А.A. Volodin, E.V. Lubentcova, A.A. Evdokimov, V.F. Lubentcov

For processing of experimental data classical methods of mathematical statistics are traditionally used. For the solution of a problem of classification by means of methods of mathematical statistics build dividing surfaces between classes in priznakovy space. However authentic statistical estimation demands large volume of data, or very strong assumptions of a type of functions of distribution, and works usually only at normal or close to normal functions of distribution. Therefore at calculation of a conditional population is required to check a hypothesis about distribution of empirical data on the normal law to distribution or to use the device of the nonparametric statistics restoring estimates of density of distribution of probabilities. Usually when using empirical methods it is supposed that the user receiving will satisfy simply result of classification. Further on the basis of the professional knowledge it will draw a conclusion on applicability of the received result, having compared a machine forecast with the expert estimates. For the solution of problems of classification and a forecast various mathematical methods are used: mathematical statistics, klasterny, regression, factorial analysis etc. In comparison with these methods the approach based on application of artificial neural networks, possesses a number of advantages. First, the computing paradigm is created uniform for all tasks. Using neural networks with rather small number of neurons, it is possible to solve rather complex problems of classification and a forecast. Neural networks represent being trained models that allows simply «finish teaching» them at receipt of new data. Use of small training samples that is characteristic for microbiological systems which don\'t provide receiving statistically authentic results by classical methods is possible. In neural networks it is possible to use any quantity of independent and dependent signs. The number of examples for different classes (at the solution of a problem of classification) can be various. The decision one network at the same time several problems of classification or a forecast is also possible. In a neural network there is a procedure of calculation of the importance of independent signs and possibility of minimization of their number. The analysis of works on neuronetwork methods of processing of information in problems of recognition of images allowed to allocate for classification of processes by extent of limitation with a substratum Kokhonen\'s neural network with disorder neurons (Kokhonen\'s layer) which application is expedient for the solution of problems of classification in the conditions of small volume of a training sample. When training a neural network entrance vectors without the indication of desirable exits are shown and corrected weight according to the algorithm of classification developed in this work.