Rule-based systems with unreliable conditions

Abstract

This paper deals with the problem of inference under uncertain information. This is a generalization of a paper of Cardona et al. (1991a) where rules were not allowed to contain negations. In contrast, this paper discusses inference with rules involving negations. This provides more flexibility in the modeling process of knowledge, but it introduces the possibility of contradictions and the reasoning is therefore not necessarily monotone. It is well known that when the rules can be organized in a tree, computations can be performed by a simple local propagation scheme. However, when the graph of rules is not a tree, an alternative procedure to the usual Markov tree cover method called factorization is presented. This method is new in the context of evidential reasoning but it is widely used in the reliability theory of complex systems and in Bayesian networks. The whole model is placed under the unifying Dempster-Shafer theory of evidence. It has also tight connections with de Kleer's assumption-based truth maintenance systems (ATMS). The method is illustrated on an example borrowed from Pearl (1988) and translated into an evidential framework rather than a Bayesian one.