Use of Successful Policies to Relearn for Induced States of Failure in Reinforcement Learning

Abstract

In this paper, we propose a method to reduce relearning costs by ap-plying successful policies to estimated failing states in reinforcement learning. When an environment is changed to another, relearning is needed in order to acquire an appropriate policy in the new environment. In order to reduce re-learning costs, an algorithm has been proposed to estimate failing states using a decision tree C4.5, and it relearns new policies only for the estimated failing states in the new environment. We try to reduce failing states furthermore by applying successful policies to the estimated failing states. Computer simulations show that our method can reduce relearning costs and improve the successful rate in reinforcement learning.