Libratus Wins HPCWire Award for Defeating Poker Pros

The poker playing AI software Libratus won the HPCWire Award in the “Best Use of AI” category at the SC17 supercomputing conference in Denver, CO. In January, Prof. Toumas Sandholm’s and Noam Brown’s bot played a combined 120,000 hands of heads-up NLHE against four poker pros and won decisively.

Toumas Sandholm, professor of computer science at Carnegie Mellon University first tested the poker skills of his AI creation in 2015. Named Claudico back then, the computer took on a team of four poker players - Doug Polk, Dong Kim, Bjorn Li and Jason Les. Each player played 20,000 hands against the bot over the course of 13 days. At the end of the challenge the humans came out on top, having a 732,713-chip difference in their favor, but somewhat controversially the scientists declared it a “statistical tie”. 

Two years later, this year in January Claudico's "big brother" Libratus had a new match-up against four human players. Jason Les and Dong Kim returned, but Polk and Li were replaced by Jimmy Chou and Daniel McAulay. Prof. Sandholm and his student, Noam Brown managed to significantly improve their AI during the two years of development, as it was attested by Les in a video the university posted on their YouTube channel halfway through their experiment.

“It’s quite a bit better than Claudico in 2015. It made pretty large mistakes at times, you don’t see that with Libratus. It is much more calculated and much more tough, you really have to pry every chip you can out of Libratus’s hand, and you have to do it with skill, not counting on his mistake” - said Les.

Libratus eventually beat its human opponents by $1,766,250 in chips, a statistically relevant difference this time. A few months later Lengpudashi, Prof. Sandholm’s latest poker AI also won against a team of six Chinese poker players heads-up. With these results, one may conclude that computers surpassed humans in one-on-one poker - full-ring games remain unsolved by the bots, however.

Libratus was recently awarded the HPCWire Award for the “Best Use of AI” in the 2017 SC17 event, the world’s biggest annual conference and exhibition for high-performance computing, likely because its achievements have implications beyond poker.

An AI that can solve incomplete information problems like poker can be applied in other areas of life, such as negotiation, cyber security or medical treatment planning according to Sandholm, since all of those fields deal with hidden information.