FCM-type co-clustering transfer reinforcement learning for non-Markov processes

Abstract

In applying reinforcement learning to continuous space problems, discretization or redefinition of the learning space can be a promising approach. Several methods and algorithms have been introduced to learning agents to respond to this problem. In our previous study, we introduced an FCCM clustering technique into Q-learning (called QLFCCM) and its transfer learning in the Markov process. Since we could not respond to complicated environments like a non-Markov process, in this study, we propose a method in which an agent updates his Q-table by changing the trade-off ratio, Q-learning and QL-FCCM, based on the damping ratio. We conducted numerical experiments of the single pendulum standing problem and our model resulted in a smooth learning process.