A Novel Nonlinear Discriminant Classifier Trained by an Evolutionary Algorithm

Abstract

In recent work, we developed a classifier inspired by reliability engineering, in which the boundaries between classes was constrained based on parameterized formulae involving single feature probabilities. This produced competitive results on the 'Iris Flower' dataset, however it was not suitable for learning tasks with highly nonlinear class boundaries (where class membership correlates very poorly with any individual feature value). The 'Balance Scale' dataset is an example of the latter type of dataset, where individual features can only influence collectively on the decision of class membership. Keeping this in mind, in this paper we describe a new nonlinear discriminant classifier, in which an evolutionary algorithm learns a probabilistic model based on a constrained, parameterized combination of all feature values. We test this method on both the 'Iris Flower' and 'Balance Scale' datasets. The results are highly competitive.