Game theory is widely present in various fields of life and can be divided into complete and incomplete information games. Incomplete information games are difficult to solve due to information uncertainty. Although artificial intelligence has made remarkable achievements in the field of complete information games, its progress in incomplete information games remains limited. Mahjong, as a typical incomplete information game, has interestedmany researchers; however, current Mahjong AI still faces numerous challenges and struggles to effectively handle information uncertainty and complex decision-making problems.
This paper proposes a novel neural network, MJ-Net, which integrates ResNet-CBAM, LSTM, and attention mechanisms to construct an opponent model for Mahjong games. Based on MJ-Net, a residual tile prediction model (for the available tiles in the current field) and a fan-type prediction model (for the possible information on the opponent's hand) are developed. By perceiving changes in the game state and sequences of opponent actions, these models dynamically reassess tile-drawing probabilities and opponents' potential hand information to optimize expected decision values.
Additionally, leveraging domain knowledge and opponent modeling, a defense model is constructed to capture opponents' strategic information, enabling the prediction of their possible actions and fan types. This allows for dynamic adjustment of strategies to reduce risks and increase returns. Finally, a hybrid decision-making framework is established by integrating the game tree with the opponent model and the defense model derived from it. This composite framework optimizes the game tree's search strategy, path evaluation, and pruning, thereby improving the overall decision quality and achieving a balance between offense and defense.
Experimental results demonstrate that MJ-Net performs effectively in hidden information prediction, significantly improving the accuracy of residual tile and fan-type predictions. The decision-making system built upon MJ-Net enhances winning rates and strengthens defensive capabilities in Mahjong games, achieving a balance between defense and offense.
