Research on improving Mahjong model based on deep reinforcement learning

Mahjong is a popular incomplete information game. There are many scholars dedicated to Mahjong research. To improve the game ability of existing Mahjong models. A method based on deep learning and reinforcement learning is proposed. Firstly, a Mahjong program (MPRE) is designed. MPRE is used to generate training data for deep learning and as a comparison program for MPRE_RL, respectively. Secondly, with the feature extraction capability of deep learning, the game ability of MPRE is transformed into a deep learning model. Thirdly, the deep learning model is continuously improved by reinforcement learning. To improve the training speed and stability of reinforcement learning, some improvements are made in the environments and rewards. Finally, the results show that MPRE_RL improved by using the proposed method get a certain enhancement in offensive (27.1% of winning rate) and defensive (19.5% of win by discard rate) aspects compared with MPRE.