Scholarlink

<p>Coordinated traffic signal control along urban corridors is critical for reducing stop-and-go waves, vehicle delay, and queue spillback, yet remains challenging under large-scale topology, partial observability, and stochastic demand. Existing MARL controllers often rely on dense or static inter-intersection communication and locally scoped rewards, which weakens corridor-level credit assignment and leads to attention diffusion and unstable coordination as network scale grows. We propose the Graph Transformer Q-Network (GTQN) that learns a state-dependent sparse interaction graph for coordination and couples it with centralized governance during training so that decentralized controllers can reliably form and sustain corridor progression. Concretely, GTQN (i) learns which upstream/downstream peers are causally influential at each decision step via a two-stage mechanism (discrete peer gating followed by soft relevance weighting), yielding sparse, interpretable coordination without attention diffusion; (ii) uses a unified graph-transformer backbone to jointly encode network topology and temporal history so delayed upstream-to-downstream effects that determine green-wave formation are represented within the value function; and (iii) introduces a Collaborative Governance Graph (CTDE) trained with a progression-oriented two-level reward that couples junction efficiency with a corridor hindrance penalty, improving credit assignment and learning stability. Experiments on synthetic and real-world urban networks, including a 100-intersection Chengdu network, show that GTQN consistently improves progression quality, achieving the lowest Average Number of Stops (e.g., ANS <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(=0.62\)</EquationSource> </InlineEquation> on SQ1 and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(=1.78\)</EquationSource> </InlineEquation> on SQ3) while reducing waiting time and queues and increasing throughput. Ablations confirm that learned sparsity, unified spatiotemporal encoding, and centralized governance are each necessary for scalable and progression-aware coordination. Robustness study demonstrates that GTQN maintains acceptable performance under moderate communication perturbations such as message dropout and delays. Code for this study is archived at <a href="https://doi.org/10.5281/zenodo.18885573">https://doi.org/10.5281/zenodo.18885573</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph transformer Q-network for collaborative governance and decentralized decision-making in multi-intersection networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月31日

查看全文
  • He Zhang

摘要

Coordinated traffic signal control along urban corridors is critical for reducing stop-and-go waves, vehicle delay, and queue spillback, yet remains challenging under large-scale topology, partial observability, and stochastic demand. Existing MARL controllers often rely on dense or static inter-intersection communication and locally scoped rewards, which weakens corridor-level credit assignment and leads to attention diffusion and unstable coordination as network scale grows. We propose the Graph Transformer Q-Network (GTQN) that learns a state-dependent sparse interaction graph for coordination and couples it with centralized governance during training so that decentralized controllers can reliably form and sustain corridor progression. Concretely, GTQN (i) learns which upstream/downstream peers are causally influential at each decision step via a two-stage mechanism (discrete peer gating followed by soft relevance weighting), yielding sparse, interpretable coordination without attention diffusion; (ii) uses a unified graph-transformer backbone to jointly encode network topology and temporal history so delayed upstream-to-downstream effects that determine green-wave formation are represented within the value function; and (iii) introduces a Collaborative Governance Graph (CTDE) trained with a progression-oriented two-level reward that couples junction efficiency with a corridor hindrance penalty, improving credit assignment and learning stability. Experiments on synthetic and real-world urban networks, including a 100-intersection Chengdu network, show that GTQN consistently improves progression quality, achieving the lowest Average Number of Stops (e.g., ANS \(=0.62\) on SQ1 and \(=1.78\) on SQ3) while reducing waiting time and queues and increasing throughput. Ablations confirm that learned sparsity, unified spatiotemporal encoding, and centralized governance are each necessary for scalable and progression-aware coordination. Robustness study demonstrates that GTQN maintains acceptable performance under moderate communication perturbations such as message dropout and delays. Code for this study is archived at https://doi.org/10.5281/zenodo.18885573.