Belief Network Inference Algorithms: a Study of Performance Based on Domain Characterisation

In recent years belief networks have become a popular representation for reasoning under uncertainty and are used in a wide variety of applications. There are a number of exact and approximate inference algorithms available for performing belief updating, however in general the task is NP-hard. To overcome the problems of computational complexity that occur when modelling larger, real-world problems, researchers have developed variants of stochastic simulation approximation algorithms, and a number of other approaches involve approximating the model or limiting belief updating to nodes of interest. Typically comparisons are made of only a few algorithms, and on a particular example network. We survey the belief network algorithms and propose a system for domain characterisation as a basis for algorithm comparison. We present performance results using this framework from three sets of experiments: (1) on the Likelihood Weighting (LW) and Logic Sampling (LS) stochastic simulation algorithms; (2) on the performance of LW and Jensen's algorithms on state-space abstracted networks, (3) some comparisons of the time performance of LW, LS and the Jensen algorithm. Our results indicate that domain characterisation may be useful for predicting inference algorithm performance on a belief network for a new application domain. 1