Reinforcement Learning-Based Handover for LEO Non-Terrestrial Networks

In low earth orbit (LEO) non-terrestrial networks (NTN), the high orbital speed of satellites and their limited visibility time over a given user can lead to frequent handovers, increasing signaling overhead, handover failures, outage probability, and service instability. This makes satellite selection a relevant problem for NTN mobility management, especially in dense mega-constellation scenarios. To address this problem, this paper presents a reinforcement learning-based agent for satellite selection in LEO NTN systems. This work uses CAVIAR with ns-3 NTN channel simulation to generate datasets for handover studies. With that, a deep Q-learning (DQN) agent is trained to select serving satellites while considering spectral efficiency, handover attempts, handover failures, outages, and satellite load. Results show that the DQN agent reduces handover failures without a loss of spectral efficiency performance, indicating the potential of reinforcement learning for NTN handover management.