Sensitivity analysis in Bayesian networks

For systems based on Bayesian networks, evidence is used to compute posterior probabilities for some hypotheses. Sensitivity analysis is concerned with questions on how sensitive the conclusion is to the evidence provided. After the basic definitions and an example we conclude that the heart of sensitivity analysis is to compute probabilities for the hypotheses given various subsets of the evidence. We show how some of these probabilities come out as a side-effect of the HUGIN propagation method. Through a modification of the HUGIN method even more of the probabilities are achieved. Finally, we give methods for answering ''what if''-questions, for determining the set of crucial findings, and determining minimal sufficient sets of findings.