Reinforcement Learning-Based Robust Resource Scheduling for Dynamic LEO Satellite Networks

INTRODUCTION: Low Earth Orbit (LEO) satellite communication extends Narrowband Internet of Things (NB-IoT) coverage for global 6G IoT services. However, long propagation delays and high mobility cause outdated channel state information (CSI), degrading system performance.                                                                                                                      OBJECTIVES: This study aims to design a robust resource scheduling strategy that mitigates CSI outdatedness, improves resource utilization, and supports large-scale IoT connectivity in dynamic LEO satellite environments.                                   METHODS: We propose a reinforcement learning–based robust scheduling framework. The Chebyshev inequality transforms probabilistic signal-to-noise ratio (SNR) constraints into deterministic bounds using statistical moments. A multi-objective optimization problem is formulated to maximize served terminals and minimize resource fragmentation. The Proximal Policy Optimization (PPO) algorithm enables intelligent allocation across network slices under dynamic conditions.                                                                                                                                                                                     RESULTS: Simulation results demonstrate that the proposed approach achieves higher scheduling success rates and better resource utilization compared with baseline methods. The reinforcement learning agent adapts effectively to environmental variations, maintaining stable performance even under severe CSI outdatedness.                                                                       CONCLUSION: The robust reinforcement learning–based scheduling provides an effective solution for NB-IoT over LEO satellites, enhancing reliability and scalability of future 6G global IoT networks.