Solving the playing strategy of Dou Dizhu using convolutional neural network: A residual learning approach

摘要

Poker is the typical game of incomplete information, and remains a longstanding challenge problem in artificial intelligence (AI). The game of Dou Dizhu has been viewed as a thorny topic in AI since it is featured with hidden information and large branching factors, and the cooperation and competition should also be handled. In this article, deep learning is adopted to train a supervised learning playing strategy network (PSN) for Dou Dizhu directly from expert human playing. Through experiments, it was found that the sample design with the appropriate historical playing hand sequence and more features of the playing situation, can help the PSN learn more competitive and accurate playing strategies faster. In the online game platform, the strategy network-based game agent reaches an average winning rate of 52.22% against the human players. In addition, the analysis of the gameplay data against human players shows that the playing strategy network has learned the rules of playing and the characteristics of card recognition and reasonable demolition, cooperation and reasoning. Finally, we improve the performance of the PSN in the aspect of sample design. Then, the experimental results show that with proper marking of the number of remaining hands, the performance of the PSN can be enhanced.