Energy Optimization for Hierarchical Federated Learning based on Over-the-Air Computing
Kang Zheng, Zhuo Li · 2023
Hierarchical Federated Learning (HFL) employs Device-to-Device (D2D) communication to facilitate collaborative model training among mobile nodes. But the communication and storage costs still limit the number of nodes involved in simultaneous training. Applying Over-the-Air Computing (AirComp) to enable a portion of these nodes to act as cluster heads can efficiently reduce energy consumption, the amount of data stored on the cloud server and storage costs. In this paper, we introduce node collaboration via AirComp to HFL, and define an energy optimization problem, which is proved to be NP-Complete. We design an nodes cooperation algorithm, which has an approximation ratio of m. Through experiments, we compare the energy consumption of using AirComp with the existing method FedAvg, resulting in 57% reduction in energy consumption. Additionally, we also compare the NCA algorithm with the greedy nodes collaboration algorithm and show that the NCA algorithm reduces energy consumption by 15% compared to the greedy nodes collaboration algorithm.