Crossover in grammatical evolution: A smooth operator?

Abstract

Grammatical Evolution is an evolutionary algorithm which can produce code in any language, requiring as inputs a BNF grammar definition describing the output language, and the fitness function. The usefulness of crossover in GP systems has been hotly debated for some time, and this debate has also arisen with respect to Grammatical Evolution. This paper serves to analyse the crossover operator in our algorithm by comparing the performance of a variety of crossover operators. Results show that the standard one point crossover employed by Grammatical Evolution is not as destructive as it might originally appear, and is useful in performing a global search over the course of entire runs. This is attributed to the fact that prior to the crossover event the parent chromosomes undergo alignment which facilitates the swapping of blocks which are more likely to be in context.