Designing a procedure for the acquisition of probability constraints for bayesian networks

Abstract

Among the various tasks involved in building a Bayesian network for a real-life application, the task of eliciting all probabilities required is generally considered the most daunting. We propose to simplify this task by first acquiring qualitative features of the probability distribution to be represented; these features can subsequently be taken as constraints on the precise probabilities to be obtained. We discuss the design of a procedure that guides the knowledge engineer in acquiring these qualitative features in an efficient way, based on an in-depth analysis of all viable combinations of features. In addition, we report on initial experiences with our procedure in the domain of neonatology. © Springer-Verlag Berlin Heidelberg 2004.