Simulation optimization of car-following models using flexible techniques

Car-following behavior is a key component of microscopic traffic simulation. Numerous models based on traffic flow theory have been developed for decades in order to represent the longitudinal interactions between vehicles as realistically as possible. Nowadays, there is a shift from conventional models to data-driven approaches. Data-driven methods are more flexible and allow the incorporation of additional information to the estimation of car-following models. On the other hand, conventional car-following models are founded on traffic flow theory, thus providing better insight into traffic behavior. The integration of data-driven methods in applications of intelligent transportation systems is an attractive perspective. Towards this direction, in this research an existing data-driven approach is further validated using another training dataset. Then, the methodology is enriched and an improved methodological framework is suggested for the optimization of car-following models. Machine learning techniques, such as classification, locally weighted regression (loess) and clustering, are innovatively integrated. In this chapter, validation of the proposed methods is demonstrated on data from two sources: (i) data collected from a sequence of instrumented vehicles in Naples, Italy, and (ii) data from the NGSIM project. In addition, a conventional car-following model, the Gipps' model, is used as reference in order to monitor and evaluate the effectiveness of the proposed method. Based on the encouraging results, it is suggested that machine learning methods should be further investigated as they could ensure reliability and improvement in data driven estimation of car-following models.