Unsupervised Learning for User Scheduling in Multibeam Precoded GEO Satellite Systems

Future generation SatCom multibeam architectures will extensively exploit full-frequency reuse schemes together with interference management techniques, such as precoding, to dramatically increase spectral efficiency performance. Precoding is very sensitive to user scheduling, suggesting a joint precoding and user scheduling design to achieve optimal performance. However, the joint design requires solving a highly complex optimization problem which is unreasonable for practical systems. Even for suboptimal disjoint scheduling designs, the complexity is still significant. To achieve a good compromise between performance and complexity, we investigate the applicability of Machine Learning (ML) for the aforementioned problem. We propose three clustering algorithms based on Unsupervised Learning (UL) that facilitate the user scheduling decisions while maximizing the system performance in terms of throughput. Numerical simulations compare the three proposed algorithms (K-means, Hierarchical clustering, and Self-Organization) with the conventional geographic scheduling and identify the main trade-offs.