UCB-VG: Virtual-Grid-Based Decentralized Learning for Link Management in LEO Networks

Abstract

Unlike traditional bent-pipe architectures, modern LEO systems employ inter-satellite links (ISLs) to support multi-hop satellite-to-satellite forwarding. These mega-constellations are a key candidate for 6G non-terrestrial networks (NTN), but their rapidly time-varying topology and highly non-uniform, time-dependent traffic demand make efficient link management challenging, often leading to long routes and congestion. We propose UCB-VG, a fully decentralized learning-based scheme that combines an Upper Confidence Bound (UCB) criterion with a Virtual Grid of fixed virtual nodes that satellites temporarily embody. This abstraction enables autonomous link-selection decisions without global topology knowledge and without assuming static ISLs. Using a realistic global simulation based on a Starlink-scale constellation and population-driven traffic, we show that UCB-VG reduces end-to-end penalties by up to 60%, limits delay growth under increasing load, and significantly lowers packet drop rates compared to baseline schemes.