Reinforcement Learning for Freeway Lane-Change Regulation via Connected Vehicles

Lane change decision-making is challenging due to complex vehicle-vehicle and vehicle-infrastructure interactions. Existing lane-change control methods often rely on vehicles with some level of autonomy, limiting their applicability at low penetration rates of automated vehicles. To address this issue, we propose a lane-change regulation framework based on multi-agent reinforcement learning (MARL) to improve freeway traffic efficiency via connected vehicles. Regulation signals, such as allowing or prohibiting left or right lane changes, are computed at a traffic management center and broadcast to connected vehicles, while human-driven vehicles remain uncontrolled. Compared with vehicle-level maneuver control, the framework reduces communication and positioning requirements and avoids direct trajectory intervention. It combines microscopic traffic simulation with a macroscopic lane-grid representation: vehicle-level trajectories generate the realized traffic dynamics, while aggregated lane-grid states support low-cost grid-agent control. Based on a multi-lane macroscopic traffic model represented by partial differential equations (PDEs), lane changes are modeled as source-term exchanges between adjacent lanes and regulated through the MARL actions. Experiments across multiple traffic scenarios, demand levels, and connected-vehicle penetration rates show that the proposed method improves overall traffic efficiency with limited additional energy consumption while maintaining comparable driving safety.