On the exploitation of neuroevolutionary information

The final outcome of neuroevolutionary processes commonly is the best structure found during the search, and a good amount of residual information from which valuable knowledge that can be extracted is usually omitted. We propose an approach that extracts this information from neuroevolutionary runs, and use it to build a Bayesian network-based metamodel that could positively impact future neural architecture searches. The metamodel is learned from the best found solutions in previous GAN structural searches and it is used to improve subsequent neuroevolutionary searches.