Generally speaking, to make agents which play board games such as Chess, Syogi and Othello is associated with difficulty in high order of dimensions, e.g., the number of states to describe the face of a board. The difficulty is well known as "Curse of dimensionality", and it becomes a barrier to create game agents. To overcome this difficulty, we propose to reduce the dimensions in a board game by Self-Organizing Maps(SOMs). SOMs have an ability that maps high dimension vectors to low one with preserving the topology among high one. That ability makes it possible to reduce the dimensions in a board game drastically. In this paper, we prefer "Othello" game because it has simple rule and the number of huge states, and we adopt lazy SOM(LSOM) which extends the conventional SOMs. The results show that our proposal succeeded in reduce the dimensions in a board game because it recalled about average 75% for matching of the face of a board. Moreover, proposal method which combined with LSOM improved the rate of discovering for common moves about 4% than with conventional SOM.%ボードゲームなどを行うエージェントを考える際,その学習時にボードゲームが持つ状態数や情報の次元の大きさが問題となることがある.そういった情報の次元に関する問題は,例えば強化学習では,次元の呪いと呼ばれて問題視されるほど解決するべき重要な課題である.そこで,我々はボードゲームにおける盤面情報の低次元化に自己組織化マップ(SOM)を用いることを提案する.SOMを用いて腰上のマス1つ1つを覚えることなく,1つの盤面情報として記憶するこノとが可能になれば,盤面情報を表現する際もSOM内のニューロンの座標を指定するのみで盤面情報を表現することができ,大幅な情報次元の削減が見込まれる.本研究では,実際にオセロを例に挙げ盤面情報の低次元化を行った.さらに,SOM よりも入力データの分布を正確に反映しやすいとされるLazy Self-Organizing Map を用いることで,より正確た盤面情報の記憶や復元を行えるようにした.
展开▼