Region-based Q-learning using convex clustering approach

Abstract

For continuous state space applications, a novel method of Q-learning is proposed, where the method incorporates a region-based reward assignment being used to solve a structural credit assignment problem and a convex clustering approach to find a region with the same reward attribution property. Our learning method can estimate a current Q-value of an arbitrarily given state by using effect functions, and has the ability to learn its actions similar to that of Q-learning. Thus, our method enables robots to move smoothly in a real environment. To show the validity of our method, the proposed Q-learning method is compared with conventional Q-learning method through a simple two dimensional free space navigation problem, and visual tracking simulation results involving a 2-DOF SCARA robot are also presented.