Genetic Approach for Optimizing Ensembles of Classifiers

An ensemble of classifiers is a set of classifiers whose predictions are combined in some way to classify new instances. Early research has shown that, in general, an ensemble of classifiers is more accurate than any of the single classifiers in the ensemble. Usually the gains obtained by combining different classifiers are more affected by the chosen classifiers than by the used combination. It is common in the research on this topic to select by hand the right combination of classifiers and the method to combine them, but the approach presented in this work uses genetic algorithms for selecting the classifiers and the combination method to use. Our approach, GA-Ensemble, is inspired by a previous work, called GA-Stacking. GA-Stacking is a method that uses genetic algorithms to find domain-specific Stacking configurations. The main goal of this work is to improve the efficiency of GAStacking and to compare GA-Ensemble with current ensemble building techniques. Preliminary results have show that the approach finds ensembles of classifiers whose performance is as good as the best techniques, without having to set up manually the classifiers and the ensemble method.