Evidence absorption : experiments on different classes of randomly generated belief networks

More and more real-life applications of the belief network framework begin to emerge. As applications grow larger, the networks involved increase in size accordingly. For large belief networks, the computations involved in probabilistic inference tend to become rather time consuming, even so to an unacceptable extent. To address this problem, we have proposed in a previous paper to incorporate the method of evidence absorption into Pearl's algorithms for probabilistic inference. In the present paper, the ability of this method to improve on the average-case computational expense of probabilistic inference is illustrated by means of experiments performed on different classes of randomly generated belief networks. Both the set-up of the experiments and the results obtained are detailed. The results from our experiments are shown to reflect to a large extent the use of belief networks incorporating a randomly generated digraph. We comment on this observation by addressing the suitability of using of randomly generated belief networks in this type of experiment