On-Line Probability Estimation for Bayesian Distributed Detection Systems in Non-Stationary Environments

Abstract In order to synthesize a binary target/no target decision, Bayesian distributed detection systems need the target probabilities and the local-detector performance probabilities. In this paper we discuss the use of three possible approaches to estimating these probabilities online, when they are not available a priori or when they change during an experiment. These techniques are: (i) decreasing-gain stochastic approximation, (ii) constant- gain stochastic approximation-like estimation, and (iii) decreasing-gain stochastic approximation enhanced by a disruption-detection algorithm. We demonstrate the operation of the algorithms in numerical simulations and conclude that both (ii) and (iii) can be useful for practical distributed detection systems which encounter jump-changes in the estimated probabilities.