Neural Discriminant Models, Bootstrapping, and Simulation

This paper considers the feed-forward neural network models for dataof mutually exclusive groups and a set of predictor variables. We takeinto account the bootstrapping based on information criterion when selectingthe optimum number of hidden units for a neural network modeland the deviance in order to summarize the measure of goodness-of-fit onfitted neural network models. The bootstrapping is also adapted in orderto provide estimates of the bias of the excess error in a prediction ruleconstructed with training samples. Simulated data from known (true)models are analyzed in order to interpret the results using the neuralnetwork. In addition, the thyroid disease database, which compares estimatedmeasures of predictive performance, is examined in both a puretraining sample study and in a test sample study, in which the realizedtest sample apparent error rates associated with a constructed predictionrule are reported. Apartment house data of the metropolitan area stationwith four-class classification are also analyzed in order to assess the bootstrappingby comparing leaving-one-out cross-validation (CV).