Analysis of schema frequencies in genetic programming

Abstract

Genetic Programming (GP) schemas are structural templates equivalent to hyperplanes in the search space. Schema theories provide information about the properties of subsets of the population and the behavior of genetic operators. In this paper we propose a practical methodology to identify relevant schemas and measure their frequency in the population. We demonstrate our approach on an artificial symbolic regression benchmark where the parts of the formula are already known. Experimental results reveal how solutions are assembled within GP and explain diversity loss in GP populations through the proliferation of repeated patterns.