advanced artificial intelligence lecture 3b: game theory
TRANSCRIPT
Game Theory
• For games where two players move simultaneously (or, without knowledge of each other)
• Often, single-move games• Two problems are studied– Agent design– Mechanism design
Prisoner’s Dilemma
• Dominant Strategy: testify(does better no matter what other player does)
• Pareto Optimal: no other outcome that all players would prefer
• Equilibrium: no player can benefit from switching (assuming other players stay the same)Every game has at least one (Nash)
Game Console Game
• Dominant Strategy?• No• Equilibrium?• blu,blu; dvd,dvd• Pareto Optimal?• Blu,blu: +9, +9
Sample Game
• Two finger Morra: players E and O show 1 or 2 fingers; total f. E wins f if even; O if odd.
• What is best strategy for each player (solution)?– Pure strategy: single move– Mixed strategy: probability distribution on moves
Convert to Normal Form
Problem: number of strategies exponential in number of information setsTexas Hold’em: 1018 states.
Extensive Games
• Sequence form (Koller): Up to 25,000 states• Abstraction– Suits; Hi/mid/lo; bet amounts; only some deals
• Can handle– PO, multi-agent, stochastic, sequential, dynamic
• Can’t handle (very well)– Unknown actions; continuous actions– Irrational opponents– Unknown utilities
Mechanism Design
• Game Theory:– Given a game, find rational policy
• Mechanism Design:– Given utility functions Ui, design a game such that
the rational strategies maximize ∑ Ui
• Examples:– Google ads, airplane tickets, radio spectra, TCP
packets, dating, doctor internships
Auction• Auction is globally better with more bidders• Easier if bidders have dominant strategy– Strategy-proof; truth-revealing; incentive compatible
• Sealed bid auction– Strategy-proof?
• Second-price (Vickrey) auction– Payoff for bid b with value v, best other bid c:
U = (v – c) if (b > c) else 0• Optimal bid:– Bid v