| Resumo: | Manual design of self-organized behavioral control for swarms of robots is a complex task. Neuroevolution has proved a viable alternative given its capacity to automatically synthesize controllers. In this paper, we introduce the concept of Genome Variations (GV) in the neuroevolution of behavioral control for robotic swarms. In an evolutionary setup with GV, a slight mutation is applied to the evolving neural network parameters before they are copied to the robots in a swarm. The genome variation is individual to each robot, thereby generating a slightly heterogeneous swarm. GV represents a novel approach to the evolution of robust behaviors, expected to generate more stable and robust individual controllers, and bene t swarm behaviors that can deal with small heterogeneities in the behavior of other members in the swarm. We conduct experiments using an aggregation task, and compare the evolved solutions to solutions evolved under ideal, noise-free conditions, and to solutions evolved with traditional sensor noise.
|
|---|