対戦ゲームの強化学習における状態表現モデルの効果

抄録

Reinforcement learning is a promising approach for various applications such as complicated task acquisition and game AI. In reinforcement learning for games, there are necessary issues of combinatorial explosion of agent states and learning speed deterioration caused by sparse distribution of states in the state space for reinforcement learning. In this paper, we address state expression for reinforcement learning in games. According to characteristics of games, a state expression model is structured considering rotation, mirror translation, and symmetric translation of game boards. Through numerical experiments of Othello game, we found the proposed model is promising for game AI.