Traffic Signal Optimization by Integrating Reinforcement Learning and Digital Twins

Kumarasamy, Vijayalakshmi K; Saroj, Abhilasha; Liang, Yu; Wu, Dalei; Hunter, Michael; Guin, Angshuman; Sartipi, Mina · 2023 · OpenAlex-citations

DOI: 10.1109/swc57546.2023.10448974

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of optimizing traffic signal timing in large-scale intelligent transportation systems, where centralized machine learning methods often lack scalability and require costly real-world deployment for training. The authors propose a novel framework integrating Decentralized Graph-based Multi-Agent Reinforcement Learning (DGMARL) with a Digital Twin (DT) to reduce traffic congestion and network-wide fuel consumption associated with vehicle stops. The primary motivation is to leverage the DT as a safe, efficient testbed for training RL agents using realistic traffic data, thereby avoiding the risks and expenses of field trials while improving adaptability in complex networks. The methodology employs a DT architecture driven by PTV-Vissim, a microscopic traffic simulation engine, which replicates the MLK Smart Corridor in Chattanooga, Tennessee, comprising 11 signalized intersections. The DT processes real-time and historical data, including vehicle counts, detector occupancy, and Signal Phasing and Timing (SPaT) data, to dynamically simulate traffic conditions. The DGMARL model treats each intersection as a local agent within a graph structure, where agents share state information with neighbors via message passing to account for upstream and downstream traffic impacts. Agents utilize an Advantage Actor-Critic algorithm with Graph Neural Networks to learn optimal signal control policies. The reward function is based on the Eco PI metric, which penalizes stops and stop delays to minimize fuel consumption. Actions are constrained by physical safety rules, such as minimum green times and pedestrian recall requirements. Experimental results demonstrate the efficacy of the proposed approach compared to a baseline actuated signal control plan. In a 24-hour simulation scenario, the DGMARL-integrated DT reduced the Eco PI performance measure by 44.27%. During the PM peak-hour scenario, the reduction averaged 29.88%. These findings indicate that the decentralized multi-agent approach, assisted by the high-fidelity Digital Twin, significantly outperforms traditional actuated control in reducing stops and associated fuel consumption. The integration allows agents to learn dynamic traffic patterns and make coordinated decisions that enhance overall network efficiency. The significance of this work lies in its demonstration of a scalable, data-driven solution for traffic management that bridges the gap between simulation and real-world deployment. By using a Digital Twin to facilitate RL training, the approach offers a robust method for optimizing traffic signals without disrupting live traffic operations. The study highlights the potential of combining DTs with decentralized RL to handle heterogeneous data and complex network constraints, providing a pathway for more adaptive, efficient, and environmentally friendly intelligent transportation systems. This framework supports future real-time implementations by establishing a reliable interface between physical infrastructure and AI-driven optimization algorithms.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• traffic density