Triangulation Heuristics for BN2O Networks

A BN2O network is a Bayesian network having the structure of a bipartite graph with all edges directed from one part (the top level) toward the other (the bottom level) and where all conditional probability tables are noisy-or gates. In order to perform efficient inference, graphical transformations of these networks are performed. The efficiency of inference is proportional to the total table size of tables corresponding to the cliques of the triangulated graph. Therefore in order to get efficient inference it is desirable to have small cliques in the triangulated graph. We analyze existing heuristic triangulation methods applicable to BN20 networks after transformations using parent divorcing and tensor rank-one decomposition and suggest several modifications. Both theoretical and experimental results confirm that tensor rank-one decomposition yields better results than parent divorcing in randomly generated BN2O networks that we tested.

• Type: Conference
• Year: 2009
• Where: ECSQARU
• Authors: Petr Savický, Jirí Vomlel
• Keywords: Artificial Intelligence, ECSQARU 2009, Parent Divorcing, Tensor Rank-One Decomposition, Triangulated Graph