Opponent Modeling in Poker Games

Texas Hold'em poker is a popular game worldwide and it attracts increasing attention from community of artificial intelligence as a typical decision-making problem in non-deterministic and incomplete information environment. One of the key tasks to deal with the poker game is the opponent modeling, which aims to exploit the opponent weakness based on history behaviors. In this paper, we study mixed-method opponent modeling, one is Bayesian probabilistic model, one is the neural network (NN)-based prediction model, and the last is opponent type identifying model. Then, we combine these three methods to generate an integrated agent for opponent modeling. The opponents are categorized into 4 types according to their risk preference of strategies. The main step in Bayesian method is calculating the posterior distribution over opponent's strategy space and selecting the maximum probability. The main idea of NN-based method is using observation data to improve the prediction accuracy of opponent's hand. The main idea of opponent type identifying model is building a classifier with two factors. Finally, we design a simplified poker game to conduct experiment and demonstrate the effectiveness of our methods.