Effective Learning of Hyper-Redundant Systems by Dividing State Space

Abstract

QDSEGA is one of efficient learning algorithm for hyper-redundant systems. However, to accomplish the learning process, it required many iterations of trials. One of the most significant cause of the problem is lack of efficiency of learning, which is observed at the beginning of the learning.In this paper, we consider the mechanism of the lack of efficiency and solve the problem, we propose a method that the state space is devided into several classes based on physical dynamics. We improve QDSEGA by embedding the proposed dividing method into the QDSEGA, and demonstrate the effectiveness of the improved QDSEGA by applying it to a locomotion task of real snake robot.