Decentralized learning protocols suffer from severe communication overheads when agents possess unbalanced data allocations. We formulate a mathematically bounded communication-efficient workload balancer that reallocates computational stress across adjacent graph neighbors, delivering a 26% reduction in slow-agent compute burdens while guaranteeing consensus convergence.
@inproceedings{sajjadi2024icdcs,
title = {Communication-Efficient Training Workload Balancing for Decentralized Multi-Agent Learning},
author = {Sajjadi Mohammadabadi, Seyed Mahmoud and Yang, Lei and Yan, Feng and Zhang, Junshan},
booktitle = {2024 IEEE 44th International Conference on Distributed Computing Systems (ICDCS)},
pages = {680--691},
year = {2024},
organization = {IEEE},
doi = {10.1109/ICDCS60910.2024.00062}
}