Introducing Novelty Search in Evolutionary Swarm Robotics

Abstract. Novelty search is a recent and promising evolutionary tech-nique. The main idea behind it is to reward novel solutions instead of progress towards a fixed goal, in order to avoid premature convergence and deception. In this paper, we use novelty search together with NEAT, to evolve neuro-controllers for a swarm of simulated robots that should perform an aggregation task. In the past, novelty search has been ap-plied to single robot systems. We demonstrate that novelty search can be applied successfully to multirobot systems, and we discuss the chal-lenges introduced when moving from a single robot setup to a multirobot setup. Our results show that novelty search can outperform the fitness-based evolution in swarm robotic systems, finding (i) a more diverse set of successful solutions to an aggregation task, (ii) solutions with higher fitness scores earlier in the evolutionary runs, and (iii) simpler solutions in terms of the topological complexity of the evolved neural networks.