ReconLight: Reconciling Individual Optimization and Global Coordination in Multi-Agent Traffic Signal Control
Authors:
- Xiaoruo ZhangXZXiaoruo Zhang
- Chengwei ZhangCZChengwei Zhang
- Ziyu ZhouZZZiyu Zhou
- Wanting LiuWLWanting Liu
- Na ZhaoNZNa Zhao
- Furui ZhanFZFurui Zhan
Preprint
- 15 August 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Multi-Agent Reinforcement Learning (MARL) has shown great promise for Adaptive Traffic Signal Control (ATSC). A core objective of ATSC is to minimize network-wide average travel time while satisfying practical constraints such as traffic fairness. However, achieving this in multi-intersection settings is difficult, primarily because the credit assignment problem makes it hard to attribute global outcomes to individual intersection decisions. To enable effective learning, existing methods resort to dense surrogate rewards, such as queue lengths or the number of waiting vehicles at individual intersections, which are immediately observable and readily optimized. However, there is no guaranty that these locally defined auxiliary objectives are consistently aligned with the global goal: reducing queue length at one intersection may simply shift congestion downstream rather than genuinely improve network-wide efficiency. Moreover, per-intersection surrogate rewards naturally overlook network-level concerns such as fairness, which requires balancing congestion across intersections and is inherently a global objective that cannot be captured by purely local metrics. To address these intertwined challenges, we propose ReconLight, a novel MARL framework that enables a smooth transition from locally responsive optimization to globally coordinated optimization through model aggregation across actor networks. ReconLight employs a dual actor-critic architecture that optimizes individual and group objectives separately, with a shared Graph Attention Network (GAT) capturing spatial dependencies among intersections. A time-varying parameter fusion mechanism progressively blends individual and global actor parameters using cosine-annealed weights and an initial-parameter anchor, enabling efficient and fair network-wide control through a stable model-aggregation process. Experiments on both synthetic and real-world datasets demonstrate that ReconLight consistently outperforms competitive baselines across all tested scenarios, achieving up to a 3.34\% reduction in average travel time.Keywords
This publication has 15 references indexed in Scilit:
- Deep Reinforcement Learning for Cooperative Intelligent Transportation Systems: A Survey on Architecture, Use Cases, and Future DirectionsIEEE Transactions on Intelligent Transportation Systems, 2026
- Using a single actor to output personalized policy for different intersectionsNeural Networks, 2026
- Neighbor-Advised Multiagent Reinforcement Learning for Traffic Signal Control in Partially Sensored NetworksIEEE Transactions on Automation Science and Engineering, 2026
- Distributed Machine Learning for Autonomous Agent Swarm: A SurveyIEEE Communications Surveys & Tutorials, 2025
- Sequential Decision MARL for Adaptive Traffic Signal Control With Different Intersections PrioritiesIEEE Transactions on Intelligent Transportation Systems, 2025
- Vehicle-Level Fairness-Oriented Constrained Multi-Agent Reinforcement Learning for Adaptive Traffic Signal ControlIEEE Transactions on Intelligent Transportation Systems, 2025
- A General Scenario-Agnostic Reinforcement Learning for Traffic Signal ControlIEEE Transactions on Intelligent Transportation Systems, 2024
- STMARL: A Spatio-Temporal Multi-Agent Reinforcement Learning Approach for Cooperative Traffic Light ControlIEEE Transactions on Mobile Computing, 2020
- Multi-Agent Deep Reinforcement Learning for Large-Scale Traffic Signal ControlIEEE Transactions on Intelligent Transportation Systems, 2019
- IntelliLight2018