An Adaptive Strategy Model for Opponent's Characteristics based on Reinforcement Learning

摘要

In order to create a robot brain having intelligent action strategies, we proposed a model for making strategy for winning a game. During a game, It can make several strategies, and adaptively select/switch them to opponent feature change. For strategy making algorithm, Q-PSP reinforced learning are used because of faster learning speed. Selection and switching of the formed strategies are done based on the similarity between two kinds of Q-functions: (1) Q{sub}x is obtained at each strategy learning, and (2) Q{sub}m is used to recognize features of an opponent. We made a simulation program for an air hockey game based on the proposed strategy model. As the results of simulation, we confirmed the operations of strategy making and selection/switching, and evaluate the effectiveness of the proposed model.