THE STRUCTURE OF THE PARAMETER SPACE OF CAR-FOLLOWING MODELS

Kurzfassung

Typically, the calibration of microscopic traffic flow models seeks to minimize the speed or the distance error, but not both of them simultaneously. By analyzing the car-following part of a couple of microscopic traffic flow models with data from car-following situations, it is demonstrated that the two objectives depend on each other in a complicated way: Minimizing the speed error does not lead to a minimum distance error and vice versa. This means the existence of a Pareto front in the speed-error-distance-error space. The Pareto front in particular, and the parameter space in general, can be probed in a systematic manner by exploring the parameter space of a given model with quasi-random sequences. Furthermore, this systematic scanning of the parameter space allows for a ranking of the importance of the parameters of the model, albeit in the context of the data given. It is demonstrated that only a sub-set of the parameters of a given model are actually needed. So far, this is only theoretical work, since it relies strongly on a fast implementation of the microscopic models. However, with the advance of faster and more parallel computers such an approach can even be implemented with ease in realistic settings and with any traffic simulation program that allows for an external setting of its parameters.