AI system beats humans at the poker table, in complex game

Researchers have created an artificial intelligence system that outperforms humans at the heads-up no-limit version of Texas Hold'em, a complex game involving 10 raised to the power of 160 decision points.

In games such as chess, players have equal access to information. What makes poker much more challenging for artificial intelligence to master is that each player has asymmetrical information about the state of the game.

Matej Moravcík and colleagues developed an AI system called DeepStack that outperformed, with statistical significance, 10 of 11 professional poker players after 3,000 games each of heads-up no-limit poker (the system also outperformed the 11th player, but that result is not statistically significant).

At each point in the game, DeepStack recomputes the game strategy; instead of calculating possible moves to the very end, it limits the depth and the type of moves allowed. This reduces the number of possible decision points from more than 10 to the power of 160 down to approximately 10 to the power of 7,  the authors say, which DeepStack's computer chip can solve in 5 seconds.

Such AI systems could enable advances in real-world problems that involve information asymmetry, such as defending strategic resources and making important medical treatment recommendations.