Validation sets for evolutionary curtailment with improved generalisation

Abstract

This paper investigates the leveraging of a validation data set with Genetic Programming (GP) to counteract over-fitting. It considers fitness on both training and validation fitness, combined with with an early stopping mechanism to improve generalisation while significantly reducing run times. The method is tested on six benchmark binary classification data sets. Results of this preliminary investigation suggest that the strategy can deliver equivalent or improved results on test data. © 2011 Springer-Verlag.