P.I. BIDJUK, A.S. GASANOV, S.H. ABDULLAYEV
STRUCTURAL AND PARAMETRIC SYNTHESIS OF PREDICTIVE RBF NEURAL NETWORKS USING ARTIFICIAL IMMUNE SYSTEMS


The purpose of the study is in development of methodology for the structural and parametric syntheses of predictive radial basis function neural networks using the ideas from artificial immune systems. The settings for the RBF neural networks were determined using appropriately constructed immune system, and combined method for forecasting time series with controlled parameters of model is proposed. It was established that increase in the population size slows down the neural network learning process, but on the other hand it resulted in improvement of the models quality. The combined forecasting algorithm and wavelet neural network shows higher accuracy of prediction than the combined algorithm and RBF network, while the latter has a higher rate of training. It was also established that for a higher level of mutation, which implies a high variability of clones of the population, the training is faster, but stability of the process is lower, which decreases the probability of finding the global optimum.

Keywords: artificial immune systems, neural networks, radial basis functions, forecasting
© Copyright.
Copyright © 1997- Institute of Control Systems of Azerbaijan National Academy of Sciences. e-Mail: [email protected]