Neural networks for classification of ARMA models: an experimental study

Abstract

The authors present a set of extensive experiments with alternative neural network, learning algorithms. These neural network configurations were tested on the problem of discriminating signals generated by an autoregressive moving-average (ARMA) linear system driven by white noise. These ARMA signals model a wide variety of signals arising in the ocean environment. The authors tested the various network models for their classification accuracy and speed of learning. The models investigated were back propagation, quickprop, Gaussian node networks, radial basis functions, the modified Kanerva methods, and networks without hidden units. For comparison, nearest-neighbor classifiers were also tested. Classification performance and learning time results are presented.<>