Communication-Aware Optimizers for FedMeta Learning in Aerial Wireless Networks

Federated Meta (FedMeta) learning integrates meta learning with Federated Learning (FL) towards addressing heterogeneity challenges in edge intelligence. Implementing FedMeta on uncrewed aerial vehicles (UAVs)-assisted aerial networks offers compound benefits, including optimal trajectory design and resource allocation. The quick adaptability of FedMeta to new edge devices with minimal data is challenged in adverse UAV-based wireless networks due to frequent update losses that lead to model bias and overfitting towards data from devices with better channels. In this letter, we propose an optimizer for the global FedMeta model update that is suitable for adverse channel conditions. We show that the proposed optimizer outperforms the state-of-the-art AdamW optimizer for FedMeta learning. By leveraging tools from stochastic geometry, particularly in UAV-orchestrated networks, we gain insights into channel behavior and integrate them into our algorithm. We also introduce a novel hybrid update rule combining our optimized strategy with AdamW, achieving superior convergence speed and overall accuracy. Extensive simulations on LEAF datasets under unreliable channel conditions validate the effectiveness of our methods.