Abstract—Progress in computing technology allows using simpler, nonparametric models compared to parametric ones due to the use of a training sample not only during the training period, but also when using the trained model later. The statement is demonstrated by comparing the neural network method for solving the classification problem and the Anderson discriminant function approximation method. It is proposed to select the neural network structure taking into account the range of errors in the irreproducibility of the neural network training results.