AI on Bridges Overcomes Top Pros in Multi-Player Poker
July 16, 2019
Artificial intelligence (AI) research took a big step forward when a CMU AI program overcame the world’s best professional players in a series of six-player poker games. Developed at the Carnegie Mellon School of Computer Science, the Pluribus program runs on PSC’s Bridges system.
The victory marks a major improvement from the CMU team’s 2017 AI, called Libratus. That earlier AI surpassed four of the world’s best in one-on-one poker while also running on Bridges. The 2017 victory had been the first in which an AI overcame top players in an incomplete-information game. In such games, players have private information not available to their competitors, and actively work to deceive each other. Such games are useful for AI research because they are more like real-world problems than set-piece games in which players have the same information.
The transition from head-to-head to multi-player poker required a stronger AI approach. Pluribus taught itself to play Texas Hold’em poker before facing the pros. In doing so and learning from subsequent competition with humans, Pluribus discovered strategies that humans do not normally employ. Its ever-changing random approach to strategy overcame humanity’s best.
The CMU team plans to apply Pluribus’s insights far afield of poker, with promising possible uses in business negotiations, medical treatment planning and intelligence. They reported their results in the prestigious journal Science last week.