Title :
Variations over the message computation algorithm of lazy propagation
Author :
Madsen, Anders L.
Author_Institution :
HUGIN Expert, Aalborg, Denmark
fDate :
6/1/2005 12:00:00 AM
Abstract :
Improving the performance of belief updating becomes increasingly important as real-world Bayesian networks continue to grow larger and more complex. In this paper, an investigation is done on how variations over the message-computation algorithm of lazy propagation may impact its performance. Lazy propagation is a junction-tree-based inference algorithm for belief updating in Bayesian networks. Lazy propagation combines variable elimination (VE) with a Shenoy-Shafer message-passing scheme in an attempt to exploit the independence properties induced by evidence in a junction-tree-based algorithm. The authors investigate, the use of arc reversal (AR) and symbolic probabilistic inference (SPI) as alternative algorithms for computing clique-to-clique messages in lazy propagation. The paper presents the results of an empirical evaluation of the performance of lazy propagation using AR, SPI, and VE as the message-computation algorithm. The results of the empirical evaluation show that no single algorithm outperforms or is outperformed by the other two alternatives. In many cases, there is no significant difference in the performance of the three algorithms.
Keywords :
belief networks; inference mechanisms; message passing; Shenoy-Shafer message-passing scheme; clique-to-clique messages; junction-tree-based algorithm; lazy propagation; message computation algorithm; real-world Bayesian network; symbolic probabilistic inference; variable elimination; Bayesian methods; Belief propagation; Clustering algorithms; Helium; Inference algorithms; Inference mechanisms; Knowledge representation; Sampling methods; Tree graphs; Uncertainty; Bayesian networks; belief updating; inference mechanisms; knowledge representation; Algorithms; Artificial Intelligence; Bayes Theorem; Computer Simulation; Decision Support Techniques; Models, Statistical; Pattern Recognition, Automated;
Journal_Title :
Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on
DOI :
10.1109/TSMCB.2005.862488