Comparison between Genetic Programming and full model selection on classification problems

Genetic Programming (GP) has been shown to be a competitive classification technique. GP is generally enhanced with a novel crossover, mutation, or selection mechanism, in order to compare the performance of this improvement with the performance of a standard GP. Although these comparisons show the capabilities of GP, it also makes harder, for a new comer, to figure out whether a traditional GP would have a competitive classification performance, when compared to state-of-the-art techniques. In this work, we try to fill this gap by comparing a standard GP, a GP with minor modifications and a ensemble of GP with two competitive techniques, namely support vector machines and a procedure that performs full model selection (Particle Swarm Model Selection). The results show that GP has better performance on problems with high dimensionality and large training sets and it is competitive on the rest of the problems tested. The former result is interesting because while Particle Swarm Model Selection is tailored to perform a data preprocessing and feature selection, GP is automatically performing these tasks and producing better classifiers.