Multi-UAV Deployment and Resource Allocation: A federated deep reinforcement learning approach

As various communication methods evolve, the demand for communication services has surged. Unmanned aerial vehicles (UAVs), due to their flexibility and low cost, have emerged as ideal platforms for providing convenient communication services. However, traditional centralized algorithms consume significant resources under limited communication capacity, while distributed algorithms often struggle to achieve effective collaboration. To address this challenge, this study proposes a multiagent federated reinforcement learning (RL) algorithm aimed at optimizing the deployment and resource allocation of multi-UAV to enhance network utility. The approach employs personalized proximal federated learning, allowing each agent to learn and train independently based on local observations, while periodically uploading model parameters to a central server. The central server maintains a global model and returns updated parameters with proximal terms, thereby improving cooperation among agents while reducing communication overhead. Additionally, the algorithm incorporates a reward network competition mechanism locally to provide effective target values for model training. To thoroughly validate the proposed algorithm’s effectiveness, a multi-UAV environment is constructed that comprehensively accounts for signal interference and transmission loss. Simulation results demonstrate that, across various complex environments, this algorithm outperforms traditional methods in terms of network utility.