Joint Client Scheduling and Quantization Optimization in Energy Harvesting-Enabled Federated Learning Networks

Authors:
Publisher website
Google Scholar
Abstract
A vital challenge in the deployment of federated learning (FL) over wireless networks is the high energy consumption incurred for the local computation and model update upload on energy-constrained devices such as IoT sensors. Equipping with energy harvesting (EH) modules is a promising solution that allows the devices to work in a self-sustainable manner. Moreover, quantizing the model updates can further improve the energy efficiency during the upload. In this paper, we propose an EH-enabled FL system with model quantization in which EH devices act as clients and client scheduling, model quantization, and transmit energy are jointly optimized to minimize the training loss while satisfying energy causality constraints and guaranteeing fairness in client selection. We formulate a non-convex mixed-integer nonlinear programming (MINLP) problem for the optimization. Then, by recasting the product of a continuous variable and a 0-1 variable in an equivalent linear form, we transform this non-convex MINLP problem into a convex problem and solve it. We present numerical evaluations on various datasets to show that our proposed system is stable and achieves high performance regardless of whether the loss function is convex or non-convex and whether the data distributions are independent and identically distributed (i.i.d.) or non-i.i.d.
Funding Information
  • Zhejiang Provincial Natural Science Foundation of China (LQ22F010008)
  • National Natural Science Foundation of China (U20A20158, 61725104)
  • National Key Research and Development Program of China (2020YFB1807101, 2018YFB1801104)
  • National Research Foundation, Singapore
  • Infocomm Media Development Authority under its Future Communications Research and Development Program
  • DSO National Laboratories under the AI Singapore Program (AISG2-RP-2020-019, FCP-ASTAR-TG-2022-003)
  • MOE Tier 1 (RG87/22)
  • National Research Foundation of Korea
  • Korean government (2021R1A2C2007638)
  • National Natural Science Foundation of China (62302446, 62276231)
  • Provincial Universities Basic Research Project (3090JYN9923001G-021)

This publication has 45 references indexed in Scilit: