1A1-X07 Reinforcement Learning with multiplex learning spaces : Consideration of the learning inefficiency with adding unavailable learning spaces(Evolution and Learning for Robotics)

Abstract

We have confirmed that a reinforcement learning becomes more efficient with multiplex learning spaces, a whole and one or two of partial learning spaces. In this paper, we extend the number of partial spaces to N, and study about the learning efficiency. We investigate the learning inefficiency in a case of increasing the number of unavailable partial learning spaces in experimental simulations. We confirmed that increasing did not influence the learning in an inefficient way, and that only one available partial learning space influenced the learning in an efficient way. As a result, we found that more multiplexing becomes more efficient for the learning, even if the partial learning spaces include some unavailable ones.