基于时空特征提取的智能网络切片算法  

Intelligent Network Slicing Algorithm Based on Temporal-Spatial Feature Extraction

作　　者：肖柏狄 李荣鹏[1] 赵志峰[1,2] 张宏纲 XIAO Baidi;LI Rongpeng;ZHAO Zhifeng;ZHANG Honggang(College of Information Science&Electronic Engineering,Zhejiang University,Hangzhou 310027,China;Zhejiang Lab,Hangzhou 311121,China)

机构地区：[1]浙江大学信息与电子工程学院,浙江杭州310027 [2]之江实验室,浙江杭州311121

出　　处：《无线电通信技术》2022年第1期74-80,共7页Radio Communications Technology

基　　金：国家自然科学基金(61731002,62071425);浙江省重点研发计划(2019C01002,2019C03131);华为合作项目;之江实验室开放课题(2019LC0AB01);浙江省自然科学基金(LY20F010016);工业和信息化部产业技术基础公共服务平台项目(2019-00891-2-1)。

摘　　要：5G网络作为新一代的移动通信网络,提供了差异多样的各类业务服务,而通过网络切片可以将5G网络的资源虚拟化并更为有效地分配给各类服务。然而,当用户的服务需求发生变化时,需要及时地进行切片资源的重新分配和管理。为了准确适配业务需求的时空变化,提升切片性能,考虑采用深度强化学习的人工智能算法对切片资源进行实时管理,提出了基于时空特征提取的智能网络切片算法。算法采用了图注意力网络(GAT)以及长短期记忆网络(LSTM)进行数据的预处理,并使用深度Q网络(DQN)进行决策。在多基站和多用户的无线接入网场景下的仿真测试中,该算法在性能上相比于传统的网络切片算法实现了较大的提升。As a new generation of communication network,5G network supports diversified services.Using network slicing,network resource is able to be virtualized and allocated to these services more efficiently.However,when users’requirements for services change,it’s important to reallocate and manage the resource of slices promptly.Aimed at the adaption to the temporal and spatial variability of services for enhancement of slice management,deep reinforcement learning,a kind of artificial intelligence algorithm,is adopted to manage the resource of slices in real time thus the intelligent network slicing algorithm based on temporal-spatial feature extraction is proposed.The algorithm uses Graph Attention Network(GAT)and Long Short-Term Memory(LSTM)for pre-process.Besides,it takes Deep Q-Network(DQN)for decision-making.Considering a radio access network scenario that contains several network slices over multiple base stations with several users,the superiority of this algorithm compared to conventional methods is verified through extensive simulations.

关 键 词：网络切片 深度强化学习 时空特征提取 资源分配 

分 类 号：TN929.5[电子电信—通信与信息系统]