Compiling Multiply Sectioned Bayesian Networks: A Comparative Study

Abstract

Inference with multiply sectioned Bayesian networks (MSBNs) can be performed on their compiled representations. The compilation involves cooperative moralization and cooperative triangulation. In earlier work, agents perform moralization and triangulation separately and the moralized subgraphs need to be made consistent to be the input of the triangulation. However, the set of moralized subnets is only an intermediate result, which is of no use except as the input to the triangulation. On the other hand, combining moralization and triangulation won’t make the compilation complex but simpler and safer. In this paper, we first propose a change to the original algorithm (the revised algorithm), which is supposed to provide higher quality compilation, then we propose an algorithm that compiles MSBNs in one process (the combined compilation), which is supposed to provide lower quality compilation, however. Finally, we empirically study the performance of all these algorithms. Experiments indicate that, however, all 3 algorithms produce similar quality compilations. The underlying reasons are discussed.