NEURAL CLASSIFICATION TECHNIQUE FOR BACKGROUND REJECTION IN HIGH-ENERGY PHYSICS EXPERIMENTS

A comparative study of Bayesian and neural classification was done. Th e mathematical models of neural networks, trained in an evolutionary w ay, and Bayesian decision rules with Parzen-window multivariate densit y estimation were applied for background rejection in gamma-ray astron omy experiments. A weight function was introduced in classification sc ore to control the relative learning 'quality' of alternative classes. The use of a new quality function, instead of classification score, a llows: to avoid usage of Monte Carlo events with inherent misleading s implifications and incorrectness; to directly optimize the desired qua ntity: the significance of source detection; to obtain the complicated nonlinear boundaries of gamma-cluster. The proposed technique can be used for background rejection in the constructing experiments of high- energy neutrino point sources identification.