| In this thesis, we propose a method for modelling Bayesian network (BN) inference, as well as an efficient method for implementing BN inference. The modelling technique is a simple graphical representation of BN inference. In our method, the probability information propagated in a network is illustrated using black and white vertices. Our method possesses two salient characteristics. First, this purely graphical approach has the potential for being used as a pedagogical tool to introduce BN inference to beginners. Second, our method potentially provides a more accurate description of BN inference, as our approach only involves conditionals, a special case of potentials, which are traditionally used to describe inference.;The second main contribution of this thesis is that we develop the first join tree propagation algorithm involving prioritized messages. Current join tree propagation algorithms treat all propagated messages as being of equal importance. In real-world BNs, however, it is often the case that only some of the messages propagated from one join tree node to another are relevant to subsequent message computation at the receiving node. Our prioritized approach identifies and computes the relevant messages first. Therefore, by not forcing a join tree node to wait for irrelevant messages, our approach performs its work sooner. The efficiency improvement offered by our method is shown by using an empirical evaluation conducted on seven real-world and one benchmark BNs. As is usually done, inference is performed in each network with varying amounts of evidence, namely, zero, nine and eighteen percent. Our prioritized join tree propagation approach finished inference faster than the current state-of-the-art exact Bayesian inference algorithm in all twenty four cases, which show the practical significance of our approach. |
