Learn to Select: Efficient Cross-device Federated Learning via Reinforcement Learning

Chunlin Tian, Zhan Shi, Li Li

Keywords: Cross-device Federated learning, Learning-based, Reinforcement Learning.

Abstract: Federated Learning (FL) is a collaborative training method that provides data privacy in the age of big data. However, it is often ineffective on edge devices due to their heterogeneous and constrained resources. The primary challenge is to identify devices with useful training data and available compute capability, which are both time-varying. In this paper, we propose FedRank, a novel federated learning approach based on reinforcement learning. Our approach addresses the device selection problem by casting it as a ranking problem and employing a pairwise training scheme. Furthermore, we leverage imitation learning from state-of-the-art algorithms to eliminate the cold start phenomenon that offsets the benefits of previous learning-based approaches. Experimental results show that FedRank improves model accuracy by 2.59\%-12.75\%, and accelerates the training process by up to 2.25$\times$ and 1.76$\times$ on average at the same time.