Reinforcement Learning Agent Embedded in Roadside Units to Implement an Adaptive Variable Speed Limit System in a Cooperative Intelligent Transportation Systems Environment

Freeway shockwaves remain a major source of recurrent congestion, yet most reinforcement learning (RL) approaches to variable speed limit (VSL) control remain theoretical, relying on fixed detectors, predefined congestion zones and full observability. This paper introduces an RL agent embedded in a Roadside Unit (RSU) within a Cooperative Intelligent Transportation System. Operating solely through V2I communication, the agent collects speed and position data from Cooperative Awareness Messages, reconstructs traffic state as a space-time matrix built from sparse connected vehicle (CV) observations, and broadcasts speed limits, requiring no additional sensing infrastructure. Training proceeds in three stages: exhaustive dataset generation in a microscopic simulator, reward selection and hyperparameter calibration on a learned surrogate model, and few-shot refinement in the simulator. The transfer operates along two axes: (i) increasing environment fidelity, and (ii) the release of the simulation seed, so that a policy acquired deterministically is refined under stochastic conditions in 400 episodes, against the 4,000 required from a random policy. Across 30 seeds, the agent selects a consistent mitigation strategy from 30% CV penetration, where attenuation is partial; the physics-based benchmark performs better above 50%, while at full penetration the two are statistically indistinguishable, with lower inter-seed variance for the RL agent. Two coordinating RSUs with 2 km coverage manage a 10 km corridor.