Maximizing entropy for inference in a class of multiply connected networks

Bayesian networks were developed by Pearl, Lauritzen, and others in the late 1980s and now constitute one of the leading technologies for applying AI to real world problems. In many applications, it is necessary to work with multiply connected Bayesian networks. In this paper it is shown that minimally prejudiced estimates of missing information may be calculated, for certain classes of multiply connected Bayesian networks, using the maximum entropy principle. The loop cutset conditioning method for updating, as devised by Pearl and Jensen, is utilized. We discuss the set of independencies required by the maximum entropy model and provide an example of the theoretical work.