Modeling oscillatory car following using deep reinforcement learning based car following models

In this work, we use reinforcement learning (RL) to train a car following model for vehicle jerk. The learned model is specifically trained for car following in low-speed oscillatory driving conditions such as stop-and-go traffic typical in congested urban centers. This driving is of particular interest since it is difficult to model and substantially contributes to urban air pollution. The proposed model is calibrated using experimental data and the model performance is compared to a baseline calibrated intelligent driver model (IDM). The proposed RL model is able to outperform the IDM in some metrics, while the IDM has lower error in others. This indicates that the proposed RL model is able to capture the general car following behavior in low-speed oscillatory driving conditions without overfitting to the training data and represents a first step toward realistic car following models that capture the full range of driver behavior.