Belief Network Algorithms: A Study of Performance

Belief Updating Algorithms There are several algorithms for exact belief updating, for example, the polytree algorithm, clustering (Pearl 1988) or the Jensen tree method (Jensen, Lauritzen, & Olesen 1989). However, approximate methods are often preferred because the complexity of exact updating is NP-hard. Approximate updating is usually done by stochastic simulation (Pearl 1988). Variants include likelihood weighting, survival-of-the-fittest and Markov Chain Monte Carlo methods. Another approach to complexity reduction is to approximate the model by simplifying the network. Some of the existing methods do this by state-space abstraction, removal of weak links, replacing small probabilities with zero and graph pruning. Such procedures may be applied individually or in combination.