Belief Networks and Local Computations

This paper is one of many attempts to introduce graphical Markov models within Dempster-Shafer theory of evidence. Here we take full advantage of the notion of factorization, which in probability theory (almost) coincides with the notion of conditional independence. In Dempster-Shafer theory this notion can be quite easily introduced with the help of the operator of composition. Nevertheless, the main goal of this paper goes even further. We show that if a belief network (a D-S counterpart of a Bayesian network) is to be used to support decision, one can apply all the ideas of Lauritzen and Spiegelhalter’s local computations.