Decentralized graph-based multi-agent reinforcement learning for traffic signal optimization

Signalized intersections are persistent bottlenecks where inefficient operations contribute to congestion, delays, safety risks, and environmental impacts. Conventional control strategies provide stability under predictable demand but lack the adaptability required to manage stochastic and heterogeneous traffic conditions. This dissertation develops a decentralized graph-based multi-agent reinforcement learning (DGMARL) framework for adaptive traffic signal control. The framework advances the state of the art by (i) embedding operational constraints, including minimum/maximum green durations, pedestrian recalls, and clearance intervals, directly into the learning process; (ii) modeling intersections as decentralized agents that exchange direction-specific states through multi-head graph attention to capture asymmetric flows and upstream inflows, thereby enabling scalable coordination across large networks; and (iii) incorporating contextual pedestrian demand via point-of-interest weighting. Control policies are optimized within a constrained Markov decision process, where modular phase selection and fairness-aware rewards jointly balance vehicle efficiency and pedestrian accessibility. The framework is validated using a high-fidelity digital twin–based simulator with real-world traffic data and further demonstrated through preliminary on-street field testing on the MLK Smart Corridor. In simulation, the proposed approach reduced pedestrian waiting times by up to 24.7% and vehicle delays by 22.6%, while decreasing emissions (CO, CO_2, NO_x, and PM_10) by an average of 9.6% and increasing vehicle throughput by more than 22%. These improvements were achieved while ensuring compliance with safety-critical signal timing rules. Analysis of graph attention weights highlights interpretable coordination across intersections, confirming the robustness and scalability of the decentralized design under varied traffic conditions. In field operation, the decentralized agents correctly interpreted real-time traffic demand, switched signal phases adaptively, and responded to pedestrian push-button activations within approximately 10-15 s along with proper recall logic, maintaining Safety and Priority of Timing (SPaT) compliance and end-to-end processing latency between 80 and 120 ms. Together, these contributions establish a pathway toward deployment-ready, equitable, and sustainable traffic signal control across diverse network settings.

Grantor dc:publisher

University of Tennessee at Chattanooga

Year dc:date.available

2027

Author dc:creator

• K Kumarasamy, Vijayalakshmi

Contributors dc:contributor

• Liang, Yu
• Wu, Dalei; Sartipi, Mina; Sun, Pengyuan
• College of Engineering and Computer Science

• Intelligent transportation systems
• Reinforcement learning
• Traffic signs and signals--Control systems--Automation

Licence

• http://rightsstatements.org/vocab/InC/1.0/

Language dc:language

English, eng

Repository record dc:identifier

https://scholar.utc.edu/theses/1034

OAI identifier oai:identifier

oai:scholar.utc.edu:theses-2222