隨著衛星通訊與天地一體化網路（Space-Air-Ground Integrated Network, SAGIN）的快速發展，異質節點間的高效協作與能量最佳化仍然充滿挑戰，特別是在低軌衛星（LEO）、高空平台（HAPs）以及地面基站（BSs）組成的多層架構中，大規模部署、不均勻分布以及動態鏈路條件使得資源分配與學習穩定性變得更加複雜。為解決這些問題，本研究提出一種結合水平與垂直聯邦學習的階層式混合深度強化學習框架，用於跨層自適應資源管理。具體而言，基於深度確定性策略梯度（DDPG）的動態分群機制能依據位置、訊噪比以及傳輸條件選擇聯邦節點；同時，基於多智能體 DDPG（MADDPG）的策略則能在動態環境下聯合最佳化頻寬、功率與運算頻率，以提升能量效率。模擬結果顯示，所提方法在能量效率、收斂速度與延遲方面均優於傳統基準方法，進而能在高度動態的 SAGIN 環境中確保穩定的學習與高效的資源利用。;With the rapid development of satellite communications and Space-Air-Ground Integrated Networks (SAGIN), efficient collaboration and energy optimization across heterogeneous nodes remain challenging due to large-scale deployment, uneven distribution, and dynamic link conditions among Low Earth Orbit (LEO) satellites, High-Altitude Platforms (HAPs), and Base Stations (BSs). To address these challenges, this study proposes a hierarchical federated hybrid deep reinforcement learning framework that integrates horizontal and vertical federated learning for adaptive cross-layer resource management. Specifically, a Deep Deterministic Policy Gradient (DDPG)-based dynamic clustering mechanism selects federated nodes according to location, signal-to-noise ratio, and transmission conditions, while a Multi-Agent DDPG (MADDPG)-based strategy jointly optimizes bandwidth, power, and computing frequency to improve energy efficiency under dynamic environments. Simulation results demonstrate that the proposed method outperforms conventional baselines in terms of energy efficiency, convergence speed, and latency, thereby ensuring stable learning and efficient resource utilization in highly dynamic SAGIN scenarios.
