A physics-informed deep learning paradigm for car-following models

Car-following behavior has been extensively studied using physics-based models, such as Intelligent Driving Model (IDM). These models successfully interpret traffic phenomena observed in the real world but may not fully capture complex cognitive process of driving. Deep learning on other hand, have demonstrated their power capturing require a large amount driving data to train. This paper aims develop family neural network based car-following that are informed by which leverage advantage both (being data-efficient and interpretable) deep generalizable) models. We design physics-informed model (PIDL-CF) architectures encoded with 4 popular - IDM, Optimal Velocity Model, Gazis-Herman-Rothery model, Full Difference Model. Acceleration is predicted for regimes: acceleration, deceleration, cruising, emergency braking. The generalization PIDL method further validated two representative models: artificial networks (ANN) long short-term memory (LSTM) model. Two types PIDL-CF problems studied, one predict acceleration only jointly discover parameters. also demonstrate superior performance Next Generation SIMulation (NGSIM) dataset over baselines, especially when training sparse. results physics those without.