Federated Learning Optimization Algorithm for Automatic Weight Optimal.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Federated learning (FL) offers privacy-preserving distributed machine learning but suffers from system and statistical heterogeneity.
• Heterogeneity in FL systems hinders global model performance and efficiency.

Purpose of the Study:

• To address resource allocation and efficiency challenges in federated learning.
• To develop an optimization algorithm that mitigates the impact of heterogeneity on global model quality.

Main Methods:

• Introduced the FedAwo algorithm, combining adaptive and federated learning for optimal client weight calculation.
• Proposed the FedAwo+ algorithm with adaptive local convergence criteria to reduce client training costs.
• Evaluated algorithms on MNIST and Fashion-MNIST datasets against baseline methods.

Main Results:

• FedAwo and FedAwo+ significantly minimize the detrimental effects of statistical and system heterogeneity.
• FedAwo+ dynamically adjusts client epochs based on local convergence, reducing redundant computations.
• Both FedAwo and FedAwo+ demonstrated faster convergence and higher accuracy compared to FedAvg, FedProx, and FedAdp.

Conclusions:

• FedAwo and FedAwo+ effectively enhance federated learning efficiency and global model performance.
• Adaptive local convergence criteria are crucial for optimizing training costs in heterogeneous FL environments.
• The proposed algorithms offer a promising solution for improving federated learning in real-world applications.