MAB-Based Online Client Scheduling for Decentralized Federated Learning in the IoT

Area of Science:

• Computer Science
• Machine Learning
• Wireless Communications

Background:

• Conventional federated learning (FL) relies on central servers, limiting scalability and robustness.
• Decentralized federated learning (DFL) enhances FL by enabling peer-to-peer model exchange among edge servers.
• Deploying DFL in the Internet of Things (IoT) is hindered by limited wireless resources and the need for efficient client scheduling.

Purpose of the Study:

• To address the challenge of client scheduling and resource optimization in DFL for IoT environments without prior client information.
• To develop an online learning algorithm capable of accurately estimating client participation delays despite heterogeneous resources and time-varying wireless channels.
• To improve the convergence rate and model accuracy in DFL systems by optimizing client selection.

Main Methods:

• Reformulated the client scheduling and resource optimization problem as a multi-armed bandit (MAB) program.
• Proposed an online learning algorithm employing contextual multi-arm slot machines for delay estimation and client scheduling.
• Conducted theoretical analysis to establish the asymptotic optimality of the proposed algorithm.

Main Results:

• The proposed algorithm achieves asymptotic optimal performance in theoretical analysis.
• Experimental results demonstrate the algorithm's capability for asymptotic optimal client selection.
• The method significantly outperforms existing algorithms in reducing the cumulative system delay.

Conclusions:

• The developed online learning algorithm effectively tackles client scheduling and resource optimization in DFL for IoT.
• Accurate client delay estimation and scheduling are crucial for efficient DFL deployment in resource-constrained environments.
• This approach offers a superior solution for enhancing DFL performance by minimizing system-wide delays.