Challenges in Applying Calibration Methods to Stochastic Traffic Models

This paper evaluates calibration and validation as a means to understand traffic flow models better. The paper concentrates on the car-following part of these models and demonstrates that the calibration of stochastic models under certain circumstances can become difficult. Three types of stochasticity are distinguished for microscopic traffic flow models: the stochasticity coming from noisy data, the stochasticity coming from the distribution of the parameters describing the driver’s behavior, and the stochasticity coming from the model itself when a noise component is added to a deterministic differential equation governing the vehicle’s movements. By using four submodels with four noise terms and an identical deterministic part, this paper shows that a calibration with synthetic and, therefore, reproducible data can lead to results that are awry. The parameters fitted by the calibration procedure were significantly different for the deterministic and stochastic models. It was concluded that stochasticity was the reason why the parameter estimation of stochastic models sometimes failed. Up to now, the authors have not been able to propose a solution to cope with this intrinsic pitfall of genuine stochastic models.