模型驅動之混合深度強化學習於聯合運算與 通訊於多功能 RIS 輔助之星空地整合網路;Twin Model-Driven Hybrid Deep Reinforcement Learning for Joint Computing and Communications in Multi-Functional RIS-Aided Space-Air-Ground Integrated Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/98069

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98069

Title:	模型驅動之混合深度強化學習於聯合運算與通訊於多功能 RIS 輔助之星空地整合網路;Twin Model-Driven Hybrid Deep Reinforcement Learning for Joint Computing and Communications in Multi-Functional RIS-Aided Space-Air-Ground Integrated Networks
Authors:	黃雋喆;Jyun-Jhe Huang
Contributors:	通訊工程學系
Keywords:	多功能重構智慧表面;星空地一體化網路;資源管理;能源效率;深度強化學習;Multi-functional reconfigurable intelligent surface;space-air-ground inte grated Network;Resource Management;Energy efficiency;Deep Reinforcement Learning
Date:	2025-06-30
Issue Date:	2025-10-17 12:19:19 (UTC+8)
Publisher:	國立中央大學
Abstract:	在本研究中，我們提出一種創新的網路架構，將多功能可重構智慧表面（MF-RIS）部署於星空地網路（SAGIN）中。與僅具備訊號反射能力的傳統 RIS 不同，MF-RIS 能夠反射、折射、放大訊號，並從無線訊號中收集能量。此架構旨在解決低地球軌道（LEO）衛星在陰影區域無法獲得太陽能時的高能耗問題，同時考慮 SAGIN 節點在計算和通訊過程中的能量消耗。透過在每個 SAGIN 節點部署 MF-RIS，可增強通道多樣性、訊號強度，並提升整體能源效率（EE）。為了實現上述目標，我們將長期能源效率最大化問題建模為一個最佳化問題，決定最佳的 MF-RIS 配置，包括放大和相位偏移參數、能量收集比例以及主動元件的數量，同時優化 LEO、高空平台（HAPS）和基地台（BS）節點的參數，如傳輸波束成形、HAPS 部署、連線和 CPU 運算能力。為了解決該問題高維度的離散與連續優化變數，以及其複雜的非凸和非線性特性，我們設計了一種雙模型增強的多代理混合深度壓縮強化學習（Twin-Model Enhanced Multi-Agent Hybrid Deep Compressed Reinforcement Learning）方案。該方法結合了壓縮技術以降低計算複雜度，並採用專門設計的混合深度強化學習策略，使每個代理能夠在連續與離散決策空間中進行最優互動。數值結果顯示，與其他基準方法（如集中式深度強化學習和分散式多代理深度確定性策略梯度（DDPG））相比，該方法在能源效率方面有顯著提升。此外，所提出的 SAGIN-MF-RIS 架構在能源效率表現上優於固定/無能量收集的 MF-RIS、僅具反射功能的傳統 RIS，以及未部署 RIS/MF-RIS 的情境，證明了其有效性。;In this thesis, we propose a novel space-air-ground integrated network (SAGIN) architecture enhanced by a multi-functional reconfigurable intelligent surface (MF-RIS). MF-RIS supports signal reflection, refraction, amplification, and wireless energy har vesting. This design mitigates energy shortages for LEO satellites in shadowed regions and considers energy consumption from both communication and computation at each SAGIN node. We formulate a long-term EE maximization problem that jointly opti mizes MF-RIS parameters (amplification, phase shift, harvesting ratio, and active ele ments) and SAGIN configurations (beamforming, HAPS deployment, user association, and computing cycles). To tackle the mixed discrete-continuous, non-convex optimiza tion challenge, we propose a twin-model enhanced multi-agent hybrid deep compressed Reinforcement Learning framework. This framework leverages model compression and a hybrid RL strategy to efficiently explore large action spaces. Simulation results show our method outperforms centralized and multi-agent deep deterministic policy gradient (DDPG) baselines in EE and outperforms fixed or no energy harvesting MF-RIS, tradi tional RIS, and no-RIS setups.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	17	View/Open

社群 sharing

Loading...