Pattern Discrimination Based on an Exponential Error Criterion.

Introduces a criterion for the classifier portion of a pattern recognition system, based on an exponential weighting of discriminant function values. This exponential error criterion combines the ability to separate separable training patterns with the extrapolative performance of classifiers based on pattern statistics. A gradient-descent method is presented for finding discriminants that satisfy the criterion, with conditions for assuring its convergence. It is shown that the exponential-optimal discriminant of unrestricted form yields a Bayes-optimal decision rule whenever the probability densities describing the classes are continuously differentiable. The exponential-optimal linear discriminant for Gaussian pattern classes is found to be equivalent in its clustering properties to the Fisher discriminant. Empirical investigations of the performance of the exponential error criterion on selected separable and nonseparable pattern sets and on patterns derived from Gaussian distributions confirm the theoretical results. Examples of its application to the classification of electrocardiograms are presented. (A Stanford doctoral dissertation in engineering, with support from the National Science Foundation and NASA.) 98 pp. Ref.