Genetic programming: parametric analysis of structure altering mutation techniques

We hypothesize that the relationship between parameter settings, specifically parameters controlling mutation, and performance is non-linear in genetic programs. Genetic programming environments have few means for a priori determination of appropriate parameters values. The hypothesized nonlinear behavior of genetic programming creates difficulty in selecting parameter values for many problems. In this paper we study three structure altering mutation techniques using parametric analysis on a problem with scalable complexity. We find through parameter analysis that two of the three mutation types tested exhibit nonlinear behavior. Higher mutation rates cause a larger degree of nonlinear behavior as measured by fitness and computational effort. Characterization of the mutation techniques using parametric analysis confirms the nonlinear behavior. In addition, we propose an extension to the existing parameter setting taxonomy to include commonly used structure altering mutation attributes. Finally we show that the proportion of mutations applied to internal nodes, instead of leaf nodes, has a significant effect on performance.