Simulating Car-following Behavior for Heteregeneous Drivers: the Need for Driver Specific Model Parameters

Abstract

Car-following models are the core component of microscopic traffic simulation. Most of the deterministic models take fixed parameter values for different drivers. However, considerable behavioral differences exist between individual drivers. Simulating car-following behaviors of different drivers thus poses a challenge for microscopic traffic simulation. In this study, three approaches to calibrating car-following models for a group of heterogeneous drivers (calibrating an `average' driver, calibrating at an individual-driver level, calibrating based on clustered drivers' data) were tested with real-world driving data. Specifically, twenty randomly selected drivers' car-following trajectories extracted from the Safety Pilot Model Deployment (SPMD) project were used to calibrate the intelligent driver model (IDM) with the abovementioned three calibration approaches. The errors of replicating drivers' behavior in the validation datasets were used to evaluate the performances of the three calibration approaches.Results show that 1) calibrating at the individual level (i.e., each driver has its own model parameters) has the best performance in replicating a group of drivers' car-following behavior; 2) calibrating an `average' driver based on all drivers' data performs worst; 3) calibrating at the cluster level achieves intermediate performance; and 4) simply averaging calibrated individual drivers' parameters is not a good way to simulate a group of heterogeneous drivers' car-following behavior. The results suggest that inter-driver heterogeneity in car-following should not be neglected in microscopic traffic simulation, and also that there is a need to develop archetypes of a variety of drivers to build a traffic mix in simulation.