Research on the DouDiZhu's playing strategy based on XGBoost

Abstract

In recent years, imperfect information games have received extensive attention in computer game research. As a three-person imperfect information game in China, DouDiZhu not only has competitive relationship but also has cooperative relationship between the players, which makes the model more complicated, hence it has high research value. In this paper, the DouDiZhu's playing strategy is converted into 182 legal playing multi-classification problems, by extracting the characteristics of winning data of different players, using the XGBoost model and setting reasonable model parameters. In addition, to make the strategy of playing cards more reasonable through phased training, character training, and rule correction methods. In this paper, a new data representation method of DouDiZhu game is proposed. The XGBoost model is innovatively used to solve the poker game problem. By adjusting the parameters and multi model training, a better card playing strategy is obtained. The experimental results show that the playing strategy predicted by the playing strategy model proposed in this paper is basically consistent with the human playing strategy, and has a good playing strategy for different situations. The method in this paper has achieved a third-place excellent result in the Chinese University Student Computer Game Contest 2019.