Evolutionary Strategies for Novelty-Based Online Neuroevolution in Swarm Robotics

Neuroevolution in robot controllers throughobjective-based genetic and evolutionary algorithms is a wellknownmethodology for studying the dynamics of evolution inswarms of simple robots. A robot within a swarm is able toevolve the simple neural network embedded as its controllerby also taking into account how other robots are performingthe task at hand. In online scenarios, this is obtained throughinter-robot communications of the best performing genomes (i.e.representation of the weights of their embedded neural network).While many experiments from previous work have shown thesoundness of this approach, we aim to extend this methodologyusing a novelty-based metric, so to be able to analyze differentgenome exchange strategies within a simulated swarm of robotsin deceptive tasks or scenarios in which it is difficult to modela proper objective function to drive evolution. In particular, wewant to study how different information sharing approachesaffect the evolution. To do so we developed and tested threedifferent ways to exchange genomes and information betweenrobots using novelty driven evolution and we compared themusing a recent variation of the mEDEA (minimal EnvironmentdrivenDistributed Evolutionary Algorithm) algorithm. As thedeceptiveness and the complexity of the task increases, ourproposed novelty-driven strategies display better performancein foraging scenarios.