Comparison of multiagent inference methods in multiply sectioned Bayesian networks

摘要

As intelligent systems are being applied to larger, open and more complex problem domains, many applications are found to be more suitably addressed by multiagent systems. Multiply sectioned Bayesian networks provide one framework for agents to estimate what is the true state of a domain so that the agents can act accordingly. Existing methods for multiagent inference in multiply sectioned Bayesian networks are based on linked junction forests. The methods are extensions of message passing in junction trees for inference in single-agent Bayesian networks. Many methods other than message passing in junction trees have been proposed for inference in single-agent Bayesian networks. It is unclear whether these methods can also be extended for multiagent inference. This paper presents the first investigation on this issue. In particular, we consider extending loop cutset conditioning, forward sampling and Markov sampling to multiagent inference. They are compared with the linked junction forest method in terms of off-line compilation, inter-agent messages during communication, consistent local inference, and preservation of agent privacy. The results reveal issues to be considered in investigating other single-agent oriented inference methods. The analysis provides those who implement multiagent probabilistic inference systems with a guide on the pros and cons of alternative methods.