A strategy for improving performance of Q-learning with prediction information

Abstract

Nowadays, learning of agents gets more and more useful in game environments. It takes a long learning time, however, to produce satisfactory results in games. Thus, we need a good method of shortening the learning time. In this paper, we present a strategy for improving learning performance in Q-learning with predictive information. This refers to the chosen action at each status in the Q-learning algorithm. It stores the referred value in the P-table of the prediction module, and then searches some high-frequency values in the table. The values are used to renew the second-compensation value from the Q-table. Our experiments show that our approach yields an efficiency improvement of 9% on the average after the middle point of the learning experiments, and that the more actions are executed in a status space, the higher the performance would be.