Resource-Aware Personalized Federated Learning Based on Reinforcement Learning

Abstract

Federated learning is an effective solution to protect data privacy, but the efficiency and performance of the entire federated system are challenging to balance due to the heterogeneity of client resources and data. To alleviate this dilemma, we propose a new training intensity allocation framework based on model structure. First, we design an optimization model that considers both time and energy consumption constraints, minimizing energy consumption under time constraints and accelerating the convergence of federated training. Then we construct a reinforcement learning-based model allocation strategy for realistic dynamic environments, automatically allocating appropriate model sizes to clients under complex communication conditions and heterogeneous computing resources. Finally, a large number of experiments demonstrate the feasibility and effectiveness of the proposed framework.