The 2006 AAAI Poker Championship (American Association of Artificial Intelligence) is the first of an annual championship which pits the best poker-playing programs in the world against eachother. This event is the most intense competition available for poker programs. In it's inaugural year, the 2006 AAAI was held at Carnegie Mellon University in Pittsburgh,USA.
In order 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, the matches featured an interesting twist: 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.
Strong poker play does not necessarily mean the player has to be better than all opponents. In addition to beating other competition, it also includes how much money a player can extract from average opponents. Instead of attempting to assign a distribution to this importance, the 2006 AAAI competition consisted of two events to completely separate the analysis of these abilities.
5 bots competed in the two events. 2 of the teams were unable to compete in both events, and so there were 4 competitors for each competition.
BluffBot: written by Teppo Salonen in Irvine, US
GS2: submitted by Carnegie Mellon University in Pittsburgh, US
Hyperborean: submitted by Univiersity of Alberta in Edmonton, Canada
Monash BPP: submitted by Monash University in Victoria, Australia
Teddy: written by Morten Lynge in Ikast, Denmark
This event tests the importance of extracting the most money from average opponents. The competition consists of 40,000 hands of limit heads-up hold'em, with 7 seconds to act. The bots play in a round-robin, and the memory is reset after the 1,000 hands. After 20 rounds, each match is duplicated with opponents sitting in opposite seats. The winner is the bot with the highest bankroll at the end of the round robin.
 |  |  |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
Each graph contains two sets of 20 matches. Since the memory was reset after each match, it shows the dominance a bot had gained by 1000 hands. The most interesting match was Hyperborean vs BluffBot.
The green bar indicates the original match, and the orange one is the duplicate. As you can see, the winnings of the duplicate round was nearly a complement of the first round, as each bot would capitalize with the luckier cards. However, over the course of the rounds, Hyperborean would win a little more and lose a little less, resulting in an overall win rate.
The play between Teddy and Monash was also of interest. Against BluffBot and Hyperborean, Teddy faired significantly better than Monash. Yet when these two strategies went up against eachother Monash BPP wiped the table with the most convincing win of the competition. Winning 1.17 small bets per hand equates to taking an average of almost $12 per hand in this $10-$20 limit match. When looking at the detailed graph, each round starts with a level plateau before drastically climbing. So clearly, Monash figured Teddy's strategy fairly soon in the match and looked to exploit it.
Overall, each of the bots finished in distinctly different win rates. Hyperborean won all of the match-ups and took first with 0.39 small bets per hand. While BluffBot held it's own against Hyperborean, it placed second as it did not do as well against Monash BPP, and actually lost against Teddy. Monash placed third due to a large part from it's crushing defeat over Teddy.
The series competition pits the bots against eachother in a win-loss format. This more traditional competition featured longer decision times and a longer period for the opponents to exploit eachother. The competitor with the higher bankroll at the end of the matches is given a win. For this event, the memory was only reset at the end of the match, with each match consisting of 6,000 hands. Each match is then duplicated with opponents sitting in opposite seats. The time limit to act in this event is 60 seconds for the 2006 AAAI Poker event. The competitor with the best overall record wins.
| | |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
 |  |  |  |
The winner of each matchup could be determined rather early in these matches. The most interesting matchup occurred between BluffBot and GS2, perhaps more by luck than by the AI matchup. GS2 held the lead in the first round for over 4000 hands, before BluffBot gained the slight lead. In the end of the first round it was almost a dead heat. However, once the duplicate match was played, it was clear which seat had the more favourable cards, as BluffBot proceeded to take a lead in the round and not look back. Still, the match was close, and it goes to show the value in the duplicate match system.
The 2006 AAAI Poker title definately belonged to Hyperborean. The program went undefeated in the 2006 series event and defended the reputation the University of Alberta has earned as leading the field in poker AI.
Once again, BluffBot had a strong showing and finished in second place. This entry was all the more impressive as it was completed by a single programmer and managed to beat the highly touted entry from Carnagie Mellon University.
The Carnagie Mellon entry failed to meet expectations with its third place finish. The well-publicized GS2 bot had very high expectations and only won 1 match, against the fourth place finishing Monash entry.
The Monash BPP (Bayesian Poker Player) placed last in this event. The highlight for Monash University in the 2006 AAAI Poker championship would definately have come in the Bankroll competition where Monash recorded the most lobsided defeat in the event.
Teddy was another single programmer entry. This bot did not fare as well in the competition, and was not entered into the Series tournament.
An earlier version of the winning program is included in the commercial software 'Poker Academy Pro', under the name 'Sparbot'. Owners of the software can connect their own bots to Poker Academy using a published API for testing and evaluation.
"We thought that would be a huge disadvantage, because the other teams could practice against our programs 24 hours a day, every day", said professor Michael Bowling, who heads up the Alberta research group.
The components of Hyperborean itself are expected to be added to Poker Academy Pro in a later release.
Naturally the advanced poker AI developed in cooperation with the University of Alberta is a great sense of pride in our product. Two of the more intriguing quotes from the competition come from the UofA team following the event:
"We've been writing good poker programs for many years", said 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".
"We used some of our older technology", added Neil Burch, the lead programmer for the Alberta team. "We didn't have much time to prepare for the competition, so we used programs that have been thoroughly tested, rather than our strongest bots".
So you know that the best is definately yet to come. The UofA games group has been pushing the limits in poker AI for some time now. It is exciting to see the expectations and projects they have on themselves, and we look forward to the new levels of AI that can be achieved with co-operation between the AI that comes from this fine University and the feedback and user-base that Poker Academy can provide.
