Research map / Image generation / Autoregressive and latent generation
DiffLight: A Partial Rewards Conditioned Diffusion Model for Traffic Signal Control with Missing Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffLight, an offline conditional diffusion model that jointly imputes missing traffic data and makes signal-control decisions by conditioning on partial (observable) rewards, using a spatial-temporal transformer and a diffusion communication mechanism.
Problem
Traffic signal control when sensor data (observations and rewards) from intersections is missing, in an offline RL setting where missing data must be imputed while learning a policy.
Approach
Partial Rewards Conditioned Diffusion (PRCD) conditions only on observable rewards (masking missing ones instead of zero-padding), uses a STFormer noise model for spatial-temporal dependencies, and a Diffusion Communication Mechanism that shares generated neighbor observations at inference, with an inverse dynamics model producing actions.
Key results
- Under random missing (10%-50%), DiffLight achieves the lowest ATT in most settings, e.g., 286.17 on Hangzhou1 at 10% vs 289.38 for the next best (DD).
- Under kriging missing on Hangzhou1 at 6.25%-25%, DiffLight reaches ATT 294.18-302.16 while baselines range roughly 300-836.
- Ablation on Hangzhou1 (RM 50%): U-Net 668.38, STFormer 350.60, STFormer+PRCD 304.71 ATT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffLight rather than the strongest listed baseline | TSC with missing traffic data · Random missing across five datasets and rates | ATT | Mixed |
| Use DiffLight rather than the strongest listed baseline | TSC with missing traffic data · Kriging missing across five datasets and rates | ATT | Mixed |
| Replace U-Net with STFormer (ablation) | DiffLight · Hangzhou_1 RM (50%) | ATT: 668.38 -> 350.60 | Helped |
| Replace U-Net with STFormer (ablation) | DiffLight · Hangzhou_1 KM (25%) | ATT: 363.80 -> 318.21 | Helped |
| Replace U-Net with STFormer (ablation) | DiffLight · Jinan_1 RM (50%) | ATT: 509.64 -> 374.41 | Helped |
| Replace U-Net with STFormer (ablation) | DiffLight · Jinan_1 KM (25%) | ATT: 454.90 -> 374.23 | Helped |
| Use partial rewards conditioning instead of zero-padding missing rewards (ablation) | DiffLight with STFormer · Hangzhou_1 RM (50%) | ATT: 350.60 -> 304.71 | Helped |
| Use partial rewards conditioning instead of zero-padding missing rewards (ablation) | DiffLight with STFormer · Hangzhou_1 KM (25%) | ATT: 318.21 -> 302.16 | Helped |
| Use partial rewards conditioning instead of zero-padding missing rewards (ablation) | DiffLight with STFormer · Jinan_1 RM (50%) | ATT: 374.41 -> 290.02 | Helped |
| Use partial rewards conditioning instead of zero-padding missing rewards (ablation) | DiffLight with STFormer · Jinan_1 KM (25%) | ATT: 374.23 -> 329.67 | Helped |
| Add DCM (ablation) | DiffLight · Hangzhou_1 RM (50%) | ATT: 304.71 vs 303.13 | No clear effect |
| Add DCM (ablation) | DiffLight · Hangzhou_1 KM (25%) | ATT: 307.14 -> 302.16 | Helped |
Surprising findings
- Diffuser-style baselines fed with SFM-imputed data degrade sharply under kriging missing; DD reaches ATT up to 1049 at 25% kriging on Jinan2, due to error accumulation from imputation followed by diffusion and inverse dynamics.
- Zero-padding missing rewards as a condition hurts performance relative to conditioning only on observable rewards (STFormer vs STFormer+PRCD).
Limitations
- Only random and kriging missing patterns are considered; mixed or more complex patterns are not studied.
- Baselines use rule-based store-and-forward imputation rather than stronger imputation models.
- Conditioning on partial rewards instead of returns may cause short-sightedness of agents.
Experimental setup
Datasets: Hangzhou1 Hangzhou2 Jinan1 Jinan2 Jinan3 New York (48 intersections)
Benchmarks: CityFlow Hangzhou1 CityFlow Hangzhou2 CityFlow Jinan1 CityFlow Jinan2 CityFlow Jinan3 New York
Metrics: average travel time (ATT)
Baselines: BC CQL TD3+BC Decision Transformer Diffuser Decision Diffuser CQL (model-based)
Builds on: Diffuser Decision Diffuser classifier-free guidance DDIM MissLight
Closest papers by meaning
- RobustLight: Improving Robustness via Diffusion Reinforcement Learning for Traffic Signal Control (ICML 2025 · similarity 0.90)
- π-Light: Programmatic Interpretable Reinforcement Learning for Resource-Limited Traffic Signal Control (AAAI 2024 · similarity 0.87)
- Prompt to Transfer: Sim-to-Real Transfer for Traffic Signal Control with Prompt Learning (AAAI 2024 · similarity 0.86)
- VLMLight: Safety-Critical Traffic Signal Control via Vision-Language Meta-Control and Dual-Branch Reasoning Architecture (NeurIPS 2025 · similarity 0.85)
- RLFTSim: Realistic and Controllable Multi-Agent Traffic Simulation via Reinforcement Learning Fine-Tuning (CVPR 2026 · similarity 0.85)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.