Generative sequence modeling for action prediction in Chinese standard Mahjong

This study proposes a generative sequence modeling approach for the action prediction task in Chinese Standard Mahjong games. Inspired by the similarity in sequential structure between natural language and Mahjong games, each game is converted into a discrete symbol sequence consisting of tile faces and actions. The Transformer decoder architecture is employed for autoregressive modeling to capture global causal relationships and long-term dependencies. Through a "pretraining-finetuning" transfer learning paradigm, a general game language representation is first learned from large-scale game data, followed by task-specific adaptation and optimization for actions such as discard, chow, pong, kong, and winning. Experimental results show that the proposed method achieves an average accuracy of 79.14% across multiple action prediction tasks in Chinese Standard Mahjong, representing a 5.74 percentage point improvement over the ResNet-50 model, which performed well in Japanese Mahjong. Ablation experiments indicate that autoregressive pretraining enhances the average accuracy by 8.10 percentage points, validating the effectiveness of global modeling and transfer learning for Mahjong action prediction. This work not only provides a novel modeling approach for Chinese Standard Mahjong games but also demonstrates the feasibility and advantages of transferring large model training paradigms from natural language processing to complex game domains.