Multi-Agent Reinforcement Learning for Autonomous Decision-Making Systems: A Survey

Abstract

The autonomous systems are emerging in complex and dynamic contexts in which several intelligent agents have to coordinate, adapt and make decisions in the face of uncertainty. Single-agent reinforcement learning typically performs poorly in such environments due to issues with agent interactions, decentralized information, and non-stationary learning conditions. One promising framework that could enable multiple agents to learn cooperative, competitive, or mixed strategies, while continuously interacting with their environment and with each other, is Multi-Agent Reinforcement Learning (MARL). In this paper, a wide survey of MARL for autonomous decision-making systems is presented. It examines the core principles and models of MARL, such as cooperative and competitive learning paradigms, centralized training and decentralized execution (CTDE), and key algorithmic categories like value-based, policy-based, and actor–critic. In addition, the survey analyzes the use of MARL in autonomous decision-making architectures. It identifies its use in autonomous vehicles, unmanned aerial vehicle (UAV) swarms, multi-robot systems, smart grids, and industrial automation. Some of the main difficulties, such as scalability, coordination, credit assignment, safety, reliability, communication limitations, and partial observability, are also examined. Finally, the paper discusses current research trends and future directions to improve the scalability, robustness, explain ability, and real-world deployment of MARL-enabled autonomous systems. The survey provides an organized view of today's progress and challenges in this ever-changing field.