A classifier system for reinforcement learning control of autonomous robots

The paper deals with reinforcement learning control of autonomous mobile robots. Based on the sensory state space analysis, we define a learning strategy and propose a classifier system operating with action rules. The system works with continuous state and action space representation of the action rules. The system consists of action selection, rules reproduction, and credit assignment mechanisms. The action signal is generated randomly. The control is specified at the joint level. For a two-legged mobile robot we have four local controllers. They work autonomously, and with respect to each other they represent a dynamic environment. Feasibility of the proposed control scheme is tested under experiment. It is shown that periodic gaits patterns of global behavior can emerge from the local controllers.