Physics-informed deep reinforcement learning-based car-following model and simulation analysis for different vehicle Types

Car-following models describe how vehicles dynamically follow a leader in traffic. These models are essential for simulating traffic flow and analyzing its characteristics. Traditional car-following models often have simple structures and exhibit low prediction accuracy, whereas data-driven models improve prediction accuracy but often lack interpretability and show low simulation stability due to their ‘black-box’ nature. To overcome these limitations, this paper introduces a physics-informed deep reinforcement learning-based car-following model. The model includes a spatio-temporal information extraction module that captures spatial, temporal, and vehicle-level features during car-following to enhance decision-making. Additionally, we apply a deep reinforcement learning framework alongside a variable time step simulation method to more accurately replicate consistent driver behavior. We conducted traffic flow simulation experiments, confirming that our model effectively reproduces natural stochastic driving behaviors while accurately reflecting real traffic conditions. The proposed model accurately represents car-following behavior, ensuring that traffic flow simulations closely reflect objective reality.