For the last two weeks, a team of top cash game players has been taking on the best poker playing artificial intelligence program developed by Carnegie Mellon University.
Claudico, as the artificial intelligence program is known, played a total of 80,000 hands of $50/$100 No Limit Hold’em heads up against Douglas Polk, Bjorn Li, Dong Kim and Jason Les, four of the top ten online heads up cash game players.
The human players beat the machine for a total of $732,713. The winnings from the game were play money, but the four also received $25,000 each put up as a prize by Microsoft and the Rivers Casino in Pittsburgh.
The game was structured so that two players were dealt one set of hands which they played against Claudico, and then the same hands were played by Claudico against the other two players. The structure, similar to duplicate bridge, was designed to reduce the effects of variance. The online video stream of the competition on Twitch soon became one of the top poker streams.
Even though the players won at a rate of 9.16 big blinds per hundred hands (bb/100), the researchers at Carnegie are celebrating,
Their program was developed to implement game theory in an artificial intelligence environment where the computer would learn and improve as it gained experience. Claudico was not designed to be a “poker bot,” poker just provides a good example of a game theory situation where decisions have to be made with incomplete information.
“It would have been no shame for Claudico to lose to a set of such talented pros, so even pulling off a statistical tie with them is a tremendous achievement,” said Tuomas Sandholm, the CMU professor of computer science in charge of developing Claudico.
The idea of a statistical tie is based on the total value of the bets—over $170 million—against the total winnings of $730k. However, this means of calculation looks to be at odds with the way poker players understand variance in the game.
To win at an average rate of over 9 bb/100 over 80,000 hands looks like a huge margin of victory. Perhaps the Carnegie Mellon researchers have underestimated the skill differential that a win rate of that magnitude implies.
The chance that a heads up player facing an opponent of equal skill wins at 9bb/100 over such a sample is very small—less than 5% probability, according to calculations by PokerDope’s variance calculator.
Even so, there is no doubting the scale of the achievement that Claudico represents for Sandholm and his team. Claudico is a development of their previous program Tartanian7 which won last year’s Association for the Advancement of Artificial Intelligence (AAAI) poker competition.
“The advances made in Claudico over Tartanian7 in just eight months were huge,” said Jason Les.
Computers have been highly proficient at Heads Up Limit Hold’em for years. Researchers at the University of Alberta announced earlier this year that their Cepheus computer program has learned Heads Up Limit Hold’em so well that the game has effectively been solved.
The Rivers Casino also enjoyed the benefit of a huge amount of publicity which the event generated. “Thanks to the online stream, the pros had fans rooting for them from all over the world throughout the challenge, in addition to the local players visiting our gaming floor,” said Craig Clark, general manager of Rivers Casino.