Traffic congestion is one of the most critical challenges in modern urban mobility, leading to significant delays, increased emissions, and economic losses worldwide. Despite decades of research, traditional traffic signal control (TSC) methods often struggle to adapt to the complex, dynamic, and stochastic nature of real-world traffic environments. This paper presents a comprehensive review of TSC approaches, aiming to bridge the gap between conventional control theories and next-generation AI-driven models. The review screens more than 400 publications, from which 148 studies demonstrating meaningful methodological innovation were selected as the analytical corpus. Unlike prior reviews that primarily focus on reinforcement learning (RL)-based techniques, this study offers a unified framework that integrates deep learning (DL) for representation learning, graph neural networks (GNNs) for spatial modeling, evolutionary algorithms (EAs) and genetic algorithms (GAs) for policy optimization, and fuzzy logic for uncertainty handling. Furthermore, we explore the emerging role of large language models (LLMs) and transformer-based architectures in enabling adaptive, context-aware, and multi-agent traffic control systems. Through both quantitative and qualitative synthesis, this review highlights emerging trends, identifies key research gaps including scalability, transferability, interpretability, and real-world deployment challenges and discusses the potential of next-generation AI models toward more autonomous and cooperative traffic management, while acknowledging that significant deployment challenges remain. The paper concludes with a road map for future research, emphasizing the integration of LLMs, self-supervised learning, and multi-agent coordination as promising but largely unevaluated directions for future smart city applications.
Patil et al. (Thu,) studied this question.