CARTS: Cooperative Reinforcement Learning for Traffic Signal Control and Carbon Emission Reduction

ICLR 2026 Conference Submission14367 Authors

Keywords: Traffic Signal Control (TSC), Reinforcement Learning, Multiagent systems

Abstract: Existing traffic signal control systems often rely on overly simplistic, rule-based approaches. Even reinforcement learning (RL)-based methods tend to be suboptimal and unstable due to the inherently local nature of control agents. To address the potential conflicts among these agents, we propose a cooperative architecture named CARTS (CooperAtive Reinforcement learning for Traffic Signal control). CARTS introduces multiple reward terms, weighted with an age-decaying mechanism, to optimize traffic signal control at a global scale. Our framework features two types of agents: local agents that focus on optimizing traffic flow at individual intersections, and a global agent that coordinates across intersections to enhance overall throughput. Importantly, the system is designed to reduce both vehicle waiting time and carbon emissions. We evaluated CARTS using real-world traffic data obtained from traffic cameras in an Asian country. Despite incorporating a global agent during training, CARTS remains decentralized at inference time, requiring no centralized coordination during deployment. Experimental results show that CARTS consistently outperforms state-of-the-art methods across all evaluated performance metrics. Moreover, CARTS effectively links carbon emission reduction with global agent coordination, providing an interpretable and practical approach to sustainable traffic signal control.

Primary Area: reinforcement learning

Submission Number: 14367