Co-Evolving Communicating Classifier Systems for Tracking

In this paper we suggest a general approach to using the genetic algorithm (GA)[1] to evolve complex control systems. It has been shown [2] that although the GA may be used to evolve simple controllers, it is not able to cope with the evolution of controllers for more complex problems. We present an architecture of co-evolving communicating classifier systems [3] as a general solution to this, where the only restriction is that each classifier system is responsible for one simple behaviour. Thus the ecology of subproblems evolves its own organisational structure at the same time its constituents evolve their solutions. Whether this structure ends up as a democratic soup, a hierarchy, or something in between, is determined by co-evolution rather than prescribed a priori by a creator. We use the trail following “tracker task” to compare the performance of a single classifier, responsible for the control of the whole system, evolved for this task with the performance of a co-evolved controller using our approach. The resulting interactions of the classifier systems are also examined.