A Rule Embbeding Method of Winrate Approximation for Texas Hold’em

Expected winrate is one of the most important features in Texas Hold’em, and it is the foundation of state estimation and decision-making. Calculation of the expected winrate needs to consider the unknown opponent’s hole cards and the undealt board cards. The huge combination space of these unknown cards makes the enumeration method time-consuming. A solution is using neural network to approximate the true winrate, but the calculation of the input features from cards information is fairly complex. In order to avoid feature engineering and improve the computing speed, a new neural network method based on rule embedding is proposed. With the elaborately designed convolution/max-pool kernel and a new proposed hypo-max-pool operation, the cards category rules are efficiently embedded into the winrate fitting network with several nodes. The result of experiments indicates that our rule embedding neural network method greatly outperforms the feature engineering method on the prediction accuracy and the computation efficiency.