Machine learning-driven integration of terrestrial and non-terrestrial networks for enhanced 6G connectivity

Abstract

Non-terrestrial networks (NTN)s are essential for achieving the persistent connectivity goal of sixth-generation networks, especially in areas lacking terrestrial infrastructure. However, integrating NTNs with terrestrial networks presents several challenges. The dynamic and complex nature of NTN communication scenarios makes traditional model-based approaches for resource allocation and parameter optimization computationally intensive and often impractical. Machine learning (ML)-based solutions are critical here because they can efficiently identify patterns in dynamic, multi-dimensional data, offering enhanced performance with reduced complexity. ML algorithms are categorized based on learning style—supervised, unsupervised, and reinforcement learning—and architecture, including centralized, decentralized, and distributed ML. Each approach has advantages and limitations in different contexts, making it crucial to select the most suitable ML strategy for each specific scenario in the integration of terrestrial and non-terrestrial networks (TNTN)s. This paper reviews the integration architectures of TNTNs as outlined in the 3rd Generation Partnership Project, examines ML-based existing work, and discusses suitable ML learning styles and architectures for various TNTN scenarios. Subsequently, it delves into the capabilities and challenges of different ML approaches through a case study in a specific scenario.