Behavior Acquisition of an Autonomous Robot by Reinforcement Learning Based on Globally Coupled Chaotic System : 1st Report, Empirical Examination of Credit Assignment and Generalization

Abstract

We describes a framework for the acquisition of perception-based navigational behaviors in autonomous mobile robots. Globally coupled chaotic system is applied to learn reactive action-rules. The control architecture of a robot consists of chaotic elements each coding an action-rule. The elements have the dynamics, which are designed so that the elements can collectively execute reinforcement learning. We carried out simulations on a navigation task in a static environment. Then, we observed the reproduction process and utility transition of the action-rules to examine how the robot acquire behaviors. Simulation results demonstrate that the robot successfully acquired behaviors such as goal-reaching, wall-following, and collision avoidance without any prior knowledge of the task space. The observation indicates that the stability of the acquired behavior depends on the types of the credit assignment and that the progress of the generalization of the state space correlates to the learning performance in a trial.