Games have long been important testing grounds for artificial intelligence research. In particular, games like chess and Go quickly become too complex to solve by brute force alone, leading researchers to develop programs that learn a form of intuition in order to master them. Poker, specifically six-player no limit Texas Hold’em, introduces a new challenge because there is more than one opponent and each player has a hidden card, as opposed to chess, in which all pieces are visible.
Noam Brown and Tuomas Sandholm, from Carnegie Mellon, recently published the results from their artificial intelligence, or AI, “Pluribus”. The Pluribus algorithm works in two stages. First, the AI learns a general strategy guide called a “blueprint” that glosses over some of the details of the game and does not consider its opponents’ tendencies. Pluribus learned the blueprint largely without human intervention by playing against copies of itself for thousands of games. Second, when playing against human opponents, Pluribus refined its strategy by simulating all the ways in which each hand could play out, while accounting for the possibility that each player could change strategies at each turn. In this way, Pluribus makes decisions in real time based on intuition gained from its past games instead of simply memorizing a catalog of moves to make for every possible scenario. Finally, Pluribus consistently beat professional poker players, all of whom had previously won at least $1 million in tournaments.
Pluribus represents a major advance in artificial intelligence that has implications in other game-theory fields, such as cybersecurity and finance (trading stocks is not so fundamentally different from poker). Although Pluribus has mastered online poker, it is still removed from the human element of the game: its “body language” will never betray when it’s bluffing. Likewise, it is also incapable of reading the non-verbal cues of its opponents, as many professionals can. Considering this technology has caught the attention of the Pentagon, let’s hope it can learn some human compassion.
Managing Correspondent: Julian Segert
Press Articles: Hold ’Em or Fold ’Em? This A.I. Bluffs With the Best New York Times
It’s Hard To Win At Poker Against An Opponent With No Tell FiveThirtyEight
Original Journal Article: Superhuman AI for multiplayer poker Science
Image credit: Flickr