A MARL-FL framework for intelligent acoustic cooperative communication in AUV swarms

Abstract Addressing the constraints imposed by the complexity of underwater acoustic channels on the cooperative communication performance of autonomous underwater vehicle (AUV) swarms, this paper constructs a cross-layer collaborative optimization framework based on Multi-Agent Reinforcement Learning and Federated Learning. By jointly optimizing the communication protocol stack and the distributed learning process, the system performance is enhanced. To support the efficient operation of this framework, two key technologies are proposed: (i) a Reinforcement Learning–based Dynamic Modulation Switching algorithm, which enhances channel adaptability through real-time optimization of physical layer parameters, and (ii) a Channel State Information–aware Lightweight Federated Learning mechanism, which utilizes dynamic compression techniques to reduce communication overhead and improve the efficiency of distributed training. These methods aim to overcome the challenges of underwater acoustic communication, achieving efficient, reliable, and intelligent cooperative operations for AUV swarms.