Research on DouDiZhu Model Based on Deep Reinforcement Learning
Chuanyu Jiang, Yajie Wang, Song Liu, Xinghui Zhang, Yuhan Yi · 2023
Computer game is the carrier of artificial intelligence (AI) research. DouDiZhu is a multi-player imperfect-information game, which has problems of enormous hidden information, high complexity, and coexistence of competition and collaboration. In this paper, a DouDiZhu AI based on deep reinforcement learning (DRL) is designed to solve the above problems. First, a heuristic-based DouDiZhu agent is developed and generates training data by self-play. Second, the DouDiZhu model is trained by supervised learning (SL) method. In this process, the role information is innovatively encoded, and the game strategies of various roles can be learned by only one model. This approach solves the problem of tedious training process and high model complexity caused by multi-modeling. Finally, the decision-making ability of the SL model is further enhanced by DRL method, where a new node reward function is designed to enrich reward signal. The experiment results demonstrate that the proposed method is feasible and effective. The runner-up agent of the 2022 IEEE Conference on Games is defeated by the proposed approach.