Facebook and CMU's Poker AI Triumphs Over Five Pros Simultaneously

Facebook and Carnegie Mellon University Develop Groundbreaking AI Poker Bot, Pluribus

Facebook, in collaboration with Carnegie Mellon University, has introduced a revolutionary artificial intelligence bot named Pluribus that has successfully defeated top professional poker players. Unlike previous AI, which excelled in one-on-one games, Pluribus stands out as the first bot to triumph in multiplayer scenarios, showcasing its capabilities in no-limit Texas Hold'em. The bot's remarkable feats included defeating both five AI opponents and a single human, as well as one bot facing off against five human players.

The researchers published their findings in Science, highlighting the creation of a multiplayer poker bot as a significant achievement in artificial intelligence. While games like chess and Go provide complete visibility of the playing field, poker involves hidden information—specifically, opponents' concealed cards. This complexity introduces unique strategies, including bluffing, which AI has historically struggled to master.

Bluffing adds another layer of difficulty; it can drastically influence the outcome of a game, yet excessive bluffing can render a player's strategies predictable. Therefore, Pluribus must carefully balance its bluffs against betting on genuinely strong hands.

Pluribus is an advanced evolution of Carnegie Mellon’s earlier bot, Libratus, which had previously conquered professional players in heads-up matches. Utilizing an innovative online search algorithm, Pluribus evaluates possible moves several steps ahead, rather than only at the game's conclusion. Additionally, it employs faster self-play algorithms tailored for games with hidden information, enhancing its ability to adapt and learn efficiently.

This efficiency allows Pluribus to operate on minimal resources—less than 128 GB of memory and just two GPUs—contrasting starkly with the 1,920 CPUs and 280 GPUs used by AlphaGo during its historic match against Go champion Lee Sedol. Remarkably, Pluribus plays an average of 20 seconds per hand, twice as fast as human pros.

Over 12 days and 10,000 hands, Pluribus faced several notable professionals, including World Series of Poker Main Event winners Chris Ferguson, Greg Merson, Darren Elias, and Jimmy Chou, all of whom have earned over $1 million in professional play. The stakes were high, and the Monetary incentive prompted the players to perform at their best.

"If each chip was worth a dollar, Pluribus would have averaged around $5 in winnings per hand, equating to approximately $1,000 per hour against five human players," Facebook reported, calling this margin a significant victory by poker standards.

The pros were captivated by Pluribus's strategic approach, particularly its unconventional moves—like betting immediately after calling in a previous round. "Playing against the poker bot was incredibly fascinating. Its strategies included plays that humans typically overlook, especially concerning bet sizing," remarked Michael Gagliano.

Ferguson added, "Pluribus is a challenging opponent. It's tough to predict its hand and it excels at making thin value bets on the river, effectively extracting value from its strong hands."

The emergence of Pluribus marks a pivotal moment in AI development, broadening not only the horizons of poker but also the potential applications of AI in complex decision-making environments.