An Adaptive Pattern Classification System

Abstract

Adaptive pattern classification is the assignment of patterns to classes based on typical patterns or training samples, used by the system to determine the decision procedure. The system is adaptive in the sense that the decision procedure is optimized according to some criterion of the system's performance on the training samples. An adaptive pattern classification system is described that does not require a priori knowledge of the probability density of the pattern vectors for each class, as do the classical statistical techniques. Any decision rule, consisting of a discriminant function, that is a linear combination of arbitrary scalar functions of the pattern vector, may be chosen on the basis of a priori knowledge about the classes, engineering judgment, and economic considerations. The system optimizes itself by adjustment of the decision parameters according to a weighted mean-square-error performance criterion, using a multivariable search technique. The proposed performance criterion is well suited for self-optimizing search procedures. It also has the property that, as the number of training samples approaches infinity, the resulting disciminant function belongs to the class of discriminant functions, chosen at the outset, which approximates the optimum Baye's discriminant function with minimum variance. Some results from simulation studies are presented which include comparison with classical statistical techniques.