The effect of number of queries and amount of evidence on the trade-off between clique tree propagation and variable elimination

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Two exact algorithms for inference in a Bayesian network (BN) are clique tree propagation (CTP) and variable elimination (VE). With CTP there are some pre- computations, but after that multiple queries can be performed e ciently. VE has no precomputation and irrelevant variables can be pruned. Processing evidence after the precomputations generates extra cost for CTP. With VE evidence is just taken into account during the query. So CTP is suited for multiple queries, because the cost of the precomputations can be spread over the queries, and VE is suited for a single query where the observations are all given. Within this thesis is investigated how this trade-o relates to the number of queries and evidence. The algorithms were used for queries in several networks with di erent amounts of observations. The time of the processes before and during the queries was measured and the amortization period was calculated, which is the amount of queries needed to make CTP more preferable than VE. The number of nodes, the average degree and the average number of values per variable all a ected the amortization period. When the values for these properties were below 50, 3.5 and 3 respectively, this period was only a few queries. It increased when the values increased and at some point VE would always be preferred. Having evidence coming in made CTP more favorable; only when large parts of the network became observed VE started to bene t from it. These results can give more insight in the trade-o between the algorithms.
Faculteit der Sociale Wetenschappen