Dynamics of neuronal models in online neuroevolution of robotic controllers

Abstract

In this paper, we investigate the dynamics of different neuronal models on online neuroevolution of robotic controllers in multirobot systems. We compare the performance and robustness of neural network-based controllers using summing neurons, multiplicative neurons, and a combination of the two. We perform a series of simulation-based experiments in which a group of e-puck-like robots must perform an integrated navigation and obstacle avoidance task in environments of different complexity. We show that: (i) multiplicative controllers and hybrid controllers maintain stable performance levels across tasks of different complexity, (ii) summing controllers evolve diverse behaviours that vary qualitatively during task execution, and (iii) multiplicative controllers lead to less diverse and more static behaviours that are maintained despite environmental changes. Complementary, hybrid controllers exhibit both behavioural characteristics, and display superior generalisation capabilities in simple and complex tasks. © 2013 Springer-Verlag.