Genetic Programming (GP) is an evolutionary method for generating tree structural programs. Normal subtree crossover in GP randomly selects a crossover point in each parent tree, and offspring are created by exchanging the selected subtrees. In the subtree crossover, it is difficult to control the global and local search because the similarity between the subtrees is not considered. In this paper, we propose a new crossover operation based on the semantic distance between the subtrees. By using the semantic crossover, the global search can be performed in the early stage of search, and the search property can be shifted to the local search as the search proceeds. As the results of experiments, the semantic crossover showed better performance than the conventional crossover.

2011 IEEE SMC Hiroshima Chapter若手研究会, 2011年7月9日, 広島市立大学