Improving Crossover in Genetic Programming for Image Recognition

Crossover operator is the predominant operator in most of Genetic Programming (GP) system. The empirical evidence shows that along with building blocks are constructed bigger and bigger as GP evolution proceeds, the crossover operator tends to disrupt those building blocks rather than preserve them. The traditional GP crossover primarily acts as macromutation. Looseness is used for representing how ”sticky” two nodes at the two ends of the link should be. We use looseness to control the selection of crossover points in our extended Looseness Controlled Crossover (hereafter referred as eLCC) approach. The eLCC approach somehow eliminates the disruptive effect of the crossover operator explicitly. On the other hand, brood recombination crossover is the analogy of biological crossover, which mainly focus on modelling animal species breeding and preserves the homology in an implicit way. There exists a brood-diversity point for brood recombination crossover. The brood-diversity point stops further performance improvement while increasing brood size.