AI & Poker: A Smart Bet

It's not everyday you get to give your alma mater a pat on the back. While a team from the University of Alberta didn't win the curling championship or debate competition, it did win the first AAAI Computer Poker Competition held in conjunction with this year's American Association of Artificial Intelligence conference in Boston.

The U of A computer program defeated all other programs in a two tournament format of one-on-one Texas Hold 'Em. The U of A poker bot won every match it played and amassed by far the most virtual money of any competitor. All in all, there were five competing bots, including three from collegiate teams and two from individuals:

• "Hyperborean" written by Michael Bowling, Martin Zinkevich, Darse Billings, Nolan Bard, Morgan Kan, Michael Johanson, Robert Holte, Jonathan Schaeffer, Neil Burch, Carmelo Piccione, and Finnegan Southey, of the University of Alberta.
• "GS2" written by Andrew Gilpin and Tuomas Sandholm from Carnegie Mellon University .
• "Monash BPP" designed by Ann Nicholson, Kevin Korb, and Steven Mascaro of Monash University (Australia).
• "BluffBot" written by Teppo Salonen from Irvine, California.
• "Teddy"by Morten Lynge in Ikast, Denmark.

Each bot won at least one series.

The game of one-on-one Limit Texas Hold'em was played in a pair of computer tournaments. The first event used a normal pace of about one second per decision, as seen in games played by humans. The second event allowed a much slower pace, in the hope that it might produce a higher level of play. Somewhat surprisingly, the winning program from Alberta made every betting decision instantaneously, even in the slower event.

The AAAI event is the most intensive competition there is for poker programs. To reduce the effects of luck, more than a quarter million games of poker were played in the two tournaments. To make the results even more reliable, every series of deals was played twice, with both competitors getting a chance to play each side of the cards. This is possible because programs can be restarted with no memory of past events. The "duplicate matches" ensure that both programs have nearly equal opportunities, despite the lucky outcomes that can occur in each game.

To level the playing field, every bot was run on an identical machine. The greater degree of control let participants force the bots to "forget", resulting in better results overall. The competition was run on 16 Windows machines located in the at the University of Alberta. There was one server machine, and 14 client machines ( one extra machine in case one of problems). The computers were 3.4-GHz P4 machines running Windows XP Professional, with 1 GB of RAM, and an 80-GB hard drive.

"Poker is a game that involves skill, chance, and many forms of uncertainty", said Alberta's professor Jonathan Schaeffer who I interviewed last year in this brief podcast. "It is a great problem for artificial intelligence, and we stand to learn a lot from competitions like this."

"We've been writing good poker programs for many years", added Darse Billings, the lead designer for the Alberta team, "but we weren't overly confident, because there is still a lot of room for improvement. Poker is a nice well-defined problem for studying some truly fundamental issues, like how to handle deliberate misinformation, and how to make intelligent guesses based on partial knowledge," explained Billings. "Good solutions in this domain could have an impact in many other computer applications."

Posted by Jon Erickson at 08:47 AM  Permalink