生成模型与网络优化:回顾与展望  

Generative Models and Network Optimization:Review and Perspective

作　　者：梁芮槐 杨博 陈鹏宇 李先进 薛一凡 曹雪琳 於志文[1,3] 郭斌 LIANG Ruihuai;YANG Bo;CHEN Pengyu;LI Xianjin;XUE Yifan;CAO Xuelin;YU Zhiwen;GUO Bin(School of Computer Science,Northwestern Polytechnical University,Xi'an 710129,China;School of Cyber Engineering,Xidian University,Xi'an 710126,China;School of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]西北工业大学计算机学院,陕西西安710129 [2]西安电子科技大学网络与信息安全学院,陕西西安710126 [3]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《无线电工程》2025年第4期726-738,共13页Radio Engineering

基　　金：国家自然科学基金(61960206008);国家自然科学基金优秀青年科学基金项目(海外);秦创原引用高层次创新创业人才项目(QCYRCXM-2022-358,QCYRCXM-2022-240)。

摘　　要：生成式人工智能(Generative AI,GenAI)的出现和发展为网络领域的智能化提供了新的机遇,将其用于网络优化这一类重要的基础问题中已经取得了引人瞩目的成果。GenAI具备传统机器学习方法应对高维问题的高效性,有潜力提供更强的泛化能力。其中,2种前沿的模型——扩散模型(Generative Diffusion Model,GDM)与大语言模型(Large Language Model,LLM),体现了不同的技术优势:生成式GDM通过迭代去噪过程生成较高精度的优化解,LLM凭借可扩展的自回归架构实现优化问题的语义解析。为促进相关研究,综述了当前生成模型与网络优化的结合工作,从生成式GDM和LLM这2类最前沿的生成模型的角度总结在网络领域已有的研究工作,介绍了一些相关领域的生成模型工作。在阐述生成模型用于网络优化的各种优势和成果的同时,分析了生成模型在网络领域的当前发展阶段和关键挑战,可为生成模型与网络优化的结合研究提供一定的参考。The emergence and development of Generative AI(GenAI)have provided new opportunities for the advancement of network intelligence,and its application in critical foundational problems of network optimization has yielded remarkable achievements.GenAI not only demonstrates efficiency in tackling high-dimensional problems compared to traditional machine learning methods,but also shows potential for stronger generalization performance across multiple scenarios.Two cutting-edge paradigms,Generatire Diffusion Model(GDM)and Large Language Model(LLM),show distinct technical advantages.GDM generate high-accuracy solutions through iterative denoising,while LLM,enable semantic parsing of the original optimization problems via scalable autoregressive architectures.To advance related research,the current integration of generative models with network optimization is reviewed.Existing research in network domain is summarized from the perspectives of two cutting-edge generative model categories,i.e.GDM and LLM.Relevant works of generative models in related fields are also introduced.The various advantages and achievements of generative models in network optimization are summarized,and the current development stage and key challenges of generative models in network domain are analyzed,offering some insights for future research on integrating generative models with network optimization.

关 键 词：生成模型 扩散模型 网络优化 资源管理 通信网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] V279[自动化与计算机技术—控制科学与工程]