A NEURAL-NETWORK APPROACH TO NONPARAMETRIC AND ROBUST CLASSIFICATION PROCEDURES

In this paper algorithms of neural-network type are introduced for sol ving estimation and classification problems when assumptions about ind ependence, gaussianity, and stationarity of the observation samples ar e no longer valid, Specifically, the asymptotic normality of several n onparametric classification tests is demonstrated and their implementa tion using a neural-network approach is presented, Initially, the neur al nets train themselves via learning samples for nominal noise and al ternative hypotheses distributions resulting in nearoptimum performanc e in a particular stochastic environment, In other than the nominal en vironments, however, high efficiency is maintained by adapting the opt imum nonlinearities to changing conditions during operation via parall el networks, without disturbing the classification process, Furthermor e, the superiority in performance of the proposed networks over more t raditional neural nets is demonstrated in an application involving pat tern recognition.