AI program defeats professionals in six-player poker game

An AI program created by Carnegie Mellon University researchers in collaboration with Facebook has beaten top professionals in the world’s most popular form of poker, six-player no-limit Texas hold’em poker.
The AI named Pluribus defeated two top players, Darren Elias, who has the most number of World Poker Tour Titles and Chris “Jesus” Ferguson, champion of six World Series of Poker events. Each of them separately played 5000 hands of poker against five Pluribus copies.
Pluribus also played against 13 pros in another experiment, all of whom have won more than a million dollars playing poker. It played against five pros at a time for a total of 10,000 hands and emerged as the winner.
Tuomas Sandholm, Computer Science professor at CMU created Pluribus along with Noam Brown, Computer Science Ph.D. and research scientist at Facebook AI. Sandholm said that Pluribus achieved a supreme level of performance at multi-player poker which is considered a milestone in AI and in game theory that has been open for decades. Milestones in AI in strategic reasoning have been limited to two-party contests until now. However, defeating five other players in a complicated game opens up new possibilities of using AI to solve real-world problems. The paper detailing this achievement has been published in the Science journal. Playing a six-player contest rather than two players involve fundamental changes in how AI develops its strategy. Brown said that some of the strategies used by Pluribus may even change the way professionals approach the game.

Algorithms used by Pluribus used some surprising features in its strategy. For example, most human players avoid “donk betting” – ending a round with a call but starting the other round with a bet. It is generally viewed as a weak move. However, Pluribus used a large number of donk bets in its game.
Michael “Gags” Gagliano who has won almost two million dollars in his career also played against Pluribus. He was fascinated by some of the strategies used by Pluribus and felt that several plays related to bet sizing were not being made by the human players.
Sandholm and Brown had earlier developed Libratus. It defeated four poker players playing a combined 120000 hands in a two-player version of the game. In games involving more than two players such as Poker, using Nash equilibrium(which is generally used by other two-player game AI’s) can result in a defeat. Pluribus first plays against six copies of itself and develops a blueprint strategy. It then looks ahead several moves for adopting further strategies. A newly developed limited-lookahead search algorithm enabled Pluribus to achieve superhuman results. It also tries out unpredictable moves as it is often considered that AI bets only when it has the best possible hand. Pluribus used efficient computation and computed its blueprint strategy in eight days using 12400 core hours compared to 15 million core hours used by Libratus.

Leave a Reply

Your email address will not be published. Required fields are marked *