Abstract
Bayesian networks in their Factor Graph Reduced Normal Form (FGrn) are apowerful paradigm for implementing inference graphs. Unfortunately, thecomputational and memory costs of these networks may be considerable, even forrelatively small networks, and this is one of the main reasons why thesestructures have often been underused in practice. In this work, through adetailed algorithmic and structural analysis, various solutions for costreduction are proposed. An online version of the classic batch learningalgorithm is also analyzed, showing very similar results (in an unsupervisedcontext); which is essential even if multilevel structures are to be built. Thesolutions proposed, together with the possible online learning algorithm, areincluded in a C++ library that is quite efficient, especially if compared tothe direct use of the well-known sum-product and Maximum Likelihood (ML)algorithms. The results are discussed with particular reference to a LatentVariable Model (LVM) structure.