Speeding Up Online Evolution of Robotic Controllers with Macro-neurons

In this paper, we introduce a novel approach to the online evolution of robotic controllers. We propose accelerating and scaling online evolution to more complex tasks by giving the evolutionary process direct access to behavioural building blocks prespecified in the neural architecture as macro-neurons. During task execution, both the structure and the parameters of macro-neurons and of the entire neural network are under evolutionary control. We perform a series of simulation-based experiments in which an e-puck-like robot must learn to solve a deceptive and dynamic phototaxis task with three light sources. We show that: (i) evolution is able to progressively complexify controllers by using the behavioural building blocks as a substrate, (ii) macro-neurons, either evolved or preprogrammed, enable a significant reduction in the adaptation time and the synthesis of high performing solutions, and (iii) evolution is able to inhibit the execution of detrimental task-unrelated behaviours and adapt non-optimised macro-neurons.