Scheduling and Power Control in Sustainable Federated Learning over Wireless Networks

Federated Learning (FL) was proposed in 2016 to address privacy concerns associated with potential data leaks in traditional centralized machine learning (ML) frameworks. FL operates by enabling multiple local devices to jointly train an ML model through the iterative exchange of model parameters among the devices and a central parameter server (PS). Specifically, each local device performs local training on its own dataset for multiple rounds and transmits model updates to the PS. The PS then aggregates these updates and distributes a new global model for the next iteration. This training cycle continues until satisfactory accuracy is achieved. Since private data never leaves local devices, FL is considered one of the key approaches for privacy-preserving, distributed ML systems. Despite its great potential to train global models in a distributed setting, FL encounters specific challenges when implemented over wireless networks, such as device heterogeneity, extensive communication overhead and energy consumption due to the frequent exchange of models. In addition, leveraging the energy harvesting capabilities of IoT devices to optimize FL performance in wireless networks has not been extensively explored. In this thesis, we aim to engineer a sustainable federated learning framework by applying device scheduling and power control with energy harvesting. Our solutions will be extensively validated through real-world experiments. Specifically, we propose POLISH, a power control and device scheduling scheme designed to minimize the overall communication delay and battery energy consumption while maximizing the total number of scheduled devices in each global iteration. POLISH applies a non-linear optimization technique to simultaneously minimize communication delay and energy consumption, followed by a device scheduling algorithm that selects as many devices as possible while satisfying time and energy constraints. In our second approach, we consider the stochastic nature of FL performance in wireless networks and jointly examine the effect of scheduling probability for each device on convergence speed and communication overhead. We propose Sto-POLISH, a stochastic power control and device scheduling scheme that dynamically adjusts scheduling probability and transmission power across all global iterations. The optimization scheme is designed to meet three long-term constraints: fairness, average power consumption, and minimum battery energy storage.

Department dc:contributor.department

Electrical and Computer Engineering

Year dc:date.issued

2024

Author dc:creator

• Zhang, Boyuan

Advisor dc:contributor.supervisor

• Lu, Ning

• Machine Learning
• Federated Learning
• Power Control
• Energy Harvesting

Statement dc:rights

• Attribution-NonCommercial-NoDerivatives 4.0 International

Licence dc:rights.uri

• http://creativecommons.org/licenses/by-nc-nd/4.0/

Language dc:language.iso

eng

Handle dc:identifier.uri

https://hdl.handle.net/1974/33471

OAI identifier oai:identifier

oai:queensu.scholaris.ca:1974/33471