1. Tutorial at CVPR 2011, Colorado Springs, CO, June 20, 2010
Web page: http://www.dsi.unive.it/~atorsell/cvpr2011tutorial
2. Course Outline
Part I (08:00 – 10:00) − M. Pelillo
Introduction to game theory
Polymatrix games, contextual pattern recognition (relaxation labeling), and
graph transduction
Evolutionary games and data clustering (dominant sets)
Coffee break
Part II (10:30 – 12:30) − A. Torsello
Applications
Repeated games and online learning
4. What is Game Theory?
“Game theory is to games of strategy what probability
theory is to games of chance. And just as probability
theory far transcends its role as the logical basis of rational
gambling, so does game theory transcende its original
guise as the logical basis of parlor games.”
Anatol Rapoport
Two-Person Game Theory (1966)
“The central problem of game theory was posed by von
Neumann as early as 1926 in Göttingen. It is the following:
If n players, P1,…, Pn, play a given game Γ, how must the ith
player, Pi, play to achieve the most favorable result for himself?”
Harold W. Kuhn
Lectures on the Theory of Games (1953)
5. A Few Cornerstones in Game Theory
1921−1928: Emile Borel and John von Neumann give the first modern formulation of a
mixed strategy along with the idea of finding minimax solutions of normal-form games.
1944, 1947: John von Neumann and Oskar Morgenstern publish Theory of Games and
Economic Behavior.
1950−1953: In four papers John Nash made seminal contributions to both non-cooperative
game theory and to bargaining theory.
1972−1982: John Maynard Smith applies game theory to biological problems thereby
founding “evolutionary game theory.”
late 1990’s −: Development of algorithmic game theory…
Nobel prizes in economics awarded to game theorists:
1994: John Nash, John Harsanyi and Reinhard Selten: "for their pioneering analysis of
equilibria in the theory of non-cooperative games”
2005: Robert Aumann and Thomas Schelling: "for having enhanced our understanding of
conflict and cooperation through game-theory analysis”
2007: Leonid Hurwicz, Eric Maskin and Roger Myerson: “for having laid the foundations of
mechanism design theory”
6. Normal-form Games
We shall focus on finite, non-cooperative, simultaneous-move games in
normal form, which are characterized by:
A set of players: I = {1, 2, …, n} (n ≥ 2)
A set of pure strategy profiles: S = S1 × S2 × … × Sn where each Si = {1, 2,
…, mi} is the (finite) set of pure strategies (actions) available to the player i
A payoff function: π : S ℜn, π(s) = (π1(s),…,πn(s)), where πi(s) (i=1…n)
represents the “payoff” (or utility) that player i receives when strategy
profile s is played
Each player is to choose one element from his strategy space in the absence of
knowledge of the choices of the other players, and “payments” will be made
to them according to the function πi(s).
Players’ goal is to maximize their own returns.
7. Two Players
In the case of two players, payoffs can be represented as two m1 x m2
matrices (say, A for player 1 and B for player 2):
A = (ahk ) ahk = π 1 (h,k)
€ B = (bhk )
€ bhk = π 2 (h,k)
Special cases:
€ €
Zero-sum games: A + B = 0 (ahk = −bhk for all h and k)
Symmetric games: BT = A
Doubly-symmetric games: A = AT = BT
9. Example 2:
Battle of the Sexes
Wife
Soccer Ballet
Soccer 2,1 0,0
Husband
Ballet 0,0 1,2
10. Example 3:
Rock-Scissors-Paper
You
Rock Scissors Paper
Rock 0,0 1 , -1 -1 , 1
Me Scissors -1 , 1 0,0 1 , -1
Paper 1 , -1 -1 , 1 0,0
11. Mixed Strategies
A mixed strategy for player i is a probability distribution over his set Si of
pure strategies, which is a point in the (mi-1)-dimensional standard
simplex:
mi
Δ i = x i ∈ R : ∀h = 1…mi : x ih ≥ 0, and ∑ x ih = 1
mi
h=1
The set of pure strategies that is assigned positive probability by mixed
strategy xi Δi is called the support of xi:
€
σ (x i ) = {h ∈ Si : x ih 0}
A mixed strategy profile is a vector x = (x1,…,xn) where each component
€
xi ∆i is a mixed strategy for player i I.
The mixed strategy space is the multi-simplex = ∆1 × ∆2 × … × ∆n
12. Standard Simplices
mi = 2 mi = 3
Note: Corners of standard simplex correspond to pure strategies.
13. Mixed-Strategy Payoff Functions
In the standard approach, all players’ randomizations are assumed to be
independent.
Hence, the probability that a pure strategy profile s = (s1,…,sn) will be
used when a mixed-strategy profile x is played is:
n
x(s) = ∏ x isi
i=1
and the expected value of the payoff to player i is:
€ ui (x) = ∑ x i (s)π i (s)
s∈S
In the special case of two-players games, one gets:
m1 m 2 m1 m 2
€ u2 (x) = ∑ ∑ x1h bhk x 2k = x1 Bx 2
T
u1 (x) = ∑ ∑ x1h ahk x 2k = x1 Ax 2
T
h=1 k=1 h=1 k=1
where A and B are the payoff matrices of players 1 and 2, respectively.
€ €
14. Dominance Relations
Notational shortcut.
Here, and in the sequel, if z and xi ∆i, the notation (xi,z−i) stands for the
strategy profile in which player i I plays strategy xi, while all other players
play according to z.
A strategy yi ∆i is said to weakly dominate another strategy xi ∆i if the first
strategy never earns a lower payoff than the second (and sometimes earns a
higher payoff).
Formally:
ui (y i ,z−i ) ≥ ui (x i ,z−i ) for all z ∈ Θ
with strict inequality for some z .
€
Strategy yi is said to strictly dominate xi if:
ui (y i ,z−i ) ui (x i ,z−i ) for all z ∈ Θ
16. Best Replies
Player i‘s best reply to the strategy profile x−i is a mixed strategy xi* ∆i such
that
ui(xi*,x−i) ≥ ui(xi,x−i)
for all strategies xi ∆i.
The best reply is not necessarily unique. Indeed, except in the extreme case
in which there is a unique best reply that is a pure strategy, the number of
best replies is always infinite.
Indeed:
When the support of a best reply x* includes two or more pure strategies,
any mixture of these strategies must also be a best reply
Similarly, if there are two pure strategies that are individually best replies,
any mixture of the two is necessarily also a best reply
17. Nash Equilibrium
The Nash equilibrium concept is motivated by the idea that a theory of
rational decision-making should not be a self-destroying prophecy that
creates an incentive to deviate for those who believe it.
A strategy profile x is a Nash equilibrium if it is a best reply to itself,
namely, if:
ui(xi,x−i) ≥ ui(zi,x−i)
for all i = 1…n and all strategies zi ∆i.
If strict inequalities hold for all zi ≠ xi then x is said to be a strict Nash
equilibrium.
Theorem. A strategy profile x is a Nash equilibrium if and only if for
every player i I, every pure strategy in the support of xi is a best reply to x−i.
It follows that every pure strategy in the support of any player’s equilibrium
mixed strategy yields that player the same payoff.
18. Finding Pure-strategy Nash Equilibria
Player 2
Left Middle Right
Top 3,1 2,3 10 , 2
High 4,5 3,0 6,4
Player 1
Low 2,2 5,4 12 , 3
Bottom 5,6 4,5 9,7
19. Multiple Equilibria in Pure Strategies
Wife
Soccer Ballet
Soccer 2,1 0,0
Husband
Ballet 0,0 1,2
20. Nash equilibrium!
You
Rock Scissors Paper
Rock 0,0 1 , -1 -1 , 1
Me Scissors -1 , 1 0,0 1 , -1
Paper 1 , -1 -1 , 1 0,0
21. Existence of Nash Equilibria
Theorem (Nash, 1951). Every finite normal-form game admits a mixed-
strategy Nash equilibrium.
Idea of proof.
1. Define a continuous map T on such that the fixed points of T are in
one-to-one correspondence with Nash equilibria.
2. Use Brouwer’s theorem to prove existence of a fixed point.
Note. For symmetric games, Nash proved that there always exists a
symmetric Nash equilibrium, namely a Nash equilibrium where all players
play the same (possibly mixed) strategy.
22. Finding Nash Equilibria:
Two-Player Zero-Sum Games
“As far as I can see, there could be no theory of
games […] without that theorem […]
I thought there was nothing worth publishing until
the Minimax Theorem was proved”
John von Neumann (1928)
Minimax Theorem (von Neumann, 1928). In any finite, two-player, zero-
sum game (with payoff matrices A and –A) in any Nash equilibrium each
player receives a payoff that is equal to both its maximim and his minimax
value:
max min x T Ay = min max x T Ay .
x y y x
Solved by standard linear programming methods!
€
23. The Complexity of
Finding Nash Equilibria
“Together with factoring, the complexity of finding a Nash
equilibrium is in my opinion the most important concrete
open question on the boundary of P today.”
Christos Papadimitriou
Algorithms, games, and the internet (2001)
At present, no known reduction exists from our problem to a decision problem
that is NP-complete, nor has our problem been shown to be easier.
An intuitive stumbling block is that every game has at least one Nash
equilibrium, whereas known NP-complete problems are expressible in terms
of decision problems that do not always have solutions.
24. Variations on Theme
Theorem (Gilboa and Zemel, 1989). The following are NP-complete problems,
even for symmetric games.
Given a two-player game in normal form, does it have:
1. at least two Nash equilibria?
2. a Nash equilibrium in which player 1 has payoff at least a given amount?
3. a Nash equilibrium in which the two players have a total payoff at least a
given amount?
4. a Nash equilibrium with support of size greater than a give number?
5. a Nash equilibrium whose support contains a given strategy?
6. a Nash equilibrium whose support does not contain a given strategy?
7. etc.
25. The Complexity Class PPAD
(Polynomial Parity Argument on Directed graphs)
Consider the class of (directed) graphs characterized by:
1. A finite but exponentially large set of vertices
2. Each vertex has indegree and outdegree at most one
3. Given a string, it is computationaly simple to (a) tell if it is a vertex of the
graph, and if so to (b) find its neighbors, and to (c) tell which one is the
predecessor and/or which one is the successor
4. There is one known source (the “standard source”)
PPAD is the class of all problems
whose solution space can be set up
as the set of all sinks and all
nonstandard sources in a directed
graph with the properties displayed
above.
26. NASH is PPAD-complete
Theorem (Daskalakis et al. 2005; Chen and Deng, 2005, 2006). The problem
of finding a sample Nash equilibrium of a general-sum finite game with two
or more players is PPAD-complete.
As of today, PPAD includes some 25 problems from such different areas as:
Fixed point analysis
Economics and market equilibria
Topology
Combinatorics and graph theory
Social networks
See: http://www.cc.gatech.edu/~kintali/ppad.html
It is not known whether or not P = PPAD. However, it is generally believed
that the two classes are not equivalent.
The common belief is that, in the worst case, computing a sample Nash
equilibrium will take time that is exponential in the size of the game.
27. Algorithms for Finding a Nash Equilibrium
1. Linear-complementarity approach (Lemke and Howson, 1964)
Only two players; uses pivoting; it may take exponentially many steps.
2. Simplicial subdivision (van der Laan et al., 1987)
Approximate a fixed point of a function by triangulating the multi-simplex
and traversing the triangulation along a fixed path.
3. Continuation methods (Govindan and Wilson, 2003)
Perturb a game to one that has a known equilibrium, and then trace the
solution back to the original game.
4. Enumeration of support (Mangasarian, 1964; Dickaut and Kaplan, 1991;
Porter, E. Nudelman and Y. Shoham, 2008)
Enumerate all possible supports and solve a feasibity program; equipped
with simple heuristics, outperform more sophisticated algorithms.
5. Evolutionary dynamics (Taylor and Jonker, 1978; Rota Bulò and Bomze, 2010)
Inspired by population-based interpretations of game theory.
28. References
Texts on (classical) game theory
J. von Neumann and O. Morgerstern. Theory of Games and Economic Behavior. Princeton
University Press (1944, 1953).
D. Fudenberg and J. Tirole. Game Theory. MIT Press (1991).
R. Myerson. Game Theory: Analysis of Conflict. Harvard University Press (1991).
M. J. Osborne and A. Rubinstein. A Course in Game Theory. MIT Press (1994).
Computationally-oriented texts
N. Nisan, T. Roughgarden, E. Tardos, and V. Vazirani (Eds.) Algorithmic Game Theory.
Cambridge University Press (2007).
Y. Shoham and K. Leyton-Brown. Multiagent Systems: Algorithmic, Game-Theoretic, and
Logical Foundations. Cambridge University Press (2009).
On-line resources
http://gambit.sourceforge.net/ a library of game-theoretic algorithms
http://gamut.stanford.edu/ a suite of game generators for testing game algorithms
30. Succinct Games
Describing a game in normal form entails listing all payoffs for all players and
strategy combinations. In a game with n players, each facing m pure strategies,
one need to store nmn numbers!
A succinct game (or a succinctly representable game) is a game which may be
represented in a size much smaller than its normal-form representation.
Examples.
Graphical games. The payoffs of each player depends on the actions of very few
(at most d) other players. The number of payoffs needed to describe this game is
nmd+1.
Sparse games. Most of the payoffs are zero (special case of graphical games).
Symmetric games. All players are identical, so in evaluating the payoff of a
combination of strategies, all that matters is how many of the n players play each
of the s strategies.
31. Polymatrix Games
A polymatrix game (a.k.a. multimatrix game) is a non-cooperative game in
which the relative influence of the selection of a pure strategy by any one
player on the payoff to any other player is always the same, regardless of
what the rest of the players do.
Formally:
There are n players each of whom can use m pure strategies
For each pair (i,j) of players there is an m x m payoff matrix Aij
The payoff of player i for the strategy combination s1,…,sn is given by
ui (s1,,sn ) = ∑ Asiji s j
j≠i
The number of payoff values required to represent such a game is O(n2m2).
The problem of finding a Nash equilibrium in a polymatrix game is PPAD-
€
complete.
34. What do you see?
From: A. Oliva and A. Torralba, “The role of context in object recognition”, Trends in Cognitive Sciences, 2007.
35. Context and the Brain
From: M. Bar, “Visual objects in context”, Nature Reviews Neuroscience, August 2004.
36. The (Consistent) Labeling Problem
A labeling problem involves:
A set of n objects B = {b1,…,bn}
A set of m labels = {1,…,m}
The goal is to label each object of B with a label of .
To this end, two sources of information are exploited:
Local measurements which capture the salient features of each object
viewed in isolation
Contextual information, expressed in terms of a real-valued n2 x m2
matrix of compatibility coefficients R = {rij( , )}.
The coefficient rij( , ) measures the strenght of compatibility between the
two hypotheses: “bi is labeled ” and “bj is labeled “.
37. Relaxation Labeling Processes
The initial local measurements are assumed to provide, for each object bi B,
an m-dimensional (probability) vector:
T
p(0) = ( p(0) (1),, p(0) (m))
i i i
with pi(0)( ) ≥ 0 and ∑ pi(0)( ) = 1. Each pi(0)( ) represents the initial, non-
contextual degree of confidence in the hypothesis “bi is labeled ”.
€
By concatenating vectors p1(0),…,pn(0) one obtains an (initial) weighted labeling
assignment p(0) ℜnm.
€
The space of weighted labeling assignments is
IK = Δ Δ
×
×…
m times
where each ∆ is the standard simplex of ℜn. Vertices of IK represent
unambiguous labeling assignments
€
A relaxation labeling process takes the initial labeling assignment p(0) as input
and iteratively updates it taking into account the compatibility model R.
38. Relaxation Labeling Processes
In a now classic 1976 paper, Rosenfeld, Hummel, and Zucker introduced
heuristically the following update rule (assuming a non-negative
compatibility matrix):
(t +1) p(t ) ( λ )q(t ) ( λ )
p i ( λ) = i (t ) i (t )
∑ pi (µ)qi (µ)
µ
where
€
€ q ( λ) = ∑ ∑ rij (λ, µ) p(t ) (µ)
(t )
i i
j µ
quantifies the support that context gives at time t to the hypothesis “bi is
labeled with label ”.
€
See (Pelillo, 1997) for a rigorous derivation of this rule in the context of a
formal theory of consistency.
39. Applications
Since their introduction in the mid-1970’s relaxation labeling algorithms
have found applications in virtually all problems in computer vision and
pattern recognition:
Edge and curve detection and enhancement
Region-based segmentation
Stereo matching
Shape and object recognition
Grouping and perceptual organization
€
Graph matching
Handwriting interpretation
…
Further, intriguing similarities exist between relaxation labeling processes
and certain mechanisms in the early stages of biological visual systems (see
Zucker, Dobbins and Iverson, 1989, for physiological and anatomical
evidence).
40. Hummel and Zucker’s Consistency
In 1983, Bob Hummel and Steve Zucker developed an elegant theory of
consistency in labeling problem.
By analogy with the unambiguous case, which is easily understood, they
define a weighted labeling assignment p IK consistent if:
∑ p (λ)q (λ) ≥ ∑ v (λ)q (λ)
i i i i i = 1…n
λ λ
for all labeling assignments v IK.
€
If strict inequalities hold for all v ≠ p, then p is said to be strictly consistent.
€
Geometrical interpretation.
The support vector q points away
from all tangent vectors at p (it has
null projection in IK).
41. Characterizations
Theorem (Hummel and Zucker, 1983). A labeling p IK is consistent if and
only if, for all i = 1…n, the following conditions hold:
1. qi( ) = ci whenever pi( ) 0
2. qi( ) ≤ ci whenever pi( ) = 0
for some constants c1…cn.
The “average local consistency” of a labeling p IK is defined as:
€
A( p) = ∑ ∑ pi ( λ )qi ( λ )
i λ
Theorem (Hummel and Zucker, 1983). If the compatibility matrix R is
symmetric, i.e., rij( , )=rji( , ), then any local maximizer p IK of A is
€
consistent.
42. Understanding the “1976-rule”
Using the Baum-Eagon inequality it is easy to prove the following result,
concerning the original Rosenfeld-Hummel-Zucker (RHZ) update rule.
Theorem (Pelillo, 1997). The RHZ relaxation operator is a “growth
transformation” for the average local consistency A, provided that
compatibility coefficients are symmetric. In other words, the algorithm
strictly increases the average local consistency on each iteration, i.e.,
A(p(t+1)) A(p(t))
€
for t = 0,1,… until a fixed point is reached.
Theorem (Elfving and Eklundh, 1982; Pelillo, 1997). Let p IK be a strictly
consistent labeling. Then p is an asymptotically stable equilibrium point for
the RHZ relaxation scheme, whether or not the compatibility matrix is
symmetric.
43. Relaxation Labeling and
Polymatrix Games
As observed by Miller and Zucker (1991) the consistent labeling problem is
equivalent to a polymatrix game.
Indeed, in such formulation we have:
Objects = players
Labels = pure strategies
Weighted labeling assignments = mixed strategies
Compatibility coefficients = payoffs
€
and:
Consistent labeling = Nash equilibrium
Strictly consistent labeling = strict Nash equilibrium
Further, the RHZ update rule corresponds to discrete-time multi-population
“replicator dynamics” used in evolutionary game theory (see next part).
44. Graph Transduction
Given a set of data points grouped into:
labeled data: {(x1 , y1 ), . . . , x , y )}
unlabeled data: {x +1 , . . . , xn } n
Express data as a graph G=(V,E)
V : nodes representing labeled and unlabeled points
E : pairwise edges between nodes weighted by the similarity between the
corresponding pairs of points
Goal: Propagate the information available at the labeled nodes to unlabeled ones
in a “consistent” way.
Cluster assumption:
The data form distinct clusters
Two points in the same cluster are expected to be in the same class
45. A Special Case:
Unweighted Undirected Graphs
A simple case of graph transduction in which the graph G is an unweighted
undirected graph:
An edge denotes perfect similarity between points
The adjacency matrix of G is a 0/1 matrix
The cluster assumption: Each node in a connected component of the graph
should have the same class label.
46. A Special Case:
Unweighted Undirected Graphs
This toy problem can be formulated as a (binary) constraint satisfaction problem
(CSP) as follows:
The set of variables: V = {v1, …, vn}
{y } for all 1 ! i ! l
$ i
Domains: Dvi = #
$ Y
% for all l +1 ! i ! n
Binary constraints: i,j: if aij = 1, then vi = vj ! 10 $
e.g. for a 2-class problem Rij = #
0 1 %
Each assignment of values to the variables satisfying all the constraints is a
solution of the CSP, thereby providing a consistent labeling for the unlabeled
points.
Goal: Generalize to real-valued (soft) constraints
Idea: Use consistency criterion of relaxation labeling (= Nash equilibrium)
47. The Graph Transduction Game
Assume:
the players participating in the game correspond to the vertices of the
graph
the set of strategies available to each player denote the possible
hypotheses about its class membership
- labeled players I = {I |1 , . . . , I |c }
- unlabeled players Iu
Labeled players choose their strategies at the outset.
– each player i ∈ Il|k always play its kth pure strategy.
The transduction game is in fact played among the unlabeled players to choose
their memberships according.
By assuming that only pairwise interactions are allowed, we obtain a polymatrix
game that can be solved used standard relaxation labeling / replicator algorithms.
48. In short…
Graph transduction can be formulated as a (polymatrix) non-cooperative game
(i.e., a consistent labeling problem).
The proposed game-theoretic framework can cope with symmetric, negative
and asymmetric similarities (none of the existing techniques is able to deal with
all three types of similarities).
Experimental results on standard datasets show that our approach is not only
more general but also competitive with standard approaches.
A. Erdem and M. Pelillo. Graph transduction as a non-cooperative
game. GbR 2011 (journal version under review).
49. References
J. T. Howson. Equilibria of polymatrix games. Management Science (1972).
A. Rosenfeld, R. A. Hummel, and S. W. Zucker. Scene labeling by relaxation operations.
IEEE Trans. Syst. Man. Cybern (1976).
R. A. Hummel and S. W. Zucker. On the foundations of relaxation labeling processes.
IEEE Trans. Pattern Anal. Machine Intell. (1983).
M. Pelillo. The dynamics of nonlinear relaxation labeling processes. J. Math. Imaging
and Vision (1997).
J. Kittler and J. Illingworth. Relaxation labeling algorithms – A review. Image and Vision
Computing (1986).
D. A. Miller and S. W. Zucker. Copositive-plus Lemke algorithm solves polymatrix
games. Operation Research Letters (1991).
D. A. Miller and S. W. Zucker. Efficient simplex-like methods for equilibria of
nonsymmetric analog networks. Neural Computation (1992).
S. W. Zucker. Relaxation labeling: 25 years and still iterating. In: L. S. Davis (Ed.)
Foundations of Image Understanding. Kluwer Academic Publisher (2001).
A. Erdem and M. Pelillo. Graph transduction as a non-cooperative game. GbR 2011
(journal version under review).
51. Evolution and the Theory of Games
“We repeat most emphatically that our theory is thoroughly static.
A dynamic theory would unquestionably be more complete and
therefore preferable.
But there is ample evidence from other branches of science that it
is futile to try to build one as long as the static side is not
thoroughly understood.”
John von Neumann and Oskar Morgenstern
Theory of Games and Economic Behavior (1944)
“Paradoxically, it has turned out that game theory is more readily
applied to biology than to the field of economic behaviour for
which it was originally designed.”
John Maynard Smith
Evolution and the Theory of Games (1982)
52. Evolutionary Games
Introduced by John Maynard Smith (1973, 1974, 1982) to model the evolution
of behavior in animal conflicts.
Assumptions:
A large population of individuals belonging to the same species which
compete for a particular limited resource
This kind of conflict is modeled as a symmetric two-player game, the
players being pairs of randomly selected population members
Players do not behave “rationally” but act according to a pre-programmed
behavioral pattern (pure strategy)
Reproduction is assumed to be asexual
Utility is measured in terms of Darwinian fitness, or reproductive success
53. Evolutionary Stability
A strategy is evolutionary stable if it is resistant to invasion by new strategies.
Formally, assume:
A small group of “invaders” appears in a large populations of individuals,
all of whom are pre-programmed to play strategy x ∆
Let y ∆ be the strategy played by the invaders
Let ε be the share of invaders in the (post-entry) population (0 ε 1)
The payoff in a match in this bimorphic population is the same as in a match
with an individual playing mixed strategy:
w = εy + (1 – ε)x ∈ ∆
hence, the (post-entry) payoffs got by the incumbent and the mutant strategies
are u(x,w) and u(y,w), respectively.
54. Evolutionary Stable Strategies
Definition. A strategy x ∆ is said to be an evolutionary stable strategy (ESS)
if for all y ∆–{x} there exists δ (0,1), such that for all ε (0, δ) we have:
u[x, εy + (1 – ε)x] u[y, εy + (1 – ε)x]
Theorem. A strategy x Δ is an ESS if and only if it meets the following
first- and second-order best-reply conditions:
1. u(y,x) ≤ u(x,x) for all y ∆
2. u(y,x) = u(x,x) u(y,y) u(x,y) for all y ∆–{x}
Note. From the conditions above, we have:
∆ESS ∆NE
If x ∆ is a strict Nash equilibrium, then x is an ESS
55. Existence of ESS’s
Unlike Nash equilibria existence of ESS’s is not guaranteed.
Unique Nash equilibrium is x=(1/3,1/3,1/3)T
Hence, all y ∆ are best replies to x
Let the “mutant” be y=(1,0,0)T You
But u(y,y) = u(x,y), hence ∆ESS = Ø
Rock Scissors Paper
Rock 0,0 1 , -1 -1 , 1
Scissors -1 , 1 0,0 1 , -1
Me
Paper 1 , -1 -1 , 1 0,0
56. Complexity Issues
Two questions of computational complexity naturally present themselves:
What is the complexity of determining whether a given game has an ESS
(and of finding one)?
What is the complexity of recognizing whether a given x is an ESS for a
given game?
Theorem (Etessami and Lochbihler, 2004). Determining whether a given
two-player symmetric game has an ESS is both NP-hard and coNP-hard.
Theorem (Nisan, 2006). Determining whether a (mixed) strategy x is an ESS
of a given two-player symmetric game is coNP-hard.
57. Replicator Dynamics
Let xi(t) the population share playing pure strategy i at time t. The state of the
population at time t is: x(t)= (x1(t),…,xn(t)) ∆.
Replicator dynamics (Taylor and Jonker, 1978) are motivated by Darwin’s
principle of natural selection:
˙
xi
∝ payoff of pure strategy i − average population payoff
xi
which yields:
€ x i = x i [ u(e i , x) − u(x, x)]
˙
= x i [(Ax) i − x T Ax ]
Notes.
Invariant under positive affine transformations of payoffs (i.e., u u+ ,
with 0) €
Standard simplex ∆ is invariant under replicator dynamics, namely, x(0) ∆
x(t) ∆, for all t 0 (so is its interior and boundary)
58. Replicator Dynamics and
Dominated Strategies
Theorem (Akin, 1980). Let i be a strictly dominated pure strategy of a two-
player symmetric game.
If x ∆ is the limit point of a replicator dynamics trajectory starting from int(∆),
then xi = 0.
Theorem (Samuelson and Zhang, 1992). Let i be an iteratively strictly
dominated pure strategy of a two-player symmetric game.
If x ∆ is the limit point of a replicator dynamics trajectory starting from int(∆),
then xi = 0.
59. Replicator Dynamics and
Nash Equilibria
Proposition. If x ∆ is a Nash equilibrium, then x is an equilibrium point of
replicator dynamics.
Note. The contrary need not be true (e.g., all corners of ∆ are equilibria).
Theorem (Bomze, 1986). If x ∆ is (Lyapunov) stable under replicator dynamics,
then x is a Nash equilibrium.
Theorem (Nachbar, 1990). A point x ∆ is a Nash equilibrium if and only if x is
the limit point of a replicator dynamics trajectory starting from int(∆).
60. Replicator Dynamics and ESS’s
Theorem (Taylor and Jonker, 1978). If x ∆ is an ESS, then it is an asymptotically
stable equilibrium point for the replicator dynamics.
The opposite need not be true.
0 6 −4
A = −3 0 5
−1 3 0
The point m=(1/3,1/3,1/3)T is asymptotically
stable (its eigenvalues have negative parts).
€
But e1=(1,0,0)T is an ESS.
Hence m cannot be an ESS (being in the interior,
it would have to be the unique ESS).
Proposition. If x int(∆) is an ESS, then all replicator trajectories starting from
int(∆) will eventually converge to x.
61. In summary …
Let x ∆ be a mixed strategy of a two-player symmetric game.
Then:
x ESS x asymptotically stable
x asymptotically stable x Nash
x Nash x stationary
More formally:
∆ESS ∆Asymp. Stable ∆NE ∆Stationary
62. Doubly Symmetric Games
In a doubly symmetric (or partnership) game, the payoff matrix A is symmetric
(A = AT).
Fundamental Theorem of Natural Selection (Losert and Akin, 1983).
For any doubly symmetric game, the average population payoff ƒ(x) = xTAx is
strictly increasing along any non-constant trajectory of replicator dynamics,
namely, d/dtƒ(x(t)) ≥ 0 for all t ≥ 0, with equality if and only if x(t) is a
stationary point.
Characterization of ESS’s (Hofbauer and Sigmund, 1988)
For any doubly simmetric game with payoff matrix A, the following statements
are equivalent:
a) x ∆ESS
b) x ∆ is a strict local maximizer of ƒ(x) = xTAx over the standard simplex ∆
c) x ∆ is asymptotically stable in the replicator dynamics
63. Discrete-time Replicator Dynamics
A well-known discretization of replicator dynamics, which assumes non-
overlapping generations, is the following (assuming a non-negative A):
A( x(t)) i
x i (t + 1) = x i (t)
x(t)T Ax(t)
which inherits most of the dynamical properties of its continuous-time
counterpart (e.g., the fundamental theorem of natural selection).
€
MATLAB implementation
68. One-class Clustering
“[…] in certain real-world problems, natural groupings are found
among only on a small subset of the data, while the rest of the data
shows little or no clustering tendencies.
In such situations it is often more important to cluster a small
subset of the data very well, rather than optimizing a clustering
criterion over all the data points, particularly in application
scenarios where a large amount of noisy data is encountered.”
G. Gupta and J. Ghosh. Bregman bubble clustering: A robust framework
for mining dense cluster. ACM Trans. Knowl. Discov. Data (2008).
70. The Need for Overlapping Clusters
Partitional approaches impose that each element cannot belong to more than one
cluster. There are a variety of important applications, however, where this
requirement is too restrictive.
Examples:
clustering micro-array gene expression data
clustering documents into topic categories
perceptual grouping
segmentation of images with transparent surfaces
References:
N. Jardine and R. Sibson. The construction of hierarchic and non-hierarchic
classifications. Computer Journal, 11:177–184, 1968
A. Banerjee, C. Krumpelman, S. Basu, R. J. Mooney, and J. Ghosh. Model-
based overlapping clustering. KDD 2005.
K. A. Heller and Z. Ghahramani. A nonparametric Bayesian approach to
modeling overlapping clusters. AISTATS 2007.
72. «In most visual fields the contents of particular areas “belong together” as
circumscribed units from which their surrounding are excluded.»
W. Köhler, Gestalt Psychology (1947)
«In gestalt theory the word “Gestalt” means any segregated whole.»
W. Köhler (1929)
73. «In fact, the concept “Gestalt” may be applied far
beyond the limits of sensory experience.
According to the most general functional definition
of the term, the processes of learning, of recall, of
striving, of emotional attitude, of thinking, acting
and so forth, may have to be included.»
W. Köhler, Gestalt Psychology (1947)
74. «The process of discovery is akin to the recognition of
shapes as analysed by Gestalt psychology.»
Michael Polanyi
Science, Faith and Society (1946)
«In my opinion every discovery of a complex
regularity comes into being through the function of
gestalt perception.»
Konrad Lorenz (1959)
76. By answering the question “what is a cluster?” we get a novel way of
looking at the clustering problem.
Clustering_old(V,A,k)
V1,V2,...,Vk - My_favorite_partitioning_algorithm(V,A,k)
return V1,V2,...,Vk
−−−−−−
Clustering_new(V,A)
V1,V2,...,Vk - Enumerate_all_clusters(V,A)
return V1,V2,...,Vk
Enumerate_all_clusters(V,A)
repeat
Extract_a_cluster(V,A)
until all clusters have been found
return the clusters found
77. Suppose the similarity matrix is a binary (0/1) matrix.
Given an unweighted undirected graph G=(V,E):
A clique is a subset of mutually adjacent vertices
A maximal clique is a clique that is not contained in a larger one
In the 0/1 case, a meaningful notion of a cluster is that of a maximal clique.
NCut
New approach
78. No need to know the number of clusters in advance (since we extract
them sequentially)
Leaves clutter elements unassigned (useful, e.g., in figure/ground
separation or one-class clustering problems)
Allows extracting overlapping clusters
Need a partition?
Partition_into_clusters(V,A)
repeat
Extract_a_cluster
remove it from V
until all vertices have been clustered
79. «Similarity has been viewed by both philosophers and psychologists
as a prime example of a symmetric relation. Indeed, the assumption
of symmetry underlies essentially all theoretical treatments of
similarity.
Contrary to this tradition, the present paper provides empirical
evidence for asymmetric similarities and argues that similarity should
not be treated as a symmetric relation.»
Amos Tversky
“Features of similarities,” Psychol. Rev. (1977)
Examples of asymmetric (dis)similarities
Kullback-Leibler divergence
Directed Hausdorff distance
Tversky’s contrast model
80. ESS’s as Clusters
We claim that ESS’s abstract well the main characteristics of a cluster:
Internal coherency: High mutual support of all elements within the
group.
External incoherency: Low support from elements of the group to
elements outside the group.
81. Basic Definitions
Let S V be a non-empty subset of vertices, and i S.
The (average) weighted degree of i w.r.t. S is defined as:
1
awdeg S (i) = ∑ aij
| S | j ∈S
Moreover, if j ∉ S, we define:
€
i j
φ S (i, j) = aij − awdeg S (i)
S
Intuitively, S(i,j) measures the similarity between vertices j and i, with
€
respect to the (average) similarity between vertex i and its neighbors in S.
82. Assigning Weights to Vertices
Let S V be a non-empty subset of vertices, and i S.
The weight of i w.r.t. S is defined as:
1 if S = 1
w S (i) = ∑ φ ( j,i)w S−{ i} ( j) otherwise
S−{ i}
j ∈S−{ i}
S-{i}
€ Further, the total weight of S is defined as:
j
i
W (S) = ∑ w S (i)
i∈S
S
€
83. Interpretation
Intuitively, wS(i) gives us a measure of the overall (relative) similarity between
vertex i and the vertices of S-{i} with respect to the overall similarity among the
vertices in S-{i}.
w{1,2,3,4}(1) 0 w{1,2,3,4}(1) 0
84. Dominant Sets
Definition (Pavan and Pelillo, 2003, 2007). A non-empty subset of vertices S
V such that W(T) 0 for any non-empty T S, is said to be a dominant set if:
1. wS(i) 0, for all i S (internal homogeneity)
2. wS {i}(i) 0, for all i ∉ S (external homogeneity)
Dominant sets clusters
The set {1,2,3} is dominant.
85. The Clustering Game
Consider the following “clustering game.”
Assume a preexisting set of objects O and a (possibly asymmetric) matrix
of affinities A between the elements of O.
Two players with complete knowledge of the setup play by simultaneously
selecting an element of O.
After both have shown their choice, each player receives a payoff,
monetary or otherwise, proportional to the affinity that the chosen element
has with respect to the element chosen by the opponent.
Clearly, it is in each player’s interest to pick an element that is strongly
supported by the elements that the adversary is likely to choose.
Hence, in the (pairwise) clustering game:
There are 2 players
The objects to be clustered are the pure strategies
The (null-diagonal) affinity matrix coincides with the similarity matrix
86. Dominant Sets are ESS’s
Dominant-set clustering
To get a single dominant-set cluster use, e.g., replicator dynamics (but see
Rota Bulò, Pelillo and Bomze, CVIU in press, for faster dynamics)
To get a partition use a simple peel-off strategy: iteratively find a dominant
set and remove it from the graph, until all vertices have been clustered
To get overlapping clusters, enumerate dominant sets (see Bomze, 1992;
Torsello, Rota Bulò and Pelillo, 2008)
87. Special Case:
Symmetric Affinities
Given a symmetric real-valued matrix A (with null diagonal), consider the
following Standard Quadratic Programming problem (StQP):
maximize ƒ(x) = xTAx
subject to x ∆
Note. The function ƒ(x) provides a measure of cohesiveness of a cluster (see
Pavan and Pelillo, 2003, 2007; Sarkar and Boyer, 1998; Perona and Freeman,
1998).
ESS’s are in one-to-one correspondence
to (strict) local solutions of StQP
Note. In the 0/1 (symmetric) case, ESS’s are in one-to-one correspondence to
(strictly) maximal cliques (Motzkin-Straus theorem).
88. Measuring the Degree of Cluster
Membership
The components of the converged vector give us a measure of the participation of
the corresponding vertices in the cluster, while the value of the objective function
provides of the cohesiveness of the cluster.
90. In a nutshell…
The dominant-set (ESS) approach:
makes no assumption on the underlying (individual) data representation
makes no assumption on the structure of the affinity matrix, being it able to
work with asymmetric and even negative similarity functions
does not require a priori knowledge on the number of clusters (since it extracts
them sequentially)
leaves clutter elements unassigned (useful, e.g., in figure/ground separation or
one-class clustering problems)
allows principled ways of assigning out-of-sample items (NIPS’04)
allows extracting overlapping clusters (ICPR’08)
generalizes naturally to hypergraph clustering problems, i.e., in the presence
of high-order affinities, in which case the clustering game is played by more
than two players (NIPS’09)
extends to hierarchical clustering (ICCV’03: EMMCVPR’09)
91. References
Evolutionary game theory
J. Weibull. Evolutionary Game Theory. MIT Press (1995).
J. Hofbauer and K. Sigmund. Evolutionary Games and Population Dynamics. Cambridge
University Press (1998).
Dominant sets
M. Pavan and M. Pelillo. A new graph-theoretic approach to clustering and segmentation.
CVPR 2003.
M. Pavan and M. Pelillo. Dominant sets and hierarchical clustering. ICCV 2003.
M. Pavan and M. Pelillo. Efficient out-of-sample extension of dominant-set clusters. NIPS 2004.
A. Torsello, S. Rota Bulò and M. Pelillo. Grouping with asymmetric affinities: A game-theoretic
perspective. CVPR 2006.
M. Pavan and M. Pelillo. Dominant sets and pairwise clustering. PAMI 2007.
A. Torsello, S. Rota Bulò and M. Pelillo. Beyond partitions: Allowing overlapping groups in
pairwise clustering. ICPR 2008.
S. Rota Bulò and M. Pelillo. A game-theoretic approach to hypergraph clustering. NIPS 2009.
M. Pelillo. What is a cluster? Perspectives from game theory. NIPS 2009 Workshop on
“Clustering: Science or Art?” (talk available on videolectures.net).
S. Rota Bulò, M. Pelillo and I. M. Bomze. Graph-based quadratic optimization: A fast
evolutionary approach. CVIU 2011.