Rock-Paper-Scissors.pptVIP

SA3- Game Theoretic Learning Tutorial: Learning Topics in Game -Theoretic Decision MakingMichael Littman - Rutgers University Why Game Theory at COLT? Multiple choice: MK MW are interested and got to pick the topic. Community on the hunt for new problems. It’s a natural fit… A Natural Fit Computational {Learning, Game} Theory COLT: concerned with interaction (learner distribution) worst-case assumption posits an adversary CGT: solution often defined in terms of best response thus, need to learn about other decision makers Outline Normal-form Games Repeated Games Stochastic Games A. Normal-Form Games n players, Ai action set for player i A is joint action space: A1 x … x An. Ri reward to player i as a function of the joint action A. Example: Rock-Paper-Scissors Two players. Each independently and simultaneously chooses an action: rock, paper, or scissors. Rewards: rock beats scissors scissors beats paper paper beats rock Matrices: Strategies Notation: s is a joint strategy for all players. Ri(s) = SaeA s(a) Ri(a) s-i is a joint strategy for all players except i. si , s-i is a joint strategy where i uses strategy si and everyone else s-i . Types of games: zero-sum games: R1 + R2 = 0. team games: R1 = … = Rn . general-sum games: (ex. Prisoner’s dilemma) Solutions: Minimax Consider matching pennies: Q: What do we do when the world is out to get us? A: Make sure it can’t. Play strategy with the best worst-case outcome. argmaxsieD(Ai) mina-ieA-i Ri(si, s-i ) Minimax optimal strategy. Minimax Equilibrium Back to matching pennies: R1(s*1, s2 ) = 0 for all s2 Minimax optimal guarantees the safety value. Generally, bad opponent choice can help (saddle point). Minimax: Linear Programming Finding minimax optimal equivalent to LP (so in P). Payoff vs. opponent action linear function of strategy. Since opponent minimizes, choices on lower surface. Choose best point: maxv,s1 v s.t. R1(s1, b) ? v for all i Nash Equilibria Nash E