- Article
Multi-agent reinforcement learning for traffic signal control commonly assumes fully observable intersection states. In real-world deployments, sensor faults and communication breakdowns produce missing observation data, which triggers severe policy oscillation and degrades control performance; existing methods cannot simultaneously achieve high-precision state imputation and low-latency edge deployment. This paper proposes DiffusionLight, a high-precision imputation framework that couples a gated-recurrent-unit state-evaluation module with shortcut-diffusion interpolation, and its lightweight variant DMLight. DiffusionLight estimates a missing state from a neighbourhood prior and the recent observation trajectory, starts the reverse diffusion chain from a learned latent point instead of pure noise, and is trained jointly with the multi-agent policy under a single composite loss; DMLight shares the interpolation parameters across intersections and replaces the dual-critic soft-actor-critic value network with single-branch Q-learning. Experiments on a 6 × 1 arterial corridor, a 4 × 4 Hangzhou grid, a 16 × 3 New York network and a 4 × 4 synthetic grid within the CityFlow (version 0.1)–LibSignal (DaRL Lab, Arizona State University (ASU)) platform show that, at 5–30% signal-missing ratios, DiffusionLight reduces average travel time (ATT) by 0.3–2.2% relative to DiffLight and by 32.7–40.3% relative to fixed-time control, while over the 6.25–25% equipment-fault range DMLight improves ATT by 0.1–19.2% over DiffLight and shortens the end-to-end run time by 35–81% relative to diffusion-based baselines. These simulation results provide a practical basis for deploying imputation-based signal control on edge devices and for subsequent field validation, offering an integrated route from simulation-grade accuracy to edge-oriented deployment.
20 November 2026






![Overall dimensions of the 40%-scale DrivAer fastback model used for the acceleration analysis, based on the DrivAer geometry of Heft et al. [7]; image taken from Shinde et al. [36].](https://mdpi-res.com/vehicles/vehicles-08-00218/article_deploy/html/images/vehicles-08-00218-g001-550.jpg)



