E. A. Samoilin

In this paper offered common structure of control unit for machine-tool with parallel kinematics, in which all coordinate Many developed models of neural networks for solution of various applied tasks of image processing are not optimal in sense of some criterion of the quality corresponding to physical sense of the solved task. As it is known, network optimization is subdivided on primary (definition of structure of a network, number of layers, neuron, links, sort of functions of activation etc.) and secondary (creation of current signals in a network), i.e., directly procedure of training of a neural network. According to it, the purpose of offered article is improvement of quality of solution of applied tasks of image processing on the basis of primary parametrical optimization functions activation of model of a neural network. Procedure of primary parametrical optimization functions activation of model of the neural network, grounded on usage of criterion of the ideal spectator according to which the sum of absolute probabilities of erratic solutions of a network is minimized is offered. The numerical researches spent on an example of the task of selection of impulse interferences on digital maps, testify that the optimized neural network allows to approximate quality of image processing to potentially accessible (Bayesian) level. Besides, offered optimization of model of a network allows to receive in many cases higher probability of convergence and speed of its training in comparison with not optimized network. The offered procedure can be used for primary optimization of other models of the neural networks realizing various tasks of processing of signals and pattern recognition. Thus for synthesis of optimal threshold parameters of neural transfer functions, as well as in the considered case of selection of interferences, elementary probability laws of presence and absence of recognized signals in entry observation on a range of values of a scalar product of an entry vector and a vector of weight coefficients neuron a hidden layer of a network should be known.