Inverse Reinforcement Learning for Variable Speed Limit Systems

Abstract

This study investigates the use of Maximum Entropy Inverse Reinforcement Learning (IRL) to infer the underlying objective functions of Variable Speed Limit (VSL) systems. The objectives of such systems are often unknown or not explicitly accessible, limiting systematic analysis, comparison, and further development. Using a controlled simulation environment with a Model Predictive Control-based upstream speed reduction scheme, we evaluate whether IRL can recover the underlying objective from observed control behavior. Different reward function formulations are considered to assess the approach’s robustness. The results show that IRL achieves high policy imitation accuracies across all tested settings. While the control policy can be recovered from only observed traffic-state features, incorporating domain knowledge via engineered features improves the interpretability and accuracy of the learned reward structure. Overall, the findings demonstrate that IRL is a promising tool for uncovering implicit traffic control objectives, enabling systematic analysis and simulation of VSL strategies.