The effect of extensive use of the mutation operator on generalization in genetic programming using sparse data sets

Abstract

Abstract. Ordinarily, Genetic Programming uses little or no mutation. Crossover is the predominant operator. This study tests the effect of a very aggressive use of the mutation operator on the generalization performance of our Compiling Genetic Programming System ('CPGS'). We ran our tests on two benchmark classification problems on very sparse training sets. In all, we performed 240 complete runs of population 3000 for each of the problems, varying mutation rate between 5% and 80~ We found that increasing the mutation rate can significantly improve the generalization capabilities of GP. The mechanism by which mutation affects the generalization capability of GP is not entirely clear. What is clear is that changing the balance between mutation and crossover effects the course of GP training substantially --for example, increasing mutation greatly extends the number of generations for which the GP system can train before the population converges.

Genetic Programming — Introduction, Applications, Theory and Open Issues

“…In particular, in Francone et al (1996) they introduce a GP system called Compiling GP and they compared its generalization ability with that of other machine-learning paradigms. Furthermore, in Banzhaf et al (1996) they show the positive effect of an extensive use of the mutation operator on generalization in GP using sparse data sets. (1) the selection of the best-ofrun individuals using a three-data-sets methodology, and (2) the application of parsimony pressure to reduce the complexity of the solutions.…”

A Novelty Search Approach for Automatic Test Data Generation

“…These techniques use basically a fitness function to guide the search e.g., gather fittest solutions over generations. These techniques are good since they reward individuals with high score but they do not favor diversity and the search may converge to many local optima [3], [4]. The idea of NS, introduced by Lehman and Stanley in 2008 [8], represents an alternative solution for this issue.…”

“…The issue of premature convergence to local optima has been a common problem in GAs. Many methods are proposed to avoid this problem [3], [4]. However, all these alternatives use a fitness-based selection to guide the search.…”

“…Since we haven't yet the minimum coverage value (the max fitness value from GAs), we are trying with this initial implementation, to record the maximum coverage value that may be reached by applying the NS. Initial project and results can be found at the project website 3 .…”

Open issues in genetic programming

“…A survey of the main contributions on generalization in GP has been done some years ago by Kushchu in [68]. In [34] the authors use what they called the "Compiling GP System" to compare its generalization ability with that of other ML paradigms and show in [8] the positive effect of an extensive use of the mutation operator on generalization in GP with sparse data sets. In [19], Da Costa and Landry have recently proposed a new GP model called Relaxed GP, showing its generalization ability.…”