Evolving the Morphology of a Neural Network for Controlling a Foveating Retina — and its Test on a Real Robot

The standard approach in evolutionary robotics is to evolve neural networks for control by encoding the parameters of the network in the genome. By contrast, we have evolved a neural controller based on biological principles from molecular and developmental biology. The key principles employed in our algorithms model the specific ligand-receptor interactions and gene regulation. These mechanisms were used to control the growth of the axons, the generation of synapses including the synaptic efficiencies (i.e. the synaptic weights in a neural network model). The evolved neural network was then transferred to a real robotic system with results comparable to the ones achieved the simulation. We hypothesize that the incorporation of mechanisms of gene regulation potentially leads to more adaptive neural networks, that can help bridging the “reality gap” between simulation and the real world.