Pluribus doesn’t wear sunglasses, but has learned to bluff at the poker table like the big boys.
Pluribus is actually an AI program developed by Carnegie Mellon University in collaboration with Facebook AI.
What separates Pluribus from other hoi polloi bots is that it has defeated top pros Darren Elias and Chris Ferguson who have each won numerous poker titles. Each of the pros separately played 5,000 rounds of poker against five copies of Pluribus, according to the university. The university has posted videos showing examples of how Pluribus bluffed and won.
The pros played the popular six-player No-Limit Texas Hold’em. In another experiment, Pluribus played another 13 pros, five at a time, for a total of 10,000 hands and came out the winner.
The achievement hits a milestone in AI and game theory, said Tuomas Sandholm, a computer science professor who developed Pluribus with Noam Brown, who is wrapping up a Ph.D. in computer science with the university as a research scientist at Facebook AI.
What’s different from prior AI games is that strategic reasoning had been limited to two-party competition, but Pluribus beat five other players in a complex game. “It opens up new opportunities to use AI to solve a wide variety of real-world problems,” Sandholm said.
Brown said that Pluribus’ playing strategies could even change the way pros play the game.
One way that Pluribus won was by placing “donk bets” far more often than the pros that it trashed. A donk bet is when a player ends one round with a call, but then starts another round with a bet—often considered a weak move that usually doesn’t make sense.
Elias, one of the pros, said Pluribus’ major strength is its ability to mix strategies. Humans try to do that as well, but can’t do so in a random way consistently, like Pluribus can. “Most people just can’t,” he said. It was noteworthy that Pluribus didn’t only play middle-of-the-road players, and instead played some of the best in the world, he added.
Another top pro, Michael Gagliano, said Pluribus also makes plays that humans don’t ever make related to the size of bets.
Sandholm and Brown also developed Libratus, which beat four poker players in a combined 120,000 hands of Heads-up No-Limit Texas Hold’em, a two-player game.
The university said poker is a bigger challenge than games like chess and Go because information in the game is incomplete. Players aren’t certain which cards are in play, and opponents can and will bluff, which makes it a tougher AI challenge and relevant to real-world problems.
Pluribus wins partly by computing a blueprint strategy by playing six copies of itself, good for the first round of bets. After that, Pluribus conducts a search of possible moves and looks ahead several moves, but not all the way to the end of the game. This limited-lookahead search is challenging, but a new limited-lookahead search algorithm was the main academic breakthrough that allowed Pluribus to win.
Brown and Sandholm published their research entitled “Superhuman AI for multiplayer poker” in Science.