A Supervised Attention-Based Multiclass Classifier for Tile Discarding in Japanese Mahjong

Abstract

Japanese Mahjong, an imperfect-information, multiplayer, multi-round game, has become an area of interest for the AI community due to its immense imperfect-information space and complex playing and scoring rules. In 2020 Microsoft unveiled the Mahjong AI Suphx that managed to outdo most of the top human players in Tenhou.net, the most popular platform for Japanese Mahjong. With supervised learning, Suphx’s discard model reached a prediction accuracy of 76.7% when tested on game logs from Tenhou.net. Two recurring problems with state-of-the-art Mahjong AIs, including Suphx, are their heightened architecture complexity and the sizeable data structure. These problems make it less feasible to replicate these experiments with limited hardware. We propose two models: MHA-B and MHA-S. Both use the same architecture with a multi-head attention layer but are trained on different training sets. Furthermore, we suggest a data structure that is a fraction of the size of contemporary alternatives. A smaller data structure implies fewer data values for the models to focus on making the models converge faster. When tested on game logs from Tenhou.net, MHA-B and MHA-S reach a prediction accuracy of 66.7% and 65.2%, respectively. Although somewhat subpar compared to state-of-the-art models’ results, our approach yields notable results considering the non-complex model architecture and the restricted data structure.