A two-level probabilistic approach for validation of stochastic traffic simulations: impact of drivers’ heterogeneity models

This paper shows how traffic heterogeneity, and the way it is encoded into a model, drastically affects a model ability to reproduce observed traffic. Being heterogeneity a major source of uncertainty, to correctly frame the proposed validation methodology the authors have first reviewed and adapted cross-disciplinary theoretical concepts about uncertainty modelling to traffic simulation. A number of open issues, including error compensation and model overfitting, has been interpreted and clarified through the proposed framework. A two-level probabilistic approach has been applied to run stochastic simulations of three NGSIM I-80 traffic scenarios, and quantitatively infer the impact of heterogeneity. According to this approach, both the car-following and the lane-changing models of each vehicle have been calibrated against observed trajectories. Based on the estimated parameters distributions, different models of heterogeneity have been quantitatively validated against macroscopic traffic patterns. Being traffic a collective phenomenon emerging from microscopic interactions, even models calibrated on microscopic trajectories need to be quantitatively validated on macroscopic traffic patterns too. Among other results, normal distributions of the model parameters, which are customarily applied in traffic simulation practice, have been found unable to reproduce the observed congestion patterns. Parameters correlation, being claimed as highly influential in previous works, is responsible for a model overfitting in traffic scenarios with low congestion. In the end, it has been demonstrated that a thorough characterization of parameters heterogeneity cannot be left out in traffic simulation, if an ersatz representation of traffic is to be avoided.


  • English

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01760454
  • Record Type: Publication
  • Files: TRIS
  • Created Date: Nov 26 2020 3:11PM