Completing incomplete bayesian networks

For reasoning with uncertain knowledge the use of probability theory has been broadly investigated. Two main approaches have been developed: Bayesian Networks and MaxEnt Completion. Bayesian Networks allow to describe large probability distributions, but require the specification of a graph structure on the involved variables together with a complete specification of a set of conditional probabilities according to this graph. Both the graph and the set of conditional probabilities may be cumbersome to work out. MaxEnt Completion is able to cope, in addition to uncertain knowledge, also with incomplete knowledge. Moreover, it does not depend on an ordering of knowledge (graph), which reduces the effort of specification. In this paper we investigate ways to combine these two approaches: We consider two kinds of incomplete Bayesian Networks — thus coping with incomplete (uncertain) knowledge — and study the usefulness of some variations of the MaxEnt Completion for processing or completing them. This analysis detected limits of the use of so-called update method ‘ground'.