Real time diagnosis with compiling Bayesian networks

This paper presents an approach to inference in Bayesian networks, which meets the uncertainty and real time challenges in diagnosis. The process can be described as compiling multi-linear function of a Bayesian network into Arithmetic Circuit for online inference. There are three advantages of this approach. First of all, offline compiling process decreases online inference time. Our method differs from the original approach of compiling Bayesian networks using variable elimination by omitting step of representing Bayesian network with Algebraic Decision Diagrams, which simplifies offline compiling process. Then, local structure reduces Arithmetic Circuits and results in less space consumption and online inference time. Finally, differential approach of inference has high efficiency as it calculates answers to multiple queries simultaneously without repeating compiling. Simplified differential approach of inference for particular situation, when evidences in fixed, decreases the searching time and improves efficiency of online inference.