Improving crossover of neural networks in evolution through speciation

Abstract

Crossover is an important genetic operator that re-combines beneficial genes together and rapidly traverses the fitness landscape. Unfortunately, neuro-evolution (NE) has not experienced the benefits of crossover. Indeed, observations have shown that crossover has been detrimental to NE approaches. Tangentially, speciation has become an important feature in NE for diversity maintenance; however, such speciation research has focused on what measure is driving speciation versus how the measure determines species. This research posits that appropriate speciation implementations enable effective crossover by determining an individual’s potential mating partners. Prior speciation research demonstrated the impact of restricting the mating pools of genomes on search performance. This paper investigates these concepts in the context of NE and results demonstrate; (1) the impact of speciation implementation in NE, (2) crossover’s negative effect on search in NE, and (3) a novel speciation approach that enables effective crossover in NE.