Evolution of developmental ontogeny for robustly reproducible phenotypes

Development has been used by a number of researchers as an efficient means of nonlinearly decoding genetic information is evolutionary systems. We show that developmental routines which do not utilise cell-cell interactions result in poor performance under noisy conditions. Addition of interactive rules permits self-organisation during development and produces robust mappings from genotype to phenotype even under noisy conditions. As a case study, we present the evolution of an edge-detecting artificial retina. The model is capable of creating three dimensional, multi-layer neural networks by modelling the development of neuron-to-neuron connectivity. Incorporating interactive overgrowth and pruning is shown to overcome the poor performance of intrinsic-only growth under noisy conditions. Staged evolution (speciation) of these processes is propose and demonstrated as an effective means of evolving such complex developmental programmes.