Advancing and Exploring Grid Network Abstractions for Applications in Traffic Management

This thesis proposes novel abstractions of grid road networks to evaluate decentralized Traffic Signal Controllers (TSCs) and their impacts on network stability and Macroscopic Fundamental Diagrams (MFDs). The study shows that networks consistently have a lower probability of gridlock when TSCs consider both upstream and downstream congestion in their signal plans. Additionally, we propose a methodology to train a Reinforcement Learning (RL) agent using abstractions of grid networks. The proposed training methodology leads to reduced computation. The simulation results demonstrate that the RL agent can effectively manage TSCs in networks and demands unseen during training.