Computational Evolutionary Game Theory and why I’m never using PowerPoint for another presentation involving maths ever again Enoch Lau – 5 September 2007.

Slides:



Advertisements
Similar presentations
News and Notes 4/13 HW 3 due now HW 4 distributed today, due Thu 4/22 Final exam is Mon May 3 11 AM Levine 101 Today: –intro to evolutionary game theory.
Advertisements

GAME THEORY.
This Segment: Computational game theory Lecture 1: Game representations, solution concepts and complexity Tuomas Sandholm Computer Science Department Carnegie.
Mixed Strategies CMPT 882 Computational Game Theory Simon Fraser University Spring 2010 Instructor: Oliver Schulte.
Effort Games and the Price of Myopia Michael Zuckerman Joint work with Yoram Bachrach and Jeff Rosenschein.
Evolution and Repeated Games D. Fudenberg (Harvard) E. Maskin (IAS, Princeton)
6.896: Topics in Algorithmic Game Theory Lecture 11 Constantinos Daskalakis.
Defender/Offender Game With Defender Learning. Classical Game Theory Hawk-Dove Game Hawk-Dove Game Evolutionary Stable Evolutionary Stable Strategy (ESS)
Congestion Games with Player- Specific Payoff Functions Igal Milchtaich, Department of Mathematics, The Hebrew University of Jerusalem, 1993 Presentation.
Infinitely Repeated Games. In an infinitely repeated game, the application of subgame perfection is different - after any possible history, the continuation.
An Introduction to... Evolutionary Game Theory
MIT and James Orlin © Game Theory 2-person 0-sum (or constant sum) game theory 2-person game theory (e.g., prisoner’s dilemma)
EC941 - Game Theory Lecture 7 Prof. Francesco Squintani
How Bad is Selfish Routing? By Tim Roughgarden Eva Tardos Presented by Alex Kogan.
Learning in games Vincent Conitzer
What is a game?. Game: a contest between players with rules to determine a winner. Strategy: a long term plan of action designed to achieve a particular.
Algorithms and Economics of Networks Abraham Flaxman and Vahab Mirrokni, Microsoft Research.
Maynard Smith Revisited: Spatial Mobility and Limited Resources Shaping Population Dynamics and Evolutionary Stable Strategies Pedro Ribeiro de Andrade.
An Introduction to Game Theory Part I: Strategic Games
Equilibrium Concepts in Two Player Games Kevin Byrnes Department of Applied Mathematics & Statistics.
A camper awakens to the growl of a hungry bear and sees his friend putting on a pair of running shoes, “You can’t outrun a bear,” scoffs the camper. His.
An Introduction to Game Theory Part II: Mixed and Correlated Strategies Bernhard Nebel.
Oblivious Routing for the L p -norm Matthias Englert Harald Räcke 1.
Lecture 1 - Introduction 1.  Introduction to Game Theory  Basic Game Theory Examples  Strategic Games  More Game Theory Examples  Equilibrium  Mixed.
Algorithmic and Economic Aspects of Networks Nicole Immorlica.
Evolutionary Game Theory
Beyond selfish routing: Network Formation Games. Network Formation Games NFGs model the various ways in which selfish agents might create/use networks.
1 Computing Nash Equilibrium Presenter: Yishay Mansour.
Evolutionary Game Theory Amit Bahl CIS620. Outline zEGT versus CGT zEvolutionary Stable Strategies Concepts and Examples zReplicator Dynamics Concepts.
Near-Optimal Network Design with Selfish Agents By Elliot Anshelevich, Anirban Dasgupta, Eva Tardos, Tom Wexler STOC’03 Presented by Mustafa Suleyman CIFTCI.
Potential games, Congestion games Computational game theory Spring 2010 Adapting slides by Michal Feldman TexPoint fonts used in EMF. Read the TexPoint.
DANSS Colloquium By Prof. Danny Dolev Presented by Rica Gonen
Network Formation Games. Netwok Formation Games NFGs model distinct ways in which selfish agents might create and evaluate networks We’ll see two models:
Price of Anarchy Bounds Price of Anarchy Convergence Based on Slides by Amir Epstein and by Svetlana Olonetsky Modified/Corrupted by Michal Feldman and.
Inefficiency of equilibria, and potential games Computational game theory Spring 2008 Michal Feldman.
1 Issues on the border of economics and computation נושאים בגבול כלכלה וחישוב Congestion Games, Potential Games and Price of Anarchy Liad Blumrosen ©
CPS Learning in games Vincent Conitzer
Learning dynamics,genetic algorithms,and corporate takeovers Thomas H. Noe,Lynn Pi.
Evolutionary Game Theory. Game Theory Von Neumann & Morgenstern (1953) Studying economic behavior Maynard Smith & Price (1973) Why are animal conflicts.
Tom Wenseleers Dept. of Biology, K.U.Leuven
5. Alternative Approaches. Strategic Bahavior in Business and Econ 1. Introduction 2. Individual Decision Making 3. Basic Topics in Game Theory 4. The.
Derivative Action Learning in Games Review of: J. Shamma and G. Arslan, “Dynamic Fictitious Play, Dynamic Gradient Play, and Distributed Convergence to.
Learning in Multiagent systems
1 Economics & Evolution Number 3. 2 The replicator dynamics (in general)
6.853: Topics in Algorithmic Game Theory Fall 2011 Constantinos Daskalakis Lecture 11.
Standard and Extended Form Games A Lesson in Multiagent System Based on Jose Vidal’s book Fundamentals of Multiagent Systems Henry Hexmoor, SIUC.
Boltzmann Machine (BM) (§6.4) Hopfield model + hidden nodes + simulated annealing BM Architecture –a set of visible nodes: nodes can be accessed from outside.
Game Theory in Wireless and Communication Networks: Theory, Models, and Applications Lecture 4 Evolutional Game Zhu Han, Dusit Niyato, Walid Saad, Tamer.
Beyond selfish routing: Network Games. Network Games NGs model the various ways in which selfish agents strategically interact in using a network They.
Beyond selfish routing: Network Games. Network Games NGs model the various ways in which selfish users (i.e., players) strategically interact in using.
1 What is Game Theory About? r Analysis of situations where conflict of interests is present r Goal is to prescribe how conflicts can be resolved 2 2 r.
1. 2 You should know by now… u The security level of a strategy for a player is the minimum payoff regardless of what strategy his opponent uses. u A.
Vincent Conitzer CPS Learning in games Vincent Conitzer
The Standard Genetic Algorithm Start with a “population” of “individuals” Rank these individuals according to their “fitness” Select pairs of individuals.
Replicator Dynamics. Nash makes sense (arguably) if… -Uber-rational -Calculating.
The Price of Routing Unsplittable Flow Yossi Azar Joint work with B. Awerbuch and A. Epstein.
Network Formation Games. NFGs model distinct ways in which selfish agents might create and evaluate networks We’ll see two models: Global Connection Game.
Application of Game Theory to Biology A long-standing puzzle in the study of animal behavior is the prevalence of conventional fights in the competition.
Theory of Computational Complexity M1 Takao Inoshita Iwama & Ito Lab Graduate School of Informatics, Kyoto University.
Network Formation Games. NFGs model distinct ways in which selfish agents might create and evaluate networks We’ll see two models: Global Connection Game.
Week 6 Applications of ODEs to the evolution game theory
Q 2.1 Nash Equilibrium Ben
Game Theory and Cooperation
Replicator Dynamics.
Presented By Aaron Roth
Oliver Schulte Petra Berenbrink Simon Fraser University
Vincent Conitzer Learning in games Vincent Conitzer
The Price of Routing Unsplittable Flow
Vincent Conitzer CPS Learning in games Vincent Conitzer
Week 6 Applications of ODEs to the evolution game theory
Presentation transcript:

Computational Evolutionary Game Theory and why I’m never using PowerPoint for another presentation involving maths ever again Enoch Lau – 5 September 2007

Outline  What is evolutionary game theory?  Why evolutionary game theory?  Evolutionary game theory concepts  Computational complexity of evolutionary stable strategies  Evolutionary game theory and selfish routing  Evolutionary game theory over graphs  Selection strategies  Finite populations 2

What is evolutionary game theory? Not creationism game theory 3

Evolutionary game theory (EGT)  An infinite number of agents in 2-player symmetric games  Payoffs calculate a fitness used for replication or imitation  Similarities with conventional game theory  Both concerned with the decisions made by agents in a game  Equilibria are important concepts for both  Differences from conventional game theory  Rationality of agents not assumed  Strategies selected by some force (evolution, cultural factors)  Higher fitness means more (asexual) reproduction  Other assumptions: complete mixing 4

Approaches to evolutionary game theory 5  Two approaches 1. Evolutionary stable strategy: derives from work of Maynard Smith and Price 2. Properties of evolutionary dynamics by looking at frequencies of change in strategies

Evolutionary stable strategy (ESS) 6  Incumbents and mutants in the population  ESS is a strategy that cannot be invaded by a mutant population  In an ESS, mutants have lower fitness (reproductive success) compared with the incumbent population  ESS is more restrictive than a Nash equilibrium  Not all 2-player, symmetric games have an ESS  Assumptions very important:  If we have a finite number of players, instead of an infinite number, different ESS

Evolutionary stable strategy (ESS) Finite population simulations on the Hawk-Dove game 7

History 8  First developed by R.A. Fisher in The Genetic Theory of Natural Selection (1930)  Attempted to explain the sex ratio in mammals  Why is there gender balance in animals where most males don’t reproduce?  R.C. Lewontin explicitly applied game theory in Evolution and the Theory of Games (1961)  Widespread use since The Logic of Animal Conflict (1973) by Maynard Smith and Price  Seminal text: Evolution and the Theory of Games (1984) by Maynard Smith

Example: hawks & doves  Two organisms fighting over a resource, worth V  Hawks: will fight for the resource, fighting costs C  Doves: will retreat from aggressive hawks, share resource with other doves  Example payoff matrix:  Nash equilibrium and ESS given by mixed strategy of (7/12, 5/12) 9 HD H-2550 D015

Why evolutionary game theory? Why not? 10

Equilibrium selection problem  Problems with using Nash equilibria:  Not all games have pure Nash equilibria  Prisoner’s Dilemma: sub-optimality of equilibria  Multiple Nash equilibria  How to choose between different Nash equilibria?  Introduce refinements to the concept of Nash equilibria  Then how to choose between refinements? 11

Hyper-rational agents  Humans sometimes prefer A to B, B to C, and C to A  EGT can predict behaviour of animals, where strong rationality assumptions fail  EGT better able to handle weaker rationality assumptions? 12

Lack of dynamical theory  Traditional game theory, which is static, lacks the dynamics of rational deliberation  Could use extensive form (tree form) instead of normal form  Quickly becomes unmanageable  Presupposes hyper-rational agents  Will not learn from observing opponent’s behaviour 13

Philosophical problems  Objections to EGT, mainly from application to human subjects  Measure of fitness in cultural evolutionary interpretations  Explanatory irrelevance of evolutionary game theory  Does EGT simply reinforce existing values and biases?  EGT does not provide sufficient evidence for the origin of phenomena  Historical records more useful? 14

Evolutionary game theory concepts This is where your head is meant to start hurting 15

Classical model 16  Infinite population of organisms  Each organism assumed equally likely to interact with each other organism  Fixed, 2-player, symmetric game  Fitness function F  A is set of actions  ∆(A) is set of probability distributions  F: ∆(A) x ∆(A)  R  F(s|t) = fitness of s playing t  ε proportion are mutants, 1 – ε are incumbents

Evolutionary stable strategy 17  s is an incumbent, t is a mutant  Expected fitness of an incumbent: (1 - ε ) F(s|s) + ε F(s|t)  Expected fitness of mutant: (1 - ε ) F(t|s) + ε F(t|t)  s is an ESS if there exists an ε t such that for all 0 fitness of mutant  Implies: 1. F(s|s) > F(t|s), or 2. F(s|s) = F(t|s) and F(s|t) > F(t|t)  A strategy s is an ESS for a 2-player, symmetric game given by a fitness function F, iff (s, s) is a Nash equilibrium of F, and for every best response t to s, t ≠ s, F(s|t) > F(t|t)

Example: hawks & doves 18  Generalised payoff matrix:  Note that (D, D) is not a Nash equilibrium  Cannot be an ESS either  If V > C:  H is an ESS  If V ≤ C:  Mixed strategy: H with prob V/C, D with prob 1 – V/C is ESS HD H(V – C) / 2V D0V / 2

Example: hawks & doves Map of proportions for Hawk-Dove game. Note that where the curve meets the straight line at a gradient of less than 1 (the middle point), that is a stable equilibrium. Where it meets it at a gradient greater than 1, it is an unstable equilibrium. 19

Replicator dynamics 20  Continuous dynamics for EGT  Find differential equations for the change in the proportion of each strategy over time  In some cases, such as the Prisoner’s Dilemma, stable states of replicator dynamics occur when everyone in the population follows the ESS  Roughly, true when only two pure strategies exist  Can fail to be true with more than two pure strategies

Example: Prisoner’s Dilemma 21  Generalised payoff matrix  with T > R > P > S and T’ > R’ > P’ > S’  Fitness functions CD C(R, R’)(S, T’) D(T, S’)(P, P’)

Example: Prisoner’s Dilemma 22  Proportion of C and D in next generation:  where W is the overall fitness of population (weighted by proportion)  Leads to differential equations:  Use payoff matrix to show that p’ d > 0 and p’ c < 0

Computational complexity of evolutionary stable strategies No good news here 23

Results and proof outline 24  Finding an ESS is both NP-hard and coNP-hard  Reduction from the problem of checking if a graph has a maximum clique of size exactly k  Recognising whether a given strategy is an ESS is also coNP-hard  Transform a graph G into a payoff matrix F, which will have an ESS iff the size of the largest clique in G is not equal to k  Transform adjacency matrix: replace all diagonal entries with the value ½, inserting 0 th row and 0 th column with entries 1 – 1/(2k)

Proof idea 25  For a mixed strategy s to be an ESS, incumbents should receive a relatively high payoff when playing other incumbents  When s plays itself, it must guarantee that the pure strategies chosen will correspond to two adjacent vertices  Mixed strategy with support over a clique will achieve this  When max clique is greater than k, uniform mixed strategy corresponding to clique will be an ESS  When max clique is less than k, get pure strategy ESS  No ESS in the case where max clique is exactly k

Technical lemma 26  If s is a strategy with s 0 = 0, then F(s|s) ≤ 1 – 1/(2k’), where k’ is the size of the maximum clique in G. This holds with equality iff s is the uniform distribution over a k’-clique.  Proof idea  By induction over the number of non-edges between the vertices in G  Inductive step: Find two non-adjacent vertices u and v, and construct a new strategy s’ by moving the probability in s from v to u

Lemmas If C is a maximal clique in G of size k’ > k, and s is the uniform distribution on C, then s is an ESS 2. If the maximum size clique in G is of size k’ < k, then the pure strategy 0 is an ESS 3. If the maximum size clique of G is at least k, then the pure strategy 0 is not an ESS 4. If the maximum size clique of G is at most k, then any strategy for F that is not equal to the pure strategy 0, is not an ESS for F

Proof of Lemma 1 28  By technical lemma, F(s|s) = 1 – 1/(2k’)  Any best response to s must have support over only C  F(0|s) = 1 – 1/(2k) < F(s|s) by construction  Take a u not in C:  u is connected to at most k’ – 1 vertices in C (since max clique size is k’)  F(u|s) ≤ 1 – 1/k’ (sum up the entries in the payoff matrix)  F(u|s) < F(s|s)  Also by technical lemma, payoff of s is maximised when s is uniform distribution over C  Hence, s is a best response to itself

Proof of Lemma 1 29  Now, need to show that for all best responses t to s, t ≠ s, F(s|t) > F(t|t) (note: t has support over C)  By technical lemma, F(t|t) < 1 – 1/(2k’) (note: no equality here since t ≠ s)  Using F, we can show that F(s|t) = 1 – 1/(2k’) (C is a clique, s and t are distributions with support over C)  You can get this by summing up the values in the payoff matrix  (k’ – ½)/k’ = 1 – 1/(2k’)  Hence, F(s|t) > F(t|t)

Proof of Lemma 2 30  Mutant strategy t  F(t|0) = 1 – 1/(2k) = F(0|0)  0 is a best response to itself  So need to show F(0|t) > F(t|t)  Form t* by setting the probability of strategy 0 in t to zero and then renormalising  Applying the technical lemma:  F(t*|t*) ≤ 1 – 1/(2k’) < 1 – 1(2k) = F(0|t)

Proof of Lemma 2 31  Expression for F(t|t):  By expanding out expressions for F(t|t) and F(t*|t*):  F(0|t) > F(t|t) iff F(0|t) > F(t*|t*)

Evolutionary game theory and selfish routing Ah, something related to my thesis topic 32

The model 33  Each agent assumed to play an arbitrary pure strategy  Imitative dynamics – switch to lower latency path with probability proportional to difference in latencies  Recall: at a Nash flow, all s-t paths have the same latency  If we restrict the latency functions to be strictly increasing, then Nash flows are essentially ESS  Paths with below average latency will have more agents switching to them than from them  Paths with above average latency will have more agents switching from them than to them

Convergence to Nash flow 34  As t  ∞, any initial flow with support over all paths in P will eventually converge to a Nash flow  Use Lyapunov’s direct method to show that imitative dynamics converge to a Nash flow  General framework for proving that a system of differential equations converges to a stable point  Define a potential function that is defined in the neighbourhood of the stable point and vanishes at the stable point itself  Then show that the potential function decreases with time  System will not get stuck in any local minima

Convergence to approximate equilibria 35  ε -approximate equilibrium: Let P ε be the paths that have latency at least (1 + ε )l*, and let x ε be the fraction of agents using these paths. A population is at ε -approximate equilibrium iff x ε < ε  Only a small fraction of agents experience latency significantly worse than the average latency  Potential function  Measures the total latency the agents experience  Integral: sums latency if agents were inserted one at a time

Convergence to approximate equilibria 36  Theorem: the replicator dynamics converge to an ε - approximate equilibrium time O( ε -3 ln(l max /l * ))  Proof: see handout

Evolutionary game theory over graphs Did you know? I am my neighbour’s neighbour. 37

The model 38  No longer assume that two organisms are chosen uniformly at random to interact  Organisms only interact with those in their local neighbourhood, as defined by an undirected graph or network  Use:  Depending on the topology, not every mutant is affected equally  Groups of mutants with lots of internal attraction may be able to survive  Fitness given by the average of playing all neighbours

Mutant sets to contract 39  We consider an infinite family G = {G n } (where G n is a graph with n vertices)  Examine asymptotic (large n) properties  When will mutant vertex sets contract?  Let M n be the mutant subset of vertices  |M n | ≥ ε n for some constant ε > 0  M n contracts if, for sufficiently large n, for all but o(n) of the j in M n, j has an incumbent neighbour i such that F(j) < F(i)  ε -linear mutant population: smaller than invasion threshold ε ’n but remain some constant fraction of the population (isn’t a vanishing population)

Results 40  A strategy s is ESS if given a mutant strategy t, the set of mutant strategies M n all playing t, for n sufficiently large, M n contracts  Random graphs: pairs of vertices jointed by probability p  If s is classical ESS of game F, if p = Ω (1/n c ), 0 ≤ c < 1, s is an ESS with probability 1 with respect to F and G  Adversarial mutations: At an ESS, at most o(n) mutants can be of abnormal fitness (i.e. outside of a additive factor τ )

Selection methods The art of diplomacy 41

Role of selection 42  Dynamics of EGT not solely determined by payoff matrix  Let the column vector p represent strategy proportions  F(p) is a fitness function  S(f, p) is the selection function  Returns the state of the population for the next generation, given fitness values and current proportions  p t + 1 = S(F(p t ), p t )  Different selection strategies result in different dynamics  Any S that maintains stable fixed points must obey p fix = S(c 1, p fix ), and show convergence around p fix

Selection methods 43  Some selection methods commonly used in evolutionary algorithms:  Truncation  ( μ, ƛ )-ES  Linear rank  Boltzmann selection

Example: Truncation selection 44  Population size n, selection pressure k  Sort population according to fitness  Replace worst k percent of the population with variations of the best k percent

Example: Linear rank selection 45  Often used in genetic algorithms  Agents sorted according to fitness, assigned new fitness values according to rank  Create roulette wheel based on new fitness values, create next generation  Useful for ensuring that even small differences in fitness levels are captured

References Just to prove I didn’t make the whole talk up. 46

References (not in any proper format!) 47  Suri S. Computational Evolutionary Game Theory, Chapter 29 of Algorithmic Game Theory, edited by Nisan N, Roughgarden T, Tardos E, and Vazirani V.  Ficici S, and Pollack J. Effects of Finite Populations on Evolutionary Stable Strategies  Ficici S, Melnik O, and Pollack J. A Game-Theoretic Investigation of Selection Methods Used in Evolutionary Algorithms