A multi-armed bandit framework for online optimization in green integrated terrestrial and non-terrestrial networks

Integrated terrestrial and non-terrestrial network (TN-NTN) architectures offer a promising solution for expanding coverage and improving capacity for the network. While nonterrestrial networks (NTNs) are primarily exploited for these specific reasons, their role in alleviating terrestrial network (TN) load and enabling energy-efficient operation has received comparatively less attention. In light of growing concerns associated with the densification of terrestrial deployments, this work aims to explore the potential of NTNs in supporting a more sustainable network. In this paper, we propose a novel online optimisation framework for integrated TN-NTN architectures, built on a multi-armed bandit (MAB) formulation and leveraging the Bandit-feedback Constrained Online Mirror Descent (BCOMD) algorithm. Our approach adaptively optimises key system parameters— including bandwidth allocation, user equipment (UE) association, and macro base station (MBS) shutdown—to balance network capacity and energy efficiency in real time. Extensive system-level simulations over a 24-hour period show that our framework significantly reduces the proportion of unsatisfied UEs during peak hours and achieves up to 19 % throughput gains and 5 % energy savings in low-traffic periods, outperforming

standard network settings following 3GPP recommendations.