Effect of Encoding on the Evolvability of an Embodied Neural Network

The evolvability of a neural network controller for a hexapod agent encoded directly and symmetrically is examined. The symmetric encoding imposes a structural regularity on the neural network and decreases the size of genotype space relative to the direct encoding. The symmetrically encoded neural networks are found to be more evolvable than the directly encoded neural networks, but it is unknown whether structural regularity or decreased size of the genotype is more important. To test whether structural regularity is more important than genotype size, the architecture of the neural network is manipulated to increase the genotype size of the symmetric encodings so that they are larger than the directly encoded genotypes. These symmetric encodings are still found to be more evolvable than the direct encodings despite having a larger genotype. In these experiments it is the encoding which determines evolvability more than size of genotype space.