Spontaneous Evolution of Modularity in Neural Networks for Robot Locomotion

Modular neural networks have a number of advantages when used to control robots. They reduce the number of connections and thus the computational power required to update them compared to fully connected networks. Var-ious literature also suggests that modularity increases robustness and evolv-ability of neural networks. Hence, modularity has the potential to pave the way towards adaptive and resilient controllers for complex robots that pre-viously proved computationally infeasible. To explore these issues, this work shows the evolution of modularity in neural networks, modelled as dynamically autonomous random boolean net-works, that control robots to exhibit phototaxis in a simulated environment. In experiments on a wheeled robot networks that settle into fixed point at-tractors were studied, whereas in experiments on a legged robot networks that settle into cyclic attractors were considered. In further experiments it was investigated how the results scale to larger networks and robots were excluded: desired network dynamics were selected for instead of selecting based on their influence on robot behavior. It was found that modularity emerged when selecting for cyclic attractors in one part of the network and selecting for fixed point attractors in the other part. i Declaration I hereby declare that this thesis is my own work and effort and that it has not been submitted, either in the same or different form, to this or any other university for a degree. Signature: i