摘要
【目的】随着近海区域业务量的快速增长,带宽需求呈现出指数级增长趋势,承载业务的第五代移动通信技术(5G)及后5G(B5G)下一代无线接入网络(NG-RAN)资源即将枯竭,波分复用无源光网络(WDM-PON)因其高带宽等优势,成为支持5G/B5G NG-RAN的有效承载方案。然而,近海复杂多变的环境为WDM-PON的网络部署带来了严峻挑战,如高昂的部署成本、大规模的路径损耗以及恶劣的水下环境等,亟需通过优化网络部署策略降低成本、风险和传输损耗等,以构建适应近海环境的网络。【方法】文章提出了一种多头注意力增强的多智能体深度Q网络(MHA-MAD)算法,通过多头注意力机制高效提取网络环境的关键特征,并为不同特征赋予动态权重,从而提升建模精度。同时,采用多智能体框架,使多个智能体在共享网络环境中协作与同步决策,实现网络部署的全局优化。【结果】与现有基准方法相比,MHA-MAD算法在网络部署中使性能提高了近42%,其结果接近理论最优解。此外,与未利用多头注意力机制的多智能体深度Q网络(DQN)算法相比,MHAMAD算法在最小化网络部署总成本、节点功耗、链路衰减和网络风险的联合优化目标上,性能提高了近8%。【结论】MHAMAD算法为面向近海场景5G/B5G NG-RAN的WDM-PON部署与优化提供了新思路。
【Objective】With the rapid growth of business volume in the nearshore area,the demand for bandwidth is showing an exponential growth trend.The resources of the 5th Generation Mobile Communication Technology(5G)and Beyond⁃5G(B5G)Next Generation Radio Access Network(NG⁃RAN)that carry the business are about to be exhausted.Wavelength Division Mul⁃tiplexing⁃Passive Optical Network(WDM⁃PON),with its advantages such as high bandwidth,has become an effective solution to support 5G/B5G NG⁃RAN.However,the complex and variable offshore environment poses severe challenges for the deploy⁃ment of WDM⁃PON networks.These challenges include high deployment costs,substantial path losses,and harsh underwater conditions.There is an urgent need to optimize network deployment strategies to reduce costs,risks,and transmission losses in order to build a network that is suitable for the offshore environment.【Methods】This study proposes a Multi⁃Head Attention en⁃hanced Multi⁃Agent Deep Q⁃Network(MHA⁃MAD)algorithm.It efficiently extracts key features of the network environ⁃ment using multi⁃head attention mechanism and assigns dynamic weights to different features,thereby improving modeling accura⁃cy.Simultaneously,the multi⁃agent structure allows multiple agents to collaborate and make synchronized decisions within a shared network environment,promoting global optimization in network deployment.【Results】Compared to other benchmarks,the MHA⁃MAD algorithm improves performance in network deployment by nearly 42%,with results approaching the theoretical optimum.Furthermore,compared to multi⁃agent Deep Q⁃Network(DQN)method without the multi⁃head attention,MHA⁃MAD algorithm improves the performance by nearly 8%in the joint optimization objective of minimizing overall network deployment costs,node power consumption,link attenuation,and network deployment risk probabilities.【Conclusion】MHA⁃MAD provides new insights for the deployment and optimization of WDM⁃PON to support 5G/B5G NG⁃RAN in offshore scenarios.
作者
李学华
郗童
王鑫
黄翔
LI Xuehua;XI Tong;WANG Xin;HUANG Xiang(Institute of Intelligent Communications and Computing,Beijing Information Science and Technology University,Beijing 102206,China;National Key Laboratory of Information Systems Engineering,National University of Defense Technology,Changsha 410073,China)
出处
《光通信研究》
北大核心
2025年第4期79-85,共7页
Study on Optical Communications
基金
北京市自然科学基金—海淀原始创新联合基金资助项目(L222004)
北京市教委科技计划一般资助项目(KM202311232012)
江苏省自然科学基金资助项目(BK20232028)。
关键词
网络部署
近海网络
无线和光纤接入网络
深度Q网络
多头注意力机制
network deployment
offshore networks
wireless and optical access networks
DQN
multi⁃head attention mecha⁃nism
