Research Article
Reinforcement Learning Algorithms with Graph Convolution Networks for Traffic Signal Control
@INPROCEEDINGS{10.1007/978-3-031-86370-7_12, author={Shreya Salmalge and Shalabh Bhatnagar}, title={Reinforcement Learning Algorithms with Graph Convolution Networks for Traffic Signal Control}, proceedings={Intelligent Transport Systems. 8th International Conference, INTSYS 2024, Pisa, Italy, December 5--6, 2024, Revised Selected Papers}, proceedings_a={INTSYS}, year={2025}, month={4}, keywords={Traffic signal control reinforcement learning graph convolution networks actor-critic and deep Q-network algorithms}, doi={10.1007/978-3-031-86370-7_12} }- Shreya Salmalge
Shalabh Bhatnagar
Year: 2025
Reinforcement Learning Algorithms with Graph Convolution Networks for Traffic Signal Control
INTSYS
Springer
DOI: 10.1007/978-3-031-86370-7_12
Abstract
Traffic congestion is the root cause of various social and economic problems like longer travel times, increased pollution, and fuel or energy consumption. Addressing the issue is becoming increasingly crucial with rising city traffic and limited road infrastructure. The way we change traffic signals has a significant impact on congestion in road networks. We implement reinforcement learning algorithms for controlling traffic signals adaptive to congestion in incoming roads at junctions. Road networks can be viewed as graphs with intersections as nodes and roads as edges. This motivates us to use graph convolutional networks (GCN) as function approximators in various RL algorithms applied to traffic signal control. We implement Deep Q-learning (DQN), Graph Convolutional Q-learning (GCQN), Graph Convolutional Actor-Critic (GCAC), and individual-DQN models to learn a deterministic policy for adaptive traffic signal control. We also present a comparison of the performances of these models and infer that GCQN models are better suited to work for large road networks. To the best of our knowledge, the Graph Convolutional Actor-Critic model is not used in any existing traffic signal control method. We also compare the GCQN and GCAC models against existing and state-of-the-art approaches. Experimental evaluation shows that our proposed method achieves performance levels comparable to the state-of-the-art techniques.
