Fell Omen 2 At The AAAI 08 Computer Poker Competition
The format of the 2008 computer poker competition was similar to last year. Carnegie Mellon hosted this year’s competition, and Andrew Gilpin put in a good deal of work as chair of the contest. In addition to head’s up limit, this year players were fielded for head’s up no-limit, and 6-player ring.
I again competed in the head’s up limit competition, submitting the successor to INOT, Fell Omen 2.
Fell Omen 2
Fell Omen 2 has two primary improvements over its predecessor INOT. Firstly it has improved hand bucketing, meaning that it classifies hands as similar in a more intelligent, and finer grained way. Secondly, it takes into account the flop’s public information. Fell Omen 2 treats different flop boards differently, so given the same hand strength, Fell omen might make a different decision based on whether the board on the flop is full of high cards or whether the board cards all have the same suit. Results
Nine bots were entered into the competition, and the UoA team again placed first. Interestingly, there was a three way tie for second place between Hyperborean08-on, Fell Omen, and GGValuta. Examining the win rates it is apparent that none of the comparisons between these three bots reaches the level of statistical significance.
Place Name Organization Win Rate against Fell Omen 2(dollars/hour) 1 Hyperborean-eq UoA $20.00 2 Fell Omen 2 UCSD $0.00 2 Hyperborean-on UoA $4.00 2 GGValuta UB -$3.00 5 GS4-Beta CMU -$20.00 6 PokeMinn 2 UoM -$153.00 7 PokeMinn 1 UoM -$154.00 8 GUS GSU -$467.00 9 Dr. Sahbak Ind -$532.00 * Win rate assuming 10/20 bets and 100 hands per hour
University of Alberta: Computer Poker Research Group –
The researchers at the UoA have again fielded two world class bots. Hyperborean08 plays a fixed equilibrium strategy similar to Fell Omen. Unlike Fell Omen it does not make some simplifying assumptions, causing it to take up more than 2 GB of space (vs. Fell Omen 2’s 100MB). A version of Hyperborean, dubbed POLARIS, recently won the Man vs. Machine Challenge competing against the top professional poker players. The difference between Hyperborean08-eq, and Hyperborean08-on is that Hyperborean08-on plays more aggressively, but slightly off the equilibrium.
Team: Michael Bowling (PhD), Duane Szafron (PhD), Nolan Bard, Neil Burch, John Hawkin, Michael Johanson (MS), Morgan Kan (MS), Nick Abou Risk, Dave Schnizlein, Kevin Waugh, Martin Zinkevich (PhD), Bryce Paradis
University of Bucharest -
Four talented undergraduate students fielded this tough competitor. GGValuta is based on a similar algorithm as the UoA, but uses different hand evaluation and internals.
Team: Mihai Ciucu, Stefan Popescu, Mihaita Alexandru Leoveanu, Diana Dulgheru
Carnegie Mellon University –
CMU is back with an improved version of their system. Their bot uses an automated method to adaptively abstract the game tree, and then solves this abstracted game with linear programming. CMU previously placed 3rd in 2006 and 2007.
Team: Andrew Gilpin, Tuomas Sandholm (PhD), Troels Bjerre Sørensen (PhD)
Georgia State University -
No information at this time.
Team: Nicholas Larson
University of Minnesota –
This team returns with two new versions of their poker bots. PokeMinn is a learning bot that adapts its play in response to its opponent. Using genetic algorithms, PokeMinn attempts to exploit weaknesses, and in 2007, though they place behind UoA, CMU and INOT, they exhibited much better play at the end of the match as opposed to the beginning. PokeMinn posted a respectable 4th in last year’s competition.
Team: Mohamed Elidrisi, Brett Borghetti (PhD)
Roy Fox -
No information is currently available about this independent competitor.
Team: Roy Fox