Practicable sensitivity analysis of Bayesian belief networks

By subjecting a Bayesian belief network to a sensitivity analysis with respect to its conditional probabilities, the reliability of its output can be evaluated. Unfortunately, straightforward sensitivity analysis of a belief network is highly time-consuming, as a result of the usually large number of probabilities to be investigated. In this paper, we show that the graphical independence structure of a Bayesian belief network induces various properties that allow for reducing the computational burden of a sensitivity analysis. We show that several analyses can be identi ed as being uninformative because the conditional probabilities under study cannot a ect the network's output. In addition, we show that the analyses that are informative comply with simple mathematical functions. By exploiting these properties, the practicability of sensitivity analysis of Bayesian belief networks is enhanced considerably. AMS classi cation: 60K10, 62A15, 62N05, 68Q25.