A decision-theoretic performance benchmark for neural networks trained to discriminate two autoregressive processes

A two-class signal-classification task is presented for use in evaluating neural-network-based methods for classifier development. It is shown that if the output-amplitude scale of two autoregressive systems has no classification significance, a maximum-likelihood decision rule (MLDR) can be readily implemented and serve as a benchmark for network-based solutions. Results are presented from an initial experiment in which the back propagation learning algorithm was used to train a network for such a task. A 7% performance deficit is observed for the network classifier relative to the MLDR. There is little sensitivity to the number of nodes in the hidden layer of the network