Learning-Based Reactive Handover under Line-of-Sight Loss: A UE-Centric NTN Continuity Model

The dynamics of Low Earth Orbit (LEO) constellations introduce frequent and unpredictable Line-of-Sight (LoS) interruptions that challenge conventional handover procedures in Non-Terrestrial Networks (NTNs). Previous analytical work has established that, once the residual LoS interval becomes shorter than the end-to-end control-plane delay, network-assisted preparation loses feasibility and the decision logic must shift toward the terminal. Building on this premise, this paper introduces a learning-based reactive handover model representing a UE-centric continuity mechanism under LoS-limited conditions. The proposed framework employs a lightweight Q-learning agent that selects target satellites based on locally observable indicators such as relative elevation and received signal level. Within a deterministic geometric environment, the agent converges toward stable switching behaviour and reproduces the qualitative trade-offs predicted by the analytical model of reactive continuity. The achieved 25-30% alignment improvement over a geometry-only baseline demonstrates that even simplified reinforcement learning can capture timing dependencies essential to NTN handover robustness. While not intended as a deployment result, the study provides a reproducible benchmark and a methodological foundation for future deep- or federated-learning extensions operating within realistic NTN channel and protocol constraints.