Joint Client Selection and Training optimization for Energy-Efficient Federated Learning

Federated Learning (FL) allows multiple edge devices (EDs) to collaboratively train shared models without sharing raw data, offering a prospective privacy-preserving machine learning framework. However, energy consumption has become a major challenge due to the high energy requirements of FL tasks and the limited battery capacity of EDs. Considering the high heterogeneity of resources and data among EDs, client selection strategies and control of local iteration numbers are key factors for optimizing energy efficiency in FL. In this paper, we present Global-aware Independent Proximal Policy optimization (GIPPO), a novel deep reinforcement learning method that jointly optimizes client selection and local training in heterogeneous environments. GIPPO considers each client as an individual intelligent agent and prioritizes all the clients in each training round based on their respective states. By intelligently selecting clients and assigning appropriate local iteration numbers, the method aims to minimize system energy consumption and enhance model performance. Experimental results demonstrate that our proposed method achieves significant energy cost savings, up to 77%, compared to baseline methods. Moreover, our method effectively adapts to varying degrees of data heterogeneity.