"Hyperborean" Wins Poker Bot Competition

BOSTON -- The University of Alberta's "Hyperborean" was the best poker playing computer program at the 21st National Conference on Artificial Intelligence (AAAI-06) at the Seaport Hotel and World Trade Center last week.

Hyperborean defeated four other "poker bots," including teams from Carnegie Mellon University and Monash University in two competitions. In the bankroll competition, the bots were dealt 1,000 heads up Limit Hold'em hands versus each competitor. At the end of the series of hands, the computers' memories were swept clean, and they were dealt identical hands but in opposite positions.

Each program played 20 duplicate matches against every competitor, with Hyperborean averaging a profit of 0.0514 small bets in each hand against Bluff Bot, 0.4067 small bets against Teddy and a whopping 0.7227 small bets per hand against Monash.

The bots also competed in a series competition which did not track how much money the programs won or lost, instead just keeping a win/loss record for each competition.

Before the results were announced, a monitor displayed sample hands from the competition, showing first how one bot reacted in a given situation, and then how the other did. In most cases, Hyperborean was able to maximize its wins and minimize its losses.

The competition was purely academic, says University of Alberta team leader Michael Bowling.

"We're in academia, we gave up on money a long time ago," Bowling said.

The technology used to win the competition, however, is quite similar to what you can find to learn how to play Texas Hold'em on PartyPoker.com. Bio Tools, the company which licenses the "Poker Academy" program available on PartyPoker.com, employs many former students from the University of Alberta's program and maintains a relationship with the university.

"I don't think we have to be the bad guys in the online poker world," Bowling said. "Certainly there are people who are building bots and trying to sell them to people purely to go on line and make money while they are sleeping. That isn't our desire. What we've done is actually about helping poker players play better by having these bots to learn from."

However, judging by the results, the bots themselves are only good against other bots, or against really bad humans. Hyperborean played a heads up match against Canadian pro Gautum Rao a few years ago, and while Rao adjusted to Hyperborean's play and was able to exploit its weaknesses, each hand is a clean slate for the poker bot.

"The future of this competition is to focus on opponent modeling solutions," Bowling said. "We have a program that actually can learn, but it takes about 40,000 hands to learn something about a player, which is useless when it comes to playing against humans, because no one has that kind of patience. The whole competition is geared towards expecting people in the future to start doing more opponent modeling."

The competition was also limited to heads up limit play, and while Bowling believes that the contest may expand to incorporate full ring games in the future, he doubts that No Limit Hold'em will be included.

"Although no limit has a lot of popularity, the downside to it is it amplifies the luck in the game," Bowling said. "This is an academic endeavor, so we want to make sure that all of our results are statistically significant. We don't want to just flip a coin and declare a winner."

In his previous life, Aaron Todd was a sports journalist by day and a poker player by night. He can now be found covering the poker beat for Casino City and making horrendously unsuccessful bluffs in his home game.
Write to Aaron at aarontodd@casinocity.com.