Deep Reinforcement Learning for Client Selection in Federated Learning
Wei Ye, Wenjiang Ouyang, Junsheng Mu, Jingchen Zhao · 2023
Federated learning (FL) is a distributed machine learning approach specifically developed to tackle the limitations and privacy concerns that arise from centralized data center. The problem of resource allocation in FL has generated significant attention in research. The objective is to find a way to allocate resources effectively while ensuring the accuracy of client training data. This paper proposes a scheme that employs reinforcement learning (RL) for client selection, capitalizing on the observation that the accuracy of FL improves as ascending clients are chosen with the increase of rounds. Furthermore, the effect of the number of learning rounds on the amount of client selections is taken into account. Experimental result shows that the largest gains are obtained when we select clients homogeneously with the rounds increasing. This approach presents a promising solution for optimizing resource allocation in FL.