The return coefficient task of determination of properties of a construction design by means of the device of neural networks is considered. As model the neural network in the form of a polutorasloyny predictor with possibility of iterative accumulation of volume is taken. Each subsequent step connected with addition of a new stream of neurons, is carried out only after training of the previous stream. Coefficients of synoptic communications pay off with the help of procedure of the return distribution from a condition of a minimum of function of an assessment. Algorithms of the solution of direct and return problems of multidimensional approximation are developed.
Keywords: neural network, approximation, synoptic communications, return distribution, construction designs