The multi-scale backbone neural network model for Chinese standard mahjong
DAI Junxue；LI Xiali；LIU Bo；WANG Zhaoqi；
Chinese standard mahjong involves multiple rounds，an immense state space，81 different categories of tiles，and a complex winning strategy，conventional neural networks struggle to express and fit the intricate data of Chinese standard mahjong. For the first time， a multi-scale backbone deep neural network has been used to construct mahjong AI algorithm，in order to better capture local and global features of national standard mahjong，suitable for processing complex data and making more accurate game strategies. Based on the game data from the IJCAI 2020 Championship， the training dataset undergoes data augmentation. Using the augmented data，the proposed algorithm underwent 5 days of supervised learning training on an NVIDIA GeForce RTX 3090 Laptop GPU. The trained model has 52 million parameters，achieving an action accuracy of 93.47%，discard accuracy of 83.93%， and declaration accuracy of 97.56%. The proposed model was deployed on Botzone platform developed by Peking University and it has entered the top 1% of the leaderboard.
