Multi-objective evolution of the Pareto optimal set of neural network classifier ensembles

Existing research demonstrates that classifier ensembles can improve on the performance of the single dasiabestpsila classifier. However, for some problems, although the ensemble may obtain a lower classification error than any of the base classifiers, it may not provide the desired trade-off among the classification rates of different classes. In many applications, classes are not of equal importance, but the preferred trade-off may be hard to quantify a priori. In this paper, we adopt multi-objective techniques to create Pareto optimal sets of classifiers and ensembles, offering the user the choice of preferred trade-off. We also demonstrate that the common practice of developing a single ensemble from an arbitrary (diverse) selection of base classifiers will be inferior to a large proportion of those classifiers.