Influence of Mating Mechanisms in Distributed Evolution for Collective Robotics

International audienceDistributed Embodied Evolution [1] is a compelling family of approaches to learning swarm robot behavior online that exploits the intrinsic parallelism of robot swarms: each robot in the swarm runs a separate instance of an evolutionary algorithm onboard, including an internal population, and robots locally exchange genomes when meeting, which is know as the mating operator or migration policy. Since the internal populations in different robots do not contain the same individuals, these approaches have been shown to naturally maintain diversity [2], which can improve search, especially in deceptive problems. Here, using a vanilla dEE algorithm (mEDEA with task-driven selection pressure), we evaluate the influence on the fitness of the evolved behaviors of 12 different mating operators, in terms of distance (short, medium, large, and full distance-based broadcast), and in terms of frequency (constantly broadcast every timestep, every 10 timesteps, or at a rate proportionate to the genome's rank in the local population)