FedAWT: Adaptive federated learning via dynamic epoch adjustment for heterogeneous clients in ad-hoc networks

Recent advancements in IoT devices have enabled on-device processing for tasks such as real-time inference and autonomous control. However, federated learning (FL) in such environments remains challenging due to limited computation resources, unstable connectivity, and severe data and system heterogeneity. To overcome these challenges, this paper proposes FedAWT (Federated learning with Adaptive Weighted Tree), a decentralized FL framework designed for ad-hoc IoT networks. A latency-aware system model is constructed to reflect communication delay, bandwidth constraints, and variations in device-level computational performance. FedAWT introduces two core mechanisms. The first is a dynamic epoch adjustment method that allocates training epochs according to local convergence status and training stability. The second is a minimum spanning tree based communication strategy that reduces model transfer latency. Experimental results confirm that FedAWT improves test accuracy by up to 33.86 % compared to baseline methods in heterogeneous environments. These results demonstrate the effectiveness of FedAWT in enhancing convergence stability and training efficiency under resource-constrained and decentralized environment.