TUBSTAPにおけるグラフ表現と画像表現の併用

ターン制戦略ゲームは人気のゲームジャンルであるが，未だ人間よりも強い AI は実現されていない．これは合法手の多さや地形の要素など，幾つかの扱いにくい性質があるためである．本稿ではターン制戦略ゲームの研究用プラットフォームである TUBSTAP を題材として，地形の要素を考慮できるグラフ表現を提案し，地形を含むマップで盤面評価を行う Graph Convolutional Networks(GCN) の学習を行った．さらに画像表現とグラフ表現を併用した統合モデルを複数提案し学習を行った．その結果，提案したグラフ表現を用いた GCN は，高精度な盤面評価が可能であることを確認した．提案した統合モデルはいずれも過学習の傾向があり GCN の性能に劣る結果となったが，複数の統合方法で過学習の大きさの差異を示すことができた．

While turn-based strategy (TBS) games are a popular game genre, AI that is stronger than humans has yet to be realized. This is due to some unwieldy characteristics of TBS games, such as the large number of legal moves and terrain elements. In this paper, using TUBSTAP, a research platform for TBS games, we propose a graph representation that can account for terrain elements and train Graph Convolutional Networks (GCNs) to evaluate the board on a map including terrain. In addition, we propose several integration models that combine image and graph representations and train them. As a result, we confirmed that the proposed GCNs with graph representations are capable of highly accurate board evaluation. Although all the proposed integration models showed a tendency to overfit and were inferior to the performance of GCN, we show differences in the magnitude of overfitting among the different integration methods.