1 Algorithms for Computing Approximate Nash Equilibria Vangelis Markakis Athens University of Economics and Business.

Slides:



Advertisements
Similar presentations
Approximate Nash Equilibria in interesting games Constantinos Daskalakis, U.C. Berkeley.
Advertisements

Nash’s Theorem Theorem (Nash, 1951): Every finite game (finite number of players, finite number of pure strategies) has at least one mixed-strategy Nash.
Game Theory Assignment For all of these games, P1 chooses between the columns, and P2 chooses between the rows.
This Segment: Computational game theory Lecture 1: Game representations, solution concepts and complexity Tuomas Sandholm Computer Science Department Carnegie.
Bilinear Games: Polynomial Time Algorithms for Rank Based Subclasses Ruta Mehta Indian Institute of Technology, Bombay Joint work with Jugal Garg and Albert.
Mixed Strategies CMPT 882 Computational Game Theory Simon Fraser University Spring 2010 Instructor: Oliver Schulte.
COMP 553: Algorithmic Game Theory Fall 2014 Yang Cai Lecture 21.
Computation of Nash Equilibrium Jugal Garg Georgios Piliouras.
6.896: Topics in Algorithmic Game Theory Lecture 12 Constantinos Daskalakis.
6.896: Topics in Algorithmic Game Theory Lecture 11 Constantinos Daskalakis.
Game Theoretical Insights in Strategic Patrolling: Model and Analysis Nicola Gatti – DEI, Politecnico di Milano, Piazza Leonardo.
ECO290E: Game Theory Lecture 5 Mixed Strategy Equilibrium.
MIT and James Orlin © Game Theory 2-person 0-sum (or constant sum) game theory 2-person game theory (e.g., prisoner’s dilemma)
Nash Equilibria In Graphical Games On Trees Edith Elkind Leslie Ann Goldberg Paul Goldberg.
Nash equilibria in Electricity Markets: A comparison of different approaches Seminar in Electric Power Networks, 12/5/12 Magdalena Klemun Master Student,
EC941 - Game Theory Lecture 7 Prof. Francesco Squintani
6.853: Topics in Algorithmic Game Theory Fall 2011 Constantinos Daskalakis Lecture 13.
Playing Games for Security: An Efficient Exact Algorithm for Solving Bayesian Stackelberg Games Praveen Paruchuri, Jonathan P. Pearce, Sarit Kraus Catherine.
Towards a Constructive Theory of Networked Interactions Constantinos Daskalakis CSAIL, MIT Based on joint work with Christos H. Papadimitriou.
Equilibrium Concepts in Two Player Games Kevin Byrnes Department of Applied Mathematics & Statistics.
by Vincent Conitzer of Duke
Christos alatzidis constantina galbogini.  The Complexity of Computing a Nash Equilibrium  Constantinos Daskalakis  Paul W. Goldberg  Christos H.
An Introduction to Game Theory Part II: Mixed and Correlated Strategies Bernhard Nebel.
1 Computing Nash Equilibrium Presenter: Yishay Mansour.
An Introduction to Game Theory Part III: Strictly Competitive Games Bernhard Nebel.
Approximation algorithms and mechanism design for minimax approval voting Ioannis Caragiannis Dimitris Kalaitzis University of Patras Vangelis Markakis.
An Introduction to Black-Box Complexity
Finite Mathematics & Its Applications, 10/e by Goldstein/Schneider/SiegelCopyright © 2010 Pearson Education, Inc. 1 of 68 Chapter 9 The Theory of Games.
On Bounded Rationality and Computational Complexity Christos Papadimitriou and Mihallis Yannakakis.
The Computational Complexity of Finding a Nash Equilibrium Edith Elkind, U. of Warwick.
Computing Equilibria Christos H. Papadimitriou UC Berkeley “christos”
Game Theory.
1 Issues on the border of economics and computation נושאים בגבול כלכלה וחישוב Congestion Games, Potential Games and Price of Anarchy Liad Blumrosen ©
Minimax strategies, Nash equilibria, correlated equilibria Vincent Conitzer
Computing Equilibria Christos H. Papadimitriou UC Berkeley “christos”
MAKING COMPLEX DEClSlONS
Game representations, solution concepts and complexity Tuomas Sandholm Computer Science Department Carnegie Mellon University.
The Computational Complexity of Finding Nash Equilibria Edith Elkind Intelligence, Agents, Multimedia group (IAM) School of Electronics and CS U. of Southampton.
Nash Equilibria In Graphical Games On Trees Revisited Edith Elkind Leslie Ann Goldberg Paul Goldberg (University of Warwick) (To appear in ACM EC’06)
Small clique detection and approximate Nash equilibria Danny Vilenchik UCLA Joint work with Lorenz Minder.
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.
Game-theoretic analysis tools Tuomas Sandholm Professor Computer Science Department Carnegie Mellon University.
Yang Cai Oct 08, An overview of today’s class Basic LP Formulation for Multiple Bidders Succinct LP: Reduced Form of an Auction The Structure of.
Introduction to Game Theory application to networks Joy Ghosh CSE 716, 25 th April, 2003.
Algorithmic Game Theory and Internet Computing Vijay V. Vazirani Georgia Tech Market Equilibrium: The Quest for the “Right” Model.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
Ásbjörn H Kristbjörnsson1 The complexity of Finding Nash Equilibria Ásbjörn H Kristbjörnsson Algorithms, Logic and Complexity.
Price of Anarchy Georgios Piliouras. Games (i.e. Multi-Body Interactions) Interacting entities Pursuing their own goals Lack of centralized control Prediction?
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.
Algorithms for solving two-player normal form games
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.
1 a1a1 A1A1 a2a2 a3a3 A2A Mixed Strategies When there is no saddle point: We’ll think of playing the game repeatedly. We continue to assume that.
1 Algorithms for Computing Approximate Nash Equilibria Vangelis Markakis Athens University of Economics and Business.
6.853: Topics in Algorithmic Game Theory Fall 2011 Constantinos Daskalakis Lecture 8.
5.1.Static Games of Incomplete Information
Parameterized Two-Player Nash Equilibrium Danny Hermelin, Chien-Chung Huang, Stefan Kratsch, and Magnus Wahlstrom..
By: Donté Howell Game Theory in Sports. What is Game Theory? It is a tool used to analyze strategic behavior and trying to maximize his/her payoff of.
Game representations, game-theoretic solution concepts, and complexity Tuomas Sandholm Computer Science Department Carnegie Mellon University.
6.853: Topics in Algorithmic Game Theory Fall 2011 Constantinos Daskalakis Lecture 12.
Commitment to Correlated Strategies Vince Conitzer and Dima Korzhyk.
Game theory basics A Game describes situations of strategic interaction, where the payoff for one agent depends on its own actions as well as on the actions.
Yuan Deng Vincent Conitzer Duke University
Nash Equilibrium: P or NP?
Communication Complexity as a Lower Bound for Learning in Games
Historical Note von Neumann’s original proof (1928) used Brouwer’s fixed point theorem. Together with Danzig in 1947 they realized the above connection.
Commitment to Correlated Strategies
EASTERN MEDITERRANEAN UNIVERSITY DEPARTMENT OF INDUSTRIAL ENGINEERING IENG314 OPERATIONS RESEARCH II SAMIR SAMEER ABUYOUSSEF
Normal Form (Matrix) Games
A Technique for Reducing Normal Form Games to Compute a Nash Equilibrium Vincent Conitzer and Tuomas Sandholm Carnegie Mellon University, Computer Science.
Presentation transcript:

1 Algorithms for Computing Approximate Nash Equilibria Vangelis Markakis Athens University of Economics and Business

2 Outline  Introduction to Games -The concepts of Nash and  -Nash equilibrium  Computing approximate Nash equilibria -A subexponential algorithm for any constant  > 0 -Polynomial time approximation algorithms  Conclusions

3 What is Game Theory? Game Theory aims to help us understand situations in which decision makers interact Goals: –Mathematical models for capturing the properties of such interactions –Prediction (given a model how should/would a rational agent act?) Rational agent: when given a choice, the agent always chooses the option that yields the highest utility

4 Models of Games Cooperative or noncooperative Simultaneous moves or sequential Finite or infinite Complete information or incomplete information

5 In this talk: Cooperative or noncooperative Simultaneous moves or sequential Finite or infinite Complete information or incomplete information

6 Noncooperative Games in Normal Form 2, 2 2, 2 0, 4 0, 4 4, 0 4, 0 -1, -1 Rowplayer Column Player The Hawk-Dove game

7 Example 2: The Bach or Stravinsky game (BoS) 2, 1 2, 1 0, 0 0, 0 0, 0 0, 0 1, 21, 2

8 Example 3: A Routing Game ●● st A: 5x B: 7.5x C: 10x

9 Example 3: A Routing Game 10, 10 5, 7.5 5, , 5 15, , 10 10, 5 10, 7.520, 20 ABC A B C

10 Definitions 2-player game (R, C): n available pure strategies for each player n x n payoff matrices R, C i, j played  payoffs : R ij, C ij Mixed strategy: Probability distribution over [n] Expected payoffs :

11 Solution Concept x *, y * is a Nash equilibrium if no player has a unilateral incentive to deviate: (x, Ry * )  (x *, Ry * )  x (x *, Cy)  (x *, Cy * )  y [Nash, 1951]: Every finite game has a mixed strategy equilibrium. Proof: Based on Brouwer’s fixed point theorem. (think of it as a steady state)

12 Solution Concept x *, y * is a Nash equilibrium if no player has a unilateral incentive to deviate: (x, Ry * )  (x *, Ry * )  x (x *, Cy)  (x *, Cy * )  y [Nash, 1951]: Every finite game has a mixed strategy equilibrium. Proof: Based on Brouwer’s fixed point theorem. (think of it as a steady state)

13 Solution Concept x *, y * is a Nash equilibrium if no player has a unilateral incentive to deviate to a pure strategy: (x i, Ry * )  (x *, Ry * )  x i (x *, Cy j )  (x *, Cy * )  y j It suffices to consider only deviations to pure strategies Let x i = (0, 0,…,1, 0,…,0) be the ith pure strategy

14 Example: The Hawk-Dove Game 2, 2 2, 2 0, 4 0, 4 4, 0 4, 0 -1, -1 Rowplayer Column Player

15 Example 2: The Bach or Stravinsky game (BoS) 2, 1 2, 1 0, 0 0, 0 0, 0 0, 0 1, 21, 2 3 equilibrium points: 1. (B, B) 2. (S, S) 3. ((2/3, 1/3), (1/3, 2/3))

16 Complexity issues  m = 2 players, known algorithms: worst case exponential time [Kuhn ’61, Lemke, Howson ’64, Mangasarian ’64, Lemke ’65]  If NP-hard  NP = co-NP [Megiddo, Papadimitriou ’89]  NP-hard if we add more constraints (e.g. maximize sum of payoffs) [Gilboa, Zemel ’89, Conitzer, Sandholm ’03]  Representation problems m = 3, there exist games with rational data BUT irrational equilibria [Nash ’51]  PPAD-complete even for m = 2 [Daskalakis, Goldberg, Papadimitriou ’06, Chen, Deng, Teng ’06] Poly-time equivalent to:  finding approximate fixed points of continuous maps on convex and compact domains

17 Approximate Nash Equilibria Recall definition of Nash eq. : (x, Ry * )  (x *, Ry * )  x (x *, Cy)  (x *, Cy * )  y  -Nash equilibria (incentive to deviate   ) : (x, Ry * )  (x *, Ry * ) +   x (x *, Cy)  (x *, Cy * ) +   y Normalization: entries of R, C in [0,1]

18 Searching for Approximate Equilibria [Lipton, M., Mehta ’03]: For any  in (0,1), and for every k  9logn/  2, there exists a pair of k-uniform strategies x, y that form an  -Nash equilibrium. Definition: A k-uniform strategy is a strategy where all probabilities are integer multiples of 1/k e.g. (3/k, 0, 0, 1/k, 5/k, 0,…, 6/k)

19 A Subexponential Algorithm (Quasi-PTAS) [Lipton, M., Mehta ’03]: For any  in (0,1), and for every k  9logn/  2, there exists a pair of k-uniform strategies x, y that form an  -Nash equilibrium. Definition: A k-uniform strategy is a strategy where all probabilities are integer multiples of 1/k e.g. (3/k, 0, 0, 1/k, 5/k, 0,…, 6/k) Corollary : We can compute an  -Nash equilibrium in time Proof: There are n O(k) pairs of strategies to look at. Verify  -equilibrium condition.

20 Proof of Existence Let x *, y * be a Nash equilibrium. - Sample k times from the set of pure strategies of the row player, independently, at random, according to x *  k-uniform strategy x - Same for column player  k-uniform strategy y Based on the probabilistic method (sampling) Suffices to show Pr[x, y form an  -Nash eq.] > 0

21 Proof (cont’d) Enough to consider deviations to pure strategies (x i, Ry)  (x, Ry) +   i (x i, Ry): sum of k random variables with mean (x i, Ry * ) Chernoff-Hoeffding bounds  (x i, Ry)  (x i, Ry * ) with high probability (x i, Ry)  (x i, Ry * ) ≤ (x *, Ry * )  (x, Ry) Finally when k =  (logn/  2 ) : Pr[  deviation with gain more than  ] =

22 Multi-player Games For m players, same technique: support size: k = O(m 2 log(m 2 n)/  2 ) running time: exp(logn, m, 1/  ) Previously [Scarf ’67]: exp(n, m, log(1/  )) (fixed point approximation) [Lipton, M. ’04]: exp(n, m) but poly(log(1/  )) (using algorithms for polynomial equations)

23 Outline  Introduction to Games -The concepts of Nash and  -Nash equilibrium  Computing approximate Nash equilibria -A subexponential algorithm for any constant  > 0 -Polynomial time approximation algorithms  Conclusions

24 Polynomial Time Approximation Algorithms For  = 1/2: Feder, Nazerzadeh, Saberi ’07: For  < 1/2, we need support at least  (log n) i k j Pick arbitrary row i Let j = best response to i Find k = best response to j, play i or k with prob. 1/2 R ij, C ij R kj, C kj

25 Polynomial Time Approximation Algorithms Daskalakis, Mehta, Papadimitriou (EC ’07): in P for  = 1-1/φ = (3-  5)/2  (φ = golden ratio) Bosse, Byrka, M. (WINE ’07): a different LP-based method 1.Algorithm 1: 1-1/φ 2.Algorithm 2: Βased on sampling + Linear Programming - Need to solve polynomial number of linear programs Running time: need to solve one linear program

26 Approach Fact: 0-sum games can be solved in polynomial time (equivalent to linear programming) - Start with an equilibrium of the 0-sum game (R-C, C-R) - If incentives to deviate are “high”, players take turns and adjust their strategies via best response moves 0-sum games: games of the form (R, -R) Similar idea used in [Kontogiannis, Spirakis ’07] for a different notion of approximation

27 Algorithm 1 1.Find an equilibrium x *, y * of the 0-sum game (R - C, C - R) 2.Let g 1, g 2 be the incentives to deviate for row and column player respectively. Suppose g 1  g 2 3.If g 1  , output x *, y * 4.Else: let b 1 = best response to y *, b 2 = best response to b 1 5.Output: x = b 1 y = (1 -  2 ) y * +  2 b 2 Theorem: Algorithm 1 with  = 1-1/φ and  2 = (1- g 1 ) / (2- g 1 ) achieves a (1-1/φ)-approximation Parameters: ,  2  [0,1]

28 Analysis of Algorithm 1 Why start with an equilibrium of (R - C, C - R)? Intuition: If row player profits from a deviation from x * then column player also gains at least as much Case 1: g 1     -approximation Case 2: g 1 >  for row player   2 for column player  (1 -  2 )(1 - (b 1, Cy * ))  (1 -  2 )(1 - g 1 ) = (1 - g 1 ) / (2 - g 1 )  max{ , (1 -  )/(2 -  )}-approximation Incentive to deviate:

29 Analysis of Algorithm 1

30 Towards a better algorithm 1.Find an equilibrium x *, y * of the 0-sum game (R - C, C - R) 2.Let g 1, g 2 be the incentives to deviate for row and column player respectively. Suppose g 1  g 2 3.If g 1  , output x *, y * 4.Else: let b 1 = best response to y *, b 2 = best response to b 1 5.Output: x = b 1 y = (1 -  2 ) y * +  2 b 2

31 Algorithm 2 1.Find an equilibrium x *, y * of the 0-sum game (R - C, C - R) 2.Let g 1, g 2 be the incentives to deviate for row and column player respectively. Suppose g 1  g 2 3.If g 1  [0, 1/3], output x *, y * 4.If g 1  (1/3,  ], - let r 1 = best response to y *, x = (1 -  1 ) x * +  1 r 1 - let b 2 = best response to x, y = (1 -  2 ) y * +  2 b 2 5.If g 1  ( , 1] output: x = r 1 y = (1 -  2 ) y * +  2 b 2

32 Analysis of Algorithm 2 (Reducing to an optimization question) - Let h = (x *, Cb 2 ) - (x *, Cy * ) Theorem: The approximation guarantee of Algorithm 2 is and is given by: - We set  2 so as to equalize the incentives of the players to deviate

33 Analysis of Algorithm 2 (solution) Optimization yields:

34 Graphically:

35 Analysis – tight example 0, 0 ,  0, 11, 1/2 ,  1, 1/20, 1 (R, C) =  =

36 Remarks and Open Problems Spirakis, Tsaknakis (WINE ’07): currently best approximation of –yet another LP-based method Polynomial Time Approximation Scheme (PTAS)? Yes if: –rank(R) = O(1) & rank(C) = O(1) [Lipton, M. Mehta ’03] –rank(R+C) = O(1) [Kannan, Theobald ’06] PPAD-complete for  = 1/n [Chen, Deng, Teng ’06]

37 Other Notions of Approximation  -well-supported equilibria: every strategy in the support is an approximate best response –[Kontogiannis, Spirakis ’07]: approximation, based also on solving 0-sum games Strong approximation: output is geometrically close to an exact Nash equilibrium –[Etessami, Yannakakis ’07]: mostly negative results

38 Thank You!