Towards Dynamic Bayesian Networks: State Augmentation for Online Calibration of DTA Systems

A key component of Dynamic Traffic Assignment (DTA) systems is the online calibration of simulation parameters, which is crucial in generating accurate predictions of network states. A widely used approach for online calibration is the Kalman filter which allows for the incorporation of demand and supply parameters and any type of measurement data. This paper presents a Dynamic Bayesian Network extension for traditional Kalman filters with a technique called state augmentation. Although it has been discussed in the calibration literature, the usage and applicability were not fully investigated. The state augmentation technique is particularly useful for delayed systems, for example in large networks with high travel times. In this paper, we discuss state augmentation for Kalman filtering and illustrate its modeling advantages via a Dynamic Bayesian Network (DBN) representation. These advantages are demonstrated by a case study using the Singapore expressway network. The results indicate that employing state augmentation yields better estimation and prediction accuracy of traffic states, around 10% less error than the standard extended Kalman filter.