Optimization of the Forecasting Neural Network Parameters for Quality of Service Management Tasks

The problem of designing a training sample in the task of forecasting time series in the presence of a delay in the effect of individual factors on the dependent variable associated with the uncertainty of the selection of training pairs was considered in the paper. The authors considered a consistently complicating problem of determining the optimal lag at which the shape of the response surface is most consistent with the form of real dependency. The method of accounting for time lags based on the approximation and the subsequent nonlinear optimization of the parameters of the neural network was proposed. Used two-layer homogeneous neural network with serial communications, with sigmoid transfer functions with a finite number of neurons. The algorithm using stochastic methods of minimization to solve of nonlinear optimization problems was proposed. The stochastic method allows to overcome the difficulties caused by the local minima encountered by the method of back propagation and other gradient descent methods. A generalized system model that predicts the time series, which gives optimal time lag, including: the distribution of training and validation sets, approximation for finding weights of a network, optimization, search time delay, shift the original series of observations, predicting values of the dependent variable was proposed by the authors.