The 20-day competition between the University of Carnegie Mellon’s latest poker-playing computer program, Libratus, and poker pros Dong Kim, Jason Les, Jimmy Chou and Daniel McCauley has ended in a resounding victory for the computer, which completed the 120,000 hands with a more than $1.5 million chip lead over the humans.
In a Pittsburgh land casino, the AI program and its four poker pro opponents did battle for 20 days between the hours of 11am and about 10pm, playing NLHE, and with only hours of the contest left to play the humans have had to concede defeat.
Technology experts have pointed out that computers are already clever enough to best humans in skill games like chess, but poker presented far more of a challenge because it is a game of imperfect information, in which players cannot see each others hands, and bluffing is an integral part of the game.
Earlier programs like Claudico were unable to wholly surmount this obstacle, but in Libratus it appears that scientists Tuomas Sandholm and Noam Brown have gone a long way to overcoming the hurdles despite the magnitude of the challenge.
Sandholm has admitted that he “wasn’t confident at all” that Libratus would beat the poker pros, and the bookies appeared to agree, setting odds of 4 to 1 that the humans would prevail…they were wrong.
Libratus has considerably more computing power, and an enhanced algorithmic approach to the game, particularly the way it deals with imperfect or hidden information.
“We didn’t tell Libratus how to play poker,” Brown told local reporters. “We gave it the rules of poker and said ‘learn on your own.”
That enabled the program to continually refine and adjust its tactics; at the end of each day Libratus would be connected to the Pittsburgh Supercomputer Center’s Bridges computer in order to run algorithms and improve its strategy.
The humans would gather to discuss the day’s action every night in order to fine tune their own responses on the following day; these sessions often went on into the small hours of the morning, making for a gruelling schedule for the poker pros.
“Libratus turned out to be way better than we imagined,” Jason Les acknowledged, adding that it had been a difficult experience to play for 11 hours a day and take a continual beating. But he said that the experience had probably made him a better player with its constant need for focus against an aggressive and sometimes unorthodox machine.
Brown told local reporters that he was impressed with his creation’s performance, saying:
“When I see the bot bluff the humans, I’m like, ‘I didn’t tell it to do that. I had no idea it was even capable of doing that.”
The indications are that poker is just one potential and arguably trivial application area for Libratus; its algorithms are not exclusively poker-centric, which means that it could for example be deployed in business negotiations, medical and cyber security planning and even military programs where imperfect information is a factor.
Brown had the last word; “People have this idea that poker is a very human game and that bots can’t bluff. That’s totally wrong. It’s not about reading your opponent and trying to tell if they are lying, it’s about the cards and probabilities,” he said.