2009 Computer Poker Winners:
Limit Bankroll:  MANZANA
MANZANA is a simple poker bot for Limit Texas Holdem.  It consists of a feed-forward neural net with one hidden layer and three output neurons.  The input neurons encode relevant features of the current game state such as the hole cards, board cards and previous betting, while the output neurons are trained to give P(fold), P(call), P(raise) Š the probability that the bots should fold, call or raise, as a function of the current game state.
Training is done using standard learning algorithms, while the training examples are constructed from the hand logs of the winner of last yearÕs competition.

Limit Runoff: GGValuta
The general algorithm used to compute the equilibrium is a counterfactual regret minimization one. The abstraction model used something a bit changed for every street: On the preflop we stored a node for every possible situation (no abstraction) and on the flop/turn/river we used a modified k-means algorithm to cluster the hands.
For every hand we mapped it onto a N-dimensional point with a different function for every street. These were chosen mostly by intuition and every point had coordinates like (E[HS], lowest(E[HS]), highest(E[HS])), where lowest and highest show the range in potential if you draw another card.
For the river we used E[HS] with some biased average of potentials on past rounds.

No Limit Bankroll:  Hyperborean-BR
Hyperborean-BR employed a variety of techniques designed to exploit the traditional method of translation used by agents. This involves manipulating the pot size in ways that are indistinguishable to its opponent. In order to exploit its opponent, it performs exploration at the beginning of the match to create a rough model of its opponent.

No Limit Runoff:  Hyperborean-Eqm
Hyperborean-Eqm was created using the same techniques used by the University of Alberta as in the past, with the exception that it now uses soft translation. Additionally, the methods used to create the strategy have been further optimized for no-limit play.

3 Player Bankroll and Runoff:  Hyperborean-Eqm
This entry was constructed by running the Counterfactual Regret Minimization (CFR) technique on an abstract 3-player game for several weeks.  Because the 3-player game is so large (in terms of game states), a very coarse-grained abstraction was implemented.  Namely, the abstraction consisted of 16 hand strength squared buckets per round where the agent forgets its buckets from previous rounds (i.e. imperfect recall)