【作者】 赵晨；

【导师】 黄永明；

【作者基本信息】 东南大学 ， 信号与信息处理， 2023， 硕士

【摘要】 近来,随着5G/6G技术的飞速发展,如何在未来的移动网络中,结合软件定义网络(Software Defined Network,SDN)和网络功能虚拟化(Network Function Virtualization,NFV)技术,采用网络切片的方式来满足网络中多样化业务的差异化服务需求,已成为热点研究问题。网络切片本身按需定制、灵活可扩展和提供特定服务性能的特点,对无线资源管理的灵活性提出了较高的要求。网络切片主要包含无线接入网(Radio Access Network,RAN)切片、传输网切片和核心网切片,本文关注的是RAN侧切片及其资源分配问题。为优化网络资源配置,保障业务服务等级协议(Service Level Agreements,SLA)需求,使不同应用间差异化的服务质量需求同时得到良好满足,动态无线资源配置技术起到至关重要的作用。因此,针对网络流量波动、网络状态快速变化和用户需求多样化的场景,本文提出两种通过灵活动态配置接入网资源以提高系统效用并保障业务服务质量的无线资源配置方案:(1)基于流量感知的动态无线网络切片资源分配方案和(2)面向用户级QoS保障的无线网络切片资源分配方案。针对网络流量波动、网络状态快速变化的场景,本文首先在第三章提出一种结合时间序列预测和深度强化学习的智能化带宽分配策略,使用长短期记忆(Long Short-Term Memory,LSTM)和Dueling深度Q网络(Dueling Deep Q Network,Dueling DQN)以最大限度提高RAN切片的频谱效率和SLA满意度。通过使用LSTM对网络切片中用户到达和数据流量进行预测,可以提高传统强化学习方法在网络切片资源分配问题中的及时性,有效将深度强化学习算法(Deep Reinforcement Learning,DRL)的计算周期与实际切片配置周期解耦。同时,为了在保障LSTM性能的同时降低其计算复杂度,以适应RAN中有限的计算资源,本文采用控制神经元连接比例的随机连接LSTM(Randomly Connected LSTM,RCLSTM)神经网络。仿真结果表明10%连接比例的RCLSTM网络能够将原始LSTM的计算时间降低约11%,同时达到比传统多层感知机(Multi-Layer Perceptron,MLP)、自回归整合移动平均(Autoregressive Integrated Moving Average,ARIMA)、支持向量回归(Support Vector Regression,SVR)预测方法低约30%的误差。此外,Dueling DQN模型与传统深度Q网络(DQN)相比,能够提高Q值估计精确度。仿真结果显示,与原始DQN、优势Actor-Critic(Advantage Actor Critic,A2C)和硬切片方法相比,RCLSTMDueling DQN结合方案可以通过提前感知网络性能变化、获取历史流量数据中的隐藏模式,有效降低网络环境波动对密集流量场景下无线切片资源管理的影响,达到更快的收敛速度、更高的频谱效率和几乎100%的切片SLA满意率。现有RAN切片资源分配技术通常考虑保证整个切片的性能,而忽略了切片内所有用户的需求。因此本文在第四章提出一种面向用户流量层面QoS需求的动态资源分配算法。针对具有频率选择性和时间选择性的无线信道,分析了考虑用户级需求对RAN切片资源分配算法设计的重要性。由此构建在频率选择性无线场景中,以保障各个切片内用户级QoS需求为前提的最大化系统吞吐率优化问题,提出了一种基于Lyapunov优化并引入用户最大资源块(Resource Block,RB)关联限制约束的近似最优解决方案。仿真实验表明所提出的算法能够实现优化目标,并证明了该算法在网络维度增加时的可拓展性及算法收敛时间方面的可行性。同时,与现有的面向切片层面和以资源请求为依据进行资源分配的方案相比,本文所提出的方法在保障用户QoS和最大化系统吞吐率方面具有明显优势。在相同的网络环境下,与现有方案中只有近60%的用户能够达到最小吞吐率需求的结果相比,本方案能够提供100%的用户QoS保障并提高系统吞吐率。更多还原

【Abstract】 Recently,with the rapid development of 5G/6G technologies,how to combine Software Defined Network(SDN)and Network Function Virtualization(NFV)technologies in the future mobile network,and adopt network slicing to meet different service demands of diverse services in the network has become a hot research issue.The characteristics of network slicing itself,which is on-demand,flexible and scalable,and provides specific service performance,impose high requirements on the flexibility of wireless resource management.Network slicing mainly includes Radio Access Network(RAN)slicing,transmission network slicing and core network slicing,and this paper focuses on RAN side slicing and its resource allocation.In order to guarantee the service level agreement(SLA)requirements,dynamic wireless resource allocation technology plays a crucial role,which should optimize the network resource allocation so that different service quality requirements of different applications can be well satisfied at the same time.Therefore,for scenarios of fluctuating network traffic,fast-changing network conditions and diverse user requirements,this paper proposes two wireless resource allocation schemes to improve system utility and guarantee service quality of service by flexibly and dynamically allocating access network resources:(1)a traffic-aware dynamic wireless network slice resource allocation scheme and(2)a user-level QoS-guaranteed wireless network slice resource allocation scheme.For the scenario of network traffic fluctuation and rapid network state change,this paper first proposes an intelligent bandwidth allocation strategy that combines time series prediction and deep reinforcement learning,using Long Short-Term Memory(LSTM)and Dueling Deep Q Network(Dueling DQN)to maximise the spectral efficiency and SLA satisfaction of RAN slices in Chapter 3.By using LSTM to predict user arrivals and traffic in network slices,the timeliness of traditional reinforcement learning methods in the network slice resource allocation problem can be improved,and the computation cycle of Deep Reinforcement Learning(DRL)algorithms can be effectively decoupled from the actual slice allocation cycle.Meanwhile,in order to reduce the computational complexity of LSTM while guaranteeing its performance to fit the limited computational resources in RAN,this paper adopts Randomly Connected LSTM(RCLSTM)neural network with controlled neuron connection ratio.The simulation results show that the 10% connected RCLSTM network is able to reduce the computation time of the original LSTM by about 11% while achieving a lower error than the traditional Multi-Layer Perceptron(MLP),Autoregressive Integrated Moving Average(ARIMA)and Support Vector Regression(SVR)prediction methods by about 30%.In addition,the Dueling DQN model is able to improve the accuracy of Q value estimation compared to the original Deep Q Networks(DQN).Simulation results show that the combined RCLSTM-Dueling DQN solution can effectively reduce the impact of network environment fluctuations on dense traffic scenarios by detecting network performance changes in advance and capturing hidden patterns in historical traffic data,compared to traditional DQN,Advantage Actor Critic(A2C)and hard slicing methods.The combined RCLSTM-Dueling DQN solution can effectively reduce the impact of network environment fluctuations on wireless slicing resource management in dense traffic scenarios by detecting network performance changes in advance,capturing hidden patterns in historical traffic data,achieving faster convergence,higher spectral efficiency and almost 100%SLA satisfaction.Existing RAN slice resource allocation techniques usually consider the guaranteed performance of the entire slice while ignoring the needs of all users within the slice.Therefore,this paper proposes a dynamic resource allocation algorithm for user traffic-level QoS requirements in Chapter 4.The importance of considering user-level requirements in the design of RAN slice resource allocation algorithms is analysed for frequency-selective and time-selective wireless channels.This leads to the formulation of an optimization problem for maximising system throughput in a frequency-selective wireless scenario with the premise of guaranteeing the QoS requirements at the user level within each slice,and proposes an approximate optimal solution based on Lyapunov optimization with the introduction of a constraint on user maximum Resource Block(RB)association.Simulation experiments show that the proposed algorithm can achieve the optimisation objective and demonstrate the scalability of the algorithm with increasing network dimension and the feasibility of the convergence time of the algorithm.At the same time,the proposed method has significant advantages over existing slice oriented and resourcerequest based resource allocation schemes in terms of guaranteeing user QoS and maximising system throughput.In the same network environment,this scheme can provide 100% of user QoS guarantee and improve the system throughput rate compared to the existing schemes where only nearly 60% of users can meet the minimum throughput rate requirement.更多还原