A Completely Evolvable Genotype-Phenotype Mapping for Evolutionary Robotics

Abstract

To achieve a desired global behavior for a swarm of robots where each robot has a local view and operating range in the environment is a well-known and challenging problem. Evolutionary Robotics is a self-adaptation approach which has been shown to e effectively find robot controllers for behaviors which are hard to implement by hand. There, evolvability is highly dependent on controller representation during evolution. It is known that using a genotypic controller representation which also encodes parts of the genotype-phenotype mapping (GPM) can lead to a meta-adaptation of the evolutionary operators to the search space structure, thus improving evolvability. We enhance this idea using a fully flexible GPM which is represented in the same way as the behavioral controllers are, and, therefore, can be completely evolved along with the behavior. The approach is based on finite state machines and extends an existing framework for decentralized evolution of robot behavior in swarms of mobile robots. Experiments indicate that the evolvable GPM outperforms both the extensively improved operators of the existing framework and a standard operator for the new real-valued genotypes with fixed GPM.