Evidential networks from a different perspective

Abstract

Bayesian networks are, at present, probably the most popular representative of so-called graphical Markov models. Naturally, several attempts to construct an analogy of Bayesian networks have also been made in other frameworks as e.g. in possibility theory, evidence theory or in more general frameworks of valuation-based systems and credal sets. We collect previously obtained results concerning conditioning, conditional independence and irrelevance allowing to define a new type of evidential networks, based on conditional basic assignments. These networks can be seen as a generalization of Bayesian networks, however, they are less powerful than e.g. so-called compositional models, as we demonstrate by a simple example. © 2013 Springer-Verlag.