Online Updating of Conditional Linear Gaussian Bayesian Networks

This paper presents a method for online updating of conditional distributions in Bayesian network models with both discrete and continuous variables. The method extends known procedures for updating discrete conditional probability distributions with techniques to cope with conditional Gaussian density functions. The method has a solid foundation for known cases and may be generalised by a heuristic scheme for fractional updating when discrete parents are not known. A fading mechanism is described to prevent the system being too conservative as cases accumulate over long time periods. The effect of the online updating is illustrated by an application to predict the number of waiting patients at the emergency department at Aalborg University Hospital.