Bronze Medal at Computer Poker Competition 2014

In this year's Computer Poker Competition our bot Learn2KEmpf (team KEmpfer) finished 3rd both in the total bankroll and bankroll instant run-off competition for 3-player limit Texas Hold'em poker. We were only beaten by bots implementing a near Nash equilibrium strategy, which is proven to be a non-exploitable strategy.

On the one hand, this result was expected, since our approach uses machine learning in order to mimic a certain player. Particularly, we trained our poker player on history data of this (and last) year's winning bot (Hyperborean). Please find further details on the bots at the organizers' web site and particularly on our approach in the Bachelor thesis of Theo Kischka.

On the other hand, the result is remarkable since the difference to the Nash bots was relatively small although our approach uses only a few million observed hands, is fast in training, and produces small models, whereas the Nash approaches have to simulate billions of hands, which takes long, and the produced models can need up to gigabytes.

We also participated in the heads-up no-limit competition with a baseline bot using hand-crafted playing rules as preparation for the next competition. As expected, we only achieved a low rank.