Efficient Communications for Federated Learning in Satellite Mega-Constellations

Tutor: Bho Matthiesen, Nasrin Razmi
Type of Thesis: Master's thesis (MSc)
date of issue: 06/2024
Student: Martin Ansch├╝tz
Status: in progress

Modern mega-constellations in low Earth orbit (LEO) rely increasingly on inter-satellite links (ISL) and multi-hop routing. These ISLs come in different flavors and are commonly categorized, in ascending order of implementation complexity, into intra-orbit links, inter-orbit links to nearby satellites, and inter-orbit links across crossing orbital planes [1]. In [2], it is shown that, even when limited to intra-orbit ISLs, the performance of satellite federated learning (SFL) benefits tremendously from employing ISLs to connect to the parameter server (PS) through multi-hop routing. Leveraging on the predictability of satellite movement and the unique properties of data traffic for federated learning (FL), this gain comes at no additional communication cost over SFL without ISLs.

The goal of this thesis is to evaluate the potential gain of employing inter-orbit ISLs for SFL. This involves two primary tasks:

  1.     Develop an event-based simulation framework for SFL over inter-orbit ISLs.
  2.     Extend predictive routing procedures from [2] towards taking spatio-temporal network dynamics into account.

Task 1 requires a comprehensive literature review on relevant physical layer technologies for ISLs, condensing these insights into a mathematical channel model suitable for large-scale system level simulations, and implementing it in ns-3. In combination with publicly available ns-3 extensions for satellite communications, the final simulation framework is developed. Task 2 builds upon predictive routing algorithms in SFL [2]. Steps towards completeling Task 2 include extending, implementing, and evaluating these algorithms within the simulation framework from Task 1.

[1]. A. U. Chaudhry and H. Yanikomeroglu, "Laser Intersatellite Links in a Starlink Constellation: A Classification and Analysis," in IEEE Vehicular Technology Magazine, vol. 16, no. 2, pp. 48-56, June 2021.

[2]. N. Razmi, B. Matthiesen, A. Dekorsy and P. Popovski, "On-board Federated Learning for Satellite Clusters with Inter-Satellite Links," in IEEE Transactions on Communications.

Last change on 06.06.2024 by N. Razmi
AIT ieee GOC tzi ith Fachbereich 1
© Department of Communications Engineering - University of BremenImprint / Contact