A robot is trouncing some of the world’s best poker players in a marathon 20-day session, using levels of intuition and deception that have big implications away from the gaming table.
The Libratus computer program from Carnegie Mellon University will play a total of 120,000 hands of no-limit Texas Hold’em against four professionals competing for a share of $200,000 prize money.
In the first day of the Brains vs AI tournament at the Rivers Casino in Pittsburgh, the robot collected $81,716 to the humans’ $7,228. The machine doubled its lead on the second day and was extending its advantage on the third, outfoxing its opposition with unpredictable behaviour, including small wagers and huge over-bets.
It is still all to play for, but it is clear that the robot stands a much better chance than its predecessor, which was roundly defeated two years ago. Experts said that the artificial intelligence (AI) appeared to have overcome specific weaknesses that humans had exploited in Claudico, the 2015 system. For example, the machine takes its own cards into consideration when deciding to bluff, giving stronger opportunities to exploit weak hands.
Significantly, the AI developed its knowledge by analysing the rules of the game, not by trying to copy humans. Libratus honed its strategy using about 15 million core hours on the Pittsburgh Supercomputing Centre’s Bridges computer and continues to “train” each night while its human opponents are asleep or drowning their sorrows.
Poker is more complex than the other games in which robots have taken on man, with levels of uncertainty that are found in other real-life situations that AI has not yet mastered. Players cannot see their opponents’ hands, meaning that, in contrast to games like chess or Go, not all of the information in the game is available to them.
Tuomas Sandholm, a professor at Carnegie Mellon and one of the system’s creators, stressed that Libratus was not specifically a poker program. Its algorithm could be applied to other situations that involved incomplete and misleading information, such as business negotiations, military strategy, cybersecurity and even medicine.
His team have competition. In research posted online but not yet peer reviewed, a team of Canadian and Czech academics led by Professor Michael Bowling of the University of Alberta said that their DeepStack program had played almost 45,000 hands of poker against several players, beating them easily.
They compared DeepStack’s technique to a good human player’s “gut feeling”, although it is based on the opponent’s playing history rather than body language.