Fast generic selection of features for neural network classifiers

Abstract

The task of classifiers is to determine the appropriate class name when presented with a sample from one of several classes. In forming the sample to present to the classifier, there may be a large number of measurements one can make. Feature selection addresses the problem of determining which of these measurements are the most useful for determining the pattern's class. In this paper, we describe experiments using a genetic algorithm for feature selection in the context of neural network classifiers, specifically, counterpropagation networks. We present two novel techniques in our application of genetic algorithms. First, we configure our genetic algorithm to use an approximate evaluation in order to reduce significantly the computation required. In particular, though our desired classifiers are counterpropagation networks, we use a nearest-neighbor classifier to evaluate feature sets. We show that the features selected by this method are effective in the context of counterpropagation networks. Second, we propose a method we call training set sampling, in which only a portion of the training set is used on any given evaluation. Again, significant computational savings can be made by using this method, i.e., evaluations can be made over an order of magnitude faster. This method selects feature sets that are as good as and occasionally better for counterpropagation than those chosen by an evaluation that uses the entire training set.