Device scheduling and channel allocation for energy-efficient Federated Edge Learning

Federated Edge Learning (FEEL) is a promising distributed machine learning paradigm in the era of edge intelligence, which supports to learn the knowledge in the dataset on the premise of protecting users' privacy. However, this learning paradigm has a big defect. As the training process is carried out at the user end and is very power-consuming, the learning task is a serious challenge for mobile devices with limited battery capacity, which may also hinders the implementation of FEEL. In practical applications, FEEL usually needs to comply with the requirements for training delay and model performance, and may also be affected by the inter-cell interference which is common in the cellular networks. However, the current works only consider the demand for training delay. In this paper, we consider the implementation of FEEL in a general cellular network, and propose an empirical assumption to characterize the relationship between model performance and training data, based on which, a workload constraint is added to the formulated problem to guarantee the model performance. For the formulated problem that contains a summation term of integral variables and an interference term with complex structure at the denominator of the objective function, we propose a device scheduling and channel allocation strategy, also called double-greedy strategy, to obtain its suboptimal solution with low computational complexity. Simulation results verify the advancement of our proposed strategy relative to the existing works, that is, achieving the best energy efficiency on the premise of ensuring the model performance. This advancement makes our strategy more flexible to satisfy the possible various requirements of service providers for model performance.