Robinson, Andrew Locke. “Evolving Neural Networks Through Random Augmentation and Sexual Reproduction.” PhD diss., University of Akron, 2020.
URL1 URL2
Neuroevolution methods which evolve the topology of a neural network as well as their weights are poised to be powerful tools in machine learning as they can potentially adapt to any degree of complexity. The neuroevolution method presented in this paper, Random Augmentation and Sexual Reproduction (RASR), outperforms all existing fixed-topology and augmenting topology neuroevolution methods on a difficult reinforcement learning benchmark. RASR achieves this result by implementing a novel reproduction technique which performs extensive crossover and allows the combination of networks of varying topologies. RASR also implements a novel augmentation process which promotes a stable complexification through generations without the need for speciation. Further, all the resultant networks produced by RASR started from the most minimal structure, a network without any connections and only input and output nodes. This shows that RASR can create complex solutions not only quickly but also without the bias of starting from a promising architecture.