Official code release for GTQN This release provides the official implementation accompanying the paper: Graph Transformer Q-Network for Collaborative Governance and Decentralized Decision-Making in Multi-Intersection Networks GTQN is a multi-agent reinforcement learning framework for coordinated traffic signal control in urban multi-intersection networks. The method combines: a decentralized Distributed Junction Controller (DJC) with two-stage sparse coordination using discrete peer gating and soft relevance weighting, Centralized Training with Decentralized Execution (CTDE) through a Collaborative Governance Graph (CGG), a unified spatiotemporal representation that integrates temporal Transformer encoding with graph-aware coordination, an event-driven SUMO/TraCI environment wrapper for multi-intersection control with phase-duration actuation, and reproducible configurations and ablation toggles for attention, reward design, and governance. What this release includes This release contains the codebase used to study corridor-level progression and decentralized coordination in multi-intersection traffic networks under partial observability, stochastic demand, and large-scale topology. It includes support for: training and evaluation in SUMO-based multi-intersection environments, synthetic benchmark traffic networks, Chengdu-based real-traffic evaluation scenarios, ablation studies on sparse attention, reward design, governance, and spatiotemporal modeling, and reproducible experiment configuration files. Main contribution GTQN learns a state-dependent sparse interaction graph for coordination between intersections and combines it with centralized governance during training to improve corridor-level credit assignment and decentralized decision-making. Reproducing the main results To reproduce the main results reported in the paper: Install SUMO and ensure sumo / sumo-gui are available in PATH, with SUMO_HOME configured. Create a Python environment and install dependencies from requirements.txt. Prepare a SUMO scenario under gtqn/envs/networks//. Train the model using scripts/train.py with the provided config files. Evaluate trained checkpoints using scripts/eval.py. The repository is structured to support reproduction of the main experimental results, including training on synthetic networks and evaluation on the SQ1, SQ2, and SQ3 scenarios. Reported results In the associated paper, GTQN achieves strong progression and coordination performance across benchmark networks, including: SQ1: ANS = 0.62 SQ2: ANS = 0.74 SQ3: ANS = 1.78 with consistent improvements in average waiting time, queue length, and throughput compared with baseline methods. License This code is released under the MIT License. See LICENSE for details. Citation If you use this code in your research, please cite the associated paper. Release purpose This is the first public release of the GTQN codebase and is intended to support transparency, reproducibility, and reuse of the reported research results.