Approximating Posterior Distributions in Belief Networks Using Mixtures