Self-Scaling Reinforcement Learning Algorithm for Generating Fuzzy Controller.

Abstract

In this paper, we propose a new reinforcement learning algorithm to generate a fuzzy controller for robot motions. This algorithm generates a range of continuous real-valued actions, and the reinforcement signal is self-scaled. This prevents the weights from overshooting when the system receives very large reinforcement values. Therefore this algorithm can obtain a solution in less iteration times. The proposed method is applied to the control of the brachiation robot, which moves dynamically from branch to branch like a gibbon swinging its body in a pendulum fashion. Through computer simulations, we show the fast convergence and the robustness against disturbances.