A.N. Nyrtsov, I.V. Chebotar, S.A. Yakovitsky, S.I. Andronov

The article is about the model of a neural classifier constructed on the basis of networks of the adaptive-resonance theory. This classifier is intended for splitting a mixed pulse stream, consisting of signals and simulating hindrances. To check this model at work simulated modeling of the process of splitting with the use of the mathematical apparatus of the adaptive- resonant theory has been carried out. Simulated modeling has allowed to find out a number of restrictions of the use of the ART networks for splitting similar signals and simulating hindrances. The main of them is the problem of uncertainty of decision-making as to the object identification when classes of signs intersect. This is typical of the signals and response simulating hindrances with identical characteristics. For the reduction of uncertainty of object identification it is recommended to use the bell- shaped function a standard function of a choice under Weber law. Instead of has an extremum within space of signs. The use of this proposal makes it possible to form it more exact border of decision-making in the range of the intersection of classes of signs. As a result it reduces the approximately 2 times probability of the wrong identification. The developed model of a neural classifier allows expanding possibilities of sophisticated devices of radiomonitoring by processing simulating hindrances.