Game AI: chess, Go and poker

In short: Games are a test bed for AI because the rules are clear and success can be measured. Chess, Go and poker need different ideas, though.

Chess

Classically (Deep Blue, 1997) with brute-force search and a hand-written evaluation. Modern programs such as Stockfish combine deep search with a small network (NNUE) that rates positions. AlphaZero instead learned only by playing itself, without an opening book.

Go

Far too many positions for pure search. AlphaGo (2016) combined a network that suggests moves, a network that rates positions and Monte Carlo tree search. It learned first from human games, then by self-play (reinforcement learning).

Poker

Poker is a game of hidden information: you do not know the opponents’ cards. You must weigh probabilities and sometimes bluff. Methods such as counterfactual regret minimisation approximate game-theoretically balanced play (Libratus, Pluribus).

Why it matters

Many ideas (self-play, search with learned evaluation, dealing with uncertainty) carry over to other areas, such as planning, chip design and language models with reasoning. See Reasoning models and chain of thought.

See also: Robotics and embodied AI