Bayesian Calibration of Car-Following Models

Recent research has revealed that there exist large inter-driver differences in car-following behavior such that different car-following models may apply to different drivers. This study applies Bayesian techniques to the calibration of car-following models, where prior distributions on each model parameter are converted to posterior distributions. The priors and posteriors are then used to calculate the so-called ‘evidence’, which can be used to quantitatively assess how well different models explain one driver's car-following behavior. When considered over multiple drivers, the evidence represents probabilities for different models as a whole. These model probabilities can be used in a micro simulation, where for each driver first a model is drawn according to these probabilities, after which parameters are drawn from the posterior distribution for each parameter of that model that were obtained when calibrating the model. In a test case on actual data the Bayesian evidence indeed reveals inter-driver differences and it is shown how these differences can quantitatively be assessed.