Intelligent algorithm for forecasting of optimum neurons quantity in perceptron with one hidden layer

The research performed focus on the development of methods of building-up of the intelligent neural network modeling solutions database as well as methods of approximation aiming at empirical knowledge conservation and representation to find the best structure of the artificial neural network (ANN). The learning sample is made up of solutions of approximation of one-dimensional functions defined in the uniform grid nodes with the help of perceptron type ANN with one hidden layer (single-layer perceptron - SLP). Computational experiment plan is made up of the points with uniform grid nodes abscissas, and the ordinates are defined by means of using of Sobol-Statnikov generator of the semi-uniform sequence of numbers. The training utilizes the stochastic approximation algorithm that is a modification of the backpropagation algorithm. As a result of SLP given points training the minimum number of neurons in the hidden layer is defined at which the target accuracy is achieved. Numerous solutions of neural network approximations of one-dimensional functions of different topology are used to build-up neural network database to determine the best neuron number in the hidden layer of single-layer perceptron in order to attain the required approximation quality.