A Texas Hold’em decision model based on Reinforcement Learning

Texas Hold 'em is a typical example of computer incomplete information game. The traditional machine learning method has been unable to deal with the huge search state space of Texas Hold 'em. In this paper, the value-based reinforcement learning algorithm is adopted, which can deal with the huge amount of data without manual extraction of data features, and can realize the unsupervised training of the model. However, the value based reinforcement learning algorithm has the problems of too complex acquisition process and over estimation of DQN algorithm. Therefore, this paper introduces DQN-S model, which combines reinforcement learning algorithm with Monte Carlo game search, and integrates Sarsa algorithm to solve the above problems to a certain extent. Finally, experimental data show that the over estimation effect of DQN algorithm is weakened to a certain extent after the Sarsa algorithm is incorporated. The average return of each game of DQN-S model and model is more than 5 chips, and the average return of each game of DQN-S model is more than 3 chips in the game with its strongest training version.