Abstract
In reinforcement learning, construction method of state-space influenced the capacity of learning. Fuzzy-ART is an incremental state-space construction method. It classifies an encountered state into an existed category. If the state does not belong to any category it makes a new category. The number of categories increases monotonically as more tasks are learned. It was necessary to prevent the redundant increase of the categories when many tasks had to be learned. We proposed a state-space construction method based on state value. It divides the important area of state-space into a lot of categories, and casts away insignificant categories according to state value. We classified that proposed method was effective to improvement of reinforcement learning in case of path planning for multi-link robots.
