Q-learning based on hierarchical evolutionary mechanism

Abstract

Reinforcement learning is applied to various fields such as robotics and mechatronics control. The reinforcement learning is an efficient method to control in unknown environment. This paper discusses a new reinforcement learning algorithm which is based on Genetic Algorithm and has a hierarchical evolutionary mechanism. The proposed learning algorithm introduces new adaptive action value tables and it enables sharing knowledge among agents effectively. Regarding sharing knowledge among agents, the knowledge is inherited not only across the generations, but also in one generation. As a result, the proposed algorithm achieves effective learning, and realizes robustness learning. Computational simulations using the pong simulator which executes table tennis prove the effectiveness of the proposed algorithm.