TinyFDRL-Enhanced Energy-Efficient Trajectory Design for Integrated Space–Air–Ground Networks

Space–air–ground integrated networks (SAGINs) hold immense potential for improved network coverage and dynamic service delivery. Yet, current methods often depend on separate, uncoordinated unmanned aerial vehicles (UAVs), leading to scalability issues and limited energy efficiency—challenges that persist even when applying intelligent machine learning (ML) methods. This article discusses a four-tier aerial computing (AC) system, leveraging the collective capabilities of low-altitude UAVs (LAUs), high-altitude UAVs (HAUs), and satellites to fully realize the potential of SAGINs within AC. Incorporating advancements in tiny machine learning (TinyML), this system boosts onboard intelligence for immediate data processing and adaptive decision making. Specifically, by utilizing the robust computational resources of higher layer SAGIN entities, we introduce a tiny federated deep reinforcement learning (TinyFDRL) algorithm across multiple tiers to achieve energy-efficient trajectories for multiple LAUs. This proposed TinyFDRL algorithm independently plans multi-LAU trajectories in unpredictable environments by combining the strengths of federated learning (FL) and deep reinforcement learning (DRL). Extensive simulations validate the algorithm, confirming its efficiency in creating energy-saving paths for LAUs in the integrated AC network.