Argumentation in Bayesian belief networks

This paper establishes an explicit connection between formal argumentation and Bayesian inference by introducing a notion of argument and a notion of defeat among arguments in Bayesian networks. First, the two approaches are compared and it is argued that argumentation in Bayesian belief networks is a typical multi-agent affair. Since in theories of formal argumentation the so-called admissibility semantics is an important criterion of argument validity, this paper finally proposes an algorithm to decide efficiently whether a particular node is supported by an admissible argument. The proposed algorithm is then slightly extended to an algorithm that returns the top-k of strongest admissible arguments at each node. This extension is particularly interesting from a Bayesian inference point of view, because it offers a computationally tractable alternative to the NPPP-complete decision problem k-MPE (finding the top-k most probable explanations in a Bayesian network).