In a SUMO simulation of a Changchun expressway, a hybrid traffic controller called MPC-STMAPPO produced 25 simulated collisions, which the study reported as a 91.99% reduction from a no-control run. It recorded a mainline speed of 38.5 km/h at WS6, one of the seven weaving segments used in the risk analysis, and an on-ramp queue of 21.82 vehicles. Those were the highest speed and shortest queue among 11 control schemes. The result is the study's clearest signal: fewer simulated collisions alongside faster traffic and a shorter queue.
A controller built in layers
The proposed system is hierarchical. Its upper L-METANET-based model predictive control layer, a planning method that uses a traffic model to choose baseline commands, is paired with a lower ST-MAPPO layer that supplies residual actions, or learned adjustments. Together they control lane-level variable speed limits and ramp metering while accounting for collision risks at merging and diverging movements.
The SUMO testbed reproduced an 18-km Eastern Expressway network in Changchun at 1:1 scale, with nine variable-speed-limit gantries and three ramp meters. The risk analysis covered 14 tasks across seven weaving segments, including both merging and diverging risk types. That gave the controller a coordinated network setting rather than a single isolated risk task.
The numbers behind the result
Before the control comparisons, the authors calibrated L-METANET and METANET against SUMO detector ground truth using GA-VNS, a calibration search method. They compared outputs with randomly sampled five-minute aggregates of flow and speed. L-METANET's mean absolute percentage error, or MAPE, was 6.7% for speed in Lane 1 and 19.4% in the other lanes, versus 10.8% and 27.5% for METANET. For flow in the other lanes, L-METANET's MAPE was 12.1% versus 24.9%, while its root mean square error, or RMSE, was 133 versus 256. Lower values indicate closer agreement with the reference data.
For risk assessment, the study used an XGBoost-SHAP-RPBL pipeline. It applied a SHAP importance threshold of 0.2 and allowed no more than eight variables into a smooth analytical risk equation with random parameters. Its AUC score, which describes how well a classifier separates higher-risk from lower-risk cases, was above 0.70 across all 14 tasks. Most tasks exceeded 0.80, and merging risk at WS6 reached 0.889.
The reported training curves also favored the hybrid setup on how quickly it stabilized. MPC-STMAPPO converged within 90 episodes and settled near a reward of -26. In pure ST-MAPPO training, reward rose from about -46 during episodes 0 to 100, converged around episode 280 and stabilized near -25. Here, an episode is a repeated run through the simulated task, so these figures describe learning behavior rather than travel time on the road.
Across 13 network-wide core metrics, MPC-STMAPPO ranked in the top three for six and in the bottom three for only one. That places the headline speed, queue and collision results within a broader pattern rather than a single best-case measure. The study did not report formal significance tests for these rankings.
In altered-demand tests, the hybrid controller kept an advantage on the reported measures. At 1.5 times baseline demand, MPC-STMAPPO maintained WS6 speed at 16.4 km/h and recorded 88 collisions, compared with 109 for ST-MAPPO. Under Random demand 2, it maintained WS3 speed at 60.6 km/h, a reported 50.37% improvement over pure MARL, and reduced WS6 delay by a reported 17.22%. These findings point to robustness within the tested scenarios, not proof that the policy will transfer to every road or traffic pattern.
Where the evidence stops
The evidence has clear boundaries. The controller was evaluated in a SUMO testbed, so its collision counts and traffic outcomes are simulated results rather than measurements from a live expressway. The AUC results show classification performance on the study's 14 risk tasks, but they do not validate the model against observed crashes. The cited summaries also did not report confidence intervals or formal significance tests for the main outcome comparisons, convergence speed or transfer results. Questions about live deployment and other expressway geometries, demand patterns and vehicle mixes remain open.
The research was supported by the National Natural Science Foundation of China under grant 52472313. The authors declared no known competing financial interests or personal relationships that could have influenced the work. The paper says data will be made available on request.
Paper data and sources
Original title: Coordinated Lane-Level Variable Speed Limits and Ramp Metering for Successive Weaving Segments Considering Merging/Diverging Risks: A Hybrid Model Predictive Control and Multi-Agent Reinforcement Learning Approach
Authors: Guodong Ma, Baofeng Sun, Wenyu Yang, Zhihong Yao
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text