Abstract
Traffic Signal Control (TSC) using Deep Reinforcement Learning (DRL) often struggles with the trade-off between generalizing across different urban environments and specializing in the unique characteristics of a specific intersection. We present DuaLight, a novel framework that decouples universal traffic principles from city-specific nuances. By integrating a scenario-shared "Co-Train" module with scenario-specific "Experiential Weights," DuaLight achieves state-of-the-art performance across diverse real-world and synthetic datasets, reducing average travel and delay times by 3-7%.
Introduction
Modern Traffic Signal Control is shifting from static, rule-based systems to adaptive Multi-Agent Reinforcement Learning (MARL). However, existing models face a "generalization vs. specialization" dilemma. Models trained on a single scenario often overfit to local quirks, failing when traffic patterns shift. Conversely, multi-scenario models often seek a "one-size-fits-all" policy that ignores critical structural differences between intersections, such as irregular 5-way junctions versus standard grids.
DuaLight addresses this by explicitly capturing both the universal "laws of traffic" (e.g., flow periodicity) and scenario-specific "empirical experiences" (e.g., which neighboring flows are most critical for a specific junction).
Methodology: The DuaLight Framework
DuaLight utilizes a dual-module architecture to balance global knowledge and local experience:
- Scenario-Shared Co-Train Module: This core architecture, comprising a Feature Extractor, Graph Attention Network (GAT), and MLP layers, is trained across all scenarios to capture fundamental traffic dynamics.
- Scenario-Specific Experiential Weight Module: This module uses learnable embeddings to assign Intersection-wise weights (prioritizing neighbor info vs. own state) and Feature-wise weights (focusing on relevant inputs like queue lengths).
These weights are integrated into a GAT, acting as a long-term memory of which spatial features have historically been most influential. An extension, DuaLight++, further enhances coordination by allowing intersections to aggregate information from "twin" intersections in entirely different cities based on embedding similarity.
Experiments and Results
The framework was evaluated using the SUMO platform across seven datasets, including synthetic grids and real-world regions in Germany (Cologne, Ingolstadt) and China (Fenglin, Nanshan).
- Performance: DuaLight outperformed SOTA models like CoLight and MetaLight, showing a 3-7% improvement in travel and delay metrics.
- Convergence: The model demonstrated significantly faster convergence, reaching stability in 10 steps in a 4x4 grid compared to 30 steps for previous best methods.
- Stability: Learning curves indicate that DuaLight maintains higher stability than meta-learning approaches, which often suffer from performance fluctuations during multi-scenario training.
- Visualization: Analysis of hidden embeddings confirmed that DuaLight successfully clusters agents from the same scenario together, meaning the experiential weights effectively capture the unique "signature" of each city.
Discussion
A key insight from this work is the "Twin Intersection" phenomenon: intersections in geographically distant cities can effectively "help" each other learn if they share structural similarities. Furthermore, ablation studies showed that while neighbor information is less critical in simple grids, it becomes vital in complex real-world scenarios like Nanshan.
Currently, DuaLight requires retraining to learn experiential weights for new cities. Future work will explore more flexible weight learning mechanisms to enable "zero-shot" adaptation to unseen urban environments.
Summary updated