Edge Server Association and Trajectory Planning of UAVs in Federated Edge Learning Architecture

Federated learning, due to its technical characteristic of distributed training, is widely integrated into collaborative unmanned aerial vehicle (UAV) swarms for executing various intelligent tasks such as target tracking and disaster monitoring. However, the traditional federated learning architecture, which relies on a single central server for global model aggregation, suffers from deteriorating communication delay due to the drastic changes in UAV mobility, thereby degrading the convergence performance of the model. To address this issue, a federated edge learning architecture is introduced into UAV swarms, utilizing multiple edge servers deployed in the network for local model parameter collection from neighboring UAVs to perform global model aggregation. To ensure rapid convergence of the learning model under this architecture, the association strategy between edge servers and UAVs and the trajectory planning of UAVs are jointly considered. We formulate an optimization problem to minimize the convergence delay of federated learning model and solve the problem by utilizing a deep reinforcement learning algorithm. Simulation results show that our proposed algorithm can effectively reduce the convergence delay and keep the energy consumption of model training in an acceptable range as well.