Abstract: Connected and automated vehicles (CAVs) can be beneficial for improving the operation of highway bottlenecks such as weaving sections. This paper proposes a bi-level control approach based on an upper-level deep reinforcement learning controller and a lower-level model predictive controller to coordinate the lane-changings of a mixed fleet of CAVs and human-driven vehicles (HVs) in weaving sections. The upper level represents a roadside controller that collects vehicular information from the entire weaving section and determines the control weights used in the lower-level controller. The lower level is implemented within each CAV, which takes the control weights from the upper-level controller and generates the acceleration and steering angle for individual CAVs based on the local situation. The lower-level controller further incorporates an HV trajectory predictor, which is capable of handling the dynamic topology of vehicles in weaving scenarios with intensive mandatory lane changes. The case study inspired by a real weaving section in Basel, Switzerland, shows that our method consistently outperforms state-of-the-art benchmarks.
Introduction
Highway weaving sections—segments where on-ramps and off-ramps are located in close proximity—are notorious for causing traffic bottlenecks. The intense mandatory lane-changing required as vehicles cross paths leads to a phenomenon known as capacity drop, where the actual discharge flow of a highway can plummet by 3% to 20% once congestion begins.
While Connected and Automated Vehicles (CAVs) offer a promising avenue for traffic coordination, existing solutions often face challenges. Local-only coordination systems ignore the broader highway context, while centralized optimization approaches for every vehicle on the road suffer from prohibitive computational complexity as traffic density increases. This research proposes a scalable bi-level system that provides global oversight without sacrificing real-time local responsiveness.
Methodology
The proposed hierarchical architecture decouples global strategy from local execution through two distinct layers:
Upper-Level Roadside Coordinator
A roadside unit utilizes Proximal Policy Optimization (PPO), a Deep Reinforcement Learning (DRL) algorithm, to monitor the entire weaving section. Instead of direct vehicle control, it outputs "control weights" that define the priorities for individual vehicles, such as balancing the need to reach an exit quickly versus maintaining a steady speed for the fleet.
Lower-Level On-Vehicle Controller
Each CAV runs its own Model Predictive Control (MPC). It receives the weights from the roadside unit and calculates optimal acceleration and steering angles based on its immediate surroundings. This ensures that vehicles respect physical limits and safety constraints, such as collision avoidance, while pursuing the global objectives set by the upper level.
HV Trajectory Prediction
To effectively manage mixed traffic, the system integrates an HV trajectory predictor based on EvolveGCN. Unlike standard Graph Neural Networks, EvolveGCN handles the dynamic topology of weaving sections, where the spatial relationships between vehicles change rapidly during lane-changing maneuvers. This allows the CAVs to anticipate human driver behavior more accurately.
Experimental Evaluation and Results
The model was trained on 530,000 simulated samples and evaluated using a case study of a real-world weaving section in Basel, Switzerland. The simulation was calibrated using empirical traffic data to ensure realistic vehicle behavior and flow.
The "Bi-Level-G" (Global) method consistently outperformed state-of-the-art benchmarks across all tested CAV penetration rates. As shown in the performance data, the system provides significant benefits even during the transition period of vehicle automation.
At a 20% CAV penetration rate, the system improved space-mean speed by 8.83% and exit flow by 4.81%. At 100% penetration, these improvements reached 54.73% and 12.86%, respectively. Crucially, the bi-level controller almost entirely mitigates the capacity drop, maintaining free-flow conditions even under high demand.
Conclusion
By controlling the objective function of local MPCs via a global DRL coordinator, this bi-level framework achieves efficient traffic management in complex weaving sections. The decoupling of high-level strategy from low-level safety-critical control allows for a scalable solution that outperforms traditional centralized or purely local methods. Future research will focus on human driver adaptation to CAV presence and the impact of communication reliability on system performance.
Summary updated