First-Ever Computer Versus Human Poker Competition Ends in Statistical Tie

May 8, 2015

Dean of the School of Computer Science at CMU, Andrew Moore (left) looks on as Doug Polk collects his "winnings" from two weeks of competition with a computer program.
Credit Deanna Garcia / 90.5 WESA

Its name is Claudico, and it’s the first artificial intelligence (AI) program of its kind to take on humans in poker matches: 80,000 hands to be exact.

Four of the world’s top poker players spent two weeks at Rivers Casino. Three of the four pros had higher winnings than Claudico, but their $732,713 collective lead wasn’t large enough to be considered scientifically reliable.

Still, Claudico’s performance is seen largely as win for artificial intelligence, which thus far has learned to play chess, drive cars and win at general knowledge competitions such as Jeopardy.

“This actually was a milestone event for the field of artificial intelligence,” said Andrew Moore, dean of the Carnegie Mellon University School of Computer Science. “This was the first major fields test of artificial intelligences. The kind of artificial intelligences here were the ones which can deal with the fact that humans are trying to deceive them and artificial intelligences which can actually deceive the humans back.”

That is why poker was an important test for Claudico, Moore said. As AI is progresses, he said they need to be able to deal with the fact that the world is full of misleading information and deliberately misleading information. Moore said Claudico beat all other computer AIs in the world and is part of a push at CMU to develop the technology as “AI starts to run the world.” He said right now there are AIs operating on our behalf and protecting us.

“There are tens of thousands of AIs right now busy in the background protecting everyone from e-mail spam attacks, those AIs need to improve and get better,” said Moore. “There are AIs doing negotiations between the United States and China. They need to be able to actually negotiate using the same mathematics that you use to play poker.”

Claudico’s strategy was created using algorithms rather than trying to program in human poker knowledge, and one of the developers said the algorithms used are not actually designed for poker.

“They’re for general imperfect-information games, and we’ve done a lot of work on applying artificial intelligence techniques to negotiations, to auctions, to cyber security, and I have a new project right now for which we are gathering grant funding and gifts to get started on applying this to medical applications,” said Tuomas Sandholm, a CMU professor who directed development of Claudico.

Back to poker, the final tally for the 80,000 games played were: Bjorn Li had an individual chip total of $529,033; Doug Polk had $213,671; Dong Kim had $70,049. Jason Les trailed Claudico by $80,482. The players are ranked among the world’s top 10 professionals in two-player no-limit Texas Hold’em. The games weren’t played for actual winnings, but the players each received appearance fees donated by Rivers Casino and Microsoft Research.