Optimizing Client Selection for Federated Learning: A PPO-Based Method
Baheti Azhati, Xianda Li · 2024
Federated learning (FL), as a distributed machine learning approach, is widely recognized and applied in various fields due to its advantages such as multi-party participation, privacy protection, and reduced communication burden. However, in the context of edge networks, the performance of FL is challenged by the heterogeneity of device resources and data. This underscores the crucial importance of client selection. In our paper, we address this challenge by considering factors like model accuracy and device communication latency, constructing a multi-objective trade-off optimization problem. We formulate this problem as a Markov Decision Process (MDP) and we propose a proximal policy optimization (PPO)-based node selection algorithm. This algorithm aims to obtain an optimal strategy for effectively addressing the trade-off problem by selecting an ideal set of clients for model aggregation. Simulation results demonstrate that our method reduces model training latency and improves accuracy, exhibiting resilience against problematic clients.