An artificial intelligence bot developed by Facebook defeated professionals in six-player Hold’em, marking the first time AI has beaten human experts in a complex game with more than two players or teams.
Although AI has made many advances, beating human players in chess and other games, poker has been more challenging because it involves hidden information and success can require techniques such as bluffing.
Pluribus, the name of the AI bot developed by Facebook in collaboration with Carnegie Mellon University, was able to defeat elite professionals in two different formats. According to Facebook’s blog post, if each chip had been worth a dollar, Pluribus would have won about $5 per hand and made about $1,000 per hour against the five human players
The Menlo Park, Calif. company says the AI bot uses a new online seach algorithm that can figure out its options by searching several moves ahead; Pluribus also uses “faster self-play” algorithms for games with uknown information, such as poker.
The players involved in the experiment seemed to be in awe of Pluribus.
“I was one of the earliest players to test the bot so I got to see its earlier versions. The bot went from being a beatable mediocre player to competing with the best players in the world in a few weeks,” Chris Ferguson, a World Series of Poker champion, said in the company’s blog post on the advance. “It was also satisfying to see that a lot of the strategies the bot employs are things that we do already in poker at the highest level. To have your strategies more or less confirmed as correct by a supercomputer is a good feeling.”
Another player said the AI bot made him improve his own skills.
“Whenever playing the bot, I feel like I pick up something new to incorporate into my game. As humans I think we tend to oversimplify the game for ourselves, making strategies easier to adopt and remember,” said Jimmy Chou, another elite poker player. “The bot doesn’t take any of these shortcuts and has an immensely complicated/balanced game tree for every decision.”
Unlike other similar AI projects, which can cost millions of dollars to achieve a major breakthrough, Facebook got a lot of computing bang for its buck with Pluribus — thanks to algorithmic improvements.
“We trained the blueprint strategy for Pluribus in eight days on a 64-core server and required less than 512 GB of RAM. No GPUs were used,” the company said in its blog post. “At typical cloud computing instance rates, it would cost less than $150 to train.”
Although the social network developed this experiment for poker, the research will help with building general AI that could thrive in multi-agent spaces, with humans and other AI bots, and further the company’s progress in this realm.
More information on this new AI has been published in Science.