CoSLight: Co-optimizing Collaborator Selection and Decision-making to Enhance Traffic Signal Control

TL;DR

CoSLight tackles the problem of coordinating traffic signal control across non-adjacent intersections by jointly learning a collaborator-selection policy and a signal-decision policy. It introduces a Dual-Feature Extractor that captures phase- and intersection-level information and a lightweight Multi-Intersection Collaboration module to select top-$k$ collaborators, all optimized in an end-to-end policy-gradient framework with symmetry and diagonal constraints. Empirical results on synthetic and real-world networks show consistent improvements over state-of-the-art baselines in both scenario- and intersection-level metrics, and comprehensive ablations validate the necessity of each component. The work demonstrates that strategic, learned collaboration patterns, including non-neighbor interactions, can substantially reduce delays and improve traffic flow in complex urban networks, with reasonable computational overhead. Overall, CoSLight offers a scalable, interpretable approach to dynamic, joint optimization of collaboration and control in multi-intersection TSC.

Abstract

Effective multi-intersection collaboration is pivotal for reinforcement-learning-based traffic signal control to alleviate congestion. Existing work mainly chooses neighboring intersections as collaborators. However, quite an amount of congestion, even some wide-range congestion, is caused by non-neighbors failing to collaborate. To address these issues, we propose to separate the collaborator selection as a second policy to be learned, concurrently being updated with the original signal-controlling policy. Specifically, the selection policy in real-time adaptively selects the best teammates according to phase- and intersection-level features. Empirical results on both synthetic and real-world datasets provide robust validation for the superiority of our approach, offering significant improvements over existing state-of-the-art methods. The code is available at https://github.com/bonaldli/CoSLight.

Figures (22)

• Figure 1: (a)-(b): The coordination between areas of intersections during rush hours; (c)-(d): The collaboration policy $\boldsymbol{\rho}$ and decision policy $\boldsymbol{\theta}$ should be jointly optimized to prevent suboptimal.
• Figure 2: (a) The illustration of intersection. (b) There are 12 movements: [North, South, West, East (four approaches)] $\times$ [Left, Go-through, Right (three directions)]. Usually, turning right isn't signal-controlled, so only 8 movements (index from 1-8 in (a)) are signal-controlled. (c) A phase is two non-conflicting movements that can be released together. There are 8 phases, e.g., phase-A combines movements 1 and 5. (d) The signal-control policy is to select one phase for the next time step according to the traffic condition.
• Figure 3: Gains over hops: choosing the right collaborators is important.
• Figure 4: Overview of our proposed CoSLight: integrating Dual-Feature Extractor for phase- and intersection-level features with the module of Multi-Intersection Collaboration to select teammates for cooperation.
• Figure 5: Ablation Studies