Multi-agent meta-reinforcement learning based on contrastive predictive coding—For rapid adaptive traffic control after emergencies
- 1Faculty of Intelligent Transportation, Anhui Sanlian University, Hefei, Anhui, China
- 2Faculty of Engineering, Anhui Sanlian University, Hefei, Anhui, China
Abstract
Emergency signal control has to clear emergency vehicles through intersections without creating excessive delay for ordinary traffic. Many existing controllers still depend on manually designed traffic features, and unfamiliar emergency patterns often require additional training before the control policy can be adjusted reliably. For this setting, a multi-agent meta-reinforcement learning framework is developed using Contrastive Predictive Coding (CPC) for traffic state representation. CPC compresses high-dimensional traffic sequences into compact representations that are subsequently used by intersection agents during few-sample policy adaptation. Agent communication is weighted by both distance and transmission delay, while hierarchical emergency priority control adjusts signal decisions according to event severity. Tests on the iTETRIS simulation platform with the CityFlow Emergency dataset produced an emergency response time of 25.3 s, an average regular vehicle delay of 80.5 s per vehicle, and a 98.2% emergency success rate in high-priority scenarios. In the ablation experiments, removing CPC caused the largest single performance loss, whereas the all-module configuration showed a 44.3% contribution to response-time reduction in the contribution analysis. Overall, the framework maintained emergency vehicle priority while keeping the effect on regular traffic within a comparatively limited range under the evaluated emergency conditions.
Related
- Predicting miRNA–Disease Associations Based on Contrastive Learning of Linear and Nonlinear Features
- How effective are contrastive learning-based approaches for activity-cliff prediction?
- Revealing pan-cancer clonal niches of T cells from single-cell RNA sequencing using contrastive learning
- Pocket contrastive learning and conformation fusion for robust structure-conditioned 3D molecular generation under receptor conformational variation.
- Advancing long-read metagenomic binning via single-copy-gene guided contrastive learning
BibTeX
@article{jiang2026multiagent,
title = {{Multi-agent meta-reinforcement learning based on contrastive predictive coding—For rapid adaptive traffic control after emergencies}},
author = {Wei Jiang and Xianwei Huang and Genjie Wang and Ruting Li and Shanshan Tang},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0357779},
url = {https://doi.org/10.1371/journal.pone.0357779},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Multi-agent meta-reinforcement learning based on contrastive predictive coding—For rapid adaptive traffic control after emergencies
AU - Wei Jiang
AU - Xianwei Huang
AU - Genjie Wang
AU - Ruting Li
AU - Shanshan Tang
PY - 2026
DA - 2026/09/08
JO - PLOS One
DO - 10.1371/journal.pone.0357779
UR - https://doi.org/10.1371/journal.pone.0357779
AB - Emergency signal control has to clear emergency vehicles through intersections without creating excessive delay for ordinary traffic. Many existing controllers still depend on manually designed traffic features, and unfamiliar emergency patterns often require additional training before the control policy can be adjusted reliably. For this setting, a multi-agent meta-reinforcement learning framework is developed using Contrastive Predictive Coding (CPC) for traffic state representation. CPC compresses high-dimensional traffic sequences into compact representations that are subsequently used by intersection agents during few-sample policy adaptation. Agent communication is weighted by both distance and transmission delay, while hierarchical emergency priority control adjusts signal decisions according to event severity. Tests on the iTETRIS simulation platform with the CityFlow Emergency dataset produced an emergency response time of 25.3 s, an average regular vehicle delay of 80.5 s per vehicle, and a 98.2% emergency success rate in high-priority scenarios. In the ablation experiments, removing CPC caused the largest single performance loss, whereas the all-module configuration showed a 44.3% contribution to response-time reduction in the contribution analysis. Overall, the framework maintained emergency vehicle priority while keeping the effect on regular traffic within a comparatively limited range under the evaluated emergency conditions.
SN - 1932-6203
ER -