Joint Device Participation, Dataset Management, and Resource Allocation in Wireless Federated Learning via Deep Reinforcement Learning

Federated Learning (FL) enables large-scale machine learning without uploading the private data of wireless devices. Due to the heterogeneity and limitation of the devices' resources, the FL accuracy and latency substantially depend on the device participation and training dataset size. In this...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on vehicular technology 2024-03, Vol.73 (3), p.4505-4510
Main Authors: Chen, Jinlian, Zhang, Jun, Zhao, Nan, Pei, Yiyang, Liang, Ying-Chang, Niyato, Dusit
Format: Article
Language:English
Subjects:
Computational modeling
Convexity
Dataset management
Datasets
Deep learning
deep reinforcement learning
device participation
Energy consumption
Federated learning
Heterogeneity
Machine learning
Markov processes
Multiagent systems
Network latency
Optimization
Resource allocation
Resource management
Task analysis
Training
Wireless communication
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Federated Learning (FL) enables large-scale machine learning without uploading the private data of wireless devices. Due to the heterogeneity and limitation of the devices' resources, the FL accuracy and latency substantially depend on the device participation and training dataset size. In this letter, to strike a trade-off between the FL accuracy and FL latency, a joint device participation, dataset management and resource allocation (DPDMRA) optimization problem is investigated. To solve the non-convex optimization problem, a Markov decision process is formulated for the resource-limited wireless FL. Moreover, due to the high dimensional continuous action space, a multi-agent softmax deep double deterministic policy gradients (MASD3) method is employed to obtain the optimal DPDMRA strategies. The double actor networks and softmax operator are designed to alleviate the underestimation bias. Simulation results demonstrate that the proposed DRL method can obtain the global optimal policy without complete information in the dynamic environment. Compared with the other baseline schemes, the proposed MASD3 approach can achieve the larger system utility with the better convergence performance.
ISSN:0018-9545
1939-9359
DOI:10.1109/TVT.2023.3325843