Evolutionary strategies for novelty-based online neuroevolution in swarm robotics

Neuroevolution in robot controllers through objective-based genetic and evolutionary algorithms is a well-known methodology for studying the dynamics of evolution in swarms of simple robots. A robot within a swarm is able to evolve the simple neural network embedded as its controller by also taking into account how other robots are performing the task at hand. In online scenarios, this is obtained through inter-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 the soundness of this approach, we aim to extend this methodology using a novelty-based metric, so to be able to analyze different genome exchange strategies within a simulated swarm of robots in deceptive tasks or scenarios in which it is difficult to model a proper objective function to drive evolution. In particular, we want to study how different information sharing approaches affect the evolution. To do so we developed and tested three different ways to exchange genomes and information between robots using novelty driven evolution and we compared them using a recent variation of the mEDEA (minimal Environment-driven Distributed Evolutionary Algorithm) algorithm. As the deceptiveness and the complexity of the task increases, our proposed novelty-driven strategies display better performance in foraging scenarios.