Towards improved Bayesian fusion through run-time model analysis

This paper considers the accuracy of state estimation based on classification using Bayesian networks. It presents a method to localize network fragments that (i) are in a particular (rare) case responsible for a potential misclassfication, or (ii) contain modeling errors that consistently cause misclassifications, even in common cases. We derive an algorithm that, within such fragments, can localize the probable cause of the misclassification. The approach is based on monitoring the Bayesian network's 'behavior' at runtime, specifically the correlation among sets of evidence. We suggest several applications for the algorithm's output, such as repairing or mitigating the effects of errors, or deactivating faulty information sources.