The First Man-Machine
Humans deal computer a loss in poker challenge
R. Colin Johnson
July 26, 2007
The humans won and the computer lost in The World's First Man-Machine Poker Championship, which pitted two poker masters against a computer program, called Polaris, at the Association for the Advancement of Artificial Intelligence (AAAI) conference this week, in Vancouver B.C.
Phil "The Unabomber" Laak, a mechanical engineer and winner on the World Poker Tour, and Ali Eslami, a gaming consultant turned professional poker player, beat Polaris in the last two matches, even though Polaris won one match and played to a draw in the first two matches. Despite their loss, the team that programmed Polaris at the University of Alberta (Edmonton, Canada) was upbeat about the results.
"We are delighted with winning and drawing two of the four matches. For us that was milestone--a huge step forward," said professor Jonathan Schaeffer, leader of the University of Alberta team. "For the last two sessions, the humans hunkered down and poured all their concentration into the match and they deserved to win."
The bout, which was supervised by the Poker Academy, included 2000 hands of poker played in four 500-hand sessions, with the cards being "dealt" electronically from randomly generated card decks. Two copies of Polaris simultaneously played against the two humans, who were isolated in separate rooms, with the pot shared between the humans at the end of each session. Laak and Eslami received a $50,000 prize for winning the match.
Unlike chess, in which a timer is used to keep the humans from thinking longer than the computer, the poker match allowed the humans to contemplate their hands as long as they wanted. Consequently, the human poker players made no obvious errors, the way Gary Kasparov did against IBM's Deep Blue--an error that many experts said caused Kasparov to loose to Deep Blue a decade ago.
"If we had had a time limit imposed, we probably would not have won," said Laak after the match.
In the third match, which Polaris lost to the humans, a bug in the program was detected after game play was over, prompting bloggers to report that the humans probably would have lost otherwise. Schaeffer disagreed.
"The bug meant that [Polaris] did not play as strongly as it should have--the version of Polaris used in session three was a learning program, and the bug meant that some of the data that was being gathered during the match was not learned. But the fact [that there was a] bug is completely irrelevant. The humans deserved to win," said Schaeffer.
Schaeffer managed a team of a dozen professors, staff engineers and graduate students who worked together to craft Polaris. Most of the algorithms that constitute Polaris were run ahead of time to compile a giant database of possible hands. During the match, Polaris, which was run on an Apple MacBook Pro, mainly just looked up the pre-computed results to decide what to do next. The only "learning" version of Polaris was used in session three; it contained some significant real-time code.
"To an engineer, poker might not sound like a difficult problem, and by industrial standards the program is not very large--maybe 20,000 lines of code--but the key to Polaris is not the actual software, but the ideas it was based on," said Schaeffer. "For chess, Deep Blue became a more formidable competitor when it was run on faster hardware. But Polaris is not like that--it was the ideas behind the algorithms that counted."
The programming team plans to stick together to improve Polaris, based on its performance during the match, plus new ideas that Schaeffer hopes new graduate students will be able to contribute before next year's match.
"This was a grueling experience for our team," said Schaeffer. "But after the match our team was pumped up, and we are looking forward to competing with an improved version of Polaris next year.
See the results of all four matches at the University of Alberta Web site.