In the previous section, we saw that the basic operation of the variable elimination algorithm is the manipulation of the factors. First, we create a factor by multiplying existing factors. Then, we eliminate a variable in to generate a new factor , which is then used to create another factor. From a different perspective, we can say that a factor is a data structure, which takes messages generated by the other factor , and generates a message which is used by the other factor .
Before we go into a detailed discussion of the belief propagation algorithm, let's discuss the graphical model that provides the basic framework for it, the clique tree, also known as the junction tree.
The clique tree () is an undirected graph over a set of factors , where each node represents a cluster of random variables and the edges connect the clusters, whose scope has a nonempty intersection. Thus, each edge between a pair of clusters and is associated with a sepset . For...