Communications of the ACM
Technical Perspective: Solving Imperfect Information Games
The study of games is as old as computer science itself. Babbage, Turing, and Shannon devised algorithms and hardware to play the game of chess. Game theory began with questions regarding optimal strategies in card games and chess, later developed into a formal system by von Neumann. Chess subsequently became the drosophila—or common fruitfly, the most studied organism in genetics—of artificial intelligence research. Early successes in chess and other games shaped the emerging field of AI: many planning algorithms first used in games became pillars of subsequent research; reinforcement learning was first developed for a checkers playing program; and the performance of game-playing programs has frequently been used to measure progress in AI.
Most of this research focused on perfect information games, in which all events are observed by all players, culminating in programs that beat human world champions in checkers, chess, Othello, backgammon, and most recently, Go. However, many applications in the real world have imperfect information: each agent observes different events. This leads to the possibility of deception and a wealth of social strategies. Imperfect information games provide a microcosm of these social interactions, while abstracting away the messiness of the real world.
No entries found