Researchers reveal how their AI beat human poker players

Researchers explain how they made their hard to beat poker-playing AI

Researchers at Carnegie Mellon University behind the Libratus artificial intelligence (AI) technology that beat all-comers at a poker tournament earlier this year have revealed how they did it.

Writing in the journal Science, the academics claim that they took a "three-pronged approach" that ultimately offers the AI "more decision points than atoms in the universe" - putting puny humans at a distinct disadvantage.

Earlier this year, Libratus dominated headlines when it beat four of the world's best professional poker players in a Texas-based game, winning chips worth just under $2 million in the process.

Tuomas Sandholm, professor of computer science at Carnegie Mellon University, teamed up with Phd student Noam Brown to detail how the artificial intelligence engine was able to achieve this "superhuman performance".

Their study shows that the technology was able to break "the game into computationally manageable parts and, based on its opponents' game play, fix potential weaknesses in its strategy during the competition".

The best AI is now able to trump humans in a range of challenging games, including checkers, chess and Go.

However, poker players have to contend with more hidden information, and this makes the game more complex for AI to navigate. The academics claimed that it as a groundbreaking moment for AI, being able to match humans in such scenarios.

During the 20-day competition involving 120,000 hands at Pittsburgh-based Rivers Casino in January 2017, Libratus solved "the primary benchmark and long-standing challenge problem for imperfect-information game-solving by AIs".

It went up against players individually in a two player game and went on to amass more than $1.8 million in chips. Libratus trumped the human poker players by "14.7 big blinds per game", which is fast, in layman's terms.

"The techniques in Libratus do not use expert domain knowledge or human data and are not specific to poker. Thus they apply to a host of imperfect-information games," said the researchers.

Although this experiment focused on strategic gaming, the researchers noted that the results are relevant for areas such as business negotiation, cybersecurity, finance, strategic pricing and military applications.

"The techniques that we developed are largely domain independent and can thus be applied to other strategic imperfect-information interactions, including non-recreational applications," Sandholm and Brown concluded.

"Due to the ubiquity of hidden information in real-world strategic interactions, we believe the paradigm introduced in Libratus will be critical to the future growth and widespread application of AI."