Uniform subtree mutation

Abstract

The traditional genetic programming crossover and mutation operators have the property that they tend to affect smaller and smaller fractions of a solution tree as the tree grows larger. It is generally thought that this property contributes to the ‘code bloat’ problem, in which evolving solution trees rapidly become unmanageably large, and researchers have investigated alternate operators designed to avoid this effect. We introduce one such operator, called uniform subtree mutation (USM), and investigate its performance—alone and in combination with traditional crossover—on six standard problems. We measure its behavior usingb oth computational effort and size effort, a variation that takes tree size into account. Our tests show that genetic programming using pure USM reduces evolved tree sizes dramatically, compared to crossover, but does impact solution quality somewhat. In some cases, however, a combination of USM and crossover yielded both smaller trees and superior performance, as measured both by size effort and traditional metrics. © Springer-Verlag Berlin Heidelberg 2002.