Policy-aware Optimization-based Modeling of Autonomous Vehicles’ Longitudinal Driving Behavior

Microscopic traffic models (MTMs) are widely used to evaluate the potential impacts of autonomous vehicles (AVs) deployment scenarios in our transportation network. Car-following (CF) and lane-changing (LC) models are the backbones of MTMs. Several studies attempt to accurately replicate these behaviors (especially CF behavior) using state-of-the-art modeling methods. A CF model consists of a set of relations and modifiable parameters that are calibrated by mass field driving data. Since mass field driving data of AVs are not available, researchers often assume these parameters and conduct impact assessments, leading to different conclusions on the potential effects of AVs. Meanwhile, AVs are agents and unlike human-driven vehicles, their behaviors are controllable and trainable. AVs might have safe and efficient driving behavior throughout a trip, therefore, we can train them to reach a destination optimally in a simulation environment. In this research, we develop an optimization framework that finds a set of optimized driving parameters for AVs under various scenarios, aiming to improve certain optimization targets (e.g., reducing travel time, number of conflicts) using a well-defined simulation-based objective function. The methodological framework consists of an optimization module and a simulation environment. The differential evolution (DE) method is employed within the optimization module to identify the optimized values of the CF parameters. The simulation environment is a SUMO-based platform where several simulation replications are conducted under certain scenario conditions. An experimental setup is designed to implement the proposed framework under different scenarios of mixed traffic and demand cases for the IDM (intelligent driving model), Krauss, and ACC (adaptive cruise control) models. The findings of this research reveal that safety could potentially be improved by optimized values of the CF model. For each policy where a higher weight is allocated to safety, generated optimized parameters significantly enhance safety as well as efficiency. In addition, the results show that minimum gap and desired time headway are the most sensitive parameters in regards to the policy targets, and their optimized values could replicate the potential CF behavior of AVs.