A flexible Q-relearning method to accelerate learning under the change of environments by reusing a portion of useful policies

We propose a relearning method for Q-learning algorithm, called “Q-relearning,” to accelerate learning under the change of environments by reusing a portion of useful policies. To deal with the problem that Q-learning algorithm can´t adapt to the change of environments efficiently, we developed an original Q-relearning method in the past study. However, there was little flexibility in choosing the portion to be reused. In this paper, we revise the Q-relearning method to permit any choice of the portion by adding the possibility to update the fixed Q-values under particular conditions. We show the effectiveness of the improved Q-relearning method by numerical experiments.