Abstract
Bayesian networks have many practical applications due to their capability to represent joint probability distribution over many variables in a compact way. Though there exist many algorithms for learning Bayesian networks from data, they are not satisfactory because the learned networks usually are not suitable directly for reasoning as they need to be transformed to some other form (tree, polytree, hypertree) statically or dynamically, and this transformation is not trivial [25]. So far only a restricted class of very simple Bayesian networks: trees and poly-trees are directly applicable in reasoning. This paper defines and explores a new class of networks: the Structured Bayesian Networks. Two methods of reasoning are outlined for this type of networks. Possible methods of learning from data are indicated. Similarity to hierarchical networks is pointed at.