<title>Simulation of a neural network for decentralized detection of a signal in noise</title>

Abstract

The problem of distributed detection of a deterministic signal with unknown strength using a neural network approach is considered. Since the signal level Is not known there does not exist a uniformly powerful detection scheme for the problem. The neural network architecture used Is a two layer feedforward perceptron model with N Inputs M nodes In the hidden layer and two output nodes that correspond to the signal and noise hypotheses. The back propagation algorithm Is used to train the system with a known noise distribution. For fixed false alarm probabilities the performance of the trained network with regard to a set of test signals is estimated and compared to the performance of some known detection procedures such as Fisher''s method Tlppett''s method and the optimum test. Simulation results are presented for a network with two input nodes and three or four nodes In the hidden layer. When Laplace noise Is assumed the network performs better than the Tlppett method. For the detection of weak and moderate signals in Gaussian noise the network performs reasonably close to the optimum detection scheme. The decision regions for the neural network detection scheme is comparable to those of the Fisher''s and Tlppett''s methods. However unlike the Tippett''s method which has a fixed decision region regardless of the noise distribution the neural network''s decision region changes with the noise distribution. 1. SIGNAL DETECTION