On the Detection of Conflicts in Diagnostic Bayesian Networks Using Abstraction

Abstract

An important issue in the use of expert systems is the so-called brittleness problem. Expert systems model only a limited part of the world. While the explicit management of uncertainty in expert systems mitigates the brittleness problem, it is still possible for a system to be used, unwittingly, in ways that the system is not prepared to address. Such a situation may be detected by the method of straw models, first presented by Jensen et al. [1990] and later generalized and justified by Laskey [1991]. We describe an algorithm, which we have implemented, that takes as input an annotated diagnostic Bayesian network (the base model) and constructs, without assistance, a bipartite network to be used as a straw model. We show that in some cases this straw model is better that the independent straw model of Jensen et al., the only other straw model for which a construction algorithm has been designed and implemented.