On-Line Estimation of Probabilities for Bayesian Distributed Detection

Abstract A learning binary Bayesian distributed detection system is investigated. It is comprised of a bank of local detectors and a data fusion center where each local detector chooses one of two hypotheses, based on collected measurements, to minimize a local Bayesian cost. The data fusion center combines the local decisions into a global decision, minimizing a global Bayesian cost. This architecture requires that each local detector know the a priori probabilities of the hypothesis and that the data fusion center know, in addition, the probabilities of false alarm and missed detection of each detector that communicates its decisions to it. A set of stochastic approximation rules is developed for systems which do not possess knowledge of these probabilities. These rules are reduced-bias versions of recursive mean estimators and allow “blind” operation of the distributed detection system well below signal-to-noise ratios that would be required otherwise.