A method, apparatus, and computer program product are presented for
automatically evaluating Bayesian network models. Operations performed
comprise receiving a Bayesian Network (BN) model including evidence nodes
and conclusion nodes that are linked with the evidence nodes by causal
dependency links, and where the evidence nodes have evidence states and
the conclusion nodes have conclusion states. The states of conclusion
nodes are set to desired conclusion states and corresponding
probabilities of occurrence of evidence states are determined by
propagating these states down the causal dependency links. Thus, samples
of most likely states of the evidence nodes are generated. Then, states
of the evidence nodes are set corresponding to the samples of the
evidence states. These states are propagated back up the causal
dependency links to obtain probabilities of the resulting states of the
conclusion nodes. Finally, a representation is outputted for the
probabilities of the states of the conclusion nodes.