The paper introduces a novel FEEL algorithm, FEDMEGA, designed for LEO satellite mega-constellation networks. The key highlights are:
Leveraging the high-speed and stable inter-satellite links (ISLs) within each orbit, the proposed FEDMEGA algorithm performs multiple intra-orbit model aggregation rounds per global round, significantly reducing the usage of low-data-rate and intermittent ground-to-satellite links (GSLs).
For efficient intra-orbit model aggregation, FEDMEGA utilizes the ring topology of satellites within each orbit and the full-duplex capability of laser ISLs to propose a ring all-reduce based scheme, which ensures fast convergence of the intra-orbit model.
To accelerate the global model aggregation, FEDMEGA employs a network flow-based transmission scheme that maximizes the amount of model parameters received by the ground parameter server (PS) in each time slot, thereby minimizing the overall transmission latency.
Theoretical convergence analysis is provided, demonstrating that FEDMEGA achieves linear speedup in terms of the number of local updates, the number of LEO satellites, and the number of intra-orbit aggregations, under non-convex settings and non-IID data distribution.
Extensive simulations on both synthetic and real datasets show that FEDMEGA outperforms existing satellite FEEL algorithms, exhibiting an approximate 30% improvement in convergence rate.
To Another Language
from source content
arxiv.org
Önemli Bilgiler Şuradan Elde Edildi
by Yuanming Shi... : arxiv.org 04-03-2024
https://arxiv.org/pdf/2404.01875.pdfDaha Derin Sorular