THE COMPARISON OF BAYESIAN AND NEURAL TECHNIQUES IN PROBLEMS OF CLASSIFICATION TO MULTIPLE CATEGORIES

The determination of elemental composition of Primary Cosmic Rays in t he energy range 10(15)-10(17) eV is still an unsolved problem. Modern surface installation registering many characteristics of Extensive Air Shower (EAS) initiated in atmosphere by incident particles provide th e possibility to handle data on an event-by-event basis and obtain res ults with reliability comparable with collider experiments. We use Bay esian decision making and neural approaches for data classification in to multiple categories. The Parzen window method was used for multivar iate density estimation along with evolutionary algorithms for net tra ining. The accuracies of reconstructed proportion of different nucleus in primary flux were estimated. Both methods provide close results pr oving convergence to minimal achievable Bayesian risk.