Abstract
The adaptive resonance theory (ART) neural network is an unsupervised learning system that can generate and grow the recognition categories based on the similarity between inputs and memories. Recently, several researches about Reinforcement Learning System which used ART as a construction of state-space have reported. Inthesystem, a proper vigilance parameter that decides a category space size is required for the adaptive learning in finite time period. However, the proper value of vigilance parameter is unclear on generally unknown in each problem. Therefore, this paper proposes a new learning system using hierarchical Fuzzy ART for two player games. The proposed system segments an input state space into subspaces by the fuzzy ARTs added hierarchically in proportion as the learning progress of player, and then learns pairs of input states and actions by the reinforcement learning. As results of experiments, we show through a fighting simulation game that the player could acquire the proper pairs of the input states and actions by learning using hierarchical Fuzzy ART against the opponent player.
