1. Finding equilibria in large sequential games of imperfect information CMU Theory Lunch – November 9, 2005 (joint work with Tuomas Sandholm)

2. 2 Motivation: Poker AI Poker is a wildly popular card game –2005 World Series of Poker: Total prize pool > $103M $56M for the Main Event ESPN broadcasting portions of the tournament Poker presents several challenges for AI –Imperfect information –Risk assessment and management –Deception (bluffing, slow-playing) –Counter-deception (calling a bluff, avoiding slow-play trap)

3. 3 Sneak preview of results Rhode Island Hold’em poker invented as a testbed for AI research [Shi & Littman 2001] Game tree has more than 3.1 billion nodes Previously, the best techniques did not scale to games this large Using our algorithm we have computed optimal minimax strategies for this game This is the largest poker game solved to date by over four orders of magnitude

4. 4 Rhode Island Hold’em poker: The Deal

5. 5 Rhode Island Hold’em poker: Round 1

6. 6 Rhode Island Hold’em poker: Round 2

7. 7 Rhode Island Hold’em poker: Round 3

8. 8 Rhode Island Hold’em poker: Showdown

9. 9 Outline of this talk 1.Game theory: Representation, computing equilibrium 2.Model: Ordered games 3.Abstraction mechanism: Information filters 4.Strategic equivalence: Game isomorphisms 5.Algorithm: GameShrink 6.Solving Rhode Island Hold’em

10. 10 Game Theory In multi-agent systems, an agent’s outcome depends on the actions of the other agents Consequently, an agent’s optimal action depends on the actions of the other agents Game theory provides guidance as to how an agent should act A game-theoretic equilibrium specifies a strategy for each agent such that no agent wishes to deviate –Such an equilibrium always exists [Nash 1950]

11. 11 Nash equilibrium and a simple example 0, 0-1, 11, -1 0, 0-1, 1 1, -10, 0 Rock Paper Scissors Paper 1/3 A Nash equilibrium specifies a strategy (possibly randomized) for each agent such that no agent wishes to deviate

12. 12 Complexity of computing equilibria Finding a Nash equilibrium is “…the most important concrete open question on the boundary of P today” [Papadimitriou 2001] –Even for games with only two players There are algorithms (requiring exponential-time in the worst-case) for computing Nash equilibria Good news: Two-person zero-sum matrix games can be solved in poly-time using linear programming

13. 13 Computing equilibria in zero-sum games Among all best responses, there is always at least one pure strategy Thus, player 1’s optimization problem is: This is equivalent to: By LP duality, player 2’s optimal strategy is given by the dual vars

14. 14 What about sequential games? Sequential games involve turn-taking, moves of chance, and imperfect information Convert sequential game into a matrix game: This approach leads to exponential number of strategies

15. 15 Sequence form Instead of a move for every information set, consider choices necessary for each leaf These choices are sequences and constitute the pure strategies in the sequence form S 1 = {{}, l, r, L, R} S 2 = {{}, c, d}

16. 16 Realization plans Players strategies are specified as realization plans over sequences: Prop. Realization plans are equivalent to behavior strategies.

17. 17 Computing equilibria via sequence form Players 1 and 2 have realization plans x and y Realization constraint matrices E and F specify constraints on realizations {} l r L R {} c d {} v v’ {} u

18. 18 Computing equilibria via sequence form Payoffs for player 1 and 2 are: x T Ay and x T (-A)y Creating payoff matrix: –Initialize each entry to 0 –For each leaf, there is a (unique) pair of sequences corresponding to an entry in the payoff matrix –Weight the entry by the product of chance probabilities along the path from the root to the leaf {} c d {} l r L R

19. 19 Computing equilibria via sequence form PrimalDual Holding x fixed, compute best response Holding y fixed, Compute best response Primal Dual

20. 20 Computing equilibria via sequence form: An example min p1 subject to x1: p1 - p2 - p3 >= 0 x2: 0y1 + p2 >= 0 x3: -y2 + y3 + p2 >= 0 x4: 2y2 - 4y3 + p3 >= 0 x5: -y1 + p3 >= 0 q1: -y1 = -1 q2: y1 - y2 - y3 = 0 bounds y1 >= 0 y2 >= 0 y3 >= 0 p1 Free p2 Free p3 Free end

21. 21 Sequence form summary The sequence form is an alternative, more compact representation [Romanovskii 1962], [Koller, Megiddo, von Stengel 1994] Two-player zero-sum games with perfect recall can be solved in time polynomial in the size of the game tree –Not enough to solve RI Hold’em (~10 9 nodes) or Texas Hold’em (~10 18 nodes)

22. 22 Our approach: Automated abstraction Instead of developing an equilibrium-finding algorithm per se, we introduce an automated abstraction technique that results in a smaller, equivalent game We prove that a Nash equilibrium in the smaller game corresponds to a Nash equilibrium in the original game Our technique applies to n-player sequential games with observed actions and ordered signals

23. 23 Illustration of our approach Nash equilibrium Original game Abstracted game Abstraction Compute Nash

24. 24 Game with ordered signals (a.k.a. ordered game) 1.Players I = {1,…,n} 2.Stage games G = G 1,…,G r 3.Player label L 4.Game-ending nodes ω 5.Signal alphabet Θ 6.Signal quantities κ = κ 1,…,κ r and γ = γ 1,…,γ r 7.Signal probability distribution p 8.Partial ordering ≥ of subsets of Θ 9.Utility function u (increasing in private signals) I = {1,2} Θ = {2♠,…,A♦} κ = (0,1,1) γ = (1,0,0) UniformHand rank

25. 25 Information filters Observation: We can make games smaller by filtering the information a player receives Instead of observing a specific signal exactly, a player instead observes a filtered set of signals –E.g. receiving the signal {A♠,A♣,A♥,A♦} instead of A♠ Represented as a partition of signal space Combining an ordered game and an information filter yields a filtered ordered game Prop. A filtered ordered game is a finite sequential game with perfect recall –Corollary If the filtered ordered game is two-person zero- sum, we can solve it in poly-time using LP

26. 26 Filtered signal trees Every filtered ordered game has a corresponding filtered signal tree –Each edge corresponds to the revelation of some signal –Each path corresponds to the revelation of a set of signals Our algorithm operates directly on the filtered signal tree –We never load the full game into memory

27. 27 Ordered game isomorphic relation Captures notion of strategic symmetry between nodes We define the relationship recursively: –Two leaves are ordered game isomorphic if, for each action history, the payoffs are the same at both leaves –Two internal nodes are ordered game isomorphic if they are siblings and there is a bijection between their children such that only ordered game isomorphic nodes are matched We can compute this relationship efficiently using dynamic programming and perfect matching computations in a bipartite graph

28. 28 Ordered game isomorphic abstraction transformation Transforms an existing information filter into a new filter that merges two ordered game isomorphic nodes The new filter yields a smaller, abstracted game Thm If a strategy profile is a Nash equilibrium in the smaller, abstracted game, then it’s interpretation in the original game is a Nash equilibrium in that game

29. 29 Applying the ordered game isomorphic abstraction transformation

30. 30 Applying the ordered game isomorphic abstraction transformation

31. 31 Applying the ordered game isomorphic abstraction transformation

32. 32 GameShrink: Efficiently computing ordered game isomorphic abstraction transformations Recall: We have a dynamic program for determining if two nodes of the filtered signal tree are ordered game isomorphic Algorithm: Starting from the top of the filtered signal tree, perform the transformation where applicable Approximation algorithm: Instead of requiring perfect matching, require a matching with a penalty below some threshold

33. 33 Algorithmic techniques to speed up GameShrink Union-Find data structure provides an efficient representation of the information filter –Linear memory and almost linear time Eliminate some perfect matching computations using easy-to-check necessary conditions –Compact histogram databases for storing win/loss frequencies to speed up the checks

34. 34 Solving Rhode Island Hold’em poker Without abstraction, LP has 91,224,226 rows & columns GameShrink computes all ordered game isomorphic abstraction transformations in under one second Now the LP has only 1,237,238 rows & columns Solving this LP yields optimal minimax strategies –CPLEX barrier method takes 7 days, 17 hours and 25 GB RAM Largest poker game solved to date by over four orders of magnitude

35. 35 Comparison to previous research Rule-based –Limited success in even small poker games Simulation/Learning –Do not take multi-agent aspect into account Game-theoretic approaches –Tiny games –Manual abstraction “Approximating Game-Theoretic Optimal Strategies for Full-scale Poker”, Billings, Burch, Davidson, Holte, Schaeffer, Schauenberg, Szafron, IJCAI-03 –Our approach: Automated abstraction

36. 36 Directions for future work Computing strategies for larger games –Requires approximation of solutions Tournament poker More than two players Other types of abstraction

37. 37 Summary Introduced an automatic method for performing abstractions in a broad class of games Introduced information filters as a technique for working with games with imperfect information Developed an equilibrium-preserving abstraction transformation, along with an efficient algorithm Described a simple extension that yields an approximation algorithm for tackling even larger games Solved the largest poker game to date –Playable on-line at http://www.cs.cmu.edu/~gilpin/gsi.html http://www.cs.cmu.edu/~gilpin/gsi.html Thank you very much for your interest