Bayesian Combination of Travel Time Prediction Models

Abstract

Short-term prediction of travel time is a central topic in contemporary intelligent transportation system (ITS) research and practice. Given the vast number of options, selecting the most reliable and accurate prediction model for one particular scientific or commercial application is far from a trivial task. One possible way to address this problem is to develop a generic framework that can automatically combine multiple models running in parallel. Existing combination frameworks use the error in the previous time steps. However, this method is not feasible in online applications because travel times are available only after they are realized; it implies that errors on previous predictions are unknown. A Bayesian combination framework is proposed instead. The method assesses whether a model is likely to produce good results from the current inputs given the data with which it was calibrated. A powerful feature of this method is that it automatically balances a good model fit with model complexity. With the use of two simple linear regression models as a showcase, this Bayesian combination is shown to improve prediction accuracy for real-time applications, but the method is sensitive in the event that all models are biased in a similar way. It is therefore recommended to increase the number and the diversity of the prediction models to be combined.