Fuzzy Q-Learning with the Modified ART Neural Network
H. Oeda, Naoki Hanada, Hideaki Kimoto, Takeshi Naraki, KEISUKE G. TAKAHASHI, Tetsuhiro Miyahara · IEEE/WIC/ACM International Conference on Intelligent Agent Technology · 2006
We present a method to acquire rules for agent's behavior, where continuous numeric percepts are classified into categories by fuzzy ART and fuzzy Q-learning is employed to acquire rules. To make fuzzy ART be suitable to fuzzy Q-learning, we modify fuzzy ART such that it selects some categories for a percept vector and returns them with their fitness values. For efficient learning, we also present a method that integrates two categories into one, where we define the similarity for any category pair and it is utilized for integration. Moreover, a vigilance parameter is defined for each category in order to control the size of a category, while ordinary fuzzy ART uses a common vigilance parameter for all categories. The methods shown here have been implemented and some experiments have been done.