A new design methodology for optimal interpolative neural networks with application to the localization and classification of acoustic transients

Abstract

An evolutionary design methodology for neural networks based on the theory of optimal interpolation, (OI) is presented. A limited application of the OI net to the problems of localization and classification of acoustic transients is discussed. The modified recursive least squares (RLS) learning algorithm presented provides an avenue for the acquisition of an appropriate neural network configuration to solve a given pattern classification problem. The authors show that both OI and the back-propagation (BP) of comparable configurations perform satisfactorily in the simulations. The RLS OI method is preferred, however, because BP would occasionally run into some local minima and convergence could be very slow for the more complex decision boundaries between classes. The authors demonstrate that the OI net is particularly suited for application to the localization and classification of acoustic transients.<>