An artificial intelligence bot, created by researchers at Facebook and Carnegie Mellon, can beat human poker professionals in a six-player, no-limit Hold’em game, Facebook announced Thursday.
Pluribus, as the AI bot is called, has “decisively” won against a number of pros, including two World Series of Poker Main Event winners.
The details on how the researchers have managed to make Pluribus so good at multiplayer poker — a notoriously hard problem in AI — are in a new paper published in Science.
Pluribus has been built on the shoulders of Libratus, an AI that bested human pro players in two-player poker in 2017. It learned to play by competing against itself, without any data of prior human or AI play. Since poker is incredibly complex, having Pluribus look too far into the future wasn’t viable; instead, the bot used a new search algorithm that helps it make good decisions by looking at just the few next moves (instead of trying to figure out all the moves until the end of the game). It also used new and faster self-play algorithms that helped it cope with all the hidden information present in poker.
“Combined, these advances made it possible to train Pluribus using very little processing power and memory — the equivalent of less than $150 worth of cloud computing resources,” wrote Facebook AI research scientist Noam Brown.
During one experiment, Pluribus played 10,000 hands of poker over 12 days, against a dozen professionals (who were playing for a total prize for $50,000, giving them a reason to win).
In money terms, Pluribus was so much better than people, that if the game were played with $1 chips, it would have made about $1,000 per hour competing against five human players.
Google News, Bing News, Yahoo News, 200+ publications
Pluribus’s average win rate against professional poker players over the course of the 10,000-hand experiment.
The comments from pros who play against Pluribus are reminiscent of the comments from Go players who pitted against Google’s AlphaGo AI, some of which said that the AI made very different choices from human players.
“There were several plays that humans simply are not making at all,” said poker pro Michael Gagliano.
“Unlike humans, Pluribus used multiple raise sizes preflop. Attempting to respond to nonlinear open ranges was a fun challenge that differs from human games,” said poker pro Seth Davies.
Of course, the purpose of this research is not to create the ultimate poker playing machine (or is it?), but to explore how AI can fare in settings which involve hidden information. Facebook thinks it could translate Pluribus’ success into other real-world interactions, “including ones involving fraud prevention, cybersecurity, and taking action on harmful content.”