Long-term Consistent Simulation of Car-following Behavior in Autonomous Driving Test Scenarios

Traffic modeling is crucial for simulation tests of autonomous driving. High-fidelity and low-complexity traffic models facilitate traffic simulation for effective and efficient simulation test results. Recent works have shown the effectiveness of data-driven car-following models with various neural networks. In the pursuit of model accuracy, the complexity of the model increases. However, high-complexity models restrict large-scale traffic simulation, and the simulated car-following behavior may not guarantee long-term consistency with naturalistic driving data. Based on the analysis, we pointed out that the accumulation of model errors over time during simulation is the reason for the inconsistency. To deal with it, a loss-modified training method is proposed, to improve the long-term consistency without increasing the complexity of the model. Experiments conducted on real-world datasets demonstrate that the Loss Modified training method significantly improves the long-term consistency of car-following models and avoids abnormal speed during simulation.