Marginalization without Summation Exploiting Determinism in Factor Algebra
It is known that solving an exact inference problem on a discrete Bayesian network with many deterministic nodes can be far cheaper than what would be expected based on its treewidth. In this article, we introduce a novel technique for this: to the operations of factor multiplication and factor summation that form the basis of many inference algorithms, we add factor indexing. We integrate this operation into variable elimination, and extend the minweight heuristic accordingly. A preliminary empirical evaluation gives promising results.
KeywordsBayesian networks exact inference factor algebra deterministic variables
Unable to display preview. Download preview PDF.
- 2.Corrada Bravo, H., Ramakrishnan, R.: Optimizing MPF queries: decision support and probabilistic inference. In: SIGMOD Conference, pp. 701–712 (2007)Google Scholar
- 7.Kjærulff, U.: Triangulation of graphs — algorithms giving small total state space. Tech. Rep. R-90-09, Dept. of Mathematics and Computer Science, Aalborg University (1990), http://www.cs.aau.dk/~uk/papers/R90-09.ps.gz
- 8.Larkin, D., Dechter, R.: Bayesian inference in the presence of determinism. In: Bishop, C.M., Frey, B.J. (eds.) Proceedings of Ninth International Workshop on Artificial Intelligence and Statistics, Key West, USA (2003)Google Scholar
- 10.Vomlel, J.: Exploiting functional dependence in Bayesian network inference. In: Darwiche, A., Friedman, N. (eds.) UAI 2002, Proceedings of the 18th Conference in Uncertainty in Artificial Intelligence, University of Alberta, Edmonton, Alberta, Canada, pp. 528–535. Morgan Kaufmann, San Francisco (2002)Google Scholar