An Optimization Method for Non-IID Federated Learning Based on Deep Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) enables collaborative model training without data sharing.
• Non-independent and identically distributed (non-IID) data across clients hinders FL convergence speed and accuracy.
• Existing FL methods face challenges in efficiently handling non-IID data distributions.

Purpose of the Study:

• To develop a novel federated learning framework, FedRLCS, to address the non-IID data challenge.
• To accelerate model convergence and improve accuracy in FL settings with heterogeneous data.
• To optimize client selection for efficient collaborative training.

Main Methods:

• Designed FedRLCS, a federated learning framework integrating deep reinforcement learning (DRL).
• Enhanced the double DQN (DDQN) algorithm's greedy strategy and action space for optimal client subset selection.
• Simulated non-IID data distributions across clients using partitioned datasets.

Main Results:

• FedRLCS significantly reduces communication rounds by 10-70% compared to state-of-the-art non-IID FL methods.
• The framework achieves target accuracy with fewer communication epochs across various datasets and models.
• No additional computation or storage costs are imposed on clients.

Conclusions:

• FedRLCS effectively accelerates convergence and improves performance in federated learning with non-IID data.
• DRL-based client selection is a viable strategy for overcoming data heterogeneity in FL.
• The proposed method offers an efficient solution for practical FL deployments facing data silos.