Method of Constructing Autonomous Decentralized Systems Using Classifier Systems. Application to Gait Acquisition of 6-Legged Robot.

It is difficult to control large complex systems by centralized control. Therefore, autonomous decentralized systems which have variety, flexibility and fault tolerability, have been studied recently. In this paper, we propose an autonomous decentralized system comprised of classifier systems. Classifier systems are parallel, message-passing, rule-based systems that learn through credit assignment and rule discovery (the genetic algorithm). In our method, each subsystem is controlled by an individual classifier system, and it generates production rules autonomously by exchanging information with adjacent subsystems. We test the system on a 6-legged robot, in computer simulations, which learns how to coordinate its legs so as to walk forward, as a practical example.