Device Scheduling for Energy-Efficient Federated Learning over Wireless Network Based on TDMA Mode
| Autor: | Hejiao Huang, Nuo Yiu, Youqiang Hu |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Schedule
Wireless network Computer science business.industry 05 social sciences Time division multiple access 050801 communication & media studies 020206 networking & telecommunications Cloud computing 02 engineering and technology Energy consumption Scheduling (computing) Upload 0508 media and communications Server 0202 electrical engineering electronic engineering information engineering business Mobile device Edge computing Computer network Efficient energy use |
| Zdroj: | WCSP |
| DOI: | 10.1109/wcsp49889.2020.9299815 |
| Popis: | With the help of edge computing, Federated Learning (FL) has become a promising machine learning model training paradigm. It enables users to train model on their devices without uploading private data to the cloud servers for centralized training, thus avoiding security risks. In FL, the edge server, served as the training organizer, schedules devices in its coverage to join in a model training activity. The devices receive model parameters from edge server and update this parameters based on its local dataset. However, this training paradigm may cause large energy consumption, which affects the endurance of the mobile devices. In order to optimize the energy consumption in the FL training, we formulate it as a nonlinear integer programming problem, and design a device scheduling strategy by adopting the classification method and hybrid branch-and-bound algorithm. Based on the information, such as CPU capacity and channel state of devices, a reasonable devices set is scheduled to participate in the training, which consumes lower energy with the training delay constraint obeyed. Simulation results show that our proposed strategy can reduce 43.3% of traversed nodes compared with the conventional best-first search strategy. When the number of total devices increases, it can also outperform the existing methods in terms of energy reduction rate by an average of 14.3% and 10.7% respectively. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|