On the Classification Performance of Bayesian Inference Networks (BINs)

This report summarizes the analytic performance methodology for Bayesian Inference Networks (BINs). The use of probabilistic reasoning is distinct from traditional acoustic contact correlation (ACC) algorithms that rely primarily on geometric-based track evidence. Classification performance bounds are developed for uncorrelated bearing and bearing-rate track evidence. The approach can be generalized to address more germane cluster association problems based on correlated evidence. A belief or association probability is recursively computed from a Fredholm integral equation as a function of the evidence and indirectly as a function of the signal-to-noise ratio (SNR) and BIN iteration number. Receiver Operating Characteristic (ROC) curves are determined from the association probabilities or through the use of a Green’s function. Real-world examples are used to illustrate the modeling approach.