
are (x0, y0, z0) and (x1, y1, z1) encoding the beginning and the end of each
entity in 3D space. In practice we obtain these input files from AutoCAD,
thanks to our plugin which extracts data from AutoCAD drawing. There's
also an Inkscape svg importer -- useful, but without some features. Adding
basic support for another graphics tools is not much work.
In later sections we will introduce problems of guiding evacuees throughout
the building. Those modules require the type-b-geometry. We convert from a
type-a-geometry to a type-b-geometry by duplicating the geometry, translating
the geometry and applying some logical operations. Figure 2 shows the idea.
Fig. 2 The conversion from type-a-geometry to type-b-geometry
There are three aspects of movement when it comes to evacuation mod-
eling [9]: (a) path-finding -- for the rough route out of the building, (b) local
movement -- evacuees interactions with other evacuees, with obstacles and with
environment, and (c) locomotion -- for "internal" movement of the agent (e.g.
body sway). A-evac models only (a) and (b).
a)b) A-Evac: the evacuation simulator for stochastic environment
5
3.2 Path-finding (roadmap)
The simulated evacuees need to be guided out of the building. The type-b-
geometry provides the input for path-finding. Each of the cuboids in type-
b-geometry -- representing obstacles -- is defined by the coordinates. These
coordinates represent corners of the shapes. Since we model each of the floors of
a building separately, we flatten 3D geometry into 2D and represent obstacles
as rectangles. Therefore type-b-geometry in path-finding module is represented
as set of 4-tuple coordinates(cid:0)(x0, y0), (x1, y1), (x2, y2), (x3, y3)(cid:1).
The set of 4-tuple elements is then flatten to the set of coordinates --
bag-of-coordinates. Due to the fact that majority of the obstacles share the
coordinates, we remove duplicates from the set (for the sake of performance).
Then, this bag-of-coordinates is the input for triangulation. We apply Delau-
nay triangulation [10], that represents space as a set of triangles. Figure 3
depicts the idea of triangulation.
Fig. 3 The idea of triangulation. a) original geometry, b) bag-of-coordinates, c) triangula-
tion.
The triangles are used as navigation meshes for the agents. The navigation
meshes define which areas of an environment are traversable by agents.
After triangulation of bag-of-coordinates, some of the triangles are located
inside the obstacles -- those (by definition) are not traversable so we remove
them. What is left is a traversable-friendly space.
We create then the graph of spatial transitions for the agents, based on
the adjacency of triangles obtained from the triangulation. Spatial transition
means that an agent can move from one triangle to another.
An agent on an edge of a triangle can always reach the other two edges.
For triangles which share edges it allows an agent to travel from one triangle
to another.
The pairs of all neighbouring edges are collected. We use python networkx
module [4] which creates a graph made of the above pairs. For further pro-
cessing we add agents positions to the graph, by pairing them with the neigh-
bouring edges.
The graph represents all possible routes from any node to any other node
in the graph. We can query the graph for the route from the current agent's
position to the closest exit. It means that agent will walk through the con-
secutive nodes and will finally reach the exit door. We instruct networkx that
we need the shortest distances in our routes (default is the least hops on the
ROOM1ROOM2ROOM3CORR1STAIR1D1D2D3D4a)c)b) 6
Adam Krasuski, Karol Krenski
graph) and we obtain a set of edges the agent should traverse in order to reach
the exit. Figure 4 depicts the set of edges returned by the graph for an example
query.
Fig. 4 The roadmap defined by the graph for an example query. The red line crosses centers
of edges that an agent needs to travel to reach the exit.
The set of edges returned by the graph cannot be used directly for path-
finding. Neither the vertices of the edges nor the centers of them, do define
the optimal path that would be naturally chosen by evacuees during real evac-
uation. Therefore an extra algorithm should be used to smooth the path. For
this purpose we apply funnel algorithm defined in [6]. The funnel is a simple
algorithm finding straight lines along the edges.
The input for the funnel consists of a set of ordered edges (named portals)
from the agent origin to the destination. The funnel always consist of 3 entities:
the origin (apex) and the two vectors from apex to vertices on edges -- the left
leg and the right leg.
The apex is first set to the origin of the agent and the legs are set to the
vertices of the first edge. We advance the left and right legs to the consecutive
edges in the set and observe the angle between the legs. When the angle gets
smaller we accept the new vertex for the leg. Otherwise the leg stays at the
given vertex. After some iteration one of the legs should cross the other leg
defining the new position of the apex. The apex is moved and we restart the
procedure.
As a result the path is smoothened and defined only by the points where the
changes in velocity vector are needed. Moreover, we used an improved version
of the funnel algorithm that allows for defining points keeping a distance from
the corners reflecting the size of the evacuee. This allows for modeling the
A-Evac: the evacuation simulator for stochastic environment
7
Fig. 5 The idea of funnel algorithm. a) starting point, b) advancing legs.
impaired evacuees on wheeled chairs or beds in the hospitals. Figure 6 depicts
the smoothened path by funnel algorithm.
Fig. 6 The roadmap from starting point to exit smoothened by funnel algorithm.
3.3 Local movement
Local movement focuses on the interaction with (a) other agents (b) static
obstacles (walls) and (c) environmental conditions. A-evac handles (a) and (b)
via RVO21 which is an implementation of the Optimal Reciprocal Collision
Avoidance (ORCA) algorithm proposed in [3,32]. Later in this section we
describe how we are picking the local targets which is an aspect of (b). (c) is
basically altering agent's state such as speed.
1 http://gamma.cs.unc.edu/RVO2/
apexleft legright leganglea)b) 8
Adam Krasuski, Karol Krenski
RVO2 aims at avoiding the velocity obstacle [11]. The velocity obstacle is
the set of all velocities of an agent that will result in a collision with another
agent or an obstacle. Otherwise, the velocity is collision-avoiding. RVO2 aims
at asserting that none of the agents collides with other agents within time τ .
The overall approach is as follows: each of the agents is aware of other
agents parameters: their position, velocity and radius (agent's observable uni-
verse). Besides, the agents have their private parameters: maximum speed and
a preferred velocity which they can auto-adjust granted there is no other agent
or an obstacle colliding. With each loop iteration, each agent responses to what
he finds in his surroundings, i.e. his own and other agents radiuses, positions
and velocities. The agent updates his velocity if it is the velocity obstacle with
another agent. For each pair of colliding agents the set of collision-avoiding ve-
locities is calculated. RVO2 finds the smallest change required to avert collision
within time τ and that is how an agent gets his new velocity. The agent alters
up to half of his velocity while the other colliding agent is required to take
care of his half. Figure 8 a-b depicts the idea of velocity collision avoidance.
The algorithm remains the same for avoiding static obstacles. However, the
value of τ is smaller with respect to obstacles as agents should be more 'brave'
to move towards an obstacle if this is necessary to avoid other agents.
It turned out problematic how to pick the local target from the roadmap.
Local targets need to be updated (usually advanced, but not always) near
points defined by funnel algorithm during path-fining phase -- the disks on
Figure 7, after they become visible to the agent. However, the disks can be
crowded and agents can be driven away from the correct courses by other
agents. We carefully inspected all possible states that agents can find them-
selves in. In order to have a clearer insight and control over the agents inside
the disks we use the Finite State Machine2 instead of just plain algorithm
block in our code. The state of the agent is defined by 4 binary features: (a)
is agent inside the disk? (b) are where agent is walking to and what agent is
looking at the same target? (c) can agent see what he is looking at (or are
there obstacles in-between)? (d) has agent reached the final node?
Within each iteration of the main loop we check the states of the agents.
The states can be changed by agents themselves -- e.g. agent has crossed the
border of the disk, or by our commands -- e.g. agent is ordered to walk to
another target. Consider these circumstances: the agent has managed to see
his next target and now he walks towards this next target -- he is in state S1.
But now he loses the eye contact with this new target and finds himself in state
S2. The program logic reacts to such a state by transiting to the state S3: start
looking at the previous target and walk towards this previous target. Based
on what happens next we can order the transition to another state or just
wait for the agent to change the state himself. By careful examination of all
possible circumstances we can make sure that our states and their transitions
can handle all possible scenarios.
2 https://en.wikipedia.org/wiki/Finite-state_machine
A-Evac: the evacuation simulator for stochastic environment
9
Fig. 7 The roadmap and local movement
On Figure 8c) we show how agents are passing through a HOLE. Due to
our concept of the disks (where searching for new targets takes place) and due
to the internals of RVO2 we gain the desired effect of agents not crossing the
very center of the disk. Instead, the agents can walk in parallel and advance to
another target which looks natural and doesn't create an unnecessary queue
of agents eager to cross the very center of the HOLE.
Fig. 8 RVO2 at its work of resolving collisions: (a) agents on direct collision courses and
(b) their calculated collision-avoiding courses, (c) three agents crossing a HOLE in parallel.
3.4 Evacuation under fire and smoke
Each a-evac simulation is preceded with the simulation of the fire. We have
only tested a-evac with CFAST [24,21]. CFAST writes its output to csv files.
Agent is moving towards this targetAgent is keeping an eye on this targetA1A2A4A3A5P1The global roadmap for agent A1:G_A1=[P1,P2,P3,P4]A1P2P3P4 10
Adam Krasuski, Karol Krenski
We need to query these CFAST results quite a bit, therefore we transform and
store these results in a fast in-memory relational database3. For each frame
of time we are repeatedly asking the same questions: (a) given the agent's
coordinates, which room is he in? (b) what are the current conditions in this
room?
When it comes to (b), the environment effects on the agent can be: (b.1)
limited visibility (eyes), (b.2) poisonous gases (nose) and (b.3) temperature
in the room (body). Both (b.1) and (b.2) are read from the default (but con-
figurable) height of 1.8 m. There are always two zones in CFAST, which are
separated at a known height, so we need to read the conditions from the correct
zone, based on where our 1.8 m belongs.
The value of visibility (OD -- optical density) affects agent's speed. We use
the relation proposed in [12] following the FDS+Evac [23]:
(cid:110)
(cid:0)1 +
· Ks
β
α
(cid:1)(cid:111)
vpref
n
vn,min, vpref
(Ks) = max
(1)
where: Ks is the extinction coefficient ([Ks] = m−1) calculated as OD/log10e
according to [19,20], vn,min is the minimum speed of the agent An and equals
0.1 · vpref
(agent's preferable velocity), and α, β are the coefficients defined
n
in [12].
n
Setting the minimal value of speed means that the agent does not stop in
thick smoke. They continue moving until the value of incapacitated Fractional
Effective Dose (FED) is exceeded, which is fatal to the agent. FED is calculated
from CFAST-provided amounts of the following species in the agent environ-
ment: carbon monoxide (CO), hydrogen cyanide (HCN), hydrogen chloride
(HCl), carbon dioxide (CO2) and oxygen (O2) by the equation [26,23]:
F EDtotal = (F EDCO + F EDHCN + F EDHCl) × HVCO2 + F EDO2
(2)
where HVCO2 is the hyperventilation induced by the concentration of CO2.
Following are the formulas for the terms in the above equation. FEDs are given
in ppm and time t in minutes. C stands for concentration of the species in %:
(cid:90) t
(cid:90) t
0
0
exp(cid:0) CHCN (t)
(cid:1)
43
220
F EDCO =
2.764 × 10−5(CCO(t))1.036dt
F EDHCN =
− 0.0045dt
(3)
(4)
Based on [17]
In contrast to the model applied in Evac, CFAST does not allow for proac-
tive correction of effect of nitrogen dioxide -- CCN = CHCN − CN O2 . Therefore
this effect is not included in the calculations.
3 https://www.sqlite.org/
A-Evac: the evacuation simulator for stochastic environment
(cid:90) t
CHCl(t)
dt
F EDHCl =
1900
0
11
(5)
(6)
(7)
Based on [28,17]
F EDO2 =
(cid:90) t
0
HVCO2 =
60 · exp(cid:2)8.13 − 0.54(20.9 − CO2 (t))(cid:3)
exp(cid:0)0.1903 · CCO2(t) + 2.0004(cid:1)
dt
7.1
There are few quantitative data from controlled experiments concerning
the sublethal effect of the smoke on people. In works [28, 5, 26,14,29] sublethal
effect in a form of incapacitation (IC50), escape ability (EC50), lingering health
problems and minor effects were reported. Incapacitation was inferred from
lethality data, to be about one-third to one-half of those required for lethal-
ity. The mean value of the ratios of the IC50 to the LC50 was 0.50 and the
standard deviation 0.21, respectively. In [14] a scale for effects based on FED
was introduced. The three ranges were proposed: 1 FED indicating lethality,
0.3 FED indicating incapacitation and 0.01 FED indicating no significant sub-
lethal effects should occur. We propose based on this data a scale for sublethal
effects of smoke for evacuees as presented in Table 1.
F EDtotal affects the agent's movement in the smoke. For F EDtotal > 0.3,
the smoke inhalation leads to sublethal effects [5] -- the agent is not able to
find safety from the fire and just stays where he is. For F EDtotal > 1 we model
lethal effects. We later use these effects in the final risk assessment.
Table 1 summarizes the FED effects on human health, what is our origi-
nal proposition for evaluation of sublethal effect of smoke. These ranges are
incorporeted in Aamks. It is based on the following works: [28,5,26,14,29]
Table 1 FED effects on human health in Aamks.
FED
< 0.01
[0.01 − 0.3[
[0.3 − 1[
≥ 1
Effect on human health
Minor or negligible
Low -- short period of hospitalization
Heavy -- lingering health problems or permanent disability
Lethal
3.5 Probabilistic evacuation modeling
This section presents the internals of our probabilistic evacuation model, which
we find distinct across the available, similar software.
Table 2 presents the distributions of the input parameters used in Aamks.
Each of the thousands of simulations in a single project is initialized with some
12
Adam Krasuski, Karol Krenski
random input setup according to these distributions. Aamks has a library
of the default parameters values for important building categories (schools,
offices, malls etc.). The Aamks users should find it convenient to have all the
distributions in a library, but they may choose to alter these values, which is
possible.
Most of the data in table 2 come from the standards and from other models,
mostly FDS/Evac. Following are some comments on table 2.
Aamks puts much attention to the pre-evacuation time [9], which models
how people lag before evacuating after the alarm has sounded. Positions 7. and
8. are separated, because the behaviour of humans in the room of fire origin is
distinct. We compile two regulations C/VM2 Verification Method: Framework
for Fire Safety Design [30] and British Standard PD 7974-6:2004 [2] in order
to get the most realistic, probability-based pre-evacuation in the room of fire
origin and in the rest of the rooms.
The Horizontal/Vertical speed (unimpeded, walking speed of an agent) is
based on [18,13,25,15].
Speed in the smoke is modeled by formula 1.
Table 2 Parameters of the distributions for the exemplary scenario.
Parameter
Distribution
µ/min
σ/max
Denstity in rooms [m2/humans]
Densitiy on corridors [m2/humans]
Densitiy in stairways [m2/humans]
Human location in the compartment x
Human location in the compartment y
normal
normal
normal
uniform
uniform
Time of the alarm
log-normal
Pre-evacuation time in the room of fire origin
uniform
5
20
50
0
0
0.7
0
2
3
3
room width
room depth
0.2
30
Pre-evacuation time in other compartments
log-normal
3.04
0.142
Horizontal speed
Vertical speed
α for speed in the smoke
β for speed in the smoke
normal
normal
normal
normal
1.2
0.7
0.706
-0.057
0.03
0.2
0.2
0.069
0.015
0.97
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13. Humans taking an alternative evacuation route
binomial
The randomness of the simulations comes from the random number gen-
erator's seed. We save the seed for each simulation so that we can repeat the
very same simulation, which is useful for debugging and visualisation.
We register all the random input setups and the corresponding results in
the database. We are expecting to research at some point the relationships in
these data with data mining or sensitivity analysis.
A-Evac: the evacuation simulator for stochastic environment
13
The final result of Aamks is the compilation of multiple simulations as a
set of distributions i.e. F-N curves. The F-N curves were created as in [12].
Figure 9 depicts the exemplary results.
Fig. 9 The results of evacuation modeling as F-N curves.
3.6 Visualization
In Aamks we use a 2D visualization for supervising the potential user's faults
in his CAD work (e.g. rooms with no doors (Figure 10)), for the final results,
and for our internal developing needs. We use a web based technology which
allows for displaying both static images and the animations of evacuees.
Fig. 10 2D visualization: animation of evacuees
0510152025303540Number of people0.000.050.100.150.200.250.30ProbabilityFatalities051015202530Number of people0.000.050.100.150.200.25ProbabilitySerious injuries051015202530Number of people0.000.020.040.060.080.100.120.140.160.18ProbabilityMinor injuries 14
Adam Krasuski, Karol Krenski
We also have a web based 3D visualization made with WebGL Threejs.
This subsystem displays realistic animations of humans during their evacuation
under fire and smoke. (Figure 11).
Fig. 11 3D visualization
4 Quality and the performance of a-evac
Below we evaluate the quality of a-evac as described in [9,27] as well as it's
computer performance.
4.1 Verification of a-evac
Verification and validation deals with how close the results of the simulations
to the reality are. We took care to be compliant with the general develop-
ment recommendations [9] by: (1) obeying good programming practices, (2)
verifying intermediate simulation outputs, (3) comparing simulation outputs
against the analytical results, and (4) creating debugging animations.
The are three types of errors that can be generated by our software: a)
error in deterministic modeling of single scenario, b) error of Monte Carlo
approximation, c) statistical error -- disturbance.
For the first type of error we applied the methods proposed in [27]. The
proposed tests are organized in five core components: (1) pre-evacuation time,
(2) movement and navigation, (3) exit usage, (4) route availability, and (5)
flow conditions/constraints. For each category there are detailed tests for the
geometry, the scenario and the expected results. The results are in table 3.
4 The method is not straightforward
A-Evac: the evacuation simulator for stochastic environment
15
Table 3 The results of Aamks tests
Id. Name of the tests
1.
2.
3.
4.
5.
6.
7.
8.
9.
10. Group behaviours
11.
12.
13.
14. Affiliation
15. Dynamic availability of exits
16. Congestion
17. Maximum flow rates
Test code Results
Verif.1.1
Pre-evacuation time distributions
Verif.2.1
Speed in a corridor
Verif.2.2
Speed on stairs
Verif.2.3
Movement around the corner
Assigned occupant demographics
Verif.2.4
Reduced visibility vs walking speed Verif.2.5
Verif.2.6
Occupant incapatication
Verif.2.7
Elevator usage
Verif.2.8
Horizontal counter-flows (rooms)
Verif.2.9
Verif.2.10
Verif.3.1
Verif.3.2
Verif.3.3
Verif.4.1
Verif.5.1
Verif.5.2
People with movement disabilities
Exit route allocation
Social influence
OK
OK
OK 4
OK
OK
OK
OK
--
OK
--
--
OK
--
--
OK
OK
OK
RVO2, the core library of a-evac which drives the local movement, was also
evaluated in [33]. The conclusions are that RVO2 is of the quality comparable
with the lattice gas and social force models. The social force model is commonly
used in a number of evacuation software.
The above is the evaluation of a single, deterministic simulation. However,
the final result is the compilation of the whole collection of such single simula-
tions -- this is how we get the big picture of the safety of the inquired building.
The picture is meant to present risk. Probability of risk is calculated as a share
of simulations resulted in fatalities, in the total number of simulations.
The accuracy of this evaluation depends on the method applied -- stochas-
tic simulations. The error is proportional to the square root of number sim-
ulations. Namely for the discrete Bernoulli probability distributions used for
example for evaluation of probability of scenario with fatalities, the error is
calculated as follows:
(cid:114)
σn = 1.96
pn(1 − pn)
n
;
(8)
where: pn is the probability of fatalities obtained as a number of simulations
resulted in fatalities to the total number of simulations, n is the number of
simulations.
We are aware of the third type of error which may be generated by the
application. The input for Aamks is a set of various probability distributions
what may occasionally generate unreal scenario. For example an evacuee who
moves very slowly on corridors and very fast on stairs. In most cases these
errors are related to the other parts of Aamks i.e. probabilistic fire modeling.
16
Adam Krasuski, Karol Krenski
However, fire environment impacts the evacuation. This error can be evaluated
by comparison of data generated by Aamks with real statistics. So far we do
not have idea how to tackle this problem efficiently. We consider to evaluate
this error by launching simulations for the building stock and check whether
we reconstruct historical data. This method is very laborious and not justified
at that moment, because our application still lacks some models i.e. fire service
intervention -- what has significant impact on fire.
4.2 Performance of the Model
The main loop of Aamks processes all agents in the time iteration. Table 4
summarizes how costly the specific calculations for a single agent within a
single time iteration are. The tests were performed on computer with Intel
Core i5-2500K CPU at 3.30 GHz with 8 GB of RAM.
Table 4 The costs of a single loop iteration per agent
Activity
Position update
Velocity update
Speed update
State update
Goals update
FED update
Time update
Time
8.12 × 10−6 s
1.07 × 10−5 s
8.50 × 10−5 s
7.75 × 10−8 s
9.25 × 10−8 s
1.01 × 10−4 s
1.22 × 10−7 s
Total share
3.97 %
5.26 %
41.63 %
0.03 %
0.04 %
48.98 %
0.06 %
The total time of a single step of the simulation for one agent is 2 × 10−4 s
and it grows linearly with the number of agents. The speed and FED calcu-
lations are most costly, because they both make database queries against the
fire conditions in the compartment. The time step for a-evac iteration is 0.05 s.
There is no significant change in fire conditions within this time frame. There-
fore for performance optimization we update speed and FED every 20-th step
of the simulation.
5 Discussion
Vertical evacuation is troublesome and not implemented. There is RVO2 3D,
but it is for aviation where agents can pass above each other -- clearly not for
our needs. Besides, we think things look actually better in 2D. We like the
idea that vertical evacuation can be still considered 2D, just rotated, and we
plan to move in this direction.
A-Evac: the evacuation simulator for stochastic environment
17
A-evac does not model the social or group behaviours. However, it is diffi-
cult to evaluate, how much the lack of such functionalities impacts the resulting
probability distributions.
In the workflow we run a CFAST simulation first. Then a-evac simulation
runs on top of CFAST results. This sequential procedure has it's drawbacks,
e.g. we don't know how long the CFAST simulation should last to produce
enough data for a-evac, so we run "too much" CFAST for safety. Also, evacuees
cannot trigger any events such as opening the door, etc. We considered a closer
a-evac-CFAST integration.
There seems to be lot's of space for improvement in Aamks. We work
with the practitioners and know the reality of fire engineering. We know the
limitations of our current implementations and most of them can be addressed
-- there are models and approaches that we can implement and the major
obstacle are the limits of our team resources. Therefore we invite everyone
interested to join our project at http://github.com/aamks.
6 Conclusion
Aamks is actively developed since 2016 and we are truly engaged in making it
better. The software, though not really ready for end-users has already served
as a support for commercial projects and fire engineers and scientists regard
Aamsk as having potential. The stochastic based workflow of Aamks is not a
new concept. There are opinions in the community that this approach is how
fire engineering should be done. Since no wide-used implementation has been
created so far, this is an additional motivation that drives our project.
Acknowledgements Aamks is hosted at github at: http://gitub.com/aamks. This work
was supported in part by ConsultRisk LTD and F&K Consulting Engineers LTD under
research grants CR-2016 SiMo (Simulation Modules) .
References
1. Association of Building Engineers: Part 7: Probabilistic risk assessment. In: PD 7974-
6:2004, The application of fire safety engineering principles to fire safety design of build-
ings. British Standard (2003)
2. Association of Building Engineers: The application of fire safety engineering principles
to fire safety design of buildings -- Part 6: Human factors: Life safety strategies Occupant
evacuation, behaviour and condition (Sub-system 6) (2004). PD 7974-6:2004
3. Van den Berg, J., Lin, M., Manocha, D.: Reciprocal velocity obstacles for real-time
multi-agent navigation. In: Robotics and Automation, 2008. ICRA 2008. IEEE Inter-
national Conference on, pp. 1928 -- 1935. IEEE (2008)
4. Brandes, U., Erlebach, T.: Network analysis: methodological foundations, vol. 3418.
Springer Science & Business Media (2005)
5. Bryner, R.G.G.N.P.: Combustion products and their effects on life safety. Fire Protec-
tion Handbook pp. 6 -- 11 (2008)
6. Chazelle, B.: A theorem on polygon cutting with applications. In: Foundations of Com-
puter Science, 1982. SFCS'08. 23rd Annual Symposium on, pp. 339 -- 349. IEEE (1982)
7. Christian, R.P., Casella, G.: Monte Carlo statistical methods. New York: Springer (2007)
18
Adam Krasuski, Karol Krenski
8. Colaborative work: FRAME Fire Risk Assessment for Engineering. FRAME -- Valida-
tion. Tech. rep. (2014)
9. Cuesta, A., Abreu, O., Alvear, D.: Evacuation Modeling Trends. Springer (2015)
10. Delaunay, B.: Sur la sph`ere vide. Bulletin of Academy of Sciences of the USSR pp.
793 -- 800 (1934)
11. Fiorini, P., Shiller, Z.: Motion planning in dynamic environments using velocity obsta-
cles. The International Journal of Robotics Research 17(7), 760 -- 772 (1998)
12. Frantzich, H., Nilsson, D.: Utrymning genom tat rok: beteende och forflyttning. Univ.
(2003)
13. Fruin, J.J.: Pedestrian planning and design. Tech. rep. (1971)
14. Gann, Richard G and Averill, Jason D and Butler, KM and Jones, WW and Mulholland,
GW and Neviaser, JL and Ohlemiller, TJ and Peacock, RD and Reneke, PA and Hall
Jr, JR: Sublethal effects of smoke on survival and health. In: Human Behaviour in Fire,
Proceedings of the 2 nd Int. Symposium, MIT, Cambridge, USA, pp. 285 -- 296 (2001)
15. Helbing, D., Farkas, I.J., Molnar, P., Vicsek, T.: Simulation of pedestrian crowds in
normal and evacuation situations. Pedestrian and evacuation dynamics 21(2), 21 -- 58
(2002)
16. Hostikka, S., Keski-Rahkonen, O.: Probabilistic fire simulator -- Monte Carlo simulation
tool for fire scenarios (2002)
17. Hull, T.R., Stec, A.A., Paul, K.: Hydrogen chloride in fires. Fire Safety Science 9,
665 -- 676 (2008)
18. Hurley, M.J., Gottuk, D.T., Hall Jr, J.R., Harada, K., Kuligowski, E.D., Puchovsky, M.,
Watts Jr, J.M., Wieczorek, C.J., et al.: SFPE handbook of fire protection engineering.
Springer (2015)
19. Jin, T.: Visibility through fire smoke in main reports on production, movement and
control of smoke in buildings. Japanese Association of Fire Science and Engineering pp.
100 -- 153 (1974)
20. Jin, T.: Visibility and human behavior in fire smoke. The SFPE Handbook of Fire
Protection Engineering, ed. PJ DiNenno, D. Drysdale, CL Beyler, WD Walton, RLP
Custer, JR Hall, Jr. and JM Watts, Jr., 3rd edition, Society of Fire Protection Engi-
neers/National Fire Protection Association, Quincy, Massachusetts (2016)
21. Jones, W.W., Forney, G.P., Peacock, R.D., Reneke, P.A.: A technical reference for cfast:
an engineering tool for estimating fire and smoke transport. NIST TN 1431 (2017).
URL https://github.com/firemodels/cfast/releases
22. Kahneman, D., Klein, G.: Conditions for intuitive expertise: a failure to disagree. Amer-
ican Psychologist 64(6), 515 (2009)
23. Korhonen, T.: FDS + Evac -- Technical reference + users guide. VTT Technical Research
Centre of Finland (2016). Evac 2.5.2
24. Peacock, R.D., Jones, W., Reneke, P., Forney, G.: CFAST-consolidated model of fire
growth and smoke transport (version 7) user's guide. Citeseer (2017). URL https:
//github.com/firemodels/cfast/releases
25. Predtetschenski, W., Milinski, A.: Planning for foot traffic in buildings (1978)
26. Purser, D.A.: Toxicity assessment of combustion products. SFPE handbook of fire
protection engineering 3, 2 -- 6 (2002)
27. Ronchi, E., Kuligowski, E.D., Reneke, P.A., Peacock, R.D., Nilsson, D.: The process
of verification and validation of building fire evacuation models. NIST Technical Note
1822 (2013)
28. Speitel, L.C.: Toxicity assessment of combustion gases and development of a survival
model. Tech. rep., DTIC Document (1995)
29. Stec, A.A., Hull, T.R.: Fire toxicity. Elsevier (2010)
30. The Ministry of Business, Innovation and Employment: C/VM2 Verification Method:
Framework for Fire Safety Design (2013). New Zealand
31. Tversky, A., Kahneman, D.: Judgment under uncertainty: Heuristics and biases. Science
185(4157), 1124 -- 1131 (1974)
32. Van Den Berg, J., Guy, S.J., Lin, M., Manocha, D.: Reciprocal n-body collision avoid-
ance. In: Robotics research, pp. 3 -- 19. Springer (2011)
33. Viswanathan, V., Lee, C.E., Lees, M.H., Cheong, S.A., Sloot, P.: Quantitative com-
parison between crowd models for evacuation planning and evaluation. arXiv preprint
arXiv:1401.0366 (2014)
|
1203.5583 | 1 | 1203 | 2012-03-26T05:40:28 | Graph-Theoretic Characterizations of Structural Controllability for Multi-Agent System with Switching Topology | [
"cs.MA",
"eess.SY"
] | This paper considers the controllability problem for multi-agent systems. In particular, the structural controllability of multi-agent systems under switching topologies is investigated. The structural controllability of multi-agent systems is a generalization of the traditional controllability concept for dynamical systems, and purely based on the communication topologies among agents. The main contributions of the paper are graph-theoretic characterizations of the structural controllability for multi-agent systems. It turns out that the multi-agent system with switching topology is structurally controllable if and only if the union graph G of the underlying communication topologies is connected (single leader) or leader-follower connected (multi-leader). Finally, the paper concludes with several illustrative examples and discussions of the results and future work. | cs.MA | cs |
Graph-Theoretic Characterizations of Structural
Controllability for Multi-Agent System with
1
Switching Topology*
Xiaomeng Liua, Hai Linb∗, Senior Member, IEEE, and Ben M. Chena,
Fellow, IEEE
Abstract
This paper considers the controllability problem for multi-agent systems. In particular, the structural
controllability of multi-agent systems under switching topologies is investigated. The structural control-
lability of multi-agent systems is a generalization of the traditional controllability concept for dynamical
systems, and purely based on the communication topologies among agents. The main contributions of
the paper are graph-theoretic characterizations of the structural controllability for multi-agent systems.
It turns out that the multi-agent system with switching topology is structurally controllable if and only
if the union graph G of the underlying communication topologies is connected (single leader) or leader-
follower connected (multi-leader). Finally, the paper concludes with several illustrative examples and
discussions of the results and future work.
I. Introduction
Due to the latest advances in communication and computation, the distributed control and
coordination of the networked dynamic agents has rapidly emerged as a hot multidisciplinary
research area [Lawton et al., 2002][Dunbar et al., 2006][Bliman et al., 2008], which lies at the
intersection of systems control theory, communication and mathematics. In addition, the advances
of the research in multi-agent systems are strongly supported by their promising civilian and
military applications, such as cooperative control of unmanned air vehicles(UAVs), autonomous
∗Corresponding author. Email: [email protected] tel 574-6313177 fax 574-6314393
aDept. of ECE, National University of Singapore, Singapore bDept of EE, Univ of Notre Dame, Notre Dame, IN 46556, USA
August 21, 2018
DRAFT
2
underwater vehicles(AUVs), space exploration, congestion control in communication networks,
air traffic control and so on [Tomlin et al., 1998][How et al., 2004]. Much work has been done
on the formation stabilization and consensus seeking. Approaches like graph Laplacians for
the associated neighborhood graphs, artificial potential functions, and navigation functions for
distributed formation stabilization with collision avoidance constraints have been developed.
Furthermore, inspired by the cooperative behavior of natural swarms, such as bee flocking, ant
colonies and fish schooling, people try to obtain experiences from how the group units make
their whole group motions under control just through limited and local interactions among them.
The control of such large scale complex systems poses several new challenges that fall beyond
the traditional methods. Part of the difficulty comes from the fact that the global behavior of
the whole group combined by multiple agents is not a simple summation of the individual
agent's behavior. Actually, the group behavior can be greatly impacted by the communication
protocols or interconnection topology between the agents, which makes the global behavior
display high complexities. Hence, the cooperative control of multi-agent systems is still in its
infancy and attracts more and more researchers' attention. One basic question in multi-agent
systems that attracts control engineers' interest is what is the necessary information exchanging
among agents to make the whole group well-behaved, e.g., controllable. This can be formulated as
a controllability problem for multi-agent systems under the leader-follower framework. Roughly
speaking, a multi-agent system is controllable if and only if we can drive the whole group
of agents to any desirable configurations only based on local interactions between agents and
possibly some limited commands to a few agents that serve as leaders. The basic issue is the
interplay between control and communication. In particular, we would like to investigate what
is the necessary and/or sufficient condition on the graph of communication topologies among
agents for the controllability of multi-agent systems.
This multi-agent controllability problem was first proposed in [Tanner, 2004], which formu-
lated it as the controllability of a linear system and proposed a necessary and sufficient algebraic
condition based on the eigenvectors of the graph Laplacian. Reference [Tanner, 2004] focused on
fixed topology situation with a particular member which acted as the single leader. The problem
was then developed in [Ji et al., 2006][Ji et al., 2007][Rahmani et al., 2006]
[Ji et al., 2008], and got some interesting results. For example, in [Ji et al., 2007], it was shown
that a necessary and sufficient condition for controllability is not sharing any common eigenvalues
August 21, 2018
DRAFT
3
between the Laplacian matrix of the follower set and the Laplacian matrix of the whole topology.
However, it remains elusive on what exactly the graphical meaning of these algebraic conditions
related to the Laplacian matrix. This motivates several research activities on illuminating the
controllability of multi-agent systems from a graph theoretical point of view. For example,
a notion of anchored systems was introduced in [Rahmani et al., 2006] and it was shown that
symmetry with respect to the anchored vertices makes the system uncontrollable. However, so far
the research progress using graph theory is quite limited and a satisfactory graphical interpretation
of these algebraic controllability conditions turns out to be very challenging. Besides, the weights
of communication linkages among agents have been demonstrated to have a great influence
on the behavior of whole multi-agent group (see [Moreau, 2005] as an example). However,
in the previous multi-agent controllability literature [Tanner, 2004]-[Rahmani et al., 2006], the
communication weighting factor is usually ignored. One classical result under this no weighting
assumption is that a multi-agent system with complete graphical communication topology is
uncontrollable [Tanner, 2004]. This is counter-intuitive since it means each agent can get direct
information from each other but this leads to a bad global behavior as a team. This shows that too
much information exchange may damage the controllability of multi-agent system. In contrast,
if we set weights of unnecessary connections to be zero and impose appropriate weights to other
connections so as to use the communication information in a selective way, then it is possible
to make the system controllable [Zamani et al., 2009].
In this paper, motivated by the above observation, the weighting factor is taken into ac-
count for multi-agent controllability problem. In particular, a new notion for the controlla-
bility of multi-agent systems, called structural controllability, which was proposed by us in
[Zamani et al., 2009], is investigated directly through the graph-theoretic approach for control
systems. The communication topology of whole multi-agent system is described by a weighted
graph and the system is called structurally controllable if one may find a set of weights such
that the corresponding multi-agent system is controllable in a classical sense. The structural
controllability reveals under certain topology whether it is possible to make the whole multi-
agent system well-behaved, i.e. controllable here through suitable choice of the communica-
tion weights. From another angle, it helps to bring to light the effects of the communication
topology on the controllability property of multi-agent systems without worrying about the
influence of weighting factors. It turns out that this controllability notation only depends on
August 21, 2018
DRAFT
4
the topology of the communication scheme in the case of a single leader under a fixed topology
[Zamani et al., 2009].
Another novelty in this paper is the successful investigation of impacts of switching topologies
on the multi-agent controllability property, for which there is barely graphical interpretation to
the best of our knowledge. Note that the results in [Tanner, 2004][Ji et al., 2006][Ji et al., 2007]
[Rahmani et al., 2006][Ji et al., 2008][Zamani et al., 2009] are all focused on multi-agent sys-
tems under fixed communication topologies which may restrict their impacts on real applications.
In many applications, it may become impossible to keep the communication topology fixed for
the whole period. Therefore, it is of practical importance to consider time varying communication
topologies. A natural framework to study the time variance of communication topology is through
switched systems, see e.g., [Lin et al., 2007][Lin et al., 2009][Sun et al., 2002]. In this paper,
we will focus on multi-agent systems under switching topologies in the framework of switched
systems. Some early efforts have been observed in the literature. Necessary and sufficient al-
gebraic conditions for the controllability of multi-agent systems under switching topology were
derived in [Ji et al., 2008][Liu et al., 2008] based on the developments of controllability study
in switched systems. However, these algebraic results lacks graphically intuitive interpretations,
which are important since they can provide us significant guidelines for the communication
protocol design for multi-agent systems. Therefore, this paper aims to fill this gap and propose
graphic interpretations of these algebraic conditions for the controllability of multi-agent sys-
tems under switching topology. In particular, we follow the setup in [Zamani et al., 2009] and
investigate the structural controllability of multi-agent systems with switching communication
topologies and for both single leader and multi-leader cases. It is assumed that the leaders act
as the external or control signal and will not be affected by any other group members. Based
on this structural controllability, we propose necessary and sufficient graph theoretic conditions
for the structural controllability of multi-agent system with switching topologies. It turns out
that the multi-agent system with switching topology is structurally controllable if and only if
the union graph G of the underlying communication topologies is connected (single leader) or
leader-follower connected (multi-leader). Some examples are given to underscore our theoretical
analysis.
The outline of this paper is as follows: In Section II, we introduce some basic preliminaries
and the problem formulation, followed by structural controllability study in Section III, where
August 21, 2018
DRAFT
5
a graphic necessary and sufficient condition for the structural controllability under single leader
case is given. In Section IV, graphical interpretation of structural controllability of multi-leader
multi-agent system is proposed. In Section V, some examples are presented to give the readers
deeper understanding of our theoretical results. Finally, some concluding remarks are drawn in
the paper.
II. Preliminaries and Problem Formulation
A. Graph Theory Preliminaries
A weighted graph is an appropriate representation for the communication or sensing links
among agents because it can represent both existence and strength of these links among agents.
The weighted graph G with N vertices consists of a vertex set V = {v1, v2, . . . , vN} and an edge
set I = {e1, e2, . . . , eN′}, which is the interconnection links among the vertices. Each edge in
the weighted graph represents a bidirectional communication or sensing media. Two vertices are
known to be neighbors if (i, j) ∈ I, and the number of neighbors for each vertex is its valency.
An alternating sequence of distinct vertices and edges in the weighted graph is called a path.
The weighted graph is said to be connected if there exists at least one path between any distinct
vertices, and complete if all vertices are neighbors to each other.
The adjacency matrix, A is defined as
A(i, j) =
Wi j
(i, j) ∈ I,
0
otherwise,
(1)
where Wi j , 0 stands for the weight of edge ( j, i). Here, the adjacency matrix A is V × V
and . is the cardinality of a set.
The Laplacian matrix of a graph G, denoted as L(G) ∈ RV×V or L for simplicity, is defined
as
L(i, j) =
Σ j∈Ni Wi j
i = j,
−Wi j
i , j and (i, j) ∈ I,
0
otherwise.
(2)
B. Multi-agent Structural Controllability with Switching Topology
Specifically, controllability problem usually cares about how to control N agents based on the
leader-follower framework. Take the case of single leader as example. Without loss of generality,
August 21, 2018
DRAFT
6
assume that the N-th agent serves as the leader and takes commands and controls from outside
operators directly, while the rest N − 1 agents are followers and take controls as the nearest
neighbor law.
Mathematically, each agent's dynamics can be seen as a point mass and follows
The control strategy for driving all follower agents is
xi = ui.
ui = − X j∈Ni
wi j(xi − x j) + wiixi,
(3)
(4)
where Ni is the neighbor set of the agent i (could contain i itself), and wi j is weight of the edge
from agent j to agent i. On the other hand, the leader's control signal is not influenced by the
followers and needs to be designed, which can be represented by
xN = uN.
(5)
In other words, the leader affects its nearby agents, but it does not get directly affected by
the followers since it only accepts the control input from an outside operator. For simplicity, we
will use z to stand for xN in the sequel. It is known that the whole multi-agent system under
fixed communication topology can be written as a linear system:
,
(6)
x
z
=
A B
0 0
x
z
+
0
uN
where A ∈ R(N−1)×(N−1) and B ∈ R(N−1)×1 are both sub-matrices of the corresponding graph
Laplacian matrix −L
The communication network of dynamic agents with directed information flow under link
failure and creation can usually described by switching topology. Under m switching topologies,
it is clear that the whole system equipped with m subsystems can be written in a compact form
x
z
=
Ai Bi
0
0
x
z
+
0
uN
,
x = Aix + Biz,
z = uN,
or, equivalently,
August 21, 2018
(7)
(8)
DRAFT
7
where i ∈ {1, . . . , m}. Ai ∈ R(N−1)×(N−1) and Bi ∈ R(N−1)×1 are both sub-matrices of the correspond-
ing graph Laplacian matrix −L. The matrix Ai reflects the interconnection among followers, and
the column vector Bi represents the relation between followers and the leader under corresponding
subsystems. Since the communication topologies among agents are time-varying, so the matrices
Ai and Bi are also varying as a function of time. Therefore, the dynamical system described in
(7) can be naturally modeled as a switched system (definition can be found latter).
Considering the structural controllability of multi-agent system, system matrices Ai and Bi,
i ∈ {1, . . . , m} are structured matrices, which means that their elements are either fixed zeros or
free parameters. Fixed zeros imply that there is no communication link between the corresponding
agents and the free parameters stand for the weights of the communication links. Our main task
here is to find out under what kinds of communication topologies, it is possible to make the group
motions under control and steer the agents to the specific geometric positions or formation as a
whole group. Now this controllability problem reduces to whether we can find a set of weights
wi j such that the multi-agent system (7) is controllable. Then the controllability problem of
multi-agent system can now be formulated as the structural controllability problem of switched
linear system (7):
Definition 1: The multi-agent system (7) with switching topology, whose matrix elements are
zeros or undetermined parameters, is said to be structurally controllable if and only if there exist
a set of communication weights wi j that can make the system (7) controllable in the classical
sense.
C. Switched Linear System and Controllability Matrix
In general, a switched linear system is composed of a family of subsystems and a rule that
governs the switching among them, and is mathematically described by
x(t) = Aσ(t)x(t) + Bσ(t)u(t),
(9)
where x(t) ∈ Rn are the states, u(t) ∈ Rr, are piecewise continuous input, σ : [t0, ∞) → M ,
{1, . . . , m} is the switching signal. System (9) contains m subsystems (Ai, Bi), i ∈ {1, . . . , m} and
σ(t)= i implies that the ith subsystem (Ai, Bi) is activated at time instance t
For the controllability problem of switched linear systems, a well-known matrix rank condition
was given in [Sun et al., 2004]:
August 21, 2018
DRAFT
Lemma 1: ([Sun et al., 2004]) If matrix:
[B1, . . . , Bm, A1B1, . . . , AmB1, . . . , AmBm, A2
1B1, . . . , AmA1B1,
. . . , A2
1Bm, . . . , AmA1Bm, . . . , An−1
1 B1, . . . , AmAn−2
1 B1, . . . ,
8
(10)
A1An−2
m Bm, . . . , An−1
m Bm]
has full row rank n, then switched linear system (9) is controllable, and vice versa.
This matrix is called the controllability matrix of the corresponding switched linear system
(9).
III. Structural Controllability of Multi-agent System with Single Leader
The multi-agent system with a single leader under switching topology has been modeled as
switched linear system (7). Before proceeding to the structural controllability study, we first
discuss the controllability of multi-agent system (7) when all the communication weights are
fixed.
After simple calculation, the controllability matrix of switched linear system (7) can be shown
to have the following form:
0,
1,
. . .
. . .
, 0, B1,
0,
, 1,
. . . , Bm, A1B1
0,
0,
0,
. . . , A1AN−3
m Bm,
0,
0,
. . . , AN−2
m Bm
0,
0
This implies that the controllability of the system (7) coincides with the controllability of the
following system:
x = Aix + Biz
i ∈ {1, . . . , m} .
(11)
Which is the extracted dynamics of the followers that correspond to the x component of the
equation. Therefore,
Definition 2: The multi-agent system (7) is said to be structurally controllable under leader z
if system (11) is structurally controllable under control input z.
For simplicity, we use (Ai, Bi)
sequel.
i ∈ {1, . . . , m} to represent switched linear system (11) in the
In (11), each subsystem (Ai, Bi) can be described by a directed graph [Lin, 1974]:
Definition 3: The representation graph of structured system (Ai, Bi) is a directed graph Gi,
with vertex set Vi = Xi ∪ Ui, where Xi = {x1, x2, . . . , xn}, which is called state vertex set and
August 21, 2018
DRAFT
9
Ui = {u1, u2, . . . , ur}, which is called input vertex set, and edge set Ii = IUiXi ∪ IXiXi, where
IUiXi = {(up, xq)Bqp , 0, 1 ≤ p ≤ r, 1 ≤ q ≤ n} and IXiXi = {(xp, xq)Aqp , 0, 1 ≤ p ≤ n, 1 ≤ q ≤ n}
are the oriented edges between inputs and states and between states defined by the interconnection
matrices Ai and Bi above. This directed graph (for notational simplicity, we will use digraph
to refer to directed graph) Gi is also called the graph of matrix pair (Ai, Bi) and denoted by
Gi(Ai, Bi).
For each subsystem, we have got a graph Gi with vertex set Vi and edge set Ii to represent
the underlaying communication topologies. As to the whole switched system, the representing
graph, which is called union graph, is defined as follows:
Definition 4: The switched linear system (11) can be represented by a union digraph, defined
as a flow structure G. Mathematically, G is defined as
G1 ∪ G2 ∪ . . . ∪ Gm = {V1 ∪ V2 ∪ . . . ∪ Vm;
(12)
I1 ∪ I2 ∪ . . . ∪ Im}
Remark 1: It turns out that union graph G is the representation of linear structured system:
(A1 + A2 + . . . + Am, B1 + B2 + . . . + Bm).
Remark 2: In lots of literature about controllability of multi-agent systems [Tanner, 2004]-
[Liu et al., 2008], the underlying communication topology among the agents is represented by
undirected graph, which means that the communication among the agents is bidirectional. Here
we still adopt this kind of communication topology. Then wi j and w ji are free parameters or zero
simultaneously (in numerical realization, the values of wi j and w ji can be chosen to be different).
Besides, one edge in undirected graph can be treated as two oriented edges. Consequently, even
though all the analysis and proofs for structural controllability of multi-agent systems are based
on the directed graph (the natural graphic representation of matrix pair (Ai, Bi) is digraph), the
final result will be expressed in undirected graph form.
Before proceeding further, we need to introduce two definitions which were proposed in
[Lin, 1974] for linear structured system x = Ax + Bu first:
Definition 5: ([Lin, 1974]) The matrix pair (A, B) is said to be reducible or of form I if there
exist permutation matrix P such that they can be written in the following form:
PAP−1 =
0
A11
A21 A22
, PB =
0
B22
,
August 21, 2018
(13)
DRAFT
10
where A11 ∈ Rp×p , A21 ∈ R(n−p)×p, A22 ∈ R(n−p)×(n−p) and B22 ∈ R(n−p)×r.
Remark 3: Whenever the matrix pair (A, B) is of form I, the system is structurally uncontrol-
lable [Lin, 1974] and meanwhile, the controllability matrix Q = hB, AB, . . . , An−1Bi will have at
least one row which is identically zero for all parameter values [Glover et al., 1976]. If there is
no such permutation matrix P, we say that the matrix pair (A, B) is irreducible.
Definition 6: ([Lin, 1974]) The matrix pair (A, B) is said to be of form II if there exist
permutation matrix P such that they can be written in the following form:
hPAP−1, PBi =
P1
P2
,
(14)
where P2 ∈ R(n−k)×(n+r) , P1 ∈ Rk×(n+r) with no more than k − 1 nonzero columns (all the other
columns of P1 have only fixed zero entries).
Here we need to recall a known result in literature for structural controllability of multi-agent
system with fixed topology [Zamani et al., 2009]:
Lemma 2: The multi-agent system with fixed topology under the communication topology G
is structurally controllable if and only if graph G is connected.
This lemma proposed an interesting graphic condition for structural controllability in fixed
topology situation and revealed that the controllability is totally determined by the communication
topology. However, how about in the switching topology situation? According to lemma 1, once
we impose proper scalars for the parameters of the system matrix (Ai, Bi) to satisfy the full rank
condition, the multi-agent system (11) is structurally controllable. However, this only proposed
an algebraic condition. Do we still have very good graphic interpretation for the relationship
between the structural controllability and switching interconnection topologies? The following
theorem answers this question and gives a graphic necessary and sufficient condition for structural
controllability under switching topologies.
Theorem 1: The multi-agent system (11) with the communication topologies Gi, i ∈ {1, . . . , m}
is structurally controllable if and only if the union graph G is connected.
Proof: Necessity: Assume that the multi-agent switched system is structurally controllable,
we want to prove that the union graph G is connected, which is equivalent to that the system
has no isolated agents in the union graph G [Zamani et al., 2009].
August 21, 2018
DRAFT
11
Suppose that the union graph G is disconnected and for simplicity, we will prove by contradic-
tion for the case that there exits only one disconnected agent. The proof can be straightforwardly
extended to more general cases with more than one disconnected agents. If there is one isolated
agent in the union graph, there are two possible situations: the isolated agent is the leader or one
of the followers. On one hand, if the isolated agent is the leader, it follows that B1 + B2 + . . . + Bm
is identically a null vector. So every Bi is a null vector. Easily we can conclude that the
controllability matrix for the switched system is never of full row rank N − 1, which means
that the multi-agent system is not structurally controllable. On the other hand, if the isolated
agent is one follower, we get that the matrix pair (A1 + A2 + . . . + Am, B1 + B2 + . . . + Bm) is
reducible. By definition 5, the controllability matrix
[B1 + B2 + . . . + Bm,
(A1 + A2 + . . . + Am)(B1 + B2 + . . . + Bm),
, · · · ,
(A1 + A2 + . . . + Am)N−2(B1 + B2 + . . . + Bm)]
always has at least one row that is identically zero. Expanding the matrix yields
[B1 + B2 + . . . + Bm,
A1B1 + A2B1 + . . . + AmB1 + A1B2 + A2B2
+ . . . + AmB2 + . . . + A1Bm + A2Bm . . . + AmBm
, . . . ,
AN−2
1 B1 + A2AN−3
1 B1 + . . . + AN−2
m Bm].
The zero row is identically zero for every parameter. This implies that every component in this
matrix, such as Bi, AiB j and Ap
j Br, has the same row always to be zero. As a result, the
controllability matrix
i Aq
[B1, . . . , Bm, A1B1, . . . , AmB1, . . . , AmBm, A2
1B1, . . . , AmA1B1,
. . . , A2
1Bm, . . . , AmA1Bm, . . . , An−1
1 B1, . . . , AmAn−2
1 B1, . . . ,
A1An−2
m Bm, . . . , An−1
m Bm]
always has one zero row. Therefore, the multi-agent system (11) is not structurally controllable.
Until now, we have got the necessity proved.
August 21, 2018
DRAFT
Sufficiency: If the union graph G is connected, we want to prove that the multi-agent system
(11) is structurally controllable.
According to lemma 2, the connectedness of the union graph G implies that the corresponding
system (A1 + A2 + A3 + . . . + Am, B1 + B2 + B3 + . . . + Bm) is structurally controllable. Then there
exist some scalars for the parameters in system matrices that make the controllability matrix
12
[B1 + B2 + . . . + Bm,
(A1 + A2 + . . . + Am)(B1 + B2 + . . . + Bm),
, . . . ,
(A1 + A2 + . . . + Am)N−2(B1 + B2 + . . . + Bm)],
has full row rank N − 1. Expanding the matrix, it follows that the matrix
[B1 + B2 + . . . + Bm,
A1B1 + A2B1 + . . . + AmB1 + A1B2 + A2B2
+ . . . + AmB2 + . . . + A1Bm + A2Bm . . . + AmBm
, . . . ,
AN−2
1 B1 + A2AN−3
1 B1 + . . . + AN−2
m Bm],
has full rank N − 1. Next, we add some column vectors to the above matrix and get
[B1 + B2 + . . . + Bm, B2, . . . , Bm,
A1B1 + A2B1 + . . . + AmB1 + A1B2 + A2B2 + . . . + AmB2
+ . . . + A1Bm + A2Bm + . . . + AmBm, A2B1, A3B1, . . . , AmBm
, . . . ,
AN−2
1 B1 + A2AN−3
1 B1 + . . . + AN−2
m Bm, A2AN−3
1 B1, . . . , AN−2
m Bm].
This matrix still have N − 1 linear independent column vectors, so it has full row rank. Next,
subtract B2, . . . , Bm from B1 +B2 +. . .+Bm; subtract A2B1, . . . , AmBm from A1B1 +A2B1 +. . .+AmBm
and subtract A2AN−3
m Bm. Since this column
fundamental transformation will not change matrix rank, the matrix still has full row rank. Now
1 B1 + . . . + AN−2
m Bm from AN−2
1 B1, . . . , AN−2
1 B1 + A2AN−3
the matrix becomes
August 21, 2018
DRAFT
[B1, . . . , Bm, A1B1, . . . , AmB1, . . . , AmBm, A2
1B1, . . . , AmA1B1,
13
. . . , A2
1Bm, . . . , AmA1Bm, . . . , An−1
1 B1, . . . , AmAn−2
1 B1, . . . ,
A1An−2
m Bm, . . . , An−1
m Bm]
which is the controllability matrix of system (11) and has full row rank N − 1. Therefore, the
multi-agent system is structurally controllable.
IV. Structural Controllability of Multi-agent System with Multi-Leader
In the above discussion, we assume the multi-agent system has totally N agents and the N-th
agent serves as the leader and takes commands and controls from outside operators directly,
while the rest N − 1 agents are followers and take controls as the nearest neighbor law. In the
following part, we will discuss the situation that several agents are chosen as the leaders of the
whole multi-agent systems, which is actually an extension of single leader case.
Similar to the single leader case, the multi-agent system with multiple leaders is given by:
= ui,
xi
xN+ j = uN+ j,
i = 1, . . . , N
j = 1, . . . , nl
(15)
where N and nl represent the number of followers and leaders, respectively. xi indicates the state
of the ith agent, i = 1, . . . , N + nl.
The control strategy ui, i = 1, . . . , N for driving all follower agents is the same as the single
leader case. The leaders' control signal is still not influenced by the followers and we are allowed
to pick uN+ j, j = 1, . . . , nl arbitrarily. For simplicity, we use vector x to stand for the followers'
states and z to stand for the leaders' states.
Then the whole multi-agent system equipped with m communication topologies can be written
in a compact form
or, equivalently,
x
z
Ai Bi
0
0
x
z
+
0
u
,
x = Aix + Biz,
z = u,
(16)
(17)
=
where i ∈ {1, . . . , m}. Ai ∈ RN×N and Bi ∈ RN×nl are both sub-matrices of the corresponding
graph Laplacian matrix −L.
August 21, 2018
DRAFT
The dynamics of the followers can be extracted as
x = Aix + Biz,
i ∈ {1, . . . , m}.
14
(18)
Remark 4: Similar with the single leader case, the structural controllability of system (16)
coincides with the structural controllability of system (18). And we say that the multi-agent
system (15) with switching topology and multi-leader is structurally controllable if and only if
the switched linear system (18) is structurally controllable with z as the control inputs.
Before proceeding further, we first discuss the structural controllability of multi-agent systems
with multi-leader under fixed topology with the following dynamics:
x = Ax + Bz,
(19)
where A ∈ RN×N and B ∈ RN×nl are both sub-matrices of the graph Laplacian matrix −L.
In [Ji et al., 2008][Ji et al., 2009], a new graph topology: leader-follower connected topology
was proposed:
Definition 7: ([Ji et al., 2008]) A follower subgraph G f of the interconnection graph G is the
subgraph induced by the follower set V f (here is x). Similarly, a leader subgraph Gl is the
subgraph induced by the leader set Vl (here is z).
Denote by Gc1 , . . . , Gcγ , the connected components in the follower G f . The definition of leader-
follower connected topology is as follows:
Definition 8: ([Ji et al., 2009]) The interconnection graph G of multi-agent system (19) is
said to be leader-follower connected if for each connected component Gci of G f , there exists
a leader in the leader subgraph Gl, so that there is an edge between this leader and a node in
Gci , i = 1, . . . , γ.
Based on this new graph topology, we can derive the criterion for structural controllability for
multi-agent system (19) under fixed topology:
Theorem 2: The multi-agent system (19) with multi-leader and fixed topology under the
communication topology G is structurally controllable if and only if graph G is leader-follower
connected.
Proof:
Necessity: The idea of necessity proof is similar to the proof of lemma 3 in
[Ji et al., 2008]. We assume that there exists one connected component Gcp not connected to
August 21, 2018
DRAFT
the leader subgraph Gl. Define Ai and Bi matrices as sub-matrices of A and B, the same as
the Fi and Ri matrices in lemma 3 of [Ji et al., 2008]. Following the analysis in lemma 3 of
[Ji et al., 2008], it can be easily got that the controllability matrix of multi-agent system (19)
15
is:
C =
B1 A1B1 A2
1B1
...
...
0
0
...
...
Bγ AγBγ A2
γBγ
...
0
...
· · · AN−1
1 B1
...
0
...
· · · AN−1
γ Bγ
· · ·
· · ·
· · ·
.
(20)
Consequently, rank C = row rank C < N. The maximum rank of C is less than N, which implies
that the corresponding multi-agent system (19) is not structurally controllable.
which can be partitioned into two subgraphs: induced leader subgraph G
Sufficiency:We use the proof of theorem 1 in [Ji et al., 2009] to help us prove the suffi-
ciency. The communication graph G consists of several connected components G(i), i = 1, . . . , κ,
(i)
l and induced follower
(i)
f . For each connected components G(i), i = 1, . . . , κ, it can be modeled as a linear
system with its system matrices being sub-matrices of A and B matrices. Following the analysis
subgraph G
in theorem 1 in [Ji et al., 2009], we can get the following equation:
rank C = rank C1 + rank C2 + . . . + rank Cκ,
(21)
where C is the controllability matrix of multi-agent system (19) and Ci is the controllability matrix
of connected component G(i). The independence of these connected components guarantees
the independence of free parameters in the corresponding matrices, which correspond to the
communication weights of the linkages. Consequently, we have that
g-rank C = g-rank C1+ g-rank C2+. . .+ g-rank Cκ
where g-rank of a structured matrix M is defined to be the maximal rank that M achieves as a
function of its free parameters. Besides, if in some connected component G(i), there is more than
one leaders, we can split it into several connected components with single leader or choose one
as leader and set all weights of the communication linkages between the followers and other
leaders to be zero. After doing this, connected component G(i) is a connected topology with
single leader. According to lemma 2, Ci has full g-rank, which equals to the number of follower
August 21, 2018
DRAFT
16
agents in G(i). Moreover, there is no common follower agent among the connected components.
Consequently, g-rank C=N and multi-agent system (19) is structurally controllable.
With the above definitions and theorems, we are in the position to present the graphical
interpretation of structural controllability of multi-agent systems under switching topology with
multi-leader:
Theorem 3: The multi-agent system (16) or (18) with the communication topologies Gi, i ∈
{1, . . . , m} and multi-leader is structurally controllable if and only if the union graph G is leader-
follower connected.
Proof: As stated in remark 1, the union graph G is the representation of the linear system:
(A1 + A2 + A3 + . . . + Am, B1 + B2 + B3 + . . . + Bm). Therefore, the condition that the union graph
G is leader-follower connected is equivalent to the condition that linear system (A1 + A2 + A3 +
. . . + Am, B1 + B2 + B3 + . . . + Bm) is structurally controllable. Following the proof steps in theorem
1, this result can get proved.
V. Numerical Examples
Next we will give two examples to illustrate the results in this paper and for simplicity, we
take single leader case as examples.
1
r
3
r
r
2
1
r
r
2
1
r
(a)
r
0
3
r
(b)
r
0
3
r
(c)
Fig. 1. Switched network with two subsystems
r
2
r
0
We consider here a four-agent network with agent 0 as the leader and with switching topology
described by the graphs in Fig. 1(a)-(b) (the self-loops are not depicted, because it will not
influence the connectivity). Overlaying the subgraphs together can get the union graph G of
this example as shown in Fig. 1(c). It turns out that the union graph of the switched system is
connected. By theorem 1, it is clear that the multi-agent system is structurally controllable.
Next, the rank condition of this multi-agent system will be checked.
From Fig. 1, calculating the Laplacian matrix for each subgraph topology, it can be obtained
that the system matrices of each subsystem are (one thing we should mention here with the
August 21, 2018
DRAFT
control strategy that each agent can use its own state information, the diagonal elements always
have free parameters, so we can get the following form of sub-matrix of Laplacian matrix) :
17
A1 =
, B1 =
; A2 =
, B2 =
λ1
0 λ2
0 λ4
0
λ5
0 λ3
0
λ6
0
λ7 λ10
λ11 λ8
0
0
0
0
λ9
0
0
λ12
.
According to lemma 1, the controllability matrix for this switched linear system is:
[B1, B2, A1B1, A2B1, A1B2, A2B2, A2
2B1, A2
A2
2B2].
After simple calculation, we can find three column vectors in the controllability matrix:
1B2, A2A1B2, A1A2B2, A2
1B1, A2A1B1, A1A2B1,
0
λ6
0
,
0
0
λ12
,
λ4λ12
0
λ3λ12
.
Imposing all the parameters scalar 1, it follows that these three column vectors are linearly
independent and this controllability matrix has full row rank. Therefore, the multi-agent system
is structurally controllable.
In the second example, we still consider a four-agent network with agent 0 as the leader
and with switching topology described by the graphs in Fig. 2(a)-(b). Overlaying the subgraphs
together can get the union graph G of this example shown in Fig. 2(c). It turns out that the
union graph of the switched system is disconnected, because agent 2 is isolated. According to
theorem 1, it is clear that the multi-agent system is not structurally controllable.
1
r
❅
❅
❅
❅
r
2
1
r
r
2
1
r
2
r
❅
❅
❅
❅
❅
r
0
3
r
(a)
❅
r
0
3
r
(b)
r
0
3
r
(c)
Fig. 2. Another switched network with two subsystems
Similarly, the rank condition of this switched linear system needs to be checked to see whether
it is structurally controllable or not.
From Fig. 2, calculating the Laplacian matrix for each graphic topology, it is clear that the
August 21, 2018
DRAFT
18
.
0
0
λ12
system matrices of each subsystem are :
A1 =
, B1 =
; A2 =
, B2 =
λ1
0 λ2
0 λ4
0
λ5
0 λ3
λ6
0
0
λ7
0
λ11
0 λ10
0
λ8
0
λ9
Computing the controllability matrix of this example yields the controllability matrix:
λ6
0
0
0
λ1λ6
. . . λ7λ10λ12 + λ9λ10λ12
0
. . .
. . .
0
λ10λ11λ12 + λ2
9λ12
.
0 λ12 λ5λ6
This matrix has the second row always to be zero for all the parameter values, which makes the
maximum rank of this matrix less than 3. Therefore, this multi-agent system is not structurally
controllable.
VI. Conclusions and Future Work
In this paper, the structural controllability problem of the multi-agent systems interconnected
via a switching weighted topology has been considered. Based on known results in the lit-
erature of switched systems and graph theory, graphic necessary and sufficient conditions for
the structural controllability of multi-agent systems under switching communication topologies
were derived. It was shown that the multi-agent system is structurally controllable if and only if
the union graph G is connected (single leader) or leader-follower connected (multi-leader). The
graphic characterizations show a clear relationship between the controllability and interconnec-
tion topologies and give us a foundation to design the optimal control effect for the switched
multi-agent system.
Some interesting remarks can be made on this result. First, it gives us a clear understanding
on what are the necessary information exchanges among agents to make the group of agents
behavior in a desirable way. Second, it provides us a guideline to design communication protocols
among dynamical agents. It is required that the resulted communication topology among agents
should somehow remain connected as time goes on, which is quite intuitive and reasonable.
Third, it is possible to reduce communication load by disable certain linkages or make them on
and off as long as the union graph is connected. Several interesting research questions arise from
this scenario. For example, what is the optimal switching sequence of topologies in the sense
August 21, 2018
DRAFT
19
of minimum communication cost? How to co-design the switching topology path and control
signals to achieve desirable configuration in an optimal way? We will investigate these questions
in our future research.
References
[Lawton et al., 2002] J. R. Lawson and R. W. Beard (2002). Synchronized multiple spacecraft rotations. Automatica, vol. 38,
no. 8, pp. 1359 -- 1364, Aug. 2002.
[Dunbar et al., 2006] W. B. Dunbar and R. M. Murray (2006). Distributed Receding Horizon Control for Multi-Vehicle
Formation Stabilization. Automatica, vol. 42, no. 4, pp. 549 -- 558, Apr. 2006.
[Bliman et al., 2008] P. -A. Bliman and G. Ferrari-Trecate (2008). Average consensus problems in networks of agents with
delayed communications. Automatica, vol. 44, no. 8, pp. 1985 -- 1995, Aug. 2008.
[Tomlin et al., 1998] C. Tomlin, G. J. Pappas, and S. Sastry (1998). Conflict Resolution for Air Traffic Management: A Study
in Multiagent Hybrid Systems. IEEE Trans. Autom. Control, vol. 43, no. 4, pp. 509 -- 521, Apr. 2008.
[How et al., 2004] J. How, E. King, Y. Kuwata (2004). Flight Demonstrations of Cooperative Control for UAV Teams. in Proc.
of AIAA 3rd "Unmanned Unlimited" Technical Conference, Workshop and Exhibit, 2004, AIAA-2004-6490.
[Tanner, 2004] H. Tanner (2004). On the controllability of nearest neighbor interconnections. in Proc. of 43rd IEEE Conference
on Decision and Control, 2004, pp. 2467-2472.
[Ji et al., 2006] M. Ji, A. Muhammad, and M. Egerstedt (2006). Leader-based multi-agent coordination: controllability and
optimal control. in Proc. of the American Control Conference, 2006, pp. 1358-1363.
[Ji et al., 2007] M. Ji and M. Egerstedt (2007). A graph-theoretic charaterization of controllability for multi-agent systems. in
Proc. of the American Control Conference, 2007, pp. 4588-4593.
[Rahmani et al., 2006] A. Rahmani and M. Mesbahi (2006). On the controlled agreement problem. in Proc. of the American
Control Conference, 2006, pp. 1376-1381.
[Ji et al., 2008] Z. Ji, H. Lin, and T. H. Lee (2008). A graph theory based characterization of controllability for nearest neighbour
interconnections with fixed topology. in Proc. of the 47rd IEEE Conference on Decision and Control, 2008, pp. 5262-5267.
[Moreau, 2005] L. Moreau (2005). Stability of multiagent systems with time-dependent communication links. IEEE Trans.
Autom. Control, vol. 50, no. 2, pp. 169 -- 182, Feb. 2005.
[Ji et al., 2008] Z. Ji, H. Lin, and T. H. Lee (2008). Controllability of multi-agent systems with switching topology. in Proc.
of the 3rd IEEE Conference on CIS-RAM, 2008, pp. 421-426.
[Ji et al., 2009] Z. Ji, Z. D. Wang, H. Lin, and Z. Wang (2009). Interconnection topologies for multi-agent coordination under
leader-follower framework. Automatica, vol. 45, no. 12, pp. 2857 -- 2857, Dec. 2009.
[Liu et al., 2008] B. Liu, T. G. Chu, L. Wang, and G. M. Xie (2008). Controllability of a Leader -- Follower Dynamic Network
With Switching Topology. IEEE Trans. Autom. Control, vol. 53, no. 4, pp. 1009 -- 1013, May 2008.
[Zamani et al., 2009] M. Zamani and H. Lin (2009). Structural controllability of multi-agent systems. in Proc. of the American
Control Conference, 2009, pp. 5743 -- 5748.
[Lin et al., 2007] H. Lin and P. J. Antsakis (2007). Switching stabilizability for continuous-time uncertain switched linear
systems. IEEE Transaction on Automatic Control, vol. 52, no. 4, pp. 633-646, April 2007.
[Lin et al., 2009] H. Lin and P. J. Antsaklis (2009). Stability and stabilizability of switched linear systems: a survey of recent
results. IEEE Trans. Autom. Control, vol. 54, no. 2, pp. 308 -- 322, Feb. 2009.
August 21, 2018
DRAFT
20
[Sun et al., 2002] Z. Sun, S. S. Ge, and T. H. Lee (2002). Controllability and reachability criteria for switched linear systems.
Automatica, vol. 38, no. 5, pp. 775-786, May 2002.
[Sun et al., 2004] Z. Sun and S. S. Ge (2004). Switched Linear Systems -- Control and Design. Springer, New York, 2004.
[Lin, 1974] C. T. Lin (1974). Structural controllability. IEEE Transactions on Automatic Control, vol. 19, No. 3, pp. 201-208,
June 1974.
[Glover et al., 1976] K. Glover and L. M. Silverman (1976). Characterization of structural controllability. IEEE Trans. Autom.
Control, vol. 21, no. 4, pp. 534 -- 537, Aug. 1976.
[Mayeda, 1981] H. Mayeda (1981). On structural controllability theorem. IEEE Transactions on Automatic Control, vol. 26,
No. 3, pp.795 - 798, June 1981.
August 21, 2018
DRAFT
|
1811.10981 | 1 | 1811 | 2018-11-27T13:49:47 | Modelling Agents Endowed with Social Practices: Static Aspects | [
"cs.MA"
] | To understand societal phenomena through simulation, we need computational variants of socio-cognitive theories. Social Practice Theory has provided a unique understanding of social phenomena regarding the routinized, social and interconnected aspects of behaviour. This paper provides the Social Practice Agent (SoPrA) model that enables the use of Social Practice Theory (SPT) for agent-based simulations. We extract requirements from SPT, construct a computational model in the Unified Modelling Language, verify its implementation in Netlogo and Prot\'eg\'e and show how SoPrA maps on a use case of commuting. The next step is to model the dynamic aspect of SPT and validate SoPrA's ability to provide understanding in different scenario's. This paper provides the groundwork with a computational model that is a correct depiction of SPT, computational feasible and can be directly mapped to the habitual, social and interconnected aspects of a target scenario. | cs.MA | cs |
MODELLING AGENTS ENDOWED WITH SOCIAL PRACTICES:
STATIC ASPECTS
A PREPRINT
Rijk Mercuur1, Virginia Dignum1, Catholijn M. Jonker2,3
1Faculty of Technology, Policy and Management, Delft University of Technology
2Faculty of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology
3Leiden Institute of Advanced Computer Science, Leiden University
November 28, 2018
ABSTRACT
To understand societal phenomena through simulation, we need computational variants of socio-
cognitive theories. Social Practice Theory has provided a unique understanding of social phenomena
regarding the routinized, social and interconnected aspects of behaviour. This paper provides the
Social Practice Agent (SoPrA) model that enables the use of Social Practice Theory (SPT) for
agent-based simulations. We extract requirements from SPT, construct a computational model in the
Unified Modelling Language, verify its implementation in Netlogo and Prot´eg´e and show how SoPrA
maps on a use case of commuting. The next step is to model the dynamic aspect of SPT and validate
SoPrA's ability to provide understanding in different scenario's. This paper provides the groundwork
with a computational model that is a correct depiction of SPT, computational feasible and can be
directly mapped to the habitual, social and interconnected aspects of a target scenario.
Keywords Social Practice Theory · Social Agents · Habits · Routines · Social Intelligence · Interconnected Actions ·
Activity Hierarchies
1
Introduction
To understand societal phenomena through simulation, we need computational variants of socio-cognitive theories.
This paper provides the Social Practice Agent (SoPrA) model that enables the use of Social Practice Theory (SPT) for
agent-based simulations. SPT studies our daily 'doings and sayings', such as commuting, dining or greeting. SPT
focusses on attributes of activities, namely related values, activities, agents, resources or locations. SPT studies the
similarity of these attributes over time (addressing habits), between people (addressing social intelligence) and between
different actions (addressing interconnected behaviour). SoPrA uses this perspective to increase our understanding of
social phenomena.
SPT has provided a unique understanding of social phenomena in sociology. Giddens (1984) showed how SPT captures
that our actions are part of a routine and often not conscious, voluntary or intentional. Reckwitz (2002) explained
how SPT captures that the social world is formed and shaped by shared practices and not an additional layer of joint
intentions or conventions. Shove et al. (2012) showed how our life is full of bundles of social practices: interconnected
actions that are similar and influence each other and that cannot be seen in isolation. By using these aspects of SPT in
social simulation we can provide potential explanations that increase our understanding of social phenomena (Grabner
2018). Although there are several useful conceptual frameworks and implementations (Holtz 2014; Narasimhan et al.
2017; Dignum & Dignum 2015), there is no computational model of SPT that can explain social phenomena in terms of
the habitual, social and interconnected aspects of a SP.
To show that SPT can provide potential explanations through simulation we need a computational model that is validated.
This paper focuses on the pre-requisite steps to validation: constructing a computational model that is a correct depiction
of SPT, computational feasible and clearly maps the model onto a target scenario (Grabner 2018). We approach this by
A PREPRINT - NOVEMBER 28, 2018
surveying the literature on SPT and summarizing the core aspects of SPT in requirements. We construct a computational
model in the Unified Modelling Language (UML) that fulfills these requirements and implement the model in both
Netlogo (Wilensky 1999) and Prot´eg´e (Musen 2015). By showing how our UML fulfills the requirements extracted
from SPT and, in turn, our implementations correctly depict our UML, we verify that we made a computational model
that is both feasible and a correct depiction of SPT. Modelling choices throughout this process are communicated and
aimed at the computational feasibility or the correct depiction of SPT. A use case on the societal problem of CO-2
emission due to car commuting shows how SoPrA clearly maps onto habitual, social and interconnected aspects of a
social problem. Future work is to model the dynamic aspect of SPT and validate SoPrA. The computational model
provided can be used to provide potential explanations of social situations or, possibly outside social simulation, to
guide human-agent interaction in a socially aptitude and efficient way.
The remainder of the paper is structured as follows. Section 2 explains the scenario that will be used as a use case
throughout the paper. Section 3 captures SPT in requirements and uses supporting literature from psychology and
agent theory to correctly model the core aspects of SPT. Section 4 and 5 present a high level description and UML
version of the SoPra Model. Section 7 shows how we can make a clear mapping from SoPrA to the habitual, social and
interconnected aspects of our scenario.
2 Use Case & Descriptive Statements
This section describes the use case of a modeller who aims to explain the societal problem of CO-2 emission due to
car commuting. The core mechanisms of SPT are essential in car commuting. First, social intelligence is used when
people coordinate their commute: carpool, bring their kids to school. Second, a meta-study found that habit is the
strongest predictors of car use; stronger than other frequently studied cognitive measures such as attitude, norm or
values (Hoffmann et al. 2017). Third, Cass & Faulconbridge (2016) found in a series of interviews that an important
consideration in taking the car is the interconnectedness of the work-commute, sport-commute and school-commute.
For example, some of the interviewed took the car to work, because this allowed them to combine that action with
bringing their kids to school. The following describes a scenario of how Bob can come to the decision to take the train
instead of the car.
Bob and Alice are at home. Bob usually takes the car to brings the kids to school and go to work.
He recently started to value the environment more, but does not really reflect on this habit when
at home with his kids. Furthermore, he equates commuting with car commuting. For him, they
are alike. Today the kids are not home. Alice asks Bob if he wants to go by train with her? She
beliefs Bob wants to commute, that biking is not very comfortable with the rain outside and that
people that go by car usually value comfort. She beliefs that normally when others brings the kids
to school by car, they will also go to work by car. However, now that the situation is different, he
will reconsider. Bob agrees to go by train. He does not have a habit in this situation and that activity
promotes environmentalism.
Section 7 shows how SoPrA can be used to capture this scenario.
3 Literature and Requirements
This section describes related literature and empirical work on SPT (Section 3.1) and on agent and psychology literature
that explores key properties of SP (Section 3.2). We end each section with a list of requirements. These requirements
summarize the section and extract what we find most important. Section 5 presents a UML model that meets these
requirements.
3.1 Social Practice Theory
3.1.1 Short History of Social Practice Theory
Social practice theory has its roots in the work of Wittgenstein and Heidegger who both ascribe a central role to praxis
or practices. This starts a tradition where the central unit with which one describes the world are our daily doings (e.g.,
breakfast, work, meetings, commuting). This is in contrast with the traditional view in sociology to see the world in
terms of agency (e.g., an employee) and social structures (e.g., an organization). For Giddens (1984) in particular,
practices are a way to bridge these two levels of abstraction as a practice exist on both levels (e.g., a meeting influences
the employee and the organization). Bourdieu (1977) and Schatzki (1996) popularized the concept of practices in the
1980s and 1990s, where Bourdiea focussed on the practice as a personal 'habit', Schatzki focussed on the practice as
2
A PREPRINT - NOVEMBER 28, 2018
something social. Reckwitz (2002) further emphasized the importance of SPs as a new way to capture the social world
that breaks with old paradigms: the social is formed and shaped by shared practices and not an additional layer of joint
intentions or conventions. Currently, the most common conceptualization of SP comes from Shove et al. (2012) who
sees the concept of practices as a collection of ever changing interconnected elements. For example, the practice of
driving is a connection of the material car, the meaning of transport and the competence to drive it. Recently, Dignum
& Dignum (2015) used SP as a parsimonious concept around which one can order relevant social knowledge for an
agent. They divert from the original line by bringing agency back on the table and linking SP to standard agent concepts
such as goals, norms and beliefs. SPT thus started as a focus on 'doing' and continued to conceptualize that 'doing' as
routines, that are shared between multiple people and relate different elements.
3.1.2 Key Properties of Social Practice
We focus on three key properties of SPs that reflect three aspects of human behaviour: habits, sociality and interconnected
behaviour. These key properties provide an aim and scope for our computational agent model of SP.
SPs are first of all practices. They are habitual enacted routines (Shove et al. 2012; Reckwitz 2002; Digmun 2004).
Such routines are repeated over time and exist by virtue of this reproduction (Shove et al. 2012). Reckwitz (2002)
emphasized that this routine is both physical and mental: there is repetitiveness in both the bodily movement and mental
elements that are associated with the SPs.
SPs are social. Shove et al. (2012) and Reckwitz (2002) interpret social here as that the SP is similar for multiple people.
For example, most people share the belief that one needs the know-how to control a car to do the SP of car driving.
Reckwitz (2002) sees SPs as a new way to capture the social world that breaks with old paradigms: the social is captured
in our shared SPs instead of in a shared mental world (e.g., as described by BDI-agents), in texts (e.g.,in scientific
papers) or interactions (e.g., in a discussion). Reckwitz (2002) this similarity means that the SP is something collective,
"but not in the sense of a mere sum of the content of single minds", but in a time-space transcending non-subjective way.
We assume that Reckwitz (2002) refers here to the SP as something macro. A separate entity that emerges out of micro
enactments, but is more than just the sum of them. Note that common knowledge dictates that SPs are often not fully
shared (i.e., similar for all people). For example, Bob and Alice might share many beliefs about the SP of commuting,
but Alice associate car-driving with efficiency and Bob with fun. As becomes clear in the next section, modelling this
tension between the individuality and sharedness of a SP is one of the main challenges this paper tackles.
There is another interpretation of sociality in SP: sociality as interaction. This interpretation becomes clear when
Dignum & Dignum (2015) zoom in on the micro aspects of SPs. For example, greeting, meeting, discussing and
teaching are all SPs that can only be enacted by multiple people. This meaning of sociality is closer to the layman
perspective on a social activity: an activity done in the presence of other people. SPs having a guiding role in this
interaction, allowing one to form expectations about the actions of others. This view contrasts with the view of Shove
et al. (2012) and Reckwitz (2002), who deem a SP social if its shared, even if there is no interaction when enacting it.
For example, one of Shove et al. (2012)'s canonical examples of a SP is showering, a SP mostly done alone. We claim
these two interpretations of sociality are reciprocally related. First, a SP only becomes shared through interaction. In
the case of greeting, the SP becomes shared through direct interaction. You see how others greet and adapt your ideas
of a SP. In the case of showering, the SP becomes shared through indirect interactions (judging remarks, sighs, sniffles).
Second, a SP can guide the interaction because it is shared. One forms expectations of the actions of the other based on
the assumption that the actions, meanings, habits (i.e., elements of a SP) are shared.
Different practices are similar and influence each other. Shove et al. (2012) called such similar practices: bundles of
practices that are interconnected. They argued that if practices are similar in some aspects, then they become similar
in other aspects too. We argue that SPs are interconnected in terms of common elements, time and space. First,
SPs connect when they are enacted at the same time or in sequence. For example, in a series of interviews, Cass &
Faulconbridge (2016) found that one main reason people take the car to work is because this makes it easier to combine
this SP with SPs of leisure, healthcare or shopping. Second, SPs connect when they are enacted in the same space. For
example, a supermarket at the train station facilitates the connection of the practice of groceries and commuting. Third,
SPs might be connected, because they share an element. For example, Shove et al. (2012) mentions how in the early
days of driving, cars easily broke down. To be able to drive a car one had to have the competence to repair it. The
practice of driving thus became connected with other practice that related to the competence of repairing, for instance,
plumbing or carpeting. The meaning of these practices as something masculine influenced the meaning of car driving.
Car driving and plumbing now share this element of meaning as something primarily masculine.
SPT.1 SPs are habitually enacted routines.
SPT.2 SPs are social:
(a) SPs are similar for multiple agents.
3
A PREPRINT - NOVEMBER 28, 2018
(b) SPs influence and are influenced by agent interactions.
SPT.3 SPs are interconnected: in terms of time, space and common elements.
3.1.3 The elements of a Social Practice
What comprises a SP? Although short names, such as "greeting" are given to SPs, SPs are collections of interconnected
elements. This idea started with Taylor (1973) for who "meanings and norms implicit in [...] practices are not just in the
minds of the actors but are out there in the practices themselves". For Taylor a practice is not merely a set of actions, but
also comprises other concepts. For (Reckwitz 2002) this means the SP's existence necessarily depends on the existence
and specific interconnectedness of these elements, but cannot be reduced to any single one of them. The practice of
skateboarding thus only exists because a group of elements comes together: materials like the skateboard, street spaces,
the competences to ride the board, rules and norms that define the practice and the meaning different groups attribute to
the practice (Shove et al. 2012). For example, as a recreational activity, an art form or a means of transport. We continue
by describing the literature on those elements we think are necessary and sufficient to model the key properties.
Although A SP is thus not merely (a set of) action(s), an action is a commonly agreed aspect of a SP. Actions here refer
to the bodily movement itself without all the mental connotations. We discern between two views in the literature. First,
according to Reckwitz (2002) a SP represents a pattern which can be filled out by a multitude of single and often unique
actions reproducing the SP. For example, the SP of commuting can consist of taking your kids to school and then going
to work. Second, for Shove et al. (2012) actions are not a part, but instead another view on the SP. She views the SP
as either a collection of elements or as something that is performed. For example, commuting can be seen as either a
collection of elements such as, the meaning of transport and the material car or as a series of actions, such as, getting
in your car and going to work. We follow Reckwitz (2002) as (1) it allows us to retain ourselves to the concept of
'elements' (instead of introducing an auxiliary concept 'perspective') in line with our aim for parsimony (2) modelling
SPs as a hierarchy of actions allows us to link SPT to agent models that focus on actions as well.
There are symbolic and social elements associated with a SP, such as, the symbolic element 'health' to the practice of
riding the bike. These symbolic and social associations are conceptualized differently by different authors. Schatzki
(1996) emphasized the 'teleoaffective' social elements, such as, embracing ends, purposes, projects and emotions. He
described how snapshots of what people do have a history and a future that are united in the moment of performance. In
our words, actions connect the now with past mental connotations and determine future connotations. Reckwitz (2002)
focused on mental, emotional and motivational knowledge. For Reckwitz it's important that the mental is part of the
routinized SP. We could not imagine football without the associated aim (to win the game), understanding (of others
players behaviour) and emotions (a particular tension). Shove et al. (2012) simplified these concepts in a bucket term
called meaning. This has the advantage that one can easily refer to all symbolic and social element of a SP, but the
disadvantage that the term is not well-defined or measurable. (In fact, Dignum & Dignum (2015) unfolds this concept
of meaning again and differs between social interpretation, roles or norms.) We follow Shove et al. (2012) here for
the sake of simplicity: it allows us to model the key properties, cost less time to work out and serves computationally
feasibility. Given the key properties we aim to model we highlight two aspects of the meaning element. First, the
element of meaning has a teleological role: it represents an end or purpose. Second, this end or purpose is social: its
shared by others and has a role in guiding interaction.
Whereas in the past SPT focussed on these classical social aspects (meanings, norms), more recently Schatzki (1996);
Latour (1996); Reckwitz (2002); Shove et al. (2012) have focussed on the physical aspects of a SP. They introduce
an element called 'materials', which helps in closing the 'gap' between the physical and social world (Shove et al.
2012). By abstracting over the actor, SPT shows that for multiple people the same SPs are connected to the same
type of physical objects. For example, almost everyone associates a car, road and gas station with the SP of driving.
These physical objects play a central role in our social world in that it provides a shared common ground (Searle 1995).
Dignum & Dignum (2015) connects the physical dimension of a SP to the decision making of an agent by emphasizing
that certain materials can trigger a SP. If the physical context triggers a SP, it means the agent will typically do the SP.
As we will show in the next section, triggering a SP is closely connected to what social psychologist call a habit.
In sum, this paper focuses on the following elements of a SP: an action element, a meaning element and a physical
element. We argue that this set of components is necessary and sufficient to model the aforementioned key properties of
a SP. There are more elements that can be routinized, socialized or interconnected. Other commonly described elements
of SP are emotions, competences, affordances, norms, goals and plans. Our view is that these are valuable concepts that
could improve SoPrA in its ability to model behaviour. However, to show how the mechanisms of routines, sociality
and interconnectedness in SP these elements are necessary and sufficient. After introducing SoPrA, we will show how
SoPrA can be clearly mapped on this aspects of our use case 7.
SPT.4 SPs consists of actions that represent the bodily movement.
4
A PREPRINT - NOVEMBER 28, 2018
SPT.5 SPs consist of a meaning element that represents a shared end or purpose.
SPT.6 SPs consist of a physical element that
(a) can trigger an action
(b) relates the physical world to the social.
3.2 Literature on Key Properties
3.2.1 Habits
A key aspect of SPs is that they are repeated over time and exist by virtue of this reproduction (Reckwitz 2002).
Repetitive behaviour is also studied from the individual perspective by social psychologist as 'habits'. They use the
term 'habit' to refer to a phenomenon whereby behaviour persists because it has become an automatic response to
particular, regularly encountered, context (Kurz & Gardner 2015). The model presented in this paper links the notion of
habitual behaviour to the notion of SP.
Habits are automatic. Moors & De Houwer (2006) compares several theoretical views and concludes that automaticity
entails unintentional, uncontrollable, goal independent, autonomous, purely stimulus driven, unconscious, efficient,
and fast behaviour. Habits are thus contrasted with predictors of intentional actions, such as values, motives and goals.
When a context triggers the habit these intentional concepts do not mediate the relation.1 For example, even if one
values the environment higher than efficiency one habitually takes the car to work. Habits can be measured with the
Self-Reported Habit Index, a 12-item index that includes items such as 'I do frequently' and 'I do without thinking'
(Verplanken & Orbell 2003). This measure has a medium-to-strong (r+ = 0.44) relation with behaviour (Gardner et al.
2011). Habits thus provide a fast automatic heuristic for action that derives its meaning from being contrasted to a slow
intentional mode of action.
Habits are context-dependent (Kurz & Gardner 2015). For example, a habit at home might not be a habit at work.
One might drink coffee every two hours at work, but not at home. The immediate setting, or context, of an agent thus
triggers an action. The literature on habits differs in what they consider as context. A narrow definition focusses on
physical tangible elements. For example, nearby cigarettes can trigger a habit of smoking. Wider definitions of context
allow time, other activities (Verplanken 2005), location and/or other people to trigger habits. Wood & Neal (2007)
nuance the context-action relation by conceptualizing habits as 'context cue'-action relations. For example, it is not so
much work that triggers drinking coffee, but a combination of cue's at work such as a colleague and the coffee machine.
The strength of these 'context cue'-action relations is continuous instead of discrete (e.g., a coffee machine is a slightly
stronger habitual trigger than a colleague) (Moors & De Houwer 2006).
The actor influences the habit. Lally et al. (2010) studied how habits are formed in the real world. Subjects were asked
to do the same action daily in the same context and report on automaticity (calculated using a sub-set of SRHI items).
She found that humans differ in the maximum automaticity they reported and also in the time to reach this maximum.
There are two aspects of habits that we consider out of scope of this paper. Different contexts can be similar and thus
trigger similar habits. Mercuur (2015) shows that although there is little known about how habits in one context relate
to habits in another context this is crucial to model habits in realistic settings. To our knowledge, there is no theoretical
or empirical work on this topic. This topic demands a study in itself and is postponed to future work. Habits are related
to the concept of attention (Wood & Neal 2007). The more attention an individual can muster the lower the chance he
or she falls into a habit. Attention becomes important when modelling habits over time. Different levels of attention
over time give rise to different decisions. As this paper focuses on static aspects this topic is postponed to future work.
H.1 Habits provide a fast automatic heuristic for action that derives its meaning from being contrasted to a slow
intentional mode of action.
H.2 Habits exist by virtue of continuous action-'context cue' associations.
H.3 Context cue's that trigger habits include physical elements and could extent to concepts such as time, other
activities, location or other people.
H.4 Humans differ in
(a) the height of the maximum automaticity (or strength of the habit) they experience
(b) the time to reach this maximum
5
A PREPRINT - NOVEMBER 28, 2018
3.2.2 Social Intelligence
The second key aspect of SPs is that they are social. The social world is also studied from the individual perspective as
social intelligence. Social intelligence is our ability to act wisely in in a world that we share with others (Thorndike
1920). The model presented in this paper links the notion of social intelligence to the notion of SP.
As mentioned, sociality in SPT has two faces: sociality as a shared world or sociality as interaction. Social intelligence
can also be seen from these two perspectives.
Castelfranchi (1997) describes social intelligence from the first perspective: as our ability to reason in a shared world.
The agent literature provides different answers on what this shared world consists of. A first series of papers uses
agents that only take into consideration the actions of other. For example, in the Consumat model (Jager et al. 2000)
an agent takes into consideration what most other agents do. Castelfranchi (1997) emphasized that we need to extend
such models to also take the mental state of other agents into consideration. The notion of social action cannot be a
behavioural notion - just based on an external description, because what makes the action social is that it is based on
certain mental states. A second series of papers focusses on such a representation of the mind of other agents based
on individual notions such as beliefs, desires and intentions (Dignum et al. 2000; Felli et al. 2014). Sociality is then
introduced as a secondary notion. For example, as the ability to form beliefs about others goals (Felli et al. 2014) or as
a mechanism to filter its intention to a socially desired set (Dignum et al. 2000). Dignum et al. (2014); Hofstede (2017)
argued that humans are at the core social beings and thus use social concepts as a primary concept. We view social
concepts here as referring to reasoning concepts that depend on being shared. For example, a notion of culture is hard to
imagine without multiple agents. A third series of paper focuses on these social concepts. For example, values, norms
(Dignum 1999), trust (Jonker & Treur 1999), culture and identity. What characterizes this work is to bring a collective
(social) notion back to the reasoning of the individual. This paper follows this last view in using social concepts as core
mechanisms for agents and the basis for our social world.
From the perspective of the individual, one beliefs in a shared social world, but also a personal view. For example,
one beliefs that car-driving is usually seen as a means for transport, but beliefs him or herself its a fun activity. These
personal and shared aspects of a view are independent: one can belief in a personal view on something without believing
this view is shared or one can belief something is the shared view without believing in this view. In other words, there
are multiple views on the world that one can all to a certain extent agree with or beliefs others agree with.
Theory of mind considers such reasoning about beliefs of others (Woodruff 1978). In addition, it treats two more
aspects: beliefs about specific others and chains of beliefs. For example, one can belief John beliefs car-driving is fun or
belief that John beliefs that I believe that John believes car driving is fun. These aspects are out of the scope of SoPrA.
SP are a heuristic that considers only two agents: itself or the group. For example, when greeting someone in most
cases it suffices to know that most people view greeting as polite and see shaking hands as a part of greeting. SP focuses
on the social intelligence that works in most situation, in contrast, theory of mind treats particular cases where more
in-depth reasoning is needed. For a model of SP, it's thus sufficient to require two aspects of our view on the shared
social world: to what extent one beliefs this view him or herself and to what extent one beliefs others hold this view.
Goleman (2011) describes social intelligence from the second perspective: as our ability to be aware and reason about
this common ground in interactions. He divides social intelligence in social awareness and social influence. Social
awareness allows one to form exceptions of other people and social influence to act upon those. As this paper focuses
on the static aspects of SPs we postpone the dynamics of interactions to future work.
SI.1 Social intelligence gives the ability to represent and reason about the shared world between agents.
SI.2 This shared world needs to be captured in terms of actions, individual concepts and foremost social concepts,
that is, concepts that depend on being shared among multiple humans.
SI.3 There are multiple views possible on this shared social world.
(a) Agents can belief to a certain extent themselves in a particular view.
(b) Agents can belief that others to a certain extent belief in a particular view.
(c) These two beliefs can be independent: one can belief something without believing others do and vice
versa.
SI.4 This common ground forms the basis of wise social interactions by allowing social awareness and social
influence.
3.2.3 Interconnected Actions
The third key aspect of SPs is that they are interconnected. This subsection uses the work on interconnected actions to
model this key aspect.
6
A PREPRINT - NOVEMBER 28, 2018
In neuroscience, Metzinger & Gallese (2003) brought together empirical evidence that shows that humans use actions
as a basic building block to represent the world. Actions here are the bodily movements conceptualized in the motor
system. Metzinger & Gallese (2003) shows how actions are used by the remainder of the brain as basic constituents of
the world it interprets. Humans form ontologies of how these actions are related to each other. Metzinger & Gallese
(2003) further shows how these action ontologies serve as a building block for social cognition: action-associations
abstract over the subject, therefore they can be used to map on a first-person or second-person perspective.
Action ontologies have been studied in the context of smart homes to be able to recognize activities (van Kasteren
et al. 2008; Storf et al. 2009; Chen et al. 2012; Okeyo et al. 2014). Okeyo et al. (2014) makes a difference between
actions, simple activities and composite activities. Actions are atomic. A simple activity is an ordered sequence of
actions. Composite activities are multiple actions within a time interval, but with no strict order to the actions and
can thus overlap. For example, driving might consist of overlaps of the activity of taking a turn and listening to the
navigation. From this point on we will separate between activities and actions. Activities refer to any bodily movement
(i.e., actions, simple activities and composite activities). Actions to refer to the subset of activities that is atomic.
Okeyo et al. (2014) separate two types of relations between activities: an ontological and temporal relation.2 The
ontological part describes relations between actions such as subsumptions, equivalence or disjointness. For example,
taking the train to school is a kind of commuting. A temporal relation encodes qualitative information regarding time.
For example, the user performs two activities after another. Humans are naturally able to make ontological and temporal
inferences about activities. For example, humans infer that if taking the car to work is environmentally unfriendly then
taking the car to school might be as well. Or they might induce one is doing the activity of commuting as one is doing
several actions in a sequence: walking to the car, going in the the car, driving to work and working.
In the area of agent communication, interconnected actions have been studied under the name of conversation protocols
(Kumar et al. 2002). A central aim in this work is to coordinate the agents by identifying important shared fixed points
(called landmarks), while allowing the agents freedom in their path to reach these states. The result of this is an activity
hierarchy with partial ordered landmarks that represents the conversation protocol.
IA.1 Actions are basic building blocks of how agents represent the world.
IA.2 Agents can relate these actions into meaningful activity hierarchies.
IA.3 Activity hierarchies serve as a building block for social cognition.
IA.4 Activities can consists of multiple other activities:
(a) sequential activities
(b) overlapping activities.
IA.5 Activity hierarchies have ontological and temporal aspects.
IA.6 Agents are able to make inference about these ontological and temporal properties.
IA.7 Activity hierarchies have shared fixed points, but also allow the freedom to have personal routes towards those
points.
3.3 Overview of Requirements
Table 1 provides an overview of the requirements extracted from the literature. The requirements are simplified for
display purposes.
Table 1: An overview of the requirements extracted from the literature.
Social Practices
are habitual
Habits
are fast and not slow
are social: similar and interactive
are action- context cue links
are interconnected
triggered by resources, time, etc.
consists of actions
consists of a meaning element
consists of a physical element
different per human
1
2
3
4
5
6
7
Social Intelligence
gives the ability to reason
in a shared world
requires shared actions,
individual and social mental concepts
requires an independent
personal and shared view
provides the basis for social
awareness and social influence
Interconnected Activities
are basic building blocks
form activity hierarchies
form the basis for social cognition
consist of other activities
can relate ontologically or temporally
allows agents to make inferences
are fixed points or personal deviations
7
A PREPRINT - NOVEMBER 28, 2018
4 The SoPrA Model: High-level Description
The SoPrA model aims to capture the habitual, social and interconnected aspects of SPT. We use the elements as
presented in the SPT literature as a basis: actions (or in our new terminology: activities), a meaning element and a
physical element. To adequately capture habits, social intelligence and interconnected actions, we use the presented
theory from other disciplines. This section presents a high level overview of SoPrA, the next section shows how SoPrA
meets the requirements presented in Table 1.
Figure 1 presents a high-level overview of SoPrA. We use activities as the central component of SoPrA, because they
are fundamental to each core aspect we aim to model: social practices consists of activities, habits are relations between
context and activities, activities are the basis for social cognition and the basis for the activity hierarchies humans use to
represent the world.
Figure 1: An abstract overview of SoPrA depicting how it captures the habitual, social and interconnected aspects of
social practices.
These activities are firstly used to capture the habitual aspects of SPs by relating the activities to context elements. An
activity will be habitually enacted when in the (performance) context of an agent there will be enough context elements
that habitually trigger the action. One takes the car to work, because one is at home, just had breakfast and sees the
keys on the table. As such, a SP can become a routine: in the same context the same activities, in the same order are
performed.
A routine only gains significance when it can be contrasted with a non-routine mode of acting. For this reason, an
activity is associated with a meaning element. The meaning element represent a shared end or purpose for the activity.
It provides a reason for the agent to deviate from the routine and intentionally chose an action.
These activities are secondly used to capture the interconnected aspect of SP by relating them to other activities. These
activities form a hierarchy that provides a structure that the agent follows when making decisions. To implement one
activity one needs to do another activity. After taking the car to work, one needs to take the car to sports. The structure
can also be used to make inferences: given that one activity implements the other they will share some properties.
These activities are thirdly used to capture the social aspects of SP by relating them to the beliefs of an agent. Agents
have a personal take on an activity and an idea of how others view the activity. These views often overlap. One thinks
car driving is efficient, one's partner thinks car driving is efficient and so, without saying, both assume they will carpool
to work with the car. Agents base their social intelligence on their shared view on activities. As such, a SP can be used
as a shared social world to guide interactions.
A practices truly becomes social when multiple agents have the same view on an activity. What constitutes a social
practice, instead of a personal practice, can thus only be determined in a bottom-up descriptive fashion. Every practice
exists on an continuum ranging from personal to shared. The extent to which a practice is shared changes over time:
daily social dynamics influence which views become prominent or dwindle. Moreover, there is no one cut-off point
8
A PREPRINT - NOVEMBER 28, 2018
where a practice is shared 'enough' to become social: every interaction requires a different threshold of sharedness.
For example, to go through a door requires less of a shared social world then playing soccer. This paper provides
groundwork by expressing for every agent in which view they belief. Future work will explore how the dynamics of
social interactions are dependent on the level of sharedness a social practice provides.
SoPrA thus associates elements (e.g., meaning, context elements) with these activities and not -- as is common
in SPT literature -- with the SP as a whole. For example, an agent does not associate the SP of commuting with
environmentalism. Instead, an agent associates an implementation of commuting -- taking the train to work -- with
environmentalism, while associating another implementation -- taking the car to work -- with the meaning of efficiency.
We give three advantages of this modelling choice. First, it allows to capture the beliefs an agent has about a SP in a
nuanced way (e.g., make a difference between forms of commuting). Second, it allows to use these nuances to guide
the decision-making of the agent (e.g., choose car-driving over the train because its more efficient). Third, it allows one
to feed the model with a selection of facts about the actions and let the agent make inferences about more complex
activities (e.g., commuting is boring, because all of its implementations are boring). This inheritance from activity to
activity will get more attention in Section 5.5.
5 The SoPrA Model: UML Model
This section provides a detailed overview of SoPrA in the Unified Modelling Language and several modelling choices.
The model is presented in parts and linked to the requirements. For the full model see 9.
5.1 Activity Tree
The basic building block of SoPrA are activities. They provide the core of a SP (SPT.4). These activities connect to each
other both temporally and ontologically (IA.5). SoPrA enables one ontological relation known as the generalization.
The generalization association allows the agent to make inferences about the properties of an activity based on its
parent (IA.6) as will be further discussed in Section 5.5. For pragmatic reasons we simplify the temporal relation to
the use of the parthood association. The parthood association specifies that one activity is composed of several other
activities. In other words, all the child activities need to be completed to complete the parent activity. The parthood
association allows us to model sequential activities. For example, car-commuting consists of bringing the kids to school
and going to work. Note that we do not specify the order in which these activities should be done nor if they can be
done simultaneously. In terms of Okeyo et al. (2014), we can model sequential activities (IA.4a), but not overlapping
activities (IA.4b). By enabling temporal and ontological associations between activities SoPrA provides an activity
hierarchy that agents use to order and represent the world (IA.1,IA.2).
Figure 2 presents the Activity class and the Implementation association class. The Implementation association
class can have the allOf type, which represents the aforementioned generalization relation, and the partOf type,
which represents the aforementioned temporal relation. To easily refer to the right level of abstraction of an activity we
differ between three types: actions (atomic), abstract actions and the top action (root). Figure 3 depicts an instance of
these classes for our use case of commuting. The activity hierarchy should be seen as the basis for the decision-making
of the agent. It allows a mechanism where the agent stepwise decides which activity it will do: from the activity of
commuting to the activity of taking the car to work.
Figure 2: An UML diagram depicting the Activity class, the Implementation association class connecting activities and
the different types one can attribute to these classes.
5.2 Activity Associations
Figure 4 present how we associate a physical element (SPT.6) and a meaning element (SPT.5) with activities to model
SPs as habitual routines (SPT.1) that can be contrasted with a slow intentional side (H.1).
9
A PREPRINT - NOVEMBER 28, 2018
Figure 3: An activity tree relating top actions, abstract actions and actions for commuting. The white arrows specify
kind-of associations and the black diamonds part-of associations.
Figure 4: An UML diagram depicting activity specific assocations such as the RelatedValue association class and the
HabitualTrigger association class.
SoPrA has a HabitualTrigger association class that associates activities with a ContextCue class. The strength
attribute of HabitualTrigger association class represents to what extent the context element is connected with the
activity. Habits are thus modelled as continious action-'context cue' associations (H.2). These context-cue's include
10
A PREPRINT - NOVEMBER 28, 2018
objects, locations, people, other activities and time (H.3) and are modelled in different ways. We model objects and
locations as type of context cue's (i.e., they do not have a seperate class). They represent the physical aspect of a SP
(SPT.6). Agents are modelled as a separate class that specializes (is generalized by) a ContextCue. We reserve a
separate class for agents as this class will have associations that other context cue's will not have (as opposed to object
and location). Activities are already part of the model and are also a specialization of Context Cue. We do not include
time as a possible habitual trigger of activities. Time is a special case as it semantics interplay with other parts of
our model (e.g., the parthood association). Due to time and space limitations we postpone the modelling of time as a
habitual trigger to modelling the dynamic aspects. The HabitualTrigger association class allows us to model SP as
habitually enacted routines (SPT.1).
SoPrA requires a meaning elements that represents a shared end or purpose (SPT.5). We aim to model the concept of
meaning as clearly distinguishable, measurable, and understandable in terms of empirical observations (i.e., operational-
ized) as this serves our aim of creating a validated model of SPT. The literature on SPT provides several candidates that
could be used to model meaning: emotions, motives, values, norms or goals.
We argue that the concept of human values (e.g., environmentalism, tradition, achievement) provides a good candidate
to model meaning as its (1) teleological, (2) shared and (3) operationalized and (4) trans-situational. First, values can
be defined as 'ideals worth pursuing' (Dechesne et al. 2012). Values are ends that people what to achieve: they are
unreachable, but can be pursued (Weide 2011). Values thus provide the teleological element we require of our model
of meaning. Second, Schwartz (2012) developed several instruments (e.g. surveys) to measure and categorize values.
Values thus provide an operational concept that has been made measurable and understandeble in terms of empirical
observations. Third, these findings on intervalue comparison have been extensively empirically tested and shown to be
consistent across 82 nations representing various age, cultural and religious groups (Schwartz 2012; Schwartz et al.
2012; Bilsky et al. 2011; Davidov et al. 2008; Fontaine et al. 2008). Values thus provide a shared concept to which
different people can relate. This satisfies our requirement to use shared social concepts as the basis of agent reasoning
(SI.2). Fourth, values are trans-situational (Weide 2011). This allows us to interconnect actions from different domains
(SPT.3). For example, environmentalism can relate both to the practice of commuting as well as dining.
Figure 4 shows how we model the connection between values, context and activities. In addition to an activity
and a value class, there is an association class called RelatedValue. The RelatedValue association class has a
strength attribute with type Double that represents how strongly an activity is attached to a value. Analogue the
HabitualTrigger association class has a strength attribute that represents how strongly a context element triggers an
activity.
5.3 Connecting Agents and Activities
To enable SPT for agent-based simulations, SoPrA needs to model the connection between SP and an agent (see Figure
5).
SoPrA connects agents and activities with the Belief association class. If multiple agents belief in the same activity
the SP is similar for them (SPT.2a). A SP can then form a common ground between agents that gives them the ability
to represent and reason about a shared social world (SI.1). The activity class here represents one view on an activity.
This allow for multiple views (SI.3), that is, multiple instantiations of the activity class for the same bodily movement.
For example, drive-car-to-school1 represents the view that car-driving is fun, while drive-car-to-school2
represents the view that car-driving is not fun.
The Belief association class has an attribute personalStrength that represents to what extent the agent itself agrees
with this view (SI.3a). The attribute sharedStrength represents to what extent an agent beliefs others agree with this
view (SI.3b). This allows the agent to have independent beliefs how others view an activity and how itself views this
activity (SI.3c). For example, I belief drive-car-to-school1 (i.e., car-driving is fun), but most authors belief that
drive-car-to-school2 (i.e., car-driving is not fun). Note that in a simpler model, one that connects these attributes
with the Activity class, these attributes are necessarily shared over multiple agents (e.g., violates SI.3c).
The Activity class has a reflexive same association that represents that two instances of the activity class (e.g.,
drive-car-to-school1 and drive-car-to-school2) refer to the same bodily-movement. This allows agents to
recognize which instances represent different views on the same bodily-movement. This will help agents in making
inferences on how different views on an activity connect. For example, Bob wants to enact drive-car-to-school1.
He beliefs that most others view the same bodily movement as drive-car-to-school2: driving the car to school
promotes the value of social interaction. Therefore Bob reasons that his neighbour will agree when he asks him to take
his kids to school as well. These inferences thus enable the agent to use the common ground SP provide to represent
and reason about the social world (SI.1).
11
A PREPRINT - NOVEMBER 28, 2018
Figure 5: An UML diagram presenting how agents and activities are connected with the belief association class. The
same association connects activities that represent the same bodily movement.
5.4 Agent Associations
Figure 6 depicts the Agent class with its attributes that are needed to deal correctly with habitual behaviour and contrast
this with intentional behaviour (SPT.1).
Figure 6: An UML diagram depicting agent specific elements such as the habitRate and the AdheredValue assocation
classs.
Agents have a attribute that ensures heterogeneity in how they acquire habits. This attribute habitRate represent
how much the strength of the HabitualTrigger association class increases when an agent experiences an action
in relation to a ContextElement. As shown in (Mercuur et al. 2017), a difference in habitRate suffices to acquire
heterogeneity in both the learning rate (H.4b) as well as the maximum habit strength the agents report (H.4a). Future
work will describe the dynamics of learning habits.
We model an association class between agents and values called AdheredValue with an attribute strength. This
represent to what extent an agent values a value. This enables agents to intentionally choose one action over another.
12
A PREPRINT - NOVEMBER 28, 2018
For example, an agent might value environmentalism highly and relate the activity Non-Car Commuting1 to environ-
mentalism, therefore, it chooses Non-Car Commuting1 over Car Commuting1. We require SoPrA to have a basis for
intentional actions as habits become only meaningful when contrasted with intentional behaviour (H.1). We chose
values to represent the intentional (or teleological) in SoPrA for the reasons we explained in section 5.2: values are
operationalized, shared and trans-situational.
5.5 Additional Formal Rules
SoPrA contains a number of logical rules that are necessary to correctly depict SPT, that is, fulfill the requirements
extracted from the literature (see Table 1). We view these logical rules in two ways. First, as constrains that limit the
number of possible instantiations of SoPrA to those that correctly depict SPT. Second, as reasoning rules for agents to
make inferences about the SP. We use first-order predicate logic to describe these rules as it is a well-known logic. In
Mercuur et al. (2018) we show that these rules can be translated to description logic such that we can use automatic
reasoners to enforce the aforementioned constrains and inferences.
The activity hierarchy is a directed tree. This implies the activity hierarchy has a unique root. This choice is made such
that the root can be used as a starting point in the decision-making process. A directed tree can be formally defined as a
connected graph where there is one more node then there are edges. Thus, the activity hierarchy H is a directed tree
H =< A, r, I >, where A is the set of instances of the Activity class, with root activity r ∈ A, and where I : A × A
is the set of instances of Implementation associations between activities, I + is its transitive closure, such that
(∀a ∈ A\{r} : (a, r) ∈ I +) ∧ (I = A − 1)
(1)
An activity can have three different types: action, abstract action and a top action. An action is an atomic leaf in the
activity tree, an abstract action is in the middle and a top action is the root node of the tree. These types are defined in
SoPrA to allow the agent to have a starting point (top action) and an ending point in the decision-making (action). We
can formally define these types as follows:
∀x ∈ A(¬∃y ∈ A I(y, x) ↔ Action(x))
∀x ∈ A((r = x) ↔ TopAction(x))
∀x ∈ A(¬(Action(x) ∨ T opAction(x)) ↔ AbstractAction(x))
(2)
(3)
(4)
The implementation association can have the partOf type, which should represent that one activity is composed of
several other activities. In other words, if an activity implements part of another activity, then there should be another
activity that also implements part of this parent activity. We express this formally as,
∀c1, p ∈ A( ImplementsPartOf(c1, p) → (∃c2 ∈ A ImplementsPartOf(c2, p) ∧ c1 (cid:54)= c2))
(5)
The activity hierarchy implies a logical relation between the activity associations.3 For example, if all implementations
of non-car commuting (taking a train, bike or walk) are strongly associated with environmentalism, then non-car
commuting is also strongly associated with environmentalism. In general, the strength with which the values are
associated with a parent activity is the average of the strength with which the values are associated with the children. To
express this formally we need to define two functions. First the function C : A (cid:55)→ An where n ∈ N, that maps a parent
activity p to its set of children C(p)
C(p) = {c ∈ A I(c, p)}.
(6)
We define the RelatedValue association class such that it is a function that maps the set of values V and activities A
to the strength attribute of the RelatedValue association. For all a ∈ A, v ∈ V let r(a, v) : A × V (cid:55)→ R denote
the strength of the relation between value v and activity a. This strength is determined as the average strength of the
relation between the same value and each of the children of the activity. If a is an action, then this strength is specified
in a seperately defined function u. Formally:
(cid:40)(cid:80)
u(a, v)
∀a ∈ A,∀v ∈ V : s(a, v) =
c∈C(a) s(c,v)
C(a)
if C(a) (cid:54)= ∅
otherwise
(7)
where u(a, v) : Action × V (cid:55)→ R is a separately defined function that denotes the strength of the relation between v and
the actions. This function depends on the domain and use case and can for example be based on empirical data.
13
A PREPRINT - NOVEMBER 28, 2018
Figure 7: An instantation of SoPrA in Netlogo for one agent that has one view on commuting. The red arrows represent
activities and the red lines implementation associations. The circle of white tree's and blue flags represent respectively
context elements and values. The remainder of the links represent beliefs, related value, habitual trigger, or adhered
value associations.
Note that the antecedent C(p) (cid:54)= ∅ ensures p has children it can inherit the values from. We do not enforce any logical
rules on the inheritance of the habitual trigger association. Humans can have a habit for a parent activity without having
a habit for its children and vice versa. Even allowing for two conflicting habits can lead to a theoretically sound model
(as shown in (Mercuur et al. 2017)).
Recall that SoPrA enables agents to make inferences about which activities refer to the same bodily movement with
the same association. The same relation represents an equivalence class. This means that the association is reflexive,
symmetric and transitive.
6
Implementation & Verification
6.1
Implementation
We implemented SoPrA in Netlogo (Wilensky 1999) and Prot´eg´e (Musen 2015). The code is available online.4
An implementation of SoPrA in Netlogo allows researchers to infer new knowledge over time. This paper provides
the static groundwork for these simulation. Netlogo has been chosen as our agent-based simulation platform for its
accessibility for researchers interested in modelling social phenomena. The UML classes are implemented as breeds in
Netlogo and associations are links. Links can have variables in Netlogo corresponding to the attributes of association
classes in UML. Netlogo does not have built-in mechanisms to represent an object hierarchy (such as presented in the
UML in Figure 8) or formal rules (such as represent in Section 5.5). However, these relations can be captured in manual
procedures. Figure 7 depicts a visualization of SoPrA in Netlogo. In future work such visualization can be used to
visualize how SPs and links between them evolve over time.
The implementation of SoPrA in Prot´eg´e allows researchers to author formal ontologies and enable formal reasoners
that aid the modeller. First, the formal reasoners can infer new static knowledge. For example, one can use the formal
reasoner to apply formal rule (7) and automatically infer that if riding a bike, taking the train to work and walking
promote environmentalism then non-car commuting also promotes environmentalism. Second, it can check if SoPrA is
satisfiable. For example, if the formal rules of Section 5.5 would be inconsistent Prot´eg´e would detect and report this.
Third, the formal reasoner can check if SoPrA is properly instantiated. For example, if one would assert that an activity
is partly implemented by only one child then Prot´eg´e would report this is inconsistent with rule (5).
Prot´eg´e makes this possible by its use of description logic: a logic that trades its expressivity to acquire decidability
(Horrocks 2007). The UML classes correspond to classes in Prot´eg´e and associations to object properties. In Mercuur
14
A PREPRINT - NOVEMBER 28, 2018
et al. (2018), we show how to capture association classes and the formal rules in Prot´eg´e. Consequently, this shows that
description logic is expressive enough to capture the semantics of SoPrA.
6.2 Verification
Verification is the act of testing whether the computational model does what it is supposed to be doing (Grabner 2018).
This amounts to testing if it adequately implements the underlying conceptual model. In our case, we want test if the
conceptual model in Section 5 is adequately implemented by (1) the formal model in Prot´eg´e and (2) the computational
model in Netlogo.
The two implementations are cross-checked. The conceptual model in Netlogo can be translated to Prot´eg´e models
using the method described in Polhill (2015). A Netlogo model should be satisfiable and properly instantiated in Prot´eg´e.
If the models hold up this means either both models are correct implementations or both models are wrong in exactly
the same way. A positive result thus lower the chance of one of the two implementations being wrong.
In Netlogo we run unit tests. We check if certain properties of the conceptual model that are not directly implemented
still hold. For example, the amount of belief associations should not exceed the amount of agents multiplied by the
amount of activities. A positive result lowers the chance that the implementation has properties that do not represent the
conceptual model.
Prot´eg´e can check if the model is satisfiable and correctly instantiated. Clearly we belief the conceptual model is
satisfiable and correctly instantiated. It thus checks if the implementation has properties we belief the conceptual model
has. A positive results implies a higher chance that the model is correctly implemented or that there is something wrong
with the conceptual model. Prot´eg´e automatically generates explanations for possible inconsistencies that can help to
detect which of the two is the case.
7 Applying SoPrA to our Use Case
This section presents how SoPrA can be used to capture the habitual, social and interconnected aspects of our use case
on CO2-emission due to car commuting. We specify the assignment of SoPrA, which part of the model should be
mapped on which part of the target (Weisberg 2012). The assignment of SoPrA is a necessary step to do validation
(Grabner 2018). Given that this paper focus on the static part, we map states of the model on the target (and not for
example, mechanism). It is always easier to validate aspects of a model that are a direct and explicit representation of
real-world objects (Schulze et al. 2017). This section demonstrates there is a direct and explicit mapping of SoPrA to
the habitual, social and interconnected aspects of the scenario in Section 2.
We capture the scenario in terms of SoPrA in Table 2. We go through the scenario step by step and explain how SoPrA
represents this step in Table 2.
Bob values the environment Line 28 shows that Bob adheres to the value of enviromentalism with a high strength.
Bob has a habit of taking the car to bring the kids to school and go to work Line 25-27 show that there is
HabitualTrigger relation between Kid1 and these activities. Note how SoPrA forces one to specify that it is
specifically the kid that triggers the habit. Furthermore, line 16, 18 and 20 show how Bob personally beliefs in
this view on these activities.
Alice asks Bob if he wants to go by train with her. This is a consequence of the following statements.
Alice beliefs Bob wants to commute Line 13 shows that Alice beliefs that Commuting1 is a view on
commuting that is shared. Line 24 shows that Commuting1 is an activity that is usually triggered at home.
Thus she infers that Bob (who is at home) would like to commute.
Alice beliefs that people that go by car usually value comfort Line 15 and 34 show that Alice beliefs that
Car Commuting1 is usually seen as an implementation of Commuting 1. Line 29 shows that Commuting1
promotes the value of comfort. Alice thus infers that Bob who usually does Car Commuting1 is looking
for an implementation of Commuting1 that promotes comfort.
Alice sees the rain outside and thinks Bob will find riding a bike uncomfortable, but the train comfort-
able Line 31 and 32 show that the views Alice beliefs to be shared on these activities state that the train
promotes comfort more than the bike.
Alice beliefs that normally when others bring the kids to school by car, they will also go to work by car
Line 15, 17, 19, 36 and 37 show that Alice beliefs that Bringing Kids to School1 is sharedly seen as
an implementation of Car Commuting1 and that this activity has another part: Going to Work with
the Car1.
15
A PREPRINT - NOVEMBER 28, 2018
.
Instance 2
Attribute 2
Table 2: A table showing a possible model that describes our use case. Only relevant instances are mentioned (e.g.,
possible RelatedValue and Habitualtrigger that do not play a role in the scenario are omitted.)
Number Class
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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
Instance 1
Commuting 1
Activity
Car Commuting 1
Activity
Bring Kids To School With Car 1
Activity
Go to Work with Car 1
Activity
Non-Car Commuting 1
Activity
Go To Work 1
Activity
Go To Work With Train 1
Activity
Go To Work With Bike 1
Activity
Bob
Agent
Alice
Agent
Agent
Kid1
ContextElement Home
Commuting 1
Alice
Belief
Commuting 1
Bob
Belief
Car Commuting 1
Belief
Alice
Car Commuting 1
Bob
Belief
Bring Kids To School With Car 1
Alice
Belief
Bring Kids To School With Car 1
Bob
Belief
Go to Work with Car 1
Alice
Belief
Go to Work with Car 1
Bob
Belief
Non-Car Commuting 1
Belief
Alice
Go to Work 1
Alice
Belief
Go to Work With Train 1
Belief
Alice
Home
HabitualTrigger Commuting 1
HabitualTrigger Car Commuting 1
Kid1
HabitualTrigger Bring Kids To School With Car 1 Kid1
HabitualTrigger Go to Work with Car 1
Kid1
Environment
AdheredValue
Comfort
RelatedValue
Comfort
RelatedValue
Comfort
RelatedValue
Comfort
RelatedValue
RelatedValue
Environment
Commuting 1
Implementation
Implementation Non-Car Commuting 1
Commuting 1
Implementation
Implementation Go to Work with Car 1
Implementation Go To Work 1
Implementation Go To Work With Train 1
Implementation Go To Work With Bike 1
Attribute 1
type: TopAction
type: AbstractAction
type: Action
type: Action
type: AbstractAction
type: AbstractAction
type: Action
type: Action
habitRate: 0.8
habitRate: 0.5
habitRate: 0.8
type: location
personalStrength: 0.1
personalStrength: 1.0
personalStrength: 0.1
personalStrength: 1.0
personalStrength: 0.1
personalStrength: 1.0
personalStrength: 0.1
personalStrength: 1.0
personalStrength: 1.0
personalStrength: 1.0
personalStrength: 1.0
strength: 1.0
strength: 1.0
strength: 1.0
strength: 1.0
strength: 0.8
strength: 0.9
strength: 0.9
strength: 0.7
strength: 0.0
strength: 1.0
type: allOf
type: allOf
type: partOf
type: partOf
type: allOf
type: allOf
type: allOf
Bring Kids To School With Car 1 Car Commuting 1
Car Commuting 1
Non-Car Commuting 1
Go to Work 1
Go to Work 1
sharedstrength: 0.8
sharedStrength: 0.8
sharedstrength: 0.8
sharedStrength: 0.8
sharedstrength: 0.8
sharedStrength: 0.8
sharedstrength: 0.8
sharedStrength: 0.8
sharedStrength: 0.8
sharedStrength: 0.8
sharedStrength: 0.8
Bob
Commuting 1
Car Commuting 1
Go To Work With Train 1
Go To Work With Bike 1
Go To Work With Train 1
Car Commuting 1
However, this situation is different Alice beliefs that the current situation could call for another activity
as there is no need for the activity Bring Kids To School 1. Line 22 shows that Alice beliefs (both
personally and wrt to others) in another implementation of Commuting 1. Namely, one with only one
part named Going To Work 1 (i.e., there is no bringing the kids to school). As Going To Work 1 is an
implementation of Non-Car Commuting1, she asks Bob to reconsider his travel mode.
Bob agrees to go by train, because it promotes environmentalism and he does not have a habit in this situation
There are no context-elements that could trigger Car-Commuting 1 over Non-Car Commuting 1 (or vice
versa). Line 33 shows that Go To Work with Train 1 promotes environmentalism.
8 Conclusion
This paper presented the SoPrA model that enables the use of Social Practice Theory for agent based simulations. SPT
provided new insights in sociology, but to show that SPT can provide potential explanations through simulation we
need a computational model that is validated. This paper provides the groundwork with a computational model that is a
correct depiction of SPT, computational feasible and can be directly mapped to the habitual, social and interconnected
aspects of a target scenario.
Section 3 extracted requirements from the literature that summarize the core aspects of SPT. We made the choice to
focus on habits, social intelligence and interconnected actions and therefore dropping elements such as competences
and affordances that were not directly instrumental to these core aspects. To correctly model these core aspects we used
supporting literature from social psychology, neuroscience and agent theory. Section 5 presented an UML model (added
with formal rules) that fulfills these requirements. We made the choice to model the temporal relations between action as
16
A PREPRINT - NOVEMBER 28, 2018
the parthood association such that we can model sequential activities, but not overlapping activities. Section 6 presented
publicly available implementations in Netlogo and Prot´eg´e that can be cross-checked to verify a proper implementation
of the UML. Section 7 showed how one can model a scenario regarding the social practice of commuting. The section
demonstrated how one maps the model on a system and how SoPrA can capture statements that combine habits,
social intelligence and interconnected element. For example, it can capture how Alice understands that Bob makes an
exception to join her in the train commute, because he doesn't have to bring the school. We thus showed how SoPrA is a
correct depiction of SPT through extracting and fulfilling requirements, the computational feasibility through presenting
a model in UML and making a verified implementation in Netlogo and Prot´eg´e and a clear mapping from target to
phenomena through a use case of commuting.
The next step is to extend the model with dynamic mechanisms that can generate output from input and validate
this extended model against empirical data. By validating the model one can show that SoPrA can provide potential
explanations and increase our understanding of social phenomena. Future work aims to explore possibilities to use
SoPrA outside agent-based simulations to guide human-agent interaction in socially aptitude and effecient way and to
create transparant and explainable AI.
17
9 Appendix A: A Complete Overview of SoPrA in UML
A PREPRINT - NOVEMBER 28, 2018
Figure 8: A Complete Overview of SoPrA in UML.
18
A PREPRINT - NOVEMBER 28, 2018
Notes
1There is debate about the exact relation between habitual and intentional actions. Gardner et al. (2011) shows in a
meta-analysis that in 8 out of 21 studies habits moderate the intention-action relation.
2Note that other authors use the term ontology to refer to any kind of relation between two objects. Temporal
relations are thus a subset of ontological relations. However, Okeyo et al. (2014) uses the term ontological to refer to
inferences one can easily make in description logic, whereas he uses the term temporal to refer to relations he can make
in Allan's temporal logic.
3The inheritance of activity associations implies that there are multiple versions of an activity based on their place
in the activity tree. For example, bringing kids to school as part of non-car commuting is different then bringing kids to
school as part of car commuting; the first one promotes environmentalism, the last one does not.
4For a Netlogo implementation for our trafic choice use case, see: https://github.com/rmercuur/SoPraTraffic-Netlogo.
For the Prot´eg´e implementation, see: https://github.com/PCSan/SOPRA.
References
Bilsky, W., Janik, M. & Schwartz, S. H. (2011). The Structural Organization of Human Values-Evidence from
Three Rounds of the European Social Survey (ESS). Journal of Cross-Cultural Psychology, 42(5), 759 -- 776. doi:
10.1177/0022022110362757
URL http://dx.doi.org/10.1177/0022022110362757
Bourdieu, P. (1977). Outline of a theory of practice, vol. 16. Cambridge University Press
URL papers://cfc50b6a-2d9e-4feb-87e5-d6012043bd5a/Paper/p2118
Cass, N. & Faulconbridge, J. (2016). Commuting practices: New insights into modal shift from theories of social
practice. Transport Policy, 45, 1 -- 14. doi:10.1016/j.tranpol.2015.08.002
URL http://dx.doi.org/10.1016/j.tranpol.2015.08.002
Castelfranchi, C. (1997). Modeling social action for AI agents. IJCAI International Joint Conference on Artificial
Intelligence, 2(c), 1567 -- 1576. doi:10.1016/S0004-3702(98)00056-3
URL http://dx.doi.org/10.1016/S0004-3702(98)00056-3
Chen, L., Nugent, C. D. & Wang, H. (2012). A Knowledge-Driven Approach to Activity Recognition in Smart Homes.
IEEE Transactions on Knowledge and Data Engineering, 24(6), 961 -- 974. doi:10.1109/TKDE.2011.51
URL http://dx.doi.org/10.1109/TKDE.2011.51
Davidov, E., Schmidt, P. & Schwartz, S. H. (2008). Bringing values back in: The adequacy of the European social
survey to measure values in 20 countries. Public Opinion Quarterly, 72(3), 420 -- 445. doi:10.1093/poq/nfn035
URL http://dx.doi.org/10.1093/poq/nfn035
Dechesne, F., Di Tosto, G., Dignum, V. & Dignum, F. (2012). No Smoking Here: Values, Norms and Culture in
Multi-Agent Systems. Artificial Intelligence and Law, 21(1), 79 -- 107. doi:10.1007/s10506-012-9128-5
URL http://dx.doi.org/10.1007/s10506-012-9128-5
Digmun, V. (2004). A model for organizational interaction: based on agents, founded in logic. SIKS
Dignum, F. (1999). Autonomous agents with norms. Artificial Intelligence and Law, 7(1), 69 -- 79. doi:10.1023/A:
1008315530323
URL http://dx.doi.org/10.1023/A:1008315530323
Dignum, F., Morley, D., Sonenberg, E. A. & Cavedon, L. (2000). Towards socially sophisticated BDI agents.
Proceedings - 4th International Conference on MultiAgent Systems, ICMAS 2000, (pp. 111 -- 118). doi:10.1109/
ICMAS.2000.858442
URL http://dx.doi.org/10.1109/ICMAS.2000.858442
Dignum, F., Prada, R. & Hofstede, G. (2014). From Autistic to Social Agents. In Proceedings of the 2014 international
conference on Autonomous Agents and Multi-Agent Systems, (pp. 1161 -- 1164)
URL http://dl.acm.org/citation.cfm?id=2617431
19
A PREPRINT - NOVEMBER 28, 2018
Dignum, V. & Dignum, F. (2015). Contextualized Planning Using Social Practices. In A. Ghose, N. Oren, P. Telang &
J. Thangarajah (Eds.), Coordination, Organizations, Institutions, and Norms in Agent Systems, (pp. 36 -- 52). Cham:
Springer International Publishing
Felli, P., Miller, T., Muise, C., Pearce, A. R. & Sonenberg, L. (2014). Artificial social reasoning: Computational
mechanisms for reasoning about others. Lecture Notes in Computer Science (including subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics), 8755, 146 -- 155. doi:10.1007/978-3-319-11973-1{\ }15
URL http://dx.doi.org/10.1007/978-3-319-11973-1{ }15
Fontaine, J. R. J., Poortinga, Y. H., Delbeke, L. & Schwartz, S. H. (2008). Structural Equivalence of the Values Domain
Across Cultures: Distinguishing Sampling Fluctuations From Meaningful Variation. Journal of Cross-Cultural
Psychology, 39(October), 345 -- 365. doi:10.1177/0022022108318112
URL http://dx.doi.org/10.1177/0022022108318112
Gardner, B., De Bruijn, G. J. & Lally, P. (2011). A systematic review and meta-analysis of applications of the self-
report habit index to nutrition and physical activity behaviours. Annals of Behavioral Medicine, 42(2), 174 -- 187.
doi:10.1007/s12160-011-9282-0
URL http://dx.doi.org/10.1007/s12160-011-9282-0
Giddens, A. (1984). The constitution of society: Outline of the theory of structuration. Cognitive Therapy and Research,
12(4), 448. doi:10.1007/BF01173303
URL http://dx.doi.org/10.1007/BF01173303
Goleman, D. (2011). Social Intelligence. Random House
Grabner, C. (2018). How to Relate Models to Reality? An Epistemo- logical Framework for the Validation and Veri-
fication of Computational Models. Journal of Artificial Societies and Social Simulation, 21(3)
URL http://jasss.soc.surrey.ac.uk///////.html
Hoffmann, C., Abraham, C., White, M. P., Ball, S. & Skippon, S. M. (2017). What cognitive mechanisms predict travel
mode choice? A systematic review with meta-analysis. Transport Reviews, 0(0), 1 -- 22. doi:10.1080/01441647.2017.
1285819
URL http://dx.doi.org/10.1080/01441647.2017.1285819
Hofstede, G. J. (2017). GRASP agents: social first, intelligent later. AI and Society, 0(0), 1 -- 9. doi:10.1007/
s00146-017-0783-7
URL http://dx.doi.org/10.1007/s00146-017-0783-7
Holtz, G. (2014). Generating Social Practices. Journal of Artificial Societies and Social Simulation, 17(1), 17
Horrocks, I. (2007). Description Logic : A Formal Foundation for Ontology Languages and Tools
Jager, W., Janssen, M. A., De Vries, H. J. M., De Greef, J. & Vlek, C. A. J. (2000). Behaviour in commons dilemmas:
Homo economicus and Homo psychologicus in an ecological-economic model. Ecological Economics, 35(3),
357 -- 379. doi:10.1016/S0921-8009(00)00220-2
URL http://dx.doi.org/10.1016/S0921-8009(00)00220-2
Jonker, C. M. & Treur, J. (1999). Formal Analysis of Models for the Dynamics of Trust based on Experiences. In
European Workshop on Modelling Autonomous Agents in a Multi-Agent World, vol. 1647, (pp. 221 -- 231). Berlin:
Springer. doi:10.1007/3-540-48437-X
URL http://dx.doi.org/10.1007/3-540-48437-X
Kumar, S., Huber, M. J., Cohen, P. R. & McGee, D. R. (2002). Toward a formalism for conversation protocols using
joint intention theory. Computational Intelligence, 18(2), 174 -- 228. doi:10.1111/1467-8640.00187
URL http://dx.doi.org/10.1111/1467-8640.00187
Kurz, T. & Gardner, B. (2015). Habitual behaviors or patterns of practice? Explaining and changing repetitive
climate-relevant actions. Wiley Interdisciplinary Reviews: Climate Change, 6(1), 113 -- 128. doi:10.1002/wcc.327
URL http://dx.doi.org/10.1002/wcc.327
Lally, P., Van Jaarsveld, C. H. M., Potts, H. W. W. & Wardle, J. (2010). How are habits formed: Modelling habit
formation in the real world. European Journal of Social Psychology, 40(6), 998 -- 1009. doi:10.1002/ejsp.674
URL http://dx.doi.org/10.1002/ejsp.674
20
A PREPRINT - NOVEMBER 28, 2018
Latour, B. (1996). On actor-network theory: A few clarifications. doi:10.2307/40878163
URL http://dx.doi.org/10.2307/40878163
Mercuur, R. (2015). Interventions on Contextualized Decision Making : an Agent-Based Simulation Study. Ph.D. thesis,
Utrecht University
URL https://dspace.library.uu.nl/handle/1874/323482
Mercuur, R., Dignum, V. & Jonker, C. (2017). Modeling Social Preferences with Values. In The Proceedings of the 3nd
International Workshop on Smart Simulation and Modelling for Complex Systems @ IJCAI, (pp. 17 -- 28)
Mercuur, R., Larsen, J. B. & Dignum, V. (2018). Modelling the Social Practices of an Emergency Room to Ensure Staff
and Patient Wellbeing. In 30th Benelux Conference on Artificial Intelligence, (p. (in press)). Den Bosch
Metzinger, T. & Gallese, V. (2003). The emergence of a shared action ontology: Building blocks for a theory.
Consciousness and Cognition, 12(4), 549 -- 571. doi:10.1016/S1053-8100(03)00072-2
URL http://dx.doi.org/10.1016/S1053-8100(03)00072-2
Moors, A. & De Houwer, J. (2006). Automaticity: a theoretical and conceptual analysis. Psychological bulletin, 132(2),
297 -- 326. doi:10.1037/0033-2909.132.2.297
URL http://dx.doi.org/10.1037/0033-2909.132.2.297
Musen, M. A. (2015). The prot´eg´e project: a look back and a look forward. {AI} Matters, 1(4), 4 -- 12. doi:
10.1145/2757001.2757003
URL http://dx.doi.org/10.1145/2757001.2757003
Narasimhan, K., Roberts, T., Xenitidou, M. & Gilbert, N. (2017). Using ABM to Clarify and Refine Social Practice
Theory. In W. Jager (Ed.), Advances in Social Simulation 2015, vol. 528. Springer International Publishing AG 2017.
doi:10.1007/978-3-319-47253-9
URL http://dx.doi.org/10.1007/978-3-319-47253-9
Okeyo, G., Chen, L. & Wang, H. (2014). Combining ontological and temporal formalisms for composite activity
modelling and recognition in smart homes. Future Generation Computer Systems, 39, 29 -- 43. doi:10.1016/j.future.
2014.02.014
URL http://dx.doi.org/10.1016/j.future.2014.02.014
Polhill, J. G. (2015). Extracting OWL Ontologies from Agent-Based Models: A Netlogo Extension. Journal of Artificial
Societies and Social Simulation, 18(2), 15. doi:10.18564/jasss.2810
URL http://dx.doi.org/10.18564/jasss.2810
Reckwitz, A. (2002). Toward a theory of social practices: A development in culturalist theorizing. European Journal of
Social Theory, 5(2), 243 -- 263. doi:10.1177/13684310222225432
URL http://dx.doi.org/10.1177/13684310222225432
Schatzki, T. R. (1996). Social Practices. A Wittgensteinian approach to human activity and the social. Cambridge
University Press
Schulze, J., Muller, B., Groeneveld, J. & Grimm, V. (2017). Agent-based modelling of social-ecological systems:
Achievements, challenges, and a way forward. Jasss, 20(2). doi:10.18564/jasss.3423
URL http://dx.doi.org/10.18564/jasss.3423
Schwartz, S. H. (2012). An Overview of the Schwartz Theory of Basic Values. Online Readings in Psychology and
Culture, 2, 1 -- 20. doi:http://dx.doi.org/http://dx.doi.org/10.9707/2307-0919.1116
URL http://dx.doi.org/http://dx.doi.org/http://dx.doi.org/10.9707/2307-0919.1116
Schwartz, S. H., Cieciuch, J., Vecchione, M., Davidov, E., Fischer, R., Beierlein, C., Ramos, A., Verkasalo, M.,
Lonnqvist, J.-E., Demirutku, K., Dirilen-Gumus, O. & Konty, M. (2012). Refining the theory of basic individual
values. Journal of Personality and Social Psychology, 103(4), 663 -- 688. doi:10.1037/a0029393
URL http://dx.doi.org/10.1037/a0029393
Searle, J. (1995). The Construction of Social Reality
Shove, E., Pantzar, M. & Watson, M. (2012). The Dynamics of Social Practice: Everyday Life and How it Changes.
London: SAGE Publications. doi:10.4135/9781446250655.n1
URL http://dx.doi.org/10.4135/9781446250655.n1
21
A PREPRINT - NOVEMBER 28, 2018
Storf, H., Becker, M. & Riedl, M. (2009). Rule-based activity recognition framework: Challenges, technique and
learning. Proceedings of the 3d International ICST Conference on Pervasive Computing Technologies for Healthcare.
doi:10.4108/ICST.PERVASIVEHEALTH2009.6108
URL http://dx.doi.org/10.4108/ICST.PERVASIVEHEALTH2009.6108
Taylor, C. (1973). Interpretation and the Sciences of Man. In Explorations in Phenomenology, vol. 25, (pp. 47 -- 101).
Philosophy Education Society Inc. doi:10.1007/978-94-010-1999-6{\ }3
URL http://dx.doi.org/10.1007/978-94-010-1999-6{ }3
Thorndike, E. (1920). Intelligence and its uses. Harper's Magazine, 140, 227 -- 235
van Kasteren, T., Noulas, A., Englebienne, G. & Krose, B. (2008). Accurate activity recognition in a home setting.
Proceedings of the 10th international conference on Ubiquitous computing - UbiComp '08, (p. 1). doi:10.1145/
1409635.1409637
URL http://dx.doi.org/10.1145/1409635.1409637
Verplanken, B. (2005). Habits and implementation intentions. In The ABC of behavioural change., (pp. 99 -- 109).
Oxford, UK: Elsevier
URL http://opus.bath.ac.uk/9438/
Verplanken, B. & Orbell, S. (2003). Reflections on Past Behavior: A Self-Report Index of Habit Strength. Journal of
Applied Social Psychology, 33(6), 1313 -- 1330. doi:10.1111/j.1559-1816.2003.tb01951.x
URL http://dx.doi.org/10.1111/j.1559-1816.2003.tb01951.x
Weide, T. v. d. (2011). Arguing to motivate decisions. SIKS Dissertation Series
URL http://www.narcis.nl/publication/Language/EN/id/170/RecordID/oai:dspace.library.uu.
nl:1874/210788
Weisberg, M. (2012). Simulation and similarity: Using models to understand the world. Oxford University Press
Wilensky, U. (1999). Netlogo itself
URL http://ccl.northwestern.edu/netlogo/
Wood, W. & Neal, D. T. (2007). A new look at habits and the habit-goal interface. Psychological review, 114(4),
843 -- 863. doi:10.1037/0033-295X.114.4.843
URL http://dx.doi.org/10.1037/0033-295X.114.4.843
Woodruff, G. (1978). Premack and Woodruff : Chimpanzee theory of mind. Behavioral and Brain Sciences, (1978),
515 -- 526
22
|
1210.4383 | 1 | 1210 | 2012-10-16T12:26:33 | Distributed Formation of Balanced and Bistochastic Weighted Diagraphs in Multi-Agent Systems | [
"cs.MA"
] | Consensus strategies find a variety of applications in distributed coordination and decision making in multi-agent systems. In particular, average consensus plays a key role in a number of applications and is closely associated with two classes of digraphs, weight-balanced (for continuous-time systems) and bistochastic (for discrete-time systems). A weighted digraph is called balanced if, for each node, the sum of the weights of the edges outgoing from that node is equal to the sum of the weights of the edges incoming to that node. In addition, a weight-balanced digraph is bistochastic if all weights are nonnegative and, for each node, the sum of weights of edges incoming to that node and the sum of the weights of edges out-going from that node is unity; this implies that the corresponding weight matrix is column and row stochastic (i.e., doubly stochastic). We propose two distributed algorithms: one solves the weight-balance problem and the other solves the bistochastic matrix formation problem for a distributed system whose components (nodes) can exchange information via interconnection links (edges) that form an arbitrary, possibly directed, strongly connected communication topology (digraph). Both distributed algorithms achieve their goals asymptotically and operate iteratively by having each node adapt the (nonnegative) weights on its outgoing edges based on the weights of its incoming links (i.e., based on purely local information). We also provide examples to illustrate the operation, performance, and potential advantages of the proposed algorithms. | cs.MA | cs | Distributed Formation of Balanced and Bistochastic
Weighted Diagraphs in Multi-Agent Systems
Themistoklis Charalambous and Christoforos N. Hadjicostis
2
1
0
2
t
c
O
6
1
]
A
M
.
s
c
[
1
v
3
8
3
4
.
0
1
2
1
:
v
i
X
r
a
Abstract
Consensus strategies find a variety of applications in distributed coordination and decision making
in multi-agent systems. In particular, average consensus plays a key role in a number of applications
and is closely associated with two classes of digraphs, weight-balanced (for continuous-time systems)
and bistochastic (for discrete-time systems). A weighted digraph is called balanced if, for each node
vj, the sum of the weights of the edges outgoing from that node is equal to the sum of the weights of
the edges incoming to that node. In addition, a weight-balanced digraph is bistochastic if all weights
are nonnegative and, for each node vj, the sum of weights of edges incoming to that node and the sum
of the weights of edges out-going from that node is unity; this implies that the corresponding weight
matrix is column and row stochastic (i.e., doubly stochastic). We propose two distributed algorithms:
one solves the weight-balance problem and the other solves the bistochastic matrix formation problem
for a distributed system whose components (nodes) can exchange information via interconnection links
(edges) that form an arbitrary, possibly directed, strongly connected communication topology (digraph).
Both distributed algorithms achieve their goals asymptotically and operate iteratively by having each
node adapt the (nonnegative) weights on its outgoing edges based on the weights of its incoming
links (i.e., based on purely local information). We also provide examples to illustrate the operation,
performance, and potential advantages of the proposed algorithms.
I. INTRODUCTION
A distributed system or network consists of a set of subsystems (nodes) that can share
information via connection links (edges), forming a directed communication topology (digraph).
Themistoklis Charalambous is with the School of Electrical Engineering, Royal Institute of Technology (KTH), Stockholm,
Sweden. E-mail: [email protected].
Christoforos N. Hadjicostis is with the Department of Electrical and Computer Engineering at the University of Cyprus,
Nicosia, Cyprus. E-mail: [email protected].
The successful operation of a distributed system depends on the structure of the digraph which
typically proves to be of vital importance for our ability to apply distributed strategies and
perform various tasks. Cooperative distributed control algorithms and protocols have received
tremendous attention during the last decade by several diverse research communities (e.g., biol-
ogy, physics, control, communication, and computer science), resulting in many recent advances
in consensus-based approaches (see, for example, [1] -- [8] and references therein).
In general, the objective of a consensus problem is to have the agents belonging to a group
agree upon a certain (a priori unknown) quantity of interest. When the agents reach agreement,
we say that the distributed system reaches consensus. Such tasks include network coordination
problems involving self-organization, formation patterns, parallel processing and distributed
optimization. One of the most well known consensus problems is the so-called average consensus
problem in which agents aim to reach the average of their initial values. The initial value
associated with each agent might be, for instance, a sensor measurement of some signal [9],
Bayesian belief of a decision to be taken [10], or the capacity of distributed energy resources for
the provisioning of ancillary services [11]. Average consensus is closely related to two classes of
graphs: weight-balanced (for continuous-time systems) and bistochastic graphs (for discrete-time
systems). A weighted graph is balanced if for each node, the sum of the weights of the edges
outgoing from that node is equal to the sum of weights of the edges incoming to that node. A
bistochastic graph is a weight-balanced graph for which the weights are nonnegative and the
sum of the weights for both outgoing and incoming edges (including self weights at each node)
is equal to one. In both cases, all edge weights are typically required to be nonzero.
It is shown in [2] that average consensus is achieved if the information exchange topology is
both strongly connected and balanced, while gossip algorithms [12], [13] and convex optimization
[14] admit update matrices which need to be doubly stochastic. These methods have applicability
to a variety of topics, such as multi-agent systems, cooperative control, and modeling the
behaviour of various phenomena in biology and physics, such as flocking. Since their operation
requires weight-balanced and bistochastic digraphs, it is important to be able to distributively
transform a weighted digraph to a weight-balanced or bistochastic one, provided that each node
is allowed to adjust the weight of its outgoing links accordingly.
In this paper, we address the problem of designing discrete-time coordination algorithms
that allow a networked system to distributively obtain a set of weights that make it weight-
balanced or bistochastic. This task is relatively straightforward when the underlying communi-
cation topology forms an undirected graph (i.e., when communication links are bi-directional)
but more challenging when dealing with a digraph (i.e., when some communication links might
be uni-directional). The paper proposes two algorithms that can be used in distributed networks
with directed interconnection topologies; the first algorithm leads to a weight-balanced digraph.
Even though there exist some earlier approaches in the literature for weight balancing (e.g., [15]
presents a finite-time algorithm but does not characterize the number of steps required in the worst
case, whereas [16] presents an asymptotic algorithm whose rate is bounded explicitly based on
the structure of the graph), this is the first time, to the best of our knowledge, that an asymptotic
algorithm of this nature is shown to admit geometric convergence rate. Work in the literature
appears for bistochastic matrix formation as well (e.g., [17] proposes an asymptotic algorithm
with an unspecified rate of convergence). Our second proposed algorithm, a modification of the
weight-balancing algorithm, leads to a bistochastic digraph with asymptotic convergence. Under
some minor assumptions, this second algorithm can also be shown to admit a geometric rate.
The remainder of the paper is organised as follows. In Section II, we provide necessary notation
and background on graph properties. In Section III, the problem to be solved is formulated,
and Sections IV and V present our main results in which we propose two algorithms, one for
weight-balancing and one for bistochastic matrix formation. Then, in Section VI, the derived
algorithms are demonstrated via illustrative examples and their performance is compared against
existing approaches in the literature. Finally, Section VII presents concluding remarks and future
directions.
II. NOTATION AND PRELIMINARIES
The sets of real, integer and natural numbers are denoted by R, Z and N, respectively; their
positive orthant is denoted by the subscript + (e.g. R+). The symbol N0 denotes the set of non-
negative integers. Vectors are denoted by small letters whereas matrices are denoted by capital
letters; A−1 denotes the inverse of matrix A. By I we denote the identity matrix (of appropriate
dimensions), whereas by 1 we denote a column vector (of appropriate dimension) whose elements
are all equal to one. A matrix whose elements are nonnegative, called nonnegative matrix, is
denoted by A ≥ 0 and a matrix whose elements are positive, called positive matrix, is denoted
by A > 0. σ(A) denotes the spectrum of matrix A, λ(A) denotes an element of the spectrum of
matrix A, and ρ(A) denotes its spectral radius. Notation diag(xi) is used to denote the matrix
with elements in the finite set {x1, x2, ..., xi, ...} on the leading diagonal and zeros elsewhere.
Let the exchange of information between nodes be modeled by a weighted digraph (directed
graph) G(V,E, W ) of order n (n ≥ 2), where V = {v1, v2, . . . , vn} is the set of nodes, E ⊆
V × V − {(vj, vj) vj ∈ V} is the set of edges, and W = [wji] ∈ Rn×n
is a weighted n × n
adjacency matrix where wji are nonnegative elements. A directed edge from node vi to node vj
is denoted by εji = (vj, vi) ∈ E, which represents a directed information exchange link from
node vi to node vj, i.e., it denotes that node vj can receive information from node vi. Note
that the definition of G excludes self-edges (though self-weights are added when we consider
bistochastic digraphs). A directed edge εji ∈ E if and only if wji > 0. The graph is undirected if
and only if εji ∈ E implies εij ∈ E. Note that a digraph G(V,E) can be thought of as a weighted
digraph G(V,E, W ) by defining the weighted adjacency matrix W with wji = 1 if εji ∈ E, and
wji = 0 otherwise.
+
A digraph is called strongly connected if for each pair of vertices vj, vi ∈ V, vj
(cid:54)= vi,
there exists a directed path from vj to vi, i.e., we can find a sequence of vertices vj ≡ vl0,
vl1, ..., vlt ≡ vi such that (vlτ +1, vlτ ) ∈ E for τ = 0, 1, ..., t − 1. All nodes that can transmit
information to node vj directly are said to be in-neighbors of node vj and belong to the set
j = {vi ∈ V εji ∈ E}. The cardinality of N −
N −
j , is called the in-degree of j and it is denoted
by D−
j . The nodes that receive information from node j comprise its out-neighbors and are
denoted by N +
j , is called the out-degree of vj and
it is denoted by D+
j . Given a weighted digraph G(V,E, W ) of order n, the total in-weight of
node vj is denoted as S−
wji, whereas the total out-weight of
wlj.
node vj is denoted by S+
j = {vl ∈ V εlj ∈ E}. The cardinality of N +
j and is defined by S−
j and is defined as S+
j =(cid:80)
j =(cid:80)
Definition 1: A weighted digraph G(V,E, W ) is called weight-balanced if the total in-weight
equals the total out-weight for every node vj ∈ V , i.e., S−
j . A weight-balanced digraph
is also called doubly stochastic (bistochastic) if each of its weighted adjacency matrix rows and
columns sums to 1.
j = S+
j
vi∈N −
vl∈N +
j
For the discrete-time setup we investigate, we conveniently define the time coordinate so that
j [k] will denote the value of the
unity is the time between consecutive iterations. For example S+
out-weight of node vj at time instant k, k ∈ N0.
III. PROBLEM FORMULATION
Given a digraph G(V,E), we want distributed algorithms that allow the nodes to obtain a
weight matrix W = [wji such that the following are achieved.
(i) The weighted digraph becomes balanced in a distributed fashion; i.e., a weight matrix
W is found such that wji > 0 for each edge (vj, vi) ∈ E, wji = 0 if (vj, vi) /∈ E, and
j = S−
S+
for every vj ∈ V .
j
(ii) The weighted digraph becomes bistochastic in a distributed fashion; i.e., a weight matrix
W is found with nonnegative diagonal elements wjj ≥ 0, such that wji > 0 if (vj, vi) ∈ E,
wji = 0 if (vj, vi) /∈ E, vj (cid:54)= vi), and wjj + S+
j = 1 for every vj ∈ V .
j = wjj + S−
IV. DISTRIBUTED ALGORITHM FOR WEIGHT-BALANCING
A. Description of the algorithm
Balancing a weighted digraph can be accomplished via a variety of algorithms. We introduce
and analyse a distributed cooperative algorithm that exhibits asymptotic convergence and outper-
forms existing algorithms suggested in the literature [18]. The algorithm achieves weight-balance
as long as the underlying digraph is strongly connected (or is a collection of strongly connected
digraphs, a necessary and sufficient condition for balancing to be possible [19]). The rate of
convergence of the algorithm is geometric and depends exclusively on the structure of the given
diagraph and some constant parameters chosen by the nodes.
Algorithm 1 is an iterative algorithm in which the nodes distributively adjust the weights of
their outgoing links such that the digraph becomes asymptotically weight-balanced. We assume
that each node observes but cannot set the weights of its incoming links. Given a strongly
connected digraph G(V,E), the distributed algorithm has each node initialize the weights of all
of its outgoing links to unity, i.e., wlj[0] = 1, ∀vl ∈ N +
j . Then, it enters an iterative stage where
each node performs the following steps:
1) It computes its weight imbalance defined by
xj[k] (cid:44) S−
j [k] − S+
j [k].
(1)
2) If it is positive (resp. negative), all the weights of its outgoing links are increased (resp.
decreased) by an equal amount and proportionally to xj[k].
We discuss why the above distributed algorithm asymptotically leads to weights that balance the
graph (and also characterize its rate of convergence) after we describe the algorithm in more
detail. For simplicity, we assume that during the execution of the distributed algorithm, the nodes
update the weights on their outgoing links in a synchronous1 fashion. Also note that 2) above
implies that wlj for vl ∈ N +
j will always have the same value.
Algorithm 1 Weight balancing algorithm
Input: A strongly connected digraph G(V,E) with n = V nodes and m = E edges (and no
self-loops).
Initialization: Each node vj ∈ V
1) Sets wlj[0] = 1, ∀vl ∈ N +
j .
2) Chooses βj ∈ (0, 1).
Iteration: For k = 0, 1, 2, . . ., each node vj ∈ V updates the weights of each of its outgoing
links wlj, ∀vl ∈ N +
j , as
(cid:32)
wlj[k + 1] = wlj[k] + βj
.
(2)
(cid:33)
S−
j [k]
D+
j
− wlj[k]
The intuition behind the proposed algorithm is that we compare S−
j [k] > S−
j wij[k].
j [k]), then the algorithm reduces (resp. increases) the weights
j [k] (resp. S+
j [k] < S−
j [k] with S+
j [k] = D+
If S+
on the outgoing links.
Proposition 1: If a digraph is strongly connected or is a collection of strongly connected
digraphs, Algorithm 1 reaches a steady state weight matrix W ∗ that forms a weight-balanced
digraph, with geometric convergence rate equal to
where
R∞(P ) = − ln δ(P ),
δ(P ) (cid:44) max{λ : λ ∈ σ(P )), λ (cid:54)= 1}.
Proof: First, we observe from equation (2) that all the outgoing links have the same weight,
(because they are equal at initialization and they are updated in the
i.e., wl(cid:48)j = wlj, ∀vl(cid:48), vl ∈ N +
j
1Even though we do not discuss this issue in the paper, asynchronous operation is not a problem.
same fashion). Hence, from hereafter, we will denote the weight on any outgoing link of node
vj as wj. In order to study the system with update formula (2) for each node in the graph, we
define w = (w1 w2
j ), and thus we can write the evolution
of the weights in matrix form as follows.
. . . wn)T with wj = wlj (vl ∈ N +
where w0 = 1 and
Pji (cid:44)
if i = j,
if vi ∈ N −
j .
w[k + 1] = P w[k], w[0] = w0
1 − βj,
βj/D+
j ,
(3)
(4)
It should be clear from the above update equation that the weights remain nonnegative during
the execution of the algorithm.
Matrix P can be written as P = I−B +BD−1A, where I is the identity matrix, B = diag(βj),
j ) and A is the adjacency matrix with aji = 1 if εji ∈ E, and aji = 0 otherwise.
D = diag(D+
Since σ(D−1A) = σ(AD−1), then ρ(D−1A) = ρ(AD−1). In addition, ρ(AD−1) = 1 because
matrix AD−1 is column stochastic. As a result, ρ(D−1A) = 1. Also, note that ¯P (cid:44) I − B +
AD−1B is column stochastic and therefore ρ( ¯P ) = 1. Furthermore,
ρ( ¯P ) = ρ( ¯P B−1DD−1B) = ρ(D−1B ¯P B−1D)
= ρ(D−1B(I − B + AD−1B)B−1D)
= ρ(I − B + BD−1A) = ρ(P ) = 1.
Since P is a nonnegative matrix, we can ask whether P is primitive,2 i.e., whether P m > 0
for some m ≥ 1. Since the digraph is strongly connected for 0 < βj < 1, ∀vj ∈ V and all the
main diagonal entries are positive, we easily conclude that m ≤ n − 1 [20, Lemma 8.5.5] and
P is primitive. Hence, there is no other eigenvalue with modulus equal to the spectral radius.
A sufficient condition for primitivity is that a nonnegative irreducible matrix A has at least one
positive diagonal element [21, Example 8.3.3], which means that some βj can also be set at
unity (as long as at least one is strictly smaller than unity). Hence, iteration (3) has a geometric
convergence rate R∞(P ) = − ln δ(P ), where δ(P ) (cid:44) max{λ : λ ∈ σ(P )), λ (cid:54)= 1}. In other
2A nonnegative matrix is said to be primitive if it is irreducible and has only one eigenvalue of maximum modulus [20].
words, δ(P ) is the second largest of the moduli of the eigenvalues of P (see also [1], [2], [4]).
Remark 1: If we change the coordinates of w[k] by introducing z[k] = B−1Dw[k], then
equation (3) becomes z[k + 1] = B−1DP D−1Bz[k] = ¯P z[k], which has total mass (1T z[k]) that
remains invariant throughout the iteration (and equal to 1T z[0] = 1T B−1D1).
B. Illustrative Example
In this illustrative example (borrowed from [19]), we demonstrate the validity of the proposed
algorithm in the network depicted in Figure 1. In our plots we are typically concerned with the
j=1 xj[k], where xj[k] is given in (1). If weight balance
absolute balance defined as ε[k] =(cid:80)n
is achieved, then ε[k] = 0 and xj[k] = 0, ∀vj ∈ V.
Fig. 1. A simple digraph which is weight-balanceable, but not bistochasticable due to the absence of self-loops. This example
is given in [19] in order to indicate that not all strongly connected digraphs are bistochasticable.
Figure 2 shows the absolute balance of Algorithm 1 when βj = 0.1, 0.5 and 0.9 for all vj ∈ V
as it evolves during the execution of the algorithm. These plots agree with the claims in Proposi-
ion 1 and validate that the algorithm convergences to a weight-balanced digraph with geometric
convergence rate. In this particular example, with the choice of βj = 0.1, 0.5 and 0.9, the rate
discussed in the proof of Proposition 1 is given by R∞(cid:0)P (βj)(cid:1) = 0.1204, 0.5180 and 0.2524,
respectively. The rates of convergence are also illustrated in Figure 3, where the convergence
rate of Algorithm 1 as obtained in Proposition 1 and the actual convergence rate of the algorithm
characterized by matrix (4). The theoretical and actual convergence rates coincide.
v1v2v3v4 Fig. 2. Absolute balance for weight-balancing algorithm (Algorithm 1) for the digraph depicted in Figure 1.
Fig. 3. Comparison of the theoretical rate of convergence as obtained in Proposition 1 and the actual convergence rate of the
weight-balancing algorithm (Algorithm 1) for the digraph depicted in Figure 1, for βj = 0.1, 0.5 and 0.9. The dashed lines
show the theoretical convergence rate, while the solid lines show the actual convergence rate of the algorithm.
The final weight-balanced digraph W (cid:63) is the same for all three cases mentioned above (because
0102030405000.10.20.30.40.5Weight−balancing with Algorithm 1 for different values of βjNumber of iterationsAbsolute balance βj=0.1 for all vjβj=0.5 for all vjβj=0.9 for all vj0102030405010−1510−1010−5100Rate of convergence of Algorithm 1 for different values of βjNumber of iterationsAbsolute balance Algorithm 1, βj=0.1Theoretical, βj=0.1Algorithm 1, βj=0.5Theoretical, βj=0.5Algorithm 1, βj=0.9Theoretical, βj=0.9 all the βj are identical -- but equal to a different constant each time) and is given by
0
0
0
1.4286
W (cid:63) =
.
0
0
0
0.7143 0.7143
0
0
1.4286
0
0
0
0.7143
Remark 2: If βj = 1, ∀vj ∈ V, then the weighted adjacency matrix P is not necessarily
primitive and hence the algorithm does not converge to weights that form a weight-balanced
digraph. We can see this in the counterexample depicted in Figure 4. We run Algorithm 1 for
Fig. 4. A simple digraph which is weight-balanceable, but the adjacency matrix is not primitive.
three different cases: (a) βj = 1, ∀vj ∈ V, (b) βj = 0.9, ∀vj ∈ V and (c) β1 = 0.9 and βj = 1,
∀vj ∈ V − {v1}. It can be seen from Figure 5 that for the first case the matrix is not primitive
and it does not converge, whereas for the other two cases it asymptotically converges (because
as long as one of the nodes has βj < 1, then the update matrix is primitive and the algorithm
forms a weight-balanced digraph).
V. DISTRIBUTED ALGORITHM FOR BISTOCHASTIC MATRIX FORMATION
A. Description of the algorithm
An algorithm is proposed, herein called Algorithm 2, with which a bistochastic adjacency
matrix is formed in a distributed fashion. One extra requirement for Algorithm 2, however, is
to maintain column stochasticity of the weighted adjacency matrix W [k] for all times k, so
v1v2v3v4 Fig. 5. A simple digraph which is weight-balanceable, but when βj = 1, ∀vj ∈ V, the weighted adjacency matrix is not
primitive and hence the algorithm does not converge to weights that form a weight-balanced digraph.
that it can be used for consensus problems simultaneously from the beginning without the need
to form the bistochastic matrix before running an algorithm. More specifically, we obtain a
sequence of column stochastic matrices W [0], W [1], W [2], . . . , W [k] such that limk→∞ W [k] =
W is bistochastic and thus the iteration x[k + 1] = W [k]x[k + 1], x[0] = x0, reaches average
consensus asymptotically [17].
Digraphs that are weight-balanceable do not necessarily admit a doubly stochastic assignment
[19, Theorem 4.1]. However, if self-weights are added then any strongly connected graph is
bistochasticable after adding enough self-loops [19, Corollary 4.2]. Algorithm 2 overcomes this
problem with the introduction of nonzero self weights wjj at each node vj ∈ V, and their
appropriate adjustment in a distributed manner. Algorithm 2 is described in detail below.
Proposition 2: If a digraph is strongly connected or is a collection of strongly connected
digraphs, then Algorithm 2 reaches a steady state weight matrix W ∗ that forms a bistochastic
digraph. Furthermore, the weights of all edges in the graph are nonzero.
Proof: As before, all the outgoing links have the same weight, i.e., wl(cid:48)j = wlj, ∀vl(cid:48)i, vl ∈
j , can be written in matrix form as
N +
j . Thus, the evolution of the weight wj (cid:44) wlj, ∀vl ∈ N +
follows.
w[k + 1] = P [k]w[k], w[0] = w0
(8)
0102030405000.20.40.60.81Weight−balancing problem − counterexampleNumber of iterationsAbsolute balance Algorithm 1 for βj=0.9 for all vjAlgorithm 1 for β1=0.9 and βj=1 for all vj≠ v1Algorithm 1 for βj=1 for all vj Algorithm 2 Bistochastic matrix formation algorithm
Input: A strongly connected digraph G(V,E) with n = V nodes and E edges (and no self-
loops).
Initialization: Set wlj[0] = 1/(1 + D+
Iterations: For k = 0, 1, 2, . . ., each node vj ∈ V updates the weights wlj, vl ∈ N +
performing the following steps:
1) It chooses βj[k] as follows:
j ), ∀vl ∈ N +
j ∪ {j}.
j , by
βj[k] =
where αj ∈ (0, 1).
2) It updates
αj
αj,
1−S+
j [k]
S−
j [k]−S+
j [k]
, S−
j [k] > S+
j [k],
otherwise,
(cid:32)
(cid:33)
,
S−
j [k]
D+
j
− wlj[k]
(5)
(6)
(7)
(9)
wlj[k + 1] = wlj[k] + βj[k]
j . [This is the same update as Algorithm 1, with the difference that
for all vl ∈ N +
the
proportionality constant βj can be adapted at each time step k, and is chosen so that
j [k + 1] ≤ 1 (so as to ensure that wjj can be chosen in Step 3 to be nonnegative and satisfy
S+
wjj + S+
3) Then, wjj ≥ 0 is assigned so that
stochasticity; i.e.,
the weighted adjacency matrix retains its column
j [k + 1] = 1).]
wjj[k + 1] = 1 − (cid:88)
wij[k + 1], ∀vj ∈ V.
where
i∈N +
j
1 − βj[k]
βj[k]/D+
j
Pji[k] (cid:44)
if i = j,
if i ∈ N −
j .
In order to make sure that the sum of each column can be made one by choosing a nonnegative
self-weight wjj, we need to establish (for all k) that S+
j , for
all vl ∈ N +
j . In our updates, there are two cases: (a) S−
j [k];
j [k + 1] ≤ 1 or wlj[k + 1] ≤ 1/D+
j [k] ≤ S+
j [k] > S+
j [k], and (b) S−
j [k], then βj[k] can be chosen to be any value αj, αj ∈ (0, 1). Then, the
j [k] ≤ S+
we analyze both cases below.
(a) When S−
algorithm is equivalent to Algorithm 1 and admits geometric convergence rates.
(b) When S−
j [k + 1] > 1 or wlj[k + 1] > 1/D+
S+
(5). In this case, for any initial S+
j [k], then wlj[k + 1], ∀vl ∈ N +
for any vl ∈ N +
j [0] for which S−
j [k] > S+
j
j , are increased. In order to avoid having
j , we choose βj[k] as shown in equation
j [k], from (6) we obtain that
j [k] > S+
j [k] + αj(1 − S+
j [k])
S+
j [k + 1] = S+
= (1 − αj)S+
j [k] + αj.
Hence, it is guaranteed that S+
shown by perfect induction that after n steps (for which S−
r ∈ Z+, r ≤ n) we have
j [k + 1] < 1, and as a result, βj[k] (cid:54)= 0 for all k. It can be easily
j [k + r] holds for all
j [k + r] > S+
j [k + n] = 1 − (1 − αj)n(1 − S+
S+
j [k]),
that approaches 1 as n → ∞. Note that someone can choose αj ∈ (0, 1) closer to 1 and guarantee
that βj[k] > , > 0, for all the iterations k.
Unlike the case of Algorithm 1, in Algorithm 2 we have a discrete-time switching dynam-
ical system whose stability condition and convergence rate are not as easily characterized. As
indicated by the various simulations we have tried (as well as the special case identified by
the remark below), it is very likely that the rate of convergence of Algorithm 2 is geometric,
however, this has not been formally established thus far.
Proposition 3: If a digraph is strongly connected or is a collection of strongly connected
j , m ≥ V, reaches
digraphs, then Algorithm 2 with initial condition wLj[0] =
a steady state weight matrix W ∗ that forms a bistochastic digraph, with geometric convergence
rate equal to R∞(P ) = − ln δ(P ), where
, ∀vL ∈ N +
m(1+D+
j )
1
if i = j,
if i ∈ N −
j .
Furthermore, the weights of all edges in the graph are nonzero.
Pji[k] (cid:44)
αj/D+
j
1 − αj
Proof: If it is possible for each node to know the number of nodes in the graph, or at least
j , where m ≥ V. Hence,
for all vL ∈ N +
1
an upper bound, then we can set wLj[0] =
we establish that S−
j [0] < 1 and S+
m(1+D+
j )
j [0] < 1 for all vj ∈ V. By equation (6) we have that
j [k + 1] = (1 − βj[k])S+
S+
j [k] + βj[k]S−
j [k]
j [k] < 1 and that βj[k] = αj, for all k. The self-weights are chosen so
which guarantees that S+
that we make each column sum up to one (and consequently the row sums, once the algorithm
converges). Therefore, with these initial conditions we can choose βj[k] to be constant (and
equal to αj ∈ (0, 1)) throughout the execution of the algorithm, and hence, admit geometric
convergence rate (because we are essentially back to Algorithm 1 with the only difference that
we also calculate the value of the self-weight (equal to 1 − D+
j wj[k]) each time). Therefore,
Algorithm 2 has a geometric convergence rate R∞(P ) = − ln δ(P ), as in Algorithm 1.
Remark 3: As before, if we change the coordinates of w[k] by introducing z[k] = B−1Dw[k],
then equation (8) becomes z[k + 1] = B−1DP [k]BD−1z[k] = ¯P [k]z[k], which has total mass
(1T z[k]) that remains invariant.
B. Illustrative Example
We consider a random strongly connected graph consisting of 50 nodes. The quantity of
interest in this case is the absolute balance of the graph at time step k; the absolute balance
Ab[k] is defined as
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 − (cid:88)
vi∈N −
j
(cid:88)
vj∈V
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) .
Ab[k] =
wji[k]
Since the weight matrix at each time step is column stochastic by construction, Ab[k] effectively
measures the distance of the weight matrix at time step k from a bistochastic matrix. As it can
be seen from Figure 6, the algorithm asymptotically converges to a bistochastic adjacency matrix
for different values of αj (in the simulations αj is chosen to be 0.9, 0.5 and 0.1).
A. Weight-Balanced Matrix Formation
VI. COMPARISONS
Here we run the algorithm for larger graphs (of size n = 50) and we compare the performance
of our algorithm against the imbalance-correcting algorithm in [15] in which every node vj adds
all of its weight imbalance xj to the outgoing node with the lowest weight w.
Fig. 6. Absolute balance for bistochastic formation algorithm (Algorithm 2) for different values of αj ∀vj ∈ V. It can be
observed that Algorithm 2 converges to a bistochastic adjacency matrix asymptotically.
The suggested weight-balancing algorithm (Algorithm 1) shows geometric convergence and
outperforms the algorithm suggested in [15], as shown in Figure 7.
B. Bistochastic Matrix Formation
Here we run the algorithm for larger graphs (of size n = 50) and we compare the performance
of our algorithm against a distributed algorithm suggested in [17] in which every node vj ∈ V
first chooses a particular weight wjj[k] ∈ (0, 1) for its self weight and it sets the weights of its
outgoing links to be wlj[k] = clj(1 − wjj[k]), where clj > 0, ∀vl ∈ N +
j , and(cid:80)
clj = 1.
vl∈N +
j
The suggested bistochastic formation algorithm (Algorithm 2) shows asymptotic convergence
and slightly outperforms the algorithm suggested in [17] for αj = 0.95, ∀vj ∈ V, as shown in
Figure 8.
VII. CONCLUSIONS AND FUTURE DIRECTIONS
In this paper we have developed two iterative algorithms: one for balancing a weighted
digraph and one for forming a bistochastic adjacency matrix in a digraph. Both algorithms
are distributed and the second one is a direct extension of the first one. The weight-balancing
algorithm is asymptotic and is shown to admit geometric convergence rate. The second algorithm,
010203040506070809010000.020.040.060.080.10.120.14Bistochastic matrix formation for a random graph of 50 nodes for various values of αjNumber of iterationsAbsolute balance αj=0.9 for all vjαj=0.5 for all vjαj=0.1 for all vj Top figure: Absolute balance plotted against the number of iterations for a random graph of 50 nodes for the distributed
Fig. 7.
weight-balancing (Algorithm 1) algorithm, the imbalance-correcting algorithm [15], and the weight-balancing algorithm proposed
in [16]. Bottom figure: Average balance plotted against the number of iterations for a 1000 random graphs of 50 nodes for the
distributed weight-balancing (Algorithm 1), the imbalance-correcting algorithm [15], and the weight-balancing algorithm proposed
in [16].
a modification of the weight-balancing algorithm, leads to a bistochastic digraph with asymptotic
convergence admitting geometric convergence rate for a certain set of initial values. The two
algorithms perform very well compared to existing algorithms, as illustrated in the examples.
Future work includes the analysis of the behavior of the suggested algorithms in the presence
of delays and changing topology.
0102030405060708090100050100150200250Average error vs Iteration for a random graph of 50 nodesNumber of iterationsAverage error Algorithm 1Imbalance−correcting algorithm in [14]Weight−balancing algorithm in [15]0102030405060708090100050100150200Average error vs Iteration for a 1000 graphs of 50 nodesNumber of iterationsAverage error Algorithm 1Imbalance−correcting algorithm in [14]Weight−balancing algorithm in [15] Fig. 8. Top figure: Absolute balance for bistochastic formation algorithms (Algorithm 2) and a distributed algorithm suggested in
[17] for a random strongly connected graph consisting of 50 nodes. Bottom figure: Average balance plotted against the number
of iterations for a 100 random strongly connected graphs of 50 nodes for the distributed bistochastic formation algorithms
(Algorithm 2) and a distributed algorithm suggested in [17].
REFERENCES
[1] A. Jadbabaie, J. Lin, and A. Morse, "Coordination of groups of mobile autonomous agents using nearest neighbor rules,"
IEEE Transactions on Automatic Control, vol. 48, no. 6, pp. 988 -- 1001, June 2003.
[2] R. Olfati-Saber and R. Murray, "Consensus problems in networks of agents with switching topology and time-delays,"
IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1520 -- 1533, September 2004.
[3] W. Ren and R. Beard, "Consensus seeking in multiagent systems under dynamically changing interaction topologies,"
IEEE Transactions on Automatic Control, vol. 50, no. 5, pp. 655 -- 661, 2005.
[4] L. Moreau, "Stability of multiagent systems with time-dependent communication links," IEEE Transactions on Automatic
Control, vol. 50, no. 2, pp. 169 -- 182, February 2005.
010203040506070809010000.050.10.150.20.25Average error vs Iteration for a random graph of 50 nodesNumber of iterationsAverage error Algorithm 2 with αj=0.95Algorithm suggested in [16]010203040506070809010000.050.10.150.20.25Average error vs Iteration for a 1000 graphs of 50 nodesNumber of iterationsAverage error Algorithm 2 with αj=0.95Algorithm suggested in [16] [5] D. Angeli and P. Bliman, "Stability of leaderless discrete-time multi-agent systems," Mathematics of Control, Signals, and
Systems, vol. 18, no. 4, pp. 293 -- 322, 2006.
[6] F. Xiao and L. Wang, "State consensus for multi-agent systems with switching topologies and time-varying delays,"
International Journal of Control, vol. 79, no. 10, pp. 1277 -- 1284, 2006.
[7] R. Olfati-Saber, J. Fax, and R. Murray, "Consensus and cooperation in networked multi-agent systems," Proceedings of
the IEEE, vol. 95, no. 1, pp. 215 -- 233, January 2007.
[8] D. Varagnolo, G. Pillonetto, and L. Schenato, "Distributed parametric and nonparametric regression with on-line
performance bounds computation," Automatica, vol. 48, no. 10, pp. 2468 -- 2481, October 2012.
[9] M. Rabbat and R. Nowak, "Distributed optimization in sensor networks," in Third International Symposium on Information
Processing in Sensor Networks (IPSN), April 2004, pp. 20 -- 27.
[10] K. Greene, J. Kniss, G. Luger, and C. Stern, "Satisficing the masses: Applying game theory to large-scale, democratic
decision problems," in International Conference on Computational Science and Engineering (CSE), vol. 4, August 2009,
pp. 1156 -- 1162.
[11] C. N. Hadjicostis and T. Charalambous, "Asynchronous coordination of distributed energy resources for the provisioning
of ancillary services," in 47th Annual Allerton Conference on Communication, Control, and Computing, September 2011,
pp. 1500 -- 1507.
[12] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Gossip algorithms: design, analysis and applications," in 24th Annual
Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), vol. 3, March 2005, pp. 1653 -- 1664.
[13] A. Nedic, A. Olshevsky, A. Ozdaglar, and J. Tsitsiklis, "On distributed averaging algorithms and quantization effects,"
IEEE Transactions on Automatic Control, vol. 54, no. 11, pp. 2506 -- 2517, November 2009.
[14] B. Johansson, M. Rabi, and M. Johansson, "A randomized incremental subgradient method for distributed optimization in
networked systems," SIAM J. on Optimization, vol. 20, no. 3, pp. 1157 -- 1170, August 2009.
[15] B. Gharesifard and J. Cort´es, "Distributed strategies for making a digraph weight-balanced," in 47th Annual Allerton
Conference on Communication, Control, and Computing, September 2009, pp. 771 -- 777.
[16] C. Hadjicostis and A. Rikos, "Distributed strategies for balancing a weighted digraph," in 20th Mediterranean Conference
on Control Automation (MED), July 2012, pp. 1141 -- 1146.
[17] A. Dominguez-Garcia and C. N. Hadjicostis, "Distributed matrix scaling and application to average consensus in directed
graphs," IEEE Transactions on Automatic Control, 2012 (to appear).
[18] J. Cort´es, "Distributed algorithms for reaching consensus on general functions," Automatica, vol. 44, no. 3, pp. 726 -- 737,
March 2008.
[19] B. Gharesifard and J. Cort´es, "When does a digraph admit a doubly stochastic adjacency matrix?" in American Control
Conference (ACC), July 2010, pp. 2440 -- 2445.
[20] R. A. Horn and C. R. Johnson, Matrix Analysis. Cambridge, CB2 2RU, UK: Cambridge University Press, 1985.
[21] C. D. Meyer, Matrix Analysis and Applied Linear Algebra. SIAM, 2000.
|
cs/0211029 | 1 | 0211 | 2002-11-22T12:10:55 | Modelling intracellular signalling networks using behaviour-based systems and the blackboard architecture | [
"cs.MA",
"q-bio.CB"
] | This paper proposes to model the intracellular signalling networks using a fusion of behaviour-based systems and the blackboard architecture. In virtue of this fusion, the model developed by us, which has been named Cellulat, allows to take account two essential aspects of the intracellular signalling networks: (1) the cognitive capabilities of certain types of networks' components and (2) the high level of spatial organization of these networks. A simple example of modelling of Ca2+ signalling pathways using Cellulat is presented here. An intracellular signalling virtual laboratory is being developed from Cellulat. | cs.MA | cs | Mode l l ing in trace llu lar s igna l ling network s u sing
behav iour-based systems and the b lackboard arch itecture
PEDRO PABLO GONZÁLEZ PÉREZ
CARLOS GERSHENSON
MAURA CÁRDENAS GARC ÍA
JAIME LAGUNEZ OTERO
Instituto de Química
Universidad Nacional Autónoma de México
Ciudad Universitaria, 04510, MÉX ICO, D.F.
[email protected] http://132.248.11.4
Abstract: This paper proposes to model the intracellular signalling networks using a fusion of behaviou r-bas ed sys tems
and the blackboard architecture. In virtue of this fusion, the model developed by us, which has been named Cellulat,
allows to take account two essential aspects of the intracellular signalling networks: (1) the cognitive capabilities of
certain types of networks’ components and (2) the high level of spatial organization of these networks. A simple
example of modelling of Ca2 + signalling pathways using Cellulat is presented here. An intracellular signalling virtual
laboratory is being developed from Cellulat.
Key-Words: behaviour-based systems, autonomous agents, blackboard architecture, intracellular signalling networks.
1 Introduction
Each cell in a multicellular organism receives specific
combinations of chemical signals generated by other
cells or by themselves. The final effect of the signals
received by a cell can be translated in the regulation of
the cell metabolism, in the alteration or maintenance of
its differentiation state, in the beginning of cellular
division and in its death. Once the extracellular signals
bind to the receptors, different process of signalling are
activated, which generate complex networks of
intra cellu lar signalling.
The more experimental data about cellular
functioning we have, the more useful the computational
models are to understand these signalling processes. The
computational models allow the visualization of the
components that integrate the network, in addition to
being able to predict or to suggest the effect of
disturbances caused on determined component or
sections of the s ignall ing pathway. A great variety of
information process ing computa tional models have been
inspired by the functioning of biological systems.
Examples of these are the artificial neural networks,
genetic algorithms and cellular automata. On the other
hand, several information p roces sing syst ems have
constituted different useful approaches to model the
functioning of diverse biological systems [16].
Within computer sciences, the artificial intelligence
area has constituted one of the main scenarios to model
biological systems. This fact responds to the great
variety of models, techniques and methods that support
this research area, many of which are inherited of
disciplines such as psychology, cognitive sciences and
neuroscience. Between the main techniques of artificial
intelligence and computer sciences commonly used to
model cellular signalling networks are artificial neural
networks [1 and 18], boolean networks [4 ], petri nets
[7], rule-based systems [3], cellular automata [4], and
multi-agent systems [5, 17 and 19].
The high complexity level of the intracellular
communication networks causes them to be difficult to
model when is considered some of these techniques
isolated. Nevertheless, when integrating the most
relevant and necessary elements of these techniques in
a singl e computa tional system, then it should be possible
to obtain a model much more robust of the intracellular
signalling pathway, and therefore, a better visualization,
understanding and prediction of the processes and
components that integrate the networks.
The theory of behaviour-based systems must
constitute a good approach to the intracellular signalling
networks. However, a powerful intracellular signalling
model should be achieved when the communication
between agents takes place through a shared data
structure, in which other cellular structures and elements
related with signalling pathways can be explicitly
represented. In this sense, the blackboard architecture
[15] results widely appropriate.
In this paper, we postulate that an effective and
robust model of intracellular signalling can be obtained
when the main structural and functional characteristics
of behaviour-based systems and
the blackboard
architecture are joined. That is, a cell can be seen as a
society of autonomous agents, where each agent
communicates with the others through the creation or
modification of signals on a shared data structure, named
“blackboard” . The au tonomous agen ts model
determinated functional components of the intracellular
signalling pathways, such as signalling proteins,
enzymes and other mechanisms. The blackboard levels
represent different cellular structures committed with the
signalling pathway, whereas the different objects created
on the blackboard represent secondary messengers,
activation or inhibition signals or others elements
belonging to intracellular medium.
In this way, when the autonomous agents are used in
an
intracellu lar signalling model,
the cogni tive
capabilities of certain components of the signalling
pathway can be taken in count; whereas the use of
blackboard architecture allows to capture the high level
of spatial organization exhibited by the intracellular
signalling networks.
This paper is structured as follows: the next section
presents an overview of the information processes of
signalling intr acel lular ne tworks. In sect ion 3 some
questions about behaviour-based systems are discussed.
Section 4 describes the basic components of blackboard
architecture and their functioning. In section 5 the
intracellular signalling model proposed by us is
presented and discussed. In section 6 an example of
modelling the Ca2 + signalling pathway using the
proposed model is shown. Some comments about current
and future work and conclusions are made in section 7.
2 An overview of signalling intracellular
networks
Each cell receives a great number of chemical signals
generated by other cells or by themselves. These
r e g u l at e
t h e i r m e t a b o l i sm, d e t e rm i n e
t h e ir
differentiation, and indicate when to divide or to start a
death process. In general the external signals are
transmitted to the interior of the cells through membrane
receptors activating diverse transduction pathways,
which can follow a same way and to generate a final
answer or to branch themselves to give rise to others.
For example activation of most membrane-associated
hormone receptors generates a diffusible intracellular
signal called a second messenger. Many different
receptors gener ate the same second messenger . Some
intracellular messengers are: cAMP (cyclic AMP),
cGMP (cycl ic GMP), I3P (Inositol triphosphate), DG
(diacylglycerol) and Ca 2 + (calcium).
The second messenger activation triggers multiple
signal transduction pathways like: the adenilate cyclase,
the hosphoinositide-calcium, the Mitogen activated
protein kinase
the
the JAK/STAT ,
(MAPK),
sphingomyelin-ceramide, and the ionic channel function
receptors pathways.
In particular i f we take the Ca2 + signall ing pathway,
the pathway looks like an interconnected netwo rk,
because the answer is to different stimuli, but the
different pathways cross and gener ate al terna tive
trajectories not easily visualized. The field of signal
transduction has centred on the discovery of increasing
crosstalk among signal ling pathways. Indeed, our
understanding of
the
topology of signalling
is
continuously evolving. Early work in the signalling field
featured the discovery of branched linear pathways,
which, resembled trees with a cell surface receptor at the
root.
3 Behav iour-based sys tems
The theory of behaviour-based systems [2 ] provides a
new philosophy for the construction of autonomous
agents, inspired by ethology. The goal of behaviour-
based systems (BBS)
is
to provide control
to
autonomous agents. An autonomous agent interacts
directly with its problem domain, it perceives its
environment through its sensors and acts on it through
its actuators. The autonomous agent environment is
commonly a dynamic, complex and unpredictable
environment, in which the autonomous agent tries to
satisfy a set of goals or motivations, varying in time. An
autonomous agent decides by itself how to relate its
external and internal inputs with its motor actions, in
such way that its goals can be satisfied [13]. Adaptation
is one of the desirable characteristics in autonomous
agents. An autonomous agent is adaptive if it has
abilities that allow him to improve its performance in
time.
A cell can be seen as an adaptive autonomous agent
or as a society of adaptive autonomous agents, where
each agent could exhibit a particular behaviour
depending on his cognitive capabilities. As an
autonomous agent, the cell perceives its external
environment through its surface receptors and acts on
this one by means of the generation of new signals,
which will be able to affect the behaviour of other cells.
The cell's external environment is also dynamic and
complex, in which an extensive range of signals and
combinations of these can exist. Furthermore, like in an
autonomous agent, the cell by itself decides how to
(signalling
received external signals
relate
the
in terna l
its
molecules)
to
s ignals
(secondary
messengers), so
that
their goals (differentiation,
proliferation, survival, among others) can be satisfied.
Final ly, the cell is able also to adapt to its environment.
That is, the cell is const antly adapti ng its behaviour to
the changes that take place in the environment and to the
signals perceived.
4 Blackboard architecture
The blackboard architecture concept was conceived by
artificial intelligence researchers in the 1970’s. The goal
of this research was to handle the problem of shared
information between multiple expert agents in problem
solving. The architecture was developed for the first
time in the syst em for language understanding, Hearsay
II [15]. La ter it has been used in a great variety of
problem domains and abstracted in many environments
for the construction of systems [10 and 14].
Blackboard architecture is defined in terms of three
basic components: (1) a set of independent modules,
named knowledge sources, which contain specific
knowledge about a problem domain, (2) a shared data
structure, named blackboard,
through which
the
knowledge sources communicate to each other, and (3)
a control mechanism, which determines the order in
which the knowledge sources will operate on the
blackboard. Figure 1 shows the blackboard architecture
component s and the rela tionships among them.
Fig. 1. Components of the Blackboard Architecture
The knowledge sources are modular software
subsystems that represent different points of view,
different strategies and different knowledge types, about
how to solve a problem or pa rt of a problem.
The blackboard is used as a central storage for all
shared information. The information on a blackboard
represents facts and deduct ions done by the knowledge
sources during the problem solving. The knowledge
sources produce changes on a blackboard, which
inc rementally lead to the formation of a solution or
acceptable set of solutions for the problem to be solved.
Because
the
knowledge
sources
respond
opportunistically to changes on the blackboard, a
mechanism is necessary that controls these changes and
decides, at every moment, which actions should be
taken. The control mechanism handles the interaction
between the blackboard, the knowledge sources and the
outsourcing such as users and control or data acquisition
subsystems.
5 Cellulat: an intracellular signalling
network model
Our proposal consists in modelling the cell as an
autonomous agent , which in turn is composed by a
society of autonomous agents, which communicates
through blackboard with others. That is, in the proposed
model the knowledge sources of the blackboard
architecture are considered as autonomous agents too.
The model proposed here constitutes a refinement and
adaptation of an action selection mechanism structured
on a blackboard architecture previously developed by us
named Internal Behaviou rs Network (IBeNet) [8, 9, 11
and 12]. Although the IBeNet was initially built to action
selection in autonomous agents (physical robots,
animats, or artificial creatures simulated on a computer),
it constitutes a working environment for the bottom-up
modelling of
information proces s ing sy s tems
characterized by: (1) coordination and opportunistic
integration of several tasks in real time, (2) use of
several abstraction or context levels for the different
types of information that participate in the processing
network, (3) decision making, (4) action selection and
(5) adaptation.
Although,
intracellular signalling model
the
proposed can be seen as the assignation of a new
semantic to the componen ts of the IBeNet, we assume
here that this fact is equivalent to the assignation of this
same semantic to the components of the blackboard
architecture, when the two following considerations are
done: (1) knowledge sources are seen as either internal
autonomous agents or interface autonomous agents and
(2) the blackboard architecture control is distributed now
between these two types of autonomous agents. The term
“internal autonomous agents” has been used to identify
to the autonomous agents which tasks are related with
the creation or modification of signals on
the
blackboard. An internal autonomous agent gets a signal
or combination or signals from a determinated
blackboard level and transduces these in other signal on
the same or other blackboard level. The way in which a
signal is transduced depends of the cognitive capabilities
of the internal autonomous agent On the other hand, the
function of an interface autonomous agent is to establish
the communica tion be tween the blackboard and the
external medium (they are similar to sensors or actuators
in BBS). Not all external signals or combinations of
these are recognized by an interface autonomous agent,
this
recognition depends both of
the signal
characteristics and the cognitive capabilities of the
interface autonomous agent.
In Figure 2 the architecture of intracellular signalling
model can be appreciated. This is a translation of
blackboard architecture shown in Figure 1 taking
account the considerations (1) and (2). In thi s way, it is
not necessary to explain the structural and functional
details of IBeNet here to understand the intracellular
signalling model proposed. It is only necessary to bear in
mind the blackboard architecture functioning explained
in section 3 and the new considerations mentioned
above. The in trace llula r signal ling model has been
named Cellulat (a kind of animat which behaves as a
cell).
Fig. 2. Architecture of the Cellulat
Three main components define the Ce llulat structure:
the blackboard, the internal autonomous agents and
interface autonomous agents.
The blackboard represents the cell’s internal
medium. The blackboard levels correspond to different
cellular structures through which occur the signal
transduction. In this way, the cellular membrane, the
citosol and the nucleus could be represented as different
blackboard levels. The solution elements recorded on the
blackboard represent two main types of intracellular
s igna ls :
second ary messen ge r molecules and
activation/inactivation signals. Both types of signals are
synthesized or created by internal autonomous agents
and these, either directly or indirectly, promote the
activation/inactivation of other internal autonomous
agents. Other types of cellular elements or structures can
be represented on the blackboard too.
The internal autonomous agents model components
of the intracellular signalling network such as proteins,
enzymes and other mechanisms necessary to carry out
the signal transduct ion. On the other hand, the interface
autonomous agents model the cell surface receptors and
the mechanisms for
the secretion of signalling
molecules. Each agent, independently of his type, has a
condition part and an action part; the way in which both
parts are linked depends on the complexity of the
intracellular component modeled by agent. For that
reason, agents who model complex components could
use more advanced techniques, such as neural networks,
genetic algorithms, or any combination of other
techniques, to link both parts. Agents which model less
complex components could use more sophisticated but
useful techniques, such as production rules, boolean
networks or others. The work of both types of agents is
event-directed. This
intracellular signal
is, each
registered on the blackboard or each extracellular signal
perceived constitutes an event, which could activate or
inactivate one or more autonomous agents. When an
internal autonomous agent is activated then this one
executes his action, which consists in the synthesis or
modification of a signal on the blackboard.
5.1 Cognitive capabilities of Cellulat’s agents
It is known that certain types of proteins exhibit several
cognit ive capabilities such as pattern recognition,
memory and handling fuzzy data [5] . The evolution,
learning and emergence of properties have also been
suggested in proteins networks [18]. In Cellulat, both
types of autonomous agents exh ibit several cognitive
capabilities including patter recognition, handling
uncertainty and fuzzy data, adaptive action selection,
memory and learning; which allow to the autonomous
agents to exhibit an adaptive behaviour. These cognitive
capabilities are supported by several artificial
intelligence paradigms
including
the
following
approaches:
reasoning, probab i l is t ic
rule-based
reasoning, artificial neural networks, genetic algorithms,
boolean networks, and fuzzy logic systems.
types of signals by the proteins, enzymes and others
mechanisms that operate on it. Examples of these signals
are the followings: PI-biphosphate, inositol triphosphate,
dyaclyglycerol, Ca2 + , and others signals indicating the
activation state of the proteins and enzymes. In this
representation, the blackboard levels have not been
identified.
5.2 Modelling spatial organization in Cellulat
Another important aspect to consider in the modelling of
intracellular signall ing networks
their spatial
is
organization. Experimental data recently obtained
suggest clearly that many intracellular signalling
networks exhibi t a high level of spatial organization.
Intracellular signalling models developed recently
approach to this question [6]. Cellulat allows to model
the
spatia l o rgan iza t ion
tak ing account
two
organizational criteria of the blackboard architecture.
One is the horizontal organization, given by the different
abstraction levels of the blackboard, which allow an
intralevel signal processing . The other is the vertical
organization, given by columns that cross vertically
different blackboard levels. These columns arise as
result of the adjoining work of several internal
autonomous agents that operate at a same section of
blackboard, which cover different blackboard levels. We
have named
these columns “agency columns”.
Convergence and divergence of agency columns could
occur, and these processes could be related with
evolution and learning of the signalling network [11] . In
this way, the model proposed allows to model
intracellular signalling pathways taking account their
spatial organization. That is, the two information
processing levels present in Cellulat (horizontal and
vertical) allow to estab lish a “topology preserving map”.
6 Modelling the Ca2+ signalling pathway
using Cellulat: A first approach
Figure 3 illustrates a first approach of the intracellular
signalling pathway modelling using Cellulat. Ins this
case, we have modelled the Ca2 + signalling pathway to
a still heavy resolution level. This is, this first model not
yet reveals struc tural detai ls of th e signal ling network
components.
As can be appreciated in Figure 3, the interface
autonomous agents model the G-protein-linked surface
receptors, whereas the internal autonomous agents
model components of the signalling network such as G
protein, phospholipase C-beta, prote in kinase C,
mechanism to release calcium, and others proteins and
enzymes. On the blackboard can be created different
Fig. 3. Cell ula t mod el o f the Ca2+ signal lin g pat hway.
In the model shown in Figure 3 it is necessary to
consider the following viewpoint: Since an internal or
interface autonomous agent can be structured by other
autonomous agents, then an internal autonomous agent
representing to a protein kinase C (or to others proteins
or enzymes) could encapsulate many proteins kinase C,
where each one of these is an internal autonomous agent
too. In this way, components of a same type could
j o i n t l y b e e n c a p s u l a t e d , w h i c h s im p l i f i e s
comprehension.
7 Conclusions
In this paper we have discussed why the fusion of the
behaviour-based systems theory and the blackboard
architecture constitutes a very suitable approach for the
modelling of cellular signalling networks, from a model
developed by us which has been named Cellulat. As a
result of this fusion, Cellulat allows to model two basic
aspects of the cellular signalling networks: (1) the
cognitive capabilities of certain types of proteins and
enzymes, and (2) the high level of spatial organization fo
the signalling networks. (1) was obtained when
considering
the adaptive autonomous agents as
functional components of the model, and using different
techniques of artificial intelligence to support the
cognit ive capabilities mentioned in Section 5.1; whereas
(2) was provided by the shared data structure, named
blackboard, which allows to model horizontal, vertical
and spa tial informat ion pro cessing.
A simple example of modelling of Ca2 + signalling
pathways using Cellulat was shown. Although still much
work to develop is left, as it is the case of the modelling
of the signalling network components at a fine resolution
level, we believe the proposed model constitutes a new
and useful approach for the modelling and understanding
of intracellular signalling networks.
Presently, Cellulat is being used in the modelling of
the signal transduction related with cellular senescence.
Our future work is directed towards the creation of a
cellular signall ing virtual laboratory, from the developed
computational model. This virtual laboratory must allow
to effect lesions on certain components or sections of the
signalling pathways, and to visualise the obtained
cellular behaviour as consequences of such lesions;
helping to understand how the different cellular
signalling pathways interact.
References:
[1] Bray, D. and Lay, S., Computer simulated evolution
of a network of cell signall ing molecules.
Biophysical Journal, Vol. 66, No. 4, 1994, pp.
972-977.
[2] Brooks, R. A., A robust layered control system for
a mobile robot, IEEE Journal of Robotics and
Automation, RA-2, 1986, April, pp. 14-23.
[3]Cárdenas, M., Búsqueda de blancos terapéuticos
empleando un modelo de señales intracelulares
mediadas por Ras, Tesis Doctoral, UNAM, México,
2000.
[4] Edwards, C.F., Computational models for cellular
information processing systems, unpublished
undergraduate thesis, University of Liverpool, 1995.
[5] Fisher, M.J., Paton, R.C. and Matsuno, K.,
Intracellular signalling proteins as 'smart' agents in
parallel distributed processes, BioSystems , 50, 1999,
pp 159-171.
[6] Fisher, M.J., Malcom, G. and Paton, R.C., Spatio-
logical processes
in
in trace llula r signal ling,
BioSystems , 55, 2000, pp 83-92.
[7] Fuss, H., Simulation of Biological Systems with
Petri Nets - Introduction to Modelling of Distributed
Systems. In Advances in System Analysis, Moller D.
(ed.), Braunschweig, Wiesbaden: Vieweg, 1-
12,1987.
[8] Gershenson, C., González, P.P. and Negrete, J.,
Action Selection Properties in a Software Simulated
Agent, Lecture Notes in Artificial Intelligence, Vol.
1793, 2000, pp 634-648.
[9] Gershenson, C., González, P. P. and Negrete, J.,
Thinking Adaptive: Towards a Behaviours Virtual
Labora tory. To be published in SAB'2000, Paris,
France.
[10] González, P.P. and J. Negrete., REDSIEX: A
cooper ative network of exper t systems with
blackboard architectures, Expert Systems , Vol. 14,
No. 4, 1997, pp. 180-189.
[11] González, P.P., Redes de Conductas Internas como
Nodos-P izarrón: Selección de Acciones y
Aprendizaje en un Robot Reactivo, Tesis Doctoral,
Instituto de Investigaciones Biomédicas/UNAM,
México, 2000.
[12] González, P.P., Negrete, J., Barreiro, A.J. and
Gershenson, C., A Model for Combination of
External and Internal Stimuli in the Action Selection
of an Autonomous Agent, Lecture Notes in Artificial
Intelligence, Vol. 1793, 2000, pp 621-633.
[13] Maes, P. , Adaptive Autonomous Agents, Journal
of Artificial Life, Vol. 1, No. 1 and 2, 1994.
[14] Negrete, J. and González, P.P., Net of multi-agent
expert systems with emergent control, Expert
Systems with Applications, Vol. 14, No. 1, 1998, pp.
109-116.
[15] Nii, H. P. , Blackboard systems. In A. Barr, P. R.
Cohen and E. A. Fe igenbaum (ed.), The Handbook
o f A r t i f i c i a l
I n t e l l i g e n c e , v o lumen
IV ,
Addison-Wesley Publishing Company, 1989.
[16] Paton, R.C ., Some Computational Models at the
Cellular Level, BioSystems , 29, 1993, pp 63-75.
[17] Paton, R.C., Staniford, G. and Kendall, G.,
Specifying Logical Agents in Cellular Hierarchies,
Proceedings of IPCAT (Information Processing in
Cells and Tissues) , Paton, R.C., Holcombe, M. y
Staniford, G. (Eds.), 1995, pp 302-317.
[18] Pritchard, L. and Dufton, M.J., Do Proteins Learn
to Evolve? The Hopfield Network as a Basis for the
Understanding of Protein Evolution, Journal of
Theoretical Biology, 202, 2000, pp 77-86.
[19] Schwab, E.D. and Pienta, K.J., Modeling signal
transduction in normal and cancer cells using
complex adaptive systems, Medical Hypotheses, 48,
1997, 111-123.
|
1511.03144 | 2 | 1511 | 2016-11-16T15:27:05 | Asynchronous Decentralized 20 Questions for Adaptive Search | [
"cs.MA",
"cs.IT",
"eess.SY",
"cs.IT",
"stat.ML"
] | This paper considers the problem of adaptively searching for an unknown target using multiple agents connected through a time-varying network topology. Agents are equipped with sensors capable of fast information processing, and we propose a decentralized collaborative algorithm for controlling their search given noisy observations. Specifically, we propose decentralized extensions of the adaptive query-based search strategy that combines elements from the 20 questions approach and social learning. Under standard assumptions on the time-varying network dynamics, we prove convergence to correct consensus on the value of the parameter as the number of iterations go to infinity. The convergence analysis takes a novel approach using martingale-based techniques combined with spectral graph theory. Our results establish that stability and consistency can be maintained even with one-way updating and randomized pairwise averaging, thus providing a scalable low complexity method with performance guarantees. We illustrate the effectiveness of our algorithm for random network topologies. | cs.MA | cs | Decentralized Adaptive Search using the Noisy 20
Questions Framework in Time-Varying Networks
1
Theodoros Tsiligkaridis, Member, IEEE
Abstract
This paper considers the problem of adaptively searching for an unknown target using multiple agents connected
through a time-varying network topology. Agents are equipped with sensors capable of fast information processing,
and we propose a decentralized collaborative algorithm for controlling their search given noisy observations.
Specifically, we propose decentralized extensions of the adaptive query-based search strategy that combines elements
from the 20 questions approach and social learning. Under standard assumptions on the time-varying network
dynamics, we prove convergence to correct consensus on the value of the parameter as the number of iterations
go to infinity. The convergence analysis takes a novel approach using martingale-based techniques combined with
spectral graph theory. Our results establish that stability and consistency can be maintained even with one-way
updating and randomized pairwise averaging, thus providing a scalable low complexity method with performance
guarantees. We illustrate the effectiveness of our algorithm for random network topologies.
I. INTRODUCTION
Consider a set of agents that try to estimate a parameter, e.g., estimate a target state or location, collectively. The
agents are connected by a time-varying information sharing network and can periodically query one of their local
neighbors about the target location. In this paper we adopt a generic observation model based on query-response
models where the queries are functions of agents' local information and successive queries are determined by a
feedback control policy. Specifically, in the 20 questions-type model considered in this paper, the observation of
each agent is coupled with the query region chosen by that agent, which is a function of its current local belief.
A centralized collaborative 20 questions framework was proposed and studied in [1], where a global centralized
controller jointly or sequentially formulates optimal queries about target location for all agents. This work was
later extended in [2] to the decentralized setting, in which each agent formulates his own query based on his local
information and exchanges beliefs with its neighbors in a synchronous fashion (i.e, updating the beliefs of all agents
simultaneously at each update step). The proposed decentralized algorithm therein consisted of two stages: 1) local
belief update; and 2) local information sharing. In stage 1 each agent implements the bisection query policy of [1]
to update their local belief function. In stage 2 the local belief functions are averaged over nearest neighborhoods
in the information sharing network. This two-stage algorithm was proven to converge to a consensus estimate of
the true state, assuming synchronous updating of all agents' beliefs in the network and irreducibility of the social
interaction graph.
The decentralized collaborative 20 questions problem is applicable to large scale collaborative stochastic search
applications where there is no centralized authority. Examples include: object tracking in camera networks [3]; road
tracking from satellite remote sensing networks [4]; and wide area surveillance networks [5]. Other applications
may include extending active testing approaches in the decentralized setting for classification problems, for instance
in vision, recommendation systems, and epidemic networks. The 20 questions paradigm is motivated by asking the
correct type of questions in the correct order and is applicable to various other domains where computational effort
and time are critical resources to manage.
In this paper, we consider a variant of the two-stage decentralized collaborative algorithm of [2] by relaxing the
assumption of a fixed network topology. We consider time-varying network topologies in which two randomly chosen
agents interact at each update step, giving rise to asynchronous updates (i.e., beliefs of all agents are not updated at
each update step). We analyze the convergence properties of the two-stage asynchronous decentralized collaborative
6
1
0
2
v
o
N
6
1
]
A
M
.
s
c
[
2
v
4
4
1
3
0
.
1
1
5
1
:
v
i
X
r
a
The material in this paper was presented in part at the 2016 IEEE International Symposium on Information Theory (ISIT), Barcelona,
Spain.
T. Tsiligkaridis is with MIT Lincoln Laboratory, Lexington, MA 02421 USA (email: [email protected]).
2
20 questions algorithm under appropriate conditions. This asynchronous model is applicable to practical sensor
networks where agents may be located in large geometric distances, and as a result their wireless communications
are unreliable or intermittent due to path occlusions or other environmental effects, and lead to time-varying network
topologies. Our analysis is based on smoothing techniques and martingale convergence theory in similar spirit to
[2]. However, the randomness due to the time-varying network topology and lack of strong connectivity at each
time instant introduce additional complications in the analysis that were absent when analyzing the static network
case of adaptive agents in [2].
In addition to theoretical analysis of the convergence of the proposed algorithm, numerical studies of per-
formance are provided showing interesting information behavior that information sharing yields. The benefit of
our asynchronous approach is that the asynchronous decentralized algorithm attains similar performance as its
synchronous counterpart introduced and analyzed in [2]. The lack of synchronization our approach offers is of
great practical interest because synchronization of a large number of agents can be difficult [6]. Furthermore, the
popular time-division-multiple-access (TDMA) communication protocol for distributed networks is only applicable
to synchronized networks. Even though the synchronous update scheme in [2] is proven to be convergent to the
correct limit, it is still an open question whether the same algorithm converges with an asynchronous implementation.
Counterexamples showing that asynchronous updates do not converge although synchronous updates converge are
presented in [7] for the standard consensus problem.
A. Prior Work
The noisy 20 questions problem, also known as Ulam's game, was introduced by Renyi [8] and was later
rediscovered by Ulam [9]. The first work making the connection between communication with feedback and noisy
search appeared in [10]. The probabilistic bisection algorithm dates back to the work of Horstein [11], where it
was originally proposed and analyzed heuristically in the contest of communication with noiseless feedback over
the binary symmetric channel. This algorithm was shown to achieve capacity for arbitrary memoryless channels in
[12], [13]. The probabilistic bisection algorithm was generalized to multiple players in [1] in the centralized setting,
and decentralized algorithms for probabilistic bisection search were proposed in [2].
Our work also differs from the works on 20 questions/active stochastic search of Jedynak, et al., [14], Castro &
Nowak [5], Waeber, et al., [15], and Tsiligkaridis, et al., [1] because we consider intermediate local belief sharing
between agents after each local bisection and update. In addition, in contrast to previous work, in the proposed
framework each agent incorporates the beliefs of its neighbors in a way that is agnostic of its neighbors' error proba-
bilities. The analysis of [16], [17] does not apply to our model since we consider controlled observations, although
we use a form of the social learning model of [16], [17]. While a randomized distributed averaging/consensus
problem was analyzed in [18], the convergence analysis is not applicable because we consider new information
injected in the dynamical system at each iteration (controlled information gathering) in addition to randomized
information sharing.
Although consensus to the true target location holds for the degenerate case of no agent collaboration by using
results in the existing literature, collaboration improves the rate of convergence of the estimation error. As shown in
the numerical results in this paper, the error decays faster as a function of iterations. This is the primary motivation
for studying such collaborative signal processing algorithms. However, proving convergence to the correct consensus
is the first step in analyzing such algorithms, and it is by no means a trivial one. In fact, even for the simple single-
agent case, only very recently, Waeber, et al [15] were able to prove the first rigorous convergence rate result for
the continuous probabilistic bisection algorithm. The focus of this paper is to establish convergence of decentralized
algorithms for probabilistic bisection in time-varying networks.
II. NOTATION
We define X∗ the true parameter, the target state in the sequel, and its domain as the unit interval X = [0, 1].
Let B(X ) be the set of all Borel-measurable subsets B ⊆ X . Let N = {1, . . . , M} index the M agents in an
interaction network, denoted by the vertex set N and the directed edges joining agents at time t ∈ N are captured
by E(t). Let At = {ai,j(t)} denote the interaction matrix at time t, which is a stochastic matrix (i.e., nonnegative
entries with rows summing to unity). At each time t, the time-varying network structure is modeled by the directed
graph (N , E(t)), where
E(t) = {(j, i) : [At]i,j > 0}
3
each x ∈ X . For any B ∈ B(X ), define Pt(B) as the vector of probabilities with i-th element equal to(cid:82)
region Ai,t = [0, Xi,t]. We let Fi,t(a) = Pi,t([0, a]) = (cid:82) a
Let Ai→j denote the interaction matrix when agent i performs a Bayesian update based on its query and averages
beliefs with agent j. In our model, a random agent i is chosen with probability qi and a collaborating agent j is
chosen with probability Pi,j at each update step. Thus, the interaction matrix At = Ai→j is chosen with probability
qiPi,j, and nodes i and j collaborate. The matrix P = {Pi,j} contains the probabilistic weights for collaboration
between agents and are zero when there is no edge in the information sharing network at any time; if Pi,j = 0,
then agent i cannot collaborate with agent j at any time.
Define the probability space (Ω,F, P) consisting of the sample space Ω generating the unknown state X∗ and
the observations {Yi,t+1} at times t = 0, 1, . . ., an event space F and a probability measure P. The expectation
operator E is defined with respect to P.
Define the belief of the i-th agent at time t on X as the posterior density pi,t(x) of target state x ∈ X based on all
of the information available to this agent at this time. Define the M × 1 vector pt(x) = [p1,t(x), . . . , pM,t(x)]T for
B pi,t(x)dx.
We define the query point/target estimate of the i-th agent as Xi,t. The query point is the right boundary of the
0 pi,t(x)dx denote the cumulative distribution function
associated with the density pi,t(·).
We assume that a randomly chosen agent i constructs a query at time t of the form "does X∗ lie in the region
Ai,t ⊂ X ?". We indicate this query with the binary variable Zi,t = I(X∗ ∈ Ai,t) to which agent i responds with
a binary response Yi,t+1, which is correct with probability 1 − i, and without loss of generality i ≤ 1/2. This
error model is equivalent to a binary symmetric channel (BSC) with crossover probability i. The query region Ai,t
depends on the accumulated information up to time t at agent i. Define the nested sequence of event spaces Ft,
Ft−1 ⊂ Ft, for all t ≥ 0, generated by the sequence of queries and responses. The queries {Ai,t}t≥0 are measurable
with respect to this filtration. Define the canonical basis vectors ei ∈ RM as [ei]j = I(j = i). The notation i.p.
denotes convergence in probability and a.s. denotes almost-sure convergence.
III. ASYNCHRONOUS DECENTRALIZED 20 QUESTIONS
Motivated by the work of [1], [2] and [16], we proceed as follows. As in the fixed-topology decentralized
algorithm in [2], starting with a collection of prior distributions {pi,0(x)}i∈N on X∗, the goal is to reach consensus
across the network through repeated querying and information sharing. Our proposed asynchronous decentralized
collaborative 20 questions algorithm consists of two stages. Motivated by the optimality of the bisection rule for
symmetric channels proved by Jedynak, et al., [14], the first stage bisects the posterior of a randomly chosen agent
i ∈ N (say with probability qi) at Xi,t and refines its own belief through Bayes' rule 1. In the second stage, agent
i collaborates with an agent j with probability Pi,j by averaging their beliefs (see Algorithm 1).
X pi,t(x)li(yx, Xi,t)dx
and can be shown to be equal to 1/2 (see proof of Lemma 1 in Appendix A). The bisection query points are
medians Xi,t = F −1
Some simplifications occur in Algorithm 1. The normalizing factor Zi,t(y) is given by(cid:82)
i,t (1/2) and the observation distribution is:
li(yx, Xi,t) = f (i)
1 (y)I(x ≤ Xi,t) + f (i)
0 (y)I(x > Xi,t).
z (·) are defined in (6). We note that the conditioning on the query region Ai,t (or query
where the distributions f (i)
point Xi,t in one-dimension) is necessary as the binary observation y is linked to the query in the 20 questions
model, in which the correct answer is obtained with probability 1 − i and the wrong answer is obtained with
probability i. We remark that the density li(yx, Xi,t) depends on the query point Xi,t, which is time-varying and
as a result, the density li(yx, Xi,t) is time-varying.
We remark that Algorithm 1 is fully decentralized, i.e., only local information processing and local information
sharing is needed. Furthermore, it operates in an asynchronous fashion as the belief updates are not occurring
simultaneously for all agents in the network. At each step, two agents collaborate with each other and update their
beliefs. This belief averaging leads to a non-Bayesian social learning scheme in similar spirit to [16], [17].
1In the asynchronous time model of Boyd, et al., [18], each agent has a clock that ticks according to a rate-one Poisson process. As a
result, the inner ticks at each agent i are distributed according to a rate-one exponential distribution, independently across agents and over
time. This corresponds to a single clock ticking according to a rate-M Poisson process at times {tk : k ≥ 1}, where {tk = tk+1 − tk}
are i.i.d. exponential random variables of rate M. Let ik ∈ N denote the agent whose clock ticked at time tk. It follows that ik are i.i.d.
Unif(N ), i.e., qi = 1/M.
Algorithm 1 Asynchronous Decentralized Bisection Search Algorithm for Time-Varying Networks
1: Input: N , P = {Pi,j : (i, j) ∈ N × N},{i : i ∈ N}
2: Output: { Xi,t : i ∈ N}
3: Initialize pi,0(·) to be positive everywhere.
4: repeat
5:
6:
7:
8:
9:
Bisect posterior density at median: Xi,t = F −1
Obtain (noisy) binary response yi,t+1 ∈ {0, 1}.
Choose an agent i ∈ N randomly (e.g., with probability qi).
Choose a collaborating agent j ∈ N with probability Pi,j.
Belief update:
i,t (1/2).
pi,t+1(x) = αipi,t(x)
li(yi,t+1x, Xi,t)
Zi,t(yi,t+1)
+ (1 − αi)pj,t(x),
pj,t+1(x) = pi,t+1(x),
pl,t+1(x) = pl,t(x),∀l (cid:54)= i, l (cid:54)= j,
x ∈ X .
where the observation p.m.f. is:
4
(1)
li(yx, Xi,t) = f (i)
1 (y)I(x ≤ Xi,t) + f (i)
0 (y)I(x > Xi,t),
and f (i)
1 (y) = (1 − i)I(y=1)I(y=0)
i
10: until convergence
y ∈ Y
0 (y) = 1 − f (i)
, f (i)
1 (y).
Collaboration becomes essential in certain scenarios when some agents in the network are completely unreliable
i = 1/2, and helps tremendously when agents are unreliable, i.e., i ≈ 1/2. When an agent is completely unreliable,
its learning capacity is zero and cannot localize the target on its own. With collaboration, neighboring agents may
steer the belief of unreliable agents closer to the true location (also see Section V for several experiments).
IV. CONVERGENCE OF ASYNCHRONOUS DECENTRALIZED ESTIMATION ALGORITHM
In this section convergence properties of Algorithm 1 are established under the assumptions below. The two
main theoretical results, Thm. 1 and Thm. 2, establish that the proposed algorithm attains asymptotic agreement
(consensus) and asymptotic consistency, respectively. Several technical lemmas are necessary and are proven in the
appendices. A block diagram showing the interdependencies between the lemmas and theorems in this section is
shown in Fig. 1.
Fig. 1.
consensus limit.
The flow of the analysis for establishing convergence of the asynchronous decentralized 20-questions algorithm to the correct
According to Algorithm 1, the density evolution described by (1) can be written in matrix form as:
pt+1(x) = (At + Dt(x))pt(x)
(2)
Lemma 1Lemma 2Lemma 3Lemma 4Lemma 5Lemma 6Theorem 1(Asymptotic Agreement)Theorem 2(Consistency) where At accounts for the averaging action of the iteration and Dt(x) is the innovation. According to Algorithm
1, it follows that with probability qiPi,j, agents i and j collaborate, resulting in:
At = Ai→j = IM + (ei + ej)(αiei + (1 − αi)ej)T
− (eieT
(cid:32)
i + ejeT
j )
li(Yi,t+1x, Xi,t)
Zi,t(Yi,t+1)
− 1
(cid:33)
(ei + ej)eT
i
(3)
(4)
Dt(x) = Di→j(x) = αi
A. Assumptions
To simplify the analysis of Algorithm 1, we make the following assumptions, which are comparable to those
made in [1], [2] and [16].
Each agent's response is governed by the conditional distribution:
li(yix, Ai,t)
def
= P (Yi,t+1 = yiAi,t, X∗ = x)
(cid:40)
=
f (i)
z (yi) =
1 (yi), x ∈ Ai,t
f (i)
0 (yi), x /∈ Ai,t
f (i)
(cid:40)
1 − i, yi = z
yi (cid:54)= z
i,
5
(5)
(6)
Assumption 1. (Memoryless Binary Symmetric Channels) We model the agents' responses as independent (mem-
oryless) binary symmetric channels (BSC) [19] with crossover probabilities i ∈ (0, 1/2]. The probability mass
function f (i)
z (Yi,t+1) = P (Yi,t+1Zi,t = z) is:
for i = 1, . . . , M, z ∈ {0, 1}. Define the set of agents I1 = {i ∈ N : i < 1/2} and assume the set I1 is nonempty.
The condition i < 1/2 implies that the response of an agent i is 'almost correct'. Agents i /∈ I1 do not have
the ability to localize the target on their own.
Assumption 2. (Average Strong Connectivity) As in [16], we assume that the network is strongly connected on
i,j qiPi,jAi→j is irreducible. Furthermore, we assume α = mini αi > 0 to ensure
average, i.e. A = E[At] =(cid:80)
collaboration.
The following standard assumption about the network connectivity over time [20], [21], [22], [23], [24] will be
made to prove the correctness of the asymptotic limit of CDF's in Thm. 2. We remark that this assumption is not
needed to prove asymptotic consensus (Thm. 1).
Assumption 3. (Strong Connectivity over Interval) There exists R such the graph (N , E(t) ∪ E(t + 1) ∪ ··· ∪
E(t + R − 1)) is the same strongly connected graph for all t.
This condition is true with high probability as R gets large.
B. Analysis
The principal component that enables the proofs of convergence of Algorithm 1 is Equation (1), that propagates
the vector of belief functions forward in time. In (2), At = Ait→jt is the time-varying interaction matrix between
agents it and jt and Dt(x) = Dit→jt(x) is a diagonal time-varying matrix dependent on the response of agent i,
yi,t+1, the query region Ait,t ⊂ X and the state x ∈ X . For intuition, (2) can be written as a sum of two terms:
(cid:18)
pt+1(x) = pt(x) − eipi,t(x) − ejpj,t(x)
li(yi,t+1x, Ai,t)
+ (ei + ej)
αipi,t(x)
Zi,t(yi,t+1)
(cid:19)
+ (1 − αi)pj,t(x)
The first term simply zeroes out the i and j components and the second term fills them in with the average between
the updated belief of agent i and the current belief of agent j. The rest of the components are left intact.
M(cid:88)
l=1
Proposition 1 provides bounds on the dynamic range of Ax, where x is any arbitrary vector (see Theorem 3.1
in [25]). The coefficient of ergodicity of an interaction matrix A is defined as [25], [26]:
τ1(A)
def
=
1
2
max
i(cid:54)=j
(cid:107)AT (ei − ej)(cid:107)1 =
1
2
max
i(cid:54)=j
ai,l − aj,l
(7)
This coefficient satisfies τ1(·) ∈ [0, 1]. The most non-ergodic interaction matrix is the identity matrix IM , for which
τ1(IM ) = 1 and there is no information sharing, and the other extreme is the full-rank matrix τ1( 1
Proposition 1. (Contraction Property of A) Assume A = {ai,j} is a M × M stochastic matrix. Let x be an
arbitrary non-negative vector. Then, we have for all pairs (i, j):
1
M 1
¯
¯
T ) = 0.
(cid:16)
(cid:17)
[Ax]i − [Ax]j ≤ τ1(A)
xi − min
i
xi
max
i
Although At is not irreducible, A = E[At] is, which is used in Lemmas 2,3. Next, we recall a tight smooth
approximation to the non-smooth maximum and minima operators. Similar results have appeared in Prop. 1 in [27]
and p. 72 in [28].
Proposition 2. (Tight Smooth Approximation to Maximum/Minimum Operator) Let a ∈ RM be an arbitrary vector.
Then, we have for all γ > 0:
and
min
i
ai ≥ − 1
γ
log
e−γai
≥ min
i
ai − log M
γ
Lemma 1. Consider Algorithm 1. Let B ∈ B(X ). Then, we have:
max
i
ai ≤ 1
γ
log
eγai
≤ max
i
ai +
log M
γ
(cid:32) M(cid:88)
(cid:32) M(cid:88)
i=1
i=1
(cid:33)
(cid:33)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
(cid:35)
Dt(x)pt(x)dx
= 0.
(cid:34)(cid:90)
E
6
(8)
(9)
where Dt(x) was defined in (4).
Proof: See Appendix A.
B
Lemmas 1 and 2 imply that the term(cid:82)
Proof: See Appendix B.
Lemma 2. Consider Algorithm 1. Let B ∈ B(X ). Then, we have E[vT Pt+1(B)Ft] = vT Pt(B) for some positive
vector v (cid:31) 0, and limt→∞ vT Pt(B) exists almost surely.
Lemma 3. Consider Algorithm 1 and let d = 1. Let B = [0, b] ∈ B(X ). Let v denote the positive left eigenvector
of A = Ei,j[Ai→j]. Define the variable:
B vT Dt(x)pt(x)dx is a martingale difference noise term.
(cid:32) M(cid:88)
i,j=1
Λt(B, P, )
def
=
1
evT Pt(B)
qiPi,jevT Ai→jPt(B)
(cid:33)
× cosh((vi + vj)ai,i(1 − 2i)µi,t(B))
(10)
where
Then, we have Λt
a.s.−→ 1 as t → ∞.
Proof: See Appendix C.
µi,t(B)
def
= min{Pi,t(B), 1 − Pi,t(B)}.
Lemma 4. Consider the same setup as Lemma 3. Then, Λt(B) a.s.−→ 1 implies µi,t(B) a.s.−→ 0 for all i ∈ N as
t → ∞.
7
Proof: See Appendix D.
Define the dynamic range (with respect to all agents in the network) of the posterior probability that X∗ lies in
set B ⊂ X :
Also, define the innovation:
di,t+1(B)
def
=
Vt(B)
def
= max
i
Pi,t(B) − min
Dt(x)pt(x)dx
=
i
B
(cid:21)
i
(cid:90)
Pi,t(B)
(11)
[Dt(x)]i,ipi,t(x)dx
We next prove a lemma that shows that the dynamic range Vt(B) has a useful upper bound.
Lemma 5. Consider Algorithm 1. Let B = [0, b] with b ≤ 1. Then, for all R ∈ N:
Vt+R(B) ≤ τ1(At+R−1 ··· At)Vt(B)
(cid:17)
di,t+R−k(B) − min
i
max
i
di,t+R−k(B)
(12)
(cid:20)(cid:90)
B
R−1(cid:88)
(cid:16)
+
Proof: See Appendix E.
k=0
To show convergence of the integrated beliefs of all agents in the network to a common limiting belief, it suffices
i.p.→ 0. Theorem 1 shows convergence of asymptotic beliefs to a common limiting belief. The structure
to show Vt(B)
of the limiting belief is given in Theorem 2.
Theorem 1. Consider Algorithm 1 and let the Assumptions 1-2 hold. Let B = [0, b], b ≤ 1. Then, consensus of the
agents' beliefs is asymptotically achieved across the network:
Vt(B) = max
i
Pi,t(B) − min
i
Pi,t(B)
i.p.−→ 0
as t → ∞.
Proof: See Appendix F.
Theorem 1 establishes that Algorithm 1 produces belief functions that become identical over all agents. This
establishes asymptotic consensus among the beliefs, i.e., that as time goes on all agents come to agreement about
the uncertainty in the target state. It remains to show the limiting belief is in fact concentrated at the true target
state X∗ (Thm. 2).
Lemma 6. Consider Algorithm 1. Assume pi,0(X∗) > 0,∀i ∈ N . Then, the posteriors evaluated at the true target
state X∗ have the following asymptotic behavior:
lim inf
m→∞
1
m
for some ci > 0.
Proof: See Appendix G.
M(cid:88)
i=1
ci log(pi,mR(X∗)) ≥ K (a.s.)
Now, we are ready to prove the main consistency result of the asymptotic beliefs. The proof is based on the
consensus result of Theorem 1.
Theorem 2. Consider Algorithm 1 and let Assumptions 1-3 hold. Let B = [0, b], b ≤ 1. Then, we have for each
i ∈ N :
Fi,t(b) = Pi,t(B)
i.p.−→ F∞(b) =
(cid:26) 0,
1,
b < X∗
b > X∗
as t → ∞. In addition, for all i ∈ N :
Xi,t
def
=
(cid:90) 1
x=0
xpi,t(x)dx
i.p.−→ X∗
(13)
Proof: Theorem 1 implies that for each agent i,
8
i.p.→ F∞(b)
Fi,t(b)
the constant I(b > X∗). Lemma 6 implies that(cid:80)M
(14)
as t → ∞, where F∞(b) is a common limiting random variable. To finish the proof, we show F∞(b) is equal to
i=1 ci log pi,mR(X∗) → ∞ almost surely as m → ∞. Thus, there
exists an agent i0 ∈ I1 such that pi0,mR(X∗) → ∞. Lemma 4 implies µi0,t(b(cid:48)) = min{Fi0,t(b(cid:48)), 1− Fi0,t(b(cid:48))} a.s.→ 0
for any b(cid:48) ∈ [0, 1]. This asymptotic result, combined with the monotonicity of the CDF operator Fi0,t(·) and
pi0,mR(X∗) a.s.→ ∞ imply Fi0,mR(b) → I(b > X∗). It then follows from (14) that Fi,t(b)
i.p.→ F∞(b) = I(b > X∗)
for all i ∈ N . The second part (i.e., (13)) of the proof is identical to the one of Theorem 2 in [2].
V. EXPERIMENTAL VALIDATION
This section presents simulations that validate the theory in Section IV and demonstrate the benefits of the
proposed asynchronous decentralized 20 questions algorithm. We compare the performance of the asynchronous
algorithm proposed in this paper with the synchronous counterpart studied in [2], the centralized fully Bayesian
estimator implemented via the basic equivalence principle derived in [1] and the standard query-based estimator
with no information sharing. The root mean-squared error (RMSE) was chosen as a performance metric and we
consider the performance for strongly connected geometric random graphs [29] to model ad-hoc wireless network
topologies. An example geometric random graph is shown in Fig. 2.
Fig. 2. Geometric random graph with M = 20 nodes over unit square. This topology defines the zero pattern of the stochastic interaction
matrix P. The nonzero probabilities Pi,j for each row/node i are chosen to be uniform, i.e., each agent is equally likely to communicated
with any of its neighbors.
To ensure fairness in the comparisons, each algorithm iteration consists of M queries. Thus, the performance at
each iteration of the synchronous setup (in which all agents are queried about the target location) should be compared
with the performance at every M iterations of the asynchronous setup. This is denoted as "effective iteration" in the
figures. We average performance over T = 200 Monte Carlo trials for each random graph realization. Performance
is further averaged over 10 geometric random graphs to obtain the ensemble-average RMSE's. The average and
worst-case RMSE metrics were calculated as:
M(cid:88)
i=1
1
M
(cid:118)(cid:117)(cid:117)(cid:116) 1
(cid:118)(cid:117)(cid:117)(cid:116) 1
T
T
T(cid:88)
T(cid:88)
t=1
t=1
RMSEavg =
RMSEmax =
( Xi,t − X∗)2
( Xi,t − X∗)2
max
i
00.20.40.60.8100.10.20.30.40.50.60.70.80.91 Figures 3 and 4 show the average and worst-case RMSE performance of the network as a function of iteration
for the case of M = 20 homogeneous agents, i.e., all agents are unreliable with error probability = 0.45. We
observe that the asynchronous and synchronous algorithms both uniformly outperform the case of no information
sharing over all iterations. We remark that the asynchronous algorithm seems to have a slower asymptotic rate
of convergence as compared to the synchronous algorithm, while it seems to improve the RMSE more than the
synchronous counterpart for the first few iterations. This may be due to biasing effects that occur in the synchronous
algorithm; i.e., the belief is perturbed by multiple unreliable neighbors at each update step. In the asynchronous
algorithm on the other hand, there are less perturbations in the initial learning stage since it consists of pairwise
belief averaging at each update step.
9
Fig. 3.
= 0.45. The average RMSE is lower for the case of information sharing vs. the case of no information sharing.
Average RMSE performance across the network. Homogeneous network of M = 20 unreliable agents with error probability
Fig. 4. Worst-case RMSE performance across the network. Homogeneous network of M = 20 unreliable agents with error probability
= 0.45. The worst-case MSE across the network is lower for the case of information sharing vs. the case of no information sharing. The
asynchronous algorithm outperforms the synchronous algorithm in the initial learning stage.
Figures 5 and 6 show the average and worst-case RMSE performance of the network for the case of M = 20
heterogeneous agents, i.e., three agents are reliable with error probability = 0.05 and the rest are unreliable with
error probability = 0.45. Here, the reliable agents speed up the convergence of the unreliable agents through belief
averaging. We observe the interesting result that the asynchronous algorithm uniformly outperforms the synchronous
algorithm over all iterations, both in terms of average and worst-case RMSE. This can be attributed to the fact that
Effective Iteration20406080100120140160180200Average RMSE00.10.20.30.40.50.60.70.80.91SynchronousAsynchronousCentralizedNo info sharingEffective Iteration20406080100120140160180200Max RMSE00.10.20.30.40.50.60.70.80.91SynchronousAsynchronousCentralizedNo info sharing one-way updating and pairwise averaging induce less bias when combining beliefs at each node at each update step,
effectively spreading the good information around the network in a simplified fashion without being influenced too
much by multiple neighbors. Thus, asynchronous implementations of the non-Bayesian decentralized 20 questions
algorithm have the potential to improve network-wide estimation performance and getting closer to the centralized
Bayesian performance, in addition to requiring significantly less infrastructure and computational complexity, in
comparison to synchronous implementations.
10
Fig. 5.
Average RMSE performance across the network. Heterogeneous network of M = 20 agents, three of which are reliable with
error probability = 0.05 and the remaining ones are unreliable with error probability probability = 0.45. The asynchronous algorithm
outperforms the synchronous decentralized estimation algorithm with information sharing, and the algorithm with no information sharing.
Fig. 6. Worst-case RMSE performance across the network. Heterogeneous network of M = 20 agents, three of which are reliable with
error probability = 0.05 and the remaining ones are unreliable with error probability probability = 0.45. The asynchronous algorithm
outperforms the synchronous decentralized estimation algorithm with information sharing, and the algorithm with no information sharing.
VI. CONCLUSION
We introduced an asynchronous version of decentralized 20 questions with noise based on one-way updating and
pairwise belief averaging, and analyzed its convergence properties. We also illustrated several benefits of information
sharing as compared to no information sharing, and asynchronous vs. synchronous implementations. Asymptotic
convergence properties of the agents' beliefs were derived, showing that they reach consensus to the true belief.
Numerical experiments were presented to validate the convergence properties of the algorithm.
Effective Iteration20406080100120140160180200Average RMSE00.10.20.30.40.50.60.70.80.91SynchronousAsynchronousCentralizedNo info sharingEffective Iteration20406080100120140160180200Max RMSE00.10.20.30.40.50.60.70.80.91SynchronousAsynchronousCentralizedNo info sharing APPENDIX A
11
PROOF OF LEMMA 1
Proof: Note the following:
E
Decomposing the inner term, we obtain:
(cid:21)(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
Di→j(x)pt(x)dx
(cid:21)
(cid:20)(cid:90)
B
E
(cid:20)
= Ei,j
Dt(x)pt(x)dx
(cid:12)(cid:12)(cid:12)Ft
(cid:20)(cid:90)
(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
(cid:18) li(yi,t+1x, Ai,t)
(cid:20)(cid:90)
Zi,t(yi,t+1)
B
− 1
(cid:19)
Di→j(x)pt(x)dx
(cid:20)(cid:90)
E
= E
(cid:20)(cid:90)
B
αi
B
= αi(ei + ej)E
(2li(yi,t+1x, Ai,t) − 1)pi,t(x)dx
(ei + ej)pi,t(x)dx
Zi,t(y)
(cid:90)
B
(cid:16)
(cid:90)
B
= 2
1 (y)I(x ∈ Ai,t) + f (i)
f (i)
0 (y)I(x /∈ Ai,t)
dx
=
pi,t(x)
X
1 (y)Pi,t(Ai,t) + f (i)
= f (i)
0 (y)(1 − Pi,t(Ai,t)) = 1/2
(cid:16)
(2li(yi,t+1x, Ai,t) − 1)pi,t(x)dx
1 (yi,t+1)Pi,t(B ∩ Ai,t) + f (i)
f (i)
(cid:20)(cid:90)
− Pi,t(B)
E
(2li(yi,t+1x, Ai,t) − 1)pi,t(x)dx
0 (yi,t+1)Pi,t(B ∩ Ac
i,t)
(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
= 0
1 (yi,t+1)] = 1/2, we obtain:
(cid:17)
(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
(cid:17)
where we used the fact that Zi,t(y) = 1/2 for all y ∈ Y. This follows from the probabilistic bisection property:
where we used the fact f (i)
1 (y) + f (i)
0 (y) = 1. From the definition of li(yx, Ai,t), it follows that:
Taking the conditional expectation, and using E[f (i)
This concludes the lemma.
B
APPENDIX B
PROOF OF LEMMA 2
Proof: From the tower property of conditional expectation, we obtain:
(At + Dt(x))pt(x)dx
= Ei,j
(At + Dt(x))pt(x)dx
E[Pt+1(B)Ft]
= E
(cid:20)(cid:90)
B
E
(cid:20)(cid:90)
(cid:20)
M(cid:88)
M(cid:88)
i,j=1
=
=
(cid:20)
B
i,j=1
= APt(B)
qiPi,jAi→j
(cid:21)
(cid:12)(cid:12)(cid:12)Ft
(cid:90)
Pt(B) + E
B
qiPi,jE
Ai→jPt(B) +
(cid:21)(cid:12)(cid:12)(cid:12)Ft
(cid:12)(cid:12)(cid:12)Ft, (i, j)
(cid:20)(cid:90)
Di→j(x)pt(x)dx
(cid:21)
(cid:12)(cid:12)(cid:12)Ft, (i, j)
(cid:21)
(cid:12)(cid:12)(cid:12)Ft
Dt(x)pt(x)dx
B
(cid:21)
(15)
where we used the result of Lemma 1. From strong connectivity (i.e., Assumption 2), it follows that A is an
irreducible stochastic matrix. Thus, there exists a left eigenvector v ∈ RM with strictly positive entries corresponding
to a unit eigenvalue-i.e., vT = vT A [30]. Multiplying both sides of (15) from the left by vT , we obtain
E[vT Pt+1(B)Ft] = vT Pt(B). Thus, the process {vT Pt(B) : t ≥ 0} is a martingale with respect to the filtration
Ft. We note that it is bounded below by zero and above by (cid:107)v(cid:107)1 almost surely. From the martingale convergence
theorem [31], it follows that it converges almost surely.
12
APPENDIX C
PROOF OF LEMMA 3
Proof: Define the tilted measure variable ζt(B)
def
= exp(vT Pt(B)). From Lemma 2 and Jensen's inequality, it
follows that
E[ζt+1(B)Ft] ≥ ζt(B)
so the process {ζt(B) : t ≥ 0} is a submartingale with respect to the filtration Ft. From the proof of Lemma 2, it
follows that ζt(B) is bounded a.s., so by the martingale convergence theorem [31], it follows that limt→∞ ζt(B)
exists and is finite almost surely. Define the ratio βt+1(B)
ζt(B) . As a result, we have from Lemma 2:
def
= ζt+1(B)
Since the variables in the limit on the LHS are bounded a.s., i.e., βt+1(B) ≤ e(cid:107)v(cid:107)1, the dominated convergence
theorem for conditional expectations [32] implies:
lim
t→∞ βt+1(B)
= lim
t→∞
a.s.= 1
(cid:34)
evT Pt(B)
B Dt(x)pt(x)dx(cid:1)
evT AtPt(B) exp(cid:0)vT(cid:82)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
(cid:35)
E(cid:104)
evT AtPt(B)evT(cid:82)
(cid:88)
a.s.−→ 1
βt+1(B)
E
1
B Dt(x)pt(x)dxFt
E[βt+1(B)Ft]
=
1
evT Pt(B)
(16)
(cid:105)
as t → ∞. The conditional expectation can be expanded as:
=
qiPi,j
evT Pt(B)
× evT Ai→jPt(B)E[evT(cid:82)
i,j
Next, we analyze the ratio of exponentials for two separate cases. First, consider the case Pi,t([0, b]) =(cid:82) b
B Dt(x)pt(x)dxFt, (i, j)]
1/2. Using the definition of Xi,t, it follows that b ≤ Xi,t. This implies that li(yx, Ai,t) = f (i)
Using this fact and P (Yi,t+1 = yFt, (i, j)) = 1/2:
(17)
0 pi,t(x)dx ≤
1 (y) for all x ≤ b.
E[evT(cid:82)
(cid:16)
1
2
B Dt(x)pt(x)dxFt, (i, j)]
e(vi+vj)αi(1−2i)Pi,t(B) + e−(vi+vj)αi(1−2i)Pi,t(B)(cid:17)
=
= cosh((vi + vj)αi(1 − 2i)Pi,t(B))
(18)
where we used the fact that (ea +e−a)/2 = cosh(a). Second, considering the complementary case Pi,t([0, b]) > 1/2,
it follows that b > Xi,t, and:
Combining the two cases (18) and (19) and using the definition of µi,t(B), we obtain:
(19)
E[evT(cid:82)
B Dt(x)pt(x)dxFt, (i, j)]
= cosh((vi + vj)αi(1 − 2i)Pi,t(Bc))
E[evT(cid:82)
B Dt(x)pt(x)dxFt, (i, j)]
= cosh (viαi(1 − 2i)µi,t(B))
The proof is completed by substituting this expression into (17).
APPENDIX D
PROOF OF LEMMA 4
13
Proof: Using the definition of Λt, Jensen's inequality and the eigenrelation vT = vT A:
evT Ai→jPt(B) cosh((vi + vj)αi(1 − 2i)µi,t(B))
evT Ai→jPt(B)+(vi+vj)αi(1−2i)µi,t(B)
(cid:105)
(cid:104)
(cid:34)
1
2
Λt(B)
=
=
≥
1
evT Pt(B)
1
evT Pt(B)
Ei,j
Ei,j
+
1
2
+
1
2
1
evT Pt(B)
(cid:32)
1
2
(cid:32)
(cid:35)
(cid:33)
(cid:33)
evT Ai→jPt(B)−(vi+vj)αi(1−2i)µi,t(B)
evT APt(B)+Ei,j[(vi+vj)αi(1−2i)µi,t(B)]
evT APt(B)−Ei,j[(vi+vj)αi(1−2i)µi,t(B)]
=
1
2
1
evT Pt(B)
evT Pt(B)e
Ei,j[(vi+vj)αi(1−2i)µi,t(B)]
+ evT Pt(B)e−Ei,j[(vi+vj)αi(1−2i)µi,t(B)]
= cosh(Ei,j[(vi + vj)αi(1 − 2i)µi,t(B)])
≥ cosh(Ei[viαi(1 − 2i)µi,t(B)]) ≥ 1
We also used the fact that ea+e−a
above that cosh(Ei[viαi(1 − 2i)µi,t(B)]) a.s.−→ 1. From the property of cosh(·), it follows that
= cosh(a) ≥ 1 for all reals a. Since Λt(B) a.s.−→ 1, it follows from the bound
2
M(cid:88)
qiviαi(1 − 2i)µi,t(B) a.s.−→ 0
which further implies (1 − 2i)µi,t(B) a.s.−→ 0 for all i, since min vi > 0. Furthermore, µi,t(B) a.s.−→ 0 for i ∈ I1.
The proof is complete.
i=1
Proof: Integrating both sides of the recursion (2):
APPENDIX E
PROOF OF LEMMA 5
Unrolling (20) over R steps:
Pt+1(B) = AtPt(B) + dt+1(B)
(20)
Pt+R(B) = At+R−1 ··· AtPt(B)
+
At+R−1 ··· At+R−kdt+R−k(B)
R−1(cid:88)
k=0
Since the product of stochastic matrices is also a stochastic matrix, Proposition 1 implies:
Vt+R(B)
= max
≤ τ1(At+R−1 ··· At)Vt(B)
Pi,t+R(B) − min
i
i
Pi,t+R(B)
(cid:32)
R−1(cid:88)
k=0
+ max
i,j
(cid:32)
R−1(cid:88)
k=0
+
max
i
[At+R−1 ··· At+R−kdt+R−k(B)]i
(cid:33)
(cid:33)
− [At+R−1 ··· At+R−kdt+R−k(B)]j
≤ τ1(At+R−1 ··· At)Vt(B)
[At+R−1 ··· At+R−kdt+R−k(B)]i
− min
[At+R−1 ··· At+R−kdt+R−k(B)]i
≤ τ1(At+R−1 ··· At)Vt(B)
i
R−1(cid:88)
(cid:16)
+
di,t+R−k(B) − min
i
max
i
(cid:17)
di,t+R−k(B)
k=0
Proof: From Lemma 5, we obtain:
APPENDIX F
PROOF OF THEOREM 1
14
(21)
(22)
We next show that the remainder is asymptotically negligible-i.e.,
Re-writing the remainder term:
(cid:34)
R−1(cid:88)
(cid:34)
k=0
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
E[Vt+R(B)Ft] ≤ E[τ1(At+R−1 ··· At)Ft]Vt(B)
+
E
di,t+R−k(B) − min
i
max
i
di,t+R−k(B)
k=0
R−1(cid:88)
(cid:34)
R−1(cid:88)
R−1(cid:88)
k=0
E
i
E
=
k=0
max
(cid:34)
E(cid:104)
E(cid:104)
E
di,t+R−k(B) − min
i
max
i
di,t+R−k(B)
→ 0.
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
(cid:12)(cid:12)(cid:12)Ft+R−k−1
(cid:12)(cid:12)(cid:12)Ft+R−k−1
(cid:105)
(cid:35)
(cid:105)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft
di,t+R−k(B) − min
di,t+R−k(B)
i
di,t+R−k(B) − min
i
max
i
di,t+R−k(B)
max
≤ 4 max
di,t+R−k(B) − min
di,t+R−k(B)
{αi(1 − 2i)µi,t+R−k−1(B)}
i
i
In order to show (22), it suffices to show:
(23)
Note that if (23) holds, then Lemma 4 implies that µi,t+R−k−1(B) a.s.→ 0 as t + R − k − 1 → ∞, and as a result,
(22) follows. Thus, we next focus on proving the bound (23).
i
Using Proposition 2, we obtain for any γ > 0:
di,t+R−k(B) − min
di,t+R−k(B)
i
log
exp(γdl,t+R−k(B))
+ log
exp(−γdl,t+R−k(B))
E(cid:104)
(cid:34)
max
i
E
(cid:34)
≤ 1
γ
≤ 1
γ
l=1
(cid:32) M(cid:88)
(cid:32) M(cid:88)
(cid:32) M(cid:88)
(cid:32) M(cid:88)
l=1
l=1
(cid:12)(cid:12)(cid:12)Ft+R−k−1
(cid:105)
(cid:33)
(cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft+R−k−1
(cid:33)
(cid:35)
(cid:33)(cid:35)
log
E[exp(γdl,t+R−k(B))Ft+R−k−1]
+ log
E[exp(−γdl,t+R−k(B))Ft+R−k−1]
15
(24)
where we used Jensen's inequality and the linearity of expectation.
Next, note that the (conditional) moment generating functions of the innovation terms can be written as a weighted
average of hyperbolic cosines:
l=1
E[eβdl,t+R−k(B)Ft+R−k−1]
M(cid:88)
=
qiPi,j cosh (βαi(1 − 2i)[ei + ej]lµi,t+R−k−1(B))
(25)
for any β ∈ R. To prove (25), set t(cid:48) = t + R − k − 1, and proceed similarly as in the proof of Lemma 3:
i,j=1
E[eγdl,t(cid:48)+1(B)Ft(cid:48)]
= Ei,j[E[eγdl,t(cid:48)+1(B)Ft(cid:48), (i, j)]Ft(cid:48)]
=
B[Di→j(x)pt(cid:48) (x)]ldx(cid:105)
qiPi,jE(cid:104)
eγ(cid:82)
=
qiPi,j cosh (γαi(1 − 2i)[ei + ej]lµi,t(cid:48)(B))
Plugging (25) into (24) and using Proposition 2 again, we obtain:
E
max
i
di,t(cid:48)+1(B) − min
i
di,t(cid:48)+1(B)
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Ft(cid:48)
(cid:88)
(cid:88)
i,j
i,j
l=1
M(cid:88)
M(cid:88)
(cid:32) M(cid:88)
(cid:32) M(cid:88)
i=1
i=1
i=1
(cid:88)
(cid:88)
i,j
i,j
(cid:34)
≤ 2
γ
≤ 2
γ
≤ 2
γ
≤ 2
γ
≤ 4
log
log
log
log
(cid:18)
(cid:33)
qiPi,j cosh (γαi(1 − 2i)[ei + ej]lµi,t(cid:48)(B))
qiPi,j exp (γαi(1 − 2i)[ei + ej]lµi,t(cid:48)(B))
max
i,j
(cid:16)
exp (γαi(1 − 2i)[ei + ej]lµi,t(cid:48)(B))
{αi(1 − 2i)µi,t(cid:48)(B)}(cid:17)(cid:33)
exp
2γ max
i
(cid:19)
γ
We have now proven (23) by taking γ → ∞ to tighten the bound.
i
max
{αi(1 − 2i)µi,t(cid:48)(B)} +
log M
Plugging (23) into (21):
Note that from Assumption 3, it follows that E[τ1(At+R−1 ··· At)Ft] ≤ 1 − < 1 for some ∈ (0, 1).
Lemma 4 implies that µi,t+R−k−1(B) a.s.→ 0, for all i ∈ I1. Note that here we used the positivity of the weights
αi along with the fact that i < 1/2. Define the non-negative sequence
E[Vt+R(B)Ft] ≤ E[τ1(At+R−1 ··· At)Ft]Vt(B)
+ 4
{αi(1 − 2i)µi,t+R−k−1(B)}
max
i
R−1(cid:88)
k=0
R−1(cid:88)
k=0
δ(R)
t
def
= 4
{αi(1 − 2i)µi,t+R−k−1(B)}
max
i
16
(26)
The above implies δ(R)
t
a.s.→ 0 as t → ∞. Taking the unconditional expectation of both sides in (26):
E[Vt+R(B)] ≤ (1 − )E[Vt(B)] + E[δ(R)
t
]
(27)
where E[δ(R)
t
] → 0 by the dominated convergence theorem. Using induction on (27), we obtain for all t = kR:
E[Vt(B)] ≤ (1 − )t/RE[V0(B)] +
(1 − )lE[δ(R)
t−(l+1)R]
Taking limits of both sides and using E[V0(B)] < ∞:
E[VkR(B)] ≤
lim sup
k→∞
E[V0(B)]
(1 − )lE[δ(R)
(k−1−l)R] = 0
t/R−1(cid:88)
(cid:19)
l=0
(cid:18)
lim
k→∞(1 − )k
k−1(cid:88)
+ lim
k→∞
l=0
Thus, the subsequence E[VkR(B)] converges to zero since Vt(B) ≥ 0. Since E[Vt+1(B)] ≤ E[Vt(B)] + E[δ(1)
all t ∈ N and E[δ(1)
further implies Vt(B)
] for
] → 0, it follows that the whole sequence E[Vt(B)] converges to zero. Markov's inequality
i.p.→ 0. The proof is complete.
t
t
APPENDIX G
Proof: Evaluating the density update at x = X∗:
PROOF OF LEMMA 6
(cid:88)
pt+1(X∗) = Atpt(X∗) + Dt(X∗)pt(X∗)
=
+ [(1 − αit)pjt,t(X∗) + αit2P (Yit,t+1Zit,t)pit,t(X∗)]
pl,t(X∗)el
l(cid:54)=it,jt
× (eit + ejt)
where Zi,t = I(X∗ ∈ Ai,t) is the query input to the noisy channel and P (Yi,t+1Zi,t) models the binary symmetric
channel for the ith agent. Taking the logarithm of both sides and using Jensen's inequality, we obtain for a
collaborating agent pair (it, jt):
log pit,t+1(X∗) = log[(1 − αit)pjt,t(X∗)
+ αit(2P (Yit,t+1Zit,t))pit(X∗)]
≥ (1 − αit) log pjt,t(X∗) + αit log pit,t(X∗)
+ αit log(2P (Yit,t+1Zit,t))
Writing this in vector form with the understanding that the logarithm of a vector is taken component-wise:
log pt+1(X∗) (cid:23) At log pt(X∗) + α log(2P (YtZt))(eit + ejt)
17
(28)
where we defined the response and query variables at time t as Yt = Yit,t+1 and Zt = Zit,t.
def
= At2At2−1 ··· At1 for t1 ≤ t2, and Φt1:t2 = I for t1 > t2. Using induction
Define the matrix product Φt1:t2
on (28), we obtain:
log pt+R(X∗) (cid:23) Φt:t+R−1 log pt(X∗)
R(cid:88)
+ α
Φt+R+1−k:t+R−1(eit+R−k + ejt+R−k)
k=1
× log(2P (Yt+R−kZt+R−k))
Let c (cid:31) 0 be the left-eigenvector of Φ = Φt:t+R−1 (cf. Assumption 3). Define the positive constants c
and αR
= mini (min{αi, 1 − αi})R. Left-multiplying (29) by cT :
cT log pt+R(X∗) ≥ cT log pt(X∗)
def
R(cid:88)
+ α · c · αR
log(2P (Yt+R−kZt+R−k))
(29)
def
= mini ci,
where we used the bound cT Φt+R+1−k:t+R−1ei ≥ c · αR,∀i, for k = 1, . . . , R. By induction, we have for m ∈ N:
k=1
cT log pmR(X∗) ≥ cT log p0(X∗)
m(cid:88)
R(cid:88)
+ α · c · αR
log(2P (Y(m−l)R−kZ(m−l)R−k))
l=1
k=1
Then, using the strong law of large numbers (LLN):
cT log pmR(X∗)
R(cid:88)
m(cid:88)
lim inf
m→∞
≥ α · c · αR lim
m→∞
1
m
log(2P (Y(m−l)R−kZ(m−l)R−k))
(cid:35)
l=1
k=1
log(2P (YR−kZR−k))
1
m
(cid:34) R(cid:88)
(cid:18)
k=1
(cid:19)
max
i∈I1
i
=: K > 0
= α · c · αR · E
≥ α · c · αR · C
where C() is the channel capacity of a BSC with crossover probability .
REFERENCES
[1] T. Tsiligkaridis, B. M. Sadler, and A. O. Hero, "Collaborative 20 Questions for Target Localization," IEEE Transactions on Information
Theory, vol. 60, no. 4, pp. 2233 -- 2252, April 2014.
[2] -- -- , "On Decentralized Estimation with Active Queries," IEEE Transactions on Signal Processing, vol. 63, no. 10, pp. 2610 -- 2622,
[3] R. Sznitman and B. Jedynak, "Active testing for face detection and localization," IEEE Transactions on Pattern Analysis and Machine
May 2015.
Intelligence, vol. 32, no. 10, pp. 1914 -- 1920, 2010.
and Machine Intelligence, vol. 18, no. 1, January 1996.
[4] D. Geman and B. Jedynak, "An Active Testing model for Tracking roads in Satellite Images," IEEE Transactions on Pattern Analysis
[5] R. Castro and R. Nowak, "Active learning and sampling," in Foundations and Applications of Sensor Management. Springer, 2007.
[6] Y.-C. Wu, Q. Chaudhari, and E. Serpedin, "Clock Synchronization of Wireless Sensor Networks," IEEE Signal Processing Magazine,
vol. 28, no. 1, pp. 124 -- 138, January 2011.
[7] W. Xia and M. Cao, "Sarymsakov Matrices and Asynchronous Implementation of Distributed Coordination Algorithms," IEEE
Transactions on Automatic Control, vol. 59, no. 8, pp. 2228 -- 2233, August 2014.
[8] A. Renyi, "On a problem of information theory," MTA Mat. Kut. Int. Kozl. B, vol. 6, pp. 505 -- 516, 1961.
18
[9] S. Ulam, Adventures of a Mathematician. New York: Charles Scribners Sons, 1976.
[10] M. V. Burnashev and K. S. Zigangirov, "An interval estimation problem for controlled observations," Problems in Information
[11] M. Horstein, "Sequential Transmission using Noiseless Feedback," IEEE Transactions on Information Theory, vol. 9, no. 3, pp. 136 -- 143,
[12] O. Shayevitz and M. Feder, "Communication with Feedback via Posterior Matching," in IEEE International Symposium on Information
[13] -- -- , "Optimal Feedback Communication via Posterior Matching," IEEE Transactions on Information Theory, vol. 57, no. 3, pp. 1186
[14] B. Jedynak, P. I. Frazier, and R. Sznitman, "Twenty questions with noise: Bayes optimal policies for entropy loss," Journal of Applied
[15] R. Waeber, P. I. Frazier, and S. G. Henderson, "Bisection search with noisy responses," SIAM Journal of Control and Optimization,
[16] A. Jadbabaie, P. Molavi, A. Sandroni, and A. Tahbaz-Salehi, "Non-bayesian Social Learning," Games and Economic Behavior, vol. 76,
Transmission, vol. 10, pp. 223 -- 231, 1974.
July 1963.
Theory, June 2007, pp. 391 -- 395.
-- 1222, March 2011.
Probability, vol. 49, pp. 114 -- 136, 2012.
vol. 53, no. 3, pp. 2261 -- 2279, 2013.
pp. 210 -- 225, 2012.
[17] P. Molavi, A. Jadbabaie, K. R. Rad, and A. Tahbaz-Salehi, "Reaching Consensus with Iincreasing Information," IEEE Journal of
Selected Topics in Signal Processing, vol. 7, no. 2, pp. 358 -- 369, April 2013.
[18] S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah, "Randomized gossip algorithms," IEEE Transactions on Information Theory, vol. 52,
no. 6, pp. 2508 -- 2530, June 2006.
[19] T. D. Cover and J. A. Thomas, Elements of Information Theory. Wiley, 2006.
[20] J. N. Tsitsiklis, D. P. Bertsekas, and M. Athans, "Distributed asynchronous deterministic and stochastic gradient optimization algorithms,"
IEEE Transactions on Automatic Control, vol. 31, no. 9, pp. 803 -- 812, Sept. 1986.
[21] V. D. Blondel, J. M. Hendrickx, A. Olshevsky, and J. N. Tsitsiklis, "Convergence in multiagent coordination, consensus, and flocking,"
in Proc. Joint 44th IEEE Conf. Decisions Contr. Eur. Contr. Conf., December 2005, pp. 2996 -- 3000.
[22] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multiagent optimization," IEEE Trans. Autom. Control, vol. 54, no. 1,
[23] A. I. Chen and A. E. Ozdaglar, "A fast distributed proximal-gradient method," in Proc. Allerton Conf. Commun., Contr., Comput.,
pp. 48 -- 61, January 2009.
October 2012, pp. 601 -- 608.
[24] S. Patterson, Y. C. Eldar, and I. Keidar, "Distributed Compressed Sensing for Static and Time-Varying Networks," IEEE Trans. Signal
Processing, vol. 62, no. 19, pp. 4931 -- 4946, October 2014.
[25] E. Seneta, Non-negative Matrices and Markov Chains, 2nd ed. New York: Springer, 1981.
[26] I. C. F. Ipsen and T. M. Selee, "Ergodicity Coefficients defined by Vector Norms," SIAM J. Matrix Anal. Appl., vol. 32, no. 1, pp.
153 -- 200, 2011.
[27] M. Chen, S. C. Liew, Z. Shao, and C. Kai, "Markov Approximation for Combinatorial Network Optimization," IEEE Transactions on
Information Theory, vol. 59, no. 10, October 2013.
[28] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge University Press, 2004.
[29] P. Gupta and P. R. Kumar, "The capacity of wireless networks," IEEE Trans. Inf. Theory, vol. 46, no. 2, pp. 388 -- 404, March 2000.
[30] A. Berman and R. J. Plemmons, Nonnegative matrices in the Mathematical Sciences. Academic Press, New York, 1979.
[31] P. Billingsley, Probability and Measure.
[32] R. Durrett, Probability: Theory and Examples, 3rd ed. Duxbury Press, Belmont, CA, 2005.
John Wiley & Sons, Inc., Hoboken, NJ, 2012.
|
1706.00968 | 1 | 1706 | 2017-06-03T15:24:28 | Precious Time: Understanding Social Stratification in the Knowledge Society Through Time Allocation | [
"cs.MA"
] | The efficient use of available resources is a key factor in achieving success on both personal and organizational levels. One of the crucial resources in knowledge economy is time. The ability to force others to adapt to our schedule even if it harms their efficiency can be seen as an outcome of social stratification. The principal objective of this paper is to use time allocation to model and study the global efficiency of social stratification, and to reveal whether hierarchy is an emergent property. A multi-agent model with an evolving social network is used to verify our hypotheses. The network's evolution is driven by the intensity of inter-agent communications, and the communications as such depend on the preferences and time resources of the communicating agents. The entire system is to be perceived as a metaphor of a social network of people regularly filling out agenda for their meetings for a period of time. The overall efficiency of the network of those scheduling agents is measured by the average utilization of the agent's preferences to speak on specific subjects. The simulation results shed light on the effects of different scheduling methods, resource availabilities, and network evolution mechanisms on communication system efficiency. The non-stratified systems show better long-term efficiency. Moreover, in the long term hierarchy disappears in overwhelming majority of cases. Some exceptions are observed for cases where privileges are granted on the basis of node degree weighted by relationship intensities but only in the short term. | cs.MA | cs |
Precious Time: Understanding Social Stratification in the
Knowledge Society Through Time Allocation
Michal Kakol, Radoslaw Nielek, Adam Wierzbicki
October 17, 2018
Abstract
The efficient use of available resources is a key factor in achieving success on both personal
and organizational levels. One of the crucial resources in knowledge economy is time. The
ability to force others to adapt to our schedule even if it harms their efficiency can be seen
as an outcome of social stratification. The principal objective of this paper is to use time
allocation to model and study the global efficiency of social stratification, and to reveal whether
hierarchy is an emergent property. A multi-agent model with an evolving social network is
used to verify our hypotheses. The network's evolution is driven by the intensity of inter-
agent communications, and the communications as such depend on the preferences and time
resources of the communicating agents. The entire system is to be perceived as a metaphor of
a social network of people regularly filling out agenda for their meetings for a period of time.
The overall efficiency of the network of those scheduling agents is measured by the average
utilization of the agent's preferences to speak on specific subjects. The simulation results
shed light on the effects of different scheduling methods, resource availabilities, and network
evolution mechanisms on communication system efficiency. The non-stratified systems show
better long-term efficiency. Moreover, in the long term hierarchy disappears in overwhelming
majority of cases. Some exceptions are observed for cases where privileges are granted on the
basis of node degree weighted by relationship intensities but only in the short term.
Introduction
1
Although most of us probably tire of endless meetings, in reality, a substantial part of added value
produced by knowledge workers is somehow related to meetings. Passing or gathering information,
evaluating progress of team members, assigning tasks, selling ideas or asking people for help -- such
activities require not only our time but also synchronization of people participating in meetings.
The question as to who should synchronize with whom is determined by social hierarchy. If someone
wants to sell you a new insurance policy, you can pick the meeting place. On the other hand, if
you want to convince your boss or supervisor to allocate some resources for your project, you must
adjust to his or her daily plans.
Some signs of hierarchy are visible in nearly all communities. Existing exceptions, such as
Anonymous, hippie communes, or many anarchists, have taken their lack of hierarchy as their
hallmark. A variety of processes can influence the order of hierarchy in a particular community.
Sometimes, especially in companies, a formal title held by a person is translated almost directly
into a position in the hierarchy.
In other cases, individual trails or particular skills are more
important. Sometimes, hierarchy is related to race, skin color, place of birth, or wealth.
Hierarchy in a society results from a stratification system. Its components, e.g., social processes
and institutions, generate inequalities in control over resources among its members [Gru94]. The
privileged enjoy a disproportionate share of valued resources, thereby constituting a hierarchy
based on social traits. Eventually, as a feedback loop, the hierarchy recreates the inequalities.
In this article, we focus on the existence of hierarchies in the knowledge society, where time and
exchange of ideas during (virtual or real, but equally time-consuming) meetings are a basis of work
and added value creation. In its most developed form, the knowledge society is exemplified by the
Silicon Valley. The motto of this paper is a quote from Vivek Wadhwa, a well-known American
technology entrepreneur and academic. The quote inspires several of our hypotheses, which are
related to a decreased effectiveness and stability of hierarchies in the modern knowledge society.
The relationship between time and social position in the knowledge society seems evident, but
is still not sufficiently understood. Broadly speaking, agents' behaviors and privileges influence
1
their ability to manage time (described in detail in the Model section). Insight into the use of
time, a key economic component affected by social stratification, can have practical implications.
Productivity of a company can be determined largely by rigidness of organizational hierarchy and
enforced strategies of managing workers' time. This is especially true for companies profiting from
the knowledge economy. As a matter of fact, decreasing hierarchies and reducing constraints on
individual time management has been advocated as a successful recipe for knowledge companies
[TW08].
Our approach assumes that agents do not care about other agents' utilities or performance on
the societal level. Agents are greedy, and attempt to maximize only their own utility. However,
we are concerned with the global efficiency of a social system. Measuring the efficiency of social
systems is a difficult task that can be addressed from various perspectives. In our research, we
use a social simulation approach that enables us to investigate the emergent effect of individual
agent's strategies on the global welfare.
The following hypotheses, regarding our model of the knowledge society, are tested in this
article:
• H1: Egalitarian knowledge societies and communities are more efficient;
• H2: Application of advanced procedures of time allocation (more intelligent algorithms) on
the agent level helps to overcome the influence of stratification in knowledge society;
• H3: Social stratification is an emergent property, but does not persist in the long term in
stable knowledge societies.
In section Related works, we present related work in sociology and, specifically, in stratification,
including related work on its modeling. Later sections cover the description of our multi-agent
model (section Model), the measures used to evaluate it (section Measures), and the results in the
order of tested hypotheses (section Results). Finally, we conclude by discussing the results and
future work in Discussion.
Related works
It is widely agreed that differentiation of social groups into hierarchical layers is present in all
developed societies. It is a worldwide phenomenon and except for utopian visions, no fully egal-
itarian social system has yet been found to work. However, since the underlying reasons can be
disputed or interpreted through the lens of ideology, we venture to relate a couple of examples.
These inequalities are so deeply embedded in our culture that not only do the privileged legitimize
the mistreatment of some groups, but paradoxically the mistreated also accept and justify this
mistreatment. [JBN04] Parents show us our position in society and teach us to expect that same
level. A study by [Hur11] has shown that legacy students (whose parents attended a particular
college) are more likely to be admitted to elite colleges. We expect this, and we obtain it, according
to the saying that "the rich get richer." [MBH+09] found that today's technological advances make
it easier to transfer wealth, and thus also the associated status, between generations.
Are inequalities that bad after all? The evidence is mixed and often ideologically underpinned.
[WPC09], with a self-explaining title, "Why more equal societies almost always do better," linked
inequality with various social problems (e.g., obesity, homicide, and higher infant mortality). De-
spite the existence of sound evidence confirming that link, there is an ongoing discussion regarding
the study's methodology. [Sno10] [Sau10] In contrast to the mentioned findings, [Row11] attempted
to prove that no correlation exists between income equality and social problems. [Atk96] went even
further and claimed that inequality fuels economic growth. Although the issue of social stratifi-
cation and inequality has been a subject of extensive research and debate over many years, the
relation between inequality and prosperity still remains unsettled.
Stratification is most often considered in terms of economic inequality. Nevertheless, let us
consider a hypothetical example in which the president of the United States finds an exceptionally
urgent matter to discuss with the author of this article. Even under such conditions, the invited
author would obviously need to adapt to the president's schedule. This example depicts well the
obvious link between time availability and social status.
Time usage and its availability can also be a consequence of stratification -- those on the bottom
of a hierarchy have less control over how they use their time and must adapt to those at the top
2
of the ladder. One can argue that there is hardly any relation, but since economics in general is
the study of how agents allocate scarce resources, time allocation is of economic interest, as it is
one of the key factors in understanding economic phenomena. Moreover, time is a component of
nearly every economic undertaking. It is a part of market labor and a key input to consumption, as
consumed commodities are produced with it, as much as market goods and production technology
and the consumer's time [Bec65]. The classic literature on time allocation in the United States can
be found in [GB75] and [JS85], which examined trends of time allocation based on time diaries,
and an extension of it can be found in [RG10]. Moreover, [AH06, AH08] reported an interesting
link between time usage patterns and socio-economic status according to which, over the last 20
years, less educated men increased the time they allocated to leisure, while more educated men
reported a decrease.
To the authors' best knowledge, there are several computational models of societies incorpo-
rating hierarchy that focus more or less directly on stratification.
[Sma99] proposed a model of
Tongan society that simulated the relationship of warfare and traditional Tongan rules of mar-
riage. Another model of Pacific island societies, proposed by [You10], modeled the relationship
between limited resource usage and different leadership models.
[RCME07] introduced a simu-
lation of promotion competition between male and female workers in a corporate environment,
specifically modeling the disparity of female representation at the highest corporate management
level. Although the proposed simulation included the temporal dimension, it modeled the problem
on a company level, and its resolution could not cover workers' social relations or schedules.
[FS07] described a model of time-use behavior. The model was fed with an empirical dataset
gathered in a time-use diaries survey, but ignored the social aspect of time-allocation and the
role of interaction between agents. Models in which interactions between agents are used for time
allocation have been researched by [WNK09], but for different purposes, namely emergence of
time-usage patterns and infection spreading [NWK08]. [SW12] also introduced a simulation of the
emergence of complex social networks based on the intensity of relations and provided an overview
of similar works, i.e., stratification into layers on an individual level by intensity of relations. The
approach presented in the aforementioned paper has been implemented in our model.
Model
The problem we have modeled consists of two components, the relation between the agents and
the communication occurring between them, making it not only a social network but also a com-
munication system. Formally, we move the problem to a social space with a defined topology, with
points representing agents and metrics of their distance, e.g., their relation strength (i.e. social
network). The relations of the actors can potentially be represented as vectors and the network
as a summation of these vectors, i.e., as a two-dimensional matrix. Such matrices in [Klu00]
are called socio-matrices or Moreno matrices; moreover, [Ley94] introduces an analogical formula-
tion for communication systems. Communication is modeled by an algorithm that searches for a
communication partner based on the agent relation topology.
The model described in this paper is a multi-agent system implemented in NetLogo and based
purely on a social network approach (no spatial dimension). The agents are embedded in a network
of relations between them, communicating with each other to fulfill their need to "talk" on various
subjects. The flowchart in Fig. 1 depicts a single model step in which the agents are executed
one after another (in a particular order) trying to fill in their "weekly" schedules. The details
on each component of the model step are presented in the following sections. Note here that
the order in which agents fill in their schedules may be fixed at the beginning of the simulation,
during model initialization (for the scenarios that model a strict hierarchy) or may be determined
in the beginning of each model step (for scenarios that model an egalitarian or mobile approach
to scheduling agent meetings).
The size of the agent population is constant over the simulation. Removing or introducing
new agents is not implemented in this model, and thus a single simulation can be referred to as
presenting the population in a short time frame, sufficiently short to say that the population does
not change. A single agent corresponds to a person, and the edge between agents represents a
symmetrical acquaintanceship -- both agents are linked with equal strength, and therefore both are
equally capable of arranging a meeting.
According to the social brain hypothesis and the findings of Dunbar, human relationships are
characterized by different levels of intensity [Dun96, Dun98], which in the model is a link's strength
3
attribute. The strength of the relationship between agents depends on the frequency of interactions
between them and has been modeled according to the study of [Hay89] - -- agents communicate
only with the agents they are directly connected to, the unused connections get weaker until they
break up, and those frequently used are strengthened by analogy. The procedure of edge strength
modification is described in detail in the Network evolution section.
On the population level, there is a finite set of all possible topics/interests parameterized in the
model as culture. Every agent has a set of interests, a subset (equal size for all the agents) drawn
from the culture with a probability following a normal distribution. Therefore, some subjects
are quite rare in the population (e.g., Persian poetry) and some very popular (e.g., football).
Additionally, every agent possesses a willingness attribute for each subject, which is assigned
randomly from zero to one with a uniform distribution for every interest, thus representing an
agent's inclination for a given subject. An agent's willingness expresses any individual agents
creativity and ability to contribute new ideas on a topic. Finally, every agent is equipped with the
same number of time slots to arrange its meetings in.
The source code of the presented model, as well as a comprehensive list of model parameters,
together with additional resources, can be found on the model's webpage http://users.pjwstk.
edu.pl/~s8811/papers/stratification/.
Meeting scheduling procedures
The modeled agents' principal objective is a kind of rational choice, namely, to fulfill maximally
their need to communicate. This need, i.e., "willingness," is an attribute of each subject in a set of
interests of a particular agent and is reduced each time an interaction between neighboring agents
sharing the same interest occurs. The payoff of their actions is the sum of the used instances of
willingness, which they try to maximize. At every turn for every agent, a scheduling procedure
is invoked that is responsible for finding partners for meetings and filling agents' time slots. A
meeting occurs when two agents agree on a particular common free time slot and subject. Every
scheduled meeting consumes a time slot and willingness to discuss a particular subject. Letting
s1 be willingness to discuss a subject 1 for agents A and B, respectively, the remaining
W A
willingness for agent A after the meeting can be calculated using the following formula:
s1 and W B
(cid:16)
(cid:17)
W (cid:48)A
s1 = W (cid:48)A
s1 − min
W (cid:48)A
s1, W (cid:48)B
s1
(1)
Figure 1: Simulation loop elements.
4
InitilizemodelMaximumstepreached?Executea modelstepFill in agents'schedulesReport model and agentsstatusNoStopYesUpdateedge weightsAdd new/remove Expired edgesModifynetworkbased on informationwhether vertices arrangeda meeting in this step Notice that the result of eq. 1 for one of the two agents participating in the meeting will always
equal zero, thereby excluding this agent as a communication partner for other agents in the current
step. A meeting consumes a single time slot, with a length that is limited and might be interpreted
as days of the week planned in advance by each agent.
The egoistic goal of the agent is to maximize its own willingness usage; hence, we use an average
value of used willingness as an overall measure of network efficiency in the model, i.e., the extent
to which the agents are utilizing their preferences. The average willingness usage expresses the
global utilization of ideas generated by all agents in a knowledge society or community. Maximizing
average willingness usage is a combinatory problem and can be solved in many ways, but finding
an optimal solution requires algorithms with exponential complexity. As users have only limited
resources, they must rely only on approximation algorithms (heuristics). Two heuristics have been
implemented and tested:
(cid:1); every agent repeats this
s1, W B
s1
1. Simple -- agent A chooses a random agent B from its acquaintances and if they share a free
time slot, selects a subject S, in order to maximize min(cid:0)W A
procedure a limited number of times (model's parameter),
2. Intelligent -- agent A iterates over the list of its' subjects sorted in descending order by
willingness to talk; for each of the iterated subjects all days of the week are used to find a
partner B with maximal willingness to talk; this heuristic can be summarized as "trying until
success".
For a detailed flowchart, see Appendix I.
Stratifying the system
The stratification in the proposed model manifests in three agent initialization orders followed by
schedule filling, simulating different levels of inequality and rigidity as defined in [Gru94].
1. Egalitarian -- The baseline used in the model is a random order of executing the agents
and is treated as a metaphor of an egalitarian system with equal opportunities to make
appointments; the inequalities and rigidity in such a system are low.
2. Hierarchical -- In the hierarchical scheduling method, each agent in the system has a fixed
place in the hierarchy, and thus the agents are executed in the same order throughout the
entire model run, at every turn. This can be referred to as the highest level of inequality and
rigidity in the system.
3. Mobile -- In the mobile method, differentiation of the agents occurs at every turn. At each
turn, the agents (nodes) are ordered according to their network degrees weighted by their
relations' strengths, resulting in a medium level of inequality and rigidity in the system. A
node's degree is interpreted (in accordance with much research in social network theory) as
the nodes social capital or influence, which determines the node's ability to schedule meetings
ahead of less influencial nodes.
Network evolution
In the initial network, relations between the agents are generated with a given density (we arbi-
trarily used settings of 0.2% or 0.6%). However, the network changes based on the outcome of
interactions between agents. The evolution of the network is shaped by three procedures that are
invoked after every turn:
1. Edge strength updating -- basically, this procedure looks up all edges in the model and
updates their strengths according to the activity at the last turn, namely, the strength of every
edge that has been used for a successful meeting arrangement is increased and the strengths of
unused edges are decreased; two separate updating functions have been implemented, linear
and logarithmic, binding the pace of relation change with their strength,
2. Edge creation -- according to the simulation parameters, new edges are created using two
heuristics, triadic closure ("my friend's friend") [Oys00] approach, in which there is some
probability that a new edge appears between two agents that have at least one common
5
friend, and the "random meeting" approach, which is simulated by the creation of an edge
between two random agents in the population. However, an additional parameter in the
model was added controlling the choice of a new neighbor. By default, the neighbors are
chosen randomly with equal probabilities, but a preferential mode can be activated to choose
more likely new neighbors of a greater network degree, i.e., preferring more popular agents.
3. Edge removing -- edges that are not used sufficiently frequently (i.e., their strength reaches
zero) are removed.
Measures
All measurements were taken after each turn, as all the agents' schedules were filled and the network
had changed accordingly. Along with the well-known network metrics (e.g., network density and
average degree), other measures were also used - measures of ranking comparison and two measures
expressing efficiency of different stratification procedures and heuristics for arranging meetings (all
normalized to one).
Schedule usage represents the percentage of all days in the schedules used in one turn and is
calculated using eq. 2,
ScheduleU sage =
i=1 Ti
N ∗ weekdays
where N is the number of agents, and Ti is the number of used time slots of the i-th agent.
Willingness usage represents the percentage of "consumed" willingness of all agents in one
turn in relation to all available willingness and is calculated using eq. 3,
W illingnessU sage =
where N is the number of agents, SUi is the sum of used willingness points in the i-th agent's
schedule, and Si is the sum of willingness points of all subjects in the ith agent's interest set.
Kendall Tau distance
To measure the differences between agents' execution order lists for consecutive simulation steps,
we arbitrarily decided to use the Kendall Tau distance. This distance is a measure between two
full rankings (over the same set of n elements) that counts the number of pairwise disagreements
between those two rankings [BFB09]. We used the normalized Kendall Tau distance, i.e., multiplied
by the maximum value of the metric: n(n−2)
.
(cid:88)
{i,j}∈P
K (σ1, σ2) =
K i,j (σ1, σ2)
KN (σ1, σ2) =
2
n (n − 2)
K (σ1, σ2)
n
(cid:40)
where σ is a list or ranking, n is the number of elements in the list, P is a set of unordered
pairs of distinct list elements, and K i,j (σ1, σ2) is a comparison of single elements i and j in both
compared lists:
K i,j (σ1, σ2) =
0 ⇔ i, j
1 ⇔ i, j
are in the same order
are in reverse order
(6)
Order list cardinality We introduced this measure to monitor the stability of the set of
privileged agents in consecutive model steps, i.e., the agents executed first. Precisely, we used the
top 100 agents from the order list representing the execution order of all 1,000 agents. To measure
the differences between those incomplete rankings, we used the simplest measure available.
In
favor of distance measures for partial rankings [BFB09], we used the cardinality of the intersection
of the partial rankings, thus monitoring the number of common agents remaining in the "elite" in
consecutive steps.
σ1 ∩ σ2
6
(7)
(cid:80)N
(cid:80)N
(cid:80)N
i=1 SUi
i=1 Si
(2)
(3)
(4)
(5)
Results
Egalitarian societies are more efficient
The results of the 72 parameter combinations (four points in time for each parameter set are
considered, i.e., the 200th step, 300th step, 500th step, and the last 1,000th step) prove that
the Egalitarian execution method, which represents a system with no hierarchy at all, results in
higher willingness usage in the majority of cases. The egalitarian scheduling method achieved the
highest willingness usage mean in 76% of tested runs; see Fig. 2. Most importantly, the pairwise
comparison of all three stratification methods in all of the runs showed statistically significant
differences in the means (p < 0.0001, Bonferroni-Dunn test for multiple comparisons). In 19.1%
of the cases, a single compared pair of the mean values showed no significant difference. However,
these were mostly pairs of Hierarchical and mobile (76.3%). Moreover 74.5% of the cases with an
insignificantly different pair were those for which Egalitarian still showed the best performance in
terms of average willingness usage value.
Obviously, there are cases in which Egalitarian does not show a better performance than strati-
fied mobile and Hierarchical. Apparently, the performance of the egalitarian system improves over
time, as the random execution order shows the greatest willingness usage values in only approx-
imately 60% of cases for the first 300 turns, but 82% for 500 turns, and eventually 100% in the
1,000th turn (details on Fig. 3). This can be interpreted as indicating that random executions
lead to the best long-term performance. Additionally, the advantage of random order in terms
of willingness usage is most common for systems of scarce time resources (five time slots). The
percentage of cases in which Egalitarian shows better efficiency is inversely proportional to the
number of available time slots. Eventually, for 25 timeslots, these cases amount to only 68%.
The advantage of the privileged agents (first to be executed) is clearly visible in the data,
and apparently this effect is common. The agents executed in first place surpass the rest of the
agents in terms of willingness usage. The increase of efficiency on average for those first 10%
agents compared to the remainder 90% of the system amounts to 20.3%, 24.4%, and 32.0% for
Egalitarian, Hierarchical, and mobile, respectively (calculated for all tested runs). This effect exists
for all tested parameter sets, and under these conditions, the agents privileged by execution with
higher priority are in the best position.
Fig. 4 and 5 depict the model willingness usage output for two time-allocation strategies, i.e.,
Intelligent and Simple. For both methods, the random execution Egalitarian shows better model
efficiency in the area of model stability. For Intelligent time allocation, however, the stratified
execution method by weighted degree has an advantage over random execution in some of its first
hundred steps. This effect is temporary and involves only the warm-up phase. What happens
exactly is that for the Intelligent partner finding heuristic, the Mobile method for ordering agents
surpasses the Egalitarian one in terms of the willingness usage value. For a simple conversation
Figure 2: Percentage of cases in which a parameter set results in the greatest Willingness Usage value for
a particular Execution Order by simulation step number.
7
25030050010000%10%20%30%40%50%60%70%80%90%100%NoneDegreeFixedSimulation step # (turn in which value was measured)% cases Mobile Hierarchical Egalitarian arrangement scheme, Egalitarian resulted in the greatest willingness values in the overwhelming
majority of cases. In contrast to what happens when Intelligent arrangements are used, the share
of cases in which Egalitarian is the most efficient drops to roughly 22% for the 250th step. However
in the long term, that is at the 1,000th step, Egalitarian random execution again results in the
best efficiency in 100% of cases.
The introduction of the Intelligent partner finding heuristic might be a metaphor for the digital
revolution. The Intelligent heuristic leading to more sophisticated schemes of arranging meetings
with other agents results in more efficient search for conversation partners. This can be referred
easily to contemporary advances in communication, such as Internet usage, with which we are
eventually going to find a "soul mate" sharing our interests. In light of the simulation results (see
Table 1 and Table 2), introduction of "intelligent" meeting arrangement schemes helps to boost
a stratified system's efficiency. However this effect applies to short-term results only as, in the
long term, i.e., at approximately the 1,000th step, random execution approximating an egalitarian
system shows the best efficiency after all. In other words, application of advanced procedures of
time allocation at the agent level does not overcome the consequences of stratification.
Figure 3: Percentage of cases in which a parameter set results in the greatest Willingness Usage value for
a particular Execution Order by number of time slots.
Figure 4: Willingness usage values for the entire model run with the "Intelligent" partner finding heuristic.
8
515250%10%20%30%40%50%60%70%80%90%100%NoneDegreeFixedNumber of weekdays (timeslots)% cases Mobile Hierarchical EgalitarianSimulation step #01002003004005006007008009001000 MobileHierarchical EgalitarianExecution order0.70.60.50.40.30.20.1Mean WillingnessusageSelected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10005001515Intell.*True0.002 Table 1: Percentage of simulation runs in which particular execution method results in
the highest willingness usage values for particular simulation step. Results for simple
partner finding heuristic.
Step
Egalitarian
Mobile
Hierarchical
100.00% 97.22% 94.44% 100.00%
0.00%
0.00%
Agentset switch
0.00%
5.56%
0.00%
2.78%
500
1000
0.00%
0.00%
300
250
Table 2: Percentage of simulation runs in which particular execution method results
in the highest willingness usage values for particular simulation step. Results for
intelligent partner finding heuristic.
Agentset switch
Step
Egalitarian
Mobile
Hierarchical
250
300
500
1000
22.22% 27.78% 69.44% 100.00%
0.00%
77.78% 72.22% 25.00%
0.00%
5.56%
0.00%
0.00%
Emergence of social stratification
While analyzing the model results, we wanted to know whether such loosening of system rigidity
would lead towards randomness or back to strict order. To determine the answer, each model
performed a dump of the agents' identification number list in the order in which they were executed
in the particular simulation step. We assume that comparing those execution order lists will provide
the answer.
As described in the Measures section, the execution order list comparison involves a)computing
the Kendall Tau distance between consecutive lists, and b)computing the order list cardinality of
two consecutive execution order lists limited to only the first 100 agents. Note that the follow-
ing section focuses on describing the results corresponding to the Mobile execution method only.
Another reason for such a choice is the fact that distributions for the remainder of the execution
methods are constant and can be summed up as follows.
In contrast to the results of the Egalitarian (random ordering) and hierarchical (fixed ordering)
methods, the Kendall Tau distance distribution of Mobile shows more interesting features. Fig. 6
and 7 depict a typical run of the model and the values of the distance measure. The characteristic of
most importance is the sharp decline in the first 200 to 300 steps and the steady climb afterwards.
The plummeting of the Kendall Tau distance values in the warm-up phase of the model run
exists in almost 67% of tested cases. We take it as a sign of the progressive organization of the
manner in which the agents are executed. Literally, the distance between compared consecutive
In other words, each consecutive list of agents'
order lists decreases over the warm-up phase.
Figure 5: Willingness usage values for the entire model run with the "Simple" partner finding heuristic.
9
Simulation step # MobileHierarchical EgalitarianExecution order01002003004005006007008009001000 Selected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10005001515Simple*True0.002Mean Willingnessusage0.60.50.40.30.2 execution order is more similar to the previous one, i.e., is more similarly ordered. After this
decrease, the value of the distance climbs steadily, i.e., the consecutive order lists become less
commonly ordered, and the system loosens its rigidity. This described upward trend exists in
nearly 71% of cases.
Apart from the visual identification of the trends, we measured the linear correlation between
the distance value and the model step number. Tables 3 and 4 present the binned correlation values
for different parameters separately, as well as for all the model runs. The advantage of negative
correlation in the first steps and positive correlation in general over the entire model run is clearly
visible. Also depicted in Figure 8 is where the correlation value distribution is compared for the
initial warm-up model phase and all 1,000 steps.
An interesting pattern in changes of distance values can be observed for different numbers of
available time slots. The discussed decrease in the initial model phase is more prominent when
more time slots are available. For 25 weekdays, 100% of cases are affected by the initial decrease,
compared to only 12.5% for five weekdays. A similar relation between number of weekdays and the
cardinality of the intersection of two consecutive agent execution order lists can also be observed.
The intersections of the consecutive order lists (first 100 agents only) are measured in order
to examine whether the set of 10% of agents to be executed first, the "privileged" caste, remains
constant over the entire model run. The intersection cardinality value indicates the number of
Figure 6: Distribution of Kendall Tau distance for consecutive model steps for a sample set of parameters
and Mobile execution only, raw data.
Figure 7: Distribution of Kendall Tau distance for consecutive model steps for a sample set of parameters
and Mobile execution only, moving average smoothing.
10
Simulation step #Kendall Tau distanceSelected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10002501515Intell.MobileFalse0.002200040060080010000.500.480.460.440.42Simulation step #Kendall Tau distance(moving average, smoothed)Selected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10002501515IntellMobileFalse0.002200040060080010000.460.450.470.48 ranges of correlation value
Table 3: Summary of the Kendall Tau distance -- step number correlation for different
parameter sets. Percentage of runs for given set of parameters for which correlation
is within one of three ranges. Results for first 100 steps.
(0.1; 1(cid:105)
0.00%
2.78%
4.17%
0.00%
0.00%
0.00%
2.78%
0.00%
2.78%
1.39%
100 steps
(cid:104)−0.1; 0.1(cid:105)
33.33%
30.56%
83.33%
12.50%
0.00%
33.33%
30.56%
30.56%
33.33%
31.94%
(cid:104)−1;−0.1)
66.67%
66.67%
12.50%
87.50%
100.00%
66.67%
66.67%
69.44%
63.89%
66.67%
Number of
weekdays
(timeslots)
New links
algorithm
Preferential
Uniform
Intelligent
Simple
250
500
5
15
25
Culture size
Arrangements
Table 4: Summary of the Kendall Tau distance -- step number correlation for different
parameter sets. Percentage of runs for given set of parameters for which correlation
is within one of three ranges. Results for first 100 steps.
ranges of correlation value
(cid:104)−1;−0.1)
Culture size
Number of
weekdays
(timeslots)
New links
algorithm
Arrangements
250
500
5
15
25
Preferential
Uniform
Intelligent
Simple
2.78%
2.78%
0.00%
0.00%
8.33%
0.00%
5.56%
0.00%
5.56%
2.78%
1000 steps
(cid:104)−0.1; 0.1(cid:105)
25.00%
27.78%
58.33%
12.50%
8.33%
22.22%
30.56%
22.22%
30.56%
26.39%
(0.1; 1(cid:105)
72.22%
69.44%
41.67%
87.50%
83.33%
77.78%
63.89%
77.78%
63.89%
70.83%
11
common agents in two consecutive order lists, e.g., 100: all agents the same in both lists, 70: 70
common agents in two lists. The common apparent effect is that the intersection value steadily
decreases over the modeled 1,000 turns. Similarly to Kendall Tau distance, we measured the
correlation between the intersection value and the model step number. The majority of tested
parameter sets, viz., 72%, reveal negative correlations.
The steady decrease in intersection values (see Fig. 9) is common, and we assume it is a sign
of progressive loosening of the system rigidity for the Mobile execution method. However, there
are also cases in which this decrease speeds up significantly, as is visible in Figure 10. Around step
numbers 300 to 400, the intersection cardinality might plummet. This effect is tightly linked to
the number of available time resources. The more weekdays available, the more cases are affected
with the sharp decline in the last 600 to 700 steps of the model runs, as shown in Tables 5 and 6.
Preferential attachment strengthens the emerging hierarchy
Preferential attachment, as one of the possible heuristics for creating new edges, was implemented
in the model, as it can be expected to make the effect of emergent hierarchy more visible. No
significant differences in willingness usage or overall efficiency were found between model runs with
preferential attachment edge building vs. non-preferential building, but there was one particularly
Figure 8: Distribution of Kendall Tau distance-step number correlation values for the first 100 steps and
an entire model run, for all tested parameter sets in general. Correlation values are on the x-axis.
Figure 9: Intersections cardinality for the entire model run for a sample of parameter sets, slight downward
trend -- rigidity loosening.
12
-0,7-0,6-0,5-0,4-0,3-0,2-0,100,10,20,30,40,50,600,050,10,150,20,250,3100 steps1000 stepsSimulation step #100 60 20 40 0 80Intersection cardinalitySelected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10005001515SimpleMobileFalse0.00220004006008001000 ranges of correlation value
250
500
5
15
25
Table 5:
Summary of the intersection cardinality vs. step number correlation for
different parameter sets. Percentage of runs for given set of parameters for which
correlation is within one of three ranges.
from 500th to 1000th step
(0.1; 1(cid:105)
(cid:104)−1;−0.1)
(cid:104)−0.1; 0.1(cid:105)
5.56%
52.78%
41.67%
52.78%
5.56%
41.67%
4.17%
20.83%
75.00%
4.17%
54.17%
41.67%
8.33%
83.33%
8.33%
50.00%
5.56%
44.44%
5.56%
55.56%
38.89%
5.56%
52.78%
41.67%
5.56%
52.78%
41.67%
62.50%
4.17%
33.33%
8.33%
33.33%
58.33%
33.33%
62.50%
4.17%
52.78%
41.67% 5.56%
Number of
weekdays
(timeslots)
New links
algorithm
Simple
1850
5018
5050
Preferential
Uniform
Intelligent
Culture size
Arrangements
Relations update
interesting relation.
Results for preferential attachment procedure are collected in Table 7. For all model runs with
preferential attachment turned on, the average Kendall Tau distance value amounts to 0.471, while
the model runs with non-preferential attachment during new edge building resulted in distance
average values of 0.476. We can conclude that introduction of preferential attachment to the
model results in bringing more structure to the order in which the agents are being executed.
This is on the condition that we consider only the Mobile execution method, that is, execution in
the order of weighted node degree. We also examined the difference in Kendall Tau distance for
preferential and non-preferential new edge building for each single parameter set, together with
the significance of this difference.
It turns out that in nearly 81% of cases, not only did preferential attachment show lower values
of Kendall Tau distance, but also the difference was statistically significant. Moreover, this scheme
is also linked with the number of time resources available in the model. The percentage of cases
in which preferential attachment results in significantly lower distance grows from 50% to 100%
with the number of weekdays available in the agent's schedule. We conclude that preferential
attachment as a mechanism of building new edges in the graphs not only prompts the emergence
Figure 10:
plummeting in the last model phase on the right.
Intersections cardinality for the entire model run for a sample of parameter sets, sudden
13
Simulation step #Intersection cardinalitySelected modelparameters#agents#culture#timeslots#subjectsPart. searchOrderPref. Attacg.Initial density10002502515Intell.MobileFalse0.00220004006008001000100 60 20 40 0 80 ranges of correlation value
250
500
5
15
25
Table 6:
Summary of the intersection cardinality vs. step number correlation for
different parameter sets. Percentage of runs for given set of parameters for which
correlation is within one of three ranges.
from 500th to 1000th step
(0.1; 1(cid:105)
(cid:104)−1;−0.1)
(cid:104)−0.1; 0.1(cid:105)
5.56%
77.78%
16.67%
75.00%
13.89%
11.11%
12.50%
66.67%
20.83%
8.33%
87.50%
4.17%
8.33%
75.00%
16.67%
72.22%
11.11%
16.67%
8.33%
80.56%
11.11%
13.89%
75.00%
11.11%
5.56%
77.78%
16.67%
79.17%
8.33%
12.50%
20.83%
62.50%
16.67%
12.50%
87.50%
0.00%
76.39%
13.89% 9.72%
Number of
weekdays
(timeslots)
New links
algorithm
Relations update
Preferential
Uniform
Intelligent
Simple
1850
5018
5050
Culture size
Arrangements
of the hierarchy but also is tightly dependent on the resources. Observed effect is statistically
significant.
Table 7: Percentage of runs for which Kendall-Tau distance is lower when preferential
attachment algorithm for creating new edges is used.
Culture size
Number of
weekdays
(timeslots)
Arrangements
Relations update
250
500
5
15
25
Intelligent
Simple
1850
5018
5050
% of runs
83.33%
77.78%
50.00%
91.67%
100.00%
83.33%
77.78%
91.67%
66.67%
83.3%
80.56%
Discussion
Our metaphor for an egalitarian system, with equal access to time resources, showed the best long-
term performance. For a longer simulation run, the superiority of the egalitarian agent execution
order over hierarchical order in terms of global effectiveness (average Willingness Usage values) is
more visible, reaching 100% of cases when considering the last simulation step (1,000th). Moreover
the performance of the egalitarian system is inversely proportional to the number of available
resources, reaching the best efficiency when the time is scarce. The obtained results strongly
support the first hypothesis (H1) for a broad range of different scenarios and parameters.
Since, in the long run, an egalitarian system performs better anyway, we can only expect that
the intelligent method of arranging meetings mentioned in the second hypothesis (H2) will have the
best performance for the initial phase of simulation. Indeed, intelligent methods, i.e., mimicking
the use of modern technology to increase the probability of finding a valuable partner and utilizing
more of the agent's preferences, show greater values of average willingness usage, but this effect is
temporary and occurs in the initial phase of the simulation. This effect is the strongest observed
for societies stratified by weighted node degree (i.e. mobile method). On the other hand, for the
14
hierarchical execution order, intelligent time management methods seem to contribute most to an
improvement of global effectiveness.
To verify the third hypothesis (H3), we have used the implemented Mobile execution order,
which was referred to as a stratified system with added mobility for agents to move on the hierarchy
ladder. We attempted to determine whether, despite some degree of the flexibility, the hierarchy
can arise and sustain. The results show that the hierarchy can indeed be emergent but only in
very specific condition and in the short term. The Mobile model running with resources set to
high availability yields signs of a crystalizing hierarchy in the short term. Until approximately half
of a single simulation run, the order in which the agents are executed stabilizes. This tendency is
strong when the resources are in plentiful supply and the preferential attachment mode is enabled.
However, in the long term this effect does not persist, and the hierarchy disappears, as the agents
are using the enabled mobility in the model. Therefore, the third hypothesis was also confirmed.
All three hypotheses raised in the introduction have been confirmed. Sensitivity analysis shows
that the obtained results are quite stable for a broad range of parameter combinations. Social
stratification, as an emergent property of a model, can itself be a broad topic for research and
an innovative means of comprehending the existence of inequality in real systems. Moreover, the
developed model provides an opportunity to study methods for overcoming social stratification,
such as asynchronous methods of communication (e.g., publication of agents' calendars and current
interests).
The results presented in this paper do not provide an ultimate answer to the question of whether
a particular level of stratification is optimal for society in general. However, obtained results
shed light on the reason for the decreased effectiveness and long-term stability of hierarchies in
the knowledge society. This reason is that hierarchies decrease the global effectiveness of time
management, measured by the degree to which agents manage to exchange ideas and discuss with
each other. Our model, based on a relatively simple process of making appointments to discuss
various topics by agents embedded in an evolving social network, is quite suitable to advanced
knowledge societies such as the Silicon Valley. The model is also able to shed insights on the
emergent properties of an advanced knowledge society. While there may be many ways to improve
the model and generalize or validate our findings, we consider them interesting areas for further
study, enabled by the contributions of this article to the research community.
References
[AH06]
Mark Aguiar and Erik Hurst. Measuring trends in leisure: The allocation of time over
five decades. Technical report, National Bureau of Economic Research, 2006.
[AH08]
[Atk96]
[Bec65]
Mark Aguiar and Erik Hurst. The increase in leisure inequality. Technical report,
National Bureau of Economic Research, 2008.
Anthony Barnes Atkinson. Public economics in action: the basic income/flat tax
proposal. OUP Catalogue, 1996.
Gary S Becker. A theory of the allocation of time. The economic journal, pages
493 -- 517, 1965.
[BFB09] Mukul S Bansal and David Fernández-Baca. Computing distances between partial
rankings. Information Processing Letters, 109(4):238 -- 241, 2009.
[Dun96]
Robin IM Dunbar. Determinants of group size in primates: A general model. Oxford
University Press, 1996.
[Dun98]
Robin IM Dunbar. The social brain hypothesis. brain, 9(10):178 -- 190, 1998.
[FS07]
[GB75]
[Gru94]
Ilan Fischer and Oriel Sullivan. Evolutionary modeling of time-use vectors. Journal of
Economic Behavior & Organization, 62(1):120 -- 143, 2007.
Gilbert Ghez and Gary S Becker. The allocation of time and goods over the life cycle.
NBER Books, 1975.
David B Grusky. The contours of social stratification. Social Stratification in Sociolog-
ical Perspective, 1994.
15
[Hay89]
[Hur11]
[JBN04]
[JS85]
[Klu00]
[Ley94]
Robert B Hays. The day-to-day functioning of close versus casual friendships. Journal
of Social and Personal Relationships, 6(1):21 -- 37, 1989.
Michael Hurwitz. The impact of legacy status on undergraduate admissions at elite
colleges and universities. Economics of Education Review, 30(3):480 -- 492, 2011.
John T Jost, Mahzarin R Banaji, and Brian A Nosek. A decade of system justification
theory: Accumulated evidence of conscious and unconscious bolstering of the status
quo. Political Psychology, 25(6):881 -- 919, 2004.
F Thomas Juster and Frank P Stafford. Time, goods, and well-being. 1985.
Juergen Kluver. The dynamics and evolution of social systems: new foundations of a
mathematical sociology, volume 29. Springer Science & Business Media, 2000.
Loet Leydesdorff. The evolution of communication systems. Loet Leydesdorff, The Evo-
lution of Communication Systems, Int. J. Systems Research and Information Science,
6:219 -- 230, 1994.
[MBH+09] Monique Borgerhoff Mulder, Samuel Bowles, Tom Hertz, Adrian Bell, Jan Beise, Greg
Clark, Ila Fazzio, Michael Gurven, Kim Hill, Paul L Hooper, and others. Intergen-
erational wealth transmission and the dynamics of inequality in small-scale societies.
science, 326(5953):682 -- 688, 2009.
[NWK08] Radoslaw Nielek, Aleksander Wawer, and Romuald Kotowski. Two antigenically indis-
tinguishable viruses in a population. In Proceedings of the World Congress on Engi-
neering, volume 3, 2008.
[Oys00]
Carol K Oyster. Groups: A user's guide. McGraw-Hill, 2000.
[RCME07] James F Robison-Cox, Richard F Martell, and Cynthia G Emrich. Simulating gender
stratification. Journal of Artificial Societies and Social Simulation, 10(3):8, 2007.
[RG10]
John Robinson and Geoffrey Godbey. Time for life: The surprising ways Americans
use their time. Penn State Press, 2010.
[Row11]
Karen Rowlingson. Does income inequality cause health and social problems? 2011.
[Sau10]
[Sma99]
[Sno10]
[SW12]
[TW08]
Peter Saunders. Beware false prophets: equality, the good society and the spirit level.
Policy Exchange, 2010.
Cathy Small. Finding an invisible history: A computer simulation experiment (in
virtual polynesia). Journal of Artificial Societies and Social Simulation, 2(3):6, 1999.
Christopher Snowdon. The spirit level delusion: fact-checking the left's new theory of
everything. Little Dice London, 2010.
Alistair Sutcliffe and Di Wang. Computational modelling of trust and social relation-
ships. Journal of Artificial Societies and Social Simulation, 15(1):3, 2012.
Don Tapscott and Anthony D Williams. Wikinomics: How mass collaboration changes
everything. Penguin, 2008.
[WNK09] Aleksander Wawer, Radoslaw Nielek, and Romuald Kotowski. Patterns in time: Mod-
elling schedules in social networks. In 1st International Workshop on Social Informatics,
pages 80 -- 85. IEEE, 2009.
[WPC09] Richard Wilkinson, Kate Pickett, and Molly Scott Cato. The spirit level. why more
equal societies almost always do better, 2009.
[You10]
Stephen Younger. Leadership in small societies. Journal of Artificial Societies and
Social Simulation, 13(3):5, 2010.
16
Appendix I
Figure 11: Simple and Intelligent schedule-filling procedure flowcharts.
17
Make agents'arrangementsWhat is theschedule fillingstrategy?Choose a randomneighborrandomAre there anyother agents?Summonan agentYesStopNoSimpleChoose firstcommon free time slotChoose a subjectmaximizing agent'swillingness usageMaximumnumberof attemptsreached?Summonan agentChoose a neighbormaximizing willingness usageAny commonsubjects?NoYesYesAny sparewillingnessand timeslots?IntelligentIterate overagent's subjectsIterate overagent'stime slotsAre there anyother agents?YesYesNoNoNo |
0804.4336 | 1 | 0804 | 2008-04-28T07:26:31 | Counterflow Extension for the F.A.S.T.-Model | [
"cs.MA",
"physics.comp-ph"
] | The F.A.S.T. (Floor field and Agent based Simulation Tool) model is a microscopic model of pedestrian dynamics, which is discrete in space and time. It was developed in a number of more or less consecutive steps from a simple CA model. This contribution is a summary of a study on an extension of the F.A.S.T-model for counterflow situations. The extensions will be explained and it will be shown that the extended F.A.S.T.-model is capable of handling various counterflow situations and to reproduce the well known lane formation effect. | cs.MA | cs |
Counterflow Extension for the F.A.S.T.-Model
Tobias Kretz1,3 and Maike Kaufman2,3 and Michael Schreckenberg3
1 PTV AG
Stumpfstrasse 1 – D-76131 Karlsruhe – Germany
E-mail: [email protected]
2 Robotics Research Group
Department of Engineering Science
University of Oxford
Parks Road – OX1 3PJ Oxford – UK
E-mail: [email protected]
3 Physics of Transport and Traffic
University of Duisburg-Essen
Lotharstrasse 1 – D-47057 Duisburg – Germany
E-mail: [email protected]
March 26, 2018
Abstract
The F.A.S.T. (Floor field and Agent based Simulation Tool) model
is a microscopic model of pedestrian dynamics [1], which is discrete
in space and time [2, 3].
It was developed in a number of more or
less consecutive steps from a simple CA model [4, 5, 6, 7, 8, 9, 10].
This contribution is a summary of a study [11] on an extension of
the F.A.S.T-model for counterflow situations. The extensions will be
explained and it will be shown that the extended F.A.S.T.-model is
capable of handling various counterflow situations and to reproduce
the well known lane formation effect.
1
Introduction
Counterflow situations [12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22] pose a special
problem to models of pedestrian dynamic as well as building construction.
To avoid deadlocks a certain degree of self-organization is necessary, which
typically leads to the formation of lanes, where people in the same lane fol-
low each other and different lanes have opposing movement directions. CA
1
models of pedestrian motion are particularly susceptible to deadlocks, if the
main orientation axis of a corridor is aligned to one of the discretization
axes. The reason for this is the given lane structure which leads to many
head-on collision situation which lack any asymmetry to decide about the
side of mutual passing. One modeling ansatz is to let opposing pedestrians
exchange their cells with a certain probability [23]. While this method is
capable of reproducing some elements of reality, one still might be unhappy,
as in reality people just do not walk through each other. A major tempta-
tion in modelling pedestrian counterflow using discrete models is to make
use of the lanes preset by the CA lattice and align the two opposing groups'
movement along one lattice axis. While such simulations can increase the
understanding of a model or method, they are merely of academic interest
and not sufficient for general use in safety, traffic or city engineering. The
other big difficulty is combining counterflow situations with speeds larger
than one cell per round. The method presented here aims at being indepen-
dent of the underlying lattice and it includes speeds larger 1.
2 Extension of the F.A.S.T.-Model by a Comoving
Dynamic Potential
The probability pxy for an agent i to select a cell (x, y) as desired cell is
multiplied by an additional factor pf
¯pxy = pxy · pf
(1)
xy:
xy
kf
N(cid:88)
j∈NN i
σ(i, j)
σ(i, j) P j
xy
xy = exp
pf
(2)
where kf is the coupling parameter that determines the strength of the effect.
NN i is the set of nearest neighbors of agent i. N is the maximum number of
nearest neighbors which is considered. NN i only consists of agents within
a distance rmax, which are visible to agent i (i.e. not hidden by walls and in
the field of view π/2 to the left and right of the direction of motion of agent
i. σ(i, j) = (cid:126)vi · (cid:126)vj is the scalar product of the velocities (cid:126)vi and (cid:126)vj of agent i
and agent j, so the fraction in equation 2 is +1, if the agents rather move
into the same direction and −1, if the rather move into opposite direction.
Finally P j
xy is the value of the comoving potential induced by agent j at
position (x, y). Assuming motion of agent j at position (x0, y0) along the
x-axis, P j
xy has the following form (for any other direction of motion, the
2
potential has to be rotated accordingly):
(cid:32)
(cid:32)
xy = 2h
P j
xy = 2h
P j
vj
vmax
j
vj
vmax
j
+ δ
+ δ
(cid:33)(cid:18)
1 − y − y0
(cid:33)(cid:18)
1 − x − x0
a
(cid:19)
(cid:19)
b
, if
, if
y − y0
x − x0 ≤ a
y − y0
x − x0 >
b
a
b
(3)
(4)
where h is the strength (height) of the potential. The content of the first
brackets models the velocity dependence of the potential, a is the basewidth
of the potential orthogonal to the direction of motion, b is the baselength
alongside the direction of motion of agent j.
The values N = 12, h = 4.0, δ = 0.2, a = 2, b = rmax = 15, and kf = 0.8
have been established as a reasonable choice of parameters. Interestingly
the simulation results worsened not only, if one reduced N, but also as it
was increased. All simulations were done with vmax = 3. Due to larger
fluctuations the method works less well for speeds vmax ≥ 5.
3 Simulation Results
Figure 1 clearly shows, how lanes are formed and by that deadlocks are
avoided and figure 2 the effect on the fundamental diagram by variation of
kf and NN i is shown.
Figure 1: Lane formation in a corridor and on an area.
4 Conclusions and Outlook
This contribution presented a model extension to the F.A.S.T.-model for the
simulation of counterflow. It was shown that lane formation was reproduced
and deadlocks were avoided for speeds up to 3 cells per round and at densities
well above the deadlock density of the earlier model. For even larger speeds
3
Figure 2: Fundamental diagrams in dependence of kf and NN i.
the model needs to be stabilized, i.e. fluctuations must be restricted to
some maximal value. An interesting empirical question would be to find the
number of maximally considered nearest neighbors for real pedestrians.
References
[1] A. Schadschneider, W. Klingsch, H. Klupfel, T. Kretz, C. Rogsch, and
A. Seyfried. Evacuation Dynamics: Empirical Results, Modeling and
Applications. In R.A. Meyers, editor, Encyclopedia of Complexity and
System Science. Springer, Berlin Heidelberg New York, 2009. (to be
published in April 2009) urn:ISBN:978-0-387-75888-6. arXiv:0802.
1620v1.
[2] T. Kretz and M. Schreckenberg. The F.A.S.T.-Model. In El Yacoubi
et al. [27], pages 712–715. DOI:10.1007/11861201_85. arXiv:0804.
1893v1.
[3] T. Kretz.
Pedestrian Traffic - Simulation and Experiments.
urn:nbn:de:hbz:
PhD thesis, Universitat Duisburg-Essen, 2007.
464-20070302-120944-7.
[4] C. Burstedde, K. Klauck, A. Schadschneider, and J. Zittarz. Simula-
tion of pedestrian dynamics using a 2-dimensional cellular automaton.
Physica A, 295:507, 2001.
DOI:10.1016/S0378-4371(01)00141-8,
arXiv:cond-mat/0102397v1.
[5] A. Kirchner and A. Schadschneider. Cellular Automaton Simulations
In Fukui et al.
http://www.thp.uni-koeln.de/~as/Mypage/
of Pedestrian Dynamics and Evacuation Processes.
[25], pages 531–536.
PSfiles/pede2.ps.
[6] A. Kirchner and A. Schadschneider. Simulation of Evacuation Pro-
cesses Using a Bionics-inspired Cellular Automaton Model for Pedes-
trian Dynamics. Physica A, 312(1–2):260–276, 2002. DOI:10.1016/
S0378-4371(02)00857-9,arXiv:cond-mat/0203461v1.
4
[7] A. Schadschneider. Cellular Automaton Approach to Pedestrian Dy-
In Schreckenberg and Sharma [24], pages 76–85.
namics – Theory.
arXiv:cond-mat/0112117v1.
[8] A. Schadschneider. Bionics-Inspired Cellular Automaton Model for
In Fukui et al. [25], pages 499–509. http:
Pedestrian Dynamics.
//www.thp.uni-koeln.de/~as/Mypage/PSfiles/pede1.ps.
[9] A. Kirchner, K. Nishinari, and A. Schadschneider. Friction Effects
and Clogging in a Cellular Automaton Model for Pedestrian Dynamics.
Phys. Rev. E, 67(056122), 2003. DOI:10.1103/PhysRevE.67.056122,
arXiv:cond-mat/0209383v1.
[10] K. Nishinari, A. Kirchner, A. Namazi, and A. Schadschneider. Ex-
tended Floor Field CA Model for Evacuation Dynamics. IEICE Trans.
Inf. & Syst., E87-D:726–732, 2004. arXiv:cond-mat/0306262v1.
[11] M. Kaufman. Lane Formation in Counterflow Situations of Pedestrian
Traffic. Master's thesis, Universitat Duisburg-Essen, 2007.
[12] M. Muramatsu, T. Irie, and T. Nagatani. Jamming transition in pedes-
DOI:10.1016/
trian counter flow. Physica A, 267:487–498, 1999.
S0378-4371(99)00018-7.
[13] D. Helbing, I.J. Farkas, and T. Vicsek. Freezing by Heating in a Driven
Mesoscopic System. Phys. Rev. Lett., 84:1240–1243, 2000. DOI:10.
1103/PhysRevLett.84.1240, arXiv:cond-mat/9904326v2.
[14] M. Muramatsu and T. Nagatani.
Jamming transition in two-
Physica A, 275:281 – 291,
DOI:10.1016/S0378-4371(99)00447-1. http://llacolen.
dimensional pedestrian traffic.
2000.
ciencias.uchile.cl/%7Evmunoz/download/papers/mn00.pdf.
[15] V.J. Blue and J.L. Adler. Cellular Automata Microsimulation of Bi-
Directional Pedestrian Flows. Transportation Research Record, Journal
of the Transportation Research Board, 1678:135–141, 2000. http://
www.ulster.net/~vjblue/CA-TRB99.pdf.
[16] A. Schadschneider, A. Kirchner, and K. Nishinari. CA Approach to
Collective Phenomena in Pedestrian Dynamics. In Bandini et al. [26],
pages 239–248. urn:ISBN:978-3-540-44304-9.
[17] Y. Tajima, K. Takimoto, and T. Nagatani. Pattern formation and
jamming transition in pedestrian counter flow. Physica A, 313:709–
723, 2002. DOI:10.1016/S0378-4371(02)00965-2.
[18] M. Isobe, T. Adachi, and T. Nagatani. Experiment and simulation of
pedestrian counter flow. Physica A, 336:638–650, 2004. DOI:10.1016/
j.physa.2004.01.043.
5
[19] A. John, A. Schadschneider, D. Chowdhury, and K. Nishinari. Collec-
tive effects in traffic on bi-directional ant trails. Journal of Theoret-
ical Biology, 231:279–285, 2004. DOI:10.1016/j.jtbi.2004.06.022.
arXiv:cond-mat/0409458v1.
[20] R. Nagai, M. Fukamachi, and T. Nagatani. Experiment and simulation
for counterflow of people going on all fours. Physica A, 358:516–528,
2005. DOI:10.1016/j.physa.2005.04.024.
[21] T. Kretz, M. Wolki, and M. Schreckenberg. Characterizing correla-
tions of flow oscillations at bottlenecks. Journal of Statistical Mechan-
ics: Theory and Experiment, P02005, 2006. DOI:10.1088/1742-5468/
2006/02/P02005. arXiv:cond-mat/0601021v2.
[22] T. Kretz, A. Grunebohm, M. Kaufman, F. Mazur, and M. Schreck-
enberg. Experimental study of pedestrian counterflow in a corridor.
Journal of Statistical Mechanics: Theory and Experiment, P10001,
2006. DOI:10.1088/1742-5468/2006/10/P10001. arXiv:cond-mat/
0609691v1.
[23] V.J. Blue and J.L. Adler. Cellular Automata Microsimulation For Mod-
eling Bi-Directional Pedestrian Walkways. Transportation Research B,
35(293), 2001. DOI:10.1016/S0191-2615(99)00052-1.
[24] M. Schreckenberg and S.D. Sharma, editors. Pedestrian and Evacuation
Dynamics, Duisburg, 2002. Springer-Verlag Berlin Heidelberg. urn:
ISBN:3-540-42690-6.
[25] M. Fukui, Y. Sugiyama, M. Schreckenberg, and D.E. Wolf, editors.
Traffic and Granular Flow '01, Nagoya, 2003. Springer-Verlag Berlin
Heidelberg. urn:ISBN:3-540-40255-1.
[26] S. Bandini, B. Chopard, and M. Tomassini, editors. Cellular Automata
- 5th International Conference on Cellular Automata for Research and
Industry, ACRI 2002, Geneva, Switzerland, October 2002. Springer-
Verlag Berlin Heidelberg. urn:ISBN:978-3-540-44304-9.
[27] S. El Yacoubi, B. Chopard, and S. Bandini, editors. Cellular Automata
- 7th International Conference on Cellular Automata for Research and
Industry, ACRI 2006, Perpignan, France, September 2006. Springer-
Verlag Berlin Heidelberg. urn:ISBN:3-540-40929-7.
6
|
1909.05232 | 1 | 1909 | 2019-09-11T17:42:14 | On Memory Mechanism in Multi-Agent Reinforcement Learning | [
"cs.MA",
"cs.AI"
] | Multi-agent reinforcement learning (MARL) extends (single-agent) reinforcement learning (RL) by introducing additional agents and (potentially) partial observability of the environment. Consequently, algorithms for solving MARL problems incorporate various extensions beyond traditional RL methods, such as a learned communication protocol between cooperative agents that enables exchange of private information or adaptive modeling of opponents in competitive settings. One popular algorithmic construct is a memory mechanism such that an agent's decisions can depend not only upon the current state but also upon the history of observed states and actions. In this paper, we study how a memory mechanism can be useful in environments with different properties, such as observability, internality and presence of a communication channel. Using both prior work and new experiments, we show that a memory mechanism is helpful when learning agents need to model other agents and/or when communication is constrained in some way; however we must to be cautious of agents achieving effective memoryfulness through other means. | cs.MA | cs | On Memory Mechanism in Multi-Agent Reinforcement Learning
Yilun Zhou,1 Derrik E. Asher,2 Nicholas R. Waytowich,2 Julie A. Shah1
1MIT CSAIL, 2US Army Research Lab
[email protected]
9
1
0
2
p
e
S
1
1
]
A
M
.
s
c
[
1
v
2
3
2
5
0
.
9
0
9
1
:
v
i
X
r
a
Abstract
Multi-agent reinforcement learning (MARL) extends (single-
agent) reinforcement learning (RL) by introducing additional
agents and (potentially) partial observability of the environ-
ment. Consequently, algorithms for solving MARL problems
incorporate various extensions beyond traditional RL meth-
ods, such as a learned communication protocol between co-
operative agents that enables exchange of private information
or adaptive modeling of opponents in competitive settings.
One popular algorithmic construct is a memory mechanism
such that an agent's decisions can depend not only upon the
current state but also upon the history of observed states and
actions. In this paper, we study how a memory mechanism
can be useful in environments with different properties, such
as observability, internality and presence of a communica-
tion channel. Using both prior work and new experiments,
we show that a memory mechanism is helpful when learning
agents need to model other agents and/or when communica-
tion is constrained in some way; however we must to be cau-
tious of agents achieving effective memoryfulness through
other means.
1
Introduction
Compared to (single-agent) reinforcement learning (RL),
multi-agent reinforcement learning (MARL) includes other
agents and potentially partial observability into the environ-
ment. Due to the partial observability of an environment
from each agent, emergent communication (Mordatch and
Abbeel, 2018; Sukhbaatar and Fergus, 2016; Foerster et al.,
2016; Gauthier and Mordatch, 2016), in which agents must
come up with a communication protocol to efficiently ex-
change information, has been studied. Another learning ob-
jective is for a few agents to learn the policies and then
get deployed in an environment with other agents. We call
these learning agents as internal agents, while the rest are
called external agents. In this setting, internal agents should
be adaptive to the different behaviors exhibited by external
agents, regardless of whether the agents are in cooperative,
competitive, or mixed relations.
A recurring theme within MARL algorithms is the use
of a memory mechanism, which is most commonly im-
plemented by a recurrent neural network for deep learning
methods, or sometimes by state augmentation if the system
designer determines that a bounded memory is sufficient.
However, sometimes the use of such a construct is not well-
motivated, with both memoryful and memoryless policy rep-
resentations having been tested in prior work (Sukhbaatar
and Fergus, 2016). Furthermore, there has not been a sys-
tematic study of whether or how such a memory mechanism
can be helpful. In this paper, we group multi-agent envi-
ronments according to three properties, and study the ben-
efit that a memory mechanism may provide. Throughout the
study, we ground our propositions with conclusions derived
from existing literature and new experiments, and consider
both theoretical and practical feasibility.
At a high level, we classify MARL environments along
three axes (detailed in Section 4). First, we consider the ob-
servability axis, or whether each agent is able to receive the
full state information at each time. Several simple bench-
marks, such as the predator-prey task and the cooperative
navigation task introduced by Lowe et al. (2017), are fully
observable. As special cases, classical game theory exam-
ples (such as rock-paper-scissors and prisoners' dilemma)
are also fully observable since they are stateless. In stark
contrast, a common source of non-observability comes from
the limited sensing range of agents, as in the traffic junction
domain used by Sukhbaatar and Fergus (2016).
Second, we consider the internality of the environment, or
whether the environment contains only internal agents (over
which we have control during test time), or also includes ex-
ternal agents (which are provided by other parties, during
test time). If all agents are internal (referred to as fully inter-
nal environments), all agents are essentially optimizing for a
shared reward within a fully cooperative setting. In partially
internal environments, agents may be engaged in fully co-
operative tasks (e.g., human-robot teaming for rescue), fully
competitive tasks (e.g., a predator-prey scenario), or mixed
tasks (e.g., traffic, in which vehicles compete for road re-
sources while also cooperating to avoid collision).
Third, we consider an axis for communication in fully in-
ternal environments: specifically, whether an explicit com-
munication channel is present. If a channel is present, agents
can learn to exchange information effectively, which may
lead to improved performance.
In this work, we study the utility of a memory mechanism
for environments according to the above three axes, and re-
port the following:
1. Within partially internal environments, regardless of re-
ward structure, memoryful policies are helpful because
they can better model opponents and inherently adapt to
potentially unseen opponent behaviors.
2. Although the presence of a communication channel with-
out restrictions does not necessitate a memory mecha-
nism, it can be helpful for agents to evolve compositional
and/or sparse communications.
3. In fully internal environments without explicit communi-
cation, a memory mechanism can typically help for agents
to engage in behavior-based communication.
4. An emergent phenomenon, which we refer to as "state-
based bookkeeping", can sometimes mitigate or remove
the duty of an explicit memory mechanism, which may or
may not be desirable.
2 Background
Similar to MDP representations in RL, a MARL en-
vironment with n agents can be specified by a tu-
ple (cid:104)S, A1, ..., An, T, R1, ..., Rn, γ(cid:105). S represents the state
space (which includes the state of all agents and additional
environmental information). Ai denotes the set of actions for
agent i. T : S × A1 × ... × An → PS represents the transi-
tion function, assuming that agents actions are simultaneous
rather than occurring in turns. Ri : S×A1×...×An×S → R
represents the reward of agent i, which is a function of the
current state, all agents' actions, and the next state. γ ∈ [0, 1]
denotes the discount factor.
is
example,
if
agents
The
the goal
is
objective
dependent
optimization
settings. For
E[(cid:80)∞
t=1 γt(cid:80)n
dividual rewards (i.e., ¯R =(cid:80)n
upon
are
environment
fully cooperative,
to identify policies
that maximize the total expected discounted reward,
i=1 Ri(st, π1(st), ..., πn(st), T (st, π1(st), ...,
πn(st))]. In this case, we can also formulate the reward
functions to be identical and equal to the average of all in-
i=1 Ri/n represents the new
reward that each agent is maximizing individually). This
is the formulation used in most literature on cooperative
agents (Sukhbaatar and Fergus, 2016; Foerster et al., 2016).
A selfish agent trying to maximize its own reward may
want to guarantee a worst-case optimal reward, resulting in
a policy that receives the highest reward regardless of other
agents' policies. Alternatively, the selfish agent may want
to achieve an average-case optimal policy, which yields the
highest expected reward assuming some distribution of op-
ponent policy.
Notably, in contrast to single-agent RL, a deterministic
policy is often not the best policy (and can even be the
worst). For example, in an iterative version of the rock-
paper-scissors game, any deterministic policy is exploitable
and should result in repetitive losses. The worst-case optimal
policy is a perfectly randomized policy that achieves equal
proportion of win, loss, and draw in expectation.
3 Memoryful Policies
It is known (Sutton and Barto, 2018) that the optimal policy
in an MDP is Markov. In other words, choosing (distribution
of) actions based only on the current state can lead to the
optimal policy, and maintaining state and/or action history
is not further advantageous. However, due to the presence of
other agents, it may be helpful for the policy to be memory-
ful. As a simple example, in a repeated rock-paper-scissors
game, against an imperfectly playing agent with a tendency
toward a specific action, a memoryful agent can exploit op-
ponent imperfection by calculating the probability of each
action and taking corresponding counter-actions. Note that
this kind of "learning" is performed on the fly, within each
episode, and thus will adapt to a distinctly different oppo-
nent next time without issue.
general form as πi : (cid:81)τ
(cid:81)τ−1
We define a memoryful policy for the i-th agent in the
t=1 Aj → Ai. In
terms of implementation, a general memoryful policy can be
represented by a recurrent neural network. However, if only
a certain length of past history is necessary, state augmen-
tation technique can also be used, as in the state representa-
tion used by Foerster et al. (2018a) in the iterative prisoners'
dilemma.
t=1 S ×(cid:81)n
j=1
4 A Taxonomy of MARL Problems
For the purpose of this work, we classify MARL problems
in terms of three attributes, each of which has different im-
plications on the requirement for memory mechanisms.
1. Observability: whether each agent individually receives
the full state information of the environment, or only a
correlated observation. Partial observability is more com-
mon in multi-agent systems, but full observability is typ-
ical in many game theory domains (such as rock-paper-
scissors or the prisoners' dilemma).
2. Internality: whether all agents are internal, or some ex-
ternal agents exist. Specifically, as discussed in the intro-
duction, we define an internal agent to be one that we can
control at test time, while an external agent is supplied
by another party during test time. If all agents are inter-
nal, we refer to that environment as fully internal; other-
wise the environment is partially internal. We assume a
perfectly shared reward structure for fully internal envi-
ronments (i.e. fully cooperative), because any source of
competition would come from external agents in realistic
situations1.
Note that partial internality does not necessarily imply a
competitive setting. For example, in human-robot team-
ing for search and rescue, the reward is still shared.
Rather, it is simply that the robot must view humans as
1Note that despite rewards being shared, the training algorithm
is not restricted to using the same reward for all agents: it can
reward agents that perform the bulk of the work while penaliz-
ing non-participating agents. The method credit assignment has re-
ceived much attention (Foerster et al., 2018b; Nguyen, Kumar, and
Lau, 2018); however, for the sake of simplicity, we do not consider
it in this paper.
external. A traffic junction represents a mixed coopera-
tive/competitive environment, since vehicles must com-
pete for road usage while cooperating to avoid collisions.
3. Communication: whether explicit message-based com-
munication is allowed among internal agents. Specifi-
cally, the directed communication channel from agent i
to j is represented by a function cij : Rdi → Rdj , where
agent i can send a di-dimensional message mij, and agent
ij = cij(mij) ∈ Rdj .
j will receive an encoded version m(cid:48)
Typically, the channel encoder is the identity function
cij(m) = m, but the encoder can also be a discretiza-
tion function to enforce discrete communication signals
(Mordatch and Abbeel, 2018; Foerster et al., 2016) or a
function that suppresses the signal after a certain number
of times to implement budgeted communication (Wang,
Everett, and How, 2019). We assume that for agent i the
outbound communication message mij is computed from
individual state observation, and the action is computed
from both the state observation and inbound communica-
tion messages m(cid:48)
ki. We do not consider a communication
channel between internal and external agents in this work.
5 Fully Observable/Fully Internal
For fully observable environments with fully internal agents,
the optimal policy is still memoryless (i.e. Markovian), at
least in theory. Specifically, we can treat the problem as
single-agent RL in which an agent learns a policy that maps
from the state space to the joint action space. Since the envi-
ronment is represented by a MDP, the optimal policy for this
agent is also Markovian (and deterministic). At test time in
the MARL problem, then, each individual agent calculates
the joint action using the original policy, and takes its in-
dividual action correspondingly. Communication is also not
required in this case, because the learned joint policy con-
tains any necessary coordination among agents.
In practice, however, scalability can be an issue for envi-
ronments containing too many agents, since the state space
and joint action spaces will grow exponentially with an in-
creasing number of agents. Therefore, intentional decentral-
ization can be helpful: with such a formulation, each agent
only observes local state information. Moreover, given ho-
mogeneous agents and parameter sharing, learned policies
also have the potential to adapt to an indefinite number of
agents (Sukhbaatar and Fergus, 2016). In this case, depend-
ing on requirements for communication, a memory mecha-
nism can be beneficial. We discuss this in greater detail in
Section 7.
6 Fully Observable/Partially Internal
As noted in the prior section, a collection of internal agents
can function effectively as one agent at both training and test
time assuming full observability: thus, we only consider the
case of a single internal agent deployed into an environment
with other external agents. In addition, we use the word "op-
ponent" to refer to the other agent, even in partially or fully
cooperative situations.
There has been a significant amount of research on learn-
ing competitive behaviors in game-theoretic settings. The
methods for doing so can be classified according to the ways
they model opponents.
The first way to address opponent modeling is to sim-
ply not model them. Obviously this is potentially problem-
atic and vulnerable. However, Littman (1994) proposed a
method to learn the worst-case optimal policy with a guar-
anteed reward against all possible type of opponents. That
paper further indicates that such a policy is memoryless, but
may be stochastic. In addition, Li et al. (2019) incorporated
such worst-case consideration into a deep MARL algorithm
proposed by Lowe et al. (2017).
The second approach is to model opponents as static or
slowly changing, so that agents can learn to track oppo-
nent policies during test time. Most recent deep MARL ap-
proaches have adopted this assumption and use co-learning
at training time, which simulates an opponent policy and
learns to better deal with that policy, while simultaneously
updating the opponent policy in order to be stronger in re-
sponse. For example, Lowe et al. (2017) proposed a general-
purpose MARL algorithm that can learn from arbitrary re-
ward structures. Foerster et al. (2018a) added a lookahead
step to affect opponent gradient update with opponent mod-
eling. Al-Shedivat et al. (2017) used meta-learning for fast
adaptation with respect to opponent evolution. Although
these methods do not explicitly require a memory mech-
anism, more interesting behaviors can emerge with some
form of memory, as demonstrated by Foerster et al. (2018a),
who used state augmentation in their iterated prisoners'
dilemma setting.
However, when the opponent is using a very different pol-
icy from the one modeled during training, or is changing
policy abruptly across episodes, the agent may receive very
bad rewards. Figure 1 sketches a possible failure case for this
kind of co-learning based method. In fact, many co-learning
approaches lack a guarantee that the current agent policy re-
mains effective against a previous opponent policy, which is
analogous to the problem of forgetting in both discrimina-
tive (Kemker et al., 2018) and generative (Wu et al., 2018)
models.
Figure 1: Training dynamics of co-learning algorithms.
In addition, Lowe et al. (2019) found that in policies
trained using such co-learning methods, while positive sig-
naling is present, positive listening is not. In other words,
the receiving agent does not seem to base its actions upon
the sending agent's communication. This suggests that, in a
stateless environment such as a matrix game with commu-
nication, the receiving agent's policy implicitly learns the
best response to the sending agent's policy (with which the
message is correlated, as observed with positive signaling),
but would poorly adapt to any policy change by the sending
agent, even with intention signaling.
The last final approach, which has been relatively less
studied, is to model opponents explicitly at test time. Specif-
ically, within an episode, an explicit memory mechanism is
used to simultaneously infer and adapt to the current oppo-
nent policy. This method places relatively few assumptions
on the opponent and can adapt to drastic changes to oppo-
nent policies across episodes.
Theory of mind (Leslie, 1987) is a popular framework for
such reasoning. For example, Rabinowitz et al. (2018) used
a mental state network to model opponent policy from past
behavior. Raileanu et al. (2018) learned to infer the goals of
opponent in an online manner and correspondingly adapt the
agent's own behavior. Yang et al. (2019) assumed switching
behaviors among several stationary policies for the opponent
and inferred the acting policy.
Another advantage is that this method can exploit any im-
perfection or weakness of a given opponent. To see this,
consider again the (stateless) game of iterated rock-paper-
scissors, wherein the policy is simply parametrized by the
probabilities of outputting rock, paper, or scissors (π =
[θr, θp, θs]). A worst-case optimal method (Littman, 1994)
ignoring the opponent's policy would compute the policy to
be π = [1/3, 1/3, 1/3], with an expected reward of 0. Co-
learning algorithms would produce a pair of policies that os-
cillate on this probability simplex. Specifically, for two agent
policies parametrized by π and π(cid:48), the expected average re-
ward for the π agent is as follows:
E[Rπ,π(cid:48)] = θrθ(cid:48)
r + θsθ(cid:48)
For the π(cid:48) agent in this case, the expected average reward
r −
is −E[Rπ,π(cid:48)]. Thus, the policy gradient is (θ(cid:48)
r) for π, and (θs − θp, θr − θs, θp − θr) for π(cid:48). The
s, θ(cid:48)
θ(cid:48)
left diagram of Figure 2 dipicts the oscillation of the policy
update.
s − θsθ(cid:48)
r.
p − θpθ(cid:48)
p − θrθ(cid:48)
s + θpθ(cid:48)
s − θ(cid:48)
p − θ(cid:48)
p, θ(cid:48)
Nevertheless, at any given time, if the learned agent is de-
ployed in an environment, even with an opponent whose pol-
icy is uniformly sampled from this simplex, the expected re-
ward is 0. Use π to denote the agent policy, and π(cid:48) to denote
√
the opponent policy. The probability density function for the
uniform opponent policy is p(π(cid:48)) =
s=1.
Thus Eπ(cid:48)[E[Rπ,π(cid:48)]] = 0 by symmetry.
3/2 · 1θ(cid:48)
By comparison, a simple memoryful policy would esti-
mate the opponent policy parameters by action history, and
take the corresponding counter-action, as implemented by
the following hand-crafted recurrent neural network (rock,
paper, and scissors have the index of 0, 1, and 2 respec-
tively).
p+θ(cid:48)
r+θ(cid:48)
h0 = [0, 0, 0],
ht = γht−1 + xt,
it = arg max ht,
yt = [1, 2, 0][it].
With γ = 1, ht represents the count for opponent actions,
xt the current opponent action, and yt the current agent ac-
tion. If we expect the opponent policy to be not stationary, a
discount factor γ ≤ 1 can be added in order to favor more
recent opponent actions.
The simple procedure exemplifies the idea that using a
memory mechanism to explicitly model opponents and de-
rive belief about the opponent policy (ht), upon which the
corresponding decision is made (yt). In some sense, this
idea is similar to the RL2 approach proposed by Duan et al.
(2016) in which variations of a single agent RL environment
are learned.
Ideally, the recurrent calculation is learned from data
with general-purpose function approximators such as LSTM
(Gers, Schmidhuber, and Cummins, 1999) or GRU (Cho
et al., 2014) cells. For example, Figure 2 plots the reward
for training a GRU network with a policy gradient, along
with the average performance for the above manually speci-
fied policy (which is statistically optimal). Our learned pol-
icy is able to reach the optimal policy within hundreds of
training episodes, while a memoryless policy based on the
co-learning approach is unable to learn against drastically
changing opponents across episodes.
Figure 2: Left: trajectories of agent and opponent poli-
cies with a co-learning approach exhibit oscillation and
non-convergence. Right: agent reward for the repeated RPS
game, with uniformly distributed opponents.
A more real-life domain is autonomous driving. Human
driver modeling is an important aspect of it, which is why
pure reinforcement learning approaches are not very com-
mon, but often combined with supervised or imitation learn-
ing from human driving data (Xu et al., 2017; Zhang and
Cho, 2016). In addition, co-learning driving policies are
not likely to be successful during deployment because the
learned opponent behavior may not be similar to actual hu-
man behaviors2.
As a concrete example, consider driving on a two-lane
road with a leading vehicle in front, represented as a grid
world as shown in Figure 3. At each time step, each vehicle
decides to either move forward (right) one position, change
lane, or stay still. Our agent is rewarded for fast traversal
through the road while avoiding collision. Therefore, its op-
timal policy for our agent depends on the other agent: if the
other agent is driving slowly, our agent may wish to pass it
on another lane, but not if the other vehicle is changing lanes
2Such observation is captured by the argument that autonomous
vehicles (AVs) driving alongside other AVs is actually easier than
driving alongside human drivers.
erratically. Specifically, we model four types of behavior for
the other agent, whose action probabilities are summarized
in Table 1.
Figure 3: Traffic domain.
type
forward
lane-change
passive-fast (PF)
passive-slow (PS)
aggressive-fast (AF)
aggressive-slow (AS)
90%
20%
30%
20%
10%
10%
70%
70%
stay
0%
70%
0%
10%
Table 1: Proportion of each action for different behaviors.
Additionally, we consider three policies for our agent. The
greedy policy will always try to pass the other agent. The
conservative will always try to stay behind the other agent
and avoid collision. Finally, the adaptive policy will infer
the other agent's policy using maximum-likelihood estima-
tion and choose the greedy strategy for passive opponent and
conservative strategy for aggressive opponent.
With total road length of 30, time penalty of 1, crash
penalty of 30, and initial separation of the two vehicles of
10 (opponent in front), Table 2 shows the reward of the three
policies against the four type of behaviors, as well as a uni-
form mixture of behavioral types.
policy
greedy
conservative
adaptive
PF
-29.0
-29.0
-29.0
PS
-33.4
-94.3
-33.3
AF
-660.3
-70.1
-70.9
AS
-237.3
-94.5
-95.3
Mix
-227.5
-71.1
-57.3
Table 2: Reward for each policy against different behaviors.
We can see that while memoryless greedy and conserva-
tive policies can each achieve high reward on different static
opponent behaviors, both perform poorly if the opponent is
randomizing its behavior per episode. On the other hand,
the adaptive policy performs opponent modeling within each
episode and can therefore achieve high reward regardless of
opponent behaviors.
7 Partially Observable/Fully Internal
In this section we assume that all agents have joint access
to the full state information for a given environment. If all
agents (combined) do not receive the full state information,
this problem is at least as hard as a partially observable
Markov decision process (POMDP) (Cassandra, Kaelbling,
and Littman, 1994), and a memory mechanism is neces-
sary (e.g. as done recently with deep recurrent Q-learning
(Hausknecht and Stone, 2015)).
With (explicit) communication
Much recent research has been focused on the use of com-
munication to complete cooperative tasks (mostly through
reinforcement learning (Gauthier and Mordatch, 2016)). For
example, Sukhbaatar and Fergus (2016) provided an end-
to-end model for learning agent communication between
an indefinite amount of agents. Foerster et al. (2016) pro-
posed a differentiable communication channel so that back-
propagation could be performed across time steps. Lowe et
al. (2017) demonstrated that arbitrary communication chan-
nels can be incorporated. None of the above works made
use of a memory mechanism, since the channel capacity is
greater than that required to encode the necessary informa-
tion exchange.
In general, if the communication channel capacity is suf-
ficient and agents can jointly observe the full state informa-
tion, they can exchange such information through commu-
nication. This is the setting studied in most literature related
to learning communication strategies (Lowe et al., 2017;
Sukhbaatar and Fergus, 2016). Since actions will be taken
upon receiving the messages, a memory mechanism would
not be helpful in such scenario.
However, if the communication channel can only output
a finite and discrete set of tokens, a memory mechanism
is necessary in order to learn a compositional semantic of
messages. For example, Mordatch and Abbeel (2018) found
that interpretable compositional language can emerge from
cooperative memoryful agents. Havrylov and Titov (2017)
conducted a similar study, but with the target of learning
to refer to pictures. Das et al. (2017) showed that dialog
communication can also emerge in a similar image guess-
ing game. Evtimova et al. (2017) studies a similar variant of
image referral game by incorporating multi-modality. Choi,
Lazaridou, and de Freitas (2018) and Bogin, Geva, and Be-
rant (2018) both used the obverter technique (Oliphant and
Batali, 1997) to learn language, based on the assumption
that the model for the receiving agent is the same as that
for the sending agent, but for different environments: a grid
world environment and a referral game. Without exception,
all these works implemented recurrent policies.
Another constraint is if the channel only allows a certain
amount of messages to pass before being non-transmitting.
This circumstance requires a memory mechanism on both
the sender and receiver ends to memorize the last message
being exchanged, so that limited amount of communication
attempts can be put into best possible use. Wang, Everett,
and How (2019) confirm that agents are unlikely to learn
good policies in the absence of such mechanism.
Without (explicit) communication
In the absence of explicit communication, agents can still
find ways to communicate their specific behaviors. This is
perhaps unsurprising, because such behavior-based commu-
nication is also common among human interactions, with
body language serving as a prominent example. However,
a memory mechanism is typically required to both generate
and parse communicative behaviors, which are temporally
extended.
As an illustrative example, consider the environment de-
picted in Figure 4. Three agents start at the left of three lanes,
while a target will appear randomly on the right of one lane.
The agents are not allowed to cross lanes and reward is re-
ceived when the target is reached. At each time step, each
agent can choose to move to the left or right for one posi-
tion, or stay still, and will incur a cost for movement (e.g.
fuel cost). Only the center (red) agent can observe the target
location, but all agents can observe other agents.
often temporally extended since primitive behaviors (i.e. ac-
tions) are not necessarily expressive enough (Mordatch and
Abbeel, 2018). Therefore, both the sending and receiving
agents could benefit from an explicit memory mechanism.
8 Partially Observable/Partially Internal
In this case, we not only have external agents, but also have
several internal agents, each of which cannot fully observe
the environmental state. From the previous two sections, it
was shown that a memory mechanism is needed to model
opponents and to communicate with teammates under con-
straints. Therefore, a memory mechanism is also expected
to help in this case, although it is studied by relatively few
amount of work. For example, Lowe et al. (2017) proposed
an algorithm that can work with this type of environment but
is based on co-learning, which is prone to drastically chang-
ing external agents.
9 State-Based Bookkeeping
Figure 4: The environment for the three-lane target pursuit
task, without explicit communication.
Without cost for agent movement, the optimal policy for
this problem is for all three agents to independently head off
to pursue the potential target (with only one agent ultimately
arriving at that target). However, in the presence of such a
cost, a better policy would be for the center agent to com-
municate target's location with its behavior, which is in turn
parsed by the two other agents to receive the corresponding
message.
target location
top
center
bottom
first two steps
→,←
→,→
→,·
Table 3: Example of behavior-based communication to solve
the domain in Figure 4.
More concretely, consider the behavior-based communi-
cation of the red agent during the first two time steps, as
shown in Table 3. The red agent uses a RIGHT-LEFT be-
havior sequence to convey that the target is in the top lane, a
RIGHT-NOP sequence to signal that the the target is in the
bottom lane, and a RIGHT-RIGHT sequence to indicate that
the target is in the center lane. Given a memory mechanism,
the blue and green agents are able to parse such temporally
extended behavioral communication and take corresponding
actions. If fuel cost is greater than time cost, it is easy to
see that such a policy achieves better reward than the naive
policy of every agent moving to the right at the same time.
Since the signaling agent must trade off using behavior
for either task completion or communication, any communi-
cation must be sparse (Wang, Everett, and How, 2019), and
Throughout this work, we assume that the agent achieves the
goal of remembering past history only through the provided
memory mechanism. However, this is not necessarily the
case, as the agent can encode memory in unexpected places:
specifically, the agent can store past history (i.e. opponent
actions or teammate communication messages) within its
own state.
As an example in a discrete domain, consider a target-
reaching task with cooperative communication in one-
dimensional discrete space that includes four targets located
at locations from 11 to 14. The mover agent starts at location
0, and must reach the correct pre-specified target (which is
not observed), with a maximum movement of five grids each
time. Its speaker teammate can observe the target location
and send a discrete communication token from a set of two
tokens {A, B} each time, but cannot observe the mover's
location. The episode is terminated when either of the four
targets is reached, and reward of 10 is given for reaching
the correct target. A cost of 1 is applied at each time step to
encourage faster completion.
Apparently, compositional language (i.e. phrases) is nec-
essary to communicate the target location using the limited
set of available tokens. In addition, when observing only the
target as the state, the speaker requires a memory mecha-
nism to generate a sequence of tokens. However, we show
that the mover does not require an explicit memory mecha-
nism in order to parse the communication.
Table 4 depicts the speaker's emitted phrases correspond-
ing to each target location, which we assume to be fixed. At
each time step, the mover's policy is to move right a certain
number of steps depending on its location and the received
token at that time step: a = π(s, c), where c ∈ {A, B}.
Policy search finds the policy illustrated in Figure 5. Upon
receiving message token A, the agent moves to position 4,
and moves to position 5 when receiving B. Therefore, its
current position serves as a way to memorize the previous
token. Then from position 4, it moves to 7 when receiving
another A, and to 8 when receiving another B. Similarly,
when from position 5, it moves to 9 when receiving A and
target location message
11
12
13
14
A, A
A, B
B, A
B, B
Table 4: Emitted message for each target location. Any token
emitted beyond the 2nd time step is null.
Figure 5: A memoryless policy incorporating state-based
bookkeeping
10 when receiving B. Again, it uses its state to memorize the
communication history (i.e. positions 7, 8, 9, and 10 clearly
disambiguate messages AA, AB, BA, and BB). Then it
will move a constant four steps to the desired target loca-
tions. Overall three time steps will be taken, which is also
optimal even with a memory mechanism. Moreover, an op-
timal policy can be derived for any communication protocol
as long as different targets are disambiguated by messages,
facilitating the discovery of such solutions.
Such a technique is also possible in continuous domains.
In particular, Wang, Everett, and How (2019) found that
when using an actor-critic algorithm with a memoryful
critic, memoryless actors are sufficient to achieve limited-
budget communication, which requires agents to remember
past messages: their finding may be explained by such phe-
nomenon. Notably, similar phenomena have been discov-
ered in other applications as well, such as neural network
capable of hiding high-frequency information inside images
(Chu, Zhmoginov, and Sandler, 2017). Nevertheless, one
way of mitigating such phenomenon is to make the transition
stochastic, so that any high-frequency information would be
destroyed by the transition function.
In a similar vein, agents in communication-free environ-
ments can also employ this kind of state-based communica-
tion, as discussed above. However, unlike state-based mem-
ory mechanisms, such communications are to a large extent
robust to environment noise, especially if such behaviors are
temporally extended, with generation and parsing performed
by a dedicated memory mechanism. However, when evaluat-
ing emergent communication protocols, it is critical to con-
firm whether agents are indeed using only the communicated
messages or other agents' behaviors as well.
10 Conclusion and Future Work
In this paper, we identified situations in multi-agent rein-
forcement learning can benefit from an explicit memory
mechanism. The two major uses of such a mechanism is to
model opponents and to carry out communicative behaviors.
First, we argue that opponent modeling is an indispens-
able component of learning with external agents, regardless
of the reward structure being shared or not. We found that
the more popular co-learning approach employed by most
algorithms will only keep an implicit model of the oppo-
nent and will poorly adapt to changing opponents, which is
shown both experiments and confirmed by findings of Lowe
et al. (2019).
Second, we argue that a memory mechanism allows for
more flexible communication protocols, both implicit and
explicit. Most literature on communication focus on explicit
communication, in which a dedicated communication chan-
nel is set up between the sending and the receiving agents.
In this scheme, a memory mechanism has been shown to en-
able both temporally extended and sparse communications.
In addition, we also show a memory mechanism can help
emerge implicit, behavior-based communication, which is
often temporally extended.
Last, we identify a potential way for agents to bypass such
a memory mechanism through what we call a state-based
bookkeeping mechanism. Specifically, since state is persis-
tently accessible by the agent during the Markovian transi-
tion, the agent may find a policy that uses state to implicitly
encode memory content.
One direction for future work is to design algorithms that
can model more sophisticated opponents, including those
that also adapt to our agent's behavior in an online man-
ner (similar to (Foerster et al., 2018a) for co-learning algo-
rithms). In addition, most studies of emergent communica-
tion is on explicit communication, but a memory mechanism
also enables us to measure implicit communication. More-
over, the most general form of communication combines im-
plicit and explicit ones, in that with a memory mechanism,
the "meaning" as understood by the receiving agent is de-
pendent not only on the explicit message, but also on the
sending agent's movement, and potentially as well as the
receiving agent's own state history. This may hinder inter-
pretability, and it may be desirable to disentangle the contri-
bution of each component to the perceived meaning of the
message.
References
Al-Shedivat, M.; Bansal, T.; Burda, Y.; Sutskever, I.; Mor-
datch, I.; and Abbeel, P. 2017. Continuous adaptation via
meta-learning in nonstationary and competitive environ-
ments. arXiv preprint arXiv:1710.03641.
Bogin, B.; Geva, M.; and Berant, J. 2018. Emergence
of communication in an interactive world with consistent
speakers. arXiv preprint arXiv:1809.00549.
Cassandra, A. R.; Kaelbling, L. P.; and Littman, M. L. 1994.
Acting optimally in partially observable stochastic do-
mains. In Aaai, volume 94, 1023 -- 1028.
Cho, K.; Van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.;
Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learn-
ing phrase representations using rnn encoder-decoder
arXiv preprint
for statistical machine translation.
arXiv:1406.1078.
Choi, E.; Lazaridou, A.; and de Freitas, N. 2018. Compo-
sitional obverter communication learning from raw visual
input. arXiv preprint arXiv:1804.02341.
Chu, C.; Zhmoginov, A.; and Sandler, M.
clegan, a master of steganography.
arXiv:1712.02950.
2017. Cy-
arXiv preprint
Das, A.; Kottur, S.; Moura, J. M.; Lee, S.; and Batra,
2017. Learning cooperative visual dialog agents
D.
In Proceedings of
with deep reinforcement learning.
the IEEE International Conference on Computer Vision,
2951 -- 2960.
Duan, Y.; Schulman, J.; Chen, X.; Bartlett, P. L.; Sutskever,
I.; and Abbeel, P. 2016. RL2: Fast reinforcement learn-
arXiv preprint
ing via slow reinforcement learning.
arXiv:1611.02779.
Evtimova, K.; Drozdov, A.; Kiela, D.; and Cho, K. 2017.
Emergent communication in a multi-modal, multi-step
referential game. arXiv preprint arXiv:1705.10369.
Foerster, J.; Assael, I. A.; de Freitas, N.; and Whiteson, S.
2016. Learning to communicate with deep multi-agent
reinforcement learning. In Advances in Neural Informa-
tion Processing Systems, 2137 -- 2145.
Foerster, J.; Chen, R. Y.; Al-Shedivat, M.; Whiteson, S.;
2018a. Learning with
Abbeel, P.; and Mordatch, I.
In Proceedings of the
opponent-learning awareness.
17th International Conference on Autonomous Agents
and MultiAgent Systems, 122 -- 130.
International Foun-
dation for Autonomous Agents and Multiagent Systems.
Foerster, J. N.; Farquhar, G.; Afouras, T.; Nardelli, N.; and
Whiteson, S. 2018b. Counterfactual multi-agent policy
gradients. In Thirty-Second AAAI Conference on Artificial
Intelligence.
Gauthier, J., and Mordatch, I. 2016. A paradigm for sit-
uated and goal-driven language learning. arXiv preprint
arXiv:1610.03585.
Gers, F. A.; Schmidhuber, J.; and Cummins, F. 1999. Learn-
ing to forget: Continual prediction with lstm.
Hausknecht, M., and Stone, P. 2015. Deep recurrent q-
learning for partially observable mdps. In 2015 AAAI Fall
Symposium Series.
Havrylov, S., and Titov, I. 2017. Emergence of language
with multi-agent games: Learning to communicate with
sequences of symbols. In Advances in neural information
processing systems, 2149 -- 2159.
Kemker, R.; McClure, M.; Abitino, A.; Hayes, T. L.; and
Kanan, C. 2018. Measuring catastrophic forgetting in
In Thirty-second AAAI conference on
neural networks.
artificial intelligence.
Leslie, A. M. 1987. Pretense and representation: The origins
of" theory of mind.". Psychological review 94(4):412.
Li, S.; Wu, Y.; Cui, X.; Dong, H.; Fang, F.; and Russell, S.
2019. Robust multi-agent reinforcement learning via min-
imax deep deterministic policy gradient. In AAAI Confer-
ence on Artificial Intelligence (AAAI).
Littman, M. L. 1994. Markov games as a framework for
multi-agent reinforcement learning. In Machine learning
proceedings 1994. Elsevier. 157 -- 163.
Lowe, R.; Wu, Y.; Tamar, A.; Harb, J.; Abbeel, O. P.; and
Mordatch, I. 2017. Multi-agent actor-critic for mixed
In Advances in
cooperative-competitive environments.
Neural Information Processing Systems, 6379 -- 6390.
Lowe, R.; Foerster, J.; Boureau, Y.-L.; Pineau, J.; and
Dauphin, Y. 2019. On the pitfalls of measuring emergent
communication. In Proceedings of the 18th International
Conference on Autonomous Agents and MultiAgent Sys-
tems, 693 -- 701. International Foundation for Autonomous
Agents and Multiagent Systems.
Mordatch, I., and Abbeel, P. 2018. Emergence of grounded
compositional language in multi-agent populations.
In
Thirty-Second AAAI Conference on Artificial Intelligence.
Nguyen, D. T.; Kumar, A.; and Lau, H. C. 2018. Credit as-
signment for collective multiagent rl with global rewards.
In Advances in Neural Information Processing Systems,
8102 -- 8113.
Oliphant, M., and Batali, J. 1997. Learning and the emer-
gence of coordinated communication. Center for research
on language newsletter 11(1):1 -- 46.
Rabinowitz, N. C.; Perbet, F.; Song, H. F.; Zhang, C.; Es-
lami, S.; and Botvinick, M. 2018. Machine theory of
mind. arXiv preprint arXiv:1802.07740.
Raileanu, R.; Denton, E.; Szlam, A.; and Fergus, R. 2018.
Modeling others using oneself in multi-agent reinforce-
ment learning. arXiv preprint arXiv:1802.09640.
Sukhbaatar, S., and Fergus, R. 2016. Learning multiagent
In Advances in
communication with backpropagation.
Neural Information Processing Systems, 2244 -- 2252.
Sutton, R. S., and Barto, A. G. 2018. Reinforcement learn-
ing: An introduction. MIT press.
Wang, R. E.; Everett, M.; and How, J. P. 2019. R-maddpg for
partially observable environments and limited communi-
cation.
Wu, C.; Herranz, L.; Liu, X.; van de Weijer, J.; Raducanu,
B.; et al. 2018. Memory replay gans: Learning to generate
new categories without forgetting. In Advances In Neural
Information Processing Systems, 5962 -- 5972.
Xu, H.; Gao, Y.; Yu, F.; and Darrell, T. 2017. End-to-
end learning of driving models from large-scale video
datasets. In Proceedings of the IEEE conference on com-
puter vision and pattern recognition, 2174 -- 2182.
Yang, T.; Hao, J.; Meng, Z.; Zheng, Y.; Zhang, C.; and
Zheng, Z. 2019. Bayes-tomop: A fast detection and best
response algorithm towards sophisticated opponents. In
Proceedings of the 18th International Conference on Au-
tonomous Agents and MultiAgent Systems, 2282 -- 2284.
International Foundation for Autonomous Agents and
Multiagent Systems.
Zhang, J., and Cho, K. 2016. Query-efficient imitation learn-
ing for end-to-end autonomous driving. arXiv preprint
arXiv:1605.06450.
|
1908.08787 | 1 | 1908 | 2019-08-13T22:29:57 | Swarm Intelligence for Morphogenetic Engineering | [
"cs.MA",
"nlin.PS",
"q-bio.TO"
] | We argue that embryological morphogenesis provides a model of how massive swarms of microscopic agents can be coordinated to assemble complex, multiscale hierarchical structures. This is accomplished by understanding natural morphogenetic processes in mathematical terms, abstracting from the biological specifics, and implementing these mathematical principles in artificial systems. We have developed a notation based on partial differential equations for artificial morphogenesis and have designed a prototype morphogenetic programming language, which permits precise description of morphogenetic algorithms and their automatic translation to simulation software. Morphogenetic programming is illustrated by two examples: (1) use of a modified flocking algorithm to route dense fiber bundles between regions of an artificial cortex while avoiding other bundles; (2) use of the clock-and-wavefront model of spinal segmentation for the assembly of the segmented spine of an insect-like robot body and for assembling segmented legs on the robot's spine. Finally, we show how a variation of smoothed particle hydrodynamics (SPH) swarm robotic control can be applied to the global-to-local compilation problem, that is, the derivation of individual agent control from global PDE specifications. | cs.MA | cs |
SWARM INTELLIGENCE FOR MORPHOGENETIC ENGINEERING
A PREPRINT
Bruce J. MacLennan
Allen C. McBride
Dept. of Electrical Engineering and Computer Science
Dept. of Electrical Engineering and Computer Science
University of Tennessee, Knoxville
University of Tennessee, Knoxville
[email protected]
[email protected]
August 26, 2019
ABSTRACT
We argue that embryological morphogenesis provides a model of how massive swarms of micro-
scopic agents can be coordinated to assemble complex, multiscale hierarchical structures. This is
accomplished by understanding natural morphogenetic processes in mathematical terms, abstracting
from the biological specifics, and implementing these mathematical principles in artificial systems.
We have developed a notation based on partial differential equations for artificial morphogenesis and
have designed a prototype morphogenetic programming language, which permits precise description
of morphogenetic algorithms and their automatic translation to simulation software. Morphogenetic
programming is illustrated by two examples: (1) use of a modified flocking algorithm to route dense
fiber bundles between regions of an artificial cortex while avoiding other bundles; (2) use of the
clock-and-wavefront model of spinal segmentation for the assembly of the segmented spine of an
insect-like robot body and for assembling segmented legs on the robot's spine. Finally, we show
how a variation of smoothed particle hydrodynamics (SPH) swarm robotic control can be applied
to the global-to-local compilation problem, that is, the derivation of individual agent control from
global PDE specifications.
Keywords active matter · artificial morphogenesis · morphogenetic engineering · smoothed particle hydrodynamics ·
swarm intelligence · swarm robotics
1 Self-assembly of Complex Hierarchical Structures
We are applying swarm intelligence to the coordination of microrobot swarms to assemble complex, hierarchically
structured physical systems. We are interested in multiscale systems that are intricately and specifically structured
from the microscopic level up through the macroscopic level: from microns to meters. On one hand, conventional
manufacturing techniques, including additive manufacturing, work well at the macroscopic level, but are difficult to
scale down below the millimeter scale. On the other, technologies that are effective at the microscopic scale, such as
photolithography and molecular self-assembly, do not scale well to complex macroscopic objects.
Nevertheless, it would be valuable to be able to assemble automatically systems structured across many length scales.
For example, the human brain has complex structures spanning more that six orders of magnitude, from micron-
scale synapses to decimeter-scale functional regions and interconnections [17].
It is not unreasonable to suppose
that a neuromorphic computer with capacities and functions similar to a human brain would have similar complexity.
How would we assemble a neuromorphic computer with 100 billion neurons and perhaps 100 trillion nonrandom
connections? Similarly complex future sensors and actuators, with capabilities similar to animal sense organs and
effectors, will span scales from the microscopic to the macroscopic.
Automatic assembly of such complex structures might seem to be an unobtainable goal, but we know it is possible,
for nature does it. Embryos develop from a single cell to a complex organism with many trillions of cells, each
itself a complex hierarchical system. During development, cells communicate, coordinate, and cooperate, behaving
as a massive swarm, to differentiate and rearrange into the various interrelated tissues that constitute a complete
A PREPRINT - AUGUST 26, 2019
organism. Morphogenesis refers to the developmental process that assembles and organizes three-dimensional forms
and structures. The process is massively parallel, distributed, and robust: all goals for us too.
Morphogenetic engineering or artificial morphogenesis takes inspiration from biological morphogenesis, seeking new
ways to create complex structures that share some ideas from the development of embryos in multicellular organisms.
Because of the diversity of mechanisms in biological morphogenesis from which inspiration might be taken, morpho-
genetic engineering is diverse. It can overlap several fields including swarm robotics, modular robotics, amorphous
computing, cellular automata, and synthetic biology. Bodies developed through morphogenetic engineering may take
form through the birth and death of sessile elements, or through the rearrangement of motile ones. Design approaches
also vary. Processes may be designed by hand in an ad hoc fashion to illustrate particular hypotheses or principles.
Designs may also be generated by hand in more systematic ways, for example through global-to-local compilation.
Evolutionary algorithms are also popular for automatic design of processes to achieve well-defined goals. Finally, a
spectrum exists from biological morphogenesis with limited human control to fully artificial systems [43].
Embryological morphogenesis has inspired a variety of approaches to swarm intelligence [16, 33, 32, 41, 7, 2, 11];
our approach to artificial morphogenesis tends to adhere more closely to natural morphogenesis than do most of these
others. reference [8] provide a rigorous taxonomy of morphogenetic engineering, including additional examples, and
reference [35] provide a more recent review of similar work.
Morphogenetic engineering is a specific approach to the goal of programmable matter, that is, the ability to system-
atically control the properties and behavior of material systems at a fine level [13, 27], and artificial morphogenesis
exhibits the properties of active matter [34, 6, 45]. Morphogenetic engineering also has some similarities to amorphous
computing [1], but artificial morphogenesis processes may begin with simple, structured preparations (see for example
Sec. 5 below), and even if they do begin in an unorganized state, it is characteristic of morphogenetic processes to
quickly self-organize.
Although natural morphogenesis is a complex and intricate process -- still incompletely understood -- biologists have
identified about twenty fundamental processes [9, pp. 158 -- 9][39]. Not all are applicable to artificial morphogenesis,
and some (such as cell division) may be difficult to implement, nevertheless these processes provide an agenda for
morphogenetic engineering, since they are in principle sufficient for assembling something as complex as an animal
body [20, 22, 25, 26]. Processes that cannot be directly implemented (such as cell division) may need to have their
function accomplished by alternative means (e.g., providing components from an external source and moving them to
the growth zone; see Sec. 5.1 for an example).
During biological morphogenesis, cells emit and receive chemical signals, migrate and differentiate in response to
those signals, and participate in both reproduction (cell proliferation) and programmed cell death (apoptosis).
In
the process they emit and absorb molecules that serve both as communication media (morphogens) and as structural
elements. Similarly, in our approach to artificial morphogenesis we distinguish active components and passive com-
ponents [27]. Active components are microscopic agents (microrobots or genetically engineered micro-organisms)
that can emit and respond to simple signals, implement simple (primarily analog) control processes, and move and
transport other (active or passive) components. Passive components are all the rest, including signaling molecules
and structural components. They do not move under their own power, but are moved by external forces (including
Brownian motion and active components). In a typical artificial morphogenesis process, agents (active components)
might transport and assemble passive components into a desired structure, or they might assemble themselves into the
structure, as cells do in biological development.
2 Describing Artificial Morphogenesis with PDEs
The challenge of artificial morphogenesis is to control very large swarms of active components (microscopic agents) to
interact with each other and with the passive components to assemble a desired structure. This requires an appropriate
level of abstraction that allows the process to be described in sufficient detail for implementation without becoming
obscured by details of a specific implementation technology. Here again we have taken our inspiration from biology,
for biologists often use partial differential equations (PDEs) to describe morphogenetic processes. At this level of ab-
straction developing tissues and diffusing signaling chemicals are described at a macroscopic level largely independent
of individual cells. It also allows the methods of continuum mechanics to be applied, which are especially appropriate
for embryological development, which takes place in the domain of "soft matter" (viscoelastic materials) [4, 9, 42].
Therefore we are operating in the domain of well-understood mathematical methods in which continuum models are
applied to macroscopic volumes of materials with a discrete molecular structure.
Scalability is a principal advantage of using PDEs for morphogenetic engineering. This is because a PDE, especially
when used in continuum mechanics or fluid dynamics, treats a material as a phenomenological continuum; that is, it
2
A PREPRINT - AUGUST 26, 2019
is treated as though composed of an infinite number of infinitesimal particles, which is a usable approximation of a
material composed of a very large number of very small particles. Therefore, if our goal in artificial morphogenesis
is to use a very large swarm of very small agents, PDEs take this goal to the continuum limit. Instead of worrying
whether our algorithms will scale up to larger swarms of smaller agents, we have the complementary, but easier, task
of scaling down from an infinite number of infinitesimal agents to a very large number of very small agents.
Another advantage of expressing morphogenetic processes in PDEs is that they are largely independent of agent size.
This is because the processes are expressed in terms of intensive quantities, such as agent density, rather than extensive
quantities, such as numbers of agents [20, 22, 28]. Therefore, an algorithm that produces structures of a particular
size will continue to be correct even if the size of the agents is changed. (There are obviously limits to this size
independence; the continuum approximation has to be good.)
In summary, using PDEs allows us to describe the behavior of massive swarms, by which we mean swarms that
can be treated as a continuous mass. We treat massive swarms the same way tissues are treated in embryological
morphogenesis and the same way fluids and solid masses are treated in continuum mechanics.
3 A Morphogenetic Programming Language
In order to test algorithms for the coordination of microrobot swarms for morphogenetic engineering, we have devel-
oped a morphogenetic programming notation based on mathematical notation for PDEs [20, 21, 23, 24, 25]. In order
to facilitate simulations, this notation has been formalized into a morphogenetic programming language, tentatively
named "Morphgen." We have a prototype implementation of Morphgen by means of a syntax macroprocessor, which
translates Morphgen programs into MATLAB® or compatible GNU Octave [29]. This approach imposes some syntac-
tic limitations on the Morphgen language, which would not be required with a conventional compiled implementation,
but it permits rapid prototyping and experimentation with language features. There are two slightly different dialects
of the language for describing either two-dimensional or three-dimensional morphogenetic systems.
(A complete
grammar for a previous version of Morphgen is published in reference [30]; here we use Morphgen version 0.13.)
3.1 Substances
A morphogenetic program is organized into a number of substances with common properties; typical substances
include diffusing morphogens and massive swarms of agents. Substances are similar to classes in object-oriented
programming languages, in that a substance defines the common properties of an unlimited number of particular
instances (called bodies in Morphgen, analogous to objects in object-oriented programming). Like subclasses in
object-oriented programming, more specific substances may be derived from more general substances by specification
of parameters that were unspecified in the more general substance, and by the addition of properties and behaviors in
the derived substance. For example, we might have a substance representing a general diffusible substance with an
unspecified diffusion rate, and one or more specific diffusible substances, with specified diffusion rates, derived from
the general substance.
We distinguish physical substances and controllable substances, but this is more a matter of degree than kind, and
applies primarily to active components. The idea is that the properties and behavior of physical substances are rela-
tively fixed. For example, a particular chemical morphogen will have specific diffusion and decay rates within a given
medium, and a particular agent swarm will be composed of agents, such as microrobots, with particular masses and
specific sensors and actuators. Changing these properties entails making a new substance, new kinds of agents, and
different hardware.
In contrast, a controllable substance has properties that are relatively easy to control. For example, a programmable
agent will permit its sensors and actuators to be controlled so that it behaves in a desired way; its software can be
changed. That is, a controllable substance is in some sense programmable, but there are degrees of programmability.
The current Morphgen language does not distinguish between physical and controllable substances, since controlla-
bility is a matter of degree and even economics. However, the distinction might be expressed in a substance hierarchy.
For example, a substance definition might reflect the physical properties and behavior of a certain kind of microrobot,
and then various substances derived from it could represent the microrobots programmed for distinct functions and
behaviors.
This is an example of a simple substance definition (a diffusible morphogen), which illustrates its parts:
3
A PREPRINT - AUGUST 26, 2019
/ /
c o n c e n t r a t i o n
substance morphogen :
s c a l a r
behavior :
f i e l d C
param d C = 0 . 3
param t C = 10
D C = d C * de l 2 C − C / t C / /
/ /
/ / decay t i m e
d i f f u s i o n c o n s t a n t
c o n s t a n t
d i f f u s i o n & decay
As illustrated by this example, the syntactic extent of many Morphgen constructs is indicated by indenting. (This
substance definition is terminated by a line indented less or equal that of the word substance.) A substance definition
has two parts: field declarations (preceding behavior) and behavioral specifications (following behavior).
3.2 Field Declaration
A substance is characterized by one or more continuous fields defined throughout the space. These may be scalar
fields, such as concentration or density fields, or vector fields, such as velocity or flux fields. The preceding definition
of "morphogen" illustrates the declaration of a scalar field called "C"; an example vector field declaration is:
v e c t o r
f i e l d V
Multiple fields of the same type can be declared on indented lines, as illustrated by this example:
s c a l a r
v e c t o r
C
S
U
V
swarm d e n s i t y
f i e l d s :
/ /
/ / magnitude o f morphogen g r a d i e n t
f i e l d s :
/ / morphogen g r a d i e n t
/ /
swarm v e l o c i t y
3.3 Behavior Specification
Fundamental to our approach is the use of partial differential equations to describe the behavior of massive swarms
of microscopic agents. Such agents and the materials that they control will move continuously in time, but we also
simulate this behavior on ordinary computers in discrete time. Therefore we describe the behavior of substances
by change equations which can be interpreted ambiguously as either ordinary PDEs or as temporal finite difference
equations. The derivation rules of this calculus respect both interpretations. We write change equations with the
notation -- DX = F (X, Y, . . .), which means either the differential equation ∂tX = F (X, Y, . . .) or the difference
equation ∆X/∆t = F (X, Y, . . .). A change equation such as
-- DC = dC∇2C − C/tC
is written as follows in the Morphgen language:
D C = d C * de l 2 C − C / t C
The Morphgen prototype implementation does not include a full expression parser, and so there are some notational
concessions to allow expressions to be handled by the syntax macroprocessor. For the most part, scalars and fields
(both scalar and vector) can be combined using standard arithmetic operators (+, −, *, /). However, products or
quotients of two scalar fields, products of a scalar field and a vector field, quotients of a vector field and a scalar field,
and powers of scalar fields must be surrounded by square brackets, for example, [C * V]. Morphgen includes several
vector operators, including the gradient del X, the Laplacian del2 X, the divergence div X, and the pointwise L2
norm of a vector field, X.
Some change equations are stochastic, either to model indeterminacy and uncertainty in the physical systems, or
to introduce randomness into the swarm control (e.g., to break deadlocks and symmetry). In order to respect both
continuous- and discrete-time stochastic change equations, the morphogenetic programming notation interprets -- DW n
to be an n-dimensional random vector field distributed N (0, 1) in each dimension; it represents n independent sources
of randomness at each point in the (two- or three-dimensional) space [20, 21, 23]. For a d ∈ {2, 3} dimensional
morphogenetic process, we usually use M -- DW n, where M is a d× n matrix that weights and sums the random inputs
into a d-dimensional vector field. M can be a 1 × n matrix to generate a scalar field, or a scalar when n ∈ {1, d}. All
of these cases are written [M DWn] in the Morphgen language.
4
A PREPRINT - AUGUST 26, 2019
Control of morphogenetic swarms often requires them to change their behavior when some quantity exceeds a thresh-
old or is within some range. To accommodate these requirements, Morphgen permits bracketed conditional factors
(actually, Heaviside step functions), which are 0 or 1 depending on whether or not the condition is true. For example,
the following describes a scalar field A in which any value greater than θ increases exponentially until it saturates at
A = 1; values less than or equal to θ decay to zero (with time constant τ):
D A = [A > t h e t a ] A * (1−A) − [A <= t h e t a ] A/ t a u
Sometimes the behavior of a field is described by several partial change equations distributed among several substance
definitions. Partial change equations are written in either of these forms:
D (cid:104)name(cid:105) += (cid:104)expr(cid:105)(cid:104)newline(cid:105)
D (cid:104)name(cid:105) −= (cid:104)expr(cid:105)(cid:104)newline(cid:105)
(This notation was first used in morphogenetic programming by reference [?]p. 20]Fleischer-PhD.) A typical applica-
tion would be describing the behavior of a morphogen. The definition of the morphogen substance might define its
diffusion and decay with a partial change equation such as this:
D M += d M * d el 2 M − M /
t a u C
This is part of the definition of the morphogen because it is a physical property of the substance. Elsewhere in the
program, in the definition of a substance representing the agent swarm (with swarm density S), another partial change
equation could describe the emission of the morphogen at a rate kM :
D M += k M * S
This is part of the behavior of the swarm substance because the morphogen is produced by the agents constituting the
swarm.
It is often convenient to name additional scalars or fields within a substance behavior definition, which is accomplished
by a statement of the form:
let (cid:104)name(cid:105) = (cid:104)expr(cid:105)(cid:104)newline(cid:105)
(In the current prototype implementation, declared fields have global scope, but let-defined variables are local to the
substances in which they are defined.)
The behavior of substances is determined also by certain fixed parameters, such as diffusion and decay rates. These
are specified by a parameter definition, which has the syntax:
param (cid:104)name(cid:105) = (cid:104)expr(cid:105)(cid:104)newline(cid:105)
Parameter definitions usually occur at the beginning of the behavior part of a substance definition, but can also be
global as part of the simulation parameters (see Sec. 3.5 below). A sequence of parameter definitions can be
indented under the word params.
The following behavioral rules [30] illustrate many of the Morphgen features we have discussed (and also line contin-
uation indicated by " ... "):
params :
v = 1
lambda = 0 . 0 3
eps = 1e−100
k W = 0 . 1
k P = 30
t D = 5
r e g u l a t i o n
/ / base swarm s p e e d
/ / d e n s i t y
/ / minimum g r a d i e n t norm
/ / d e g r e e o f
/ / p a t h d e p o s i t i o n r a t e
/ / d e l a y p e r i o d
random motion
l e t U = d el A
l e t S = U
l e t V = [ ( v*U ) / ( S+ eps ) ] − lambda * d el [ ( C−1)2]
D C = [ t >t D ] −div [C*V]
change i n d e n s i t y
/ /
D P += [ t >t D ] k P * [C*(1−P ) ]
+ [ k W DW 2 ]
/ / p a t h d e p o s i t i o n
. . .
5
A PREPRINT - AUGUST 26, 2019
3.4 Bodies
The substance definitions of a morphogenetic program describe the general properties of the materials or substances
(including massive agent swarms) that are involved in the process. A complete morphogenetic program also requires
specification of the initial conditions, which in the case of morphogenesis means an initial preparation of the sub-
stances. Since the goal of morphogenesis is for complex structure to emerge by means of self-assembly, the initial
preparation should be simple in structure: simple spatial arrangements of materials and of sources and sinks.
In our case, the initial conditions are specified by defining a small number of bodies belonging to defined substances.
A body definition has a header followed by a sequence of (indented) initializations, for example:
body O b s t a c l e s of p a t h m a t e r i a l :
f o r ( x , y ) within 0 . 0 6 of
f o r ( x , y ) within 0 . 0 6 of
/ / p l a c e o b s t a c l e s
( −0.1 , 0 . 2 2 5 ) : P = 1
( 0 . 1 , −0.225): P = 1
Initializations currently have the following forms:
(cid:104)initialization(cid:105)
(cid:104)region(cid:105)
(cid:104)init(cid:105)
for (cid:104)region(cid:105) : (cid:104)newline(cid:105) (cid:104)init(cid:105)∗ (cid:104)dedent(cid:105)
::= for (cid:104)region(cid:105) : (cid:104)init(cid:105)
::= (cid:104)expr(cid:105) < (cid:104)name(cid:105) < (cid:104)expr(cid:105),(cid:104)expr(cid:105) < (cid:104)name(cid:105) < (cid:104)expr(cid:105)
((cid:104)name(cid:105),(cid:104)name(cid:105)) within (cid:104)expr(cid:105) of ((cid:104)expr(cid:105),(cid:104)expr(cid:105))
::= (cid:104)name(cid:105) = (cid:104)expr(cid:105)(cid:104)newline(cid:105)
((cid:104)dedent(cid:105) represents delimitation by an indent level less than or equal to the the beginning of the construct.) This
syntax is for the 2D version of Morphgen; the 3D version has similar initializations. The second kind of for definition
permits multiple fields to be initialized in a region, e.g.:
body S t a r t of
swarm
f o r −0.5<x <0.5 , 0<y <0.1:
S = 1
M = 0 . 0 5
/ /
/ /
i n i t i a l
i n i t i a l morphogen c o n c e n t r a t i o n
swarm d e n s i t y
(Uninitialized regions are assumed to be zero.)
3.5 Simulation
The preceding morphogenetic programming constructs apply as well to simulations as to specifying real morpho-
genetic processes executed by physical microrobots. For simulation purposes, the Morphgen language provides addi-
tional facilities. The overall structure of a Morphgen program has the syntax:
morphogenetic program (cid:104)name(cid:105) :
(cid:104)sim params(cid:105)
(cid:104)substance(cid:105)∗
(cid:104)body(cid:105)∗
(cid:104)visualization(cid:105)
end program
The parameters for the simulation are followed by the substance definitions, which are followed by the body defini-
tions, which are followed by visualization commands.
6
The block of simulation parameters have the syntax:
A PREPRINT - AUGUST 26, 2019
(cid:104)sim params(cid:105)
(cid:104)sim par(cid:105)
(cid:104)log params(cid:105)
::= simulation parameters : (cid:104)sim par(cid:105)∗ (cid:104)dedent(cid:105)
::= duration = (cid:104)num(cid:105)(cid:104)newline(cid:105)
::= log params (cid:104)name(cid:105) [, (cid:104)name(cid:105)]∗ (cid:104)newline(cid:105)
temporal resolution = (cid:104)num(cid:105)(cid:104)newline(cid:105)
space (cid:104)num(cid:105) < x < (cid:104)num(cid:105),(cid:104)num(cid:105) < y < (cid:104)num(cid:105)(cid:104)newline(cid:105)
spatial resolution = (cid:104)num(cid:105)(cid:104)newline(cid:105)
save (cid:104)name(cid:105)+ to (cid:104)filename(cid:105)(cid:104)newline(cid:105)
load (cid:104)name(cid:105)+ from (cid:104)filename(cid:105)(cid:104)newline(cid:105)
(cid:104)log params(cid:105)
log note (cid:104)characters(cid:105) (cid:104)newline(cid:105)
The space specification defines the (two- or three-dimensional) region in which the morphogenetic process takes
place; duration determines how long (in simulated time) the simulation runs. The spatial and temporal resolution
of the simulation are defined by the spatial resolution and temporal resolution specifications. The save and
load directives allow scalar and vector fields to be saved and restored. Finally, log params records the specified
parameter values in a time-stamped text file, and log note puts a note in it.
Morphgen provides visualization commands to display fields in a variety of formats either during the simulation
(running) or at its end (final):
(cid:104)visualization(cid:105)
(cid:104)command(cid:105)
::= visualization : (cid:104)command(cid:105)+ (cid:104)dedent(cid:105)
::= display (cid:104)time(cid:105) (cid:104)primitive(cid:105) as (cid:104)kind(cid:105) (cid:104)options(cid:105)
make movie (cid:104)filename(cid:105) of (cid:104)primitive(cid:105) as (cid:104)kind(cid:105) (cid:104)options(cid:105)
(cid:104)stability report(cid:105)
::= running final
::= {mesh contours colors} [limits ((cid:104)expr(cid:105),(cid:104)expr(cid:105))]
::= (cid:104)characters(cid:105) (cid:104)newline(cid:105)
quivers [(cid:104)primitive(cid:105) mesh]
(cid:104)time(cid:105)
(cid:104)kind(cid:105)
(cid:104)options(cid:105)
For example, the following command displays a scalar field C as a heat map during the simulation:
d i s p l a y running C as c o l o r s
The following displays a vector field V as an array of quivers on a 0.2 × 0.2 mesh at the completion of the simulation:
d i s p l a y f i n a l V as quivers 0 . 2 mesh
A running display of a field can be converted to a movie, e.g.,
make movie Cvid . mp4 of C as contours
Finally, Morphgen provides several commands for assessing the numerical stability of the simulation:
(cid:104)stability report(cid:105)
::= report diffusion number for (cid:104)primitive(cid:105)
report Courant number for (cid:104)primitive(cid:105)
report Peclet number for (cid:104)primitive(cid:105) and (cid:104)primitive(cid:105)
4 Example: Routing Neural Pathways
Suppose we wanted to assemble a neuromorphic computer with a complexity comparable to a mammalian brain. Such
a neurocomputer might have billions of artificial neurons, each with many thousands of connections, and it would be
organized into functional regions, each comprising many millions of neurons, with remote regions connected by dense
bundles of millions of neural fibers (artificial axons). In a developing mammalian brain, axons grow toward their
destinations by following chemical signals (morphogens) that indicate way stations and the final destinations. The
stepwise development of a morphogenetic program to solve this problem illustrates our approach to swarm intelligence.
7
A PREPRINT - AUGUST 26, 2019
4.1 Discrete Agents
We begin with a simple path routing process. An agent will be placed at the connection's origin (e.g., the originating
point for a neural fiber), and a source of a diffusing morphogen (the attractant) will be placed at its destination. If the
agent moves up the attractant gradient, then it will find its way to the destination, and it can create in its wake the path
from the origin to the destination. However, we want to create many such paths, and so they cannot go in a straight
line to their destinations, but must weave their way around already created paths. There are at least two ways to
accomplish this. One is to have existing paths emit a repellant morphogen, which the agents try to avoid while seeking
the attractant [26]. An alternative, simpler solution is to have the existing paths absorb or degrade the morphogens, so
that they become sinks for it [25, 28]. Although both solutions work, in this case the simpler solution works better,
since the concentration of diffusing morphogens decreases exponentially with distance from the source [30]. (For
a 3-dimensional diffusion-decay process C = D∇2C − C/τ, the steady-state concentration at a distance r from a
point source with rate k is C(r) = k
Dτ ).) Since the concentration of attractant varies significantly
with distance from the destination, it is difficult, with the first approach, to properly balance it against the repellant
throughout the space. This balance is achieved in the second approach because the existing paths absorb a fraction of
the attractant in their vicinity.
√
4πDr exp(−r/
Figure 1: Neural fiber bundle routing by modified flocking algorithm. There are five bundles, each comprising 5000
fibers, joining randomly selected origins and destinations on the lower and upper surfaces.
8
A PREPRINT - AUGUST 26, 2019
Our goal, however, is not to have single neural fibers connecting (artificial) brain regions, but to have dense bundles of
point-to-point connections. Therefore, rather than having a single agent follow the attractant gradient from the origin
to the destination, we want a large swarm of agents to do so, creating in its wake a bundle of fibers that is tight but not
too tight. Flocks of birds and schools of fish move in an organized way in compact groups, and so we have explored
a modified flocking algorithm as a way of coordinating a swarm to lay down fiber bundles from an origin region to
a destination region [26, 28]. This algorithm has been shown to scale up from five agents (and hence five fibers per
bundle) up to 5000 agents and hence 5000 fibers per bundle, with no change of parameters and a small number of
errors (acceptable in neural computation): scaling over four orders of magnitude [28]. See Fig. 1.
4.2 Continuous Swarm
Next, we take the number of agents in the swarm to the continuum limit. The development and refinement of the
morphogenetic algorithm is described in reference [30]; here we summarize the final algorithm.
The path routing algorithm makes use of four substances: an attractant morphogen, the goal material marking the
destination, the path material laid down by the agents, and the swarm substance, which is composed of agents. The
morphogen substance is defined by a concentration field and a diffusion-decay equation, which describes its physical
behavior:
substance morphogen :
s c a l a r
behavior :
f i e l d A
/ / morphogen c o n c e n t r a t i o n
param d A = 0 . 0 3
param tau A = 100
D A += d A * d el 2 A − A/ tau A
/ /
/ / decay t i m e
d i f f u s i o n c o n s t a n t
c o n s t a n t
d i f f u s i o n + decay
/ /
The change in A is described by a partial change equation because this reflects only its physical diffusion and decay;
it is also emitted by the goal material and absorbed by the path material.
The goal material emits the attractant morphogen (and in this sense is active), but otherwise does not change:
substance g o a l m a t e r i a l :
s c a l a r
behavior :
f i e l d G
param k G = 100
D G = 0
D A += k G * [G*(1−A) ]
/ / d e n s i t y o f g o a l m a t e r i a l
/ /
a t t r a c t a n t
/ / G f i e l d i s
/ / g o a l e m i t s a t t r a c t a n t
r e l e a s e
f i x e d
r a t e
The partial change equation for A describes its emission by the goal material up to saturation at A = 1 and therefore
acts as a source term.
The path material P , which is laid down by the swarm, has two properties: (1) It absorbs the attractant morphogen
(so that new paths avoid existing paths), which is defined by a partial change equation for the morphogen, -- DA −=
P A/τP . Therefore the complete PDE for the morphogen is:
A = dA∇2A − A/τA + kGG(1 − A) − P A/τP .
(2) P increases its concentration autocatalytically if its concentration is above a threshold θP , and it decays if it is
below the threshold, which is described by the partial change equation:
-- DP += aP [P > θP ]P (1 − P ) − [P ≤ θP ]P/tP .
(1)
This autocatalytic behavior sharpens up the paths and ensures they have consistent density, which the equation drives
to P = 1 (path present) or P = 0 (path absent). It is a partial equation because the swarm also deposits path material.
The foregoing behavioral specifications are combined in the path material substance definition:
9
A PREPRINT - AUGUST 26, 2019
/ / p a t h d e n s i t y
substance p a t h m a t e r i a l :
f i e l d P
s c a l a r
behavior :
params :
/ / a u t o c a t a l y t i c
/ / a u t o c a t a l y t i c
a P = 20
t h e t a P = 0 . 3
t h e t a C = 0 . 0 2 / / quorum t h r e s h o l d
/ / p a t h decay t i m e
t P = 1
/ /
t a u P = 0 . 2
r a t e
t h r e s h o l d
c o n s t a n t
a t t r a c t a n t a b s o r p t i o n t i m e
/ / a u t o c a t a l y s i s :
D P = a P * [ P>t h e t a P ] [ P*(1−P ) ] − [ P<=t h e t a P ] P / t P
D A −= [ P*A ] / t a u P
/ / p a t h a b s o r b s a t t r a c t a n t
Finally we define the swarm substance, which controls the mass of agents to lay down a path while approaching the
goal and avoiding existing paths. The principal variable is the swarm density C, which is a real number reflecting the
continuum approximation. The swarm substance also has a vector field V, which defines the velocity of the swarm
throughout space. The change in swarm density, then, is simply the negative divergence of the flux CV:
-- DC = −[t > tD] div CV.
The conditional factor [t > tD] prevents swarm movement until tD time units have passed, which allows the mor-
phogen to diffuse throughout the space.
The velocity field is a weighted combination of three influences: the normalized morphogen gradient, a density control
term, and random (Brownian) motion to break symmetries. The normalized morphogen gradient (or gradient versor)
V1 is ∇A/(cid:107)∇A(cid:107), but to avoid possible division by zero, we compute it: V1 = U/((cid:107)U(cid:107)+), where U = ∇A and is
a small number. To control the density, we define a potential function (C − 1)2 that is minimized at the desired density
C = 1. The density control velocity is the negative normalized gradient of this potential: V2 = −W/((cid:107)W(cid:107) + ),
where W = ∇[(C − 1)2]. A regularization parameter λ controls the relative magnitude of these two velocity fields,
(1 − λ)V1 + λV2, which are combined with Brownian motion kW -- DW 2 and an overall speed v to compute the final
velocity field:1
V = v[(1 − λ)V1 − λV2 + kW DW 2)].
(2)
In Morphgen the substance definition begins like this:
substance swarm :
s c a l a r
v e c t o r
f i e l d C
f i e l d s :
U
V
W
/ /
swarm d e n s i t y
/ / morphogen g r a d i e n t
/ /
/ / d e n s i t y g r a d i e n t
swarm v e l o c i t y
behavior :
params :
v = 1
lambda = 0 . 1
eps = 1e−100
k W = 0 . 3
t D = 5
r e g u l a t i o n
/ / base swarm s p e e d
/ / d e n s i t y
/ / minimum g r a d i e n t norm
/ / d e g r e e o f
/ /
l e t U = d el A
/ / morphogen g r a d i e n t
l e t W = d el [ ( C−1)2]
/ / d e n s i t y g r a d i e n t
l e t V = v * ( ( 1 − lambda ) * [U / ( U + eps ) ]
− lambda * [W/ ( W + eps ) ] + [ k W DW 2 ] )
D C = [ t >t D ] −div [C*V]
t i m e d e l a y
random motion
. . .
Finally and most importantly the swarm deposits path material at a relative rate kP but saturates at P = 1. Therefore
the swarm contributes to the change in P by -- DP += [t > tD]kP C(1 − P ), which begins after tD time units. This is
expressed in Morphgen:
param k P = 30
D P += [ t >t D ] k P * [C*(1−P ) ]
/ / p a t h d e p o s i t i o n r a t e
/ / p a t h d e p o s i t i o n
1 Alternately, the Brownian motion can be described by a diffusion term.
10
A PREPRINT - AUGUST 26, 2019
(a) 2D simulation of path formation, origin at lower right, des-
tination at upper left, 9 obstacles.
(b) 3D simulation of path formation, origin in foreground, des-
tination on back surface, four obstacles. The figure shows re-
gions of path density P > 0.5.
Figure 2: Simulations of neural path formation. In both simulations it can be seen that a swarm sometimes splits to go
around an obstacle. For more information, see reference [30].
To test the algorithm we can define some obstacles and an origin and destination for a new path. Everything we have
discussed so far works for either a two- or three-dimensional morphogenetic system, but there are minor differences
between a 2D and 3D simulation.
For a 2D simulation, the obstacles (representing previously generated paths) are just circular regions of path material.
These can be created by a body definition such as shown in Section 3.4. For 3D simulations there are additional ways
to create obstacles, or the algorithm itself can be used to generate multiple paths, as in the earlier simulations with
discrete swarms.
To initialize the creation of a path, the swarm needs to be placed at the origin, and goal material needs to be placed at
the destination. Here are typical body definitions:
body Cohort of
swarm :
f o r −0.05 < x < 0 . 0 5 , −0.95 < y < −0.9: C = 1
body Goal of g o a l m a t e r i a l :
f o r −0.05 < x < 0 . 0 5 , 0 . 9 < y < 0 . 9 5 : G = 1
A series of simulations exploring algorithm variants and parameter values is described in reference [30]; here we
present typical results in Figure 2.
Figure 2a shows the result of a 2D simulation in which the origin was placed near the lower right corner and the
destination was placed near the upper left. The nine disks represent previous paths, which are obstacles to be avoided.
The new path finds its way around the obstacles, splitting to do so (similar to the bundles in Fig. 1). The degree of
allowed splitting can be controlled by the swarm density regularization parameter λ (Eq. 2). The edges of the bundle
are sharp and the density is constant due to the autocatalysis of path material (Eq. 1).
Figure 2b shows the result of a 3D simulation in which the origin was placed near the foreground and the destination
was placed on the rear surface. The path found its way around four obstacles representing previously generated paths.
The 2D and 3D morphogenetic programs are essentially identical.
5 Example: Body and Leg Segmentation
Our second example illustrates how a swarm of microscopic agents can be coordinated to produce a simple insect-like
robot body with a segmented "spine" and segmented "legs." (A typical 2D result is shown in Fig. 3.) The algorithm
uses a process inspired directly by embryological spinal development but it uses it in two different ways: to segment
11
final P-1-0.8-0.6-0.4-0.200.20.40.60.81x-1-0.8-0.6-0.4-0.200.20.40.60.81y A PREPRINT - AUGUST 26, 2019
the robot's spine (similar to its function in vertebrate development) and to segment the robot's legs (which is not how
insects' or other animals' legs are segmented). Thus it shows how natural morphogenetic processes can be redeployed
for different purposes in artificial morphogenesis.
The process in question is the clock-and-wavefront model of spinal somatogenesis, which was first proposed in 1976
and finally confirmed in 2008 [3, 5]. As the vertebrate embryo develops, a pacemaker in its tailbud periodically
produces a pulse of chemical, a segmentation morphogen. This chemical pulse propagates toward the head of the
embryo, transmitted through the tissue, as each cell is stimulated by its neighbors to produce its own pulse. As each
pulse passes through undifferentiated tissue, it causes the differentiation of one more somite or spinal segment. This
takes place in a sensitive region defined by a low concentration of two morphogens: a caudal morphogen, which
diffuses towards the head from the tailbud, and a rostral morphogen, which diffuses toward the tail from already
differentiated segments. Therefore, the segments differentiate one by one from the head end toward the tail as the
embryo grows. The ratio of the growth rate to the pacemaker frequency determines the length of the segments, while
the product of the frequency and growth duration determines the number of segments (which is characteristic of a
species).
5.1 Growth Process
Our morphogenetic program to assemble the robot body will make use of three different substances. The most im-
portant one is called medium and represents the mass of active components (the swarm) that does most of the work.
This substance has a variety of properties, represented by scalar and vector fields; the most important is its density M;
others will be introduced as required below. When it is part of a defined segment, medium is in a differentiated state,
and the density of differentiated tissue is represented by a scalar field S > 0.
As in biology, differentiation means a change of internal state variables that affects the behavior and other properties of
a substance. In this case, the differentiation into segment tissue could cause the agents to couple with their neighbors,
resulting in a rigid structure, or to secrete some structural substance.
The second substance represents terminal tissue: either the tailbud of the developing spine or the "feet" of the devel-
oping legs; its principal property is a scalar field T , its density. Terminal tissue is composed of active components,
and it is responsible both for growth processes and for generating the pacemaker signals. During the morphogenetic
process, terminal tissue has a velocity: away from the head during spinal development, and away from the spinal axis
during leg development.
Figure 3: Simulation of swarm assembly of insect-like robot body. The body is growing toward the right; the head
segment is visible on the left, and the tailbud is at the extreme right but not shown. Between them, eight segments
have been assembled, each with a pair of segmented legs (the last two pairs incomplete at this point in the simulation,
T = 60).
12
A PREPRINT - AUGUST 26, 2019
(a) Terminal tissue T .
(b) Attractant H.
(c) Undifferentiated tissue M.
Figure 4: Simulation of spine growth process in progress (t = 30). (a) Density T of terminal tissue (moving to
right). (b) Concentration H of attractant emitted by terminal tissue. (c) Density M of assembled undifferentiated
tissue. Agents are introduced by the sources along the upper and lower edges and migrate up the attractant gradient to
assemble into the spine.
The initial state of the morphogenetic process requires the preparation of two bodies: a head segment already in a
differentiated state, and a tail segment adjacent to it. For a 2D simulation, this could be written:
body Head of medium :
f o r 0 . 0 1 < x < 1 , −0.5 < y < 0 . 5 :
M = 1
S = 1
/ / o t h e r
/ / medium t i s s u e
/ /
i n i t i a l i z a t i o n
segment
t i s s u e
body T a i l of
t e r m i n a l :
f o r 1 < x < 2 , −0.75 < y < 0 . 7 5 :
T = 1
/ / o t h e r
i n i t i a l
/ /
i n i t i a l i z a t i o n
t a i l bud
As the terminal tissue moves away from medium tissue, the intervening space is filled with new, undifferentiated
medium tissue. In biological development this is a result of cell proliferation, which expands the mass of undifferenti-
ated tissue and pushes the tailbud further away from the head. In previous versions of our morphogenetic program, we
have given the tail an initial velocity away from the head segment, and as it has moved it has produced undifferentiated
tissue in its wake [22, 24, 25, 26, 28]. This approach makes sense if the agents are able to reproduce themselves,
for example, if they are implemented by genetically engineered microorganisms. If the agents are microrobots, how-
ever, and presumably unable to reproduce, then we need an alternative approach to growth. In this case the agents
can be supplied from an external source and routed to the growth region. As tail tissue moves, it emits an attractant
morphogen into the growth area (the growing gap between the existing body and the tail). Available agents from the
external source follow the attractant gradient into the growth area until they reach a desired density (M = 1 in our
case). Figure 4 shows a simulation of growth in progress.
This approach requires a source of free agents (particles of medium), and so we define a substance source to represent
this process; it has a scalar field J representing the density of source material (J ∈ {0, 1}). The source provides agents
at rate κJ up to saturation (M = 1); therefore, as agents move out of the source regions, more will be introduced to
maintain M = 1 density at the source. This provision of free agents takes place only so long as growth continues,
which is represented by the condition G > ϑG. Therefore the source of new agents is described -- DM + = [G >
ϑG]κJ J(1 − M ). The source substance is defined like this in Morphgen:
substance s o u r c e :
s c a l a r
behavior :
f i e l d J
/ /
/ /
s o u r c e d e n s i t y
param k J = 1
D M += [G > t h e t a G ] k J * [ J * ( 1 − M) ]
i n p u t
r a t e
Source regions must be prepared as part of the initialization of the morphogenetic process. For example, the simulation
in Figure 4 uses the following sources:
body A g e n t S o u r c e s of
s o u r c e :
f o r 5 < x < 6 , 1 . 8 < y < 1 . 9 5 :
f o r 5 < x < 6 , −1.95 < y < −1.8:
J = 1
J = 1 / /
/ / upper
l o w e r
s o u r c e
s o u r c e
13
A PREPRINT - AUGUST 26, 2019
Terminal tissue has the following behavior. Let u be a unit vector field pointing in the direction of desired growth,
and let r be the desired growth rate, which is the same as the speed of terminal tissue movement. Then the change in
terminal concentration is the negative divergence of its flux:
-- DT = − div(T u[G > ϑG]r).
The condition G > ϑG holds so long as growth continues; when it fails, the movement rate is effectively zero and the
terminal tissue stops moving. Figure 4a shows the massive swarm of terminal agents moving to the right.
Terminal tissue emits a growth attractant, with concentration H, up to saturation, which diffuses and decays at specified
rates:
-- DH += κH T (1 − H),
-- DH += DH∇2H − H/τH .
Figure 4b shows the diffusion of the attractant from the terminal tissue.
The swarm of agents M from the sources follows the attractant gradient ∇H at a speed controlled by v and aggregating
up to a maximum density of M = 1:
(3)
Once the agents have assembled in the spine (where M = 1) they do not move again. Figure 4c shows the growth of
undifferentiated spinal tissue in progress.2
-- DM = −v[M < 1] div[M∇H].
5.2 Spine Assembly
The foregoing explains the growth of undifferentiated spinal tissue; we turn now to the clock-and-wavefront process,
which causes the spine to differentiate into segments of specified number and size.
The concentrations of the caudal and rostral morphogens are represented by scalar fields C and R, respectively (Fig. 5).
The rostral morphogen is produced by already differentiated segment tissue (represented by segment density S > 0),
from which it diffuses and decays:
-- DR = κRS(1 − R) + DR∇2R − R/τR.
(4)
2 Other simulation results in this chapter use the original growth process.
Figure 5: Diffusion of C and R morphogens after seven new spinal segments have differentiated (tan color). Blue
color at far right represents C (caudal morphogen) diffusing from the tailbud. Green color around the segments
represents R (rostral morphogen) diffusing from differentiated segments. The α wave visible between the sixth and
seventh segments has just differentiated the seventh segment (which is not yet emitting R). The developing legs are
also emitting C and R morphogens.
14
A PREPRINT - AUGUST 26, 2019
The caudal morphogen concentration C is considered a property of the medium substance, and its diffusion and decay
are physical properties defined by a partial equation in the behavior of medium:
-- DC += DC∇2C − C/τC.
The caudal morphogen is produced from terminal tissue (i.e., tissue with terminal density T > 0), which is part of the
behavior of the terminal substance:
-- DC += κCT (1 − C).
The segmentation signal α is produced periodically in the tailbud and is actively propagated forward through the
medium tissue (Fig. 6). The segmentation morphogen is a property of medium, which defines its physical diffusion
and decay rates:
-- Dα += Dα∇2α − α/τα.
Terminal tissue (T > 0) in the tailbud produces a pulse of α so long as growth is continuing (represented by G > ϑG)
and the tail's pacemakers are in the correct phase (represented by K > ϑK):
-- Dα += κψ[G > ϑG ∧ K > ϑK]T (1 − α).
This partial equation is an extension of α behavior in the definition of the terminal substance. (Additional detail on
the pacemaker and growth duration can be found in reference [28] and Sec. 5.3 below.)
The medium substance also produces α when it is sufficiently stimulated by it (α > ϑα), which allows it to propagate
a wave of α stimulation. After it produces an α pulse, the tissue enters a refractory period (represented by ≥ ϑ),
which ensures that the wave is unidirectional (towards the head). The variable φ (which is only transiently nonzero)
represents the density of medium tissue in this sensitive and stimulated state, and is used both to produce a pulse of α
and to start the refractory timer:
φ = [α > ϑα ∧ < ϑ]M,
-- Dα += κφφ(1 − α),
-- D = φ − /τ.
(5)
(6)
(7)
The actual differentiation of a region of medium tissue into a segment (S > 0) takes place when a sufficiently strong
(α > αlwb) segmentation wave passes through a region with sufficiently low rostral (R < Rupb) and caudal (C <
Cupb) morphogen concentrations. This causes a rapid cς (1 − S) increase in differentiated tissue, which then triggers
autocatalytic differentiation κSS(1 − S) until it is complete (S = 1):
-- DS = κSS(1 − S) + [α > αlwb ∧ R < Rupb ∧ C < Cupb]cς (1 − S)
= (κSS + [α > αlwb ∧ R < Rupb ∧ C < Cupb]cς )(1 − S).
This can be written as follows in the current version of the Morphgen language:
Figure 6: Propagation of α morphogen. Two successive waves are propagating leftward toward the head. The decaying
α pulse in the tailbud is visible at the righthand end.
15
A PREPRINT - AUGUST 26, 2019
c h i = [ a l p h a > a l p h a l w b ] [ R < R upb ] [ C < C upb ] c s e g
l e t
D S = [ ( c h i + k S * S ) * ( 1 − S ) ]
There is a small gap between the differentiated segments because segmentation takes place in a region where
the rostral concentration R has dropped to Rupb (see Figs. 3, 5). To determine the width g of the gap, set the
threshold Rupb equal to the concentration at a distance g from the anterior differentiated tissue:
in a 2D space,
DRτR), where K0 is a modified Bessel function of the second kind (Basset function). Figure 7
Rupb = κR
2πDR
shows a 3D simulation of segmentation using the same morphogenetic program and parameters as the 2D simulation.
√
K0(g/
5.3 Leg Assembly
The process described so far is analogous to the somitogenesis or spinal segmentation that takes place during vertebrate
embryological development. Now however we exploit the process for a different morphogenetic task: the generation
of a pair of segmented legs on each spinal segment (e.g., Fig. 3).
The first task is to control the placement of the legs on the spinal segments; in the case of a 2D simulation, the
only variable is their location along the length of a segment. This can be determined by the relative concentration of
morphogens that diffuse from the anterior and posterior ends of each segment (Fig. 8), but this requires distinguishing
the anterior and posterior ends of each segment: a symmetry breaking process.
The required information is available when the segmentation α wave passes through the tissue, since at that time
the more anterior tissue is emitting the rostral morphogen R, and the more posterior tissue is emitting the caudal
morphogen C. Therefore, the anterior region of a segment further differentiates into anterior tissue (with density A)
when the segmentation wave passes through (α > αlwb) and the rostral morphogen is in the correct range (0.5Rupb >
R > 0.25Rupb in this simulation), which trigger an autocatalytic differentiation process that goes to completion:
-- DA = κASA(1 − A) + [0.5Rupb > R > 0.25Rupb ∧ α > αlwb]cA − A/τA.
Here, we expect R to reflect rostral morphogen diffusing from segments anterior to the currently differentiating seg-
ment. However, the differentiation of A tissue takes place simultaneously with the differentiation of S tissue (as α
passes through), which begins immediately to emit R morphogen. This will cause the entire segment to differentiate
into A tissue, and to prevent this we introduce a state variable B that temporarily blocks production of the rostral
morphogen until anterior tissue differentiation is complete. It is set by the the segmentation impulse (Eq. 5) and then
decays exponentially:
-- DB = cBφ − B/τB + [B > 1](1 − B)/tB.
Figure 7: Three-dimensional simulation of segmentation. Three new segments posterior to the head have been assem-
bled. The figure shows regions in which S > 0.5.
16
A PREPRINT - AUGUST 26, 2019
We choose the time constant τB to allow A differentiation to complete. We replace Eq. 4 for the production of R with:
-- DR = κR[B < θB]S(1 − R) + DR∇2R − R/τR,
which blocks production until B has decayed below θB.
Similarly, the α wave and caudal morphogen in an appropriate range (0.95Cupb > C > 0.8Cupb) trigger autocatalytic
differentiation of posterior border tissue:
-- DP = κP SP (1 − P ) + [0.95Cupb > C > 0.8Cupb ∧ α > αlwb]cP − P/τP .
(Emission of caudal morphogen does not have to be blocked since it comes from the tailbud.)
Once the anterior and posterior border tissues have differentiated, they can do their job of emitting anterior and poste-
rior border morphogens, which provide a reference frame for determining position along each segment:
-- Da = [A > ϑA]κaS(1 − a) + Da∇2a − a/τa,
-- Dp = [P > ϑP ]κpS(1 − p) + Dp∇2p − p/τp.
The thresholds (A > ϑA, P > ϑP ) ensure that only well-differentiated tissue produces these morphogens.
Our plan is for the legs to be assembled on imaginal tissue at the correct location on each segment. The anterior/-
posterior position is defined by anterior and posterior border morphogens in the correct ranges (aupb > a > alwb,
pupb > p > plwb), but these morphogens diffuse throughout the spinal tissue and we want to ensure that imaginal
tissue differentiates only on the surface of the spine. Moreover, this notion of being "on the surface" should be inde-
pendent of the agent size (which is in fact infinitesimal in our continuum model), so that the morphogenetic algorithm
is scale-invariant. Taking a cue from the natural world, in which bacteria and other organisms do quorum sensing to
determine if their population density is sufficient for some purpose, we have our agents estimate their local population
density. If it is near its maximum (M ≈ 1 in our case), then an agent knows it is in the interior; if it is near its minimum
(M ≈ 0) then the agent is relatively isolated. If however the local density is M ≈ 0.5, then the agent knows it is near
the surface; how large M − 0.5 is allowed to be will determine the effective thickness of the surface layer.
This strategy requires that an agent be able to determine its local population density. There are many ways to do this,
but perhaps the most practical is for the agents to emit a slowly diffusing, rapidly decaying morphogen N, which
serves to broadcast density information over a short range. The local morphogen concentration becomes a surrogate
for population density, more precisely, for the density convolved with a smoothing kernel. (See Sec. 6 for more on
implementation of this strategy.) Morphgen code such as the following can be used to define a scalar field E that
represents closeness to the surface (or in the case of 2D simulations, the edge).
Figure 8: Diffusion of a and p morphogens from anterior and posterior tissue, respectively, in each differentiated
segment. Orange color represents a morphogen diffusing from differentiated anterior tissue, and green color represents
p morphogen diffusing from differentiated posterior tissue. White lines on the spinal surface are imaginal tissue I from
which the legs grow. (Two α waves are also visible.)
17
A PREPRINT - AUGUST 26, 2019
D N = k N * [M*(1−N) ] + D N* d el 2 N − N/ tau N / / d e n s i t y
l e t E = [N > 0 . 1 5 ]
[N < 0 . 2 ] S
s u r r o g a t e
/ /
edge marker
Then the autocatalytic differentiation of imaginal tissue can be triggered by a combination of correct morphogen
ranges and being near to the surface:
-- DI = [aupb > a > alwb ∧ pupb > p > plwb]ES(1 − I).
(8)
This equation is written as follows in Morphgen:
D I = [ a upb > a ] [ a > a lwb ] [ p upb > p ] [ p > p lwb ]
[ E* [ S*(1− I ) ] ]
Once the imaginal tissue has fully differentiated, the next step is to initialize the clock-and-wavefront process to grow
and segment the legs. This is accomplished by having the imaginal tissue differentiate into terminal tissue with a
velocity vector directed outward from the spine; thus the imaginal tissue becomes the "foot" of a future leg. We have
used the rapid differentiation of imaginal tissue ( -- DI > ϑDI) to trigger this transformation. The velocity vector is
directed down the S gradient: −∇S/(cid:107)∇S(cid:107). The leg grows just like the spine grows, recruiting free agents from the
sources to fill in the gap between the developing leg and the outward moving foot. The exact same morphogens (H,
R, C, α, a, p) and tissue types (T , S, A, P ) are used to control leg segmentation as to control spinal segmentation.
5.4 Termination
There are two problems with using the same morphogenetic process for the spine and for the legs. The first is that with
the same parameters, the number and size of the segments will be the same in the spine and the legs, which might not
be what we want. The second is that, in the absence of a mechanism to prevent it, the leg segments will develop their
own imaginal tissue, from which "leglets" will grow, and so on, in a fractal manner, which is not our goal. Therefore
we must consider the termination of morphogenetic processes, which in fact is an interesting problem in embryology.
How big should limbs and organs get? What limits growth processes in normal development?
In our morphogenetic program we have adopted a simple mechanism. In several of the equations above, we have seen
that a process continues only so long as a condition G > ϑG is satisfied. Here G is a scalar property of terminal tissue,
which decays exponentially: -- DG = −G/τG, where τG is the growth time constant. If the initial value of G is G0, then
the ϑG threshold will be reached at time t = τG ln(G0/ϑG). If ν is the pacemaker frequency, then (cid:98)ντG ln(G0/ϑG)(cid:99)
complete segments will be generated. The length of the segments will be v/ν, where v is the growth rate (the speed of
Figure 9: Segmentation with higher pacemaker frequency. This simulation has twice the pacemaker frequency νS
as the simulation shown in Fig. 5. Both are shown at the same simulation time, but this has assembled 12 segments
(compared to six in the previous simulation) with ten pairs of legs compared to five previously; the segments are
also approximately half the length as in the previous simulation.
In this simulation νS = 1/π, the previous had
νS = 1/(2π).
18
A PREPRINT - AUGUST 26, 2019
terminal tissue movement). There are a number of parameters here that could be controlled, but if we want the same
processes operating in spine and leg segmentation, then probably the simplest are the initial timer value and pacemaker
frequency. For spinal segments of length λS, we set the pacemaker frequency to νS = v/λS. For NS spinal segments,
we initialize the timer to
(cid:18) NS
νSτG
(cid:19)
.
GS = ϑG exp
See Fig. 9 for the effect of νS on spinal segment length and number. The equations for NL leg segments of length λL
are analogous.
As mentioned above, rapid differentiation of imaginal tissue can be used as a transient impulse to initialize leg growth;
we define it ζ = [ -- DI > ϑDI]cL. This impulse can be used to trigger conversion of imaginal tissue to terminal tissue
and to initialize the leg timers and pacemaker frequencies:
(9)
(10)
(11)
κGL and κν are large so they reset the parameters while ζ > 0. Agents in imaginal tissue also reorient their velocity
vectors down the S gradient, that is, outward from the spine:
-- DT += κLIT (1 − T ) + ζ,
-- DG += κGLIζ(GL − G),
-- Dν += κνIζ(νL − ν).
-- Du += Iκu(−∇S/(cid:107)∇S(cid:107) − u).
There are several ways to prevent the formation of fractal "leglets" on the legs; the most direct is to block the differen-
tiation of imaginal tissue, as given by Eq. 8. The problem is that this equation is part of the behavior of medium tissue,
which is the same in the spine and legs, since its constituent agents come from a common source. Leg terminal tissue,
however, develops from imaginal tissue, but tail terminal tissue does not, which provides a basis for distinguishing the
legs from the spine. Therefore we introduce a variable L representing the density of tissue in the leg state. When the
differentiation of imaginal tissue triggers the differentiation of leg terminal tissue (Eq. 9), this tissue can simultane-
ously differentiate into the L state, which differentiates it from tail terminal tissue: -- DL += κLIL(1 − L) + ζ. As
medium particles arrive from the sources and assemble themselves between the foot terminal tissue and the growing
legs, they must also inherit the L property. This can be accomplished by having L tissue emit a short-range morphogen
µ that causes developing medium tissue to inherit the L property. Therefore, if the concentration of µ in medium tissue
is above a threshold (meaning that it is near "foot" terminal tissue) and it is not spinal tissue, then it enters the L state:
Then we change imaginal tissue differentiation (Eq. 8) so that it takes place only in non-leg tissue (L ≈ 0):
-- DL += κζ[µ > ϑµ ∧ S < 0.5](1 − L).
-- DI = [aupb > a > alwb ∧ pupb > p > plwb ∧ L < ϑL]ES(1 − I).
Finally, note that the spinal tissue between the last segment and the tailbud is undifferentiated because the caudal
morphogen concentration C is too high. Therefore it does not develop legs and forms the tail proper between the leg-
bearing segments and the tailbud. To determine the tail's length λ, set the caudal morphogen concentration at a distance
λ in front of the tailbud to the caudal morphogen threshold for segmentation: in 2D, Cupb = κC
DCτC).
2πDC
Also, as can be seen in Figs. 3, 7, and 9, the first spinal segment posterior to the head can have a different length
compared to the other segments; this is because its length depends on the initial phase of the pacemaker.
K0(λ/
√
6 Toward a Morphgen Compiler
6.1 Agents as Substantial Particles
Our approach to morphogenetic engineering treats agent swarms and other substances from the perspective of contin-
uum mechanics. That is, although our swarms and tissues are composed of finite numbers of discrete elements of finite
size, our intended application is to very large numbers of very small elements, and so the continuum approximation is
useful. From the continuum mechanics perspective, tissues and massive swarms are infinitely divisible into infinites-
imal material points or particles representing differential volume elements of continua [28]. Therefore, a material
point or particle does not represent a single agent (or molecule or other active or passive component), but a very tiny
volume containing many such physical elements. As a consequence, thinking about the individual behavior of parti-
cles in such a continuum is helpful in understanding how the individual agents should behave, but it is not sufficient
to equate particles and agents. For example, a particle might have a distribution of orientations or velocities, or be in
19
A PREPRINT - AUGUST 26, 2019
a mixture of differentiation states, whereas individual agents have definite orientations, velocities, and differentiation
states [20, 21, 23].
We have seen how PDEs can be used to describe the motion, differentiation, and other behavior of swarms of micro-
scopic agents to assemble complex shapes, but these PDEs cannot be used directly to control the agents. These PDEs
have been expressed in an Eulerian (spatial) reference frame, in which the derivatives are relative to fixed spatial
locations and describe the changes of variables at those locations as particles flow through them. For example, we
may be concerned with the change in temperature q(t, p) at a particular location p, that is, ∂q(t, p)/∂t, as particles
with different temperatures flow through it. More relevant for the control of agent swarms is a Lagrangian (material)
reference frame, in which derivatives are relative to fixed particles and describe how the properties of those particles
change as they move through space. Thus we might be more concerned with the temperature Q(t, P ) of a particle P
and how it changes as that particle flows through space, ∂Q(t, P )/∂t; the latter is called the material or substantial
derivative and is commonly written DQ/Dt.
Since a particle's position is a function of its velocity v, the material derivative can be expanded by the chain rule:
∂q
∂t
For a vector property of particles, the formula is analogous:
DQ
Dt
=
+ v · ∇q.
DQ
Dt
=
∂q
∂t
+ v · ∇q,
where ∇q is a second-order tensor. (In a Cartesian coordinate frame (∇q)jk = ∂qj/∂xk and v · ∇q = vT∇q.)
6.2 Morphogen-based SPH Control
Morphgen describes continuous fields of infinite numbers of infinitesimal agents. A compiler for the Morphgen
language would produce code for given numbers of agents of given finite size, and would thus realize the scalability
potential of the language's PDE approach. Such a compiler would be an example of what some authors have termed
"global-to-local (GTL)" compilation [14]. GTL compilation seeks to turn on its head the usual challenge of emergence:
It seeks not to predict collective behavior from simple rules for individuals, but to derive such rules based on desired
collective behavior. In this section we explore a novel, embodied variation of smoothed particle hydrodynamics (SPH)
swarm robotic control as a partial basis for a GTL compiler for Morphgen.
SPH itself is a meshfree Lagrangian numerical method used in physics simulations and was introduced by reference
[12] and reference [19]. SPH was first proposed as a method for controlling robot swarms by reference [37]. In SPH
robotic literature to date, researchers have used SPH to cause swarms to emulate fluids or bodies much like fluids (e.g.,
[38, 10, 31, 40]). Such use is in keeping with typical use of SPH in physics to simulate fluids. However, we believe
the true potential of SPH in robotic control lies in its ability to simulate non-physical PDEs. A step toward this goal
was taken by reference [36], who design swarms that emulate fluids, but with non-physical variation over time in fluid
parameters (see reference [44] for a more recent paper continuing this work). Our work goes further, applying SPH to
Morphgen programs that do not resemble physical fluid behavior. This is one key contribution that our work makes
with respect to the SPH robotic control literature, which we address further in Sec. 6.8. However, we first discuss our
other key contribution, a proposal for overcoming a communications challenge posed by traditional SPH robotics.
6.3 Natural Smoothing Functions
If an agent can produce a physical field around itself having certain properties, then implicit in this field is what we
term a natural smoothing function (NSF). We will refer to such physical fields as "NSF fields" to distinguish them
from the potentially more abstract fields which NSF fields may be used to implement, as described in the next section.
The NSF field around a given agent must satisfy the following:
• The strength or intensity (we will say value) of the NSF field must decrease smoothly and monotonically with
distance from the agent.
• The NSF field's value must approach zero with distance, and quickly enough that the integral of value over
all space converges.
• The NSF field must be roughly symmetrical around the agent.
• The agent must have control over the amplitude of the NSF field independently of its shape.
20
A PREPRINT - AUGUST 26, 2019
Finally, the NSF fields of different agents must sum naturally in the environment, and each agent must be able to sense
locally both the value and gradient of the summed NSF field. If these requirements are met, then groups of such agents
can implement SPH robotic control. The first three of these requirements are derived from general requirements on
SPH smoothing functions; see reference [18] for a discussion. Another key requirement is that the integral of the
smoothing function be unity, which we address in the next section.
In later sections, we explore one possibility in depth: that of agents that release and detect dilute substances that diffuse
and degrade in a surrounding aqueous medium. This scenario is inspired by the cells of a developing embryo as it
undergoes morphogenesis. We refer to the smoothing functions implicit in this mechanism as "morphogen smoothing
functions" (MSF). A second possibility, not discussed further in the present work, is that of agents producing and
sensing electromagnetic or acoustic NSF fields distinguished by frequency. Before addressing MSFs specifically, we
discuss the NSF framework generally.
6.4
Incorporating SPH Estimates into the NSF Framework
In contrast to the term "NSF field," we use the term "field" by itself to describe more general, and potentially abstract,
entities. These are the fields which traditional SPH seeks to simulate and which SPH swarm robotic control seeks to
generate. Examples in the context of artificial morphogenesis are density, velocity, or differentiation state of a tissue
or agent continuum.
A diversity of SPH-based estimates of fields and their derivatives exist. In general, an SPH estimate at a point is the sum
over neighboring particles of a smoothing function, or one of its derivatives, multiplied by some expression involving
various spatially-varying quantities. Each of these quantities (e.g., mass, density, or velocity) is associated either with
the location of the desired estimate or with the location of a given neighboring particle. Some SPH estimates require
only quantities associated with the locations of neighboring particles, and such forms permit ready incorporation into
the NSF framework. The most important example is the estimate of a field itself at location x:
(cid:104)f (x)(cid:105) =
mj
ρj
f (xj)Wj(x),
(12)
where mj, ρj, and f (xj) are the mass of, local density at, and field value at, each neighboring particle j.
We can perform this estimate in the NSF framework if we assume that each agent can produce an NSF field as described
in Sec. 6.3. If an agent j controls the amplitude of its NSF field so that it has unity integral, then the measurement by
an agent i of the value of the NSF field produced by agent j is equivalent to the evaluation of Wj(xi), where Wj(x)
is a smoothing function whose shape is determined by the physical nature of the produced NSF field. If instead agent
j controls the amplitude of its NSF field so that its integral is mj
f (xj), then the measurement by agent i of agent
ρj
j's NSF field is the entire term mj
f (xj)Wj(xi). Because the NSF fields of each neighbor j sum naturally in the
ρj
environment, the total NSF field value measured locally by agent i is the desired local estimate of a field (cid:104)f (xi)(cid:105).
(This procedure requires agents to estimate their local densities, which they can do through a similar NSF estimate
based on (cid:104)ρ(xi)(cid:105) =(cid:80)
j mjWj(xi).)
Because Eq. 12 is valid at points between particles as well as at particles, a spatial derivative of the field of SPH
estimates is an estimate of that derivative in the corresponding field function. For example,
(cid:104)∇f (xi)(cid:105) = ∇xi
mj
ρj
f (xj)Wj(xi).
(13)
In traditional SPH, where a separate calculation is needed for an SPH estimate at each point, the above fact is not
directly useful. Instead, much of the value of traditional SPH comes from the ability to rearrange expressions so that
derivatives are applied only to the smoothing function and can be taken analytically. For example,
(cid:88)
j
(cid:88)
j
(cid:88)
j
(cid:104)∇f (xi)(cid:105) =
f (xj)∇xiWj(xi).
mj
ρj
By contrast, in the NSF framework, SPH estimates are physically embodied by the total NSF field at a given point.
Therefore, if technologically feasible, an agent can make an SPH estimate of a spatial derivative by sensing that
derivative in the local NSF field. In particular, we assume agents can sense NSF field gradients locally, for example
using differences between multiple sensors at their boundary, such that Eq. 13 is directly useful.
Where such direct estimates are not practical, it may be possible to manipulate other methods from the SPH literature
into forms compatible with the NSF framework. The Laplacian operator is an important example. reference [15]
21
A PREPRINT - AUGUST 26, 2019
(cid:88)
derive an SPH form of the Laplacian operator requiring no derivatives. In two dimensions:
(cid:104)∆f (xi)(cid:105) =
2
α
(f (xj) − f (xi)) Wj(xi)
mj
ρj
,
where α is a constant that depends on an integral relating to a given smoothing function. Although it is not possible in
is just an
the NSF framework to take the pairwise difference as shown, we can rearrange and note that(cid:80)
f (xj)Wj(xi) −(cid:88)
SPH estimate of unity, and so can be replaced by 1. This gives us a form we can calculate in the NSF framework:
(cid:104)∆f (xi)(cid:105) =
j Wj(xi) mj
ρj
f (xi)Wj(xi)
j
(cid:88)
(cid:88)
j
j
mj
ρj
2
α
=
=
2
α
2
α
((cid:104)f (xi)(cid:105) − f (xi)) .
mj
ρj
j
f (xj)Wj(xi) − 1 · f (xi)
mj
ρj
(14)
6.5 Benefits of NSF Framework
NSF-based SPH robotic control addresses communication challenges and inefficiencies in traditional SPH robotic
control. Traditional methods require each agent, for each SPH estimate it makes, to communicate separately with
O(n) other agents, where n is the number of agents in its neighborhood. In contrast, NSF-based agents need only
sense an NSF field value locally, a O(1) operation. In establishing separate data links with its neighbors, a traditional
agent must avoid communication interference with its neighbors, all of which may be simultaneously trying to establish
links with their neighbors. By contrast, NSF-based agents communicate implicitly by sensing an NSF field locally and
manipulating their contribution to it; therefore, they do not require any explicit data links or even a concept of having
individual neighbors.
We can view these advantages over traditional SPH control as the benefits of embracing an embodied approach to
artificial morphogenesis. In the agent-to-agent communication of traditional SPH control, information is an abstraction
atop the physical fields chosen as communication media. Overcoming communication interference is thus a struggle
against the way these fields interact in space. By contrast, the NSF framework works with, rather than against, the
interaction (as well as diminution with distance) of the physical NSF fields available to a system. For example,
morphogen-based SPH control takes advantage, in two key ways, of the behavior of morphogens in the aqueous
environments typical of a growing embryo. First, it offloads pairwise distance estimates to the natural diminution with
distance of a dilute substance as it diffuses and degrades. Second, it offloads a summation loop across many neighbors
to the natural summation of dilute substances from multiple sources.
6.6 Morphogen Smoothing Functions
This section explores the use of morphogens to implement NSFs as a basis for SPH control. We refer to such NSFs
as MSFs, as mentioned in Sec. 6.3. We first discuss the properties of these MSFs with respect to SPH generally.
Afterward we derive the morphogen production rates needed to implement SPH control using MSFs.
Throughout this section, we make several simplifying assumptions:
• Each agent is a point.
• Each morphogen obeys Fick's law for homogeneous, isotropic diffusion (i.e., change in concentration is a
constant proportion of the Laplacian).
• Degradation of each morphogen is uniform in the medium and proportional to morphogen concentration.
• Each agent is unmoving.
• Morphogen production, diffusion, and degradation are in equilibrium.
Later, we introduce corrections for some violations of the latter two assumptions.
For our embodied SPH approach to work, we need the SPH estimate of a field variable at a point to equal the summed
concentration of an associated morphogen at that point:
mj
ρj
f (xj)Wj(x) =
(cid:104)f (x)(cid:105) =
(cid:88)
(cid:88)
cj(x),
(15)
j
j
22
A PREPRINT - AUGUST 26, 2019
Table 1: Idealized morphogen smoothing functions. The row labeled "Physical, normalized" refers to the physical part
of the general solution to the relevant PDE, normalized so that its integral over space is unity. r is the magnitude of
displacement. J0 and Y0 are Bessel functions of the first and second kind respectively, of order zero. K0 is a modified
Bessel function of the second kind, of order zero. E is diffusion rate and k is decay rate, and in the final row we make
the substitution h =
(cid:113) E
Radial form of PDE E ∂2ρ
General solution
k to obtain forms resembling the typical presentation of SPH smoothing functions.
One dimension
∂r2 − kρ = 0
√
ρ(r) = c1e−
√
Three dimensions
2 E
r
Two dimensions
E
r
∂r + E ∂2ρ
ρ(r) = c1J0
ρ(r) = c1
k
E r
∂ρ
e
r
ir
(cid:18)
(cid:19)
(cid:113) k
∂r2 − kρ = 0
(cid:19)
(cid:18)
(cid:113) k
(cid:19)
(cid:18)
(cid:113) k
−ir
E
E
+ c2Y0
r
k
E
∂ρ
∂r2 − kρ = 0
∂r + E ∂2ρ
√
−
√
k
r√
E
kr
√
−
+ c2
e
r
k
E
r
Physical, normalized ρ(r) =
ρ(r) = k
2πE K0
r
E
ρ(r) = k
4πE
e
With h =
ρ(r) = e
− r
h
h
ρ(r) =
K0( r
h )
2πh2
ρ(r) = e
− r
h
4πh2r
(cid:113) E
k
+ c2e
(cid:113) k
k
E r
√
E e−
k
E r
where cj(x) is the morphogen concentration at x attributable to neighboring agent j. Each agent must therefore choose
to produce this morphogen at a rate rj such that
Let Cj be the total amount of morphogen in the environment due to agent j, at equilibrium:
(16)
(17)
(18)
If k is the morphogen's decay rate, then Cj is related to rj by
Ω
and therefore reaches equilibrium when
cj(x) =
f (xj)Wj(x).
mj
ρj
(cid:90)
Cj =
cj(x)dx.
dCj
dt
= rj − kCj
Cj =
rj
k
.
Solving Eq. 16 for Wj(x), integrating both sides, and considering the requirement of unity for the integral of Wj(x),
we have
Wj(x)dx =
mjf (xj)
Ω
cj(x)dx =
Cj =
mjf (xj)
kmjf (xj)
ρj
ρj
rjρj
= 1.
(cid:90)
Ω
(cid:90)
We can now solve for rj to determine the correct production rate for agent j:
rj =
kmjf (xj)
ρj
.
(19)
Under these simplified assumptions, the SPH smoothing functions implied by this production rate are described in
Table 1.
6.7 Correcting for Transient Spurious Morphogen Smoothing Functions
Traditional SPH assumes no lag time in information flow, but two factors prevent morphogens from conveying infor-
mation instantaneously among agents. The first is the time needed for production and degradation processes to reach
steady state, which is a non-spatial consideration. The second is the time needed for diffusion to spread information
spatially from an agent to its neighbors. These two factors lead to transient spurious MSFs, and we discuss each in
turn. In neither case do we fully correct for transient effects; instead, we make simplifying assumptions that allow us
to correct for the most detrimental aspect of each effect.
23
6.7.1 Non-equilibrium of Production and Degradation Rates
A PREPRINT - AUGUST 26, 2019
We first address the non-spatial transient effect of morphogen production being out of equilibrium with degradation,
which occurs whenever production rates have recently changed. We make two simplifying assumptions to find a
correction term for agents to add to their sensor readings to better approximate what those readings would be under
equilibrium conditions. First, we neglect diffusion time; in other words, we assume that each agent instantaneously
injects morphogen into its entire neighborhood according to the distribution of the idealized MSF. Second, we assume
that the field variable values for all agents in a neighborhood are changing at the same rate a so that f (xj, t) = at + bj
since time t = 0, and that until t = 0 the values were not changing and the system was in equilibrium.
Based on these assumptions, we wish to find a correction term g(t) such that (cid:104)f (x, t)(cid:105) = (cid:80)
j cj(x, t) + g(t). At a
given time t > 0, from Eq. 17 and Eq. 19 the total mass of morphogen in the environment due to agent j obeys
This equation is solved by
dCj(t)
dt
= rj(t) − kCj(t) =
Cj(t) = αe−kt +
mj
ρj
(cid:16)
(cid:17)
at + bj − a
k
kmj(at + bj)
− kCj(t).
ρj
.
(20)
Because the system was in equilibrium at t = 0, we can substitute from Eq. 18 and Eq. 19:
Cj(0) =
rj(0)
k
=
mjf (xj, 0)
=
mjbj
ρj
.
After substituting into Eq. 20 and solving for α, we have:
at + bj − a
k
e−kt +
Cj(t) =
mja
ρjk
mj
ρj
(cid:16)
ρj
=
(cid:17)
which we can solve for f (xj, t):
f (xj, t) =
ρj
mj
Cj(t) +
a
k
− mja
ρjk
,
(21)
mjf (xj, t)
e−kt +
mja
ρjk
ρj
(cid:0)1 − e−kt(cid:1) .
(cid:0)1 − e−kt(cid:1) Wj(x)
(cid:88)
(cid:104)f (x, t)(cid:105) =
Substituing Eq. 21 into Eq. 12 and rearranging, a correct SPH estimate of f (x, t) under our non-equilibrium assump-
tions would therefore be
The left summation of Eq. 22 is the SPH estimate of f (x, t) under the equilibrium assumption and is also(cid:80)
j cj(x, t),
the sensed morphogen concentration at x (because we derived our production rate under the equilibrium assumption).
The right summation of Eq. 22 is just the SPH estimate of a
k
estimate is necessary for this term:
(cid:88)
(cid:0)1 − e−kt(cid:1), but this does not depend on j, so no SPH
Cj(t)Wj(x) +
mj
ρj
(22)
a
k
j
j
(cid:104)f (x, t)(cid:105) =
Thus we have our desired correction term:
(cid:0)1 − e−kt(cid:1) .
(cid:88)
a
k
Cj(t)Wj(x) +
j
g(t) =
a
k
(cid:0)1 − e−kt(cid:1) .
(23)
We wish for each agent to keep track of g(t), which is possible so long as it maintains an estimate of ∂f (x,t)
. (In
our simulation we simply have each agent remember one previous value of f (x, t) which it can divide by a known
dt = ae−kt. Note that we can rearrange Eq. 23 to get
timestep.) To see this, we first take a time derivative: dg(t)
e−kt = 1 − k
a g(t). We substitute this into the derivative along with a = ∂f (x,t)
to obtain
∂t
∂t
which implies the needed update rule for agents to follow (in our simulation, ∆g(t) = ∆f (x, t) − kg(t)∆t).
dg(t)
∂f (x, t)
=
dt
− kg(t),
(24)
∂t
24
A PREPRINT - AUGUST 26, 2019
6.7.2 Transient Effects of Diffusion Rate
The time it takes for morphogen diffusion to convey information spatially also leads to violations of steady-state
assumptions whenever the position or production rate of an agent has recently changed. Here we address only transient
effects due to agent motion, because in our simulations it appears to be more detrimental than spatial effects due to
production rate changes. As an agent moves, its contribution to the total morphogen field becomes stretched behind
and compressed in front, with more complex distortions if the agent's path curves. Even more complex effects arise
as a result of the particularities of an agent's instrument positions and the way a given medium flows around an
agent's body. Because such particularities imply implementation-dependence of distortions, we do not pursue analytic
corrections. Instead, we propose that agents be calibrated to correct for some of this distortion, an approach we test in
simulation.
The most damaging aspect of this distortion is the effect that an agent's own contribution to the morphogen field has
on its gradient estimates. In addition, an agent has no knowledge of its neighbors' individual positions and velocities,
limiting its ability to correct for distortions in neighbors' contributions to the morphogen field. For these reasons we
focus calibration on correcting for distortions in an agent's own contributions to its gradient estimates.
The central idea of calibration is to have isolated agents move at constant velocities and, once steady-state relative to
the agent is reached, to record the measured gradient. A table of gradients for different speeds can be created in this
way, and during operation these gradients (with appropriate interpolation) can be subtracted from measured gradients
during operation to partially correct for motion-related asymmetry in the MSF. For our simulation, we repeat this
procedure with different headings to average out numerical effects related to the finite difference method grid used for
diffusion.
6.8 Beyond Fluid Dynamics: SPH as a Partial Morphgen Compiler
SPH robotic control generally, and our NSF variant in particular, can be viewed as a compilation technique for Morph-
gen code. The utility of SPH as a compilation technique is limited in the present work by the requirement that swarm
∂t = −∇ · (ρu). This trivially holds when swarm density is
density fields must obey the mass continuity equation ∂ρ
specified in Morphgen implicitly by velocity or acceleration fields, but may not hold when swarm density fields are
specified directly. For example, a nonconservative equation such as ∂ρ
∂t = 5 cannot yet be compiled with our method.
Other fields representing swarm state are not limited in this way. In future work we plan to relax this requirement
through agent recruitment and removal or through cell-like division and apoptosis.
SPH compilation introduces a layer of abstraction or virtualization between what we might think of as Morphgen
"application" code and what we envision as a library of physical morphogens, which can also be described in Morph-
gen. Consider a Morphgen program that specifies a velocity or acceleration field for a swarm, along with several
other fields. These other fields may be thought of as representing tissue differentiation, gene expression, or even mor-
phogens, as seen in earlier examples. To compile such a program into agent code, all non-swarm fields are treated as
agent state, and a unique physical "library" morphogen is associated with each. (In practice, a programmer might find
it convenient to specify these associations, because physical morphogens with different ratios of diffusion to degra-
dation rates might prove more robust for implementing different fields. However, such low-level associations are not
essential in principle: Eq. 19 specifies the correct production rate regardless of diffusion and degradation rates.) As
described in Sec. 6.6, these physical morphogens implicitly provide the smoothing functions needed for each agent to
make SPH estimates of a given field and its derivatives.
To further illustrate this layer of abstraction, consider a Morphgen application program that specifies what the pro-
grammer conceptualizes as a morphogen (for example, if the Morphgen code specifies a diffusion and degradation
rate as well as sources and sinks). The SPH compilation process would nonetheless treat this field as a swarm state
variable and associate it with a physical morphogen, whose physical parameters are theoretically independent of the
abstract morphogen described by the application code. This layer of abstraction frees the programmer from the need
to find a physical substance with particular parameters. Instead, we envision engineering a small library of physical
morphogens that can be repurposed to implement a variety of fields in different Morphgen programs.
As mentioned in Sec. 6.2, SPH swarm robotic control has focused on enabling swarms to behave somewhat like fluids,
as described by the Navier-Stokes or Euler equations. Swarms have therefore been specified with acceleration fields,
and other fields have been designed to simulate such physical quantities as energy and pressure. In the context of our
current work, we prefer to reframe these applications as special cases of SPH compilation of implicit Morphgen-like
programs that happen to describe roughly physical fluid-like behavior. (This is a simplification; for example, reference
[38] derive one force using SPH, and add to it a body force derived separately from a potential field.)
25
A PREPRINT - AUGUST 26, 2019
In comparison to past fluid-related SPH swarm control, the example that follows illustrates some of this potential
generality of SPH compilation. Rather than specifying physics-inspired quantities such as energy or pressure, arbitrary
fields useful in solving a path-finding problem are specified. And rather than specifying acceleration, velocity is instead
specified, making swarm behavior more similar to flocking than to Newtonian dynamics. (Velocity-based motion is
also a natural fit for fluid media with low Reynolds numbers such that propulsion force is approximately proportional
to velocity rather than to acceleration. However, agents could integrate or differentiate on their own to achieve,
respectively, acceleration-based control where force is proportional to velocity or velocity-based control where force
is proportional to acceleration.)
6.9 Example: Path-finding
To illustrate our method, we simulate a two-dimensional environment that is inspired by, but does not rigorously
implement, an aqueous environment with low Reynolds number, as would be found in a developing embryo. Agents
are modeled as discs. Sensors are placed around the surface at 90° intervals, with morphogen production instruments
on the surface halfway between each pair of sensors. Agents have full control over their velocity, but no propulsion
mechanism is explicitly modeled; agents' locations are simply changed. As agents move, the surrounding medium
and morphogens flow around the agents' bodies. This flow is qualitatively similar to laminar flow but does not strictly
obey fluid dynamics. Motion of the aqueous medium and morphogens is calculated using a finite difference method.
Some random noise is applied to agents' motion to qualitatively simulate Brownian motion, and some bias and noise
is applied to sensor readings.
In the Morphgen program below, (cid:104)T(cid:105) refers to the SPH estimate of T , as opposed to an agent's internal stored value.
substance swarm :
scalar fields :
ρ // swarm density
A // environmental cue for lower-left square
B // environmental cue for upper-right square
S // to diffuse from lower-left square
T // to follow gradient of S back from upper-right square
vector fields :
V // swarm velocity
behavior :
params :
θρ = 0.02
θA = 0.2
θB = 0.2
θT1 = 0.1
θT2 = 0.05
κS1 = 500
κS2 = 500
κS3 = 1
κS4 = 100
κT1 = 500
κT2 = 600
κT3 = 0.002
κT4 = 60
κV = 0.0003
κVmax = 0.003
26
A PREPRINT - AUGUST 26, 2019
(cid:18) ∇S · ∇T
(cid:19)
let φ = arccos
-- DS −= [ρ < θρ ∨ B > θB]κS1S
-- DS += [A ≤ θA](cid:0)κS3∇2S − κS4 S(cid:1)
-- DS += [A > θA ∨ S > 1]κS2 (1 − S)
(cid:107)∇S(cid:107)(cid:107)∇T(cid:107)
-- DT += [B > θB]κT1(1 − T )
-- DT += [(cid:107)∇S(cid:107) > 0 ∧ (cid:107)∇T(cid:107) > 0 ∧ (cid:104)T(cid:105) > θT1]
-- DT += κT3∇2T
-- DT −= κT4T 2
-- DT += (V · ∇T )[(cid:104)T(cid:105) > θT2 ∧ T < θT1]
− ρ∇T
(cid:19)
let V0 =
κV 2T ln
1
ρ
(cid:18)
1 − φ
π
(cid:19)6
κT2
(cid:18) 8ρ
(cid:19)(cid:18)∇ρ
(cid:18)(cid:107)V0(cid:107)
(cid:19) V0(cid:107)V0(cid:107)
κVmax
T
T
T 2
V = κVmax tanh
Simulation of the above Morphgen program is shown in Table 2. All PDEs and simulation parameters, including
starting density, are the same for all three runs shown, with only agent diameter and number varying. To hold starting
density roughly constant across the eightfold difference in agent diameter, agent number varies about 64-fold. The
qualitative similarity across this range demonstrates the potential of Morphgen and SPH to facilitate scalable swarm
dynamics.
We describe some of the key components of this program, which aligns agents along the shortest path between two
distinct environmental cues. Agents that sense a cue A represented by the lower-left square set a high value for the
variable S. Other agents implement a diffusion-decay equation to establish the field distribution of S. (This equation
is independent of the diffusion-decay equation that models the physical behavior of the morphogen that implements
S in the SPH scheme.) Agents that sense a second cue B represented by the upper-right square as well as a non-zero
gradient of S set a high value for variable T . Other agents increase their value of T depending on how nearly opposite
are the gradient directions of S and T and whether a neighborhood estimate of T is above a threshold. Otherwise, T
diffuses slowly. Agents follow the gradient of a function involving both T and local density such that a target density
is achieved where T is high. Agents' speed saturates with the tanh function so that dynamics are similar to flocking
behavior when far from optimum arrangement but approach rest when closer to that optimum.
Because of the Lagrangian nature of SPH, field variables are naturally carried along with agents, which may or may
not be desired. At low speeds and with field dynamics that are somehow (as in this program) tethered to environmental
cues, the distinction is often subtle in practice. When it is desired that Morphgen PDEs refer to an Eulerian frame,
advection can be added counter to agent velocity, as reflected by -- DT + = V · ∇T in the program above. In this
program, we only add this counter-advection where it helps the growing T region to continue its growth against the
opposing motion of agents following the T gradient.
7 Conclusions
We have argued that embryological morphogenesis provides a model of how massive swarms of microscopic agents
can be coordinated to assemble complex, multiscale hierarchical structures, that is, artificial morphogenesis or mor-
phogenetic engineering. This is accomplished by understanding natural morphogenetic processes in mathematical
terms, abstracting from the biological specifics, and implementing these mathematical principles in artificial systems.
As have embryologists, we have found partial differential equations and continuum mechanics to be powerful math-
ematical tools for describing the behavior of very large numbers of very small agents, in fact, taking them to the
continuum limit. In this way we intend to have algorithms that scale to very large swarms of microrobots.
To this end we have developed a PDE-based notation for artificial morphogenesis and designed a prototype morpho-
genetic programming language. This language permits the precise description of morphogenetic algorithms and their
automatic translation to simulation software, so that morphogenetic processes can be investigated.
We illustrated the morphogenetic programming language and morphogenetic programming techniques with two exam-
ples. The first addressed the problem of routing dense bundles of many fibers between specified regions of an artificial
27
A PREPRINT - AUGUST 26, 2019
Table 2: Path-finding example. Agents use SPH to establish path between yellow and turquoise squares.
1186 agents,
diameter 0.006
10675 agents,
diameter 0.002
166 agents,
diameter 0.016
Time = 0
Time = 20
Time = 70
Time = 270
28
A PREPRINT - AUGUST 26, 2019
brain. Inspired by axonal routing during embryological development, we used a modified flocking algorithm to route
fiber bundles between origins and destinations while avoiding other bundles. Simulations showed that this algorithm
scaled over at least four orders of magnitude, with swarms of 5000 agents. Then we took the number and size of agents
to the continuum limit and showed how morphogenetic programming could be used to coordinate a massive swarm of
agents to lay down a path between designated termini while avoiding obstacles.
Our second example showed how a natural morphogenetic model -- the clock-and-wavefront model of spinal
segmentation -- could be applied in morphogenetic engineering both in a similar context -- creating the segmented
"spine" of an insect-like robot body -- and for a different purpose -- assembling segmented legs on the robot's spine. A
massive swarm of microscopic agents is supplied from external sources and guided to the assembly sites, where they
begin a process of differentiation coordinated by the emission of and response to several morphogens. As demon-
strated in simulation, this reasonably complex process can be controlled to assemble a structure with a spine and legs
with specified numbers and sizes of segments.
Finally, we showed how an embodied variation of smoothed particle hydrodynamics (SPH) swarm robotic control can
be applied to the global-to-local compilation problem, that is, the derivation of individual agent control from global
PDE specifications. By these means, available physical morphogens can be used to implement the abstract morphogens
and other fields required for a morphogenetic process. The physical morphogens define natural smoothing functions
that can be used in SPH to estimate the concentrations, gradients, and other functions of morphogenetic fields.
29
A PREPRINT - AUGUST 26, 2019
References
[1] Harold Abelson, Don Allen, Daniel Coore, Chris Hanson, George Homsy, Thomas F. Knight, Jr., Radhika Nag-
pal, Erik Rauch, Gerald Jay Sussman, and Ron Weiss. Amorphous computing. Commun. ACM, 43(5):74 -- 82,
May 2000.
[2] P. Bourgine and A. Lesne, editors. Morphogenesis: Origins of Patterns and Shapes, Berlin, 2011. Springer.
[3] J. Cooke and E. C. Zeeman. A clock and wavefront model for control of the number of repeated structures during
animal morphogenesis. J. Theoretical Biology, 58:455 -- 76, 1976.
[4] P. G. de Gennes. Soft matter. Science, 256:495 -- 497, 1992.
[5] M.-L. Dequ´eant and O. Pourqui´e. Segmental patterning of the vertebrate embryonic axis. Nature Reviews
Genetics, 9:370 -- 82, 2008.
[6] Amin Doostmohammadi, Jordi Ign´es-Mullol, Julia M. Yeomans, and Francesc Sagu´es. Active nematics. Nature
Communications, 9(1):3246, 2018.
[7] Ren´e Doursat. Organically grown architectures: Creating decentralized, autonomous systems by embryomorphic
engineering. In R. P. Wurtz, editor, Organic Computing, pages 167 -- 200. Springer, 2008.
[8] Ren´e Doursat, Hiroki Sayama, and Olivier Michel. A review of morphogenetic engineering. 12(4):517 -- 535.
[9] Gabor Forgacs and Stuart A. Newman. Biological Physics of the Developing Embryo. Cambridge University
Press, Cambridge, UK, 2005.
[10] Ryo Fujiwara, Takeshi Kano, and Akio Ishiguro. Self-swarming robots that exploit hydrodynamical interaction.
pages 1 -- 7.
[11] J.-L. Giavitto and A. Spicher. Computer morphogenesis. In P. Bourgine and A. Lesne, editors, Morphogenesis:
Origins of Patterns and Shapes, pages 315 -- 340. Springer, Berlin, 2011.
[12] Robert A. Gingold and Joseph J. Monaghan. Smoothed particle hydrodynamics: theory and application to non-
spherical stars. 181(3):375 -- 389.
[13] S. C. Goldstein, J. D. Campbell, and T. C. Mowry. Programmable matter. Computer, 38(6):99 -- 101, June 2005.
[14] Heiko Hamann. Space-Time Continuous Models of Swarm Robotic Systems, volume 9 of Cognitive Systems
Monographs. Springer Berlin Heidelberg.
[15] C. Huang, J. M. Lei, M. B. Liu, and X. Y. Peng. An improved KGF-SPH with a novel discrete scheme of
laplacian operator for viscous incompressible fluid flows. 81(6):377 -- 396.
[16] H. Kitano. Morphogenesis for evolvable systems. In E. Sanchez and M. Tomassini, editors, Towards Evolvable
Hardware: The Evolutionary Engineering Approach, pages 99 -- 117. Springer, Berlin, 1996.
[17] Jeff W. Lichtman and Winfried Denk. The big and the small: Challenges of imaging the brain's circuits. Science,
334(6056):618 -- 623, 4 November 2011.
[18] G. R. Liu and M. B. Liu. Smoothed particle hydrodynamics: a meshfree particle method. World Scientific.
OCLC: ocm52947194.
[19] Leon B. Lucy. A numerical approach to the testing of the fission hypothesis. 82:1013 -- 1024.
[20] Bruce J. MacLennan. Preliminary development of a formalism for embodied computation and morphogenesis.
Technical Report UT-CS-09-644, Department of Electrical Engineering and Computer Science, University of
Tennessee, Knoxville, TN, 2009.
[21] Bruce J. MacLennan. Models and mechanisms for artificial morphogenesis. In F. Peper, H. Umeo, N. Matsui,
and T. Isokawa, editors, Natural Computing, Springer series, Proceedings in Information and Communications
Technology (PICT) 2, pages 23 -- 33, Tokyo, 2010. Springer.
[22] Bruce J. MacLennan. Morphogenesis as a model for nano communication. Nano Communication Networks.,
1(3):199 -- 208, 2010.
[23] Bruce J. MacLennan. Artificial morphogenesis as an example of embodied computation. International Journal
of Unconventional Computing., 7(1 -- 2):3 -- 23, 2011.
[24] Bruce J. MacLennan. Embodied computation: Applying the physics of computation to artificial morphogenesis.
Parallel Processing Letters, 22(3):1240013, 2012.
[25] Bruce J. MacLennan. Molecular coordination of hierarchical self-assembly. Nano Communication Networks,
3(2):116 -- 128, June 2012.
30
A PREPRINT - AUGUST 26, 2019
[26] Bruce J. MacLennan. Coordinating massive robot swarms. International Journal of Robotics Applications and
Technologies, 2(2):1 -- 19, 2014.
[27] Bruce J. MacLennan. The morphogenetic path to programmable matter. Proceedings of the IEEE, 103(7):1226 --
1232, 2015.
[28] Bruce J. MacLennan. Coordinating swarms of microscopic agents to assemble complex structures. In Ying Tan,
editor, Swarm Intelligence, Vol. 1: Principles, Current Algorithms and Methods, PBCE 119, chapter 20, pages
583 -- 612. Institution of Engineering and Technology, 2018.
[29] Bruce J. MacLennan. The Synmac syntax macroprocessor: Introduction and manual, version 5. Faculty Publi-
cations and Other Works -- EECS. http://trace.tennessee.edu/utk elecpubs/23, University of Tennessee Depart-
ment of Electrical Engineering and Computer Science, 2018.
[30] Bruce J. MacLennan. A morphogenetic program for path formation by continuous flocking.
Journal of Unconventional Computing, 14:91 -- 119, 2019.
International
[31] Jose Martin Z. Maningo, Gerard Ely U. Faelden, Reiichiro Christian S. Nakano, Argel A. Bandala, Ryan Rhay P.
Vicerra, and Elmer P. Dadios. Formation control in quadrotor swarm aggregation using smoothed particle hy-
drodynamics. In Region 10 Conference (TENCON), 2016 IEEE, pages 2070 -- 2075. IEEE.
[32] S. Murata and H. Kurokawa. Self-reconfigurable robots: Shape-changing cellular robots can exceed conventional
robot flexibility. IEEE Robotics & Automation Magazine, pages 71 -- 78, March 2007.
[33] Radhika Nagpal, Attila Kondacs, and Catherine Chang. Programming methodology for biologically-inspired
self-assembling systems. In AAAI Spring Symposium on Computational Synthesis: From Basic Building Blocks
to High Level Functionality, March 2003.
[34] Daniel Needleman and Zvonimir Dogic. Active matter at the interface between materials science and cell biology.
Nature Reviews Materials, 2:17048 EP -- , 07 2017.
[35] Hyondong Oh, Ataollah Ramezan Shirazi, Chaoli Sun, and Yaochu Jin. Bio-inspired self-organising multi-robot
pattern formation: A review. 91:83 -- 100.
[36] Muhammed R. Pac, Aydan M. Erkmen, and Ismet Erkmen. Control of robotic swarm behaviors based on
smoothed particle hydrodynamics. In Intelligent Robots and Systems, 2007. IROS 2007. IEEE/RSJ International
Conference on, pages 4194 -- 4200. IEEE.
[37] James R. Perkinson and Bahram Shafai. A decentralized control algorithm for scalable robotic swarms based
In Proc. of the IASTED Int. Conf. on Robot. and Applications, pages
on mesh-free particle hydrodynamics.
102 -- 107, 2005.
[38] Luciano C. A. Pimenta, Guilherme A. S. Pereira, Nathan Michael, Renato C. Mesquita, Mateus M. Bosque, Luiz
Chaimowicz, and Vijay Kumar. Swarm Coordination Based on Smoothed Particle Hydrodynamics Technique.
IEEE Transactions on Robotics, 29(2):383 -- 399, April 2013.
[39] I. Salazar-Ciudad, J. Jernvall, and S. A. Newman. Mechanisms of pattern formation in development and evolu-
tion. Development, 130:2027 -- 37, 2003.
[40] Matthew B. Silic, Zhuoyuan Song, and Kamran Mohseni. Anisotropic flocking control of distributed multi-agent
In 2018 AIAA Information Systems-AIAA Infotech @ Aerospace. American
systems using fluid abstraction.
Institute of Aeronautics and Astronautics.
[41] A. Spicher, Olivier Michel, and J.-L. Giavitto. Algorithmic self-assembly by accretion and by carving in MGS.
In Proc. of the 7th International Conference on Artificial Evolution (EA '05), number 3871 in Lecture Notes in
Computer Science, pages 189 -- 200. Springer-Verlag, Berlin, 2005.
[42] Larry A. Taber. Nonlinear Theory of Elasticity: Applications in Biomechanics. World Scientific, Singapore,
2004.
[43] Brian P. Teague, Patrick Guye, and Ron Weiss. Synthetic morphogenesis. page a023929.
[44] U. Tilki and A. M. Erkmen. Fluid swarm formation control for hand gesture imitation by ellipse fitting.
51(6):473 -- 475.
[45] Wang Xi, Thuan Beng Saw, Delphine Delacour, Chwee Teck Lim, and Benoit Ladoux. Material approaches to
active tissue mechanics. Nature Reviews Materials, 4(1):23 -- 44, 2019.
31
|
1311.0758 | 1 | 1311 | 2013-11-04T16:28:31 | Observation of large-scale multi-agent based simulations | [
"cs.MA"
] | The computational cost of large-scale multi-agent based simulations (MABS) can be extremely important, especially if simulations have to be monitored for validation purposes. In this paper, two methods, based on self-observation and statistical survey theory, are introduced in order to optimize the computation of observations in MABS. An empirical comparison of the computational cost of these methods is performed on a toy problem. | cs.MA | cs |
Observation of large-scale multi-agent based
simulations
Gildas Morvan1,2 Alexandre Veremme1,3 Daniel Dupont1,3
1Univ Lille Nord de France, F-59000 Lille, France
2UArtois, LGI2A, F-62400, Béthune, France
3HEI, F-59046, Lille, France
http://www.lgi2a.univ-artois.fr/~morvan/
[email protected]
Abstract
The computational cost of large-scale multi-agent based simulations (MABS) can be ex-
tremely important, especially if simulations have to be monitored for validation purposes.
In this paper, two methods, based on self-observation and statistical survey theory, are
introduced in order to optimize the computation of observations in MABS. An empirical
comparison of the computational cost of these methods is performed on a toy problem.
Keywords:
large-scale multi-agent based simulations, observation methods, scalability
Introduction
1
Theoretical and practical advances in the field of multi-agent based simulations (MABS) allow
modelers to simulate very complex systems to solve real world problems. However the anal-
ysis and validation of simulations remain engineering problems that do not have "turnkey"
solutions. Thus, MABS users conducting such tasks face two main issues:
1. define validation metrics for the simulation,
2. compute efficiently the metrics.
The first issue is generally solved by constructing a set of ad-hoc qualitative or quantitative
rules on simulation properties. To evaluate these rules, it is then mandatory to observe the
corresponding simulation properties and thus to consider the second issue. A distinctive char-
acteristic of MABS is that global simulation properties are not necessary directly observable:
they may need to be computed from local agent properties. Fortunately, most of modern
MABS platforms come with observation frameworks and toolboxes. Basically, three types of
observation methods are generally available (Railsback et al., 2006):
1
1. interactive observation: users select observed properties during simulations, e.g., using
a point and click interface,
2. brute-force direct observation: simulation agents sharing a given property are moni-
tored; agent properties are then aggregated by a so-called observer agent that computes
the observation (fig. 1),
3. indirect observation: the observed property is inferred from the observable conse-
quences of agent actions, e.g., in the environment.
observer
observation
simulation agents
Figure 1: Brute-force direct observation method
While the first method is clearly unadapted to the observation of large-scale or batch simu-
lations, the second has an important computational cost. This issue is illustrated with a simple
case study inspired by "StupidModel" (Railsback et al., 2005): N agents move randomly in a
two dimensional environment (cid:69) discretized into 100·100 square cells with Moore neighbour-
hood during 1000 steps. An area (cid:90) ⊆ (cid:69) is defined. The number of agents Z in the area (cid:90) is
observed at each simulation step. This simulation is implemented on the MadKit/TurtleKit
platform1 (Michel et al., 2005). Figure 2 shows the CPU times needed to compute unobserved
and observed simulations on a Dell Precision 650 workstation2, using indirect and brute-force
direct observation methods, for the given expected value E(Z) = N /5, as a function of the
number of simulation agents N.
This work is based on existing implementations (by MadKit/TurtleKit and MASON) of
the direct observation method. Thus, an empirical computational complexity metric, i.e., the
CPU time needed to compute simulations, is used. This metric, denoted (cid:67) , depends, in our
case study, on the number of simulation agents, N, and on the expected value of the cardinal
of the subset of simulation agents computed by f i l t e r , E(Z).
These results show that, in this case, indirect observation has a minor impact on the
computational cost of the simulations. Kaminka et al. also note that this method is not
intrusive: simulation agents do not have to be modified or accessed during the observation
process (Kaminka et al., 2002). However, Wilkins et al. underline that the applicability of
this method is limited: the observed property might not be inferred (Wilkins et al., 2003); it
is often true in complex, i.e., in most of the real world, cases. Thus, direct observation often
remains the only available option.
The brute-force direct observation method is, for this particular and very simple toy-
problem, linear in the number of simulation agents3, i.e., (cid:67) (ob s) ∝ N, while the model is
exponential, i.e., (cid:67) (mod e l) ∝ αN , α > 1. However, complex simulations generally involve
non-linear observation problems (Veremme et al., 2010). Thus, improving direct observation
appears to be a good lead to improve the efficiency of large-scale complex MABS.
SON platform (Luke et al., 2005), leading to similar results.
1All the simulations and observation methods described in this paper have also been implemented on the MA-
2CPU: 2 × 3.06 GHz Intel XeonTM, RAM: 4 × 1 GB. Full specification:
http://www.dell.com/downloads/emea/products/precn/precn_650_uk.pdf.
All the results presented in this paper have been computed by this machine.
3Authors would like to thank anonymous reviewers for raising this issue.
2
no observation
indirect observation
brute-force observation
1000
100
10
1
)
s
(
e
m
i
t
U
P
C
2000
4000
6000
8000
10000 12000 14000 16000 18000 20000
Number of agents
Figure 2: CPU times (log scale) needed to compute observed and unobserved simulations for
E(Z) = N /5
In this paper, two non-brute-force direct observation methods, based on self-observation
and statistical survey theory are introduced. An empirical comparison of the computational
cost of these methods is performed and discussed on the presented case study.
2 Filtrated direct observation of MABS
Basically, there are two ways to compute a direct observation:
1. a set of agents (cid:65) (generally all the simulation agents that share the properties that have
2. a subset (cid:65) (cid:48) of (cid:65) , computed at runtime, is probed (fig. 3).
to be observed), statically defined, is probed by an observer agent,
ob s(cid:48)
(cid:65) (cid:48)
observation
(cid:65)
f i l t e r
Figure 3: Filtrated direct observation
Formally, we consider an observation function
(cid:65) →(cid:73) ,
ob s : 2
(1)
where (cid:65) is a set of agents and (cid:73) represents the values that can be observed. We define
(cid:65) (cid:48) ⊆(cid:65) as the minimal subset of agents, dynamically defined by a set of constraints (e.g., in
the case study, a unique constraint related to the position of the agent), needed to compute
ob s correctly. In other words, (cid:65) (cid:48) is a set of agents such as
ob s(cid:48)((cid:65) (cid:48)) = ob s((cid:65) ) and
(cid:64) (cid:65) (cid:48)(cid:48) ⊂(cid:65) (cid:48) ob s(cid:48)((cid:65) (cid:48)(cid:48)) = ob s((cid:65) ),
(2)
3
where ob s(cid:48) is a "simplified" observation function, i.e., that can be computed faster than ob s
because it is specific to (cid:65) (cid:48):
∀ (cid:65) (cid:48)(cid:48) (cid:54)=(cid:65) (cid:48) ⊆(cid:65) , ob s(cid:48)((cid:65) (cid:48)) = ob s((cid:65) (cid:48)).
(3)
Thus, in the case study, the observation function ob s((cid:65) ) observes the position of each
agent in (cid:65) to count only the ones that are situated in (cid:90) , while an observation function
ob s(cid:48)((cid:65) (cid:48)) would only returns (cid:65) (cid:48) because the agents of (cid:65) (cid:48) are by definition situated in (cid:90) .
To use ob s(cid:48), it is necessary to consider a filtering function f i l t e r able to identify the
subset (cid:65) (cid:48) (in the case study, this subset only contains the agents situated in (cid:90) ):
f i l t e r : 2(cid:65) → 2(cid:65) ∀(cid:65) (cid:48),(cid:65) (cid:48)(cid:48) ⊆(cid:65) ,
if f i l t e r ((cid:65) (cid:48)) =(cid:65) (cid:48)(cid:48), then (cid:65) (cid:48)(cid:48) ⊆(cid:65) (cid:48).
(4)
The goal is to define and implement a filtering function, such as the cost of the observation
computation is reduced, i.e.,
(cid:67) (ob s
(cid:48)(f i l t e r ((cid:65) ))) <(cid:67) (ob s((cid:65) )).
(5)
In the following section, two different implementations of this idea are presented.
Implementation of filtrated direct observation methods
3
3.1 Self-observation
The core idea of this method is to implement the filtering function in the simulation agents
themselves. Then, using an organizational structure, denoted group, allows to identify the set
of agents that has to be observed. Thus, a group is defined as the set of agents that contains the
sufficient and necessary information to compute an observation. In other words, a group defines,
for a given observation, the minimal set of agents that is mandatory to compute it. Agents
observe themselves to determine if they have to join, leave or stay in a group. A filtering
function is defined by a set of rules R, that specifies the conditions under which an agent has
to be observed, evaluated at each simulation step (fig. 4).
Thus, in our case study, we consider a group G, that contains the agents situated in (cid:90) .
• if the agent is in (cid:90) and does not belong to G, then the agent joins G,
• if the agent is not in (cid:90) and belongs to G, then the agent leaves G.
The following set of rules, defined here in natural language, is associated to each agent:
The observation system (ob s(cid:48)) only probes the agents of G. Figure 5 presents the CPU time
difference between simulations observed with self-observation and brute-force methods as a
function of the number of agents in the simulation, N, and the mean rate of observed agents,
E(Z)/N. The dashed line represents the isoline 0, i.e., the conditions for which there is no
difference between the two methods. Thus, the area below this line maps the cases for which
self-observation is faster.
3.2 Statistical survey
If the equation 2 is rewritten as follows:
ob s((cid:65) (cid:48))(cid:39) ob s((cid:65) ) and
(cid:64) (cid:65) (cid:48)(cid:48) ⊂(cid:65) (cid:48) ob s((cid:65) (cid:48)(cid:48))(cid:39) ob s((cid:65) ),
(6)
i.e., if imprecise observations are authorized, it becomes possible to filter the observed pop-
ulation on a statistical basis. As a result, we do not consider a specific observation function
anymore as the set of agents returned by the filtering function is not necessary the set of agents
that contains the sufficient and necessary information to compute the observation.
4
ob s(cid:48)
observation
group G
organisational structures
if R =(cid:62)
f i l t e r : set of rules R
(cid:65)
Figure 4: The self-observation based method
Statistical survey theory4 provides a formal ground to determine optimal sampling method
and size of observed population sample.
Let once again consider our case study; we denote n the size of the observed population,
randomly sampled at each simulation step. An estimator of Z, denoted Z is constructed from
this sample. Many estimator definitions can be found in the literature. In this case, as the pop-
ulation of simulation agents is homogeneous with respect to E(Z) (all the agents have the same
probability to be in (cid:90) ), n is determined with the Horvitz-Thompson estimator (Horvitz and
Thompson, 1952) :
n
−1 = d 2
4S2
+ 1
N
,
(7)
(8)
where d is the maximal absolute error accepted for the observation and
S2 (cid:39) (1− E(Z)
N
)· E(Z)
N
.
Impact on the computational cost is shown in figure 6. The semantic is the same than
figure 5: the left area maps the cases for which the statistical survey based method is faster
than brute-force method.
3.3 Discussion
Figure 7 sums up the previous results qualitatively: conditions for which it is preferable to
use one method over another are identified. These results are specific to our case study and
its implementation; however, they highlight that the choice of an observation method is not
trivial and that the performance of the different available methods should be analyzed on a
set of simulations before using the model in a production context.
In a given context, knowing the map of the fastest observation methods allows to dy-
namically adapt the observation method to use the most efficient one. Impact of dynamic
adaptation of the observation method on CPU time is presented in the context of the first
example (cf. fig. 2) in figure 8.
4Proofs of statistical survey theory results presented in this paper will not be given.
Interested readers may
refer to Bethlehem (2009) for an exhaustive presentation of sampling designs, estimator construction and variability
estimation methods.
5
0
2000
1500
1000
500
0
-500
-1000
)
s
(
e
m
i
t
U
P
C
0
5000
N
10000
15000
0.2
20000
1
0.6
0.4
0.8
E(Z)
N
Figure 5: CPU time difference between simulations observed with self-observation and brute-
force methods as a function of the number of agents in the simulation, N, and the mean rate
of observed agents, E(Z)/N (response surface estimation)
0
)
s
(
e
m
i
t
U
P
C
200
0
-200
-400
-600
-800
-1000
0
5000
N
10000
15000
0.2
20000
1
0.6
0.4
0.8
E(Z)
N
Figure 6: CPU time difference between simulations observed with statistical survey (d = 0.08)
and brute-force methods as a function of the number of agents in the simulation, N, and the
mean rate of observed agents, E(Z)/N (response surface estimation)
6
E(Z)
N
1
0.8
0.6
0.4
0.2
0
0
B
A
C
5000
10000
N
15000
20000
Figure 7: Map of the fastest observation methods (response surface estimation); A: self-
observation, B: brute-force, C: statistical survey (d = 0.008)
no observation
brute-force observation
filtrated observation
1000
100
10
1
)
s
(
e
m
i
t
U
P
C
2000
4000
6000
8000
10000 12000 14000 16000 18000 20000
Number of agents
Figure 8: CPU times (log scale) needed to compute observed and unobserved simulations for
E(Z) = N /5, filtrated observation method being dynamically adapted to the context
7
4 Conclusion and perspectives
Observation methods presented in this paper allow, under specific conditions identified on
a simple case study, to reduce significantly the computational cost of MABS composed of
numerous agents. However, filtering is not the only option. Considering that imprecise
observations are acceptable, while a precision level is guaranteed, the optimal observation
frequency could be determined from the observed property variation. Roughly, the more the
variation, the more the observation frequency. However, MABS are used to simulate complex
systems with nonlinear dynamics. Dynamic adaptation of observation frequency could be an
interesting lead to reduce MABS computational cost.
Moreover, in the statistical survey based method (cf. section 3.2), we consider a simple
random sampling method, assuming the population is homogeneous. Real world MABS of-
ten involve heterogeneous agents for which the distribution of observed individual properties
is not uniform. Clever sampling methods, e.g., a stratified random sampling approach, should
then be used. In very complex cases, a machine learning system should be implemented to
analyze the impact of sampling method properties on the observation quality and compu-
tational cost, and determine the optimal ones. Similarly, the organizational model used to
implement the self-observation based method is very simple: the only organizational struc-
ture that is defined is the "group". Using a more comprehensive one, e.g., AGR (Ferber and
Gutknecht, 1998), would allow to consider very complex and fine observations.
From a methodological point of view, authors experimented that setting up an observa-
tion method, generally improves the design of simulation validity metrics. Indeed, it forces
simulation designers and users to explicitly define local and global observed properties and
their sufficient and necessary conditions of observability, and then the validity constraints
over them.
While this paper focuses on reducing the complexity of observation, many published
works concentrated on agent interactions by dynamically scaling up and down simulated en-
tities or using more structured interaction artifacts (Gaud et al., 2008; Parunak, 2012; Razavi
et al., 2011). Together, these approaches should lead to the conception of highly efficient
large-scale MABS simulators.
References
Bethlehem, J. (2009). Applied Survey Methods: A Statistical Perspective. Wiley.
Ferber, J. and Gutknecht, O. (1998). A meta-model for the analysis and design of orga-
nizations in multi-agent systems. In Proceedings of the Third International Conference on
Multi-Agent Systems (ICMAS98), pages 128–135.
Gaud, N., Galland, S., Gechter, F., Hilaire, V., and Koukam, A. (2008). Holonic multilevel
simulation of complex systems : Application to real-time pedestrians simulation in virtual
urban environment. Simulation Modelling Practice and Theory, 16:1659–1676.
Horvitz, D. and Thompson, D. (1952). A generalization of sampling without replacement
from a finite universe. Journal of the American Statistical Association, 47:663–685.
Kaminka, G., Pynadath, D., and Tambe, M. (2002). Monitoring teams by overhearing: A
multi-agent planrecognition approach. Journal of Artificial Intelligence, 17:83–135.
Luke, S., Cioffi-Revilla, C., Panait, L., Sullivan, K., and Balan, G. (2005). Mason: A multia-
gent simulation environment. Simulation, 81(7):517–527.
Michel, F., Beurier, G., and Ferber, J. (2005). The turtlekit simulation platform: Application
In Proceedings of the First International Conference on Signal-Image
to complex systems.
Technology and Internet Based Systems, pages 122–127.
8
Parunak, H. (2012). Between agents and mean fields. In Multi-Agent-Based Simulation XII,
volume 7124 of Lecture Notes in Artificial Intelligence, pages 113–126. Springer.
Railsback, S., Lytinen, S., and Grimm, V. (2005). Stupidmodel and extensions: A tem-
plate and teaching tool for agent-based modeling platforms. http://condor.depaul.edu/ slyti-
nen/abm/StupidModelFormulation.pdf.
Railsback, S., Lytinen, S., and Jackson, S. (2006). Agent-based simulation platforms: Review
and development recommendations. Simulation, 82(9):609–623.
Razavi, S., Gaud, N., Mozayani, N., and Koukam, A. (2011). Multi-agent based simulations
using fast multipole method: application to large scale simulations of flocking dynamical
systems. Artificial Intelligence Review, 35(1):53–72.
Veremme, A., Lefevre, E., Morvan, G., and Jolly, D. (2010). Application of the belief function
In First international workshop on the
theory to validate multi-agent based simulations.
theory of belief functions.
Wilkins, D., Lee, T., and Berry, P. (2003). Interactive execution monitoring of agent teams.
Journal of Artificial Intelligence Research, 18:217–261.
9
|
1109.6340 | 1 | 1109 | 2011-09-28T20:15:30 | Negotiating Socially Optimal Allocations of Resources | [
"cs.MA"
] | A multiagent system may be thought of as an artificial society of autonomous software agents and we can apply concepts borrowed from welfare economics and social choice theory to assess the social welfare of such an agent society. In this paper, we study an abstract negotiation framework where agents can agree on multilateral deals to exchange bundles of indivisible resources. We then analyse how these deals affect social welfare for different instances of the basic framework and different interpretations of the concept of social welfare itself. In particular, we show how certain classes of deals are both sufficient and necessary to guarantee that a socially optimal allocation of resources will be reached eventually. | cs.MA | cs |
Journal of Artificial Intelligence Research 25 (2006) 315 -- 348
Submitted 08/05; published 03/06
Negotiating Socially Optimal Allocations of Resources
Ulle Endriss
ILLC, University of Amsterdam
1018 TV Amsterdam, The Netherlands
Nicolas Maudet
LAMSADE, Universit´e Paris-Dauphine
75775 Paris Cedex 16, France
Fariba Sadri
Department of Computing, Imperial College London
London SW7 2AZ, UK
Francesca Toni
Department of Computing, Imperial College London
London SW7 2AZ, UK
[email protected]
[email protected]
[email protected]
[email protected]
Abstract
A multiagent system may be thought of as an artificial society of autonomous software
agents and we can apply concepts borrowed from welfare economics and social choice theory
to assess the social welfare of such an agent society. In this paper, we study an abstract
negotiation framework where agents can agree on multilateral deals to exchange bundles
of indivisible resources. We then analyse how these deals affect social welfare for different
instances of the basic framework and different interpretations of the concept of social welfare
itself. In particular, we show how certain classes of deals are both sufficient and necessary
to guarantee that a socially optimal allocation of resources will be reached eventually.
1. Introduction
A multiagent system may be thought of as an artificial society of autonomous software
agents. Negotiation over the distribution of resources (or tasks) amongst the agents inhab-
iting such a society is an important area of research in artificial intelligence and computer
science (Rosenschein & Zlotkin, 1994; Kraus, 2001; Chavez, Moukas, & Maes, 1997; Sand-
holm, 1999). A number of variants of this problem have been studied in the literature.
Here we consider the case of an artificial society of agents where, to begin with, each agent
holds a bundle of indivisible resources to which it assigns a certain utility. Agents may then
negotiate with each other in order to agree on the redistribution of some of these resources
to benefit either themselves or the agent society they inhabit.
Rather than being concerned with specific strategies for negotiation, we analyse how the
redistribution of resources by means of negotiation affects the well-being of the agent society
as a whole. To this end, we make use of formal tools for measuring social welfare developed
in welfare economics and social choice theory (Moulin, 1988; Arrow, Sen, & Suzumura,
2002). In the multiagent systems literature, the utilitarian interpretation of the concept of
social welfare is usually taken for granted (Rosenschein & Zlotkin, 1994; Sandholm, 1999;
Wooldridge, 2002), i.e. whatever increases the average welfare of the agents inhabiting a
society is taken to be beneficial for society as well. This is not the case in welfare economics,
c(cid:13)2006 AI Access Foundation. All rights reserved.
Endriss, Maudet, Sadri, & Toni
for instance, where different notions of social welfare are being studied and compared with
each other. Here, the concept of egalitarian social welfare takes a particularly prominent
role (Sen, 1970; Rawls, 1971; Moulin, 1988; Arrow et al., 2002).
In this model, social
welfare is tied to the individual welfare of the weakest member of society, which facilitates
the incorporation of a notion of fairness into the resource allocation process. While the
discussion of the respective advantages and drawbacks of different notions of social welfare
in the social sciences tends to be dominated by ethical considerations,1 in the context
of societies of artificial software agents the choice of a suitable formal tool for modelling
social welfare boils down to a clear-cut (albeit not necessarily simple) technical design
decision (Endriss & Maudet, 2004).
Indeed, different applications may call for different
social criteria. For instance, for the application studied by Lemaıtre, Verfaillie, and Bataille
(1999), where agents need to agree on the access to an earth observation satellite which has
been funded jointly by the owners of these agents, it is important that each one of them
receives a "fair" share of the common resource. Here, a society governed by egalitarian
principles may be the most appropriate. In an electronic commerce application running on
the Internet where agents have little or no commitments towards each other, on the other
hand, egalitarian principles seem of little relevance. In such a scenario, utilitarian social
welfare would provide an appropriate reflection of the overall profit generated. Besides
utilitarian and egalitarian social welfare, we are also going to discuss notions such as Pareto
and Lorenz optimality (Moulin, 1988), as well as envy-freeness (Brams & Taylor, 1996).
In this paper, we study the effect that negotiation over resources has on society for a
number of different interpretations of the concept of social welfare. In particular, we show
how certain classes of deals regarding the exchange of resources allow us to guarantee that a
socially optimal allocation of resources will be reached eventually. These convergence results
may be interpreted as the emergence of a particular global behaviour (at the level of society)
in reaction to local behaviour governed by the negotiation strategies of individual agents
(which determine the kinds of deals agents are prepared to accept). The work described here
is complementary to the large body of literature on mechanism design and game-theoretical
models of negotiation in multiagent systems (see e.g. Rosenschein & Zlotkin, 1994; Kraus,
2001; Fatima, Wooldridge, & Jennings, 2004). While such work is typically concerned with
negotiation at the local level (how can we design mechanisms that provide an incentive
to individual agents to adopt a certain negotiation strategy?), we address negotiation at a
global level by analysing how the actions taken by agents locally affect the overall system
from a social point of view.
As we shall see, truly multilateral deals involving any number of agents as well as any
number of resources may be necessary to be able to negotiate socially optimal allocations
of resources. This is certainly true as long as we use arbitrary utility functions to model
the preferences of individual agents. In some application domains, however, where utility
functions may be assumed to be subject to certain restrictions (such as being additive), we
are able to obtain stronger results and show that also structurally simpler classes of deals
(in particular, deals involving only a single resource at a time) can be sufficient to negotiate
socially optimal allocations. Nevertheless, for other seemingly strong restrictions on agents'
1. A famous example is Rawls' veil of ignorance, a thought experiment designed to establish what constitutes
a just society (Rawls, 1971).
316
Negotiating Socially Optimal Allocations of Resources
utility functions (such as the restriction to dichotomous preferences) we are able to show
that no reduction in the structural complexity of negotiation is possible.
Our approach to multiagent resource allocation is of a distributed nature. In general,
the allocation procedure used to find a suitable allocation of resources could be either cen-
tralised or distributed. In the centralised case, a single entity decides on the final allocation
of resources amongst agents, possibly after having elicited the agents' preferences over al-
ternative allocations. Typical examples are combinatorial auctions (Cramton, Shoham, &
Steinberg, 2006). Here the central entity is the auctioneer and the reporting of preferences
takes the form of bidding. In truly distributed approaches, on the other hand, allocations
emerge as the result of a sequence of local negotiation steps. Both approaches have their
advantages and disadvantages. Possibly the most important argument in favour of auction-
based mechanisms concerns the simplicity of the communication protocols required to im-
plement such mechanisms. Another reason for the popularity of centralised mechanisms is
the recent push in the design of powerful algorithms for combinatorial auctions that, for
the first time, perform reasonably well in practice (Fujishima, Leyton-Brown, & Shoham,
1999; Sandholm, 2002). Of course, such techniques are, in principle, also applicable in the
distributed case, but research in this area has not yet reached the same level of maturity
as for combinatorial auctions. An important argument against centralised approaches is
that it may be difficult to find an agent that could assume the role of an "auctioneer" (for
instance, in view of its computational capabilities or in view of its trustworthiness).
The line of research pursued in this paper has been inspired by Sandholm's work on
sufficient and necessary contract (i.e. deal) types for distributed task allocation (Sandholm,
1998). Since then, it has been further developed by the present authors, their colleagues, and
others in the context of resource allocation problems (Bouveret & Lang, 2005; Chevaleyre,
Endriss, Estivie, & Maudet, 2004; Chevaleyre, Endriss, Lang, & Maudet, 2005a; Chevaleyre,
Endriss, & Maudet, 2005b; Dunne, 2005; Dunne, Laurence, & Wooldridge, 2004; Dunne,
Wooldridge, & Laurence, 2005; Endriss & Maudet, 2004, 2005; Endriss, Maudet, Sadri,
& Toni, 2003a, 2003b).
In particular, we have extended Sandholm's framework by also
addressing negotiation systems without compensatory side payments (Endriss et al., 2003a),
as well as agent societies where the concept of social welfare is given a different interpretation
to that in the utilitarian programme (Endriss et al., 2003b; Endriss & Maudet, 2004). The
present paper provides a comprehensive overview of the most fundamental results, mostly on
the convergence to an optimal allocation with respect to different notions of social welfare,
in a very active and timely area of ongoing research.
The remainder of this paper is organised as follows. Section 2 introduces the basic ne-
gotiation framework for resource reallocation we are going to consider. It gives definitions
for the central notions of allocation, deal, and utility, and it discusses possible restrictions
to the class of admissible deals (both structural and in terms of acceptability to individual
agents). Section 2 also introduces the various concepts of a social preference we are going
to consider in this paper. Subsequent sections analyse specific instances of the basic ne-
gotiation framework (characterised, in particular, by different criteria for the acceptability
of a proposed deal) with respect to specific notions of social welfare. In the first instance,
agents are assumed to be rational (and "myopic") in the sense of never accepting a deal
that would result in a negative payoff. Section 3 analyses the first variant of this model
of rational negotiation, which allows for monetary side payments to increase the range of
317
Endriss, Maudet, Sadri, & Toni
acceptable deals. As we shall see, this model facilitates negotiation processes that maximise
utilitarian social welfare. If side payments are not possible, we cannot guarantee outcomes
with maximal social welfare, but it is still possible to negotiate Pareto optimal allocations.
This variant of the rational model is studied in Section 4. Both Section 3 and 4 also in-
vestigate how restrictions to the range of utility functions agents may use to model their
preferences can affect such convergence results.
In the second part of the paper we apply our methodology to agent societies where the
concept of social welfare is given a different kind of interpretation than is commonly the
case in the multiagent systems literature. Firstly, in Section 5 we analyse our framework
of resource allocation by negotiation in the context of egalitarian agent societies. Then
Section 6 discusses a variant of the framework that combines ideas from both the utilitarian
and the egalitarian programme and enables agents to negotiate Lorenz optimal allocations
of resources. Finally, Section 7 introduces the idea of using an elitist model of social welfare
for applications where societies of agents are merely a means of enabling at least one agent
to achieve their goal. This section also reviews the concept of envy-freeness and discusses
ways of measuring different degrees of envy.
Section 8 summarises our results and concludes with a brief discussion of the concept of
welfare engineering, i.e. with the idea of choosing tailor-made definitions of social welfare
for different applications and designing agents' behaviour profiles accordingly.
2. Preliminaries
The basic scenario of resource allocation by negotiation studied in this paper is that of an
artificial society inhabited by a number of agents, each of which initially holds a certain
number of resources. These agents will typically ascribe different values (utilities) to differ-
ent bundles of resources. They may then engage in negotiation and agree on the reallocation
of some of the resources, for example, in order to improve their respective individual welfare
(i.e. to increase their utility). Furthermore, we assume that it is in the interest of the system
designer that these distributed negotiation processes -- somehow -- also result in a positive
payoff for society as a whole.
2.1 Basic Definitions
An instance of our abstract negotiation framework consists of a finite set of (at least two)
agents A and a finite set of resources R. Resources are indivisible and non-sharable. An
allocation of resources is a partitioning of R amongst the agents in A.
Definition 1 (Allocations) An allocation of resources is a function A from A to subsets
of R such that A(i) ∩ A(j) = { } for i 6= j and Si∈A A(i) = R.
For example, given an allocation A with A(i) = {r3, r7}, agent i would own resources r3 and
r7. Given a particular allocation of resources, agents may agree on a (multilateral) deal to
exchange some of the resources they currently hold. In the most general case, any numbers
of agents and resources could be involved in a single deal. From an abstract point of view,
a deal takes us from one allocation of resources to the next. That is, we may characterise
a deal as a pair of allocations.
318
Negotiating Socially Optimal Allocations of Resources
Definition 2 (Deals) A deal is a pair δ = (A, A′) where A and A′ are allocations of
resources with A 6= A′.
The set of agents involved in a deal δ = (A, A′) is given by Aδ = {i ∈ A A(i) 6= A′(i)}.
The composition of two deals is defined as follows: If δ1 = (A, A′) and δ2 = (A′, A′′), then
δ1 ◦ δ2 = (A, A′′). If a given deal is the composition of two deals that concern disjoint sets
of agents, then that deal is said to be independently decomposable.
Definition 3 (Independently decomposable deals) A deal δ is called independently
decomposable iff there exist deals δ1 and δ2 such that δ = δ1 ◦ δ2 and Aδ1 ∩ Aδ2 = { }.
Observe that if δ = (A, A′) is independently decomposable then there exists an intermediate
allocation B different from both A and A′ such that the intersection of {i ∈ A A(i) 6= B(i)}
and {i ∈ A B(i) 6= A′(i)} is empty, i.e. such that the union of {i ∈ A A(i) = B(i)} and
{i ∈ A B(i) = A′(i)} is the full set of agents A. Hence, δ = (A, A′) not being independently
decomposable implies that there exists no allocation B different from both A and A′ such
that either B(i) = A(i) or B(i) = A′(i) for all agents i ∈ A (we are going to use this fact in
the proofs of our "necessity theorems" later on).
The value an agent i ∈ A ascribes to a particular set of resources R will be modelled by
means of a utility function, that is, a function from sets of resources to real numbers. We
are going to consider both general utility functions (without any restrictions) and several
more specific classes of functions.
Definition 4 (Utility functions) Every agent i ∈ A is equipped with a utility function
ui : 2R → R. We are going to consider the following restricted classes of utility functions:
• ui is non-negative iff ui(R) ≥ 0 for all R ⊆ R.
• ui is positive iff it is non-negative and ui(R) 6= 0 for all R ⊆ R with R 6= { }.
• ui is monotonic iff R1 ⊆ R2 implies ui(R1) ≤ ui(R2) for all R1, R2 ⊆ R.
• ui is additive iff ui(R) = Pr∈R ui({r}) for all R ⊆ R.
• ui is a 0-1 function iff it is additive and ui({r}) = 0 or ui({r}) = 1 for all r ∈ R.
• ui is dichotomous iff ui(R) = 0 or ui(R) = 1 for all R ⊆ R.
Recall that, given an allocation A, the set A(i) is the bundle of resources held by agent i
in that situation. We are usually going to abbreviate ui(A) = ui(A(i)) for the utility value
assigned by agent i to that bundle.
2.2 Deal Types and Rationality Criteria
In this paper we investigate what kinds of negotiation outcomes agents can achieve by using
different classes of deals. A class of deals may be characterised by both structural constraints
(number of agents and resources involved, etc.) and rationality constraints (relating to the
changes in utility experienced by the agents involved).
Following Sandholm (1998), we can distinguish a number of structurally different types
of deals. The most basic are 1-deals, where a single item is passed from one agent to another.
319
Endriss, Maudet, Sadri, & Toni
Definition 5 (1-deals) A 1-deal is a deal involving the reallocation of exactly one resource.
This corresponds to the "classical" form of a contract typically found in the Contract Net
protocol (Smith, 1980). Deals where one agent passes a set of resources on to another
agent are called cluster deals. Deals where one agent gives a single item to another agent
who returns another single item are called swap deals. Sometimes it can also be necessary
to exchange resources between more than just two agents. In Sandholm's terminology, a
multiagent deal is a deal that could involve any number of agents, where each agent passes
at most one resource to each of the other agents taking part. Finally, deals that combine the
features of the cluster and the multiagent deal type are called combined deals by Sandholm.
These could involve any number of agents and any number of resources. Therefore, every
deal δ, in the sense of Definition 2, is a combined deal. In the remainder of this paper, when
speaking about deals without further specifying their type, we are always going to refer to
combined deals (without any structural restrictions).2
An agent may or may not find a particular deal δ acceptable. Whether or not an agent
will accept a given deal depends on the rationality criterion it applies when evaluating deals.
A selfish agent i may, for instance, only accept deals δ = (A, A′) that strictly improve its
personal welfare: ui(A) < ui(A′). We call criteria such as this, which only depend on
the utilities of the agent in question, personal rationality criteria. While we do not want
to admit arbitrary rationality criteria, the classes of deals that can be characterised using
personal rationality criteria alone is somewhat too narrow for our purposes. Instead, we are
going to consider rationality criteria that are local in the sense of only depending on the
utility levels of the agents involved in the deal concerned.
Definition 6 (Local rationality criteria) A class ∆ of deals is said to be characterised
by a local rationality criterion iff it is possible to define a predicate Φ over 2A×R×R such
that a deal δ = (A, A′) belongs to ∆ iff Φ({(i, ui(A), ui(A′)) i ∈ Aδ}) holds true.
That is, Φ is mapping a set of triples of one agent name and two reals (utilities) each to
truth values. The locality aspect comes in by only applying Φ to the set of triples for
those agents whose bundle changes with δ. Therefore, for instance, the class of all deals
that increase the utility of the previously poorest agent is not characterisable by a local
rationality criterion (because this condition can only be checked by inspecting the utilities
of all the agents in the system).
2.3 Socially Optimal Allocations of Resources
As already mentioned in the introduction, we may think of a multiagent system as a society
of autonomous software agents. While agents make their local decisions on what deals to
propose and to accept, we can also analyse the system from a global or societal point of
view and may thus prefer certain allocations of resources over others. To this end, welfare
economics provides formal tools to assess how the distribution of resources amongst the
members of a society affects the well-being of society as a whole (Sen, 1970; Moulin, 1988;
Arrow et al., 2002).
2. The ontology of deal types discussed here is, of course, not exhaustive. It would, for instance, also be of
interest to consider the class of bilateral deals (involving exactly two agents but any number of items).
320
Negotiating Socially Optimal Allocations of Resources
Given the preference profiles of the individual agents in a society (which, in our frame-
work, are represented by means of their utility functions), a social welfare ordering over
alternative allocations of resources formalises the notion of a society's preferences. Next we
are going to formally introduce the most important social welfare orderings considered in
this paper (some additional concepts of social welfare are discussed towards the end of the
paper). In some cases, social welfare is best defined in terms of a collective utility function.
One such example is the notion of utilitarian social welfare.
Definition 7 (Utilitarian social welfare) The utilitarian social welfare swu(A) of an
allocation of resources A is defined as follows:
swu(A) = Xi∈A
ui(A)
Observe that maximising the collective utility function swu amounts to maximising the
average utility enjoyed by the agents in the system. Asking for maximal utilitarian social
welfare is a very strong requirement. A somewhat weaker concept is that of Pareto optimal-
ity. An allocation is Pareto optimal iff there is no other allocation with higher utilitarian
social welfare that would be no worse for any of of the agents in the system (i.e. that would
be strictly better for at least one agent without being worse for any of the others).
Definition 8 (Pareto optimality) An allocation A is called Pareto optimal iff there is
no allocation A′ such that swu(A) < swu(A′) and ui(A) ≤ ui(A′) for all i ∈ A.
The first goal of an egalitarian society should be to increase the welfare of its weakest
member (Rawls, 1971; Sen, 1970). In other words, we can measure the social welfare of
such a society by measuring the welfare of the agent that is currently worst off.
Definition 9 (Egalitarian social welfare) The egalitarian social welfare swe(A) of an
allocation of resources A is defined as follows:
swe(A) = min{ui(A) i ∈ A}
The egalitarian collective utility function swe gives rise to a social welfare ordering over
alternative allocations of resources: A′ is strictly preferred over A iff swe(A) < swe(A′). This
ordering is sometimes called the maximin-ordering. The maximin-ordering only takes into
account the welfare of the currently weakest agent, but is insensitive to utility fluctuations
in the rest of society. To allow for a finer distinction of the social welfare of different
allocations we introduce the so-called leximin-ordering.
For a society with n agents, let {u1, . . . , un} be the set of utility functions for that
society. Then every allocation A determines a utility vector hu1(A), . . . , un(A)i of length n.
If we rearrange the elements of that vector in increasing order we obtain the ordered utility
vector for allocation A, which we are going to denote by ~u(A). The number ~ui(A) is the ith
element in such a vector (for 1 ≤ i ≤ A). That is, ~u1(A) for instance, is the utility value
assigned to allocation A by the currently weakest agent. We now declare a lexicographic
ordering over vectors of real numbers (such as ~u(A)) in the usual way: ~x lexicographically
precedes ~y iff ~x is a (proper) prefix of ~y or ~x and ~y share a common (proper) prefix of length
k (which may be 0) and we have ~xk+1 < ~yk+1.
321
Endriss, Maudet, Sadri, & Toni
Definition 10 (Leximin-ordering) The leximin-ordering ≺ over alternative allocations
of resources is defined as follows:
A ≺ A′
iff
~u(A) lexicographically precedes ~u(A′)
We write A (cid:22) A′ iff either A ≺ A′ or ~u(A) = ~u(A′). An allocation of resources A is called
leximin-maximal iff there is no other allocation A′ such that A ≺ A′.
Finally, we introduce the concept of Lorenz domination, a social welfare ordering that
combines utilitarian and egalitarian aspects of social welfare. The basic idea is to endorse
deals that result in an improvement with respect to utilitarian welfare without causing a
loss in egalitarian welfare, and vice versa.
Definition 11 (Lorenz domination) Let A and A′ be allocations for a society with n
agents. Then A is Lorenz dominated by A′ iff
k
Xi=1
~ui(A) ≤
k
Xi=1
~ui(A′)
for all k with 1 ≤ k ≤ n and, furthermore, that inequality is strict for at least one k.
For any k with 1 ≤ k ≤ n, the sum referred to in the above definition is the sum of the
utility values assigned to the respective allocation of resources by the k weakest agents. For
k = 1, this sum is equivalent to the egalitarian social welfare for that allocation. For k = n,
it is equivalent to the utilitarian social welfare. An allocation of resources is called Lorenz
optimal iff it is not Lorenz dominated by any other allocation.
We illustrate some of the above social welfare concepts (and the use of ordered utility
vectors) by means of an example. Consider a society with three agents and two resources,
with the agents' utility functions given by the following table:
u1({ }) = 0
u1({r1}) = 5
u1({r2}) = 3
u1({r1, r2}) = 8 u2({r1, r2}) = 17 u3({r1, r2}) = 7
u3({ }) = 0
u3({r1}) = 2
u3({r2}) = 6
u2({ }) = 0
u2({r1}) = 4
u2({r2}) = 2
First of all, we observe that the egalitarian social welfare will be 0 for any possible allocation
in this scenario, because at least one of the agents would not get any resources at all. Let
A be the allocation where agent 2 holds the full bundle of resources. Observe that this is
the allocation with maximal utilitarian social welfare. The corresponding utility vector is
h0, 17, 0i, i.e. ~u(A) = h0, 0, 17i. Furthermore, let A′ be the allocation where agent 1 gets
r1, agent 2 gets r2, and agent 3 has to be content with the empty bundle. Now we get an
ordered utility vector of h0, 2, 5i. The initial element in either vector is 0, but 0 < 2, i.e.
~u(A) lexicographically precedes ~u(A′). Hence, we get A ≺ A′, i.e. A′ would be the socially
preferred allocation with respect to the leximin-ordering. Furthermore, both A and A′ are
Pareto optimal and neither is Lorenz-dominated by the other. Starting from allocation A′,
agents 1 and 2 swapping their respective bundles would result in an allocation with the
ordered utility vector h0, 3, 4i, i.e. this move would result in a Lorenz improvement.
322
Negotiating Socially Optimal Allocations of Resources
3. Rational Negotiation with Side Payments
In this section, we are going to discuss a first instance of the general framework of resource
allocation by negotiation set out earlier. This particular variant, which we shall refer to as
the model of rational negotiation with side payments (or simply with money), is equivalent
to a framework put forward by Sandholm where agents negotiate in order to reallocate
tasks (Sandholm, 1998). For this variant of the framework, our aim will be to negotiate
allocations with maximal utilitarian social welfare.
3.1 Individual Rationality
In this instance of our negotiation framework, a deal may be accompanied by a number
of monetary side payments to compensate some of the agents involved for accepting a loss
in utility. Rather than specifying for each pair of agents how much money the former is
supposed to pay to the latter, we simply say how much money each agent either pays out
or receives. This can be modelled by using what we call a payment function.
Definition 12 (Payment functions) A payment function is a function p from A to real
numbers satisfying the following condition:
p(i) = 0
Xi∈A
Here, p(i) > 0 means that agent i pays the amount of p(i), while p(i) < 0 means that it
receives the amount of −p(i). By definition of a payment function, the sum of all payments
is 0, i.e. the overall amount of money present in the system does not change.3
In the rational negotiation model, agents are self-interested in the sense of only propos-
ing or accepting deals that strictly increase their own welfare (for a justification of this
approach we refer to Sandholm, 1998). This "myopic" notion of individual rationality may
be formalised as follows.
Definition 13 (Individual rationality) A deal δ = (A, A′) is called individually rational
iff there exists a payment function p such that ui(A′) − ui(A) > p(i) for all i ∈ A, except
possibly p(i) = 0 for agents i with A(i) = A′(i).
That is, agent i will be prepared to accept the deal δ iff it has to pay less than its gain in
utility or it will get paid more than its loss in utility, respectively. Only for agents i not
affected by the deal, i.e. in case A(i) = A′(i), there may be no payment at all. For example,
if ui(A) = 8 and ui(A′) = 5, then the utility of agent i would be reduced by 3 units if it were
to accept the deal δ = (A, A′). Agent i will only agree to this deal if it is accompanied by
a side payment of more than 3 units; that is, if the payment function p satisfies −3 > p(i).
For any given deal, there will usually be a range of possible side payments. How agents
manage to agree on a particular one is not a matter of consideration at the abstract level
at which we are discussing this framework here. We assume that a deal will go ahead
as long as there exists some suitable payment function p. We should point out that this
3. As the overall amount of money present in the system stays constant throughout the negotiation process,
it makes sense not to take it into account for the evaluation of social welfare.
323
Endriss, Maudet, Sadri, & Toni
assumption may not be justified under all circumstances. For instance, if utility functions
are not publicly known and agents are risk-takers, then a potential deal may not be identified
as such, because some of the agents may understate their interest in that deal in order to
maximise their expected payoff (Myerson & Satterthwaite, 1983). Therefore, the theoretical
results on the reachability of socially optimal allocations reported below will only apply
under the assumption that such strategic considerations will not prevent agents from making
mutually beneficial deals.
3.2 An Example
As an example, consider a system with two agents, agent 1 and agent 2, and a set of two
resources R = {r1, r2}. The following table specifies the values of the utility functions u1
and u2 for every subset of {r1, r2}:
u1({ }) = 0
u1({r1}) = 2
u1({r2}) = 3
u1({r1, r2}) = 7 u2({r1, r2}) = 8
u2({ }) = 0
u2({r1}) = 3
u2({r2}) = 3
Also suppose agent 1 initially holds the full set of resources {r1, r2} and agent 2 does not
own any resources to begin with.
The utilitarian social welfare for this initial allocation is 7, but it could be 8, namely if
agent 2 had both resources. As we are going to see next, the simple class of 1-deals alone are
not always sufficient to guarantee the optimal outcome of a negotiation process (if agents
abide to the individual rationality criterion for the acceptability of a deal). In our example,
the only possible 1-deals would be to pass either r1 or r2 from agent 1 to agent 2. In either
case, the loss in utility incurred by agent 1 (5 or 4, respectively) would outweigh the gain
of agent 2 (3 for either deal), so there is no payment function that would make these deals
individually rational. The cluster deal of passing {r1, r2} from agent 1 to 2, on the other
hand, would be individually rational if agent 2 paid agent 1 an amount of, say, 7.5 units.
Similarly to the example above, we can also construct scenarios where swap deals or
multiagent deals are necessary (i.e. where cluster deals alone would not be sufficient to
guarantee maximal social welfare). This also follows from Theorem 2, which we are going
to present later on in this section. Several concrete examples are given by Sandholm (1998).
3.3 Linking Individual Rationality and Social Welfare
The following result, first stated in this form by Endriss et al. (2003a), says that a deal (with
money) is individually rational iff it increases utilitarian social welfare. We are mainly going
to use this lemma to give a simple proof of Sandholm's main result on sufficient contract
types (Sandholm, 1998), but it has also found useful applications in its own right (Dunne
et al., 2005; Dunne, 2005; Endriss & Maudet, 2005).
Lemma 1 (Individually rational deals and utilitarian social welfare) A deal δ =
(A, A′) is individually rational iff swu(A) < swu(A′).
324
Negotiating Socially Optimal Allocations of Resources
'⇒': By definition, δ = (A, A′) is individually rational iff there exists a payment
Proof.
function p such that ui(A′) − ui(A) > p(i) holds for all i ∈ A, except possibly p(i) = 0 in
case A(i) = A′(i). If we add up the inequations for all agents i ∈ A we get:
Xi∈A
(ui(A′) − ui(A)) > Xi∈A
p(i)
By definition of a payment function, the righthand side equates to 0 while, by definition of
utilitarian social welfare, the lefthand side equals swu(A′) − swu(A). Hence, we really get
swu(A) < swu(A′) as claimed.
'⇐': Now let swu(A) < swu(A′). We have to show that δ = (A, A′) is an individually
rational deal. We are done if we can prove that there exists a payment function p such that
ui(A′) − ui(A) > p(i) for all i ∈ A. We define the function p : A → R as follows:
p(i) = ui(A′) − ui(A) −
swu(A′) − swu(A)
A
(for i ∈ A)
First, observe that p really is a payment function, because we get Pi∈A p(i) = 0. We also
get ui(A′) − ui(A) > p(i) for all i ∈ A, because we have swu(A′) − swu(A′) > 0. Hence, δ
must indeed be an individually rational deal.
✷
Lemma 1 suggests that the function swu does indeed provide an appropriate measure of
social well-being in societies of agents that use the notion of individual rationality (as given
by Definition 13) to guide their behaviour during negotiation. It also shows that individual
rationality is indeed a local rationality criterion in the sense of Definition 6.
3.4 Maximising Utilitarian Social Welfare
Our next aim is to show that any sequence of deals in the rational negotiation model
with side payments will converge to an allocation with maximal utilitarian social welfare;
that is, the class of individually rational deals (as given by Definition 13) is sufficient to
guarantee optimal outcomes for agent societies measuring welfare according to the utilitarian
programme (Definition 7). This has originally been shown by Sandholm (1998) in the
context of a framework where rational agents negotiate in order to reallocate tasks and
where the global aim is to minimise the overall costs of carrying out these tasks.
Theorem 1 (Maximal utilitarian social welfare) Any sequence of individually ratio-
nal deals will eventually result in an allocation with maximal utilitarian social welfare.
Proof. Given that both the set of agents A as well as the set of resources R are required
to be finite, there can be only a finite number of distinct allocations of resources. Further-
more, by Lemma 1, any individually rational deal will strictly increase utilitarian social
welfare. Hence, negotiation must terminate after a finite number of deals. For the sake
of contradiction, assume that the terminal allocation A does not have maximal utilitarian
social welfare, i.e. there exists another allocation A′ with swu(A) < swu(A′). But then, by
Lemma 1, the deal δ = (A, A′) would be individually rational and thereby possible, which
contradicts our earlier assumption of A being a terminal allocation.
✷
325
Endriss, Maudet, Sadri, & Toni
At first sight, this result may seem almost trivial. The notion of a multilateral deal without
any structural restrictions is a very powerful one. A single such deal allows for any number
of resources to be moved between any number of agents. From this point of view, it is not
particularly surprising that we can always reach an optimal allocation (even in just a single
step!). Furthermore, finding a suitable deal is a very complex task, which may not always
be viable in practice. The crucial point of Theorem 1 is that any sequence of deals will
result in an optimal allocation. That is, whatever deals are agreed on in the early stages of
negotiation, the system will never get stuck in a local optimum and finding an allocation
with maximal social welfare remains an option throughout (provided, of course, that agents
are actually able to identify any deal that is theoretically possible). Given the restriction to
deals that are individually rational for all the agents involved, social welfare must increase
with every single deal. Therefore, negotiation always pays off, even if it has to stop early
due to computational limitations.
The issue of complexity is still an important one. If the full range of deals is too large
to be managed in practice, it is important to investigate how close we can get to finding
an optimal allocation if we restrict the set of allowed deals to certain simple patterns.
Andersson and Sandholm (2000), for instance, have conducted a number of experiments on
the sequencing of certain contract/deal types to reach the best possible allocations within
a limited amount of time. For a complexity-theoretic analysis of the problem of deciding
whether it is possible to reach an optimal allocation by means of structurally simple types
of deals (in particular 1-deals), we refer to recent work by Dunne et al. (2005).
3.5 Necessary Deals
The next theorem improves upon Sandholm's main result regarding necessary contract
types (Sandholm, 1998), by extending it to the cases where either all utility functions are
monotonic or all utility functions are dichotomous.4 Sandholm's original result, translated
into our terminology, states that for any system (consisting of a set of agents A and a
set of resources R) and any (not independently decomposable) deal δ for that system, it
is possible to construct utility functions and choose an initial allocation of resources such
that δ is necessary to reach an optimal allocation, if agents only agree to individually
rational deals. All other findings on the insufficiency of certain types of contracts reported
by Sandholm (1998) may be considered corollaries to this. For instance, the fact that,
say, cluster deals alone are not sufficient to guarantee optimal outcomes follows from this
theorem if we take δ to be any particular swap deal for the system in question.
Theorem 2 (Necessary deals with side payments) Let the sets of agents and re-
sources be fixed. Then for every deal δ that is not independently decomposable, there exist
utility functions and an initial allocation such that any sequence of individually rational
deals leading to an allocation with maximal utilitarian social welfare must include δ. This
continues to be the case even when either all utility functions are required to be monotonic
or all utility functions are required to be dichotomous.
4. In fact, our theorem not only sharpens but also also corrects a mistake in previous expositions of this
result (Sandholm, 1998; Endriss et al., 2003a), where the restriction to deals that are not independently
decomposable had been omitted.
326
Negotiating Socially Optimal Allocations of Resources
Proof. Given a set of agents A and a set of resources R, let δ = (A, A′) with A 6= A′ be
any deal for this system. We need to show that there are a collection of utility functions
and an initial allocation such that δ is necessary to reach an allocation with maximal social
welfare. This would be the case if A′ had maximal social welfare, A had the second highest
social welfare, and A were the initial allocation of resources.
We first prove the existence of such a collection of functions for the case where all utility
functions are required to be monotonic. Fix ǫ such that 0 < ǫ < 1. As we have A 6= A′,
there must be an agent j ∈ A such that A(j) 6= A′(j). We now define utility functions ui
for agents i ∈ A and sets of resources R ⊆ R as follows:
ui(R) = (cid:26) R + ǫ
R
if R = A′(i) or (R = A(i) and i 6= j)
otherwise
Observe that ui is a monotonic utility function for every i ∈ A. We get swu(A′) = R+ǫ·A
and swu(A) = swu(A′) − ǫ. Because δ = (A, A′) is not individually decomposable, there
exists no allocation B different from both A and A′ such that either B(i) = A(i) or B(i) =
A′(i) for all agents i ∈ A. Hence, swu(B) ≤ swu(A) for any other allocation B. That
is, A′ is the (unique) allocation with maximal social welfare and the only allocation with
higher social welfare than A. Therefore, if we were to make A the initial allocation then
δ = (A, A′) would be the only deal increasing social welfare. By Lemma 1, this means that
δ would be the only individually rational (and thereby the only possible) deal. Hence, δ is
indeed necessary to achieve maximal utilitarian social welfare.
The proof for the case of dichotomous utility functions is very similar; we only need to
show that a suitable collection of dichotomous utility functions can be constructed. Again,
let j be an agent with A(j) 6= A′(j). We can use the following collection of functions:
ui(R) = (cid:26) 1 if R = A′(i) or (R = A(i) and i 6= j)
0 otherwise
We get swu(A′) = A, swu(A) = swu(A′)−1 and swu(B) ≤ swu(A) for all other allocations
B. Hence, A is not socially optimal and, with A as the initial allocation, δ would be the
only deal that is individually rational.
✷
We should stress that the set of deals that are not independently decomposable includes
deals involving any number of agents and/or any number of resources. Hence, by Theorem 2,
any negotiation protocol that puts restrictions on the structural complexity of deals that
may be proposed will fail to guarantee optimal outcomes, even when there are no constraints
on either time or computational resources. This emphasises the high complexity of our
negotiation framework (see also Dunne et al., 2005; Chevaleyre et al., 2004; Endriss &
Maudet, 2005; Dunne, 2005). The fact that the necessity of (almost) the full range of deals
persists, even when all utility functions are subject to certain restrictions makes this result
even more striking. This is true in particular for the case of dichotomous functions, which
are in some sense the simplest class of utility functions (as they can only distinguish between
"good" and "bad" bundles).
To see that the restriction to deals that are not independently decomposable matters,
consider a scenario with four agents and two resources. If the deal δ of moving r1 from
agent 1 to agent 2, and r2 from agent 3 to agent 4 is individually rational, then so will be
327
Endriss, Maudet, Sadri, & Toni
either one of the two "subdeals" of moving either r1 from agent 1 to agent 2 or r2 from
agent 3 to agent 4. Hence, the deal δ (which is independently decomposable) cannot be
necessary in the sense of Theorem 2 (with reference to our proof above, in the case of δ
there are allocations B such that either B(i) = A(i) or B(i) = A′(i) for all agents i ∈ A,
i.e. we could get swu(B) = swu(A′)).
3.6 Additive Scenarios
Theorem 2 is a negative result, because it shows that deals of any complexity may be
required to guarantee optimal outcomes of negotiation. This is partly a consequence of the
high degree of generality of our framework. In Section 2.1, we have defined utility functions
as arbitrary functions from sets of resources to real numbers. For many application domains
this may be unnecessarily general or even inappropriate and we may be able to obtain
stronger results for specific classes of utility functions that are subject to certain restrictions.
Of course, we have already seen that this is not the case for either monotonicity (possibly
the most natural restriction) or dichotomy (possibly the most severe restriction).
Here, we are going to consider the case of additive utility functions, which are appro-
priate for domains where combining resources does not result in any synergy effects (in the
sense of increasing an agent's welfare). We refer to systems where all agents have additive
utility functions as additive scenarios. The following theorem shows that for these additive
scenarios 1-deals are sufficient to guarantee outcomes with maximal social welfare.5
Theorem 3 (Additive scenarios) In additive scenarios, any sequence of individually ra-
tional 1-deals will eventually result in an allocation with maximal utilitarian social welfare.
Proof. Termination is shown as for Theorem 1. We are going to show that, whenever the
current allocation does not have maximal social welfare, then there is still a possible 1-deal
that is individually rational.
In additive domains, the utilitarian social welfare of a given allocation may be computed
by adding up the appropriate utility values for all the single resources in R. For any
allocation A, let fA be the function mapping each resource r ∈ R to the agent i ∈ A that
holds r in situation A (that is, we have fA(r) = i iff r ∈ A(i)). The utilitarian social welfare
for allocation A is then given by the following formula:
swu(A) = Xr∈R
ufA(r)({r})
Now suppose that negotiation has terminated with allocation A and there are no more
individually rational 1-deals possible. Furthermore, for the sake of contradiction, assume
that A is not an allocation with maximal social welfare, i.e. there exists another allocation
A′ with swu(A) < swu(A′). But then, by the above characterisation of social welfare for
additive scenarios, there must be at least one resource r ∈ R such that ufA(r)({r}) <
ufA′ (r)({r}). That is, the 1-deal δ of passing r from agent fA(r) on to agent fA′(r) would
increase social welfare. Therefore, by Lemma 1, δ must be an individually rational deal,
i.e. contrary to our earlier assumption, A cannot be a terminal allocation. Hence, A must
be an allocation with maximal utilitarian social welfare.
✷
5. This has also been observed by T. Sandholm (personal communication, September 2002).
328
Negotiating Socially Optimal Allocations of Resources
We conclude this section by briefly mentioning two recent results that both extend, in
different ways, the result stated in Theorem 3 (a detailed discussion, however, would be
beyond the scope of the present paper). The first of these results shows that rational
deals involving at most k resources each are sufficient for convergence to an allocation
with maximal social welfare whenever all utility functions are additively separable with
respect to a common partition of R -- i.e. synergies across different parts of the partition
are not possible and overall utility is defined as the sum of utilities for the different sets
in the partition (Fishburn, 1970) -- and each set in this partition has at most k elements
(Chevaleyre et al., 2005a).
The second result concerns a maximality property of utility functions with respect to
1-deals. Chevaleyre et al. (2005b) show that the class of modular utility functions, which is
only slightly more general than the class of additive functions considered here (namely, it is
possible to assign a non-zero utility to the empty bundle), is maximal in the sense that for no
class of functions strictly including the class of modular functions it would still be possible
to guarantee that agents using utility functions from that larger class and negotiating only
individually rational 1-deals will eventually reach an allocation with maximal utilitarian
social welfare in all cases.
4. Rational Negotiation without Side Payments
An implicit assumption made in the framework that we have presented so far is that every
agent has got an unlimited amount of money available to it to be able to pay other agents
whenever this is required for a deal that would increase utilitarian social welfare. Concretely,
if A is the initial allocation and A′ is the allocation with maximal utilitarian social welfare,
then agent i may require an amount of money just below the difference ui(A′) − ui(A) to be
able to get through the negotiation process. In the context of task contracting, for which
this framework has been proposed originally (Sandholm, 1998), this may be justifiable,
but for resource allocation problems it seems questionable to make assumptions about the
unlimited availability of one particular resource, namely money. In this section, we therefore
investigate to what extent the theoretical results discussed in the previous section persist
to apply when we consider negotiation processes without monetary side payments.
4.1 An Example
In a scenario without money, that is, if we do not allow for compensatory payments, we
cannot always guarantee an outcome with maximal utilitarian social welfare. To see this,
consider the following simple problem for a system with two agents, agent 1 and agent 2,
and a single resource r. The agents' utility functions are defined as follows:
u1({ }) = 0
u2({ }) = 0
u1({r}) = 4 u2({r}) = 7
Now suppose agent 1 initially owns the resource. Then passing r from agent 1 to agent 2
would increase utilitarian social welfare by an amount of 3. For the framework with money,
agent 2 could pay agent 1, say, the amount of 5.5 units and the deal would be individually
rational for both of them. Without money (i.e. if p ≡ 0), however, no individually rational
deal is possible and negotiation must terminate with a non-optimal allocation.
329
Endriss, Maudet, Sadri, & Toni
As maximising social welfare is not generally possible, instead we are going to investigate
whether a Pareto optimal outcome (see Definition 8) is possible in the framework without
money, and what types of deals are sufficient to guarantee this.
4.2 Cooperative Rationality
As will become clear in due course, in order to get the desired convergence result, we need
to relax the notion of individual rationality a little. For the framework without money, we
also want agents to agree to a deal, if this at least maintains their utility (that is, no strict
increase is necessary). However, we are still going to require at least one agent to strictly
increase their utility. This could, for instance, be the agent proposing the deal in question.
We call deals conforming to this criterion cooperatively rational.6
Definition 14 (Cooperative rationality) A deal δ = (A, A′) is called cooperatively ra-
tional iff ui(A) ≤ ui(A′) for all i ∈ A and there is an agent j ∈ A such that uj(A) < uj(A′).
In analogy to Lemma 1, we still have swu(A) < swu(A′) for any deal δ = (A, A′) that is
cooperatively rational, but not vice versa. We call the instance of our negotiation framework
where all deals are cooperatively rational (and hence do not include a monetary component)
the model of rational negotiation without side payments.
4.3 Ensuring Pareto Optimal Outcomes
As the next theorem will show, the class of cooperatively rational deals is sufficient to
guarantee a Pareto optimal outcome of money-free negotiation. It constitutes the analogue
to Theorem 1 for the model of rational negotiation without side payments.
Theorem 4 (Pareto optimal outcomes) Any sequence of cooperatively rational deals
will eventually result in a Pareto optimal allocation of resources.
Proof. Every cooperatively rational deal strictly increases utilitarian social welfare (this is
where we need the condition that at least one agent behaves truly individually rational for
each deal). Together with the fact that there are only finitely many different allocations
of resources, this implies that any negotiation process will eventually terminate. For the
sake of contradiction, assume negotiation ends with allocation A, but A is not Pareto
optimal. The latter means that there exists another allocation A′ with swu(A) < swu(A′)
and ui(A) ≤ ui(A′) for all i ∈ A.
If we had ui(A) = ui(A′) for all i ∈ A, then also
swu(A) = swu(A′); that is, there must be at least one j ∈ A with uj(A) < uj(A′). But then
the deal δ = (A, A′) would be cooperatively rational, which contradicts our assumption of
A being a terminal allocation.
✷
Observe that the proof would not have gone through if deals were required to be strictly
rational (without side payments), as this would necessitate ui(A) < ui(A′) for all i ∈ A.
Cooperative rationality means, for instance, that agents would be prepared to give away
resources to which they assign a utility value of 0, without expecting anything in return. In
6. Note that "cooperatively rational" agents are still essentially rational. Their willingness to cooperate
only extends to cases where they can benefit others without any loss in utility for themselves.
330
Negotiating Socially Optimal Allocations of Resources
the framework with money, another agent could always offer such an agent an infinitesimally
small amount of money, who would then accept the deal.
Therefore, our proposed weakened notion of rationality seems indeed a very reasonable
price to pay for giving up money.
4.4 Necessity Result
As our next result shows, also for the framework without side payments, deals of any
structural complexity may be necessary in order to be able to guarantee an optimal outcome
of a negotiation.7 Theorem 5 improves upon previous results (Endriss et al., 2003a) by
showing that this necessity property persists also when either all utility functions belong to
the class of monotonic functions or all utility functions belong to the class of dichotomous
functions.
Theorem 5 (Necessary deals without side payments) Let the sets of agents and re-
sources be fixed. Then for every deal δ that is not independently decomposable, there exist
utility functions and an initial allocation such that any sequence of cooperatively rational
deals leading to a Pareto optimal allocation would have to include δ. This continues to be
the case even when either all utility functions are required to be monotonic or all utility
functions are required to be dichotomous.
Proof. The details of this proof are omitted as it is possible to simply reuse the construction
used in the proof of Theorem 2. Observe that the utility functions defined there also
guarantee ui(A) ≤ ui(A′) for all i ∈ A, i.e. A is not Pareto optimal, but A′ is. If we were
to make A the initial allocation, then δ = (A, A′) would be the only cooperatively rational
deal (as every other deal would decrease social welfare).
✷
4.5 0-1 Scenarios
We conclude our study of the rational negotiation framework without side payments by
identifying a class of utility functions where we are able to achieve a reduction in structural
complexity.8 Consider a scenario where agents use additive utility functions that assign
either 0 or 1 to every single resource (this is what we call 0-1 functions).9 This may be
appropriate when we simply wish to distinguish whether or not the agent needs a particular
resource (to execute a given plan, for example). This is, for instance, the case for some of
the agents defined in the work of Sadri, Toni, and Torroni (2001). As the following theorem
shows, for 0-1 scenarios (i.e. for systems where all utility functions are 0-1 functions),
7. This theorem corrects a mistake in the original statement of the result (Endriss et al., 2003a), where the
restriction to deals that are not independently decomposable had been omitted.
8. As dichotomous functions are a special case of the non-negative functions, the full range of (not indepen-
dently decomposable) deals is also necessary in scenarios with non-negative functions. Interestingly, this
changes when we restrict ourselves to positive utility functions. Now the result of Theorem 5 would not
hold anymore, because any deal that would involve a particular agent (with a positive utility function)
giving away all its resources without receiving anything in return could never be cooperatively rational.
Hence, by Theorem 4, such a deal could never be necessary to achieve a Pareto optimal allocation either.
9. Recall the distinction between 0-1 functions and dichotomous functions. The latter assign either 0 or 1
to each bundle, while the former assign either 0 or 1 to each individual resource (the utilities of bundles
then follow from the fact that 0-1 functions are additive).
331
Endriss, Maudet, Sadri, & Toni
1-deals are sufficient to guarantee convergence to an allocation with maximal utilitarian
social welfare, even in the framework without monetary side payments (where all deals are
required to be cooperatively rational).
Theorem 6 (0-1 scenarios) In 0-1 scenarios, any sequence of cooperatively rational 1-
deals will eventually result in an allocation with maximal utilitarian social welfare.
Proof. Termination is shown as in the proof of Theorem 4. If an allocation A does not have
maximal social welfare then it must be the case that some agent i holds a resource r with
ui({r}) = 0 and there is another agent j in the system with uj({r}) = 1. Passing r from i
to j would be a cooperatively rational deal, so either negotiation has not yet terminated or
we are indeed in a situation with maximal utilitarian social welfare.
✷
This result may be interpreted as a formal justification for some of the negotiation strategies
proposed by Sadri et al. (2001).
5. Egalitarian Agent Societies
In this section, we are going to apply the same methodology that we have used to study
optimal outcomes of negotiation in systems designed according to utilitarian principles in
the first part of this paper to the analysis of egalitarian agent societies. The classical
counterpart to the utilitarian collective utility function swu is the egalitarian collective
utility function swe introduced in Definition 9 (Moulin, 1988; Sen, 1970; Rawls, 1971).
Therefore, we are going to study the design of agent societies for which negotiation can be
guaranteed to converge to an allocation of resources with maximal egalitarian social welfare.
Our first aim will be to identify a suitable criterion that agents inhabiting an egalitarian
agent society may use to decide whether or not to accept a particular deal. Clearly, coop-
eratively rational deals, for instance, would not be an ideal choice, because Pareto optimal
allocations will typically not be optimal from an egalitarian point of view (Moulin, 1988).
5.1 Pigou-Dalton Transfers and Equitable Deals
When searching the economics literature for a class of deals that would benefit society
in an egalitarian system we soon encounter Pigou-Dalton transfers. The Pigou-Dalton
principle states that whenever a utility transfer between two agents takes place which
reduces the difference in utility between the two, then that transfer should be considered
socially beneficial (Moulin, 1988). In the context of our framework, a Pigou-Dalton transfer
(between agents i and j) can be defined as follows.
Definition 15 (Pigou-Dalton transfers) A deal δ = (A, A′) is called a Pigou-Dalton
transfer iff it satisfies the following criteria:
• Only two agents i and j are involved in the deal: Aδ = {i, j}.
• The deal is mean-preserving: ui(A) + uj(A) = ui(A′) + uj(A′).
• The deal reduces inequality: ui(A′) − uj(A′) < ui(A) − uj(A).
332
Negotiating Socially Optimal Allocations of Resources
The second condition in this definition could be relaxed to postulate ui(A) + uj(A) ≤
ui(A′) + uj(A′), to also allow for inequality-reducing deals that increase overall utility.
Pigou-Dalton transfers capture certain egalitarian principles; but are they sufficient as
acceptability criteria to guarantee negotiation outcomes with maximal egalitarian social
welfare? Consider the following example:
u1({ }) = 0
u1({r1}) = 3
u1({r2}) = 12
u1({r1, r2}) = 15 u2({r1, r2}) = 17
u2({ }) = 0
u2({r1}) = 5
u2({r2}) = 7
The first agent attributes a relatively low utility value to r1 and a high one to r2. Further-
more, the value of both resources together is simply the sum of the individual utilities, i.e.
agent 1 is using an additive utility function (no synergy effects). The second agent ascribes
a medium value to either resource and a very high value to the full set. Now suppose the
initial allocation of resources is A with A(1) = {r1} and A(2) = {r2}. The "inequality
index" for this allocation is u1(A) − u2(A) = 4. We can easily check that inequality is in
fact minimal for allocation A (which means that there can be no inequality-reducing deal,
and certainly no Pigou-Dalton transfer, given this allocation). However, allocation A′ with
A′(1) = {r2} and A′(2) = {r1} would result in a higher level of egalitarian social welfare
(namely 5 instead of 3). Hence, Pigou-Dalton transfers alone are not sufficient to guarantee
optimal outcomes of negotiations in egalitarian agent societies. We need a more general
acceptability criterion.
Intuitively, agents operating according to egalitarian principles should help any of their
fellow agents that are worse off than they are themselves (as long as they can afford to do
so without themselves ending up even worse). This means, the purpose of any exchange of
resources should be to improve the welfare of the weakest agent involved in the respective
deal. We formalise this idea by introducing the class of equitable deals.
Definition 16 (Equitable deals) A deal δ = (A, A′) is called equitable iff it satisfies the
following criterion:
min{ui(A) i ∈ Aδ} < min{ui(A′) i ∈ Aδ}
Recall that Aδ = {i ∈ A A(i) 6= A′(i)} denotes the set of agents involved in the deal δ.
Given that for δ = (A, A′) to be a deal we require A 6= A′, Aδ can never be the empty set
(i.e. the minima referred to in above definition are well-defined). Note that equitability is
a local rationality criterion in the sense of Definition 6.
It is easy to see that any Pigou-Dalton transfer will also be an equitable deal, because
it will always result in an improvement for the weaker one of the two agents concerned.
The converse, however, does not hold (not even if we restrict ourselves to deals involving
only two agents). In fact, equitable deals may even increase the inequality of the agents
concerned, namely in cases where the happier agent gains more utility than the weaker does.
In the literature on multiagent systems, the autonomy of an agent (one of the central
features distinguishing multiagent systems from other distributed systems) is sometimes
equated with pure selfishness. Under such an interpretation of the agent paradigm, our
333
Endriss, Maudet, Sadri, & Toni
notion of equitability would, of course, make little sense. We believe, however, that it is
useful to distinguish different degrees of autonomy. An agent may well be autonomous in
its decision in general, but still be required to follow certain rules imposed by society (and
agreed to by the agent on entering that society).
5.2 Local Actions and their Global Effects
We are now going to prove two lemmas that provide the connection between the local
acceptability criterion given by the notion of equitability and the two egalitarian social
welfare orderings introduced in Section 2.3 (i.e. the maximin-ordering induced by swe as
well as the leximin-ordering).
The first lemma shows how global changes are reflected locally. If a deal happens to
increase (global) egalitarian social welfare, that is, if it results in a rise with respect to the
maximin-ordering, then that deal will in fact be an equitable deal.
Lemma 2 (Maximin-rise implies equitability) If A and A′ are allocations with
swe(A) < swe(A′), then δ = (A, A′) is an equitable deal.
Proof. Let A and A′ be allocations with swe(A) < swe(A′) and let δ = (A, A′). Any
agent with minimal utility for allocation A must be involved in δ, because egalitarian social
welfare, and thereby these agents' individual utility, is higher for allocation A′. That is,
we have min{ui(A) i ∈ Aδ} = swe(A). Furthermore, because Aδ ⊆ A, we certainly have
swe(A′) ≤ min{ui(A′) i ∈ Aδ}. Given our original assumption of swe(A) < swe(A′), we
now obtain the inequation min{ui(A) i ∈ Aδ} < min{ui(A′) i ∈ Aδ}. This shows that δ
will indeed be an equitable deal.
✷
Observe that the converse does not hold; not every equitable deal will necessarily increase
egalitarian social welfare (although an equitable deal will never decrease egalitarian social
welfare either). This is, for instance, not the case if only agents who are currently better off
are involved in a deal. In fact, as argued already at the end of Section 2.2, there can be no
class of deals characterisable by a local rationality criterion (see Definition 6) that would
always result in an increase in egalitarian social welfare.
To be able to detect changes in welfare resulting from an equitable deal we require
the finer differentiation between alternative allocations of resources given by the leximin-
ordering. In fact, as we shall see next, any equitable deal can be shown to result in a strict
improvement with respect to the leximin-ordering.
Lemma 3 (Equitability implies leximin-rise) If δ = (A, A′) is an equitable deal, then
A ≺ A′.
Proof. Let δ = (A, A′) be a deal that satisfies the equitability criterion and define α =
min{ui(A) i ∈ Aδ}. The value α may be considered as partitioning the ordered utility
vector ~u(A) into three subvectors: Firstly, ~u(A) has got a (possibly empty) prefix ~u(A)<α
where all elements are strictly lower than α. In the middle, it has got a subvector ~u(A)=α
(with at least one element) where all elements are equal to α. Finally, ~u(A) has got a suffix
~u(A)>α (which again may be empty) where all elements are strictly greater than α.
334
Negotiating Socially Optimal Allocations of Resources
By definition of α, the deal δ cannot affect agents whose utility values belong to ~u(A)<α.
Furthermore, by definition of equitability, we have α < min{ui(A′) i ∈ Aδ}, which means
that all of the agents that are involved will end up with a utility value which is strictly
greater than α, and at least one of these agents will come from ~u(A)=α. We now collect the
information we have on ~u(A′), the ordered utility vector of the second allocation A′. Firstly,
it will have a prefix ~u(A′)<α identical to ~u(A)<α. This will be followed by a (possibly empty)
subvector ~u(A′)=α where all elements are equal to α and which must be strictly shorter than
~u(A)=α. All of the remaining elements of ~u(A′) will be strictly greater than α. It follows
that ~u(A) lexicographically precedes ~u(A′), i.e. A ≺ A′ holds as claimed.
✷
Again, the converse does not hold, i.e. not every deal resulting in a leximin-rise is necessarily
equitable. Counterexamples are deals where the utility value of the weakest agent involved
stays constant, despite there being an improvement with respect to the leximin-ordering at
the level of society.
A well-known result in welfare economics states that every Pigou-Dalton utility transfer
results in a leximin-rise (Moulin, 1988). Given that we have observed earlier that every deal
that amounts to a Pigou-Dalton transfer will also be an equitable deal, this result can now
also be regarded as a simple corollary to Lemma 3.
5.3 Maximising Egalitarian Social Welfare
Our next aim is to prove a convergence result for the egalitarian framework (in analogy to
Theorems 1 and 4). We are going to show that systems where agents negotiate equitable
deals always converge towards an allocation with maximal egalitarian social welfare.
Theorem 7 (Maximal egalitarian social welfare) Any sequence of equitable deals will
eventually result in an allocation of resources with maximal egalitarian social welfare.
Proof. By Lemma 3, any equitable deal will result in a strict rise with respect to the leximin-
ordering ≺ (which is both irreflexive and transitive). Hence, as there are only a finite number
of distinct allocations, negotiation will have to terminate after a finite number of deals. So
suppose negotiation has terminated and no more equitable deals are possible. Let A be the
corresponding terminal allocation of resources. The claim is that A will be an allocation
with maximal egalitarian social welfare. For the sake of contradiction, assume it is not, i.e.
assume there exists another allocation A′ for the same system such that swe(A) < swe(A′).
But then, by Lemma 2, the deal δ = (A, A′) will be an equitable deal. Hence, there is still
a possible deal, namely δ, which contradicts our earlier assumption of A being a terminal
allocation. This shows that A will be an allocation with maximal egalitarian social welfare,
which proves our claim.
✷
From a purely practical point of view, Theorem 7 may be of a lesser interest than the
corresponding results for utilitarian systems, because it does not refer to an acceptability
criterion that only depends on a single agent. Of course, this coincides with our intu-
itions about egalitarian societies: maximising social welfare is only possible by means of
cooperation and the sharing of information on agents' preferences.
After having reached the allocation with maximal egalitarian social welfare, it may be
the case that still some equitable deals are possible, although they would not increase social
335
Endriss, Maudet, Sadri, & Toni
welfare any further (but they would still cause a leximin-rise). This can be demonstrated
by means of a simple example. Consider a system with three agents and two resources. The
following table fixes the utility functions:
u1({ }) = 0
u1({r1}) = 5
u1({r2}) = 0
u1({r1, r2}) = 5 u2({r1, r2}) = 7.5 u3({r1, r2}) = 9.5
u3({ }) = 8
u3({r1}) = 9
u3({r2}) = 8.5
u2({ }) = 6
u2({r1}) = 7
u2({r2}) = 6.5
A possible interpretation of these functions would be the following. Agent 3 is fairly well off
in any case; obtaining either of the resources r1 and r2 will not have a great impact on its
personal welfare. The same is true for agent 2, although it is slightly less well off to begin
with. Agent 1 is the poorest agent and attaches great value to r1, but has no interest in
r2. Suppose agent 3 initially holds both resources. This corresponds to the ordered utility
vector h0, 6, 9.5i. Passing r1 to agent 1 would lead to a new allocation with the ordered
utility vector h5, 6, 8.5i and increase egalitarian social welfare to 5, which is the maximum
that is achievable in this system. However, there is still another equitable deal that could
be implemented from this latter allocation: agent 3 could offer r2 to agent 2. Of course, this
deal does not affect agent 1. The resulting allocation would then have the ordered utility
vector h5, 6.5, 8i, which corresponds to the leximin-maximal allocation.
To be able to detect situations where a social welfare maximum has already been reached
but some equitable deals are still possible, and to be able to stop negotiation (assuming we
are only interested in maximising swe as quickly as possible), however, we would require a
non-local rationality criterion. No criterion that only takes the welfare of agents involved
in a particular deal into account could be sharp enough to always tell us whether a given
deal would increase the minimum utility in society (see also our discussion after Lemma 2).
We could define a class of strongly equitable deals that are like equitable deals but on top
of that require the (currently) weakest agent to be involved in the deal. This would be
a sharper criterion, but it would also be against the spirit of distributivity and locality,
because every single agent would be involved in every single deal (in the sense of everyone
having to announce their utility in order to be able to determine who is the weakest).
5.4 Necessity Result
As our next theorem will show, if we restrict the set of admissible deals to those that
are equitable, then every single deal δ (that is not independently decomposable) may be
necessary to guarantee an optimal result (that is, no sequence of equitable deals excluding δ
could possibly result in an allocation with maximal egalitarian social welfare). Furthermore,
our theorem improves upon a previous result (Endriss et al., 2003b) by showing that this
holds even when all utility functions are required to be dichotomous.10
Theorem 8 (Necessary deals in egalitarian systems) Let the sets of agents and re-
sources be fixed. Then for every deal δ that is not independently decomposable, there exist
10. This theorem also corrects a mistake in the original statement of the result (Endriss et al., 2003b), where
the restriction to deals that are not independently decomposable had been omitted.
336
Negotiating Socially Optimal Allocations of Resources
utility functions and an initial allocation such that any sequence of equitable deals lead-
ing to an allocation with maximal egalitarian social welfare would have to include δ. This
continues to be the case even when all utility functions are required to be dichotomous.
Proof. Given a set of agents A and a set of resources R, let δ = (A, A′) be any deal for this
system. As we have A 6= A′, there will be a (at least one) agent j ∈ A with A(j) 6= A′(j).
We use this particular j to fix suitable (dichotomous) utility functions ui for agents i ∈ A
and sets of resources R ⊆ R as follows:
ui(R) = (cid:26) 1 if R = A′(i) or (R = A(i) and i 6= j)
0 otherwise
That is, for allocation A′ every agent assigns a utility value of 1 to the resources it holds.
The same is true for allocation A, with the sole exception of agent j, who only assigns a
value of 0. For any other allocation, agents assign the value of 0 to their set of resources,
unless that set is the same as for either allocation A or A′. As δ is not independently
decomposable, this will happen for at least one agent for every allocation different from
both A and A′. Hence, for every such allocation at least one agent will assign a utility value
of 0 to its allocated bundle. We get swe(A′) = 1, swe(A) = 0, and swe(B) = 0 for every
other allocation B, i.e. A′ is the only allocation with maximal egalitarian social welfare.
The ordered utility vector of A′ is of the form h1, . . . , 1i, that of A is of the form
h0, 1, . . . , 1i, and that of any other allocation has got the form h0, . . .i, i.e. we have A ≺ A′
and B (cid:22) A for all allocations B with B 6= A and B 6= A′. Therefore, if we make A the
initial allocation of resources, then δ will be the only deal that would result in a strict rise
with respect to the leximin-ordering. Thus, by Lemma 3, δ would also be the only equitable
deal. Hence, if the set of admissible deals is restricted to equitable deals then δ is indeed
necessary to reach an allocation with maximal egalitarian social welfare.
✷
This result shows, again, that there can be no structurally simple class of deals (such as the
class of deals only involving two agents at a time) that would be sufficient to guarantee an
optimal outcome of negotiation. This is the case even when agents only have very limited
options for modelling their preferences (as is the case for dichotomous utility functions).11
While this negative necessity result is shared with the two other instances of our negotia-
tion framework we have considered, there are currently no positive results on the sufficiency
of 1-deals for restricted domains in the egalitarian setting (see Theorems 3 and 6). For in-
stance, it is not difficult to construct counterexamples that show that even when all agents
are using additive 0-1 functions, complex deals involving all agents at the same time may be
required to reach an allocation with maximal egalitarian social welfare (a concrete example
may be found in Endriss et al., 2003b).
11. However, observe that unlike for our two variants of the framework of rational negotiation, we do not
have a necessity result for scenarios with monotonic utility functions (see Theorems 2 and 5). Using a
collection of monotonic utility functions as in the proof of Theorem 2 would not allow us to draw any
conclusions regarding the respective levels of egalitarian social welfare of A and A′ on the one hand, and
other allocations B on the other.
337
Endriss, Maudet, Sadri, & Toni
6. Negotiating Lorenz Optimal Allocations
In this section, we are going to analyse our framework of resource allocation by negotiation in
view of the notion of Lorenz optimal allocations introduced in Definition 11. We begin with
a somewhat more general discussion of possible local rationality criteria for the acceptability
of a given deal.
6.1 Local Rationality Criteria and Separability
So far, we have studied three different variants of our negotiation framework: (i) rational ne-
gotiation with side payments (aiming at maximising utilitarian social welfare); (ii) rational
negotiation without side payments (aiming at Pareto optimal outcomes); and (iii) negoti-
ation in egalitarian agent societies. The first two instances of our framework, where agents
are either individually rational or cooperatively rational, have been natural choices as they
formalise the widely made assumptions that agents are both purely self-interested and my-
opic (for scenarios with and without monetary side payments, respectively).
The third variant of the framework, which applies to egalitarian agent societies, is at-
tractive for both conceptual and technical reasons. Conceptually, egalitarian social welfare
is of interest, because it has largely been neglected in the multiagent systems literature
despite being the classical counterpart to the widely used notion of utilitarian social wel-
fare. Technically, the analysis of egalitarian agent societies has been interesting, because
egalitarian social welfare does not admit the definition of a local rationality criterion that
directly captures the class of deals resulting in an increase with respect to this metric. This
is why we took the detour via the leximin-ordering to prove termination.
The class of social welfare orderings that can be captured by deals conforming to a
local rationality criterion is closely related to the class of separable social welfare orderings
(Moulin, 1988). In a nutshell, a social welfare ordering is separable iff it only depends on
the agents changing utility whether or not a given deal will result in an increase in social
welfare. Compare this with the notion of a local rationality criterion (Definition 6); here it
only depends on the agents changing bundle whether or not a deal is acceptable. A change
in utility presupposes a change in bundle (but not vice versa). Hence, every separable
social welfare ordering corresponds to a class of deals characterised by a local rationality
criterion (but not vice versa). This means that every separable social welfare ordering gives
rise to a local rationality criterion and proving a general convergence theorem becomes
straightforward.12 The leximin-ordering, for instance, is separable (Moulin, 1988), which is
why we are not going to further discuss this social welfare ordering in this paper.13
Similarly, the ordering over alternative allocations induced by the notion of Lorenz dom-
ination (Definition 11) is also separable. Hence, the definition of an appropriate class of
deals and the proof of a general convergence result would not yield any significant new
insights either. However, here the analysis of the effects that some of the previously intro-
12. Indeed, in the case of the framework of rational negotiation with side payments, the central argument in
the proof of Theorem 1 has been Lemma 1, which shows that individual rationality is in fact equivalent
to the local rationality criterion induced by swu.
13. A suitable rationality criterion would simply amount to a lexicographic comparison of the ordered utility
vectors for the subsociety of the agents involved in the deal in question.
338
Negotiating Socially Optimal Allocations of Resources
duced rationality criteria have on agent societies when social well-being is assessed in terms
of the Lorenz condition is rather instructive, as we are going to see next.
6.2 Lorenz Domination and Existing Rationality Criteria
We are now going to try to establish connections between the global welfare measure induced
by the notion of Lorenz domination on the one hand, and various local criteria on the
acceptability of a proposed deal that individual agents may choose to apply on the other. For
instance, it is an immediate consequence of Definitions 11 and 14 that, whenever δ = (A, A′)
is a cooperatively rational deal, then A must be Lorenz dominated by A′. As may easily be
verified, any deal that amounts to a Pigou-Dalton transfer (see Definition 15) will also result
in a Lorenz improvement. On the other hand, it is not difficult to construct examples that
show that this is not the case for the class of equitable deals anymore (see Definition 16).
That is, while some equitable deals will result in a Lorenz improvement, others will not.
Our next goal is to check whether it is possible to combine existing rationality criteria
and define a class of deals that captures the notion of Lorenz improvements in as so far
as, for any two allocations A and A′ such that A is Lorenz dominated by A′, there exists
a sequence of deals (or possibly even a single deal) belonging to that class leading from
A to A′. Given that both cooperatively rational deals and Pigou-Dalton transfers always
result in a Lorenz improvement, the union of these two classes of deals may seem like a
promising candidate. In fact, according to a result reported by Moulin (1988, Lemma 2.3),
it is the case that any Lorenz improvement can be implemented by means of a sequence of
Pareto improvements (i.e. cooperatively rational exchanges) and Pigou-Dalton transfers. It
is important to stress that this seemingly general result does not apply to our negotiation
framework. To see this, we consider the following example:
u1({ }) = 0
u1({r1}) = 6
u1({r2}) = 1
u1({r1, r2}) = 7 u2({r1, r2}) = 7 u3({r1, r2}) = 10
u3({ }) = 0
u3({r1}) = 1
u3({r2}) = 1
u2({ }) = 0
u2({r1}) = 1
u2({r2}) = 6
Let A be the allocation in which agent 3 owns both resources, i.e. ~u(A) = h0, 0, 10i and
utilitarian social welfare is currently 10. Allocation A is Pareto optimal, because any other
allocation would be strictly worse for agent 3. Hence, there can be no cooperatively rational
deal that would be applicable in this situation. We also observe that any deal involving only
two agents would at best result in a new allocation with a utilitarian social welfare of 7 (this
would be a deal consisting either of passing both resources on to one of the other agents,
or of passing the "preferred" resource to either agent 1 or agent 2, respectively). Hence, no
deal involving only two agents (and in particular no Pigou-Dalton transfer) could possibly
result in a Lorenz improvement. However, there is an allocation that Lorenz dominates
A, namely the allocation assigning to each one of the first two agents their respectively
preferred resource. This allocation A′ with A′(1) = {r1}, A′(2) = {r2} and A′(3) = { } has
got the ordered utility vector h0, 6, 6i.
The reason why the general result reported by Moulin is not applicable to our domain
is that we cannot use Pigou-Dalton transfers to implement arbitrary utility transfers here.
Moulin assumes that every possible utility vector constitutes a feasible agreement. In the
339
Endriss, Maudet, Sadri, & Toni
context of resource allocation, this would mean that there is an allocation for every possible
utility vector. In our framework, where agents negotiate over a finite number of indivisible
resources, however, the range of feasible allocations is limited. For instance, in the above
example there is no feasible allocation with the (ordered) utility vector h0, 4, 6i. In Moulin's
system, agents could first move from h0, 0, 10i to h0, 4, 6i (a Pigou-Dalton transfer) and then
from h0, 4, 6i to h0, 6, 6i (a Pareto improvement). In our system, on the other hand, this is
not possible.
6.3 Simple Pareto-Pigou-Dalton Deals and 0-1 Scenarios
As we cannot use existing rationality criteria to compose a criterion that captures the notion
of a Lorenz improvement (and that would allow us to prove a general convergence theorem),
we are going to investigate how far we can get in a scenario with restricted utility functions.
Recall our definition of 0-1 scenarios where utility functions can only be used to indicate
whether an agent does or does not need a particular resource. As we shall see next, for 0-1
scenarios, the aforementioned result of Moulin does apply. In fact, we can even sharpen it a
little by showing that only Pigou-Dalton transfers and cooperatively rational deals involving
just a single resource and two agents each are required to guarantee negotiation outcomes
that are Lorenz optimal. We first give a formal definition of this class of deals.
Definition 17 (Simple Pareto-Pigou-Dalton deals) A deal δ is called a simple
Pareto-Pigou-Dalton deal iff it is a 1-deal and either cooperatively rational or a Pigou-
Dalton transfer.
We are now going to show that this class of deals is sufficient to guarantee Lorenz optimal
outcomes of negotiations in 0-1 scenarios
Theorem 9 (Lorenz optimal outcomes) In 0-1 scenarios, any sequence of simple
Pareto-Pigou-Dalton deals will eventually result in a Lorenz optimal allocation of resources.
Proof. As pointed out earlier, any deal that is either cooperatively rational or a Pigou-
Dalton transfer will result in a Lorenz improvement (not only in the case of 0-1 scenarios).
Hence, given that there are only a finite number of distinct allocations of resources, after
a finite number of deals the system will have reached an allocation where no more simple
Pareto-Pigou-Dalton deals are possible; that is, negotiation must terminate. Now, for the
sake of contradiction, let us assume this terminal allocation A is not optimal, i.e. there
exists another allocation A′ that Lorenz dominates A. Amongst other things, this implies
swu(A) ≤ swu(A′), i.e. we can distinguish two cases: either (i) there has been a strict
increase in utilitarian welfare, or (ii) it has remained constant. In 0-1 scenarios, the former
is only possible if there are (at least) one resource r ∈ R and two agents i, j ∈ A such that
ui({r}) = 0 and uj({r}) = 1 as well as r ∈ A(i) and r ∈ A′(j), i.e. r has been moved from
agent i (who does not need it) to agent j (who does need it). But then the 1-deal of moving
only r from i to j would be cooperatively rational and hence also a simple Pareto-Pigou-
Dalton deal. This contradicts our assumption of A being a terminal allocation.
Now let us assume that utilitarian social welfare remained constant, i.e. swu(A) =
swu(A′). Let k be the smallest index such that ~uk(A) < ~uk(A′). (This is the first k for
which the inequality in Definition 11 is strict.) Observe that we cannot have k = A, as
340
Negotiating Socially Optimal Allocations of Resources
this would contradict swu(A) = swu(A′). We shall call the agents contributing the first
k entries in the ordered utility vector ~u(A) the poor agents and the remaining ones the
rich agents. Then, in a 0-1 scenario, there must be a resource r ∈ R that is owned by
a rich agent i in allocation A and by a poor agent j in allocation A′ and that is needed
by both these agents, i.e. ui({r}) = 1 and uj({r}) = 1. But then moving this resource
from agent i to agent j would constitute a Pigou-Dalton transfer (and hence also a simple
Pareto-Pigou-Dalton deal) in allocation A, which again contradicts our earlier assumption
of A being terminal.
✷
In summary, we have shown that (i) any allocation of resources from which no simple
Pareto-Pigou-Dalton deals are possible must be a Lorenz optimal allocation and (ii) that
such an allocation will always be reached by implementing a finite number of simple Pareto-
Pigou-Dalton deals. As with our earlier convergence results, agents do not need to worry
about which deals to implement, as long as they are simple Pareto-Pigou-Dalton deals. The
convergence to a global optimum is guaranteed by the theorem.
7. Further Variations
In this section, we are going to briefly consider two further notions of social preference and
discuss them in the context of our framework of resource allocation by negotiation. Firstly,
we are going to introduce the idea of elitist agent societies, where social welfare is tied to
the welfare of the agent that is currently best off. Then we are going to discuss societies
where envy-free allocations of resources are desirable.
7.1 Elitist Agent Societies
Earlier we have discussed the maximin-ordering induced by the egalitarian collective utility
function swe (see Definition 9). This ordering is actually a particular case of a class of social
welfare orderings, sometimes called k-rank dictators (Moulin, 1988), where a particular
agent (the one corresponding to the kth element in the ordered utility vector) is chosen
to be the representative of society. Amongst this class of orderings, another particularly
interesting case is where the welfare of society is evaluated on the basis of the happiest
agent (as opposed to the unhappiest agent, as in the case for egalitarian welfare). We call
this the elitist approach to measuring social welfare.
Definition 18 (Elitist social welfare) The elitist social welfare swel(A) of an allocation
of resources A is defined as follows:
swel(A) = max{ui(A) i ∈ A}
In an elitist agent society, agents would cooperate in order to support their champion (the
currently happiest agent). While such an approach may seem somewhat unethical as far
as human society is concerned, we believe that it could indeed be very appropriate for
certain societies of artificial agents. For some applications, a distributed multiagent system
may merely serve as a means for helping a single agent in that system to achieve its goal.
However, it may not always be known in advance which agent is most likely to achieve its
goal and should therefore be supported by its peers. A typical scenario could be where a
341
Endriss, Maudet, Sadri, & Toni
system designer launches different agents with the same goal, with the aim that at least one
agent achieves that goal -- no matter what happens to the others. As with egalitarian agent
societies, this does not contradict the idea of agents being autonomous entities. Agents may
be physically distributed and make their own autonomous decisions on a variety of issues
whilst also adhering to certain social principles, in this case elitist ones.
From a technical point of view, designing a criterion that would allow agents inhab-
iting an elitist agent society to decide locally whether or not to accept a particular deal
is very similar to the egalitarian case.
In analogy to the case of equitable deals defined
earlier, a suitable deal would have to increase the maximal individual welfare amongst the
agents involved in any one deal. As for the egalitarian case, there can be no class of deals
characterised by a local rationality criterion that would exactly capture the range of deals
resulting in an increase in elitist social welfare (because every agent in the system would
have to be consulted first to determine who is currently best off). To prove convergence,
we would have to resort to an auxiliary social welfare ordering (similarly to the use of the
leximin-ordering in the proof of Theorem 7).
Of course, in many cases there is a much simpler way of finding an allocation with
maximal elitist social welfare. For instance, if all agents use monotonic utility functions,
then moving all resources to the agent assigning the highest utility value to the full bundle
R would be optimal from an elitist point of view. More generally, we can always find an
elitist optimum by checking whose utility function has got the highest peak. That is, while
the highest possible elitist social welfare can easily be determined in a centralised manner,
our distributed approach can still provide a useful framework for studying the process of
actually reaching such an optimal allocation.
7.2 Reducing Envy amongst Agents
Our final example for an interesting approach to measuring social welfare in an agent society
is the issue of envy-freeness (Brams & Taylor, 1996). For a particular allocation of resources,
an agent may be "envious" of another agent if it would prefer that agent's set of resources
over its own. Ideally, an allocation should be envy-free.
Definition 19 (Envy-freeness) An allocation of resources A is called envy-free iff we
have ui(A(i)) ≥ ui(A(j)) for all agents i, j ∈ A.
Like egalitarian social welfare, this is related to the fair division of resources amongst agents.
Envy-freeness is desirable (though not always achievable) in societies of self-interested agents
in cases where agents have to collaborate with each other over a longer period of time. In
such a case, should an agent believe that it has been ripped off, it would have an incentive
to leave the coalition which may be disadvantageous for other agents or the society as a
whole. In other words, envy-freeness plays an important role with respect to the stability of
a group. Unfortunately, envy-free allocations do not always exist. A simple example would
be a system with two agents and just a single resource, which is valued by both of them.
Then whichever agent holds that single resource will be envied by the other agent.
Furthermore, aiming at agreeing on an envy-free allocation of resources is not always
compatible with, say, negotiating Pareto optimal outcomes. Consider the following example
of two agents with identical preferences over alternative bundles of resources:
342
Negotiating Socially Optimal Allocations of Resources
u1({ }) = 0
u1({r1}) = 1
u1({r2}) = 2
u1({r1, r2}) = 0 u2({r1, r2}) = 0
u2({ }) = 0
u2({r1}) = 1
u2({r2}) = 2
For this example, either one of the two allocations where one agent owns all resources and
the other none would be envy-free (as no agent would prefer the other one's bundle over
its own). However, such an allocation would not be Pareto optimal. On the other hand,
an allocation where each agent owns a single resource would be Pareto optimal, but not
envy-free (because the agent holding r1 would rather have r2).
We should stress that envy is defined on the sole basis of an agent's private preferences,
i.e. there is no need to take other agents' utility functions into account. Still, whether an
agent is envious does not just depend on the resources it holds, but also on the resources
it could hold and whether any of the other agents currently hold a preferred bundle. This
somewhat paradoxical situation makes envy-freeness far less amenable to our methodology
than any of the other notions of social welfare we have discussed in this paper.
To be able to measure different degrees of envy, we could, for example, count the number
of agents that are envious for a given allocation. Another option would be to compute for
each agent i that experiences any envy at all the difference between ui(A(i)) and ui(A(j))
for the agent j that i envies the most. Then the sum over all these differences would also
provide an indication of the degree of overall envy (and thereby of social welfare). In the
spirit of egalitarianism, a third option would be identify the degree of envy in society with
the degree of envy experienced by the agent that is the most envious (Lipton, Markakis,
Mossel, & Saberi, 2004). However, it is not possible to define a local acceptability criterion
in terms of the utility functions of the agents involved in a deal (and only those) that
indicates whether the deal in question would reduce envy according to any such a metric.
This is a simple consequence of the fact that a deal may affect the degree of envy experienced
by an agent not involved in the deal at all (because it could lead to one of the participating
agents ending up with a bundle preferred by the non-concerned agent in question).
8. Conclusion
We have studied an abstract negotiation framework where members of an agent society
arrange multilateral deals to exchange bundles of indivisible resources, and we have analysed
how the resulting changes in resource distribution affect society with respect to different
social welfare orderings.
For scenarios where agents act rationally in the sense of never accepting a deal that
would (even temporarily) decrease their level of welfare, we have seen that systems where
side payments are possible can guarantee outcomes with maximal utilitarian social welfare,
while systems without side payments allow, at least, for the negotiation of Pareto optimal
allocations. We have also considered two examples of special domains with restricted utility
functions, namely additive and 0-1 scenarios. In both cases, we have been able to prove the
convergence to a socially optimal allocation of resources also for negotiation protocols that
allow only for deals involving only a single resources and a pair of agents each (so-called
1-deals). In the case of agent societies where welfare is measured in terms of the egalitarian
collective utility function, we have put forward the class of equitable deals and shown that
343
Endriss, Maudet, Sadri, & Toni
negotiation processes where agents use equitability as an acceptability criterion will also
converge towards an optimal state. Another result states that, for the relatively simple
0-1 scenarios, Lorenz optimal allocations can be achieved using one-to-one negotiation by
implementing 1-deals that are either inequality-reducing or that increase the welfare of
both agents involved. We have also discussed the case of elitist agent societies where social
welfare is tied to the welfare of the most successful agent. And finally, we have pointed
out some of the difficulties associated with designing agents that would be able to negotiate
allocations of resources where the degree of envy between the agents in a society is minimal.
Specifically, we have proved the following technical results:
• The class of individually rational deals14 is sufficient to negotiate allocations with
maximal utilitarian social welfare (Theorem 1).
• In domains with additive utility functions, the class of individually rational 1-deals is
sufficient to negotiate allocations with maximal utilitarian social welfare (Theorem 3).
• The class of cooperatively rational deals is sufficient to negotiate Pareto optimal allo-
cations (Theorem 4).
• In domains with 0-1 utility functions, the class of cooperatively rational 1-deals is
sufficient to negotiate allocations with maximal utilitarian social welfare (Theorem 6).
• The class of equitable deals is sufficient to negotiate allocations with maximal egali-
tarian social welfare (Theorem 7).
• In domains with 0-1 utility functions, the class of simple Pareto-Pigou-Dalton deals
(which are 1-deals) is sufficient to negotiate Lorenz optimal allocations (Theorem 9).
For each of the three convergence results that apply to deals without structural restrictions
(rather than to 1-deals), we have also proved corresponding necessity results (Theorems 2, 5,
and 8). These theorems show that any given deal (defined as a pair of allocations) that is not
independently decomposable may be necessary to be able to negotiate an optimal allocation
of resources (with respect to the chosen notion of social welfare), if deals are required
to conform to the rationality criterion in question. As a consequence of these results,
no negotiation protocol that does not allow for the representation of deals involving any
number of agents and any number of resources could ever enable agents (whose behaviour
is constrained by our various rationality criteria) to negotiate a socially optimal allocation
in all cases. Rather surprisingly, all three necessity results continue to apply even when
agents can only differentiate between resource bundles that they would be happy with and
those they they would not be happy with (using dichotomous utility functions). Theorems 2
and 5 also apply in case all agents are required to use monotonic utility functions.
A natural question that arises when considering our convergence results concerns the
complexity of the negotiation framework. How difficult is it for agents to agree on a deal
and how many deals are required before a system converges to an optimal state? The latter
of these questions has recently been addressed by Endriss and Maudet (2005). The paper
14. Recall that individually rational deals may include monetary side payments.
344
Negotiating Socially Optimal Allocations of Resources
establishes upper bounds on the number of deals required to reach any of the optimal allo-
cations of resources referred to in the four convergence theorems for the model of rational
negotiation (i.e. Theorems 1, 3, 4, and 6). It also discusses the different aspects of com-
plexity involved at a more general level (such as the distinction between the communication
complexity of the system, i.e. the amount of information that agents need to exchange to
reach an optimal allocation, and the computational complexity of the reasoning tasks faced
by every single agent). Dunne (2005) addresses a related problem and studies the number
of deals meeting certain structural requirements (in particular 1-deals) that are required to
reach a given target allocation (whenever this is possible at all -- recall that our necessity
results show that excluding certain deal patterns will typically bar agents from reaching
optimal allocations).
In earlier work, Dunne et al. (2005) have studied the complexity of deciding whether
one-resource-at-a-time trading with side payments is sufficient to reach a given allocation
(with improved utilitarian social welfare). This problem has been shown to be NP-hard.
Other complexity results concern the computational complexity of finding a socially opti-
mal allocation, independently from the concrete negotiation mechanism used. As mentioned
earlier, such results are closely related to the computational complexity of the winner deter-
mination problem in combinatorial auctions (Rothkopf, Pekec, & Harstad, 1998; Cramton
et al., 2006). Recently, NP-completeness results for this optimisation problem have been
derived with respect to several different ways of representing utility functions (Dunne et al.,
2005; Chevaleyre et al., 2004). Bouveret and Lang (2005) also address the computational
complexity of deciding whether an allocation exists that is both envy-free and Pareto opti-
mal.
Besides presenting technical results, we have argued that a wide spectrum of social
welfare orderings (rather than just those induced by the well-known utilitarian collective
welfare function and the concept of Pareto optimality) can be of interest to agent-based
applications.
In the context of a typical electronic commerce application, where partic-
ipating agents have no responsibilities towards each other, a system designer may wish
to ensure Pareto optimality to guarantee that agents get maximal payoff whenever this is
possible without making any of the other agents worse off.
In applications where a fair
treatment of all participants is vital (e.g. cases where the system infrastructure is jointly
owned by all the agents), an egalitarian approach to measuring social welfare may be more
appropriate. Many applications are in fact likely to warrant a mixture of utilitarian and
egalitarian principles. Here, systems that enable Lorenz optimal agreements may turn out
to be the technology of choice. Other applications, however, may require social welfare to
be measured in ways not foreseen by the models typically studied in the social sciences.
Our proposed notion of elitist welfare would be such an example. Elitism has little room in
human society, where ethical considerations are paramount, but for a particular computing
application these considerations may well be dropped or changed.
This discussion suggests an approach to multiagent systems design that we call welfare
engineering (Endriss & Maudet, 2004). It involves, firstly, the application-driven choice (or
possibly invention) of a suitable social welfare ordering and, secondly, the design of agent
behaviour profiles and negotiation mechanisms that permit (or even guarantee) socially
optimal outcomes of interactions between the agents in a system. As discussed earlier,
designing agent behaviour profiles does not necessarily contradict the idea of the autonomy
345
Endriss, Maudet, Sadri, & Toni
of an agent, because autonomy always has to be understood as being relative to the norms
governing the society in which the agent operates. We should stress that, while we have been
studying a distributed approach to multiagent resource allocation in this paper, the general
idea of exploring the full range of social welfare orderings when developing agent-based
applications also applies to centralised mechanisms (such as combinatorial auctions).
We hope to develop this methodology of welfare engineering further in our future work.
Other possible directions of future work include the identification of further social welfare
orderings and the definition of corresponding deal acceptability criteria; the continuation of
the complexity-theoretic analysis of our negotiation framework; and the design of practical
trading mechanisms (including both protocols and strategies) that would allow agents to
agree on multilateral deals involving more than just two agents at a time.
Acknowledgments
We would like to thank J´erome Lang and various anonymous referees for their valuable
comments. This research has been partially supported by the European Commission as
part of the SOCS project (IST-2001-32530).
References
Andersson, M., & Sandholm, T. W. (2000). Contract type sequencing for reallocative nego-
tiation. In Proceedings of the 20th International Conference on Distributed Computing
Systems (ICDCS-2000), pp. 154 -- 160. IEEE.
Arrow, K. J., Sen, A. K., & Suzumura, K. (Eds.). (2002). Handbook of Social Choice and
Welfare, Vol. 1. North-Holland.
Bouveret, S., & Lang, J. (2005). Efficiency and envy-freeness in fair division of indivisible
goods: Logical representation and complexity.
In Proceedings of the 19th Interna-
tional Joint Conference on Artificial Intelligence (IJCAI-2005), pp. 935 -- 940. Morgan
Kaufmann Publishers.
Brams, S. J., & Taylor, A. D. (1996). Fair Division: From Cake-cutting to Dispute Resolu-
tion. Cambridge University Press.
Chavez, A., Moukas, A., & Maes, P. (1997). Challenger: A multi-agent system for distributed
resource allocation. In Proceedings of the 1st International Conference on Autonomous
Agents (Agents-1997), pp. 323 -- 331. ACM Press.
Chevaleyre, Y., Endriss, U., Estivie, S., & Maudet, N. (2004). Multiagent resource allocation
with k-additive utility functions. In Proceedings of the DIMACS-LAMSADE Work-
shop on Computer Science and Decision Theory, Vol. 3 of Annales du LAMSADE,
pp. 83 -- 100.
Chevaleyre, Y., Endriss, U., Lang, J., & Maudet, N. (2005a). Negotiating over small bundles
of resources. In Proceedings of the 4th International Joint Conference on Autonomous
Agents and Multiagent Systems (AAMAS-2005), pp. 296 -- 302. ACM Press.
Chevaleyre, Y., Endriss, U., & Maudet, N. (2005b). On maximal classes of utility functions
In Proceedings of the 19th International Joint
for efficient one-to-one negotiation.
346
Negotiating Socially Optimal Allocations of Resources
Conference on Artificial Intelligence (IJCAI-2005), pp. 941 -- 946. Morgan Kaufmann
Publishers.
Cramton, P., Shoham, Y., & Steinberg, R. (Eds.). (2006). Combinatorial Auctions. MIT
Press.
Dunne, P. E. (2005). Extremal behaviour in multiagent contract negotiation. Journal of
Artificial Intelligence Research, 23, 41 -- 78.
Dunne, P. E., Wooldridge, M., & Laurence, M. (2005). The complexity of contract negoti-
ation. Artificial Intelligence, 164 (1 -- 2), 23 -- 46.
Dunne, P. E., Laurence, M., & Wooldridge, M. (2004). Tractability results for automatic
contracting. In Proceedings of the 16th Eureopean Conference on Artificial Intelligence
(ECAI-2004), pp. 1003 -- 1004. IOS Press.
Endriss, U., & Maudet, N. (2004). Welfare engineering in multiagent systems. In Engineering
Societies in the Agents World IV, Vol. 3071 of LNAI, pp. 93 -- 106. Springer-Verlag.
Endriss, U., & Maudet, N. (2005). On the communication complexity of multilateral trading:
Extended report. Journal of Autonomous Agents and Multiagent Systems, 11 (1), 91 --
107.
Endriss, U., Maudet, N., Sadri, F., & Toni, F. (2003a). On optimal outcomes of nego-
tiations over resources.
In Proceedings of the 2nd International Joint Conference
on Autonomous Agents and Multiagent Systems (AAMAS-2003), pp. 177 -- 184. ACM
Press.
Endriss, U., Maudet, N., Sadri, F., & Toni, F. (2003b). Resource allocation in egalitar-
In Secondes Journ´ees Francophones sur les Mod`eles Formels
ian agent societies.
d'Interaction (MFI-2003), pp. 101 -- 110. C´epadu`es-´Editions.
Fatima, S. S., Wooldridge, M., & Jennings, N. R. (2004). An agenda-based framework for
multi-issues negotiation. Artificial Intelligence, 152 (1), 1 -- 45.
Fishburn, P. C. (1970). Utility Theory for Decision Making. John Wiley and Sons.
Fujishima, Y., Leyton-Brown, K., & Shoham, Y. (1999). Taming the computational com-
In Pro-
plexity of combinatorial auctions: Optimal and approximate approaches.
ceedings of the 16th International Joint Conference on Artificial Intelligence (IJCAI-
1999), pp. 548 -- 553. Morgan Kaufmann Publishers.
Kraus, S. (2001). Strategic Negotiation in Multiagent Environments. MIT Press.
Lemaıtre, M., Verfaillie, G., & Bataille, N. (1999). Exploiting a common property resource
under a fairness constraint: A case study. In Proceedings of the 16th International Joint
Conference on Artificial Intelligence (IJCAI-1999), pp. 206 -- 211. Morgan Kaufmann
Publishers.
Lipton, R. J., Markakis, E., Mossel, E., & Saberi, A. (2004). On approximately fair allo-
cations of indivisible goods. In Proceedings of the 5th ACM Conference on Electronic
Commerce (EC-2004), pp. 125 -- 131. ACM Press.
Moulin, H. (1988). Axioms of Cooperative Decision Making. Cambridge University Press.
347
Endriss, Maudet, Sadri, & Toni
Myerson, R. B., & Satterthwaite, M. A. (1983). Efficient mechanisms for bilateral trading.
Journal of Economic Theory, 29 (2), 265 -- 281.
Rawls, J. (1971). A Theory of Justice. Oxford University Press.
Rosenschein, J. S., & Zlotkin, G. (1994). Rules of Encounter. MIT Press.
Rothkopf, M. H., Pekec, A., & Harstad, R. M. (1998). Computationally manageable com-
binational auctions. Management Science, 44 (8), 1131 -- 1147.
Sadri, F., Toni, F., & Torroni, P. (2001). Dialogues for negotiation: Agent varieties and dia-
logue sequences. In Proceedings of the 8th International Workshop on Agent Theories,
Architectures, and Languages (ATAL-2001), pp. 405 -- 421. Springer-Verlag.
Sandholm, T. W. (2002). Algorithm for optimal winner determination in combinatorial au
ctions. Artificial Intelligence, 135, 1 -- 54.
Sandholm, T. W. (1998). Contract types for satisficing task allocation: I Theoretical results.
In Proceedings of the AAAI Spring Symposium: Satisficing Models.
Sandholm, T. W. (1999). Distributed rational decision making. In Weiss, G. (Ed.), Multia-
gent Systems: A Modern Approach to Distributed Artificial Intelligence, pp. 201 -- 258.
MIT Press.
Sen, A. K. (1970). Collective Choice and Social Welfare. Holden Day.
Smith, R. G. (1980). The contract net protocol: High-level communication and control in a
distributed problem solver. IEEE Transactions on Computers, C-29 (12), 1104 -- 1113.
Wooldridge, M. (2002). An Introduction to MultiAgent Systems. John Wiley and Sons.
348
|
1504.03809 | 4 | 1504 | 2016-09-09T01:34:09 | Unperturbed Schelling segregation in two or three dimensions | [
"cs.MA"
] | Schelling's model of segregation, first described in 1969, has become one of the best known models of self-organising behaviour. While Schelling's explicit concern was to understand the mechanisms underlying racial segregation in large cities from a game theoretic perspective, the model should be seen as one of a family, arising in fields as diverse as statistical mechanics, neural networks and the social sciences, and which are concerned with interacting populations situated on network structures. Despite extensive study, however, the (unperturbed) Schelling model has largely resisted rigorous analysis, prior results in the literature generally pertaining to variants of the model in which noise is introduced into the dynamics of the system, the resulting model then being amenable to standard techniques from statistical mechanics or stochastic evolutionary game theory. A series of recent papers (one by Brandt, Immorlica, Kamath, and Kleinberg, and two by the authors), has seen the first rigorous analysis of the one dimensional version of the unperturbed model. Here we provide the first rigorous analysis of the two and three dimensional unperturbed models, establishing most of the phase diagram, and answering a challenge from a recent paper by Brandt, Immorlica, Kamath, and Kleinberg. | cs.MA | cs |
UNPERTURBED SCHELLING SEGREGATION IN TWO OR THREE DIMENSIONS
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
Abstract. Schelling's models of segregation, first described in 1969 [18] are among the best known
models of self-organising behaviour. Their original purpose was to identify mechanisms of urban racial
segregation. But his models form part of a family which arises in statistical mechanics, neural networks,
social science, and beyond, where populations of agents interact on networks. Despite extensive study,
unperturbed Schelling models have largely resisted rigorous analysis, prior results generally focusing
on variants in which noise is introduced into the dynamics, the resulting system being amenable to
standard techniques from statistical mechanics or stochastic evolutionary game theory [25]. A series
of recent papers [6, 3, 4], has seen the first rigorous analyses of 1-dimensional unperturbed Schelling
models, in an asymptotic framework largely unknown in statistical mechanics. Here we provide the first
such analysis of 2- and 3-dimensional unperturbed models, establishing most of the phase diagram, and
answering a challenge from [6].
1. Introduction
Schelling's spatial proximity models of segregation [18] provided strikingly simple examples the emergence
of order from randomness. Subsequently, and particularly following Young [25], a range of variations of his
models have been studied by mathematicians, statistical physicists, computer scientists, social scientists,
and others, making Schelling segregation one of the best known theoretical examples of self-organising
behaviour (indeed, this was cited by the committee upon awarding Schelling the Nobel memorial prize
for Economics in 2005).
Schelling models have the following general set-up: a population of individuals of two types are initially
randomly distributed on a grid (or other graph). Each individual considers a certain region around it to
be its neighbourhood, and it has an intolerance (τ ), expressing the proportion of its neighbours it requires
to be of its own type in order to be happy.
Unhappy individuals then rearrange themselves in order to become happy. This may happen through
a variety of possible dynamics.
In Schelling's own work on 2-dimensional segregation, at each time
step a selected unhappy node would move to a vacant position where it would be happy. However,
much subsequent work on Schelling segregation has followed the breakthrough analysis of Young [25]
who studied a 1-dimensional model which dispensed with vacancies, where agents instead rearranged
themselves via pairwise swaps (at each time-step a randomly selected pair of unhappy agents of opposite
types swap positions).
This version: August 21, 2018.
Key words and phrases. Schelling Segregation and Algorithmic Game Theory and Complex Systems and Non-linear
Dynamics and Ising model and Spin Glass.
Authors are listed alphabetically. Barmpalias was supported by the 1000 Talents Program for Young Scholars from
the Chinese Government, and the Chinese Academy of Sciences (CAS) President's International Fellowship Initiative No.
2010Y2GB03. Additional support was received by the CAS and the Institute of Software of the CAS. Partial support was
also received from a Marsden grant of New Zealand and the China Basic Research Program (973) grant No. 2014CB340302.
Andy Lewis-Pye (previously Andrew Lewis) was supported by a Royal Society University Research Fellowship. The authors
thank Sandro Azaele for some helpful comments on an earlier draft of this paper.
1
2
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
The lack of vacancies is arguably a closer approximation to life in modern cities without large numbers of
uninhabited properties, and has proved itself more amenable to mathetatical analysis. Numerous variants
of Young's model have subsequently been investigated, notably in the work of Zhang [26, 27, 28], as well
as the series of papers of which the current work is a part [6, 3, 4].
Young's pairwise-swapping dynamic is known to statistical physicists as the Kawasaki dynamic. An even
simpler choice, adopted in [4] and the current paper, is the Glauber dynamic, whereby at each time-step a
single randomly selected unhappy agent switches type. Although simple, this is by no means unintuitive,
the idea is that an agent who becomes unhappy moves out of the city, and is replaced by an agent
of the opposite type. Thus our model is an open system, and we posit limitless pools of both types
beyond the system. The Glauber dynamic also, we argue, brings the model closest to others in the same
statistical-physical family, such as the Ising and Hopfield models (see 1.2 below).
Ever since Schelling's original simulations with zinc and copper coins on a checkerboard back in 1969,
the robustness of the phenomenon of segregation has been recognised: as the process unfolds, large
clusters emerge, consisting of individuals of only one type. This same phenomenon has appeared in
diverse variants of the model, strikingly including models in which agents have an active preference for
integration [27, 16]. Seen from a game theoretic perspective, this provides an example of a recurrent
theme in Schelling's research -- especially as elaborated upon in [19] -- that individuals acting according
to their interests at the local level can produce global results which may be unexpected and undesired by
all. Understanding this phenomenon has been the goal of much subsequent research.
Besides changing the dynamic, Young in [25], made another highly influential innovation by introducing
noise to the model. Much subsequent research has focussed on such perturbed variants, which can be
thought of as systems of temperature T > 0, in which agents have a small but non-zero probability of
acting against their own interests. Young used techniques from evolutionary game theory -- an analysis
in terms of stochastically stable states -- to analyse such a model. These ideas were then substantially
developed in the work of Zhang [26, 27, 28]. While the language used may differ from that of those of sta-
tistical physicists, the basic analysis is essentially equivalent: Zhang establishes a Boltzmann distribution
for the set of configurations, and then his stochastically stable states correspond to ground states.
1.1. Our contribution. The difficulty in analysing the unperturbed (or temperature T = 0) variants of
the model stems from the large number of absorbing states for the underlying Markov process. Neverthe-
less, a breakthrough came in [6], where Brandt, Immorlica, Kamath and Kleinberg used an analysis of
locally defined stable configurations, combined with results of Wormald [24], to provide the first rigorous
analysis of an unperturbed 1-dimensional Schelling model, for the case τ = 0.5 under Kawasaki dynam-
ics. In [3] the authors gave a more general analysis of the same model for τ ∈ [0, 1]. In [4], the authors
then analysed variants of the model under Glauber dynamics (as well as variations thereupon), with an
additional innovation: the two types of agent may have unequal intolerances τα, τβ ∈ [0, 1]. We extend
the analysis of this phenomenon in the current paper. This modification of the model has justification in
social research. See for example [20], where it is found that black US citizens are happier in integrated
neighbourhoods than their white compatriots. It has often been argued (including by the authors in [4]
and Schelling himself in for example [19]) that the simplicity of Schelling models make them relevant in
areas far beyond the topic of racial segregation. Another application might be that of product adoption:
which of two competing products customers choose. The peer-effect aspect of this phenomenon can be
addressed with a Schelling model. In such a context, the asymmetry of τα, τβ reflects the idea of 'diffi-
culty of engagement', which is argued in [7] to offer insights into the obstacles that innovative ecological
products face in terms of market penetration.
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
3
Figure 1. The evolving process: n = 600, w = 5, τα = 0.44, τβ = 0.42.
Throughout this series of papers, although the proofs vary widely in their details, the analysis hinges on
the use of a spread of models across which the neighbourhood radius (w) size grows large (but remains
small relative to the size of the whole system, n). This allows exact results to be obtained asymptotically,
by careful probabilistic analysis of various structures occurring in the initial configuration. (It might
be objected, at this point, that neighbourhoods of arbitrarily large size are not in the spirit of models
of racial segregation. However, this should be seen as a mathematical device allowing the outcome to
be predicted with arbitrarily high precision. In simulations of all models in this family, relatively small
values of w are generally sufficient for a clear outcome. See Figures 1, 2, 3 for examples.)
The current paper uses the same approach to extend these results to the two and three dimensional
models. A mathematical account of unperturbed Schelling segregation in higher dimensions has been
seen as a challenging problem for some time. As Brandt, Immorlica, Kamath and Kleinberg say in [6]:
Finally, and most ambitiously, there is the open problem of rigorously analyzing the Schelling model in
other graph structures including two-dimensional grids. Simulations of the Schelling model in two dimen-
sions reveal beautiful and intricate patterns that are not well understood analytically. Perturbations of
the model have been successfully analyzed using stochastic stability analysis [. . . ] but the non-perturbed
model has not been rigorously analyzed. Two-dimensional lattice models are almost always much more
challenging than one-dimensional ones, and we suspect that to be the case with Schelling's segregation
model. But it is a challenge worth undertaking: if one is to use the Schelling model to gain insight into
the phenomenon of residential segregation, it is vital to understand its behavior on two-dimensional grids
since they reflect the structure of so many residential neighborhoods in reality.
Thus the aim of this paper is to answer this challenge, providing the first rigorous analysis of the two and
three dimensional unperturbed models of Schelling segregation. Our work gives an almost comprehensive
picture of the phase diagram of such a model under Glauber dynamics with distinct intolerances τα
and τβ, revealing interesting phase transitions around certain thresholds, which are solutions to derivable
equations and which we numerically approximate. Some grey areas around these thresholds remain, where
the behaviour of the model is not proven rigorously, but these correspond to relatively small intervals.
Let us briefly compare the results of the current paper (namely Theorems 1.2 and 1.3 below) with those
of [4] where the authors considered a 1-dimensional version of the same model. This can be summed up
by comparing the two illustrations of Figure 4 (representing the 2- and 3-dimensional models) with the
final illustration within Figure 4 of [4] representing the 1-dimensional model. The overall picture is very
similar; the differences are in the precise locations of the thresholds between staticity almost everywhere
and takeover almost everywhere (around 0.3531 in the 1-dimensional case, 0.3652 in the 2-dimensional
case, and 0.3897 in the 3-dimensional case), as well as the fact that in the 2- and 3- dimensional model we
4
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
Figure 2. The evolving process: n = 600, w = 10, τα = 0.44, τβ = 0.42.
continue to have small grey areas around those thresholds where the outcome has not yet been rigorously
established. This will be explained further in Section 1.3.2 below.
Following the online release of an earlier version of the current paper, Immorlica, Kleinberg, Lucier, and
Zadomighaddam released an unpublished manuscript1 which analyses a closely related, but interestingly
distinct, two dimensional Schelling system. They show, under a continuous time Poisson clock dynamic,
and with the hypothesis that agents of both types share the same intolerance which is close to, but less
than 0.5 (i.e. τα = τβ = 1−ε
2 ), that the size of the segregated regions which emerge will, with probability
approaching 1, be of exponential size in the neighbourhood radius (eΘ(w2)).
1.2. The Schelling model across science. While Schelling's ultimate concern was to understand some
of the mechanisms underlying racial segregation in the United States, such a model may also be applied
to many contexts in which the 'individuals' represent particles or agents of a sort which might more
plausibly behave according to simple rules which govern their behaviour at the local level. In fact these
model are rightly seen as fitting within a larger family of discrete time models arising in diverse fields.
Before we give a formal description of our model, we review two other important examples.
1.2.1. Ising Models. Many authors have pointed out direct links to the Ising model, used to analyse
phase transitions in the context of statistical mechanics (see [21, 10, 17, 11, 15]). As this observation
suggests, the dynamics can fruitfully be analysed by assigning an energy level to any given configuration,
typically corresponding to some measure of the mixing of types. Typically, the Ising model is considered
at temperature T > 0, meaning that while transitions are more much likely to occur from configurations
with higher energy to those with lower energy, they may occasionally also occur in the opposite direction.
As discussed above, this corresponds to a perturbed Schelling model, and eases the analysis.
In models with an unperturbed dynamics (i.e. with T = 0), which are the focus of the current paper,
transitions which increase the energy never occur. Thus our results fit into the statistical physical theory
of rapid cooling.
There is another important remark to make. In the current paper (as in [4]), we consider models in which
individuals of different type may have unequal intolerances, τα and τβ. As discussed in Section 1.3. of [4],
this equates to an Ising model with a non-standard external field whose strength differs for spins in state
+1 spins compared to those in state -1. Assuming the usual notation for an Ising model, if we attempt
1N. Immorlica, R. Kleinberg, B. Lucier, M. Zadomighaddam Exponential Segregation in a Two-Dimensional Schelling
Model with Tolerant Individuals, preprint, arXiv: 1511.02537.
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
5
to construct a Hamiltonian function in the standard way:
H = − (cid:88)
(cid:88)
i
σiK(σi)
σiσj +
i−j∞≤w
we find that in general K(+1) (cid:54)= K(−1), meaning that H does not act as an energy function, since it
may increase as well as decrease (even in the unperturbed case).
Thus, our results below can be understood in relation to an Ising model under rapid cooling, with range
of interaction w and a non-standard external field as described.
1.2.2. Neural Networks. Hopfield networks provide an example of a model from the same family, in
which the unperturbed dynamics are of particular interest. Introduced by Hopfield [13] in 1982 these are
recurrent artificial neural networks, which have been much studied as a form of associative memory. Here,
in fact, the unperturbed dynamics would seem to be an essential aspect of the intended functionality --
corresponding to any initial state the absorbing state reached within a finite number of steps of the
dynamical process is the 'memory' which the net is considered as associating with that input. Many
authors have analysed the connections to spin-glass models [1]. A major difference with the Schelling
model though, is that the standard form of Hopfield network has no geometry: each threshold node takes
inputs from all others. From a biological perspective, however, it is of interest to study models of sparse
connectivity [9, 8]. The results of this paper can easily be adapted so as to correspond to a form of 'local'
Hopfield net, in which nodes arranged on a grid are only locally connected.
1.2.3. Cascading Phenomena. Our own avenue into these questions came via connections to the study
of cascading phenomena on networks as studied by Barab´asi, Kleinberg and many others, a good in-
troduction to which can be found in [14]. The dynamics of the Schelling process are either identical or
almost identical to versions of the Threshold Model used to model the flow of information, technology,
behaviour, viruses, opinions etc., on large real world networks. Depending on context, both the perturbed
and unperturbed dynamics may be of interest here. The principal difference with the Schelling model
is typically underlying network, instead of a regular lattice, a random graph of some form (e.g. Watts-
Strogatz graph [23], or Barab´asi-Albert preferential attachment [2]), which better reflects the clustering
coefficient and degree distribution of the real world network in question.
Given the results of this paper, an immediate question is whether the techniques developed here can be
applied to understand emergent phenomena on such random structures. Along these lines, Henry, Pra(cid:32)lat
and Zhang have described a simple but elegant model of network clustering [12], inspired by Schelling
segregation. Although an interesting system, it does not display the kind of involved threshold behaviour
that one might expect; thus there may be more to be said here.
1.3. The model. We describe the two dimensional model first, and it is then simple to extend to three
dimensions. For the two dimensional model we consider an n × n 'grid' of nodes. To avoid boundary
issues, we work on the flat torus T = [0, n− 1)× [0, n− 1), i.e. R2/ ≡ where (x, y) ≡ (x + n, y) ≡ (x, y + n).
Thus, in the context of discussing Euclidean coordinates, arithmetical operations are always performed
modulo n. We let Rn and Nn denote the reals and the natural numbers modulo n respectively. Formally,
then, 'nodes' are elements of N2
n. In the initial configuration (at stage 0) each node is assigned one of
two types, either α or β. In this initial configuration the types of nodes are independent and identically
distributed, with each node having probability 0.5 of being type α.
For each node we also consider a certain neighbourhood, with size specified by a parameter w: the
neighbourhood of the node u, denoted N (u) is the set of nodes v such that u − v∞ ≤ w. So far,
6
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
Static almost everywhere
With high probability a random node does not
change its type throughout the process.
α (or β) takeover almost everywhere
A random node has high probability of being of type
α (β respectively) in the final configuration.
α (or β) takeover totally
With high probability all nodes are of type α (β re-
spectively) in the final configuration.
Table 1. The three (or five) behaviours of the model.
then, we have specified two parameters for the model, n and w. The two remaining parameters are the
intolerance levels τα, τβ ∈ [0, 1]. At any given stage we say that a node u of type α is happy if the
proportion of the nodes in N (u) which are of type α is at least τα (note that u is included in its own
neighbourhood). Similarly a node u of type β is happy if the proportion of the nodes in N (u) which
are of type β is at least τβ. We say that a node is hopeful if it is not happy, but would be happy if it
changed type (the types of all other nodes remaining unchanged). The dynamical process then unfolds
as follows. At each stage s + 1, we consider the set of hopeful nodes at the end of stage s. Taking each
in turn we change the type, providing it is still hopeful given earlier changes during stage s + 1. The
process terminates when there are no remaining hopeful nodes.
In the description of the dynamic process above, we have stated that during each stage s + 1, each node
which was hopeful at the end of stage s is taken in turn, the type then being changed if it still remains
hopeful given earlier changes during stage s + 1. For the sake of definiteness, we may consider the nodes
to be lexicographically ordered, but this choice is essentially arbitrary, and it will be clear from what
follows, that all results apply for any ordering of the nodes (since the particular form of the ordering is
not used in the proof of any lemma). In fact the model will behave in an almost identical fashion if,
instead of taking the hopeful nodes in order at each stage, we instead pick hopeful nodes uniformly at
random. This fact is easy verified by simulation, and our expectation is that it would not be difficult
to use techniques applied in [4] to extend the results presented here to such a dynamics. The choice to
consider nodes in order at each stage, simply means that certain aspects of the dynamics become easier to
analyse. Roughly, this is because we are immediately guaranteed that the process will unfold at the same
rate at different locations on the grid, rather than having to apply a probabilistic analysis to show that
formal statements to this effect will be close to true most of the time. It is important for the analysis,
however, that the set of hopeful nodes considered during stage s + 1, is fixed at the end of stage s --
thereby avoiding 'waves' of changes traversing large distances in a single stage.
It is easy to define an appropriate Lyapunov function, establishing that the process must eventually
terminate. For example, one may consider the sum over all nodes u, of the number nodes of the same
type as u in N (u). The three dimensional model is defined identically, except that nodes are now elements
n. Once again, N (u) is the set of nodes v such that u − v∞ ≤ w, but obviously the (cid:96)∞ metric is
of N3
now understood in three dimensions.
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
7
Figure 3. Simulations showing three behaviors of the model, with α and β nodes de-
picted as red and green respectively. From left to right we have "static almost every-
where", "β takeover almost everywhere" and "β takeover totally". These are the final
states of the model with n = 600, w = 2 and (τα, τβ) equal to (0.249,0.1), (0.4,0.3) and
(0.6, 0.4) respectively.
Note that the numbers of nodes of each type do not remain stable throughout the process. Although the
initial expectation is 50% of each type, during the process it is conceivable that one type may increase
or decrease its population. In fact, intuitively one would expect that the type with the least intolerance
level has an advantage over the other type, as it is less susceptible to changes. As we are going to see in
the next section, this is largely true but there are several caveats. For example if both intolerance levels
are very low or very high, then one would expect the system to remain largely stable. In the following
section we make precise statements about what can be rigorously proved, establishing most of the phase
diagram.
1.3.1. Behaviors of the model. Our results are asymptotic in nature, and we will use the shorthand "for
0 (cid:28) w (cid:28) n" to mean "for all sufficiently large w, and all n sufficiently large compared to w". By a
scenario we mean the class of all instances of the model with fixed values of τα, and τβ, but w and n
varying. We will identify a scenario with its signature pair (τα, τβ).
The phase diagram of the model will be expressed according to five behavior types, indicating the degree
of growth of one population over the other that can occur by the time the model reaches its final state.
Our results and analysis justify this natural choice of behavioural classification. Table 1 shows the names
we give to the different behaviors of the model, along with their corresponding descriptions. The precise
definitions are given in Definition 1.1.
Definition 1.1 (Five different behaviors). We list the three (or five, taking into account the two types)
main ways in which a scenario can behave:
(i) A scenario is static almost everywhere if the following holds for every > 0: for 0 (cid:28) w (cid:28) n a node
u chosen uniformly at random has a probability > 1− of having its type unchanged throughout the
process.
(ii) Type α (β) takes over almost everywhere if the following holds for every > 0: for 0 (cid:28) w (cid:28) n
a node u chosen uniformly at random has a probability > 1 − of being of type α (β) in the final
configuration.
(iii) Type α (β) takes over totally if the following holds for every > 0: for 0 (cid:28) w (cid:28) n the probability
that all nodes are of type α (β) in the final configuration exceeds 1 − .
8
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
Figure 4. The phase diagrams of the two and the three dimensional models respectively.
Note that in the latter we have a larger grey region, corresponding to behaviour which
we have not been able to rigorously establish.
Figure 3 shows the final states of the model for three different choices of signature (τα, τβ). The first state
indicates a "static almost everywhere" behavior, with only occasional areas where nodes have switched
to type β. The second state indicates an "almost everywhere β takeover", with the occasional small areas
where nodes have succeeded in keeping the α-type. Finally the third state is a total β takeover. The
signatures for the corresponding behaviors illustrate Theorem 1.2 of the next section.
In general, when τα < τβ, we might expect the α type to be more persistent in the process. In order
to formally establish results expressing such expectations, we often need a stronger notion of inequality,
expressing that (for example) τα is sufficiently less than τβ.
In the case of the 2-dimensional model
we denote this relation by τα (cid:67) τβ (the formal definition is given in Definition 1.4). In the case of the
3-dimensional model sufficient inequality is a slightly stronger condition, and is denoted by τα(cid:50) τβ (the
formal definition is given in Definition 1.5). Note that 'sufficiently less' has different meanings for the
2-dimensional and 3-dimensional models. This will not cause confusion as the two models are dealt with
separately, and we use different symbols for the two notions of inequality. The second item of Figure 5
illustrates the relation between sufficient inequality for the 2-dimensional and 3-dimensional models with
the usual inequality.
1.3.2. Main theorems. Our main result for the two dimensional model is the following classification of
behaviors of the model, according to the signature (τα, τβ). The different cases correspond to the positions
of the two intolerance levels with respect to each other and other values like 0.5, 0.25 and a certain constant
κ (cid:117) 0.365227 that will be formally defined (as the unique solution of a certain equation) shortly. Recall
that (cid:67) indicates the 'sufficiently less' relation for the two dimensional model (see Definition 1.4). Given
the symmetric roles of α and β, it is clear that the following theorem (and all those which follow) also
holds with the roles of α and β reversed. From now on we shall not explicitly mention this fact, and will
leave it to the reader to fill in these symmetries.
Theorem 1.2 (Behavior of the two dimensional model). The behavior of the two dimensional model is
dictated by the signature (τα, τβ) as follows:
(a) If τα, τβ < 0.25 then the scenario is static almost everywhere.
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
9
(b) If κ < τα < 0.5 and τβ (cid:67) τα then β takes over almost everywhere.
(c) For τβ < 0.5 < τα, β takes over totally.
(d) If 1 − κ > τα > 0.5 and τβ (cid:66) τα then α takes over almost everywhere.
(e) If τα, τβ > 0.75 then the scenario is static almost everywhere.
where κ is the unique solution in [0, 1] to (1−2κ)1−2κ = 22(1−κ)κκ(1−κ)3(1−κ) (numerically κ (cid:117) 0.365227).
So if τα and τβ are both sufficiently small then most nodes remain unchanged throughout the process.
If at least one of τα and τβ is reasonably large while remaining in the interval (0, 0.5), however, then the
situation changes dramatically. Here 'reasonably large' means above the threshold κ. While it may not
be immediately obvious where the equation defining κ comes from, it will be derived later by comparing
the probabilities of certain structures in the initial configuration.
Figures 1 and 2 display the evolving process for simulations corresponding to the scenario of clause (b)
of Theorem 1.2.2 Here nodes of type α are depicted red, while those of type β are green. The shade
of red or green corresponds to the number of same-type nodes within N (u): the brighter the shade the
more nodes there are of the same type as u in N (u). Since τβ (cid:67) τα, most nodes are of type β in the final
configuration, with this proportion tending to 1 as w → ∞.
Using almost identical proofs, we can get slightly weaker results for the three dimensional model.
Theorem 1.3 (Behavior of the three dimensional model). The behavior of the three dimensional model
is dictated by the signature (τα, τβ) as follows:
(a) If τα, τβ < 0.25 then the scenario is static almost everywhere.
(b) If κ∗ < τα < 0.5 and τβ(cid:50) τα then β takes over almost everywhere.
(d) If 1 − κ∗ > τα > 0.5 and τβ(cid:51) τα then α takes over almost everywhere.
(c) For τβ < 0.5 < τα, then β takes over totally.
(e) If τα, τβ > 0.75 then the scenario is static almost everywhere.
where κ∗ is the unique solution in [0, 1] to (1− 2κ∗)4(1−2κ∗) = 223−8κ∗ κ19κ∗
κ∗ (cid:117) 0.3897216).
∗
(1− κ∗)27(1−κ∗) (numerically,
Note that our results for the three dimensional model are the same as for the two dimensional model
except that the thresholds are slightly different (making the gaps in the phase diagram slightly larger).
Sufficient inequality for the 2D and 3D models. In this section we formally define the notions of
sufficient inequality (denoted by (cid:67) and (cid:50)) for the two dimensional and the three dimensional models
respectively. The particulars of the definition result from comparing the probabilities of certain structures
in the initial configuration, and the motivation behind the definition will become clear in Section 2.
Definition 1.4 (Sufficient inequality for 2D). We let g(x, k) = xkx(1 − x)k(1−x). For τ0, τ1 ∈ (0, 0.5) we
say that τ0 is sufficiently less than τ1, denoted τ0 (cid:67) τ1, if g(τ0, 2) > 2g(τ1, 3). For τ0, τ1 ∈ (0.5, 1) we say
that τ0 is sufficiently greater than τ1, denoted τ0 (cid:66) τ1, if 1 − τ0 is sufficiently less than 1 − τ1.
2The C++ code for these simulations is available at http://barmpalias.net/schelcode.shtml.
10
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
Figure 5. The function g of Definitions 1.4 and 1.5, for the values of k that we use. In
the second figure the horizontal axis is τα and the perpendicular axis is τβ. Then the
τβ (cid:66) τα, τβ > τα in [0.5, 1].
three regions show the relations τβ(cid:50) τα, τβ (cid:67) τα, τβ < τα in [0, 0.5] as well as τβ(cid:51) τα,
Note that the function g is decreasing in [0, 0.5] and increasing in [0.5, 1] (see the first item of Figure 5).
Hence (cid:67) in [0, 0.5] is a strengthening of < and (cid:66) in [0.5, 1] is a strengthening of >. The requirement
that τβ (cid:67) τα is not overly strong, as the second item of Figure 5 shows. Table 2 is a chart that shows for
various values of τα in [0, 0.5] the amount by which τβ needs to be smaller in order for τβ (cid:67) τα to hold
Definition 1.5 (Sufficient inequality for 3D). Again let g(x, k) = xkx(1 − x)k(1−x). For τ0, τ1 ∈ (0, 0.5)
(and similarly for(cid:50) of Definition 1.5).
we say that τ0 is sufficiently less than τ1, denoted τ0(cid:50) τ1, if g(τ0, 8) > 219g(τ1, 27). For τ0, τ1 ∈ (0.5, 1)
we say that τ0 is sufficiently greater than τ1, denoted τ0(cid:51) τ1, if 1 − τ0 is sufficiently less than 1 − τ1.
The relation(cid:50) is significantly stronger than (cid:67), as is made clear by Figure 5 and Table 2.
1.4. Further notation and terminology. The variables u and v are used to range over nodes, while
r, x, y, z, τ range over R. We let x and y range over either R2 or R3 depending on context, while
a, b, c, d, e, i, j, k, n, m, w range over N. Often we shall use x and y (and z when working in three dimen-
sions) to specify the coordinates of a node u = (x, y) -- so while x, y and z range over R, often they will
be elements of N.
In the context of discussing a given scenario (τα, τβ), we say that an event X occurs in the limit, if for
every > 0, X occurs with probability > 1 − for 0 (cid:28) w (cid:28) n, i.e. X holds with probability > 1 −
whenever w is sufficiently large and n is sufficiently large compared to w, and where it is to be understood
that how large one has to take w and how large n must be in comparison to w may depend upon , τα
and τβ. When working in two dimensions we let Cu,r be the circle of radius r centred at u, and we
u,r be the disc of radius r centred at u, i.e. the set of all x ∈ T with u − x ≤ r (where u − x
let C†
denotes the Euclidean distance between u and x). Then we extend this notation to A ⊆ T as follows.
If A ⊂ C†
u,r to be the intersection of all convex subsets of C†
containing A. If r < 0.25n and A ⊂ C†
u,r, then we shall call A local. If A is local then A† is the same
with respect to all C†
u,r in this
case, referring simply to A†. In the context of discussing A which is local (such as N (u) for a node u)
u,r then we define A† with respect to C†
u,r
u,r such that r < 0.25n and A ⊂ C†
u,r, and we shall suppress mention of C†
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
11
τα
τβ (cid:67) τα
0.390000
0.024443
0.400000
0.410000
0.420000
0.022266
0.020076
0.017874
τβ(cid:50) τα
0.090076
0.082213
0.074255
0.066211
τα
τβ (cid:67) τα
0.430000
0.015661
0.450000
0.470000
0.490000
0.011212
0.006737
0.002247
τβ(cid:50) τα
0.058092
0.041673
0.025074
0.008370
Table 2. This chart shows, for various values of τα in [0, 0.5], the minimum amount by
which τβ needs to be less than τα in order to have τβ (cid:67) τα or τβ(cid:50) τα respectively.
and two nodes u = (x, y) and v = (x(cid:48), y(cid:48)) in A, we say that x ≤ x(cid:48) if there exists r < 0.5n with x + r = x(cid:48)
n of the form [x, x + m) × [y, y + m)
(and similarly for y and y(cid:48)). By an m-square we mean a subset of N2
n of the form [x, x + a) × [y, y + b) for
for some x, y ∈ N. By an a × b rectangle we mean a subset of N2
some x, y, a, b ∈ N.
1.5. Overview of the proof. At the start of Section 2 we shall prove Theorem 1.2 (a), which along
with Theorem 1.2 (e) is the easiest case and amounts to little more than some easy observations. The
bulk of the work will then be in the remainder of Section 2, in which we prove Theorem 1.2 (b) and
Theorem 1.2 (c). With these in place, Theorem 1.2 (d) and Theorem 1.2 (e) will follow from certain
symmetry considerations, as described in Section 3. In Section 4 we describe how to modify the proofs
for the two dimensional model, to give all of the corresponding theorems for the three dimensional model.
Throughout, certain proofs which are of a technical nature, and which the reader might profitably skip
on a first reading, are deferred to Section 5. In the remainder of this subsection, we outline some of the
basic ideas behind the proof of Theorem 1.2.
The key to the proof lies in considering certain local structures, which might exist in the initial config-
uration or else might develop during the dynamic process. Our first central notion is that of a stable
structure:
Definition 1.6 (Stable structures). We say that a set of nodes A is an α-stable structure if, for every
u ∈ A, there are at least τα(2w+1)2 many α nodes in N (u)∩A. We define β-stable structures analogously.
The point of an α-stable structure is this (as may be seen by induction on stages): no node of type
α which belongs to an α-stable structure can ever change type. We shall also be interested in certain
conditions on N (u) and variants of this neighbourhood:
• (Partial neighbourhoods.) Suppose that (cid:96) is a straight line passing through N †(u). Let A1 be
all those points in N †(u) on or above (cid:96), and let A2 be all those points in N †(u) on or below (cid:96).
†
i ) = γ(2w + 1)2 (µ denotes Lebesgue measure), then we call Ai a γ-partial neighbourhood
If µ(A
of u, with defining line (cid:96). If u is an α node, we let pnα
γ,τ (u) be the event that there exists some
γ-partial neighbourhood of u which contains at least τ (2w + 1)2 many α nodes in the initial
configuration (and similarly for β). We also let pnα
γ,τ (u)[s] be the corresponding event for the
end of stage s, rather than the initial configuration.3
3One should think of pn as p-artial n-eighbourhood, uh as u-nh-appy, and ruh as r-ight-extended neighbourhood u-nh-
appy.
12
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
• (Extended neighbourhoods.) We say that uhα
τ (u) holds if there are strictly less than τ (2w +
1)2 many α nodes in N (u) in the initial configuration (and again we let uhα
τ (u)[s] denote the
corresponding event for the end of stage s). If u = (x, y) then the right extended neighbourhood
of u, denoted N (cid:5)(u), is the set of nodes (x(cid:48), y(cid:48)) such that x − w ≤ x(cid:48) ≤ x + 2w and y − y(cid:48) ≤ w.
We say that ruhα
τ (u) holds if there are strictly less than τ (2w + 1)(3w + 1) many α nodes in
N (cid:5)(u) in the initial configuration.
So a γ-partial neighbourhood of u is a subset of N †(u) (of a particularly simple form) which has measure
γ(2w + 1)2. If pnα
(u) holds, then it is not only the case that u is happy -- it has some γ-partial neigh-
bourhood which already contains enough α nodes for u to be happy. The right extended neighbourhood
N (cid:5)(u) is a (particularly simple) superset of N (u). While it is not strictly true that for τ < 0.5, ruhα
τ (u)
implies uhα
τ (u), it is not difficult to see that the former condition is less likely than the latter.
γ,τα
While the notion of a stable structure might seem essentially passive, we can strengthen the notion to
give a form of stable structure which will tend to grow over stages:
Definition 1.7 (Firewalls). If all nodes in C†
called a β-firewall of radius rw centred at u (at that stage). We say v is on the outer boundary of C†
if v /∈ C†
and v(cid:48) ∈ C†
u,rw are of type β (at some stage) then this set of nodes is
u,rw but has an immediate neighbour belonging to this set, i.e. there exists v(cid:48) with v − v(cid:48)∞ = 1
u,rw
u,rw.
u,r∗w satisfies the condition that N (v) ∩ C†
Now suppose that τα, τβ < 0.5, that γ > 0.5 is fixed, and that we choose some r∗ ∈ N+ sufficiently large
that, for 0 (cid:28) w (cid:28) n:
(†a) Any node v ∈ C†
(†b) Any node v on the outer boundary of any C†
γ-partial neighbourhood of v.
We may now observe that if r ≥ r∗ then for 0 (cid:28) w (cid:28) n: a) any β-firewall of radius rw is a β-stable
structure, and b) any α node v on the outer boundary of a β-firewall of radius rw at stage s will be
(v)[s] does not hold. For b), our choice of r∗ implies that v will be unhappy,
hopeful, so long as pnα
and then since τα, τβ < 0.5 unhappiness automatically implies being hopeful. The point is this:
u,r∗w has all of N (v) − C†
u,r∗w ≥ τβ(2w + 1)2.
u,r∗w contained in some
γ,τα
A β-firewall of sufficient radius will grow over stages, so long as no node v on the outer
boundary satisfies pnα
(v)[s].
γ,τα
With these ideas in place, we can now sketch our approach to the proof of Theorem 1.2. First of all
suppose τα, τβ < 0.25 and consider Theorem 1.2 (a). The basic idea is to show that for 0 (cid:28) w (cid:28) n,
a node u0 chosen uniformly at random very probably belongs to a structure which is both α-stable
and β-stable in the initial configuration. From our previous observations, it then follows that no nodes
within this structure can ever change type. In fact a very simple choice of structure suffices -- if we fix
r which is sufficiently large then we can show that, for 0 (cid:28) w (cid:28) n, C†
u0,rw will very probably be such
a stable structure. This follows because all nodes within the disc have (at least) close to half of their
neighbourhood within the disc, and since (via an application of the weak law of large numbers) we can
expect nodes within the disc to be quite evenly distributed.
Now suppose that τα > 0.25, τβ < τα < 0.5 and consider Theorem 1.2 (b). In this case it still holds that
unhappy nodes can be expected to be rare in the initial configuration, and since τβ < τα we have that
unhappy β nodes are less likely than unhappy α nodes. Taking w large we have that unhappy β nodes
are much less likely than unhappy α nodes. A naive approach might then proceed as follows. Choosing
u0 uniformly at random, for 0 (cid:28) w (cid:28) n we should be able to choose a large region around u0, Q say,
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
13
which can be expected to contain unhappy α nodes but no unhappy β nodes in the initial configuration.
What then can be expected to occur in the vicinity of an unhappy α node u in the early stages? Well if u
changes type then this may cause other α nodes in N (u) to become unhappy. If these then change type
then this may cause further α nodes to become unhappy, and so on. In this manner a cascade of changes
to type β may emanate out from the (rare) initially unhappy α nodes, bringing about the formation of
large β-firewalls. We might then look to establish that u0 very probably belongs to such a β-firewall in
the final configuration (and so must ultimately be of type β). There are two immediate difficulties with
this initial plan, however.
(1) First of all, it is not actually clear whether or not unhappy α nodes are likely to give rise to
the formation of large β-firewalls. It is for this reason that we consider the events ruhα
(u) and
τα
the condition τβ (cid:67) τα. The fact that τβ (cid:67) τα is precisely what we need in order to be able to
choose Q so that in the limit there will be α nodes u in Q for which ruhα
(u) holds in the
initial configuraiton -- a less probable condition than uhα
(u) -- while maintaining the absence of
unhappy β nodes in Q. When ruhα
τα
(u) holds we are able to show that a β-firewall of sufficient
radius around u very probably results.
τα
τα
(2) The second problem is that we aren't guaranteed that such a β-firewall will spread until α-
firewalls interfere with the process -- all we observed above was that if r ≥ r∗ (with r∗ chosen
appropriately) then for 0 (cid:28) w (cid:28) n, any α node v on the outer boundary of a β-firewall of radius
rw at stage s will be hopeful, so long as pnα
(v)[s] does not hold. This is where κ comes into
play. From the fact that τα > κ, we are able to choose an appropriate γ > 0.5 and show that
one can choose Q so as to ensure the absence of nodes in Q for which pnα
(v) holds (in the
initial configuration). We can then establish that our large β-firewall, formed within Q in the
early stages of the process, will spread until u0 is contained within it.
γ,τα
γ,τα
Proving Theorem 1.2 parts (c), (d) and (e) then only involves simple modifications of the proofs for parts
(a) and (b).
2. The proof of clauses (a), (b) and (c) of Theorem 1.2
The proof of Theorem 1.2 (a). As mentioned previously, proving Theorem 1.2 (a) requires only some
simple observations. The basic idea is to show that, under the hypothesis of the theorem, a node chosen
uniformly at random will very probably belong to both α-stable and β-stable structures in the initial
configuration, and can therefore never change type.
Throughout this subsection we assume that τα, τβ < 1
4 are fixed and we work for varying w and n such
4 > τ1 > τ2 > τ3 > max{τα, τβ}. We consider a fixed node
that 0 (cid:28) w (cid:28) n. Choose τ1, τ2, τ3 such that 1
u0, which is chosen uniformly at random. We first observe that all nodes in a large disc centred at u0
(with radius some multiple of w), will have (at least) close to half of their neighbourhood in the disc.
Then the aim is to show that for large w we can expect the nodes of each type to be distributed very
evenly in the disc, meaning that nodes in the disc can be expected to have (at least) close to 0.25(2w +1)2
many α nodes within the intersection of the disc and their neighbourhood (and similarly for β). In the
limit, then, this disc will be both an α-stable and a β-stable structure.
Lemma 2.1 (Neighbourhood/disc intersections). Suppose τ(cid:48) < 0.5. There exists r such that, for 0 (cid:28)
w (cid:28) n, all nodes u in C†
u0,rw satisfy N (u) ∩ C†
u0,rw > τ(cid:48)(2w + 1)2.
u0,rw, let A be the set of all nodes in N (u) ∩ C†
Proof. Given u ∈ C†
lattice polygon. By Pick's theorem µ(A†) = i + b
u0,rw. Then the boundary of A† is a
2 − 1, where µ denotes Lebesgue measure, i is the number
14
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
of nodes in the interior of A† and b is the number of nodes on the boundary. For sufficiently large r and
for 0 (cid:28) w (cid:28) n, µ(A†) > τ(cid:48)(2w + 1)2, so the result follows directly from Pick's theorem.
(cid:3)
For the remainder of this paragraph, fix r as guaranteed by Lemma 2.1 when τ(cid:48) = 2τ1. Let u0 = (x, y)
and for u ∈ C†
u0,rw let N1(u) = N (u) ∩ C†
u0,rw. Now, to establish the even distribution of nodes of each
type within C†
u0,rw, we divide this region up into a fixed number (i.e. independent of w and n) of small
neighbourhoods. So, for some k ∈ N+, let w(cid:48) = (cid:100)w/k(cid:101), and consider small neighbourhoods of the form
[x + aw(cid:48), x + (a + 1)w(cid:48))× [y + bw(cid:48), y + (b + 1)w(cid:48)), where a, b ∈ Z and n (cid:29) aw(cid:48), n (cid:29) bw(cid:48). Let Π be the set of
u0,rw, let N2(u) be the set of
these small neighbourhoods which lie entirely within C†
nodes in N1(u) which belong to small neighbourhoods in Π entirely contained in N †(u). For sufficiently
large k, and for 0 (cid:28) w (cid:28) n, we have that N2(u) > 2τ2(2w + 1)2 for all u ∈ C†
u0,rw. So fix k satisfying
this condition. Finally, choose > 0 such that (1 − )τ2 > τ3. Then, since k is fixed, it follows by the
weak law of large numbers that in the limit, the proportion of the nodes in each small neighbourhood in
Π which are of type α is greater than 0.5(1− ) (and similarly for β). It then follows that, in the limit, all
u0,rw satisfy the condition that the number of α nodes in N2(u) is greater than τ3(2w + 1)2,
nodes u ∈ C†
and so is greater than τα(2w + 1)2 (and similarly for β). Thus C†
u0,rw will be both an α-stable and a
β-stable structure, as required.
u0,rw. Given u ∈ C†
γ,τα
The proof of clause (b) of Theorem 1.2. We suppose we are given a node u0 = (x0, y0), chosen uniformly
at random. Recalling the discussion of Section 1.5, our first aim is to establish a region Q containing u0,
of size which means that in the initial configuration (for 0 (cid:28) w (cid:28) n) we shall very likely find u ∈ Q
such that ruhα
(u) holds, while at the same time it is very unlikely that we will find any u such that
τα
(u) or uhβ
pnα
(u) hold. The following lemma is what we need in order to do this, and provides the
τβ
motivation behind our definitions of κ and τα (cid:67) τβ. The proof appears in Section 5, and simply consists
of applying standard bounds for the tail of the binomial distribution together with multiple applications
of Stirling's approximation.
Lemma 2.2 (Properties of (cid:67) and κ). Suppose that κ < τα < 0.5 and τβ (cid:67) τα. Choose τ such that
2 and ζ > 1 such that for 0 (cid:28) w (cid:28) n, and for a node u
τβ (cid:67) τ < τα and κ < τ . Then there exists γ > 1
selected uniformly at random, P(ruhα
τ (u)) > ζ w · P(pnα
(u)) and P(ruhα
γ,τα
τ (u)) > ζ w · P(uhβ
τβ
(u)).4
For the remainder of this section we assume that τα, τβ satisfying the conditions of Lemma 2.2 are fixed.
Once and for all we choose (any) τ such that τβ (cid:67) τ < τα and κ < τ , and then we let γ and ζ be as
guaranteed by Lemma 2.2. Given this choice of γ, we then let r∗ be sufficiently large that conditions
(†a) and (†b) of Section 1.5 are satisfied. We also assume we are given 0 > 0. Our aim is to show that
for 0 (cid:28) w (cid:28) n, the probability u0 will be of type β in the final configuration is > 1 − 0. The basic
approach is to establish that in the limit u0 will be consumed by a β-firewall which originates with Q.
Defining Q. To specify Q it will be useful to consider a grid of nodes with disjoint neighbourhoods,
meaning that these neighbourhoods are independently distributed in the initial configuration:
Definition 2.3 (Grid). Recall that u0 = (x0, y0). We say that v = (x, y) is on the grid if there exist
a, b ∈ Z with a,b < 0.25n/(2w + 1) and x = x0 + a(2w + 1), y = y0 + b(2w + 1). Suppose m = 2k + 1.
Then by the m-square on the grid centred at u0 we mean the set of nodes {(x0 +a(2w +1), y0 +b(2w +1)) :
a, b ∈ [−k, +k]}.
4Here it is to be understood that how large one must take w, and how large n must be compared to w, may depend on
our particular choices of τ and γ (as well as the given values τα and τβ ).
(cid:19)π(w)k0
(cid:18)
1 − 1
π(w)
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
τ (u)) = 1/π(w). Note that 1/π(w) → 0 as w → ∞, and recall that (cid:0)1 − 1
Now consider the initial configuration and let π(w) ∈ R be such that for u chosen uniformly at random,
e as x → ∞.
P(ruhα
Take k0 such that e−k0 (cid:28) 0 (with 0 as fixed previously), noting that k0 does not depend on w. Take
the least odd number m0(w) ∈ N+ such that m0(w)2 ≥ π(w)k0. Let Q0 be the m0(w)-square on the grid
centred at u0. Then the probability that not a single node u ∈ Q0 satisfies ruhα
(cid:1)x → 1
τ (u) is at most:
15
x
which is (cid:28) 0 for sufficiently large w. So far then, we have identified a set of nodes Q0, which has the
property that the following will fail to be true with probability (cid:28) 0: there exists u ∈ Q0 such that
ruhα
τ (u) holds.
We need a little more from the vicinity of u0 that we are going to work with. In order to ensure that
firewalls created inside Q†
0 will grow to include u0, we need a much larger region in which we will very
(u) hold. We let Q1 be the (8w+1)m0(w)-
likely not have any nodes u for which either of pnα
square on the grid centred at u0. Then we define Q = Q†
1. Let p be the probability that any node u ∈ Q
(u). Since the number of nodes in Q is ≤ π(w)k0(8w + 1)4, it follows that p is at most
satisfies pnα
π(w)k0(8w + 1)4 times the probability that pnα
(u) holds for u selected uniformly at random. So by
Lemma 2.2, p (cid:28) 0 for sufficiently large w. A similar argument holds for uhβ
(u). Thus we conclude
that for 0 (cid:28) w (cid:28) n, Typical vicinity fails to hold with probability (cid:28) 0:
Definition 2.4 (Typical vicinity). We say that Typical vicinity holds if all of (1) -- (3) below are true:
(u) or uhβ
τβ
γ,τα
γ,τα
γ,τα
τβ
τ (u0) does not hold;
(1) ruhα
(2) There exists u ∈ Q0 such that ruhα
(3) There does not exist any u in Q such that either of pnα
τ (u) holds;
γ,τα
(u) or uhβ
τβ
(u) hold.
Note that (1) -- (3) from Definition 2.4 all refer to the initial configuration.
Definition 2.5. Consider the initial configuration. For τ(cid:48) ∈ [0, 1] we say that juα
τ(cid:48)(u) holds if there are
less than τ(cid:48)(2w + 1)2 many α nodes in N (u), but changing the type of 2w+1 β nodes in N (u) would
τ(cid:48)(u) holds if there are less than τ(cid:48)(2w + 1)(3w + 1) many
cause this not to be the case. We say that rjuα
α nodes in N (cid:5)(u), but changing the type of 3w+1 β nodes in N (cid:5)(u) would cause this not to be the case.5
Observe that if Typical vicinity holds, then we are guaranteed the existence of u ∈ Q†
rjuα
Smoothness conditions. Our aim will be to show that for 0 (cid:28) w (cid:28) n, a large β-firewall can be expected
to form in the early stages of the process in the vicinity of u ∈ Q†
τ (u) holds. To this end,
however, we must first examine what can be expected in the initial configuration, from the vicinity of u
which is chosen uniformly at random from amongst the nodes such that rjuα
τ (u) holds. In particular we
†
are interested in the three regions: C 0(u) := C†
u,3r∗w. In this
subsection, we look to establish certain smoothness conditions -- that the types of nodes in these regions
will be nicely distributed.
†
u,2r∗w and C 2(u) := C
u,r∗w, C 1(u) := C
0 for which rjuα
0 for which
τ (u) holds.
For some large k1, which we shall specify later and which will not depend on w, the basic idea now is that
we want to cover the nodes in C 2(u) with disjoint w
-squares. This occasions the minor inconvenience
k1
5Think of ju as j-ust u-nhappy and think of rju as r-ight extended neighbourhood j-ust u-nhappy.
16
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
that k1 may not divide w. We therefore let Ik1 (u) be a pairwise disjoint set of rectangles, whose union
contains all nodes in C 2(u), and such that:
• Each element of Ik1(u) has nonempty intersection with C 2(u) and is of the form [x, x+a)×[y, y+b)
for some a, b ∈ {(cid:98)w/k1(cid:99),(cid:100)w/k1(cid:101),(cid:100)w/k1(cid:101) + 1} and x, y ∈ N;
• Each element of Ik1(u) is either entirely contained in N (cid:5)(u), or else is disjoint from N (cid:5)(u).
Definition 2.6 (Smoothness for u). Suppose given k1 ∈ N+ and 1 > 0 and u such that rjuα
We say that Smoothk1,1(u) holds if:
τ (u) holds.
• For each A ∈ Ik1(u) which is contained in N (cid:5)(u), the proportion of the nodes in A which are of
type α is in the interval [τ − 1, τ + 1].
• For each A ∈ Ik1 (u) which is disjoint from N (cid:5)(u), the proportion of the nodes in A which are
of type α is in the interval [0.5 − 1, 0.5 + 1].
0 such that rjuα
The following lemma is then almost immediate and is proved in Section 5:
Lemma 2.7 (Likely smoothness). Suppose given k1 ∈ N + and 1, > 0. Suppose that u is selected
τ (u) holds. For 0 (cid:28) w (cid:28) n, the probability
uniformly at random from amongst the nodes such that rjuα
that Smoothk1,1 (u) holds is greater than 1 − .
Definition 2.8 (The smoothness event). We let u1 be chosen uniformly at random from amongst the
nodes u ∈ Q†
τ (u) holds (so that if there exists no such node then u1 is undefined). For
any k1, 1 > 0, we let Smooth(k1, 1) be the event that u1 is defined and Smoothk1,1(u1) holds.
Our arguments so far suffice to show that, for any k1, 1 > 0, if 0 (cid:28) w (cid:28) n then the probability that
Smooth(k1, 1) fails to hold is (cid:28) 0. We now want to examine what satisfaction of Smooth(k1, 1) can tell
us about the proportion of α nodes in regions contained in C 2(u1) but which are not one of the rectangles
in Ik1(u1). In order to do so we define a couple of functions, which describe the proportion of α nodes
in a given set A, either in the initial configuration or else after all α nodes in B ⊆ A have changed type.
We also consider idealised versions of these functions which will be easier to work with most of the time.
Here and elsewhere ¯B denotes the complement of B.
Definition 2.9. Consider sets of nodes A, B ⊂ C 2(u).
(1) Let Ξ(A) be the proportion of the elements of A which are of type α in the initial configuration.
(2) Let A0 = A ∩ N (cid:5)(u) and A1 = A − A0. Define Ξ∗(A, u) = (τA0 + 0.5A1)/A.
(3) Take the initial configuration and then change all nodes in B to type β. Let Ξ(A, B) denote the
proportion of the elements of A which are now of type α.
(4) Let A0 = A ∩ B, A1 = A ∩ ¯B ∩ N (cid:5)(u) and let A2 = A − (A0 ∪ A1). We define Ξ∗(A, u, B) =
(τA1 + 0.5A2)/A.
So Ξ∗(A, u) gives the proportion of nodes in A which would be of type α, if exactly proportion τ of those
nodes in A ∩ N (cid:5)(u) were of type α, and exactly half of the nodes in the remainder of A were of type α.
On the other hand Ξ∗(A, u, B) gives the corresponding proportion if the same conditions hold, but then
we change all nodes in B to type β. One may think of Ξ∗ as an idealised version of Ξ. Satisfaction of
τ (u) and Smoothk1,1(u) for large k1 and small 1 will ensure that Ξ∗ is a reasonable approximation
rjuα
to Ξ, in a sense that we will make precise.
Now we want to work with a greater variety of sets of nodes than just those in Ik1(u), but we still only
need to consider sets which are reasonably large and of a reasonably simple form:
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
17
k2
Definition 2.10. Given a node u and k2 ∈ N +, we let I∗
(u) be the set of all sets of nodes A such that
A ⊆ C 2(u) and either (1) A is an a × b rectangle for a, b ≥ w/k2, or (2) the set of nodes inside a regular
polygon with sides of length ≥ w/k2, or (3) the union of two sets of the form (1) or (2).
So long as we restrict attention to sets of nodes in I∗
observation allows us to work with the idealised functions Ξ∗:
Obervation 2.11 (Smoothness for elements of I∗
sufficiently large and 1 > 0 is sufficiently small then, for 0 (cid:28) w (cid:28) n, satisfaction of rjuα
Smoothk1,1 (u) suffices to ensure that:
(u)). Suppose given k2 ∈ N+ and δ > 0. If k1 is
τ (u) and
(u) (for some k2 to be specified), the following
k2
k2
(ı) For all A, B ∈ I∗
(u), Ξ(A) − Ξ∗(A, u) < δ and Ξ(A, B) − Ξ∗(A, u, B) < δ.
k2
Establishing the creation of firewalls. Now we choose values of k2 and δ which will allow us to argue that
a large firewall is very probably created around u1 in the early stages of the process. With these values
specified, k1 and 1 are simply chosen to be those values guaranteed by Lemma 2.11, meaning that we
can make use of our idealised functions Ξ∗. Numerical values in the following definition are somewhat
arbitrary, but suffice for our purposes.
Definition 2.12 (Choosing k2, δ, k1 and 1). Choose k2 > (0.5 − τ )/(τα − τ ), choose δ > 0 such that
δ (cid:28) min{(τα − τ )(2τ − 0.5)/(2k2), 10−5} and choose k1 sufficiently large and 1 > 0 sufficiently small
that for 0 (cid:28) w (cid:28) n satisfaction of rjuα
τ (u) and Smoothk1,1(u) suffices to ensure satisfaction of (ı) (as
specified in Observation 2.11).
The next lemma finally establishes that a firewall of radius r∗w very probably forms around u1.
τ (u) and Smoothk1,1 (u) both hold. Let t∗ = 2r∗w + 3w and
Lemma 2.13 (Firewalls). Suppose that rjuα
suppose that there are no hopeful β nodes in C 2(u) at any stage ≤ t∗. Then all nodes in C 0(u) are of
type β at stage t∗.
Proof. In what follows it will be convenient to assume that w is even. Only small modifications are
required to deal with the case that w is odd. It is also convenient to assume that r∗ > 2. The proof
will basically consist of repeated applications of condition (ı). We shall apply this condition in order to
inductively establish a sequence of increasingly large rectangles for which all nodes become of type β. Let
2 and u∗ := (x∗, y). The following argument is illustrated in the first picture of
u = (x, y), let x∗ := x + w
Figure 6. The rectangles we consider are as follows (where a ∈ N+):
• We let R0
• We let R1
• We let R2
a be the set of nodes v = (x(cid:48), y(cid:48)) such that 0 ≤ x(cid:48) − x ≤ w and y − y(cid:48) ≤ a;
a be the set of nodes v = (x(cid:48), y(cid:48)) such that x − a ≤ x(cid:48) ≤ x + w + a and y − y(cid:48) ≤ w;
a be the set of nodes v = (x(cid:48), y(cid:48)) such that x(cid:48) − x∗ ≤ w
2 w(cid:101) and y(cid:48) − y ≤ w + a.
2−1
Working always for sufficiently large w, we show first that all nodes in R0(cid:98)w/k2(cid:99) are of type β by the end
of stage 1. Then we show inductively that:
2 +(cid:100)√
a are of type β by the end of stage a, for a ≤ w/2;
a when a ≤ w;
(†1) All nodes in R0
(†2) The same result holds for R0
(†3) All nodes in R1
(†4) All nodes in R2
a are of type β by the end of stage w + a, for a ≤ (cid:100) w√
(cid:101);
a are of type β by the end of stage 2w + a, for all a ≤ (cid:100)√
2
2 w(cid:101).
2−1
18
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
R2
c
b
c
R0
α
u∗
u
a
R1
b
u∗
Or(u∗)
Figure 6. Rectangles in the proof of Lemma 2.13 and the octagon of Definition 2.15
N (cid:5)(u)
Sr(u∗)
From there we shall be able to argue that all nodes in C 0(u) will eventually be of type β.
So our first task is to establish that all nodes in R0(cid:98)w/k2(cid:99) will be of type β by the end of stage 1. We have
that R0(cid:98)w/k2(cid:99) ∈ I∗
(u) and by our choice of k2 the following holds for all nodes v in this rectangle:
k2
Ξ∗(N (v)) ≤ τ + τα
.
2
Since δ < (τα − τ )/2, it follows from (ı) that for 0 (cid:28) w (cid:28) n, Ξ(N (v)) < τα. So for 0 (cid:28) w (cid:28) n, all α
nodes in R0(cid:98)w/k2(cid:99) are hopeful in the initial configuration and will be of type β by the end of stage 1, as
required.
For any a ∈ N, let da = a/(2w + 1). In order to inductively establish (†1) -- (†4) we then need the various
technical facts contained in the following proposition, which is easily verified by direct calculation.
Proposition 2.14. Given > 0, the following all hold for 0 (cid:28) w (cid:28) n:
(1) For all a ∈ ((cid:98)w/k2(cid:99), w/2] and all v ∈ R0
(2) For all a ∈ (w/2, w] and v ∈ R0
(3) For all a ∈ (0,(cid:100) w√
(cid:101)] and v ∈ R1
√
(4) For all a ∈ (0,(cid:100)w(
2− 1)/2(cid:101)] and v ∈ R2
a − R0
a − R1
2
√
1/
2) + .
a − R0
a−1, Ξ∗(N (v), u, R0
a−1) − (τ + (0.5 − 2τ )da) < .
a−1, Ξ∗(N (v), R0
a−1, Ξ∗(N (v), R1
a−1) < 0.25 + 0.25τ + .
√
a−1) < 3/8 − (1/2 − τ )(1 − 1/
a − R2
a−1, Ξ∗(N (v), R2
a−1 ∪ R1(cid:100) w√
2
2)/4 + .
(cid:101)) < 3/8− (1/16)(3/2−
Given (1) of Proposition 2.14, the induction to establish (†1) now goes through easily. Applying (ı) we
a−1, Ξ(N (v), Ra−1) ≤ τ + (0.5 −
have that if 0 (cid:28) w (cid:28) n then for a ∈ ((cid:98)w/k2(cid:99), w/2] and v ∈ R0
2τ )da + 2δ. Since da > 1/3k2 and δ < (2τ − 0.5)/(6k2) the induction step follows.
Similarly (2) of Proposition 2.14 gives us the induction step in establishing (†2). Applying (ı) we have
that if 0 (cid:28) w (cid:28) n then for a ∈ (w/2, w] and v ∈ R0
a−1) < 0.25 + 0.25τ + 2δ. Now
for τ ∈ (κ, 0.5), 0.25 + 0.25τ < τ − 0.02 so, since δ < 10−5, the induction step follows.
For (†3) we have that if 0 (cid:28) w (cid:28) n then for a ∈ (0,(cid:100) w√
(1/2−τ )(1−1/
so the induction step follows.
2)/4+2δ. Now for τ ∈ (κ, 0.5) we have τ −(3/8−(1/2−τ )(1−1/
a−1) < 3/8 −
2)/4) > 2·10−5 > 2δ,
a−1, Ξ(N (v), R0
a−1, Ξ(N (v), R0
(cid:101)] and v ∈ R1
a − R0
a − R0
a − R1
√
√
2
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
√
Finally, for (†4) we have that if 0 (cid:28) w (cid:28) n then for a ∈ (0,(cid:100)w(
Ξ(N (v), R2
√
(cid:101)) < 3/8 − 1/16(3/2 − 1/
a−1 ∪ R1(cid:100) w√
2) + 2δ. Then for τ ∈ (κ, 0.5) we have
2 − 1)/2(cid:101)] and v ∈ R2
19
a − R2
a−1,
2
√
τ − (3/8 − (1/16)(3/2 − 1/
2)) > 0.02 > 2δ.
Once again the induction step goes through.
2(cid:101) ∪ R2(cid:100)w(
√
√
So far we have established that all nodes in R := R1(cid:100)w/
2−1)/2(cid:101) will be of type β by the end
of stage 3w. Now we wish to extend this and argue that all nodes in C 0(u) will eventually be of type
β. Previously we observed that, for 0 (cid:28) w (cid:28) n, any α node v on the outer boundary of C 0(u) will be
hopeful, so long as pnα
(v) does not hold. In order to get to the point where we can conclude that all
nodes in C 0(u) will become of type β, we use a similar idea, but we must work with smaller regions than
C 0(u) -- this will not be a problem, since satisfaction of Smoothk1,1(u) guarantees the failure of much
(v) for v ∈ C 1(u). In fact the calculations will be simpler if we work with
weaker conditions than pnα
regular octagons rather than circles:
Definition 2.15 (Octagon). Given r ∈ R, we let Sr(u∗) be the square in R2 centred at u∗ with sides of
r (u∗), as
length r parallel to the axes. Then we let Or(u∗) be the largest regular octagon contained in S†
illustrated in the second image of Figure 6.
γ,τα
γ,τα
2), Or0 (u∗) has sides of length w and that O†
(u∗) is entirely contained in
Note that for r0 = w(1 +
R† (where R is as specified above). Now if r ≥ r0 then for 0 (cid:28) w (cid:28) n, if v ∈ C 1(u) is an α node on the
outer boundary of Or(u∗) then Ξ(N (v)) < 11/32 + 2δ < τ , and so v is hopeful. Inductively we conclude
†
that all nodes in C 1(u) ∩ O
r0+2a(u∗) are of type β by the end of stage 3w + a. Now there exists a with
C 0(u) ⊂ O
(cid:3)
†
r0+2a(u∗) ⊂ C 1(u). Since a < 2r∗w, the result follows as required.
r0
√
The following lemma completes our proof of clause (b) of Theorem 1.2.
Lemma 2.16 (Eventual conversion). Suppose that Typical vicinity and Smooth(k1, 1) both hold.
Then there exists a stage s after which u0 is always of type β.
Proof. Recall that we defined Q0 to be the m0(w)-square on the grid centred at u0, and that we defined
Q1 to be the (8w + 1)m0(w)-square on the grid centred at u0. Now we define Q2 to be the 3m0(w)-square
on the grid centred at u0. First of all we observe that we are guaranteed a large number of stages before
the existence of any unhappy β nodes in Q2. This follows because a β node which is happy in the initial
configuration, cannot become unhappy until strictly after the first stage (if such a stage exists) at which
another β node in its neighbourhood becomes unhappy. For 0 (cid:28) w (cid:28) n we have that in the initial
configuration, if m = min{u − v∞ : v ∈ Q2, u is an unhappy β node}, then m > 3wm0(w)(2w + 1).
Thus we are guaranteed not to find any unhappy β nodes in Q2 at stages prior to t∗ = 3m0(w)(2w + 1).
Applying Lemma 2.13 we conclude that for 0 (cid:28) w (cid:28) n we shall have that by stage 2r∗w + 3w all nodes
in C 0(u1) will be of type β. For 0 (cid:28) w (cid:28) n it then follows inductively, by the choice of r∗, that at each
stage < 2m0(w)(2w + 1) and after the creation of this firewall, its radius will expand by at least 1. Since
u0 − u1 < m0(w)(2w + 1) we conclude that there exists some stage after which u0 always belongs to a
(cid:3)
β-firewall of radius > r∗w centred at u1.
The proof of clause (c) of Theorem 1.2. This case is much simpler. Recall that in the proof of clause (b)
of Theorem 1.2 we chose r∗ so that certain conditions were satisfied. Now that we have τβ < 0.5 < τα,
however, we can simply choose r∗ large enough such that for 0 (cid:28) w (cid:28) n if r ≥ r∗ then for any β node
v inside a β firewall of radius rw, close enough to a half of N (v) will lie inside the firewall to ensure
20
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
that v is happy, and similarly any α node v(cid:48) on the outer boundary of such a firewall will be hopeful.
Such a firewall must then spread until every node is eventually contained inside it. For any > 0, if n is
sufficiently large then there will exist such a firewall in the initial configuration with probability > 1 − .
Whenever there exists such a firewall all nodes must eventually be of type β.
3. The proofs of Theorems 1.2 (d) and 1.2 (e)
In this section we work with τα, τβ > 0.5. The proofs are simple modifications of the proofs of Theorems
1.2(b) and 1.2(a). Rather than describing those proofs again in their entirety, with only small changes, we
describe the necessary modifications. Roughly speaking, the idea is that we now replace considerations as
to whether a β node u has less than proportion τβ many β nodes in its neighbourhood, with the question
as to whether it has proportion ≤ 1 − τ α many β nodes. If this holds then u is unhappy as type β, but
would be happy if it changed type (in fact we should take into account the effect of the type change of u
on the proportion, meaning that 1 − τα is not exactly the right proportion to consider but approaches it
for large w). The nodes of this type, then, will be the hopeful nodes -- and these are the nodes which may
initiate firewalls. Similarly, we replace the notion of an α-stable structure, with that of an α-intractable
structure:
Definition 3.1 (Intractable structures). We say that a set of nodes A is an α-intractable structure if it
contains β nodes, and for every β node u ∈ A, u would be unhappy (as an α node) if all nodes in ¯A∪{u}
were changed to type α.
So the point is that if a β node belongs to an α-intractable structure in the initial configuration then it
can never change type.
The proof of Theorem 1.2 (e). The proof goes through word for word the same as that of Theorem 1.2
(a), if one replaces τα, τβ everywhere with 1 − τα and 1 − τβ and if one replaces 'α-stable' or 'β-stable'
with 'β-intractable' or 'α-intractable'.
The proof of Theorem 1.2 (d). We define various events, which take the place of pnα
ruhα
τ (u) in the proof of Theorem 1.2 (b).
τ (u) and rjuα
(u), uhβ
τβ
(u),
γ,τα
• Consider the initial configuration.
If u is a β node, we let pnα
(u) be the event that there
exists some γ-partial neighbourhood of u, A say, such that u would be unhappy (as a node of
type α) if all nodes in ¯A ∪ {u} were changed to type α. Note that, roughly, this corresponds to
A containing at least (1 − τα)(2w + 1)2 many nodes of type β. We also let pnα
(u)[s] be the
corresponding event for the end of stage s, rather than the initial configuration.
γ,τα
γ,τα
• We say that hβ
τβ
(u) holds if u would be happy in the initial configuration if its type was changed
to β (or remains β if already of this type). Note that roughly this corresponds to N (u) containing
at most (1 − τβ) many nodes of type α. We say that ruhα
τ (u) holds if N (cid:5)(u) contains strictly
less than (1 − τ )(2w + 1)(3w + 1) many β nodes. We say that rjuα
τ (u) holds,
but this would no longer be true if any node of type α in N (cid:5)(u) changed type.
(u) holds, if ruhα
τ
In what follows we use the notation f (w) (cid:39) g(w) introduced in the proof of Lemma 2.2. The crucial
observation is that P(pnα
τ (u)) =
P(ruhα
1−τ (u)). We therefore have the following analogue of Lemma 2.2:
(u)) (cid:39) P(pnα
(u)) (cid:39) P(uhβ
(u)), while P(ruhα
(u)), and P(hβ
τβ
γ,1−τα
1−τβ
γ,τα
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
21
Lemma 3.2. Suppose that 1− κ > τα > 0.5 and τβ (cid:66) τα. Choose τ such that τβ (cid:66) τ > τα and 1− κ > τ .
2 and ζ > 1 such that for 0 (cid:28) w (cid:28) n, and for a node u selected uniformly at
Then there exists γ > 1
random, P(ruhα
τ (u)) > ζ w · P(pnα
τ (u)) > ζ w · P(hβ
(u)) and P(ruhα
(u)).
τβ
γ,τα
1−τβ
γ,1−τα
(u), uhβ
(u), ruhα
With Lemma 3.2 in place, the remainder of the proof then goes through almost identically to that for
Theorem 1.2 (b), with (for the new values τα, τβ satisfying the conditions in the statement of Theorem
1.2 (d), and for τ chosen as in Lemma 3.2), pnα
(u) taking the place of
pnα
which is chosen uniformly at random, will very probably eventually be of type α.
So we choose r∗ satisfying (†a) and (†b) of Section 1.5 and observe, in a manner entirely analogous to
what took place before, that if r ≥ r∗ then for 0 (cid:28) w (cid:28) n, any β node v on the outer boundary of a
α-firewall of radius rw at stage s will be hopeful, so long as pnα
(v)[s] does not hold. Then we define
Q0 and Q as before, but using π(w) which is now defined in terms of P(ruhα
τ (u)). Our new version of
Typical vicinity holds iff:
1−τ (u) respectively. Now, however, we look to show that u0
τ (u) and rjuα
τ
1−τ (u) and rjuα
(u), ruhα
(u), hβ
τβ
γ,τα
γ,τα
τ (u0) does not hold;
1. ruhα
2. There exists u ∈ Q0 such that ruhα
3. There does not exist any u in Q such that either of pnα
τ (u) holds;
(u) or hβ
τβ
(u) hold.
γ,τα
As before, we may conclude that if 0 (cid:28) w (cid:28) n then (our new version of) Typical vicinity fails to
hold with probability (cid:28) 0. In our new version of Definition 2.6, we replace "type α" with "type β".
In Definition 2.9 we again replace "type α" with "type β" and vice versa, and we replace τ with 1 − τ ,
so that these functions now detail the proportion of nodes of type β (rather than α as previously was
the case). In Definition 2.12 and Lemma 2.13 we replace τ and τα with 1 − τ and 1 − τα respectively
and we exchange "type α" everywhere for "type β", and vice versa. Then the proof of our new version
of Lemma 2.13 goes through as before, except that now we are inductively able to argue that all nodes
in the relevant sets Ri
j are eventually of type α, and that C 0(u) will eventually be a firewall of type α.
Finally, using the same argument as in the proof of Lemma 2.16 but using our new values of Q0, m0 and
so on, and replacing β with α, we conclude that if Typical vicinity and Smooth(k1, 1) both hold then
there exists a stage after which u0 is always of type α.
4. Modifying the proofs for the three dimensional model
First note that Theorems 1.3 (a) and 1.3 (c) can be proved exactly as in the two dimensional case, if we
simply modify all of the two dimensional notions to their three dimensional counterparts in the obvious
way. Circles are replaced by spheres and rectangles by three dimensional blocks. We also extend the
notation A† to three dimensional space in the obvious way. Then Theorems 1.3 (d) and 1.3 (e) will once
again follow by the same symmetry considerations that we applied previously in Section 3, once Theorem
1.3 (b) is established. Once again then, the bulk of the work is in establishing Theorem 1.3 (b).
To establish Theorem 1.3 (b), we can apply almost exactly the same proof as in the two dimensional
case. In that previous case, however, we worked quite hard in order to give a low value for κ, resulting
in a slightly fiddly argument for the proof of Lemma 2.13, which does not obviously extend to three
dimensional space. For the three dimensional model, we shall be a little lazy and shall work with an
analogue of the event ruhα
τ (u) which gives a value for κ∗ which could easily be improved upon, but which
allows for a very simple proof in our three dimensional counterpart to Lemma 2.13.
We redefine pnα
γ,τ (u) and uhα
τ (u) in the obvious way, and also define our analogue of ruhα
τ (u) as follows:
22
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
• Suppose that P is a plane passing through N †(u). Let A1 be all those points in N †(u) on or
†
i ) = γ(2w + 1)3, then
If u is an α node, we let
γ,τ (u) be the event that there exists some γ-partial neighbourhood of u which contains at
above P , and let A2 be all those points in N †(u) on or below P . If µ(A
we call Ai a γ-partial neighbourhood of u, with defining plane P .
pnα
least τ (2w + 1)3 many α nodes in the initial configuration (and similarly for β).
• We say that uhα
τ (u) holds if there are strictly less than τ (2w + 1)3 many α nodes in N (u) in the
initial configuration. If u = (x, y, z) then the extended neighbourhood of u is the set of nodes
(x(cid:48), y(cid:48), z(cid:48)) such that x − x(cid:48),y − y(cid:48)z − z(cid:48) ≤ (cid:100) 3
τ (u) holds if there are
strictly less than τ (3w + 1)3 many α nodes in the extended neighbourhood of u in the initial
configuration.
2 w(cid:101). We say that euhα
Arguing almost exactly as in the proof of Lemma 2.2, we then obtain the following three dimensional
analogue:
Lemma 4.1. Suppose that κ∗ < τα < 0.5 and τβ(cid:50) τα. Choose τ such that τβ(cid:50) τ < τα and κ∗ < τ . Then
2 and ζ > 1 such that for 0 (cid:28) w (cid:28) n, and for a node u selected uniformly at random,
τ (u)) > ζ w · P(uhβ
τβ
(u)).
there exists γ > 1
P(euhα
τ (u)) > ζ w · P(pnα
γ,τα
(u)) and P(euhα
τ (u), i.e. ejuα
τ (u) in terms of euhα
τ (u) was defined in terms of ruhα
The only significant task in modifying the proof to work in three dimensions is then in forming the
analogue of Lemma 2.13, which is now much easier. We define ejuα
τ (u), just as
τ(cid:48)(u) holds if there are less than τ(cid:48)(3w + 1)3 many α
rjuα
nodes in the extended neighbourhood of u, but changing the type of (3w + 1)2 β nodes in this extended
neighbourhood would cause this not to be the case. Rather than inductively building a sequence of larger
and larger rectangles for which all nodes will become of type β, we simply observe that satisfaction of
ejuα
τ (u) and (the three dimensional analogue of) Smoothk1,1(u) for appropriately large k1 and small 1,
suffices to ensure that all α nodes in a cube centred at u with sides of length w, are unhappy in the initial
configuration. In particular this certainly implies that all nodes in the interior of a sphere of diameter
w centred at u will be of type β by the end of stage 1. Now satisfaction of Smoothk1,1(u) suffices to
ensure that all α nodes on the outer boundary of this sphere are hopeful at the end of stage 1, and it
then follows that this sphere of β nodes will grow at subsequent stages as required.
5. Deferred proofs
5.1. The proof of Lemma 2.2. We previously defined γ-partial neighbourhoods and right-extended
neighbourhoods. It will be useful to consider also some related notions:
• By the lower neighbourhood of a node u = (x, y), denoted ln(u), we mean the set of nodes
{(x(cid:48), y(cid:48)) ∈ N (u) : y(cid:48) < y or (y(cid:48) = y & x(cid:48) ≤ x)}. If u is an α node and τ ∈ [0, 1], we let lnα
τ (u)
be the event that ln(u) contains at least τ (2w + 1)2 many α nodes in the initial configuration.
So lnα
(u) is the event that the lower neighbourhood already has enough α nodes to ensure that
τα
τ (u) denote the corresponding event when α is replaced everywhere by β.6
u is happy. We let lnβ
• Suppose that (cid:96) is any straight line passing through u. Let A1 be all those nodes in N (u) strictly
below (cid:96), and let A2 be all those nodes in N (u) strictly above (cid:96). Suppose that (cid:96) intersects the
boundary of N †(u) at x1 and x2 (ordered arbitrarily). Let (cid:96)i be the (closed) straight line segment
between xi and u. If L ⊂ N (u) is such that there exist i, j ∈ {0, 1}, L = Ai∪ ((cid:96)j ∩N (u)) then we
call L a rotated lower neighbourhood of u, with defining line (cid:96). If u is an α node and τ ∈ [0, 1], we
6One should think of ln as (standing for) l-ower n-eighbourhood, and rn as r-otated n-eighbourhood.
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
23
Figure 7. Lower, rotated lower and γ-partial neighbourhoods respectively.
τ (u) be the event that there exists some rotated lower neighbourhood of u which contains
let rnα
at least τ (2w + 1)2 many α nodes in the initial configuration (and similarly for β).
So the various neighbourhoods with which we are concerned are really very simple, and are depicted in
Figure 7. The lower neighbourhood is (roughly) just the bottom half of the neighbourhood, while the
rotated lower neighbourhood is just a rotation of this. Recall that a γ-partial neighbourhood of u is
just a subset of N (u) defined by a straight line, and which has measure γ(2w + 1)2. The right extended
neighbourhood results roughly from tacking on an extra half of N (u) to the right.
We prove the following for u chosen uniformly at random:
(1) τ > κ implies that there exists ζ > 1 such that for 0 (cid:28) w (cid:28) n, P(ruhα
(2) τβ (cid:67) τ implies that there exists ζ > 1 such that for 0 (cid:28) w (cid:28) n, P(ruhα
2 such that for 0 (cid:28) w (cid:28) n, P(lnα
(3) τα > τ implies that there exists γ > 1
τ (u)) > ζ w · P(lnα
τ (u)) > ζ w · P(uhβ
τ (u)) > P(pnα
γ,τα
(u)).
τ (u)).
(u)).
τβ
γ,τα
(u)).
τ (u)) > ζ w · P(pnα
2 such that for 0 (cid:28) w (cid:28) n, P(ruhα
k(cid:48)≤k b(N, k(cid:48)) and b(N,≥ k) = (cid:80)
With these facts established, (1) and (3) combined then suffice to show that, for the ζ guaranteed by (1),
there exists γ > 1
First of all we note some basic facts about the binomial distribution. Define b(N, k) = 2−N(cid:0)N
k) = (cid:80)
k(cid:48)≥k b(N, k(cid:48)). For the duration of this proof, we shall write
f (N ) (cid:39) g(N ) in order to indicate that there exist polynomials P and Q such that f (N ) · P (N ) > g(N )
and g(N ) · Q(N ) > f (N ) for all N . We also write f (N ) (cid:37) g(N ) to indicate that there exists f(cid:48) such
that f (N ) ≥ f(cid:48)(N ) for all N and f(cid:48)(N ) (cid:39) g(N ). Now it follows, by direct inspection and from standard
bounds on the tail of the binomial distribution (see for example Theorem 1.1. of [5]) that the following
hold for fixed 0 < τ1 < 0.5 < τ2 < 1 and for all integers c, d ∈ [−2, 2]:
(U1)
(U2)
b(N,≤ (cid:98)τ1N(cid:99)) (cid:39) b(N + c,≤ (cid:98)τ1N(cid:99) + d) (cid:39) b(N,(cid:98)τ1N(cid:99)) (cid:39) b(N + c,(cid:98)τ1N(cid:99) + d);
b(N,≥ (cid:98)τ2N(cid:99)) (cid:39) b(N + c,≥ (cid:98)τ2N(cid:99) + d) (cid:39) b(N,(cid:98)τ2N(cid:99)) (cid:39) b(N + c,(cid:98)τ2N(cid:99) + d).
(cid:1), b(N,≤
k
Now put N = 2w(w +1), so that (2w +1)2 = 2N +1. For τα, τ and τβ in the range given by the conditions
of the lemma, it follows from (U1) and (U2) that:
(U3)
τ (u)) (cid:39) b(N,(cid:100)2τ N(cid:101)) and P(uhβ
τ (u)) (cid:37) b(3N,(cid:100)3τ N(cid:101)), P(lnα
(u)) (cid:39) b(2N,(cid:98)τβ2N(cid:99)).
P(ruhα
τβ
Proving (1). From (U3) we get that P(ruhα
Stirling's approximation, we see that the powers of e cancel and:
τ (u))/P(lnα
τ (u))) (cid:37) b(3N,(cid:100)3τ N(cid:101))/b(N,(cid:100)2τ N(cid:101)). Applying
24
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
P(ruhα
τ (u))/P(lnα
Simplifying this we get:
τ (u))) (cid:37) 2N (3N )3N +0.5(2τ N )2τ N +0.5(N (1 − 2τ ))N (1−2τ )+0.5
23N (3N τ )3N τ +0.5(3N (1 − τ ))3N (1−τ )+0.5N N +0.5
τ (u)) (cid:37)(cid:16)
(1−2τ )1−2τ +1/(2N )
(cid:17)N
.
.
P(ruhα
τ (u))/P(lnα
(U4)
Now since τ > κ we have (1 − 2τ )1−2τ > 22(1−τ )τ τ (1 − τ )3(1−τ ). Thus there exists ζ > 1 so that for all
sufficiently large N the term inside the brackets in (U4) is > ζ, giving the result.
22−2τ−1/(2N )τ τ (1−τ )3(1−τ )+1/(2N )
Proving (2). From (U3) we get that P(ruhα
ing Stirling's approximation, and simplifying as before we get:
τ (u))/P(uhβ
τβ
(u)) (cid:37) b(3N,(cid:100)3τ N(cid:101))/b(2N,(cid:98)τβ2N(cid:99)). Apply-
(U5)
P(ruhα
τ (u))/P(uhβ
τβ
(u)) (cid:37)
2τβ +1/(2N )
β
(1−τβ )2(1−τβ )+1/(2N )
21−1/(2N )τ 3τ +1/(2N )(1−τ )3(1−τ )+1/(2N )
β
τ (u)) > ζ w · P(lnα
τ(cid:48)(u)). We then look to show that P(rnα
(1 − τβ)2(1−τβ ) > 2τ 3τ (1 − τ )3(1−τ ). Thus there exists ζ > 1 so that
Now since τβ (cid:67) τ we have that τ 2τβ
for all sufficiently large N the term inside the brackets in (U5) is > ζ, giving the result.
Proving (3). Choose τ(cid:48) with τ < τ(cid:48) < τα. Note first that there exists ζ > 1 such that for 0 (cid:28) w (cid:28) n,
P(lnα
τ(cid:48)(u))
-- meaning that once again there exists ζ > 1 such that for 0 (cid:28) w (cid:28) n, P(lnα
τ(cid:48)(u)). In
order to see this, observe first that each rotated lower neighbourhood of u contains precisely the same
number of nodes, and so is precisely as likely to contain τ(cid:48)(2w +1)2 many nodes of type α. As the defining
line (cid:96) rotates through an angle of 2π, each node leaves the corresponding rotated lower neighbourhood
precisely once and also joins precisely once. Thus there are at most 2(2w + 1)2 distinct rotated lower
neighbourhoods of u and the probability that at least one of them contains at least τ(cid:48)(2w + 1)2 many
nodes of type α is at most 2(2w + 1)2 times the probability that a given one does. Finally we show that,
for an appropriate choice of γ > 1
τ(cid:48)(u)). So
define:
τ(cid:48)(u)) is at most 2(2w + 1)2 · P(lnα
2 , there exists a constant c such that P(pnα
τ (u)) > ζ w · P(rnα
(u)) < c · P(rnα
γ,τα
(cid:18) τ
(cid:19)N
.
γ =
τα + τ(cid:48)
4τ(cid:48)
.
Suppose we are given that pnα
(u) holds. Given A which is a γ-partial neighbourhood of u containing at
least τα(2w + 1)2 many nodes of type α, with defining line (cid:96) say, consider a rotated lower neighbourhood
of u which is a subset of A, with defining line parallel to (cid:96). Call this rotated lower neighbourhood A0.
Let z0 be the expected number of α nodes in A0 (given the stated conditions on A), and let z1 be the
actual number of α nodes in A0. As w → ∞:
γ,τα
z0/(2w + 1)2 → z∗ :=
2τ(cid:48)τα
(τα + τ(cid:48))
,
and for any > 0, the probability that z1/(2w + 1)2 − z∗ > tends to 0. The result follows since
(2τα)/((τα + τ(cid:48))) > 1.
5.2. The proof of Lemma 2.7. If θ(w) is the proportion of the nodes in N (cid:5)(u) which are of type α
τ (u) holds, then θ(w) → τ as w → ∞. Since we consider k1 and 1 fixed it suffices to show that
when rjuα
for any > 0 and for a given A ∈ Ik1(u) the following occurs with probability > 1 − for 0 (cid:28) w (cid:28) n:
(1) If A ⊂ N (cid:5)(u) then the proportion of the nodes in A which are of type α is in the interval
[τ − 1, τ + 1].
UNPERTURBED SCHELLING SEGREGATION IN 2D AND 3D
25
(2) If A is disjoint from N (cid:5)(u) then the proportion of the nodes in A which is of type α is in the
interval [0.5 − 1, 0.5 + 1].
Now (2) follows simply from the weak law of large numbers. Also, (1) follows from Chebyshev's inequality
and standard results for the variance of a hypergeometric distribution. If φ is the proportion of the nodes
in A which are of type α then:
P(φ − θ(w) > 1/2) < A−2(1/2)−2Var(θ) = O(1)A−1.
References
[1] D.J. Amit, H. Gutfreund, and H. Sompolinsky, Spin-glass models of neural networks, Phys. Rev. A, 32, 1985.
[2] A-L. Barab´asi, R. Albert, Emergence of scaling in random networks, Science 286 (5439): 509-512, 1999.
[3] G. Barmpalias, R. Elwes, and A. Lewis-Pye. Digital morphogenesis via Schelling segregation. In 55th Annual IEEE
Symposium on Foundations of Computer Science, Oct. 18-21, Philadelphia, 2014. FOCS 2014.
[4] G. Barmpalias, R. Elwes, and A. Lewis-Pye. Tipping Points in 1-Dimensional Schelling Models with Switching Agents.
J. Stat. Phys., 158:806 -- 852, 2015.
[5] B. Bollob´as, Random Graphs, Cambridge Studies in Advanced Mathematics, 2001.
[6] C. Brandt, N. Immorlica, G. Kamath, R. Kleinberg, An Analysis of One-Dimensional Schelling Segregation, Proc. 44th
Annual ACM Symposium on Theory of Computing (STOC 2012).
[7] K. Byrka, A. Jdrzejewski, K. Sznajd-Weron, and R. Weron, 2016. Difficulty is critical: The importance of social factors
in modeling diffusion of green products and practices. Renewable and Sustainable Energy Reviews, 62, pp.723-735.
[8] A. Canning and J-P. Naef, Phase Diagrams and the Instability of the Spin Glass States for the Diluted Hopfield Neural
Network Model, J. Phys 1, France 2, p.1791, 1992.
[9] I. P. Castillo and N.S. Skantzos, The Little-Hopfield model on a sparse random graph, J. Phys. A: Math. Gen. 37
(2004) 9087 -- 9099.
[10] L. Dall'Asta, C. Castellano, M. Marsili, Statistical physics of the Schelling model of segregation, J. Stat. Mech, 7, 2008.
[11] L. Gauvin, J. Vannemenus, J.-P. Nadal, Phase diagram of a Schelling segregation model, European Physical Journal
B, 70, 293 -- 304, 2009.
[12] A. D. Henry, P. Pra(cid:32)lat, and C. Zhang, Emergence of segregation in evolving social networks. Proceedings of the National
Academy of Sciences, 108(21):8605-8610, May 2011.
[13] J.J. Hopfield, Neural networks and physical systems with emergent collective computational abilities, Proceedings of
the National Academy of Sciences, vol. 79 no. 8 pp. 2554 -- 2558, April 1982.
[14] J. Kleinberg, Cascading Behavior in Networks: Algorithmic and Economic Issues, Algorithmic Game Theory, N. Nisan,
T. Roughgarden, E. Tardos, V. Vazirani, eds., Cambridge University Press, 2007.
[15] G. ´Odor, Self-organising, two temperature Ising model describing human segregation, International journal of modern
physics C, 3, 393 -- 398, 2008.
[16] R. Pancs and N. Vriend, Schellings spatial proximity model of segregation revisited, Journal of Public Economics,
91(1-2):1-24, 2007.
[17] M. Pollicott, and H. Weiss, The dynamics of Schelling-type segregation models and a non-linear graph Laplacian
variational problem, Adv. Appl. Math., 27, 17-40, 2001.
[18] T. Schelling, Models of segregation, The American Economic Review, pages 488-493, 1969.
[19] T. Schelling, Micromotives and Macrobehavior, New York, Norton, 1978.
[20] Schuman, H., C. Steeh, L. Bobo, and M. Krysan. 1997. Racial Attitudes in America: Trends and Interpretations
(revised edition). Cambridge, MA: Harvard University Press.
[21] D. Stauffer and S. Solomon, Ising, Schelling and self-organising segregation, European Physical Journal B, 57, 473 -- 479,
2007.
[22] D. Vinkovi´c and A. Kirman, A Physical Analogue of the Schelling Model, Proceedings of the National Academy of
Sciences, no. 51, volume 103, 19261-19265, 2006.
[23] D.J. Watts, S.H. Strogatz, Collective dynamics of 'small-world' networks, Nature, 393 (6684): 440-442, 1998.
[24] N.C. Wormald, Differential equations for random processes and random graphs, Annals of Applied Probability, 5:1217-
1235, 1995.
[25] H.P. Young, Individual Strategy and Social Structure: An Evolutionary Theory of Institutions. Princeton, NJ: Prince-
ton University Press, 1998.
[26] J. Zhang, A dynamic model of residential segregation, Journal of Mathematical Sociology, 28(3):147 -- 170, 2004.
[27] J. Zhang, Residential segregation in an all-integrationist world, Journal of Economic Behavior & Organization,
54(4):533-550, 2004.
26
GEORGE BARMPALIAS, RICHARD ELWES, AND ANDY LEWIS-PYE
[28] J. Zhang, Tipping and residential segregation: A unified Schelling model, Journal of Regional Science, 51:167-193,
Feb. 2011.
George Barmpalias: (1) State Key Lab of Computer Science, Institute of Software, Chinese Academy of
Sciences, Beijing 100190, China and (2) School of Mathematics, Statistics and Operations Research, Victoria
University, Wellington, New Zealand
E-mail address: [email protected]
URL: http://barmpalias.net
Richard Elwes, School of Mathematics, University of Leeds, LS2 9JT Leeds, U.K.
E-mail address: [email protected]
URL: http://richardelwes.co.uk
Andy Lewis-Pye, Mathematics, Columbia House, London School of Economics, WC2A 2AE, London, U.K.
E-mail address: [email protected]
URL: http://aemlewis.co.uk
|
1107.4153 | 1 | 1107 | 2011-07-21T04:15:25 | Performance and Convergence of Multi-user Online Learning | [
"cs.MA",
"cs.LG"
] | We study the problem of allocating multiple users to a set of wireless channels in a decentralized manner when the channel quali- ties are time-varying and unknown to the users, and accessing the same channel by multiple users leads to reduced quality due to interference. In such a setting the users not only need to learn the inherent channel quality and at the same time the best allocations of users to channels so as to maximize the social welfare. Assuming that the users adopt a certain online learning algorithm, we investigate under what conditions the socially optimal allocation is achievable. In particular we examine the effect of different levels of knowledge the users may have and the amount of communications and cooperation. The general conclusion is that when the cooperation of users decreases and the uncertainty about channel payoffs increases it becomes harder to achieve the socially opti- mal allocation. | cs.MA | cs |
Performance and Convergence of Multi-user
Online Learning
Cem Tekin and Mingyan Liu
Department of Electrical Engineering and Computer Science
University of Michigan, Ann Arbor, Michigan, 48109-2122
{cmtkn, mingyan}@umich.edu
Abstract. We study the problem of allocating multiple users to a set
of wireless channels in a decentralized manner when the channel quali-
ties are time-varying and unknown to the users, and accessing the same
channel by multiple users leads to reduced quality due to interference.
In such a setting the users not only need to learn the inherent channel
quality and at the same time the best allocations of users to channels
so as to maximize the social welfare. Assuming that the users adopt a
certain online learning algorithm, we investigate under what conditions
the socially optimal allocation is achievable. In particular we examine
the effect of different levels of knowledge the users may have and the
amount of communications and cooperation. The general conclusion is
that when the cooperation of users decreases and the uncertainty about
channel payoffs increases it becomes harder to achieve the socially opti-
mal allocation.
Keywords: multi-user learning, multi-armed bandits, spectrum sharing, con-
gestion games
1
Introduction
In this paper we study the dynamic spectrum access and spectrum sharing prob-
lem in a learning context. Specifically, we consider a set of N common channels
shared by a set of M users. A channel has time varying rate r(t), and its statis-
tics are not completely known by the users. Thus each user needs to employ
some type of learning to figure out which channels are of better quality, e.g., in
terms of their average achievable rates. At the same time, simultaneous use of
the same channel by multiple users will result in reduced rate due to interference
or collision. The precise form of this performance degradation may or may not
be known to the user. Thus the users also need to use learning to avoid excess
interference or congestion. Furthermore, each user may have private information
that is not shared, e.g., users may perceive channel quality differently due to
difference in location as well as individual modulation/coding schemes.
Without a central agent, and in the presence of information decentralization
described above, we are interested in the following questions: (1) for a given
common learning algorithm, does the multiuser learning process converge, and
(2) if it does, what is the quality of the equilibrium point with respect to a
globally optimal spectrum allocation scheme, one that could be computed for a
global objective function with full knowledge of channel statistics as well as the
users' private information.
A few recent studies have addressed these questions in some special cases.
For instance, in [3] it was shown that learning using a sample-mean based index
policy leads to a socially optimal (sum of individual utilities) allocation when
channels evolve as iid processes and colliding players get zero reward provided
that this optimal allocation is such that each user occupies one of the M best
channels (in terms of average rates). This precludes the possibility that not all
users may have the same set of M best channels, and that in some cases the best
option is for multiple users to share a common channel, e.g., when N < M .
In this study we investigate under what conditions the socially optimal al-
location is achievable by considering different levels of communication (or coop-
eration) allowed among users, and different levels of uncertainty on the channel
statistics. The general conclusion, as intuition would suggest, is that when the
cooperation of users increases and the channel uncertainty decreases it becomes
easier to achieve the socially optimal welfare. Specifically, we assume that the
rate (or reward) user i gets from channel j at time t is of the form rj (t)gj(nj(t))
where rj (t) is the rate of channel j at time t, nj(t) is the number of users using
channel j at time t, and gj is the user independent interference function (IF) for
channel j. This model is richer than the previously used models [3,14,16] since
rj(t) can represent environmental effects such as fading or primary user activity,
while gj captures interactions between users. We consider the following three
cases.
In the first case (C1), each channel evolves as an iid random process in time,
the users do not know the channel statistics, nor the form of the interference, nor
the total number of users present in the system, and no direct communication
is allowed among users. A user can measure the overall rate it gets from using
a channel but cannot tell how much of it is due to the dynamically changing
channel quality (i.e., what it would get if it were the only user) vs. interference
from other users. In this case, we show that if all users follow the Exp3 algorithm
[7] then the channel allocation converges to a set of pure Nash equilibria (PNE)
of a congestion game defined by the IFs and mean channel rates. In this case a
socially optimal allocation cannot be ensured, as the set of PNE are of different
quality, and in some cases the socially optimal allocation may not be a PNE.
In the second case (C2), each channel again evolves as an iid random process
in time, whose statistics are unknown to the user. However, the users now know
the total number of users in the system, as well as the fact that the quantitative
impact of interference is common to all users (i.e., user independent), though
the actual form of the interference function is unknown. In other words the rate
of channel j at time t is perceived by user i as hj(t, nj(t)) so user i cannot
distinguish between components rj (t) and gj(nj(t)). Furthermore, users are now
allowed minimal amount of communication when they happen to be in the same
channel, specifically to find out the total number of simultaneous users of that
channel. In this case we present a sample-mean based randomized learning policy
that achieves socially optimal allocation as time goes to infinity, with a sub-linear
regret over the time horizon with respect to the socially optimal allocation.
In the third case (C3), as in case (C2) the users know the total number
of users in the system, as well as the fact that the IF is user independent and
decreasing without knowing the actual form of the IF. However, the channels are
assumed to have constant, albeit unknown, rates. We show that even without
any communication among users, there is a randomized learning algorithm that
achieves the socially optimal allocation in finite time.
It's worth pointing out that in the settings outlined above, the users are
non-strategic, i.e., each user simply follow a pre-set learning rule rather than
playing a game. In this context it is reasonable to introduce minimal amount of
communication among users and assume they may cooperate. It is possible that
even in this case the users may not know their IF but only the total rate they
get for lack of better detecting capabilities (e.g., they may only be able to detect
the total received SNR as a result of channel rate and user interference).
Online learning by a single user was studied by [1,4,6,15], in which sample-
mean based index policies were shown to achieve logarithmic regret with respect
to the best single-action policy without a priori knowledge of the statistics, and
are order-optimal, when the rewards are given by an iid process. In [5,21,22]
Markovian rewards are considered, with [22] focusing on restless reward pro-
cesses, where a process continues to evolve according to a Markov chain regard-
less of the users' actions. In all these studies learning algorithms were developed
to achieve logarithmic regret. Multi-user learning with iid reward processes have
been studied in a dynamic spectrum context by [3,11,16], with a combinatorial
structure adopted in [11], and with collision and random access models in [3,16].
In [13], convergence of multi-user learning with Exp3 algorithm to pure Nash
equilibrium is investigated under the collision and fair sharing models. In the col-
lision model, when there is more than one user on a channel all get zero reward,
whereas in the random access model one of them, selected randomly, gets all the
reward while others get zero reward. In the fair sharing model, a user's utility is
inversely proportional to the number of users who are on the same channel with
the user. Note that these models do not capture more sophisticated communi-
cation schemes where the rate a user gets is a function of the received SNR of
the form gj(n) = fj(
) = where Pt is the nominal transmit power of
all users and N0 the noise. Moreover, in the above studies the socially optimal
allocation is a rather simple one: it is the orthogonal allocation of users to the
first M channels with the highest mean rewards. By contrast, we model a more
general interference relationship among users, in which an allocation with users
sharing the same channel may be the socially optimal one. The socially optimal
allocations is not trivial in this case and additional mechanisms may be needed
for the learning algorithms to converge.
N0+(n−1)Pt
Pt
All of the above mentioned work assumes some level of communication be-
tween the users either at the beginning or during the learning. If we assume
no communication between the users, achieving the socially optimal allocation
seems very challenging in general. Then one may ask if it is possible to achieve
some kind of equilibrium allocation. Kleinberg et. al. [14] showed that it is pos-
sible for the case when the channel rates are constant and the users do not
know the IFs. They show that when the users use aggregate monotonic selection
dynamics, a variant of Hedge algorithm [10], the allocation converges to weakly
stable equilibria which is a subset of Nash equilibria (NE) of the congestion game
defined by the IFs. They show that for almost all congestion games weakly stable
equilibria is the same as PNE.
Other than the work described above [2] considers spatial congestion games, a
generalization of congestion games and gives conditions under which there exists
a PNE and best-response play converges to PNE. A mechanism design approach
for socially optimal power allocation when users are strategic is considered in
[12].
The organization of the remainder of this paper is as follows. In Sect. 2 we
present the notations and definitions that will be used throughout the paper.
In Sects. 3, 4, 5 we analyze the cases stated in (C1), (C2), (C3) and derive the
results respectively. Conclusion and future research is given in Sect. 6.
2 Preliminaries
Denote the set of users by M = {1, 2, . . . , M }, and the set of channels N =
{1, 2, . . . , N }. Time is slotted and indexed by t = 1, 2, . . . and a user can select
a single channel at each time step t. Without loss of generality let rj(t) ∈ [0, 1]
be the rate of channel j at time t such that {rj(t)}t=1,2,... is generated by an
iid process with support [0, 1] and mean µj ∈ [0, 1]. Let gj : N → [0, 1] be the
interference function (IF) on channel j where gj(n) represents the interference
when there are n users on channel j. We express the rate of channel j seen by
a user as hj(t) = rj(t)gj(nj (t)) when a user does not know the total number of
users nj(t) using channel j at time t as in cases (C1) and (C3). When a user knows
nj(t), we express the rate of channel j at time t as hj,nj (t)(t) = rj(t)gj(nj(t)) as
in case (C2). Let Si = N be the set of feasible actions of user i and σi ∈ Si be the
action, i.e., channel selected by user i. Let S = S1 × S2 × . . . × SN = N M be the
set of feasible action profiles and σ = {σ1, σ2, . . . , σM } ∈ S be the action profile
of the users. Throughout the discussion we assume that the action of player i at
time t, i.e., σπi
i (t) is determined by the policy πi. When πi is deterministic, πi(t)
is in general a function from all past observations and decisions of user i to the set
of actions Si. When πi is randomized, πi(t) generates a probability distribution
over the set of actions Si according to all past observations and decisions of user
i from which the action at time t is sampled. Since the dependence of actions to
the policy is trivial we use σi(t) to denote the action of user i at time t, dropping
the superscript πi.
Let Kj(σ) be the set of users on channel j when the action profile is σ. Let
j=1 µjKj(σ)gj (Kj(σ))
be the set of socially optimal allocations and denote by σ∗ any action profile
i=1 µσi gσi(Kσi (σ)) = arg maxσ∈SPN
A∗ = arg maxσ∈SPM
i=1 µσ∗
i gσ∗
i (Kσ∗
PM
that is in the set A∗. Let v∗ denote the socially optimal welfare, i.e., v∗ =
j denote the payoff a user gets from channel j
under the socially optimal allocation, i.e., v∗
j = µjgj(Kj(σ∗)) if Kj(σ∗) 6= 0.
Note that any permutation of actions in σ∗ is also a socially optimal allocation
since IFs are user-independent.
i (σ∗)) and v∗
For any policy π, the regret at time n is
R(n) = nv∗ − E" n
Xt=1
M
Xi=1
rσi(t)(t)gσi(t)(Kσi(t)(σ(t)))# ,
where expectation is taken with respect to the random nature of the rates and
the randomization of the policy. Note that for a deterministic policy expectation
is only taken with respect to the random nature of the rates. For any randomized
policy πi, let pi(t) = (pi1(t), pi2(t), . . . , piN (t)) be the mixed strategy of user i at
time t, i.e., a probability distribution on {1, 2, . . . , N }. For a profile of policies
π = [π1, π2, . . . , πM ] for the users let p(t) = (p1(t)T , p2(t)T , . . . pM (t)T )T be
the profile of mixed strategies at time t, where pi(t)T is the transpose of pi(t).
Then σi(t) is the action sampled from the probability distribution pi(t). The
dependence of p to π is trivial and not shown in the notation.
3 Allocations Achievable with Exp3 Algorithm (Case 1)
We start by defining a congestion game. A congestion game [17,18] is given by
the tuple (M, N , (Σi)i∈M, (hj)j∈N ), where M denotes a set of players (users),
N a set of resources (channels), Σi ⊂ 2N the strategy space of player i, and
hj : N → R a payoff function associated with resource j, which is a function of
the number of players using that resource. It is well known that a congestion
game has a potential function and the local maxima of the potential function
corresponds to PNE, and every sequence of asynchronous improvement steps is
finite and converges to PNE.
In this section we relate the strategy update rule of Exp3 [7] under assump-
tions (C1) to a congestion game. Exp3 as given in Fig. 1 is a randomized al-
gorithm consisting of an exploration parameter γ and weights wij that depend
exponentially on the past observations where i denotes the user and j denotes
the channel. Each user runs Exp3 independently but we explicitly note the user
dependence because a user's action affects other users' updates.
At any time step before the channel rate and user actions are drawn from
the corresponding distributions, let Rj denote the random variable correspond-
ing to the reward of the jth channel. Let Gij = gj(1 + K ′
j(i)) be the random
variable representing the payoff user i gets from channel j where K ′
j(i) is the
random variable representing the number of users on channel j other than user
i. Let Uij = RjGij and ¯uij = Ej[E−i[Uij]] be the expected payoff to user i by
using channel j where E−i represents the expectation taken with respect to the
randomization of players other than i, Ej represents the expectation taken with
respect to the randomization of the rate of channel j. Since the channel rate is
independent of users' actions ¯uij = µj ¯gij where ¯gij = E−i[Gij ].
Exp3 (for user i)
1: Initialize: γ ∈ (0, 1), wij (t) = 1, ∀j ∈ N , t = 1
2: while t > 0 do
3:
pij(t) = (1 − γ)
wij (t)
l=1 wil(t)
+
γ
N
PN
Sample σi(t) from the distribution on pi(t) = [pi1(t), pi2(t), . . . , piN (t)]
Play channel σi(t) and receive reward hσi(t)(t)
for j = 1, 2, . . . , N do
if j = σi(t) then
pij (t)N (cid:17)
Set wij (t + 1) = wij (t) exp(cid:16) γhσi(t)(t)
Set wij (t + 1) = wij (t)
4:
5:
6:
7:
8:
else
9:
10:
end if
11:
end for
12:
t = t + 1
13:
14: end while
Fig. 1. pseudocode of Exp3
Lemma 1. Under (C1) when all players use Exp3, the derivative of the continuous-
time limit of Exp3 is the replicator equation given by
ξij =
1
N
(µjpij)
N
Xl=1
pil(¯gij − ¯gil) .
Proof. Note that
(1 − γ)wij (t) =
N
Xl=1
wil(t)(cid:16)pij(t) −
γ
N(cid:17) .
(1)
We consider the effect of user i's action σi(t) on his probability update on channel
j. We have two cases: σi(t) = j and σi(t) 6= j. Let Aγ,t
pij (t)N(cid:17).
i,j = exp(cid:16) γUij (t)
Consider the case σi(t) = j.
pij (t + 1) =
(1 − γ)wij(t)Aγ,t
i,j
PN
l=1 wil(t) + wij (t)(cid:0)Aγ,t
i,j − 1(cid:1)
+
γ
N
.
(2)
Substituting (1) into (2)
pij(t + 1) =
=
+
γ
N
N
PN
l=1 wil(t)(cid:0)pij (t) − γ
l=1 wil(t)(cid:16)1 + pij (t)− γ
PN
N(cid:1) Aγ,t
(cid:0)pij(t) − γ
(cid:0)Aγ,t
i,j − 1(cid:1)
1 + pij (t)− γ
1−γ
1−γ
+
i,j
N
i,j
N(cid:1) Aγ,t
i,j − 1(cid:1)(cid:17)
(cid:0)Aγ,t
γ
N
.
The continuous time process is obtained by taking the limit γ → 0, i.e., the
rate of change in pij with respect to γ as γ → 0. Then, dropping the discrete
time script t,
i,j − 1(cid:1)(cid:17)
N
N
pij N Aγ,t
i,j +(cid:0)pij − γ
i,j (cid:16) pij − 1
N(cid:1) Aγ,t
(cid:16)1 + pij − γ
i,j(cid:17)(cid:16)1 + pij − γ
N(cid:1) Uij
1−γ (cid:0)Aγ,t
i,j − 1(cid:1)(cid:17)2
(cid:16)1 + pij − γ
1−γ (cid:0)Aγ,t
i,j + pij − 1
i,j(cid:1)(cid:17)
1−γ (cid:0) γ
i,j − 1(cid:1)(cid:17)2
1−γ (cid:0)Aγ,t
(1−γ)2 Aγ,t
N Aγ,t
1
N
+
N
N
N
pij = lim
γ→0
dpij
dγ
N Aγ,t
γ→0(cid:16) −1
= lim
+ (cid:0)pij − γ
Uij(1 − pij)
.
=
N
pij (t + 1) =
Consider the case σi(t) = k 6= j. Then,
Thus
pij = lim
γ→0
(1 − γ)wij (t)
−1
+
N
=
1−γ
pij(t) − γ
N
l=1 wil(t) + wik(t)(cid:16)Aγ,t
PN
1 + pik(t)− γ
(cid:16)Aγ,t
i,k − 1(cid:17)
N (cid:16)1 + pik− γ
i,k − 1(cid:17)(cid:17)
1−γ (cid:16)Aγ,t
i,k − 1(cid:17)(cid:17)2
(cid:16)1 + pik− γ
1−γ (cid:16)Aγ,t
i,k + pik− 1
N(cid:1)(cid:16) pik− 1
1−γ (cid:16) γ
(1−γ)2 Aγ,t
i,k − 1(cid:17)(cid:17)2
(cid:16)1 + pik− γ
1−γ (cid:16)Aγ,t
N
N
N
N
N
+ (cid:0)pij − γ
= −
pikUik
N
.
+
γ
N
i,k − 1(cid:17)
γ
N
.
i,k(cid:17)(cid:17)
N Aγ,t
+
1
N
(3)
(4)
Then from (3) and (4), the expected change in pij with respect to the prob-
ability distribution pi of user i over the channels is
¯pij = Ei[ pij] =
1
N
pij Xl∈N −{j}
pil(Uij − Uil).
Taking the expectation with respect to the randomization of channel rates and
other users' actions we have
ξij = Ej[E−i[¯pij ]]
=
=
1
N
1
N
pij Xl∈N −{j}
Xl=1
(µjpij)
N
pil (Ej[E−i[Uij]] − Ej[E−i[Uil]])
pil(¯gij − ¯gil) .
⊓⊔
Lemma 1 shows that the dynamics of a user's probability distribution over the
actions is given by a replicator equation which is commonly studied in evolu-
tionary game theory [19,20]. With this lemma we can establish the following
theorem.
Theorem 1. For all but a measure zero subset of [0, 1]2N from which the µj's
and gj's are selected, when γ in Exp3 is arbitrarily small, the action profile
converges to the set of PNE of the congestion game (M, N , (Si)i∈M, (µjgj)j∈N ).
Proof. Because the replicator equation in Lemma 1 is identical to the replicator
equation in [14], the proof of converge to PNE follows from [14]. Here, we briefly
explain the steps in the proof. Defining the expected potential function to be
the expected value of the potential function φ where expectation is taken with
respect to the user's randomization one can show that the solutions of the repli-
cator equation converges to the set of fixed points. Then the stability analysis
using the Jacobian matrix yields that every stable fixed point corresponds to a
Nash equilibrium. Then one can prove that for any stable fixed point the eigen-
values of the Jacobian must be zero. This implies that every stable fixed point
corresponds to a weakly stable Nash equilibrium strategy in the game theoretic
sense. Then using tools from algebraic geometry one can show that almost every
weakly stable Nash equilibrium is a pure Nash equilibrium of the congestion
game.
We also need to investigate the error introduced by treating the discrete time
update rule as a continuous time process. However, by taking γ infinitesimal
we can approximate the discrete time process by the continuous time process.
For a discussion when γ is not infinitesimal one can define approximately stable
⊓⊔
equilibria [14].
The main difference between Exp3 and Hedge [14] is that in Exp3 users
do not need to observe the payoffs from the channels that they do not select,
whereas Hedge assumes complete observation. In addition to that, we considered
the dynamic channel rates which is not considered in [14].
4 An Algorithm for Socially Optimal Allocation with
Sub-linear Regret (Case 2)
2M −1+2γ
2M
tations so k∗ = arg maxk∈KPN
µjgj(kj). Then the stability condition says that arg maxk∈KPN
In this section we propose an algorithm whose regret with respect to the so-
cially optimal allocation is O(n
) for γ > 0 arbitrarily small. Clearly
this regret is sublinear and approaches linear as the number of users M in-
creases. This means that the time average of the sum of the utilities of the play-
ers converges to the socially optimal welfare. Let K = {k = (k1, k2, . . . , kN ) :
kj ≥ 0, ∀j ∈ N , k1 + k2 + . . . + kN = M } denote an allocation of M users
to N channels. Note that this allocation gives only the number of users on
each channel. It does not say anything about which user uses which chan-
nel. We assume that the socially optimal allocation is unique up to permu-
j=1 µjkjgj(kj) is unique. We also assume the
following stability condition of the socially optimal allocation. Let vj(kj ) =
j=1 kj vj(kj ) = k∗
if vj (k) − vj (k) ≤ ǫ, ∀k ∈ {1, 2, . . . , M }, ∀j ∈ N , for some ǫ > 0, where
vj : N → R is an arbitrary function. Let T i
j,k(t) be the number of times user
i used channel j and observed k users on it up to time t. We refer to the
tuple (j, k) as an arm. Let ni
j,k(t) be the time of the tth observation of user
i from arm (j, k). Let ui
j,k(t) be the sample mean of the rewards from arm
(j, k) seen by user i at the end of the tth play of arm (j, k) by user i, i.e.,
ui
j,k(t) = (hj,k(ni
j,k(t)))/t. Then the socially optimal allo-
j,k(t). The
pseudocode of the Randomized Learning Algorithm (RLA) is given in Fig. 2. At
time t RLA explores with probability 1/(t
M ) by randomly choosing one of
the channels and exploits with probability 1 − 1/(t
M ) by choosing a channel
which is occupied by a user in the estimated socially optimal allocation.
j,k(1)) + . . . + hj,k(ni
cation estimated by user i at time t is ki∗(t) = arg maxk∈KPN
1
j=1 kjui
1
2M − γ
2M − γ
The following will be useful in the proof of the main theorem of this section.
Lemma 2. Let Xi, i = 1, 2, . . . be a sequence of independent Bernoulli random
variables such that Xi has mean qi with 0 ≤ qi ≤ 1. Let ¯Xk = 1
i=1 Xi ,
¯qk = 1
i=1 qi. Then for any constant ǫ ≥ 0 and any integer n ≥ 0,
k Pk
kPk
Proof. The result follows from symmetry and [9].
P(cid:0) ¯Xn − ¯qn ≤ −ǫ(cid:1) ≤ e−2nǫ2
.
Lemma 3. For p > 0, p 6= 1
(n + 1)1−p − 1
1 − p
<
n
Xt=1
1
tp < 1 +
n1−p − 1
1 − p
Proof. See [8].
(5)
⊓⊔
(6)
⊓⊔
RLA (for user i)
1: Initialize: 0 < γ << 1, ui
sample σi(1) uniformly from N .
j,k(1) = 0, T i
j,k(1) = 0, ∀j ∈ N , k ∈ M, t = 1,
2: while t > 0 do
3:
play channel σi(t), observe l(t) the total number of players using channel
σi(t) and reward hσi(t),l(t)(t).
Set T i
Set T i
j,l(t) for (j, l) 6= (σi(t), l(t)).
σi(t),l(t)(t) + 1.
σi(t),l(t)(t + 1) = T i
j,l(t + 1) = T i
σi(t),l(t)(t + 1) =
j,l(t + 1) = ui
Set ui
Set ui
σi(t),l(t) (t)ui
T i
σi(t),l(t) (t)+hσi(t),l(t) (t)
T i
σi (t),l(t)(t+1)
.
j,l(t) for (j, l) 6= (σi(t), l(t)).
Set ki∗(t + 1) = arg maxk∈KPN
j=1 kjui
j,kj (t + 1).
Set θ∗i(t+1) to be the set of channels used by at least one user in k∗i(t+1).
10: Draw it randomly from Bernoulli distribution with P (it = 1) =
1
4:
5:
6:
7:
8:
9:
if σi(t) ∈ θ∗(t + 1) and l(t) = ki∗
j (t + 1) then
σi(t + 1) = σi(t)
σi(t+1) is selected uniformly at random from the channels in θ∗(t+1).
Draw σi(t + 1) uniformly at random from N .
t(1/2M )−γ/M
if it = 0 then
else
11:
12:
13:
14:
15:
16:
17:
18:
end if
19:
t = t + 1
20:
21: end while
end if
else
Fig. 2. pseudocode of RLA
Theorem 2. When all players use RLA the regret with respect to the socially
optimal allocation is O(n
) where γ can be arbitrarily small.
2M −1+2γ
2M
Proof. Let H(t) be the event that at time t there exists at least one user that
computed the socially optimal allocation incorrectly. Let ω be a sample path.
Then
I(ω ∈ H(t)) ≤
n
(n,M,N,M)
Xt=1
X(t,i,j,l)=(1,1,1,1)
≤
n
M
Xt=1
Xi=1
I(k∗i(t) 6= k∗)
I(ui
j,l(T i
j,l(t)) − vj(l) ≥ ǫ)
(n,M,N,M)
+
=
(n,M,N,M)
X(t,i,j,l)=(1,1,1,1)
X(t,i,j,l)=(1,1,1,1)
j,k(t) =q a ln t
I(cid:18)ui
I(cid:18)ui
j,k(t) . Then T i
T i
j,l(T i
j,l(t)) − vj(l) ≥ ǫ, T i
j,l(t) ≥
j,l(T i
j,l(t)) − vj(l) ≥ ǫ, T i
j,l(t) <
(7)
a ln t
ǫ2 (cid:19)
ǫ2 (cid:19)
j,k(t) = ǫi
a ln t
Let ǫi
fore,
I(cid:18)ui
I(cid:18)ui
j,l(T i
j,l(t)) − vj (l) ≥ ǫ, T i
j,l(t) ≥
j,l(T i
j,l(t)) − vj(l) ≥ ǫi
j,l(T i
j,l(t)) − vj (l) ≥ ǫ, T i
j,l(t) <
j,k(t) ≥ a ln t
ǫ2 ⇒ ǫ ≥ q a ln t
T i
j,k(t). There-
a ln t
ǫ2 (cid:19) ≤ I(cid:0)ui
ǫ2 (cid:19) ≤ I(cid:18)T i
a ln t
j,l(t) <
a ln t
ǫ2 (cid:19)
j,l(t)(cid:1)
Then, continuing from (7),
n
(n,M,N,M)
I(ω ∈ H(t))
Xt=1
X(t,i,j,l)=(1,1,1,1)(cid:18)I(cid:0)ui
≤
j,l(T i
j,l(t)) − vj (l) ≥ ǫi
j,l(t)(cid:1) + I(cid:18)T i
j,l(t) <
a ln t
ǫ2 (cid:19)(cid:19)(8)
Taking the expectation over (8),
(n,M,N,M)
E" n
Xt=1
X(t,i,j,l)=(1,1,1,1)
X(t,i,j,l)=(1,1,1,1)
I(ω ∈ H(t))#
P(cid:0)ui
P(cid:18)T i
(n,M,N,M)
j,l(T i
j,l(t)) − vj(l) ≥ ǫi
j,l(t)(cid:1)
j,l(t) <
a ln t
ǫ2 (cid:19) .
(9)
≤
+
We have
j,l(t)) − vj (l) ≥ ǫi
j,l(t)) − vj(l) ≥ ǫi
j,l(t))
− vj(l) ≥ ǫi
j,l(T i
j,l(T i
j,l(T i
T i
j,l(t)
2(T i
P (cid:0)ui
= P (cid:0)ui
= P Si
≤ 2 exp −
j,l(t))2
j,l(t))2(ǫi
T i
j,l(t)
j,l(T i
j,l(t)(cid:1)
j,l(t)(cid:1) + P (cid:0)ui
j,l(t)! + P Si
! = 2 exp −
j,l(t)) − vj(l) ≤ −ǫi
j,l(T i
T i
j,l(t)
j,l(t))
− vj (l) ≤ −ǫi
j,l(t)(cid:1)
j,l(t)!
2T i
j,l(t)a ln t
T i
j,l(t) ! =
2
t2a ,
(10)
where (10) follows from the Chernoff-Hoeffding inequality.
Now we will bound P(cid:16)T i
j,l(t) be the number of time
steps in which player i played channel j and observed l users on channel j in the
time steps where all players randomized up to time t. Then
ǫ2 (cid:17). Let T Ri
j,l(t) < a ln t
{ω : T i
j,l(t) <
a ln t
ǫ2 } ⊂ {ω : T Ri
j,l(t) <
a ln t
ǫ2 },
Thus
P (cid:18)T i
j,l(t) <
a ln t
ǫ2 (cid:19) ≤ P(cid:18)T Ri
j,l(t) <
(11)
(12)
a ln t
ǫ2 (cid:19) .
Now we define new Bernoulli random variables X i
j,l(s) = 1
if all players randomize at time s and player i selects channel j and observes l
players on it according to the random draw. X i
j,l(t) =
j,l(s) = 0 else. Then T Ri
j,l(s) as follows: X i
j,l(s) = 1) = ρspl where pl =
(M −1
l−1 )(M +N −l−2
)
N −2
(M +N −1
N −1 )
and ρs =
1
s(1/2)−γ Then
s=1
j,l(t) <
a ln t
ǫ2 (cid:19)
s=1 X i
1
j,l(s). P (X i
Pt
s(1/2)−γ . Let st =Pt
P (cid:18)T Ri
= P T Ri
≤ P T Ri
j,l(t)
t
j,l(t)
t
−
−
pkst
t
pkst
t
<
<
a ln t
tǫ2 −
pkst
t !
a ln t
tǫ2 −
pk(t + 1)(1/2)+γ − 1
t((1/2) + γ)
! ,
(13)
where (13) follows from Lemma 3. Let τ (M, N, ǫ, γ, γ′, a) be the time that for
all k ∈ {1, 2, . . . , M }.
pk(t + 1)(1/2)+γ − 1
t((1/2) + γ)
−
a ln t
tǫ2 ≥ t(1/2)+γ ′
,
(14)
where 0 < γ′ < γ. Then for all t ≥ τ (M, N, ǫ, γ, γ′, a) (14) will hold since RHS
increases faster than LHS. Thus we have for t ≥ τ (M, N, ǫ, γ, γ′, a)
P T Ri
≤ P T Ri
j,l(t)
t
j,l(t)
t
−
−
pkst
t
pkst
t
≤ e−2tt2γ ′
−1
= e−2t2γ ′
<
a ln t
tǫ2 −
pk(t + 1)(1/2)+γ − 1
t((1/2) + γ)
!
< t−(1/2)+γ ′!
1
t2 .
≤ e−2 ln t =
Let a = 1. Then continuing from (9) by substituting (10) and (15) we have
E" n
Xt=1
I(ω ∈ H(t))# ≤ M 2N τ (M, N, ǫ, γ, γ′, 1) + 3
1
t2! .
n
Xt=1
(15)
(16)
Thus we proved that the expected number of time steps in which there exists
at least one user that computed the socially optimal allocation incorrectly is
finite. Note that because RLA explores with probability
t1/2M −γ/M , the expected
number of time steps in which all the players are not randomizing up to time n
is
1
n
Xt=1 1 −(cid:18)1 −
1
t(1/2M)−γ/M(cid:19)M! ≤
n
Xt=1
M
t1/2M−γ/M = O(n
2M −1+2γ
2M
).
(17)
Note that players can choose γ arbitrarily small, increasing the finite regret due
to τ (M, N, ǫ, γ, γ′, 1). Thus if we are interested in the asymptotic performance
then γ > 0 can be arbitrarily small.
Now we do the worst case analysis. We classify the time steps into two. Good
time steps in which all the players know the socially optimal allocation correctly
and none of the players randomize excluding the randomizations done for settling
down to the socially optimal allocation. Bad time steps in which there exists a
player that does not know the socially optimal allocation correctly or there is
a player that randomizes excluding the randomizations done for settling down
to the socially optimal allocation. The number of Bad time steps in which there
exists a player that does not know the socially optimal allocation correctly is
finite while the number of time steps in which there is a player that randomizes
excluding the randomizations done for settling down to the socially optimal
). The worst case is when each bad step is followed by
allocation is O(n
a good step. Then from this good step the expected number of times to settle
2M −1+2γ
2M
down to the socially optimal allocation is (cid:18)1 −
1
(M +z∗
z∗
−1 )(cid:19) /(cid:18)
−1
1
(M +z∗
z∗
−1 )(cid:19) where
−1
z∗ is the number of channels which has at least one user in the socially optimal
allocation. Assuming in the worst case the sum of the utilities of the players is
0 when they are not playing the socially optimal allocation we have
R(n) ≤
1 −
1
(M +z∗
−1 )
−1
−1 ) M 2N τ (M, N, ǫ, γ, γ′, 1) + 3
−1
z∗
1
(M +z∗
z∗
= O(n
2M −1+2γ
2M
)
1
t2! + O(n
n
Xt=1
2M −1+2γ
2M
)!
⊓⊔
Note that we mentioned earlier, under a classical multi-armed bandit prob-
lem approach as cited before [3,4,5,15,16,21,22], a logarithmic regret O(log n) is
achievable. The fundamental difference between these studies and the problem
in the present paper is the following: Assume that at time t user i selects channel
j. This means that i selects to observe an arm from the set {(j, k) : k ∈ M} but
the arm assigned to i is selected from this set depending on the choices of other
players.
Also note that in RLA a user computes the socially optimal allocation accord-
ing to its estimates at each time step. This could pose significant computational
effort since integer programming is NP-hard in general. However, by exploiting
the stability condition on the socially optimal allocation a user may reduce the
number of computations; this is a subject of future research.
5 An Algorithm for Socially Optimal Allocation (Case 3)
In this section we assume that gj(n) is decreasing in n for all j ∈ N . For
simplicity we assume that the socially optimal allocation is unique up to the
permutations of σ∗. When this uniqueness assumption does not hold we need a
more complicated algorithm to achieve the socially optimal allocation. All users
use the Random Selection (RS) algorithm defined in Fig. 3. RS consists of two
phases. Phase 1 is the learning phase where the user randomizes to learn the
interference functions. Let Bj(t) be the set of distinct payoffs observed from
channel j up to time t. Then the payoffs in set Bj(t) can be ordered in a de-
creasing way with the associated indices {1, 2, . . . , Bj(t)}. Let O(Bj (t)) denote
this ordering. Since the IFs are decreasing, at the time Bj(t) = M , the user has
learned gj. At the time ∪N
j=1 Bj(t) = M N , the user has learned all IFs. Then,
the user computes A∗ and phase 2 of RS starts where the user randomizes to
converge to the socially optimal allocation.
Random Selection (RS)
1: Initialize: t = 1, b = 0, Bj(1) = ∅, ∀j ∈ N , sample σi(1) from the uniform
distribution on N
if hσi(t)(t) /∈ Bσi(t)(t) then
Bσi(t+1)(t + 1) ← O(Bσi(t)(t) ∪ hσi(t)(t))
b = b + 1
2: Phase 1
3: while b < M N do
4:
5:
6:
7:
8:
9:
end if
Sample σi(t + 1) from the uniform distribution on N
t = t + 1
10: end while
11: find the socially optimal allocation σ∗
12: Phase 2
13: while b ≥ M N do
if hσi(t)(t) < v∗
14:
15:
16:
17:
end if
18:
t = t + 1
19:
20: end while
Sample σi(t + 1) from the uniform distribution on N
σi (t) then
else
σi(t + 1) = σi(t)
Fig. 3. pseudocode of RS
Theorem 3. Under the assumptions of (C3) if all players use RS algorithm to
choose their actions, then the expected time to converge to the socially optimal
allocation is finite.
Proof. Let TOP T denote the time the socially optimal allocation is achieved, TL
be the time when all users learn all the IFs, TF be the time it takes to reach the
socially optimal allocation after all users learn all the IFs. Then TOP T = TL +TF
and E[TOP T ] = E[TL] + E[TF ]. We will bound E[TL] and E[TF ]. Let Ti be the
first time that i users have learned the IFs. Let τi = Ti − Ti−1, i = 1, 2, . . . , M
and T0 = 0. Then TL = τ1 + . . . + τM . Define a Markov chain over all N M pos-
sible configurations of M users over N channels based on the randomization of
the algorithm. This Markov chain has a time dependent stochastic matrix which
changes at times T1, T2, . . . , TM . Let PT0 , PT1 , . . . , PTM denote the stochastic ma-
trices after the times T0, T1, . . . , TM respectively. This Markov chain is irreducible
at all times up to TM and is reducible with absorbing states corresponding to
the socially optimal allocations after TM . Let T1, T2, . . . TM be the times that
all configurations are visited when the Markov chain has stochastic matrices
PT0 , PT1 , . . . , PTM −1 respectively. Then because of irreducibility and finite states
E[ Ti] < z1, i = 1, . . . , M for some constant z1 > 0 . Since τi ≤ Ti, i = 1, . . . , M
a.s. we have E[TL] < M z1. For the Markov chain with stochastic matrix PTM
all the configurations that do not correspond to the socially optimal allocation
are transient states. Since starting from any transient state the mean time to
⊓⊔
absorption is finite E[TF ] < z2, for some constant z2 > 0.
6 Conclusion
In this paper we studied the decentralized multiuser resource allocation problem
with various levels of communication and cooperation between the users. Under
three different scenarios we proposed three algorithms with reasonable perfor-
mance. Our future reserach will include characterization of achievable perfor-
mance regions for these scenarios. For example, in case 2 we are interested in
finding an optimal algorithm and a lower bound on the performance.
References
1. Agrawal, R.: Sample Mean Based Index Policies with O(log(n)) Regret for the
Multi-armed Bandit Problem. Advances in Applied Probability 27(4), 1054–1078
(December 1995)
2. Ahmad, S., Tekin, C., Liu, M., Southwell, R., Huang, J.: Spectrum Sharing as
Spatial Congestion Games. http://arxiv.org/abs/1011.5384 (2010)
3. Anandkumar, A., Michael, N., Tang, A.: Opportunistic Spectrum Access with Mul-
tiple Players: Learning under Competition. In: Proc. of IEEE INFOCOM (March
2010)
4. Anantharam, V., Varaiya, P., Walrand, J.: Asymptotically Efficient Allocation
Rules for the Multiarmed Bandit Problem with Multiple Plays-Part I: IID Re-
wards. IEEE Trans. Automat. Contr. pp. 968–975 (November 1987)
5. Anantharam, V., Varaiya, P., Walrand, J.: Asymptotically Efficient Allocation
Rules for the Multiarmed Bandit Problem with Multiple Plays-Part II: Marko-
vian Rewards. IEEE Trans. Automat. Contr. pp. 977–982 (November 1987)
6. Auer, P., Cesa-Bianchi, N., Fischer, P.: Finite-time Analysis of the Multiarmed
Bandit Problem. Machine Learning 47, 235–256 (2002)
7. Auer, P., Cesa-Bianchi, N., Freund, Y., Schapire, R.: The Nonstochastic Multi-
armed Bandit Problem. SIAM Journal on Computing 32, 48–77 (2002)
8. Chlebus, E.: An Approximate Formula for a Partial Sum of the Divergent p-series.
Applied Mathematics Letters 22, 732–737 (2009)
9. D.W. Turner, D.M. Young, J.S.: A Kolmogorov Inequality for the Sum of Indepen-
dent Bernoulli Random Variables with Unequal Means. Statistics and Probability
Letters 23, 243–245 (1995)
10. Freund, Y., Schapire, R.: Adaptive Game Playing Using Multiplicative Weights.
Games and Economic Behaviour 29, 79–103 (1999)
11. Gai, Y., Krishnamachari, B., Jain, R.: Learning Multiuser Channel Allocations in
Cognitive Radio Networks: a Combinatorial Multi-armed Bandit Formulation. In:
IEEE Symp. on Dynamic Spectrum Access Networks (DySPAN) (April 2010)
12. Kakhbod, A., Teneketzis, D.: Power Allocation and Spectrum Sharing in Cognitive
Radio Networks With Strategic Users. In: 49th IEEE Conference on Decision and
Control (CDC) (December 2010)
13. Kasbekar, G., Proutiere, A.: Opportunustic Medium Access in Multi-channel Wire-
less Systems: A Learning Approach. In: Proceedings of the 48th Annual Allerton
Conference on Communication, Control, and Computation (September 2010)
14. Kleinberg, R., Piliouras, G., Tardos, E.: Multiplicative Updates Outperform
Generic No-Regret Learning in Congestion Games. In: Annual ACM Symposium
on Theory of Computing (STOC) (2009)
15. Lai, T., Robbins, H.: Asymptotically Efficient Adaptive Allocation Rules. Advances
in Applied Mathematics 6, 4–22 (1985)
16. Liu, K., Zhao, Q.: Distributed Learning in Multi-Armed Bandit with Multiple
Players. IEEE Transactions on Signal Processing 58(11), 5667–5681 (November
2010)
17. Monderer, D., Shapley, L.S.: Potential Games. Games and Economic Behavior
14(1), 124–143 (1996)
18. Rosenthal, R.: A Class of Games Possessing Pure-strategy Nash Equilibria. Inter-
national Journal of Game Theory 2, 65–67 (1973)
19. Sandholm, W.H.: Population Games and Evolutionary Dynamics. Manuscript
(2008)
20. Smith, J.M.: Evolution and the Theory of Games. Cambridge University Press
(1982)
21. Tekin, C., Liu, M.: Online Algorithms for the Multi-armed Bandit Problem with
Markovian Rewards. In: Proceedings of the 48th Annual Allerton Conference on
Communication, Control, and Computation (September 2010)
22. Tekin, C., Liu, M.: Online Learning in Opportunistic Spectrum Access: A Restless
Bandit Approach. In: 30th IEEE International Conference on Computer Commu-
nications (INFOCOM) (April 2011)
|
cs/0207022 | 1 | 0207 | 2002-07-07T11:26:42 | What is a Joint Goal? Games with Beliefs and Defeasible Desires | [
"cs.MA",
"cs.GT"
] | In this paper we introduce a qualitative decision and game theory based on belief (B) and desire (D) rules. We show that a group of agents acts as if it is maximizing achieved joint goals. | cs.MA | cs | Wat is a Joint Goal?
Games with Beliefs and Defeasible Desires
2
0
0
2
l
u
J
7
]
A
M
.
s
c
[
1
v
2
2
0
7
0
2
0
/
s
c
:
v
i
X
r
a
Mehdi Dastani
Leendert van der Torre
Institute of Information and Computing Sciences
Department of Artificial Intelligence
Utrecht University
email: [email protected]
Vrije Universiteit Amsterdam
email: [email protected]
Abstract
In this paper we introduce a qualitative de-
cision and game theory based on belief (B)
and desire (D) rules. We show that a group
of agents acts as if it is maximizing achieved
joint goals.
2. The decision making of a group or society of
agents is described in terms of concepts gener-
alized from those used for individual agents, such
as joint goals, joint intentions, joint commitments,
etc. Moreover, also new concepts are introduced
at this social level, such as norms (a central con-
cept in most social theories).
1
Introduction
One of the main problems in agent theory is the dis-
tinction between formal theories and tools developed
for individual autonomous agents, and theories and
tools developed for multi agent systems.
In the so-
cial sciences, this distinction is called the micro-macro
dichotomy. The prototypical example is the distinc-
tion between classical decision theory based on the
expected utility paradigm (usually identified with the
work of Neumann and Morgenstern [15] and Savage
[11]) and classical game theory (such as the work
of Nash and more recently the work of Axelrod).
Whereas classical decision theory is a kind of opti-
mization problem (maximizing the agent's expected
utility), classical game theory is a kind of equilibria
analysis.
There are several approaches in practical reasoning
(within philosophy), cognitive science and artificial in-
telligence to bring the micro and macro description
together. The basic idea is two-fold:
It is still an open problem how the micro-macro di-
chotomy of classical decision and game theory is re-
lated to the micro-macro dichotomy of these alter-
native theories. Has or can the micro and macro
level be brought together by replacing classical the-
ories by alternative theories? Before this question
can be answered, the relation between the classical
and alternative theories has to be clarified. Doyle
and Thomason [6] argue that classical decision theory
should be reunited with alternative decision theories
in so-called qualitative decision theory (QDT), which
studies qualitative versions of classical decision the-
ory, hybrid combinations of quantitative and qualita-
tive approaches to decision making, and decision mak-
ing in the context of artificial intelligence applications
such as planning, learning and collaboration. Quali-
tative decision theories have been developed based on
beliefs (probabilities) and desires (utilities) using for-
mal tools such as modal logic [2] and on utility func-
tions and knowledge [7]. More recently these beliefs-
desires models have been extended with intentions or
BDI models [5, 10].
1. The decision making of individual autonomous
agents is described in terms of other concepts than
maximizing utility. For example, since the early
40s there is a distinction between classical decision
theory and artificial intelligence based on utility
aspiration levels and goal based planning (as pi-
oneered by Simon [12]). Cognitive theories are
typically based on vague concepts from folk psy-
chology like beliefs, desires and intentions.
In this paper we introduce a rule based qualitative de-
cision and game theory, based on belief (B) and desire
(D) rules. We call an individual autonomous agent
which minimizes its unreached desires a BD rational
agent. We define goals as a set of formulas which can
be derived by beliefs and desires in a certain way, such
that BD rational agents act as if they maximize the
set of achieved goals, and agents maximizing their sets
of achieved goals are BD rational. Moreover, groups of
agents which end up in equilibria act as if they maxi-
mize joint goals.
Like classical decision theory but in contrast to sev-
eral proposals in the BDI approach [5, 10], the theory
does not incorporate decision processes, temporal rea-
soning, and scheduling.
The layout of this paper is as follows. Section 2 de-
velops a qualitative logic of decision. This logic tells
us what the optimal decision is, but it does not tell us
how to find this optimal decision. Section 3 considers
the AI solution to this problem [12, 8]: break down the
decision problem into goal generation and goal based
decisions.
1.1 QDT and NMR
Qualitative decision theory is
related to non-
monotonic logic ("qualitative") and to reasoning about
uncertainty ("decision theory" formalizes decision
making under uncertainty). However, they are based
on different disciplines. According to a distinction
made by Aristotle, non-monotonic reasoning and rea-
soning about uncertainty formalize theoretical (or con-
clusion oriented) reasoning, whereas qualitative deci-
sion theory formalizes practical (or action oriented)
reasoning.
Thomason [13] observes that classical decision theory
neglects the issue of 'correct inference', and that the
absence of a logical theory of practical reasoning is
largely due to the unavailability of appropriate infer-
ence mechanisms. To handle even the simplest cases of
practical reasoning, it is essential to have a reasoning
mechanism that allows for practical conclusions that
are non-monotonic in the agent's beliefs.
Where classical decision theory is based on probabil-
ities and utilities, qualitative decision theory is based
on beliefs and desires.
In a modal approach, where
the beliefs and desires are represented by modalities
B and D respectively, and an action operator by the
modal operator Do, we may have for example that:
B(thirsty), D(drink) ⊢ Do(go − to − pub)
2 A qualitative decision and game
theory
The qualitative decision and game theory introduced
in this section is based on sets of belief and desire rules.
We define an agent system specification, we show how
we can derive a game specification from it, and we give
some familiar notions from game theory such as Pareto
efficient decisions (choosing an optimal decision) and
Nash equilibria. First we consider the logic of rules we
adopt.
2.1 Logic of rules
The starting point of any theory of decision is a dis-
tinction between choices made by the decision maker
and choices imposed on it by its environment. For ex-
ample, a software upgrade agent (decision maker) may
have the choice to upgrade a computer system at a
particular time of the day. The software company (en-
vironment) may in turn allow/disallow such a upgrade
at a particular time. Let S = {α1, . . . , αn} be the so-
ciety or set of agents, then we therefore assume n dis-
joint sets of propositional atoms: A = A1 ∪ . . . ∪ An =
{a, b, c, . . .} (agents' decision variables [7] or control-
lable propositions [2]) and W = {p, q, r, . . .} (the world
parameters or uncontrollable propositions). In the se-
quel we consider each decision maker as entities con-
sisting of defeasible rules. Such a decision maker gen-
erates its decisions by applying subsets of defeasible
rules to its input. This results in the so-called condi-
tional mental attitudes [4]. Before we proceed some
notations will be introduced.
• LAi, LW and LAW for the propositional languages
built up from these atoms in the usual way, and
variables x, y, . . . to stand for any sentences of
these languages.
• CnAi , CnW and CnAW for the consequence sets,
and =Ai , =W and =AW for satisfiability, in any
of these propositional logics.
• x ⇒ y for an ordered pair of propositional sen-
tences called a rule.
B(thirsty), D(drink), B(pub−closed) ⊢ Do(go−to−shop)
• ER(T ) for the R extension of T , as defined in
Definition 1 below.
A drawback of such a modal logic approach is that on
the one hand modal logic is notorious for its problems
with formalizing conditionals or rules, and practical
reasoning on the other hand is usually seen as a kind
of rule based reasoning. In this paper we therefore do
not use modal logic but we use a rule based formalism.
In our framework the generation of decisions are for-
malized based on the notion of extension. In particu-
lar, the decision of an agent, which is specified by a set
of defeasible rules R and has the input T , is the exten-
sion calculated based on R and T . This is formalized
in the following definition.
Definition 1 (Extension) Let R ⊆ LAW × LAW be
a set of rules and T ⊆ LAW be a set of sentences. The
consequents of the T -applicable rules are:
R(T ) = {y x ⇒ y ∈ R, x ∈ T }
and the R extension of T is the set of the consequents
of the iteratively T -applicable rules:
ER(T ) = ∩T ⊆X,R(CnAW (X))⊆X X
We give some properties of the R extension of T in
Definition 1. First note that ER(T ) is not closed under
logical consequence. The following proposition shows
that ER(T ) is the smallest superset of T closed under
the rules R interpreted as inference rules.
Proposition 1 Let
• E0
• Ei
R(T ) = T
R(T ) = Ei−1
R (T )∪R(CnAW (Ei−1
R (T ))) for i > 0
We have ER(T ) = ∪∞
0 Ei
R(T ).
Usually, an decision making agent is required to pre-
serve its decisions under the growth of inputs. The
following proposition shows that ER(T ) is monotonic.
Proposition 2 We have R(T ) ⊆ R(T ∪ T ′) and
ER(T ) ⊆ ER(T ∪ T ′).
Monotonicity is illustrated by the following example.
Example 1 Let R = {⊤ ⇒ p, a ⇒ ¬p} and T =
{a}, where ⊤ stands for any tautology like p ∨ ¬p. We
have ER(∅) = {p} and ER(T ) = {a, p, ¬p}, i.e. the R
extension of T is inconsistent.
Of course, allowing inconsistent decisions may not be
intuitive. We are here concerned about possible de-
cisions rather than reasonable or feasible decisions.
Later we will define reasonable or feasible decisions
by excluding inconsistent decisions.
2.2 Agent system specification
An agent system specification given in Definition 2
contains a set of agents and for each agent a descrip-
tion of its decision problem. The agent's decision prob-
lem is defined in terms of its beliefs and desires, which
are considered as defeasible belief and desire rules, a
priority ordering on the desire rules, as well as a set
of facts and an initial decision (or prior intentions).
We assume that agents are autonomous, in the sense
that there are no priorities between desires of distinct
agents.
a
is
agent
system specification
Definition 2 (Agent system specification)
An
tuple
AS = hS, F, B, D, ≥, δ0i that contains a set of
agents S, and for each agent i a finite set of facts
Fi ⊆ LW (F = Sn
i=1 Fi), a finite set of belief rules
Bi ⊆ LAW × LW (B = Sn
i=1 Bi), a finite set of
desire rules Di ⊆ LAW × LAW (D = Sn
i=1 Di), a
relation ≥i⊆ Di × Di (≥= Sn
i=1 ≥i) which is a total
ordering (i.e. reflexive, transitive, and antisymmetric
and for any two elements d1 and d2 in Di, either
d1 ≥i d2 or d2 ≥i d1), and a finite initial decision
i ⊆ LA (δ0 = Sn
δ0
i ). For an agent i ∈ S we write
x ⇒i y for one of its rules.
i=1 δ0
A belief rule 'the agent αi believes y in context x' is
an ordered pair x ⇒i y with x ∈ LAW and y ∈ LW ,
and a desire rule 'the agent desires y in context x' is
an ordered pair x ⇒i y with x ∈ LAW and y ∈ LAW .
It implies that the agent's beliefs are about the world
(x ⇒i p), and not about the agent's decisions. These
beliefs can be about the effects of decisions made by
the agent (a ⇒i p) as well as beliefs about the effects of
parameters set by the world (p ⇒i q). Moreover, the
agent's desires can be about the world (x ⇒i p, desire-
to-be), but also about the agent's decisions (x ⇒i a,
desire-to-do). These desires can be triggered by pa-
rameters set by the world (p ⇒i y) as well as by deci-
sions made by the agent (a ⇒i y). Modelling mental
attitudes such as beliefs and desires in terms of defea-
sible rules results in what might be called conditional
mental attitudes [4].
2.3 Agent Decisions
The belief rules are used to determine the expected
consequences of a decision, where a decision δ is any
subset of LA that contains the initial decision δ0. The
set of expected consequences of this decision δ is the
belief extension of F ∪ δ. Moreover, we consider a
feasible decision as a decision that does not imply a
contradiction.
=
Definition 3 (Decisions) Let
h{α1, . . . , αn}, F, B, D, ≥, δ0i
system
specification. An AS decision profile δ for agents α1,
. . . , αn is δ = hδ1, . . . δni where δi is a decision of
agent αi such that
be an agent
AS
δ0
i ⊆ δi ⊆ LAi
for i = 1 . . . n
A feasible decision for agent i is δi such that
EBi (Fi ∪ δi) is consistent
A feasible decision profile is a decision profile such that
EB(F ∪ δ) is consistent
where we write EB(F ∪ δ) for Sn
i=1 EBi (Fi ∪ δi).
The following example illustrates the decisions of a
single agent.
Example 2 Let A1 = {a, b, c, d, e}, W = {p, q} and
AS = h{α1}, F, B, D, ≥, δ0i with F1 = {¬p}, B1 =
{c ⇒ q, d ⇒ q, e ⇒ ¬q}, D1 = {⊤ ⇒ a, ⊤ ⇒ b, b ⇒
p, ⊤ ⇒ q, d ⇒ q}, ≥1= {b ⇒ p > ⊤ ⇒ b}, and
δ0
1 = {a}. The initial decision δ0
1 reflects that the agent
has already decided in an earlier stage to reach the
desire ⊤ ⇒ a. Note that the consequents of all B1
rules are sentences of LW , whereas the antecedents of
the B1 rules as well as the antecedents and consequents
of the D1 rules are sentences of LAW . We have due
to the definition of ER(S):
EB(F ∪ {a}) = {¬p, a}
EB(F ∪ {a, b}) = {¬p, a, b}
EB(F ∪ {a, c}) = {¬p, a, c, q}
EB(F ∪ {a, d}) = {¬p, a, d, q}
EB(F ∪ {a, e}) = {¬p, a, e, ¬q}
. . .
EB(F ∪ {a, d, e}) = {¬p, a, d, e, q, ¬q}
. . .
Therefore {a, d, e} is not a feasible AS decision profile,
because its belief extension is inconsistent. Continued
in Example 4.
The following example illustrates that the set of feasi-
ble decisions of an agent may depend on the decisions
of other agents.
Example 3 Let A1 = {a}, A2 = {b}, W = {p} and
AS = h{α1, α2}, F, B, D, ≥, δ0i with F1 = F2 = ∅,
B1 = {a ⇒ p}, B2 = {b ⇒ ¬p}, D1 = {⊤ ⇒ p}, D2 =
{⊤ ⇒ ¬p}, ≥ is the identity relation, and δ0
2 = ∅.
We have that h∅, {b}i is a feasible decision profile, but
h{a}, {b}i is not. If δ1 = ∅, then agent α2 can decide
δ2 = {b}. However, if δ1 = {a}, then agent α2 cannot
decide so, i.e. δ2 6= {b}.
1 = δ0
2.4 Agent preferences
In this section we introduce a way to compare deci-
sions. We compare decisions by comparing sets of de-
sire rules that are not reached by the decisions. Since
only ordering on individual desire rules, and not or-
dering on sets of desire rules, are given, we first lift
the ordering on individual desire rules to an ordering
on sets of desire rules.
Definition 4 Let AS = hS, F, B, D, ≥, δ0i be an
agent system specification, D′
two subsets of Di,
i the set of Di elements which are not D′
and Di \ D′
i
elements. We have D′
i∃d′ ∈
if ∀d′′ ∈ D′′
i (cid:23) D′′
i
i \ D′
i, D′′
i
D′
D′
D′
i \ D′′
i (cid:23) D′′
i (cid:23) D′′
i such that d′ > d′′. We write D′
i, and we write D′
i and D′′
i
i and D′′
i (cid:23) D′
i.
6(cid:23) D′
i ≻ D′′
i
i ≃ D′′
i
if
if
The following propositions show that the priority re-
lation (cid:23) is reflexive, anti-symmetric, and transitive.
Proposition 3 Let D1 6⊆ D3 and D3 6⊆ D1. For fi-
nite sets, the relation (cid:23) is reflexive ( ∀D D (cid:23) D),
anti-symmetric ∀D1, D2 (D1 (cid:23) D2 ∧ D2 (cid:23) D1) →
D1 = D2, and transitive, i.e. D1 (cid:23) D2 and D2 (cid:23) D3
implies D1 (cid:23) D3.
The desire rules are used to compare the decisions.
The comparison is based on the set of unreached de-
sires and not on the set of violated or reached desires.
A desire x ⇒ y is unreached by a decision if the ex-
pected consequences of this decision imply x but not
y. The desire rule is violated or reached if these conse-
quences imply respectively x∧¬y or x∧y, respectively.
Definition 5 (Comparing decisions) Let
AS = hS, F, B, D, ≥, δ0i be an agent system specifica-
tion and δ be a AS decision. The unreached desires
of decision δ for agent αi are:
Ui(δ) =
{x ⇒ y ∈ Di EB(F ∪ δ) = x and EB(F ∪ δ) 6= y}
Decision δ is at least as good as decision δ′
agent αi, written as δ ≥U
i δ′, iff
for
Ui(δ′) (cid:23) Ui(δ)
Decision δ dominates decision δ′ for agent αi, written
as δ >U
i δ′, iff
δ ≥U
i δ′ and δ′ 6≥U
i δ
The following continuation of Example 2 illustrates the
comparison of decisions.
Example 4 (Continued) We have:
U ({a}) = {⊤ ⇒ b, ⊤ ⇒ q},
U ({a, b}) = {b ⇒ p, ⊤ ⇒ q},
U ({a, c}) = {⊤ ⇒ q},
U ({a, d}) = {⊤ ⇒ b, d ⇒ q},
U ({a, e}) = {⊤ ⇒ b, ⊤ ⇒ q},
U ({a, b, c}) = {b ⇒ p}.
. . .
We thus have for example that the decision {a, c} dom-
inates the initial decision {a}, i.e. {a, c} >U {a}.
There are two decisions for which their set of un-
reached contains only one desire. Due to the priority
relation, we have that {a, c} >U {a, b, c}.
2.5 Agent games
In this subsection, we consider agents interactions
based on agent system specifications, their correspond-
ing agent decisions, and the ordering on the decisions
as explained in previous subsections. Game theory is
the usual tool to model the interaction between self-
interested agents. Agents select optimal decisions un-
der the assumption that other agents do likewise. This
makes the definition of an optimal decision circular,
and game theory therefore restricts its attention to
equilibria. For example, a decision is a Nash equilib-
rium if no agent can reach a better (local) decision by
changing its own decision. The most used concepts
from game theory are Pareto efficient decisions, dom-
inant decisions and Nash decisions. We first repeat
some standard notations from game theory [1, 9].
As mentioned, we use δi to denote a decision of agent
αi and δ = hδ1, . . . , δni to denote a decision profile con-
taining one decision for each agent. δ−i is the decision
profile of all agents except the decision of agent αi.
(δ−i, δ′
i) denotes a decision profile which is the same
as δ except that the decision of agent i from δ is re-
placed with the decision of agent i from δ′. δ′
i δi
denotes that decision δ′
i is better than δi according to
his preferences >U
i δi if better or equal. ∆
is the set of all decision profiles for agents α1, . . . , αn,
∆f ⊆ ∆ is the set of feasible decision profiles, and ∆i
is the set of possible decisions for agent αi.
i and δ′
i >U
i ≥U
Definition 6 (Game specification) Let
AS = hS = {α1, . . . , αn}, F, B, D, ≥, δ0i be spec-
ification of agent system in S, Ai be the set of AS
feasible decisions of agent αi according to Definition
3, ∆f = A1 × . . . × An, and ≥U
i be the AS preference
relation of agent αi defined on its feasible decisions
according to definition 5. Then, the game specification
of AS is the tuple hS, ∆f , (≥U
i )i.
We now consider different types of decision profiles
which are similar to types of strategy profiles from
game theory.
Definition 7 A PS decision profile δ = hδ1, . . . , δni ∈
∆f is:
Pareto decision if there is no δ′ = hδ′
1, . . . , δ′
ni ∈ ∆f
for which δ′
i >U
i δi for all agents αi.
strongly Pareto decision if
hδ′
1, . . . , δ′
αi and δ′
ni ∈ ∆f for which δ′
j >U δj for some agents αj.
there is no δ′ =
i ≥U
i δi for all agents
dominant decision if for all δ′ ∈ ∆f and for every
agent i it holds: (δ′
i) i.e. a deci-
sion is dominant if it yields a better payoff than
−i, δi) ≥U
−i, δ′
i (δ′
any other decisions regardless of what the other
agents decide.
Nash decision if
all
agents
i
it
holds:
(δ−i, δi) ≥U
i) for all δ′
i ∈ ∆i
f
for
i (δ−i, δ′
It is a well known fact that Pareto decisions exist (for
finite games), whereas dominant decisions do not have
to exist. The latter is illustrated by the following ex-
ample.
two
1 = δ0
agents,
2 = ∅.
be
and α2
Example 5 Let α1
F1 = F2 = ∅, and initial decisions δ0
They have the following beliefs en desires:
Bα1 = {a ⇒ p, ¬a ⇒ ¬p}
Dα1 = {⊤ ⇒ p, ⊤ ⇒ q}
≥α1= ⊤ ⇒ p > ⊤ ⇒ q > ⊤ ⇒ ¬q > ⊤ ⇒ ¬p
Bα2 = {b ⇒ q, ¬b ⇒ ¬q}
Dα2 = {⊤ ⇒ ¬p, ⊤ ⇒ ¬q}
≥α2= ⊤ ⇒ ¬p > ⊤ ⇒ ¬q > ⊤ ⇒ q > ⊤ ⇒ p
Let ∆f be feasible decision profiles, EB be the
outcomes of the decisions, and U (δi) be the set of
unreached desires for agent αi.
∆
EB
Uδ1
Uδ2
ha, bi
{p, q}
∅
ha, ¬bi
h¬a, bi
h¬a, ¬bi {¬p, ¬q}
{p, −q}
{¬p, q}
{⊤ ⇒ q}
{⊤ ⇒ p}
{⊤ ⇒ p, ⊤ ⇒ q}
{⊤ ⇒ ¬p,
⊤ ⇒ ¬q}
{⊤ ⇒ ¬p}
{⊤ ⇒ ¬q}
∅
According to definition 5, for A1 :
U (ha, bi) > U (ha, ¬bi) > U (h¬a, bi) > U (h¬a, ¬bi)
and for A2 :
U (h¬a, ¬bi) > U (h¬a, bi) > U (ha, ¬bi) > U (ha, bi).
None of these decision profiles are dominant decisions,
i.e. the agents specifications has no dominant solution
with respect to their unreached desires.
The following example illustrates a typical cooperation
game.
Example 6 B1 = {a ⇒ p , b ⇒ ¬p ∧ q}, D1 = {⊤ ⇒
p ∧ q} B2 = {c ⇒ q , d ⇒ p ∧ ¬q} D2 = {⊤ ⇒ p ∧ q}.
The agents have a common goal p ∧ q, which they can
only reach by cooperation.
The following example illustrates a qualitative version
of the notorious prisoner's dilemma, where the self-
ish behavior of individual autonomous agents leads to
global bad decisions.
Example 7 Let A1 = {a} (α1 cooperates), A2 = {b}
(α2 cooperates), and AS be an agent system specifica-
tion with D1 = {⊤ ⇒ ¬a ∧ b, ⊤ ⇒ b, ⊤ ⇒ ¬(a ∧ ¬b)},
D2 = {⊤ ⇒ a ∧ ¬b, ⊤ ⇒ a, ⊤ ⇒ ¬(¬a ∧ b)}. The only
Nash decision is {¬a, ¬b}, whereas both agents would
prefer {a, b}.
Starting from an agent system specification, we can
derive the game specification and in this game specifi-
cation we can use standard techniques to for example
find the Pareto decisions. However, the problem with
this approach is that the translation from an agent
system specification to a game specification is compu-
tationally expensive. For example, a compact agent
representation with only a few belief and desire rules
may lead to a huge set of decisions if the number of
decision variables is high.
The main challenge of qualitative game theory is there-
fore whether we can bypass the translation to game
specification, and define properties directly on the
agent system specification. For example, are there
particular properties of agent system specification for
which we can prove that there always exists a domi-
nant decision for its corresponding derived game spec-
ification? A simple example is an agent system spec-
ification in which each agent has the same belief and
desire rules.
In this paper we do not further pursue these issues,
but we turn to our focus of interest: joint goals.
3 Joint goals
In this section we ask ourselves the question whether
and how we can interpret a decision profile or equilib-
rium as goal-based or goal-oriented behavior. We first
define decision rules and sets of decision profiles closed
under indistinguishable decision profiles.
3.1 Decision rule
A decision rule maps a agent system specification to a
set of possible decision profiles.
Definition 8 A decision rule is a function from agent
system specifications to sets of feasible decision pro-
files.
Decision theory prescribes a decision maker to select
the optimal or best decision, which can be defined as a
decision that is not dominated. Is there an analogous
prescription for societies of agents? A set of cooperat-
ing agents has to select an optimal or Pareto decision.
We call such cooperating agents a BD rational society
of cooperating agents.
Definition 9 A BD rational society of cooperating
agents is a set of agents, defined by an agent system
specification AS, that selects a Pareto AS decision.
In this paper we consider decision rules based on un-
reached desires. We therefore assume that a decision
rule cannot distinguish between decision profiles δ1
and δ2 such that δ1 ∼U δ2. We say that two decision
profiles δ and δ′ are indistinguishable if U (δ) = U (δ′),
and we call a set of decisions U-closed if the set is
closed under indistinguishable decision profiles.
Definition 10 Let AS = hS, F, B, D, ≥, δ0i be an
agent system specification and ∆ a set of AS decision
profiles. ∆ is U-closed if δ ∈ ∆ implies δ′ ∈ ∆ for all
AS decision profiles δ′ such that U (δ) = U (δ′).
An example of a decision rule is the function that maps
agent system specifications to Pareto decision profiles
(a BD rational decision rule). Another example is a
function that maps agent system specifications to Nash
equilibria if they exist, otherwise to Pareto decisions.
3.2 Goals
In this section we show that every society of agents
can be understood as planning for joint goals, whether
the decision is reached by cooperation or is given by
a (e.g. Nash) equilibrium. We define not only goals
which must be reached, called positive goals, but we
add negative goals. Negative goals are defined in the
following definition as states the agent has to avoid.
They function as constraints on the search process of
goal-based decisions.
Definition 11 (Goal-based decision) Let AS =
hS, F, B, D, ≥, δ0i be an agent system specification,
and the so-called positive joint goal set G+ and the
negative joint goal set G− be subsets of LAW . A deci-
sion δ is a hG+, G−i decision if EB(F ∪ δ) =AW G+
and for each g ∈ G− we have EB(F ∪ δ) 6=AW g.
Joint goals are defined with respect to a set of decision
profiles. The definition of ∆ joint goal set encodes
a decision profile together with its indistinguishable
decision profiles as a positive and negative goal set.
AS
=
Definition 12 (∆ goal set) Let
hS, F, B, D, ≥, δ0i
system specifica-
be an agent
tion and ∆ an U-closed set of feasible decisions. The
two sets of formulas hG+, G−i ⊆ (LAW , LAW ) is a ∆
joint goal set of AS if there is an AS decision δ ∈ ∆
such that
G+ = {y x ⇒ y ∈ D, EB(F ∪ δ) =AW x ∧ y}
G− = {x x ⇒ y ∈ D, EB(F ∪ d) 6=AW x}
hG+, G−i ⊆ (LAW , LAW ) is a feasible joint goal set of
AS if there is an U-closed set of feasible decisions ∆
such that hG+, G−i ⊆ (LAW , LAW ) is a ∆ joint goal
set of AS. hG+, G−i ⊆ (LAW , LAW ) is a joint goal set
of AS if there is a set D′ ⊆ D such that
G+ = {y x ⇒ y ∈ D′}
G− = {x x ⇒ y ∈ D′}
The first part of the representation theorem follows
directly from the definitions.
Proposition 4 Let AS = hS, F, B, D, ≥, δ0i be an
agent system specification and ∆ be an U-closed set of
feasible decision profiles. For a decision profile δ ∈ ∆
of AS there is a ∆ joint goal set hG+, G−i of AS such
that δ is a hG+, G−i decision.
Proof. Follows directly from the definitions.
calculate all joint goals
filter feasible joint goals
for each feasible joint goal set
find goal-based decision profiles
order these decision profiles
In other words, the goal-based representation suggests
to calculate the joint goals first. However, the problem
to calculate these joint goals is still computationally
hard. There are two ways to proceed:
• Define heuristics for the optimization problem;
• Find a fragment of the logic, such that the opti-
mization becomes easier.
For example, consider the following procedure to find
(positive) goals:
G ⊆ EB∪D(F ∪ δ0)
Proposition 5 Let AS = hS, F, B, D, ≥, δ0i be an
agent system specification and ∆ an U-closed set of
feasible decision profiles.
For a ∆ joint goal set
hG+, G−i of AS, a hG+, G−i decision is a ∆ decision.
Proof. Follows from U-closed property.
This procedure is not complete, because it does not
take effects of actions into account. Thus, in the gen-
eral case it can be used as a heuristic. Moreover, it is
complete for the fragment in which the belief rules do
not contain effects of actions, i.e. B ⊆ S × LW × LW .1
The representation theorem is a combination of Propo-
sition 4 and 5.
4 Concluding remarks
Theorem 1 Let AS = hS, F, B, D, ≥, δ0i be an agent
system specification and ∆ an U-closed set of feasible
decision profiles. A decision profile δ is in ∆ if and
only if there is a ∆ goal set hG+, G−i of AS such that
δ is a hG+, G−i decision profile.
The second theorem follows from the first one.
Theorem 2 Let AS = hS, F, B, D, ≥, δ0i be an agent
system specification. A decision profile δ is a feasible
AS decision profile if and only if there is a feasible goal
set hG+, G−i of AS such that δ is a hG+, G−i decision
profile.
Consider a society of agent that tries to determine its
Pareto decision profiles. The game specifications sug-
gest the following algorithm:
calculate all decision profiles
for all decision profiles,
calculate consequences
order all decision profiles
The goal-based representation suggest an alternative
approach:
In this paper we have defined a qualitative decision
and game theory in the spirit of classical decision and
game theory. The theory illustrates the micro-macro
dichotomy by distinguishing the optimization problem
from game theoretic equilibria. We also showed that
any group decision, whether based on optimization or
on an equilibrium, can be represented by positive and
negative goals.
We think that the method of this paper is more inter-
esting than its formal results. The decision and game
theory are based on several ad hoc choices which need
further investigation. For example, the desire rules are
defeasible but the belief rules are not (the obvious ex-
tension leads to wishful thinking problems as studied
in [14, 3]). However, the results suggest that any group
decision can be understood as reaching for goals. We
hope that further investigations along this line brings
the theories and tools used for individual agents and
multi agent systems closer together.
1We cannot add beliefs on decision variables, as will be
clear in the next section. Suppose A = {a} and B = {⊤ →
a}. Clearly all decisions should be considered. However,
all decision profiles given in Definition 3 imply a and thus
all decisions not implying a would be excluded.
[15] J. von Neumann and O. Morgenstern. Theory of
Games and Economic Behavior. Princeton Uni-
versity Press, Princeton, NJ, 1 edition, 1944.
References
[1] K. Binmore. Fun and Games : A Text on Game
Theory. 1992.
[2] C. Boutilier. Toward a logic for qualitative deci-
sion theory. In Proceedings of the KR'94, pages
75 -- 86, 1994.
[3] J. Broersen, M. Dastani, , and L. van der Torre.
Resolving conflicts between beliefs, obligations,
intentions and desires. In Symbolic and Quanti-
tative Approaches to Reasoning and Uncertainty.
Proceedings of ECSQARU'01, LNCS 2143, pages
568 -- 579. Springer, 2001.
[4] J. Broersen, M. Dastani, Z. Huang, J. Hulstijn,
and L. van der Torre. The BOID architec-
ture: Conflicts between beliefs, obligations, in-
tentions and desires. In Proceedings of he Fifth
International Conference on Autonomous Agents
(AA2001), pages 9 -- 16. ACM Press, 2001.
[5] P.R. Cohen and H.J. Levesque. Intention is choice
with commitment. Artificial Intelligence, 42:213 --
261, 1990.
[6] J. Doyle and R. Thomason. Background to qual-
itative decision theory. AI magazine, 20(2), sum-
mer 1999.
[7] J. Lang. Conditional desires and utilities - an al-
ternative approach to qualitative decision theory.
In In Proceedings of the European Conference on
Artificial Intelligence (ECAI'96), pages 318 -- 322,
1996.
[8] A. Newell. The knowledge level. Artificial Intel-
ligence, 1982.
[9] Martin J. Osborne and Ariel Rubenstein. A
Course in Game Theory. The MIT Press, Cam-
bridge, Massachusetts, 1994.
[10] A. Rao and M. Georgeff. Modeling rational agents
within a BDI architecture. In Proceedings of the
KR91, 1991.
[11] L. Savage. The foundations of statistics. 1954.
[12] H. A. Simon. The Sciences of the Artificial. MIT
Press, Cambridge, MA, second edition, 1981.
[13] R. Thomason. Practical reason. Presentation at
PRR'99 workshop at IJCAI'99.
[14] R. Thomason. Desires and defaults: A framework
for planning with inferred goals. In Proceedings
KR 2000, pages 702 -- 713, 2000.
|
1905.01357 | 2 | 1905 | 2019-05-28T23:08:07 | Teaching on a Budget in Multi-Agent Deep Reinforcement Learning | [
"cs.MA",
"cs.LG"
] | Deep Reinforcement Learning (RL) algorithms can solve complex sequential decision tasks successfully. However, they have a major drawback of having poor sample efficiency which can often be tackled by knowledge reuse. In Multi-Agent Reinforcement Learning (MARL) this drawback becomes worse, but at the same time, a new set of opportunities to leverage knowledge are also presented through agent interactions. One promising approach among these is peer-to-peer action advising through a teacher-student framework. Despite being introduced for single-agent RL originally, recent studies show that it can also be applied to multi-agent scenarios with promising empirical results. However, studies in this line of research are currently very limited. In this paper, we propose heuristics-based action advising techniques in cooperative decentralised MARL, using a nonlinear function approximation based task-level policy. By adopting Random Network Distillation technique, we devise a measurement for agents to assess their knowledge in any given state and be able to initiate the teacher-student dynamics with no prior role assumptions. Experimental results in a gridworld environment show that such an approach may indeed be useful and needs to be further investigated. | cs.MA | cs | Teaching on a Budget in
Multi-Agent Deep Reinforcement Learning
Ercument Ilhan, Jeremy Gow and Diego Perez-Liebana
School of Electronic Engineering and Computer Science
Queen Mary University of London
London, United Kingdom
{e.ilhan, jeremy.gow, diego.perez}@qmul.ac.uk
9
1
0
2
y
a
M
8
2
]
A
M
.
s
c
[
2
v
7
5
3
1
0
.
5
0
9
1
:
v
i
X
r
a
Abstract -- Deep Reinforcement Learning (RL) algorithms can
solve complex sequential decision tasks successfully. However,
they have a major drawback of having poor sample efficiency
which can often be tackled by knowledge reuse. In Multi-Agent
Reinforcement Learning (MARL) this drawback becomes worse,
but at the same time, a new set of opportunities to leverage
knowledge are also presented through agent interactions. One
promising approach among these is peer-to-peer action advising
through a teacher-student framework. Despite being introduced
for single-agent RL originally, recent studies show that it can
also be applied to multi-agent scenarios with promising empirical
results. However, studies in this line of research are currently
very limited. In this paper, we propose heuristics-based action
advising techniques in cooperative decentralised MARL, using
a nonlinear function approximation based task-level policy. By
adopting Random Network Distillation technique, we devise a
measurement for agents to assess their knowledge in any given
state and be able to initiate the teacher-student dynamics with
no prior role assumptions. Experimental results in a gridworld
environment show that such an approach may indeed be useful
and needs to be further investigated.
Index Terms -- multi-agent, reinforcement learning, deep q-
networks, action advising, teacher-student
I. INTRODUCTION
Reinforcement Learning (RL) [1] is a prominent framework
for solving sequential decision-making tasks. In recent years,
the success of deep learning has led to the emergence of a
new set of techniques called deep RL that achieved important
breakthroughs, reaching super-human level of play in many
complex domains, including Atari games [2], the game of Go
[3] and StarCraft II [4]. However, modern RL methods suffer
from sample inefficiency and long training times, resulting
in limited scalability and applicability in practical situations.
Furthermore,
in the cases of having multiple RL agents
interacting in a shared environment, which is referred to as
Multi-Agent Reinforcement Learning (MARL), these problems
are further intensified due to the domain inherent challenges
such as non-stationarity and curse of dimensionality introduced
by multi-agent dynamics.
One natural solution to accelerate learning in deep RL
through overcoming sample inefficiency is knowledge reuse. A
considerable amount of methods to leverage past knowledge
are proposed to this date [5]. On the one hand, they are mostly
limited to single-agent RL due to the difficulties of MARL.
978-1-7281-1884-0/19/$31.00 ©2019 IEEE
On the other hand, agent interactions in MARL present unique
opportunities of knowledge reuse.
As shown in [6], various classes of methods are proposed to
take advantage of multiple agents for this purpose, such as im-
itation learning, inverse RL, and learning from demonstrations.
Action advising, which involves agents exchanging advice in the
form of actions between each other to drive their exploration
is one of the most flexible approaches among these. The only
requirement it has for the agents is to have a common action
space and be able to access their local observations through
observing or communicating at the time they need an advice.
It has a strong track record in single-agent RL and seems to
be a promising direction to investigate in MARL.
As Section II shows, the research on applications of action
advising techniques in MARL is currently in its early stages
and the methods are very limited. In this study, we aim to
investigate the performance of heuristic-based teacher-student
framework in cooperative MARL with agents employing
nonlinear function approximation representations in their task-
level policies. The problem setting we address does not hold any
assumptions of student-teacher roles for agents, is completely
decentralised in training and execution stages, and assumes that
the communication between agents is limited; which we believe
is realistic. To the best of our knowledge, this is the first study
to utilise classical action advising approach in conjunction with
deep MARL, and we hope to shed a light on advantages and
shortcomings to determine further research directions.
The paper is structured as follows: Section II provides a
review of related work. Afterwards, the background information
regarding the studied techniques are provided in Section III.
Section IV describes the game environment used in this project
and Section V details the proposed approach. Section VI
presents the experimental setup and Section VII the results
obtained. Finally, Section VIII concludes the paper and outlines
directions for future work.
II. LITERATURE REVIEW
A. Advising Methods in Single-Agent Reinforcement Learning
The idea of peer-to-peer knowledge sharing via advising
has its roots in single-agent RL. [7] is one of the first studies
to adopt this method, in a form of student-initiated advising,
in which the learning agent is assisted by an expert agent
whenever it asks for advice to make a decision. In [8], it
was shown that it is possible for one of the peers to be a
human and provide feedback to accelerate the learning of the
agent. Later on, [9] proposed the teacher-student framework.
According to this method, an expert teacher constantly monitors
a student agent's learning process and provides it with action
advices at appropriate times, limited with a budget, considering
the practical concerns regarding attention and communication.
Addressing the challenge of when to advise became more
crucial with the introduction of a budget in this line of research.
Following this work, [10] treated the optimal way of spending
this budget in terms of student's learning performance as an
RL problem and attempted to learn how to teach. In [11], a
new approach called jointly-initiated is presented. As opposed
to the previous methods in which advising occurs with student
[7] or teacher [9] initiation, they claim that their approach
discards the need for student agent to be constantly monitored,
making these techniques more feasible for human-agent settings.
[12] extended teacher-student framework to take advantage
of getting advice from multiple teachers by combining the
advices using a voting based selection. Moreover, the case of
getting suboptimal advices when student surpasses teacher's
performance is also addressed. Even though this improvement
relaxed the requirement of expert teachers, it still assumes
them to follow fixed policies that are good enough to provide
advice. Finally, [13] made further investigations in learning to
teach concept based on [10], and distinguished the qualities of
being an expert and being a good teacher, claiming that best
performers are not always the best teachers.
All of these previously mentioned studies are based on agents
that operate in isolated single-agent environments. In MARL,
multiple agents simultaneously learn in the same environment
while affecting each others' policies, rendering it non-stationary.
Due to this, even if we have expert agents, they can no longer be
guaranteed to follow a fixed policy. Moreover, the assumption
of teacher and student roles within agents are not applicable in
MARL, especially with large number of agents. Despite being
challenging to adapt on, these properties reflect a more natural
and realistic way of peer-to-peer knowledge sharing.
B. Advising Methods in Multi-Agent Reinforcement Learning
The application of action advising methods in MARL is
a challenging and a fairly new subject. [14] was the first
to propose a teacher-student framework that is suitable with
multi-agent settings. They extended the heuristics from [9] by
introducing several metrics based on the number of state visits
to measure confidence in a given state, in order to overcome
the challenge of having no fixed roles of student and teacher.
They tested their methods in a cooperative team of agents
utilizing SARSA(λ) with linear function approximation as
task-level policies. In [15], teaching in MARL is approached
as advising-level meta-learning problem. They proposed a
centralised training and decentralised execution procedure using
multi-agent actor-critic for teaching-level policies and tabular
Q-learning for task-level policies. Following this work, [16]
extended the idea of teaching to agents that use deep neural
networks in their task-level policies and act in environment with
continuous state and action spaces. Despite having promising
results, both of these learning based approaches are currently
limited with 2 agents only.
III. BACKGROUND
A. Multi-Agent RL
We are interested in a cooperative multi-agent setting
where agents get local observations and act in a decentralised
fashion. Decentralised partially observable Markov decision
process (Dec-POMDP)[17], which is a generalisation of
Markov decision process (MDP) for problems with multiple
decentralised agents with local observations, is a suitable
formalisation for this purpose. A Dec-POMDP is defined by a
tuple (cid:104)I,S,A,T ,R, Ω,O, γ(cid:105) where I is the set of n agents,
S is the set of states, A = ×i∈IAi is the set of joint actions,
T is the state transition probabilities, R is the reward function,
Ω = ×i∈IΩi is the finite of set of joint observations, O is the
set of conditional observation probabilities, and γ ∈ [0, 1) is
the discount factor. At each timestep t in a state s(cid:48), each agent
t determined by O(os(cid:48), a), where
i perceives observation oi
a = (cid:104)a1,··· , an(cid:105) is the joint action that caused the state
transition from s to s(cid:48) according to T (s(cid:48)a, s), and receives
reward ri determined by R(s, at). In fully-cooperative case,
which is the setup we are interested in, every agent gets
the same shared reward as ri = ··· = rn. In our problem
framework, agents are able to observe each other at any time
and infer their local observations. However, they still receive
local observations individually from environment.
B. Deep Q-Networks
Deep Q-Networks (DQN) [18] is a deep RL algorithm which
employs deep neural networks to represent environment states
and learn approximations of state-action values similar to Q-
learning in an end-to-end fashion. It overcomes the limitations
present in tabular and linear function approximation methods
(like tile coding) which struggle to generalise across states
and deal with large state spaces. At each timestep t, an
action at is selected randomly (depending on the exploration
technique) or greedily according to maxa Q(ot, a; ω) based
on state observation ot, where ω are the parameters of the
neural network. The transition information ot, at, rt+1, ot+1
is then stored in a buffer D, called replay memory. Using
minibatches of samples drawn from D, the network parameters
ω are trained periodically with gradient descent to minimise
the loss (rt+1 + γ maxa(cid:48) Q(ot+1, a(cid:48)) − Q(ot, at))2.
There has been many improvements on DQN over the
years [19], and some of them became essential parts of
the algorithm as they greatly improve convergence with no
significant drawbacks. In this paper, we utilize double Q-
learning, prioritised replay, dueling networks and noisy nets
enhancements among these.
C. Random Network Distillation
Random Network Distillation (RND) [20] is a technique
built to provide an intrinsic curiosity reward for RL agents to
enhance their exploration capabilities in complex environments.
It involves usage of two neural networks alongside the actual
task-level RL algorithm, namely target and predictor networks,
denoted by differentiable functions G and G respectively. These
networks are identical in structure, which is defined arbitrarily,
and are able to map observations to embeddings as in G : O →
Rk and G : O → Rk. They are initialised with different random
weights, and they produce different embeddings for identical
inputs in their initial state. Over the course of training, samples
used in task-level RL algorithm are also used to train predictor
network G using gradient descent to minimise the mean squared
error (cid:107)G(x) − G(x; θ)(cid:107)2. By doing so, the predictor network
becomes more accurate at matching target network's outputs
for the samples it observes more, which is referred to as
distilling a randomly initialised network. The error between
target and predictor are expected to be higher for type of states
that are seen less frequently and it acts as a natural indicator
of state novelty. The authors of [20] state that, due to its
simplicity, this method does not have to deal with error types
generated by environment stochasticity, model misspecification
and learning dynamics, unlike its counterparts that try to predict
state transitions.
D. Teaching on a Budget
The teacher-student framework [9] is a peer-to-peer knowl-
edge transfer procedure originally developed for two agents,
namely the teacher and the student. While the student is learning
to perform a single-agent task, the teacher provides action
suggestions to assist its exploration whenever it is appropriate.
These advices are expected to accelerate the student's learning.
Teacher agents also have a form of budget bgive which is the
number of advices they can provide to limit the amount they
can exchange to reflect realistic scenarios. The challenge of
this framework comes from distributing advices in the most
efficient way to accelerate the student's learning progression.
For this purpose, several heuristics are proposed to determine
when to advise in teacher-initiated strategy:
• Early advising: The intuition is that the earlier states
are generally more important in learning, and advice is
better taken early on. Following this idea, student is given
advice by teacher in the first bgive states.
• Importance advising: Even though earlier states are
usually more important, not all of them may worth to be
given advice in. Therefore, a budget is better spent on the
states that are determined to be more important. In this
method, importance of a state s is computed as:
I(s) = max
a
Qteacher(s, a) − min
a
Qteacher(s, a)
(1)
Advice is provided in a state s if bgive > 0 and I(s) is
greater than the importance threshold τimp.
IV. THE GAME ENVIRONMENT
Our experiment domain is a gridworld game with discrete
space and time, in which the agents must spread and cover
the landmark positions cooperatively to maximise their shared
rewards. The grid is sized 10 × 10 and consists of 3 agents
Fig. 1. Two frames of the game used for this study. On the left, the initial
game state, with three landmarks (grey) and three agents (green, blue and red).
On the right, the terminal state with all agents over the landmarks. Agents'
trajectories are shown as a shaded colour.
and 3 landmarks. The objective of the agents is to move
as quickly as possible to cover all the landmarks of the
level. While the landmarks are stationary, the agents can
navigate around the grid by using a discrete set of actions
A = { Do N othing, M ove U p, M ove Down, M ove Lef t,
M oveRight }. Movement actions executed by an agent change
its position on the grid by 1 tile in the corresponding direction.
At each timestep, the agents observe the relative positions of
the other agents and landmarks in form of x-axis and y-axis
values, plus the current timestep. All values are normalised
to [−1, 1] using the maximum distance and game duration,
respectively. Agents also receive a common reward in [0, 1],
determined by how many of the landmarks have an agent
covering it. The game's reward is calculated as:
L(cid:88)
l=1
1
L
((cid:80)I
i=1
1
(cid:62) 1
1(cid:107)dli(cid:107)=0)
(2)
where I is the set of agents, L is the set of landmarks, 1
is the indicator function, and dli is the Manhattan distance
between landmark l and agent i. In other words, in order
to act optimally and maximise the total reward, the agents
must determine the lowest cost (in terms of distance) of the
agent-landmark pair set and move to the appropriate landmark
following the shortest path. Since the agents have no access to
a forward model, they are expected to learn these strategies
through interactions.
Despite this environment's representational and mechanical
simplicity, it still is capable of presenting complex behavioural
challenges for MARL. Therefore, we decided to conduct
experiments in this setting to focus on behavioural learning
while keeping the computational expense at minimum.
We tuned the environment complexity in terms of size and
reward sparsity, in order to ensure that the environment is
solvable at least sub-optimally by independent DQN agents,
and there remains to be some exploration challenge to keep
the knowledge of the peers valuable to the agent.
V. PROPOSED METHOD
Our approach aims to accelerate learning of a cooperative
team of agents in a multi-agent environment with a teacher-
student action advising approach. The problem setting we
take into consideration is fully observable, completely decen-
tralised in every stage, and most importantly, assumes that
the environment is too complex for tabular or linear function
approximation methods to be successful, unlike the majority
of related work. In this settings, every agent operates with a
local observation, however they are able to observe each other
and infer any other agent's local observations. Additionally,
they can exchange action advising requests and responses in
forms of actions. We take the budget constraints in inter-agent
interactions into consideration as introduced in [9], and we want
to achieve acceleration while keeping the agent interactions
at minimum, since we believe that such behaviour will be the
most useful in practical applications.
The general structure of our proposal can be expressed in two
parts: a task-level policy and the teacher-student framework,
which we explain in the following sections.
A. Task-Level Policy
At task-level learning and decision making, agents are de-
signed to employ a nonlinear function approximation according
to the target problem structure. Our agents employ DQN along
with its well-established improvements of double Q-learning,
prioritised replay, dueling networks and noisy nets. In our
framework, maybe the most important of these is prioritised
replay as it helps with sampling transitions produced by advices
which are likely to have larger temporal-difference error more
frequently.
The network structure is a multilayer perceptron (MLP),
which is the archetypal form of the deep learning models [21],
with a single hidden layer formed by 256 units. All layer
connections are dense and noisy. This addition removes the
need to follow an explicit exploration policy such as -greedy
while learning, which we think is more suitable for an action
advising approach.
As the multi-agent learning strategy, we follow independent
learners approach in which each agent treats other agents
as part of a non-stationary environment and behaves by
taking only its own actions into account. This is the simplest
approach in MARL without requiring the algorithm to have
any specialisations for multiple agents. Therefore, we keep
DQN as it is without any further modifications in this regard,
which is referred to as independent DQNs. Despite having no
theoretical convergence guarantees in multi-agent environments,
independent DQNs were able to exhibit promising empirical
results in previous studies and are often used as a baseline for
further improvements [22][23].
B. Teaching on a Budget
We adopt action advising in the form of the teacher-student
framework to perform inter-agent knowledge transfer. This
method requires only a common action set and minimal
similarity between agents. In addition, considering commu-
nication costs, exchanging actions instead of episodes or policy
parameters is more preferable especially when task-level policy
employs a complex model like DQNs. In our approach, agents
can broadcast requests for advices when they need it, and can
also respond to requests for advices from other agents. They
have separate budgets of asking bask and giving bgive advices,
which determine the total number of times they can perform
these interactions following either early advising or importance
advising heuristics.
Despite its simplicity, adapting the teacher-student framework
to MARL with deep RL task-level policies is not straightfor-
ward. First of all, in our problem setting, every agent is learning
simultaneously and no longer have fixed policies. Moreover,
they now have different levels of knowledge about the task,
and have no information on each other's expertise since they
can not have any access to internal information and they hold
on assumptions about roles of being teacher or student. In
order to overcome this, we follow a jointly-initiated advising
strategy [11]. Every time an agent requests an advice from
another agent, the agent in the position of teacher executes this
interaction if it thinks that it can take this role. Furthermore,
as agents have no fixed roles or knowledge levels that are set
previously, they must be capable of determining if they need
an advice based on the state they are, or if they are experienced
enough to give advice. In [14], they rely on number of visits to
measure an agent's certainty in a given state. However, this is
not applicable when state space is large and nonlinear function
approximation is used to represent states. We propose to use
RND technique as a metric to measure agents' uncertainties as
an alternative for nonlinear function approximation. Every time
an agent uses a sample of state observation to train its internal
task-level policy, this sample is also used to distill its predictor
network G. Consequently, it will be able to measure how
uncertain it is in a state with observation o by measuring the
error (cid:107) G(o; θ) − G(o)(cid:107)2 when determining if it needs advice
or is capable of giving it. This approach can be treated as
number of visits in nonlinear function approximation regime.
At each timestep, each agent i chooses its action according to
teacher-student augmented action selection procedure defined
in Algorithm 1. If it has a remaining asking budget bask, it
measures its uncertainty µ using its internal models G and G;
if µ is higher than a predefined asking threshold τask the agent
broadcasts its advice requests to other agents as described in
Algorithm 2. Agents who have any remaining advice giving
budget bgive attempt to respond to this request. If agent receives
any advice, it determines which one to follow by using majority
voting (ties broken at random), to then execute such action.
Otherwise, if no advice is received, it continues exploring by
following its own policy.
Responding to an advice request happens as described in
Algorithm 3. Upon getting a request, the agent first checks if it
has any remaining giving budget bgive. If so, it then determines
its expertise by using its internal G and G, computing (cid:107)G(o)−
G(o)(cid:107)2. It then compares this value with threshold τgive to
decide to proceed or not. Finally, if it is set to follow the Early
Advising heuristic, it broadcasts an action advice according to
its policy; if the heuristic is set to Importance Advising, it also
checks if the state is important enough to give advice for by
using Equation 1.
The general learning flow of an agent with DQN policy and
action advising mechanism is shown in Algorithm 4. As can
be seen, the task-level policy can easily be isolated from these
Algorithm 1 Teacher-student augmented action selection
Require: state observation o
1: procedure SELECTACTION
2:
3:
4:
5:
6:
µ ← (cid:107)G(o) − G(o)(cid:107)2
if µ > τask then
a ← N one
if bask > 0 then
a ← ASKFORADVICE(o)
(cid:46) Determine uncertainty
7:
8:
9:
10:
11:
if a = N one then
a ← π(o)
bask ← bask − 1
else
return a
a ← N one
A ← ∅
for every other agent i do
aadvice ← i.ADVISE
if aadvice (cid:54)= N one then
Algorithm 2 Advice seeking
Require: state observation o
1: procedure ASKFORADVICE
2:
3:
4:
5:
6:
7:
8:
9:
10:
if A (cid:54)= ∅ then
add aadvice to A
return a
a ← perform majority voting in A
a ← N one
if bgive > 0 then
Algorithm 3 Respond to advice request
1: procedure ADVISE
2:
3:
4:
5:
6:
7:
8:
9:
10:
switch teaching method do
os ← observe student's state.
µ ← (cid:107)G(os) − G(os)(cid:107)2 (cid:46) Determine uncertainty
if µ < τgive then
case Early Advising
a ← π(o)
bgive ← bgive − 1
io ← maxa Q(os, a) − mina Q(os, a)
if io > τimportance then
case Importance Advising
a ← π(o)
bgive ← bgive − 1
11:
12:
13:
14:
15:
16:
return a
enhancements except for its action selection policy. Therefore,
it can be changed for any other policy as long as it satisfies the
assumption of computing state-action values, for importance
advising to be applicable.
VI. EXPERIMENTAL SETUP
The objective of the evaluation is to understand if and
how the proposed modifications can enhance the learning
performance. We conducted experiments through multiple
Initialise DQN model
Initialise replay memory
Initialise G and G
for all training episodes do
o ← initial observation
for all episode steps do
Algorithm 4 Training of task-level policy (DQN) with teacher-
student framework
1: procedure TRAIN
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
a ← SELECTACTION(o)
Execute a and get reward r and observation o(cid:48)
Store transition (o, a, r, o(cid:48)) in replay memory
o ← o(cid:48)
Sample a minibatch from replay memory
Train DQN with minibatch
Train G with minibatch
learning sessions, each one consisting of a set fixed number
of different game episodes which are initialised with random,
yet non-overlapping, agent and landmark positions.
The performance of the agents is assessed as a team
through a learning session. They are evaluated every 100
episodes in a predefined set of 50 evaluation levels. During
evaluation, learning and teaching procedures are disabled,
and the levels used are fixed across all learning sessions.
An evaluation score is calculated by normalising the average
episode rewards obtained across 50 levels with the maximum
possible total reward (determined by a set of hand-crafted
expert agents), giving a score in [0, 1], where 1 indicates the
optimal performance.
The performance of the proposed methods can be assessed
by looking at the evaluation scores across a learning session,
according to the following two metrics:
• Asymptotic performance: This is measured directly by
looking at the evaluation scores values, and represents
how good the agents are at solving the game.
• Learning speed: This is measured by looking at area
under the curve of evaluation scores against the number
of training episodes graph.
For agents to be able to benefit from knowledge transfer,
there must be some form of knowledge heterogeneity within
the team. In MARL, such heterogeneity tends to arise when the
agents explore different parts of the state space, use different
task-level policies in terms of complexity and representation,
and are in different stages of training. Since our environment
is fully observable, and we use identical agents, only the latter
is applicable in our setting.
One objective of this study is to determine how the proposed
methods work in different types of knowledge heterogeneity.
Therefore, we design the following 2 scenarios:
• Scenario I: we train a team of agents in levels from a single
distribution of levels; then, we take agents from different
stages of pre-training to form a team to be evaluated.
• Scenario II: we train 3 sets of agents in 3 different
level distributions, in which landmarks and initial agents
Fig. 2. Level structure types in terms of possible regions for initial positions
of agents (gray) and landmarks (red).
locations are strictly limited to predefined regions (see
Figure 2). Then, we take one agent with moderate
performance arbitrarily from each level type to form a
team to evaluate. Note that the learning sessions used to
pre-train agents are generated with different seeds than
the ones where we run the final evaluation.
The following is a list of the different types of agents used
in this study. These agents are trained in the levels first to
acquire some knowledge, and then picked to form teams of 3
agents as explained above. These agents are:
• A0: Agent with no prior knowledge.
• A10: Agent taken from a team of agents trained for 10K
episodes. This agent obtained an evaluation score of 0.45
in a level distribution identical to the evaluation sessions.
• A20: Agent taken from a team of agents trained for
20K episodes. The evaluation score is 0.91 in a level
distribution identical to the learning evaluation sessions.
• A10-1: Agent taken from a team of agents trained for
10K episodes, with 0.51 evaluation score in a restricted
level distribution of type 1.
• A10-2: Agent taken from a team of agents trained for
10K episodes, with 0.41 evaluation score in a restricted
level distribution of type 2.
• A10-3: Agent taken from a team of agents trained for
10K episodes, with 0.43 evaluation score in a restricted
level distribution of type 3.
The methods we evaluate along with the agent teams used
in them are as follows:
• XP: The team to be evaluated is formed by agents A0,
A10, A20 and no advising is used. This method serves
as a baseline for the first scenario.
• XP-EA: The team to be evaluated is formed by agents
A0, A10, A20, and early advising is enabled.
• XP-IA: The team to be evaluated is formed by agents
with different knowledge levels A0, A10, A20, and
importance advising is enabled.
• XP-L: The team to be evaluated is formed by agents A10-
1, A10-2, A10-3 and no advising is used. This method
serves as a baseline for the second scenario.
• XP-L-EA: The team to be evaluated is formed by agents
A10-1, A10-2, A10-3, using early advising.
• XP-L-IA: The team to be evaluated is formed by agents
with different knowledge levels A10-1, A10-2, A10-3,
using importance advising.
The hyperparameters of the learning algorithms are deter-
mined empirically, as shown in Table I. They are kept the same
TABLE I
HYPERPARAMETERS.
Parameter name
Minimum replay memory size to start learning
Replay memory capacity
Prioritisation type
Prioritisation exponent α
Prioritisation importance sampling β
Target network update period
Minibatch size
Learning rate
Train period
Discount factor γ
Adam
Huber loss δ
Asking threshold τask
Giving threshold τgive
Importance threshold τimp
Value
10000
25000
proportional
0.6
0.5 → 1
10000
64
0.001
2
0.99
1.5 × 10−4
1
10
3
1
across different scenarios and configurations.
We run the tests for the methods with teaching enabled,
namely XP-EA and XP-IA, with different advice budgets of
1K, 2K, 5K, 10K, 20K and unlimited (∞). For Scenario I (XP,
XP-EA, XP-IA), learning is run for 10K episodes, while 20K
episodes are run for Scenario II (XP-L, XP-L-EA, XP-L-IA).
The budgets are separate for each individual agent and are
set the same for asking and giving advice initially. Moreover,
due to stability concerns with deep RL methods, especially
in multi-agent settings, every experiment is repeated 10 times
with different random seeds.
VII. RESULTS AND DISCUSSION
Tables II and III show the results for the scenarios described
in this paper. These results show the asymptotic performance
(score) and the area under the curve (AUC) values as indicators
of the learning performance of the agents. Results are reported
for different moments of the evaluated training: after 2.5K,
5K, 7.5K, 10K, 15K and 20K episodes, and they include the
standard error of the measure (10 repetitions). Results that are
significantly different than the first row (baseline) of each table,
according to Welch's t-test (p < 0.05) are denoted in bold.
Additionally, evaluation scores of the methods with highest
final AUC values from each of the scenarios are plotted against
their respective baseline as shown in Figures 3 and 4.
In the first scenario, both of the algorithms provided slight
accelerations in learning and achieved very similar final
performances with the baseline method at 10Kth episode. The
only significant difference from the baseline was seen at scores
at 5Kth episode, by XP-IA (5K). The overall best performing
agent in this scenario is XP-IA (10K). This can be seen as
an indication that the importance metric was indeed a useful
heuristic to distribute advice over more important states in this
scenario. The budget seems to have more effect on XP-IA,
achieving its best result at 10K, confirming the claim that too
much advice may have negative effects on performance [9].
In the second scenario, XP-L-IA failed to show any signif-
icant advantage over the baseline XP-L except for the score
at 5Kth episode with budget of 1K. This can be a result of
ASYMPTOTIC PERFORMANCE (SCORE) AND AREA UNDER THE CURVE (AUC) VALUES OF XP, XP-EA AND XP-IA AT LEARNING EPISODES 2.5K, 5K, 7.5K
AND 10K WITH STANDARD ERRORS IN PARENTHESES. RESULTS OF XP-EA AND XP-IA THAT ARE SIGNIFICANTLY DIFFERENT THAN XP ALGORITHM
ACCORDING TO WELCH'S t-TEST WITH p < 0.05 ARE DENOTED IN BOLD. EACH ALGORITHM IS RUN 10 TIMES.
TABLE II
Algorithm
XP
XP-EA (1K)
XP-EA (2K)
XP-EA (5K)
XP-EA (10K)
XP-EA (20K)
XP-EA (∞)
XP-IA (1K)
XP-IA (2K)
XP-IA (5K)
XP-IA (10K)
XP-IA (20K)
XP-IA (∞)
At 2500th
At 5000th
At 7500th
At 10000th
AUC
Score
AUC
Score
AUC
Score
AUC
Score
13.94 (0.04)
0.58 (0.005)
28.72 (0.12)
0.62 (0.009)
47.15 (0.44)
0.83 (0.013)
68.26 (0.6)
0.87 (0.006)
13.9 (0.05)
13.85 (0.07)
13.88 (0.03)
13.85 (0.04)
13.76 (0.05)
13.76 (0.05)
13.94 (0.04)
13.91 (0.08)
13.92 (0.03)
13.97 (0.04)
13.96 (0.04)
13.96 (0.04)
0.58 (0.002)
0.57 (0.005)
0.58 (0.004)
0.57 (0.004)
0.57 (0.004)
0.57 (0.003)
0.58 (0.003)
0.58 (0.005)
0.57 (0.004)
0.58 (0.003)
0.58 (0.003)
0.58 (0.003)
28.81 (0.14)
28.58 (0.13)
28.75 (0.11)
28.57 (0.09)
28.56 (0.09)
28.53 (0.06)
29.0 (0.15)
28.86 (0.21)
28.84 (0.12)
29.22 (0.22)
29.19 (0.2)
29.19 (0.2)
0.63 (0.014)
0.61 (0.009)
0.63 (0.018)
0.61 (0.006)
0.64 (0.016)
0.64 (0.011)
0.66 (0.016)
0.64 (0.02)
0.65 (0.008)
0.68 (0.024)
0.67 (0.022)
0.67 (0.022)
47.18 (0.54)
46.39 (0.59)
47.02 (0.56)
46.3 (0.47)
47.38 (0.37)
47.11 (0.24)
47.62 (0.59)
47.11 (0.61)
48.01 (0.38)
48.26 (0.72)
47.92 (0.58)
47.92 (0.58)
0.81 (0.013)
0.79 (0.024)
0.81 (0.024)
0.8 (0.021)
0.84 (0.005)
0.82 (0.006)
0.82 (0.012)
0.81 (0.015)
0.83 (0.006)
0.8 (0.019)
0.8 (0.012)
0.8 (0.012)
68.27 (0.73)
67.16 (0.87)
67.86 (0.86)
67.28 (0.68)
68.87 (0.45)
68.0 (0.31)
68.92 (0.71)
68.16 (0.84)
69.49 (0.41)
69.6 (0.98)
69.01 (0.75)
69.01 (0.75)
0.86 (0.01)
0.86 (0.009)
0.87 (0.007)
0.87 (0.008)
0.88 (0.003)
0.86 (0.003)
0.87 (0.006)
0.87 (0.009)
0.88 (0.004)
0.88 (0.006)
0.87 (0.004)
0.87 (0.004)
ASYMPTOTIC PERFORMANCE (SCORE) AND AREA UNDER THE CURVE (AUC) VALUES OF XP-L, XP-L-EA AND XP-L-IA AT LEARNING EPISODES 5K, 10K,
15K AND 20K WITH STANDARD ERRORS IN PARENTHESES. RESULTS OF XP-L-EA AND XP-L-IA THAT ARE SIGNIFICANTLY DIFFERENT THAN XP-L
ALGORITHM ACCORDING TO WELCH'S t-TEST WITH p < 0.05 ARE DENOTED IN BOLD. EACH ALGORITHM IS RUN 10 TIMES.
TABLE III
Algorithm
XP-L
XP-L-EA (1K)
XP-L-EA (2K)
XP-L-EA (5K)
XP-L-EA (10K)
XP-L-EA (20K)
XP-L-EA (∞)
XP-L-IA (1K)
XP-L-IA (2K)
XP-L-IA (5K)
XP-L-IA (10K)
XP-L-IA (20K)
XP-L-IA (∞)
At 5000th
At 10000th
At 15000th
At 20000th
AUC
Score
AUC
Score
AUC
Score
AUC
Score
15.38 (0.12) 0.36 (0.005)
38.34 (0.43)
0.52 (0.006)
66.97 (0.71)
0.66 (0.02)
105.46 (1.37)
0.86 (0.011)
0.37 (0.009)
38.28 (0.66)
68.55 (1.49)
15.58 (0.1) 0.39 (0.008) 39.51 (0.33) 0.54 (0.005) 69.85 (0.85)
15.4 (0.12)
15.51 (0.14) 0.39 (0.007) 39.98 (0.37) 0.54 (0.006) 71.18 (0.68) 0.75 (0.022) 112.69 (1.27) 0.89 (0.005)
0.74 (0.026) 112.14 (1.77) 0.87 (0.014)
15.53 (0.1) 0.38 (0.008) 39.53 (0.35) 0.55 (0.005) 71.3 (1.02)
0.74 (0.026) 112.14 (1.77) 0.87 (0.014)
15.53 (0.1) 0.38 (0.008) 39.53 (0.35) 0.55 (0.005) 71.3 (1.02)
15.53 (0.1) 0.38 (0.008) 39.53 (0.35) 0.55 (0.005) 71.3 (1.02)
0.74 (0.026) 112.14 (1.77) 0.87 (0.014)
110.8 (1.55)
108.4 (2.28)
0.73 (0.024)
0.7 (0.022)
0.88 (0.018)
0.86 (0.02)
0.53 (0.012)
15.39 (0.13) 0.38 (0.006)
15.27 (0.14) 0.37 (0.008)
15.37 (0.09) 0.38 (0.009)
15.37 (0.1)
0.37 (0.009)
15.37 (0.09) 0.38 (0.009)
15.37 (0.09) 0.38 (0.009)
38.54 (0.5)
38.4 (0.43)
38.67 (0.32)
38.58 (0.34)
38.67 (0.32)
38.67 (0.32)
0.52 (0.008)
0.53 (0.007)
0.52 (0.005)
0.52 (0.006)
0.52 (0.005)
0.52 (0.005)
67.95 (0.97)
67.73 (0.69)
68.13 (0.59)
68.06 (0.65)
68.13 (0.59)
68.13 (0.59)
0.68 (0.025)
0.67 (0.017)
0.69 (0.025)
0.68 (0.027)
0.69 (0.025)
0.69 (0.025)
107.6 (1.59)
106.47 (1.42)
107.32 (1.46)
107.04 (1.6)
107.32 (1.46)
107.32 (1.46)
0.87 (0.01)
0.85 (0.016)
0.86 (0.019)
0.85 (0.02)
0.86 (0.019)
0.86 (0.019)
Fig. 3. Evaluation scores versus number of learning episodes of XP, XP-EA
(5K) and XP-IA (20K). Shaded areas indicate 95% confidence intervals.
Fig. 4. Evaluation scores versus number of learning episodes of XP-L, XP-L-
EA (5K) and XP-L-IA (1K). Shaded areas indicate 95% confidence intervals.
the importance metric not being accurate at reflecting actual
relevance of states in this kind of agent knowledge setting. On
the other hand, XP-L-EA performed very well with significant
improvements in terms of asymptotic performance and learning
speed at multiple stages of learning. Moreover, it even managed
to achieve a significantly better final performance. This may
be caused by the agents starting with similar (and moderate)
amount of knowledge, so the early advices are likely to be
useful for any of them without having a need for additional
importance assessment. The identical results of XP-L with 10K
or a higher budget is caused by not having the need to make
use of it beyond some point, once the agent is certain of the
decision it is making on its own (as controlled with the advice
budgets). This can be considered as an another benefit of using
this uncertainty measurement technique if it is tuned well.
VIII. CONCLUSIONS
This paper describes the application, for the first time, of
action advising via heuristic-based teacher-student framework
on Multi-Agent Reinforcement Learning (MARL) agents,
employing policies with nonlinear function approximation.
The environment used for training agent advising is a grid-
based game in which three agents need to coordinate to place
themselves in three different landmarks. The work described
here shows that using off-policy learning can provide significant
improvements in the speed and performance of agents that learn
via advice, particularly when the team is composed of agents
with heterogeneous knowledge. Moreover, Random Network
Distillation (RND) can be a reliable metric to be utilised as a
state visit counter through nonlinear function approximation
when state space complexity is high.
Another interesting finding is that the state importance metric
may be inefficient in some cases of knowledge distribution
amongst agents, for example if they are all are experts but
on different state distributions. Additionally, it is worthwhile
highlighting that, even if it
is possible to determine the
experience of the agents for their roles in knowledge exchange
relationships, this experience is importantly biased by the other
agents that were present at the time they built their knowledge.
Further investigation on how to adapt the importance metric
for agent advising is an interesting line of future work.
Off-policy learning through replay memory may be a slowing
down factor in action advising, as it takes given advice into
consideration in a delayed way and reduces the rate they
influence the agent's current policy. Therefore, in addition
to the enhancements like prioritised experience replay, it may
be useful to implement more specific techniques like multi-step
advice and continual monitoring of agents for fixed periods of
time after advice exchange. This is similar to previous work
in the field [11][16].
Finally, another interesting line of future work is to expand
the problem to more than 3 agents, which brings interesting
aspects to the discussion such as defining a more accurate peer
selection and decision making beyond majority voting.
REFERENCES
[1] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction.
MIT press, 2018.
[2] V. Mnih, K. Kavukcuoglu, D. Silver, et al., "Human-level Con-
trol Through Deep Reinforcement Learning," Nature, vol. 518:7540,
pp. 529 -- 533, 2015.
[3] D. Silver, A. Huang, C. J. Maddison, et al., "Mastering the Game of
Go with Deep Neural Networks and Tree Search," Nature, vol. 529,
no. 7587, pp. 484 -- 489, 2016.
[4] O. Vinyals, I. Babuschkin, J. Chung, et al., AlphaStar: Mastering the
Real-Time Strategy Game StarCraft II, https://deepmind.com/blog/
alphastar-mastering-real-time-strategy-game-starcraft-ii/, 2019.
[5] M. E. Taylor and P. Stone, "Transfer Learning for Reinforcement
Learning Domains: A Survey," Journal of Machine Learning Research,
vol. 10, pp. 1633 -- 1685, 2009.
F. L. Da Silva and A. H. R. Costa, "A Survey on Transfer Learning
for Multiagent Reinforcement Learning Systems," Journal of Artificial
Intelligence Research, vol. 64, pp. 645 -- 703, 2019.
J. A. Clouse, On Integrating Apprentice Learning and Reinforcement
Learning. University of Massachusetts Amherst, 1996.
S. Griffith, K. Subramanian, J. Scholz, C. L. Isbell, and A. L. Thomaz,
"Policy Shaping: Integrating Human Feedback with Reinforcement
Learning," in Neural Information Processing Systems, 2013, pp. 2625 --
2633.
L. Torrey and M. E. Taylor, "Teaching on a Budget: Agents Advising
Agents in Reinforcement Learning," in International conference on
Autonomous Agents and Multi-Agent Systems, AAMAS '13, Saint Paul,
MN, USA, May 6-10, 2013, 2013, pp. 1053 -- 1060.
[7]
[8]
[6]
[9]
[15]
[14]
[10] M. Zimmer, P. Viappiani, and P. Weng, "Teacher-Student Framework: A
Reinforcement Learning Approach," in AAMAS Workshop Autonomous
Robots and Multirobot Systems, 2014.
[11] O. Amir, E. Kamar, A. Kolobov, and B. J. Grosz, "Interactive Teaching
Strategies for Agent Training," in Proceedings of the Twenty-Fifth
International Joint Conference on Artificial Intelligence, IJCAI 2016,
New York, NY, USA, 9-15 July 2016, 2016, pp. 804 -- 811.
[12] Y. Zhan, H. Bou-Ammar, and M. E. Taylor, "Theoretically-Grounded
Policy Advice from Multiple Teachers in Reinforcement Learn-
ing Settings with Applications to Negative Transfer," CoRR, vol.
abs/1604.03986, 2016.
[13] A. Fachantidis, M. E. Taylor, and I. P. Vlahavas, "Learning to Teach
Reinforcement Learning Agents," CoRR, vol. abs/1707.09079, 2017.
F. L. da Silva, R. Glatt, and A. H. R. Costa, "Simultaneously Learning
and Advising in Multiagent Reinforcement Learning," in Proceedings
of the 16th Conference on Autonomous Agents and MultiAgent Systems,
ser. AAMAS '17, Sao Paulo, Brazil, 2017, pp. 1100 -- 1108.
S. Omidshafiei, D. Kim, M. Liu, G. Tesauro, M. Riemer, C. Amato, M.
Campbell, and J. P. How, "Learning to Teach in Cooperative Multiagent
Reinforcement Learning," CoRR, vol. abs/1805.07830, 2018.
[16] D. K. Kim, M. Liu, S. Omidshafiei, S. Lopez-Cot, M. Riemer, G. Habibi,
G. Tesauro, S. Mourad, M. Campbell, and J. P. How, "Learning Hier-
archical Teaching in Cooperative Multiagent Reinforcement Learning,"
ArXiv preprint arXiv:1903.03216, 2019.
F. A. Oliehoek and C. Amato, A Concise Introduction to Decentralized
POMDPs, ser. Springer Briefs in Intelligent Systems. Springer, 2016.
[18] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D.
Wierstra, and M. A. Riedmiller, "Playing Atari with Deep Reinforcement
Learning," CoRR, vol. abs/1312.5602, 2013.
[19] M. Hessel, J. Modayil, H. van Hasselt, T. Schaul, G. Ostrovski,
W. Dabney, D. Horgan, B. Piot, M. G. Azar, and D. Silver, "Rainbow:
Combining Improvements in Deep Reinforcement Learning," in Pro-
ceedings of the Thirty-Second AAAI Conference on Artificial Intelligence,
(AAAI-18), 2018, pp. 3215 -- 3222.
[20] Y. Burda, H. Edwards, A. J. Storkey, and O. Klimov, "Exploration by
[17]
[21]
[22]
Random Network Distillation," CoRR, vol. abs/1810.12894, 2018.
I. J. Goodfellow, Y. Bengio, and A. C. Courville, Deep Learning, ser.
Adaptive Computation and Machine Learning. MIT Press, 2016.
J. Foerster, I. A. Assael, N. de Freitas, and S. Whiteson, "Learning
to Communicate with Deep Multi-Agent Reinforcement Learning," in
Neural Information Processing Systems, 2016, pp. 2137 -- 2145.
[23] A. Tampuu, T. Matiisen, D. Kodelja, I. Kuzovkin, K. Korjus, J. Aru,
J. Aru, and R. Vicente, "Multiagent Cooperation and Competition with
Deep Reinforcement Learning," CoRR, vol. abs/1511.08779, 2015.
|
1503.03767 | 1 | 1503 | 2015-03-12T15:24:12 | Exploiting Near Time Forecasting From Social Network To Decongest Traffic | [
"cs.MA",
"cs.SI"
] | Preventing traffic congestion by forecasting near time traffic flows is an important problem as it leads to effective use of transport resources. Social network provides information about activities of humans and social events. Thus, with the help of social network, we can extract which humans will attend a particular event (in near time) and can estimate flow of traffic based on it. This opens up a wide area of research which poses need to have a framework for traffic management that can capture essential parameters of real-life behaviour and provide a way to iterate upon and evaluate new ideas. In this paper, we present building blocks of a framework and a system to simulate a city with its transport system, humans and their social network. We emphasize on relevant parameters selected and modular design of the framework. Our framework defines metrics to evaluate congestion avoidance strategies. To show utility of the framework, we present experimental studies of few strategies on a public transport system. | cs.MA | cs | Exploiting Near Time Forecasting From Social Network
To Decongest Traffic
Deepika Pathania and Kamalakar Karlapalem
Center for Data Engineering,
International Institute of Information Technology Hyderabad, India
{[email protected],[email protected]}
Abstract. Preventing traffic congestion by forecasting near time traffic flows is
an important problem as it leads to effective use of transport resources. Social
network provides information about activities of humans and social events. Thus,
with the help of social network, we can extract which humans will attend a par-
ticular event (in near time) and can estimate flow of traffic based on it. This opens
up a wide area of research which poses need to have a framework for traffic man-
agement that can capture essential parameters of real-life behaviour and provide
a way to iterate upon and evaluate new ideas. In this paper, we present building
blocks of a framework and a system to simulate a city with its transport system,
humans and their social network. We emphasize on relevant parameters selected
and modular design of the framework. Our framework defines metrics to evalu-
ate congestion avoidance strategies. To show utility of the framework, we present
experimental studies of few strategies on a public transport system.
Keywords: multi-modal traffic management, decongestion, social networks, simula-
tion
1 Introduction
Traffic congestion is a major problem that is faced in many parts of the world. It wastes
two precious resources - time and fuel. According to Texas A&M Transportation Insti-
tute's 2012 Urban Mobility Report [25], congestion cost of extra time and fuel is $121
billion in monetary terms for 498 urban areas in 2011; and for an average commuter,
it costs more than $2 per day. This cost has increased from $40 billion to $121 billion
over last 3 decades. There can be many causes responsible for congestion: (i) traffic
(demand) more than road-rail capacity (supply); for example, increase in ridership of
public transportation like trains due to social events (ii) unexpected human movement
activities, for example, vehicle breakdowns on roads/rails, events like concerts in a city.
Public transport systems are usually prepared for routine movement of humans, but not
for unexpected events; either they do not have prior information (accidents, etc.) or not
enough information (like an estimate on number of humans that may visit a concert and
which routes would see an increased use).
Traffic prediction is a possible solution to solve congestion problem. If we have
prior knowledge about extent of traffic at particular places, we can make arrangements
5
1
0
2
r
a
M
2
1
]
A
M
.
s
c
[
1
v
7
6
7
3
0
.
3
0
5
1
:
v
i
X
r
a
2
Deepika Pathania and Kamalakar Karlapalem
beforehand and avoid congestion; like having extra buses to meet increase in ridership.
Many solutions have been studied regarding prediction of traffic. In real time sensing of
traffic, current behavior and flow of traffic is judged (using sensors, image processors,
etc.) and then congestion-prone areas are found. These methods provide information
about likelihood of congestion at a short notice and it might not be feasible to control
congestion formation. Analyzing historical data related to traffic behavior is another
approach to anticipate traffic congestions. Aforementioned techniques do not provide
movement plan and travel destination of humans. If we can estimate travel plans well
before, we know the immediate future flow of traffic and congestion avoidance plans
can be implemented in a timely and efficient manner.
1.1 Motivation
With the emergence of social networking websites, we explore another approach for
traffic prediction. Nowadays, many humans update their daily activities or plans on so-
cial websites. Further, these websites have also become a media of rapid information
dissemination. Humans notice events (like a concert or a spot-sale) and quickly spread
this information to thousands of others. Social network data can potentially (with some
limitations due to privacy) enable us to estimate which humans will attend a particular
event. We can estimate the flow of traffic using humans' source and destination loca-
tions. This will allow us to estimate traffic flow well before time and thus give us more
time to explore possibilities of actions that could be taken to control traffic congestion.
This problem domain opens up a large area of research.
In our work, we present a framework that simulates essential components of the
problem:
1. a city with its points of interests (homes, offices, train-stations, schools, theatres,
hospitals, parks, etc.);
2. a human population with their daily activities;
3. a social network of humans;
4. a transport management system that monitors traffic and plans to control conges-
tion;
5. social events which cause anomalies in traffic pattern.
Our design and framework for solving the problem is generic and components are flex-
ibly pluggable to be useful and extendible by other researchers. We use Singapore city
to explain framework parameters and present results of simulation.
We consider trains as means of transport in our framework. Areas of congestion is
limited to train stations and elements that form congestion are humans (and not any
vehicles). Trains take commuters from one train station to another. There is a waiting
period for humans at train stations. If an already full train arrives, then humans will not
be able to get into train and have to wait for the next train that can cause congestion. If
there is too much congestion at train station, it will take a long time before a human can
board a train. Hence, the transport management system aims to plan train services such
that humans have to wait as less as possible.
We have used agent-based approach to develop our framework. Humans carry out
their daily activities like going to schools, offices, attending social events, etc. Each
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
3
human takes autonomous decisions in picking up road and rail route to reach a destina-
tion. Human autonomously decides whether to attend a social event or not based on its
interactions with its friends. Transport system (section 5) consists of trains, stations, a
train inquiry system, a transport manager and one station master for each station. Sta-
tion master tracks ridership at its station and asks transport manager to take actions if
ridership increases beyond expectation. Additionally, it regulates arrival of trains at a
station when number of trains waiting to arrive at station is more than number of avail-
able platforms. Transport manager frames train schedule and routes such that waiting
time of humans at stations is as less as possible. This is standard operations research
problem and not part of this paper.
Agent-based architecture facilitates these individual behaviors to different actors in
the simulation and manages interactions among them. Our framework models a social
network composed of humans and friend relationships amongst them. Humans post
updates on social network about their activities, interests and their plans (for example,
attending some social events) and quickly spread information about events; these posts
may in turn influence more humans. Using these posts, the transport manager extracts
information about events which are going to be held in near future - their locations and
timings. The transport manager also analyzes these posts to find out humans who plan
to attend these events in order to estimate which regions of city are expected to face
increase in demand of public transportation. This approach enables transport manager
to foresee human movement and take measures to meet increase in ridership of trains
to avoid making humans wait for trains at stations. This gives more reaction time than
online prediction techniques suggested in [9], [26], [1], [3]. Since, the posts provide
knowledge of travel plans of humans, it increases the probability of accuracy in traffic
prediction at different stations.
This paper is organized as follows. Section 2 presents current state of related work.
Section 3 presents a module that generates events with same temporal and geographical
distribution as in real-life Singapore. Section 4 presents how we model humans, social
network and how a human influences its connection. Section 5 presents the components
of transport system. Finally, we present results in section 6 and conclude the paper in
following section.
2 Related Work
Several models of simulations of traffic and transport systems have been studied and
implemented focusing on finding the regions which could see traffic jams or congestion
in near future. We divide these approaches in two categories: (i) traffic simulation; and
(ii) traffic prediction.
2.1 Traffic Simulation
TRANSIMS [21] is an integrated set of tools developed to conduct regional transporta-
tion system analyses. This system models population with daily activities and their
travel movements on a detailed represented transportation network. SUMO [2] simu-
lates road network traffic. It models collision free vehicle movements on road and safe
4
Deepika Pathania and Kamalakar Karlapalem
junctions with traffic lights. DYNASMART [8] is a time-based microscopic approach
for modeling traffic flow and evaluating overall network performance under real-time
information systems. It simulates individual vehicles and each vehicle chooses its path
from k-shortest paths according to its characteristics. It can model traffic flow dynamics
such as congestion formation and studying different control strategies. It also provides
route guidance, transport network planning and traffic operation decisions. TraMas [4]
is a multi-agent based approach in road traffic simulation. It includes three layers - co-
operative, decision and control. In this system, each cross-section has a traffic agent
to control traffic at that location by cooperating with other agents. [14] presents a sys-
tem consisting of mobile human agents whose travel behavior is modeled close to real
life, i.e., strategies for human agents are activity scheduling, route choice, reaction to
congestion and choice of residence or workplace.
2.2 Traffic Prediction
Real-time sensing of traffic on road, statistical methods based on past history and sim-
ulation based techniques have been studied for prediction of traffic congestion.
Social network based approach [29] presents a study of use of social media in finding
solution to traffic congestion problem happening during golden week (Sep 30th - Oct
7th) in China. Golden week and travel were found as highly popular topics in social net-
working sites. Based on analysis of travel related topics on social media for the month
of September, authors present how on-line users' attributes and published content can
be employed to forecast the geographic distribution of potential tourists. Another ex-
ample of investigating social media data to predict traffic congestion is studied in [7].
Authors extract tweet semantics and predict traffic congestion by incorporating seman-
tics in auto-regression model. STAR-CITY (Semantic Traffic Analytics and Reasoning
for City) [12] is an innovative system for diagnosis/prediction of traffic conditions and
efficient urban planning which semantically analyzes heterogeneous data from various
sources like historical data, present road/weather information, public transport stream,
social media feeds and city events. Our work provides a framework for implementing
different approaches for analyzing data of social media and extracting events; in addi-
tion it also provides a way of experimenting with strategies that can control congestion
and improve traffic management.
Conventional sensing approach This approach deals with keeping track of count of
vehicles moving on segment of road and estimates their speed which is used for detec-
tion of congestion areas. Conventional sensors like inductive loop detectors, magnetic
sensors, video image processors, etc. are used for detection [22]. Image processing on
images captured by traffic cameras is another example of detecting state of conges-
tion [9]. GPS technology has been used for tracking movements of vehicles on road
[26]. Vehicles equipped with GPS systems form a vehicular network. They communi-
cate within the network and collect information like speed, position and direction of
moving from other participating vehicles. This information helps to detect the state of
congestion areas and further helps to avoid congestion.
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
5
Sensor based approaches require installation and maintenance of different equip-
ments for sensing. These approaches are suitable in predicting traffic of only near time
giving transport management less time to take actions.
Statistical approach Statistical methods of traffic prediction learns about the traffic
conditions from the past traffic patterns. Some of these methods are based on neural
networks [27] and Bayesian networks [28]. Clustering trajectories [5], [13] is another
way to find out routes with high traffic load. Regression analysis [17] has also been
proposed to predict future motion of individuals.
3 Event Broadcaster
Social events like spot sale or a celebrity function can occur and are identified at short
notice before their actual occurrence. Their information is published on websites, in
newspapers, popularized on social networks, etc. Humans interact with these media on
regular basis and come to know about such events. We encapsulate these media into a
single module called event broadcaster. In our simulation, humans poll this broadcaster
to know about upcoming events.
For representing a real life behavior, an event broadcaster needs to address questions
like where to create event; how often to create events; and when to start broadcasting
information about an event. We analyzed tweets from Singapore of one week to get
locational and temporal distribution of events using the work done in [18] and NLTK
[15].
Events temporal distribution: In [18], authors present a system, ET, to extract events
from tweets. ET's event extraction mechanism is based on fact that if a particular key-
word shows a sudden significant increase in its frequency over time, then it can be
inferred that some trend / event involving that keyword has started. ET takes in a set
of tweets as input and divides them in fixed-length time intervals. Each tweet is tok-
enized into words and stop words are trimmed out. Each pair of remaining consecutive
words is taken as a bi-gram and constitute a single keyword. ET extracts frequent key-
words in tweets for each interval. It trims out those keywords that are frequent in all
intervals. Keywords that show a sudden increase in their frequencies across consecutive
time blocks are considered as event representative keywords. Event representative key-
words are grouped together using hierarchical agglomerative clustering technique. The
similarity score between two keywords is defined using common co-occurring textual
features and similarity in their appearance patterns. Each resulting cluster represents
one event. ET detects events that range from small events like car accidents to big
events like golden globe awards.
Broadcast about an event in simulation starts at time interval of first occurrence of
the cluster of related event representative keywords. We use NLTK [15] library to ex-
tract precise time of event from related tweets. We select the most commonly occurring
time range as the start and end time of the event.
6
Deepika Pathania and Kamalakar Karlapalem
Events locational information: We use NLTK [15] library to extract location infor-
mation about tweets. Given a piece of text, it returns a variety of named entities like
PERSON, LOCATION, DATE, etc. For each event that was extracted; or more pre-
cisely for each cluster of event representative keywords; we select tweets that contained
them. We feed these tweets into NLTK and extract location. Next, we pick the most
commonly occurring location in these tweets as location of our event. We feed this
location to Google's Geocoding API and get equivalent latitude and longitude [6].
Determining intended age group for event: Events are often intended towards a partic-
ular age group of humans. For example, while a science exhibition is mainly organized
for students, wedding ceremonies usually involve humans of all ages. Due to privacy
issues, Twitter's public API does not provide access to both tweets and age together.
Hence, in our current implementation, we randomly select an age range for the event
and broadcast that as part of event information.
In this paper, our focus is not on approach to determine social events and influences
in them. As solutions and accuracy of solutions increase, these solutions can be incor-
porated as pluggable modules. Our focus is that our idea is viable and we can use this
for better traffic management.
Table 1: Distribution of human categories
Distribution
Category
Working-professional
Student
Home-maker
Senior-citizen
40%
30%
15%
15%
4 Humans and their social network
Humans are modeled as proactive agents. We divide humans into four different cate-
gories: working-professional, student, home-maker and senior-citizen; distribution given
in Table 1. Humans makes different type of trips based on their categories as summa-
rized in Table 2.
For example, senior citizen, home-other/none, 7:00-9:00 indicates that between 7
to 9, a senior citizen will either go from home to some other place or not leave home at
all.
We use OSRM (Open Source Routing Machine) [16] - a high performance routing
engine - for computing shortest routes in road network. It is incorporated in simulation
as a separate service. Each human queries it to find its route from its source to train
station and from train station to destination. We keep maximum speed of human on
roads as 35 km/hr - this is to incorporate the fact that humans can use a cab or bus to
reach train stations at proper time.
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
7
Table 2: Trips based on human categories
Category
Trip
Trip starts between
Working-professional
home-office
Student
Home-maker
office-restaurant
restaurant-office
office-home
home-school
school-home
home-other
other-home
home-shop
shop-home
home-shop/none
shop-home
Senior citizen
home-other/none
other-home
home-other/none
other-home
7:30-9:30
12:00-13:00
12:30-13:30
17:30-20:00
7:00-8:00
13:30-14:30
18:30-20:30
18:30-20:30
9:00-11:00
9:30-11:30
18:00-20:00
18:30-21:00
7:00-9:00
8:30-10:00
17:00-18:30
17:30-19:30
4.1 Social network of humans
Social network consists of humans and relationships between them. It is modeled as a
directed graph - nodes indicate humans and edge indicates connections(x "follows" y)
between humans. Each human in our simulation is a part of social network. Distribution
for number of friends of a human is taken from Twitter dataset [24] which contains a
network of 100,000 humans; minimum, maximum and average number of friends are
1, 5000 and 500 respectively. We find that distribution of number of friends decreases
exponentially as shown in Figure 1. Our simulation models 100,000 humans and we
use above distribution for number of friends.
4.2 Influence Model
Influence probability is the likelihood of a human influencing interest of its connec-
tion towards a social event. We consider following factors while computing influence
probability.
Age similarity in human: A human is likely to influence connections of similar age
as there is a greater chance that they share similar interests. Also on several occasions,
events are oriented towards a certain age group of humans. Groups of age and their
distribution are presented in Table 3. We compute similarity in age of two humans A
and B using following formula (where age group is computed using Table 3):
SimilarAgeInfluence(A, B) =
1 − (age group(A) − age group(B))/6
(1)
8
Deepika Pathania and Kamalakar Karlapalem
Distribution of Numer of friends in Social Network
800
750
700
650
600
550
500
450
400
350
300
250
200
150
100
50
0
200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600 2800 3000 3200 3400 3600 3800 4000 4200 4400 4600 4800
Number of friends
Fig. 1: Distribution of friends from Twitter dataset
Table 3: Age Groups
Age interval Group No. Population percentage
0-14
15-24
25-40
41-54
55-64
≥65
1
2
3
4
5
6
13.6
18.2
25.1
25.0
09.9
08.1
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
9
Similarity in class of human: Many times, events are intended for a specific class of
humans like conferences for professionals, science exhibitions for students. So we can
say that a student is likely to influence another student more than influencing a working
professional for going to a science exhibition. Re-using human classes in Table 1, we
define similarity between two humans A and B as:
SimilarClassInfluence(A, B)
= 1
= 0
if (classA == classB)
otherwise
(2)
Proximity in points of contact (homes / offices / schools): A human is likely to
influence connections that are part of its family, office co-workers, neighbours, go to
same school. Also, if the connection lives nearby, it could get an easy company to visit
the event. Hence, influence would be proportional to nearness in points of contact. We
compute proximity of two people A and B as:
Proximity(A, B) = Minimum(Distance(homeA, homeB),
Distance(officeA, officeB),
Distance(schoolA, schoolB))
(3)
If school is undefined for A or B, then we consider distance between their schools as
infinite; similarly for offices. Since A and B can have a different set of connections,
influence of A on B is not same as of B on A. So, to normalize proximity-based in-
fluence of A on its connection B, we divide Proximity(A, B) with the least proximate
connection of B. Hence, we compute proximity-based influence as follows:
LeastProximateConnection(B) =
Maximum(∀i ∈ (connections of B), proximity(i, B))
ProximityInfluence(A, B) =
1 − Proximity(A,B)/LeastProximateConnection(B)
(4)
(5)
Lastly, we combine above factors with equal weights and say that probability with
which A will successfully influence B towards an event as:
InfluenceProbability(A, B) = (
SimilarAgeInfluence(A, B)
+ SimilarClassInfluence(A, B)
+ ProximityInfluence(A, B)
)/3.0
(6)
We could have also made influence probability a constant for the network, i.e. each
human influences all connections by same probability. We present comparison in results
section 6.1 - based on these results, we found that above model mimics reality more
closely than constant influence probability.
10
Deepika Pathania and Kamalakar Karlapalem
4.3 Human decision to attend an event
On being interested to attend an event, a human may decide whether to attend it or
not. In real life this could be an outcome of many factors like distance of event from
home, financial cost of attending event, other clashing events, family obligations. In our
simulation we only consider distance factor. Human estimates time to reach the event
based on trains' schedule information available from transport manager (see section
5.2). A human considers to be "on time" if it can reach to the event within ±τ minutes
of the start-time, where τ is 1/10th of duration of event. If human finds that it can reach
to the event on time then it confirms its decision to attend it otherwise there will be no
way to attend the event even if they wanted.
4.4 Event information injection in social network
In day-to-day life, humans check for new interesting information on internet several
times a day; also they do it after certain time intervals and not check it continuously.
In the simulation, humans poll EventBroadcaster (section 3) with certain probability at
regular intervals to get details about upcoming events. Upon receiving list of events, a
human selects those events which matches its age-group. It further selects those events
where it can reach before event's start time; and tweets about selected events.
4.5 Information Diffusion using influence probability model
We now present how information about an event is spread among the connections in
social network and how the interest towards it is aroused in a human using influence
probability model defined in Section 4.2. Intuitively, interest of a human towards an
event increases as more and more of its connections in social network present their
desire to attend the event through their posts. We try to capture this motivation. We
use independent cascade model studied in [10][11] for spreading of interest towards an
event in the social network.
As mentioned earlier, we model social network as a graph where nodes represent
humans and edges represent connections. Whenever a human affirms to attend an event
and posts about it, we say that the corresponding node in the graph is active for that
event. Initially, human nodes who injected this information (section 4.4) are marked
active and rest of the nodes as inactive with respect to this event. When a node becomes
active, it tries to influence and activate its inactive neighbors (friends) towards the event.
It succeeds with certain probability which we call influence probability (section 4.2) and
is distinct for every edge. On successful activation, neighbor further propagates event
information. Hence, each active node tries to activate its neighbors and the process
continues in discrete steps till no new node is activated, i.e., there is no node which
changed from being inactive to active in the current step.
5 Transport System
In this section, we describe components that make up transport system. We use Singa-
pore's metro train network as an example to describe how we select parameters.
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
11
5.1 Station Master
In real life, a train station needs to be managed with respect to how and when trains can
arrive. Also, during certain time periods ridership of trains may increase and if the trains
are not frequent then it can cause stations to get congested. The station management
may ask the city transport system to provide appropriate measures like start new trains,
increase capacity of existing trains, etc. In order to know the state of each train station
and communicate it with city transport system, we appoint a station master agent in
each station.
Tracking ridership at station: Framework implements a simple mechanism for track-
ing number of humans present at a station at a given time. When a human arrives at a
station (either from a train or from outside), it tells station master its destination station
and asks for a token. When human leaves station or boards a train, it returns the token to
station master. Thus, the number of tokens that are not returned gives number of people
present at the station.
Deciding train arrival at station: Every station has a fixed number of platforms as-
sociated with it. So at a given time only a limited number of trains can arrive and halt
at the station. Station master keeps a count of trains that are currently standing at the
station. When a train is about to arrive, it first asks station master if there is a plat-
form available for it; if not, then train halts at its position. Station master keeps track of
all trains waiting to arrive. Whenever a platform is free, station master picks that train
which has halted for the longest time.
5.2 Transport manager
In real life, every transport system is managed by a committee which lays out plans
for train-routes, decides schedules and frequency of trains, etc. In our simulation, we
capture this notion through transport manager. Transport manager decides schedule and
capacity of trains to meet changes in ridership of trains. Its main motive is to reduce the
amount of time humans wait for trains at stations at as less as possible. It also provides
train inquiry system (section 5.3) with most recent schedule of trains.
Estimating ridership: In order to be able to frame good strategies for schedule and
route of trains, the transport manager needs to have a good estimation of ridership at
various stations during various hours of the day. Transport manager actively monitors
ridership at various stations by communicating with station masters and keeps track of
it for future estimations. It also tries to estimate ridership based on upcoming social
events. It continuously polls event broadcaster (section 3) about schedule and location
of upcoming events. For each event, it extracts data from social network about humans
that are planning to visit. Based on humans' locations and the time of event it estimates
change in ridership of train routes using procedure in Table 4. Sum of regular ridership
(as obtained from station masters) and the estimated change due to events gives final
ridership.
12
Deepika Pathania and Kamalakar Karlapalem
Construct a directed graph with train stations as nodes and train routes as edges.
for every hour of next 24 hours:
for every event in this hour:
Find location of the event.
Find train station nearest to event location and map it on the graph as destination.
for each human attending this event:
Estimate whether human will come from home / school / office based on its category
and time of day.
Find location of human based on above estimation.
Find train station nearest to human location and map it on the graph as source for
above destination.
for each train route:
Traverse the route and count number of sources that come before their destination.
Mark above count as change in ridership of this train route for this hour.
Table 4: Procedure to estimate change in ridership
Capacity of trains: Capacity of trains is defined as number of seats in it. Initially, each
train has same capacity; transport manager may strategize to change it as and when
needed. Defining strategies to change capacity is a plug-and-play module for users of
our framework. An important feature of our simulated population is that each human
rides trains at least once a day and most humans rides multiple times - this is not true
in real-life where a significant number of people do not ride trains or use trains less
often. Hence, capacity of trains is chosen based on ratio of ridership of Singapore's
metro trains in real-life to ridership in our simulation. From [20], the daily ridership is
2.3 million; and each train has capacity of 1920 ([23]). In our simulation, we have a
ridership of 370,000 on a regular day. Hence, we compute initial capacity of trains as:
370, 000 ∗ 1920/2, 300, 000 = 310
Schedule of trains: We initialize train schedule using Singapore's metro schedule
taken from [19]. Again strategizing and testing schedules is a pluggable feature of our
framework. In order to evaluate efficiency of a schedule (or capacity), our framework
computes two metrics - train-usage and average waiting-time - and presents its results
to the users.
Train usage: Schedules should be economical; for example, one may decide to increase
frequency of trains on certain routes in order to reduce waiting times; but this will lead
to increase in cost of running the trains and there may be a large number of seats that
stay unoccupied for large part of train trips. Train usage expresses above rationale. For
a given time interval T , it is defined as:
UT = (Pi tiT )/(C ∗ T )
where tiT is total time for which seat i was occupied (need not be for contiguous
duration) in time interval T , and C is capacity of the train.
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
13
Average waiting time: Average waiting time of humans for given time interval T is
defined as:
WT = Pi tiT /N
where tiT is total time for which a human waited for train at any station during time
interval T and N is total number of humans.
5.3 Train inquiry system
In real life, it is a common practice that humans inquire about train schedule on phones,
metro-websites, on bulletin boards at train stations, etc. We encapsulate all sources of
train related information into a single module which we call train inquiry system. Train
inquiry system continuously polls transport manager for getting most updated schedule
of trains. It also continuously polls station masters to know about current location and
delays of trains. Humans query this system for train information and plan their trips
accordingly.
CC − without event (left), with event (middle) and using congestion control strategy (right)
EW − without event (left), with event (middle) and using congestion control strategy (right)
3.5
3
2.5
1.5
2
1
e
g
a
s
U
n
i
a
r
T
r0
r1
r2
r3
r4
3.5
3
2.5
e
g
a
s
U
n
i
a
r
T
1.5
2
1
0.5
0
6AM
7AM
8AM
9AM 10AM 11AM 12PM 1PM
6PM
Time Intervals of 1 hour
2PM
3PM
4PM
5PM
0.5
0
6AM
7AM
8AM
7PM
8PM
9PM 10PM
9AM 10AM 11AM 12PM 1PM
6PM
Time Intervals of 1 hour
2PM
3PM
4PM
5PM
r0
r1
r2
r3
r4
7PM
8PM
9PM 10PM
Fig. 2: Comparison of Train Usage across CC and EW train lines
6 Results
We simulate train transport system of Singapore with a population of 100,000 humans.
We simulate four major metro lines of Singapore rail network - North South Line (NS),
North East Line (NE), East West Line (EW) and Circular Line (CC) with a total of
87 train stations ([19]). Trains have capacity for 310 (section 5.2) and follow real-life
schedule.
We divide results on validity of framework design and parameters based on follow-
ing aspects.
14
Deepika Pathania and Kamalakar Karlapalem
Comparing information spread in different events using influence model (left) and constant influence probability (right)
d
e
c
n
e
u
l
f
n
i
y
l
e
v
i
t
i
s
o
p
e
l
p
o
e
p
f
o
r
e
b
m
u
N
4500
4000
3500
3000
2500
2000
1500
1000
500
0
Working−professional
student
home−maker
senior−citizen
Student Event
Working−Professional Event
Common Event
Fig. 3: Comparing humans influenced positively for events
Types of Events
6.1 Influence probability model
We simulated three events: a science exhibition, a professional conference and a wed-
ding event and studied the distribution of human in terms of both age and class that
were influenced positively towards these events. Results showed that using our influence
probability model (section 4.2), ratio of students that attended science exhibition was
much higher than other class of humans; similar results were obtained for age-groups
(see Fig. 3). Next, we made influence probability constant for each connection and set
it to 0.5. Using constant probability model, we get almost same ratio of humans, both
in terms of class and age groups, that attended science exhibition. On the other hand,
for wedding we see a similar distribution of various age-groups attending the event in
both cases. Figure 3 shows the comparisons. Hence, we see that per-connection varied
influence probability model mimics reality closely and performs better than constant
influence probability in this aspect.
CC − without event (left), with event (middle) and using congestion control strategy (right)
EW − without event (left), with event (middle) and using congestion control strategy (right)
)
s
e
t
u
n
i
m
n
i
(
e
m
i
T
t
i
a
W
60
50
40
30
20
10
0
r0
r1
r2
r3
r4
6AM
7AM
8AM
9AM 10AM 11AM 12PM 1PM
2PM
3PM
4PM
5PM
6PM
7PM
8PM
9PM 10PM
Time Intervals of 1 hour
60
)
s
e
t
u
n
i
m
n
i
(
e
m
i
T
t
i
a
W
50
40
30
20
10
0
r0
r1
r2
r3
r4
6AM
7AM
8AM
9AM 10AM 11AM 12PM 1PM
2PM
3PM
4PM
5PM
6PM
7PM
8PM
9PM 10PM
Time Intervals of 1 hour
Fig. 4: Comparison of average waiting time across CC and EW train lines
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
15
6.2 Effect of events
We ran two simulations - with and without a social event. Event was held near stations
EW21 and CC22 between 8:30 to 11:30 am. In order to get fine-grained look at the
usage and wait-time metrics, we divide each train-line into 5 sections - r0, r1, r2, r3, r4
- each section has equal number of consecutive stations. For a given time interval, usage
of a section of train line is summation of usage of trains that cross this section. Similarly,
for a given time interval, waiting time of a section of train line is average of wait times
for humans standing on stations in that section of train line. Comparison of usage of
trains and average wait-time are shown in figures 2, 4. We see an increased usage on
EW and CC lines during event hours. NE and NS lines also see a small increase but
results are omitted due to space limitations. Also, we see that train usage is higher in
middle sections of lines rather than tail stations (irrespective of events) which matches
real-life behavior.
6.3 Effect of congestion control strategies
We implemented a greedy congestion control strategy as follows. Transport manager
monitored upcoming events and estimated riderships at the beginning of every hour. If
it saw that ridership increased beyond capacity of trains, then it increased capacity of
train by adding a compartment. Transport manager has only a fixed number of com-
partments and each compartment seats 31 humans. So, if transport manager is out of
compartments, it detaches the compartment from a train whose ridership has decreased.
It also assigns compartment in order of decreasing ridership of affected trains. With this
strategy, Figure 4 shows that wait-time was reduced significantly (as compared to fixed
capacity model). Though the effective train usage was slightly decreased 2 as transport
manager detaches compartment only when it needs.
6.4 Strategies involving human behavior
We implemented a strategy in which human alters its trip based on state of ridership
in trains. Initially, a human selects route which takes shortest total travel time to reach
destination. Human keeps polling station master to get the current state of ridership
of trains. If a human perceives that train involved in its route is full, then it computes
alternative routes to destination and selects one with next minimum total travel time.
Total travel time is defined as sum of time spent in traveling from source to destination
including time taken to go to and fro train stations and wait time at station. Waiting
at the same train station for next train running on original route is also considered a
valid alternate route. We incorporated human behavior exercising alternative routing
in simulation and found that 10% of humans choose to take alternates. Table 5 shows
comparison of average waiting time and average total trip time in experiment with or
without using alternative routing approach. It is noted that there is slight difference in
both of metrics.
16
Deepika Pathania and Kamalakar Karlapalem
Table 5: Alternative Routing Approach
Without
With
alternate routing alternate routing
Average waiting time
Average total travel time
34.3 mins
82.8 mins
31.7 mins
79.5 mins
7 Conclusion
Traffic congestion wastes time and money; and needs to be controlled effectively. Social
networks can provide ample information about upcoming events and locational distri-
bution of people planning to attend such events. We presented basic building blocks
of a framework that simulates train transport system of a city, human behavior and a
social network connecting them; and aims to capture real-life like behavior. It provides
researchers a platform to further experiment in this area easily. Future work includes
incorporating buses and cars as well as modes of transportation; and documenting and
open-sourcing the framework for public use.
References
1. Y. Ando, Y. Fukazawa, H. Iwasaki, and S. Honiden. Performance of pheromone model for
predicting traffic congestion. In AAMAS, 2006.
2. M. Behrisch, L. Bieker, J. Erdmann, and D. Krajzewicz. Sumo: Simulation of urban mobility.
In The Third International Conference on Advances in System Simulation - SIMUL, October
2011.
3. M. Ben-Akiva, M. Bierlaire, H. Koutsopoulos, and R. Mishalani. Dynamit: a simulation-
based system for traffic prediction. In DACCORD Short Term Forecasting Workshop, Febru-
ary 1998.
4. J. M. A. Fernandes and E. Olivei. Traffic control through behaviour-based multi-agent sys-
In International Conference on the Practical Application of Intelligent Agents and
tem.
Multi-Agent Technology, 1999.
5. S. Gaffney and P. Smyth. Trajectory clustering with mixtures of regression models.
In
Proceedings of the 5th ACM International Conference on Knowledge Discovery and Data
Mining (SIGKDD), San Diego, CA, 1999.
6. Google geocoding. https://developers.google.com/maps/documentation/geocoding/.
7. J. He, W. Shen, P. Divakaruni, L. Wynter, and R. Lawrence. Improving traffic prediction
with tweet semantics. In International Joint Conference on Artificial Intelligence, 2013.
8. T. Y. Hu, R. W. Rothery, and H. Mahmassani. Dynasmart: Dynamic network assignment-
In Center for Transportation Research,
simulation model for advanced road telematics.
1992.
9. V. Jain, A. Sharma, and L. Subramaniam. Road traffic congestion in the developing world.
In roceedings of the 2nd ACM Symposium on Computing for Development, 2012.
10. D. Kempe, J. Kleinberg, and E. Tardos. Maximizing the spread of influence through a social
network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge
Discovery and Data Mining, pages 137 -- 146, New York, USA, 2003.
11. D. Kempe, J. Kleinberg, and E. Tardos. Influential nodes in a diffusion model for social ne-
towrks. In Proc. 32nd International Colloquium on Automata, Languages and Programming
(ICALP), 2005.
Exploiting Near Time Forecasting From Social Network To Decongest Traffic
17
12. F. Lecue, S. Tallevi-Diotallevi, J. Hayes, R. Tucker, V. Bicer, M. Sbodio, and P. Tommasi.
Star-city: semantic traffic analytics and reasoning for city. In International conference on
Intelligent User Interfaces, 2014.
13. J. Lee, J. Han, and K. Whang. Trajectory clustering: A partition-and-group framework. In
Proceedings of the SIGMOD Conference, Beijing, China, 2009.
14. K. Nagel and F. Marchal. Computational methods for multi-agent simulations of travel be-
haviory. In Tenth International Conference on Travel Behaviour Research, Lucerne, August
2003.
15. Natural language toolkit. http://www.nltk.org/.
16. Open
routing machine.
http://project-osrm.org/.
source
http://map.project-osrm.org/ and
17. R. K. Oswald, W. T. Scherer, and B. L. Smith. Traffic flow forecasting using approximate
nearest neighbor nonparametric regression. In In Final project of ITS Center project: Traffic
forecasting: non-parametric regressions, December 2000.
18. R. Parikh and K. Karlapalem. ET: events from tweets. In Proceedings of the 22nd interna-
tional conference on World Wide Web companion, pages 613 -- 620, 2013.
19. Singapore metro website. http://www.smrt.com.sg.
20. Singapore land transport authority report. http://www.lta.gov.sg/content/dam/ltaweb/corp/PublicationsResearch/files/FactsandFigures/Stats_in_Brief_2013.pdf,
2013.
21. L. Smith, R. Beckman, K. Nagel, and M. Williams. Proceedings of the 5th national trans-
portation planning methods applications conference. In Proceedings of international confer-
ence on World Wide Web companion, 1995.
22. Traffic
detector
handbook
by
u.s.
department
of
transportation.
http://www.fhwa.dot.gov/publications/research/operation.
23. Siemens train information. http://en.wikipedia.org/wiki/Siemens_C651.
24. Twitter dataset for friends of users.
http://kevinchai.net/datasets and
http://strict.dista.uninsubria.it/?p=364.
25. Urban
mobility
report
by
texas
transportation
institute.
http://mobility.tamu.edu/ums, 2012.
26. S. A. Vaqar and O. Basir. Traffic pattern detection in a partially deployed vehicular ad-hoc
network of vehicles. In Appears in Wireless Communications, IEEE, December 2009.
27. Vics(vehicle
information
and
communication
system).
http://www.vics.or.jp/english/index.htm.
28. T. Yoshii, H. Akahane, and M. Kuwahara. Impacts of the accuracy of traffic information in
dynamics route guidance systems. In Proc. Of the 3rd Annual World Congress on Intelligent
Transport Systems, 1996.
29. K. Zeng, W. Liu, X. Wang, and S. Chen. Traffic congestion and social media in china. In
IEEE Intelligent Systems, vol. 28, pages 72 -- 77, 2013.
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.