A Survey on Multi-agent Reinforcement Learning for Adaptive Transportation Solutions

Reinforcement learning (RL) has become an essential tool in optimizing transportation systems, offering promising solutions to dynamic and complex decision-making challenges. This paper explores the application of RL in four representative domains of transportation systems, including road-level traffic signal control, low-altitude unmanned aerial vehicle (UAV) navigation, high-altitude air traffic management (ATM), and maritime autonomous ship (AMS) navigation. These domains span land, low airspace, high airspace, and sea, reflecting key operational environments where multi-agent coordination is crucial. While significant advances have been made in the development of RL-based systems, traditional models often face limitations when applied to diverse, real-world environments due to their reliance on single-scenario training. To address this, we focus on cross-scenario multi-agent reinforcement learning (MARL), which aims to enhance the adaptability and scalability of RL models across multiple, distinct scenarios. Specifically, we examine offline RL and meta RL techniques, which enable agents to transfer policies learned in one environment to new, unseen scenarios. Despite the progress, challenges such as policy transferability, data efficiency, and scalability remain, which need to be addressed for broader application in transportation systems. This paper provides an overview of these challenges, reviews the state-of-the-art methods, and discusses future directions for advancing RL in transportation systems.