A Comparative Study of Model Predictive Control and Deep Reinforcement Learning for Distributed Traffic Signal Control

This paper presents a comparative study of two distributed traffic signal control approaches: a cell-transmission-model-based model predictive control (MPC) method and a deep reinforcement learning (DRL) method based on latent spatio-temporal network (LST-Net). In the MPC method, each intersection predicts local traffic evolution over the prediction horizon using the cell transmission model (CTM) and exchanges predicted boundary traffic information with neighboring controllers to enhance coordination among adjacent intersections. In the DRL method, temporal local traffic observations are processed to generate local signal decisions, while a compact latent representation of the network-level spatial traffic state is incorporated during centralized training to improve value estimation and policy learning. The controllers are evaluated on a 4 × 4 grid road network across multiple demand levels and traffic realizations, including variations in temporal demand profiles and spatial demand distributions, with Fixed-Time and Max-Pressure included as reference controllers. The comparison reveals clear differences in how controller performance changes as traffic demand increases. LST-Net generally provides favorable traffic efficiency under low-to-moderate demand, whereas MPC exhibits more consistent system-level performance as demand increases. The results further show that favorable in-network performance does not necessarily imply lower total time when depart delay is considered. Under unseen traffic scenarios, MPC maintains comparatively consistent performance, whereas LST-Net remains competitive and performs particularly well in some scenarios but exhibits greater variation across different traffic conditions. These findings indicate that no single control strategy is uniformly superior across all tested traffic conditions.