Learning-Based Stochastic Driving Model for Autonomous Vehicle Testing

In the simulation-based testing and evaluation of autonomous vehicles (AVs), how background (BVs) drive directly influences AV’s driving behavior further affects test results. Most existing simulation platforms use either predetermined trajectories or deterministic models to model BV behaviors. However, cannot react AV maneuvers, are different from real human drivers because lack stochastic components errors. Both methods lead unrealistic traffic scenarios. This paper presents a learning-based that meets unique needs (i.e., interactive human-like stochasticity). The is built based on long short-term memory architecture. By incorporating concept quantile regression into loss function model, behaviors reproduced without prior assumption drivers. trained with large-scale naturalistic data (NDD) Safety Pilot Model Deployment project compared intelligent (IDM). Analysis individual shows proposed can reproduce more similar than IDM. To validate ability in generating environment, experiments implemented. results show flow parameters such as speed, range, time headway distribution match closely NDD, which significant importance for evaluation.