This paper proposes using a graph-based data representation to more efficiently learn to play chess. Usually, games work by interactions between components, and having those interactions be the way a game position is represented has the possibility to increase the efficiency of any related model. We show the result of experiments replacing the traditional CNN layers in AlphaZero by GNN layers, significantly improving performance in our limited setting, and introduce a new GNN layer, based on Graph Attention Networks (GAT), in order to handle edge features in addition to node features.
| 氏名 | コース | 研究室 | 役職/学年 |
|---|---|---|---|
| Rigaux Tomas | データ科学コース | Applied Machine Learning Laboratory | 博士3回生 |
| Hisashi Kashima | 知能情報学コース | Collective Intelligence Laboratory | 教授 |