Prediction-Loss-Guided Long-Tail Augmentation and Uncertainty-Aware Reinforcement Learning for Car-Following Control

Reinforcement learning (RL) has demonstrated great potential for autonomous car-following control by directly optimizing key performance metrics, including safety, efficiency, and ride comfort. This study proposes an integrated RL-based car-following framework that combines trajectory prediction, reshaping of long-tail distributions, and uncertainty-aware decision-making. The trajectory prediction loss is redefined as a quantitative metric for scenario rarity and risk, facilitating a distribution-level data augmentation strategy that enriches safety-critical scenarios while preserving the naturalness of driving behavior. A probabilistic Gaussian LSTM predictor is further employed to model future vehicle motion uncertainty, which is explicitly incorporated into multi-step action evaluation through variance-weighted virtual rewards. Experiments on the highD dataset demonstrate that the proposed approach significantly improves robustness and safety, achieving substantial reductions in collision rates compared with baseline RL methods. Cross-dataset evaluation on NGSIM further confirms its generalization capability.