当前位置：和泉文库 > 计算机 > 电子科技大学：《无线互联网 Principles of Wireless Networking》课程教学资源（课件讲稿）Handout 06 DL for MEC

电子科技大学：《无线互联网 Principles of Wireless Networking》课程教学资源（课件讲稿）Handout 06 DL for MEC

文件格式：PDF，文件大小：6.54MB，售价：17.46元

文档详细内容（约64页）

系统结构 2020年秋季 11 / 65 无线互联网 132 IEEE Network • May/June 2019 because they are gathered in real time. Second, the collected data vary from diverse kinds of sensor data, variable user preference, location data, and user-content relationships. Finally, the quantity of data to be handled is likely to be very high, for example, in the order of Terabytes. To facilitate pattern matching and location queries, a high performance relational database is used to efficiently store certain data, whereas a distributed database is used to optimize simple queried data for scalability. FEATURE EXTRACTION Feature extraction is the key element of the proposed framework. Two types of features have to be extracted from collected raw data, which are described as follows: User Features: The popularity of content in proactive caching might fluctuate across a user population since different content might be preferred by different users. The content preference of a user might be associated with the variety of their features as shown in Fig. 2a. User features involve the following: • Personal characteristics of the user, such as demographic information (for example, gender and age), mood or personality. • The explicit context, which involves the extraction of the explicit context, which is defined as a process of accumulating data in which the accumulation circumstances are properly described, for example, the weather conditions. • The latent context, which involves the extraction and collection of hidden or latent contextual patterns from mobile sensors to represent the user context. The addition of various types of user features leads to the expansion of its dimensionality, which results in a processing task of large amounts of training data. Therefore, an auto-encoder is used to determine the relationships between the diverse features and extract them in low-dimensional representation. Typically, the user feature extraction procedure consists of three phases: • Raw data is accumulated from the available data sources, for example, mobile sensors such as WiFi, GPS, microphone, active applications, and accelerometer. • A set of features is extracted from the raw data by using feature engineering, which consists of computing statistics such as dominant values, entropy, standard deviation, average, and so on. • An unsupervised technique called auto-encoder is applied to the extracted features to determine hidden patterns in the raw data. An auto-encoder is an unsupervised learning method that sets the target values as equivalent to the inputs by applying back-propagation. For example, for the given input of an unlabelled training dataset {a1, a2, a3, , ai }, where ai £ Rn , the aim of the auto-encoder would be to set ai = a^i . In order to extract the long-term patterns of the users and latent context, we use a simple auto-encoder, which is based on an existing netFIGURE 1. Overview architecture of DeepCachNet framework. Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on July 01,2020 at 04:24:21 UTC from IEEE Xplore. Restrictions apply

仿真结果 2020年秋季 12 / 65 无线互联网 136 IEEE Network • May/June 2019 Then, the rest of the ratings in matrix P and their features are used as a testing set, and are predicted by using SVDFeature [15], and the most popular content is stored accordingly. Caching at the SBS: In this segment, the estimated popular content at the core network is greedily stored at the SBS until no space remains, as in [3]. The performance of the proposed framework is measured in terms of the user satisfaction ratio and the backhaul load. The user satisfaction ratio is described as the fraction of the content supplied at a given target rate, and the backhaul load denotes the fraction of the traffic carried by the backhaul links over the total possible amount of traffic generated by the content requests. The detailed procedure to determine the backhaul load and the satisfaction ratio can be found in [3]. Variations of the backhaul load and the user satisfaction ratio with respect to backhaul capacity, storage size, and traffic intensity are shown in Fig. 4 and discussed as follows. Variation in the Backhaul Capacity (I’d): The total capacity of backhaul links (II’d) is considered to be lesser than the wireless links capacity (IId). Escalation of backhaul capacity clearly contributes to growth in the satisfaction ratio in all cases, since any content is provided through backhaul that is not present in the caches of the SBS. ConFIGURE 4. DeepCachNet performance in terms of user satisfaction ratio and backhaul load with variation in the: a) backhaul capacity; b) storage ratio; c) traffic intensity. (a) (b) (c) Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on July 01,2020 at 04:24:21 UTC from IEEE Xplore. Restrictions apply. 仿真实验 6个SBS; 60个用户; 220内容. 对比了“任意缓存”[黑线]; GroundTruth[绿线]; 合作滤波[红线]; 本文的合作滤波[蓝线]. 考察指标为用户满意度[越大越好] 和Backhaul的负载[越小越好]. 分别研究了backhaul容量[a]、缓存容量[b]和业务量[c]的影响. 结论都是: 本文方法最接近 GroundTruth的效果

系统模型 2020年秋季 15 / 65 无线互联网 non-uniformity and coupling in the user and BS locations, which inspires us to consider a PCP-based system model to formulate our optimization problems. 3 System model and problem description 3.1 Multi-cluster heterogeneous network model We consider a E-cluster K-tier HetNet model as in Fig. 1. The heterogeneous network comprising of K different types of BSs in each cluster is considered in this work with its storage capacity denoted by Ck and effective radius denoted by rk, 1 ! k ! K, where the effective radius means that the user can connect to the k-tier BS if and only if the distance between them is less than or equal to the effective radius rk. And this K-tier HetNets in the cluster e are modeled as K independent homogeneous Poisson point process (PPP) with their deployment density denoted as ktotal ke ; 1 ! k ! K; 1 ! e ! E. Again, user distribution on the 2D plane is assumed to be homogeneous PPP with its density denoted by kue ; 1 ! e ! E. We assumed that different clusters have different user preferences and are independent of each other, while for users in the same cluster, they share the same user preference, hence the same content popularity distribution. 3.2 Base station sleeping Because of the spatial and temporal evolution of the user density, here we introduce the base station sleeping to adapt to the time–space-varying user requests and avoid the base station resource waste at the idle time and the base station resource deficiency at the rush time [44]. We consider our BSs running on two modes: activated model and sleep model. The k-tier BSs in the cluster e running on the activated model has the distribution density denoted as kke, which is the variable we should optimize. Hence, the base station running on the sleeping mode has the distribution density denoted by ktotal ke " kke. This BS distribution density is closely related to the user density in each cluster. The energy consumption costs of the k-tier BSs running on the activated model and sleep model are denoted as ak and bk. We define tk ¼ ak " bk, for 1 ! k ! K. 3.3 Dynamic file library and cache refreshing model We assume there is a dynamic file library which contains F files with normalized unit size with its content popularity distribution evolves through time. We denote the fth most popular file at time slot t in the cluster e as ce fðtÞ, and its Fig. 1 A multi-cluster two-tier HetNet model, with user and BSs distributed as independent homogeneous PPP in each cluster Neural Computing and Applications 123 密集部署的HetNet E个cluster[簇]; K类基站[tier]. 用户的空间分布为Poisson点过程. 均值为��, � ≤ � ≤ � 每类基站的分布也是PPP. 均值为�� , � ≤ � ≤ �; � ≤ � ≤ � 第k类基站的通信半径为��, 缓存容量为��, � ≤ � ≤ � 为了节能, 各类基站都有活动模式和睡眠模式. [活动模式下的能耗为��, 睡眠模式下的能耗为��. � ≤ � ≤ �] 本文通过调整决策变量��[e簇中k类基站的活动数目均值] 来控制系统能耗

点击进入文档下载页（PDF格式）

共64页，可试读20页，点击继续阅读 ↓↓

您可能感兴趣的文档

点击购买下载（PDF）

下载及服务说明

购买前请先查看本文档预览页，确认内容后再进行支付；
如遇文件无法下载、无法访问或其它任何问题，可发送电子邮件反馈，核实后将进行文件补发或退款等其它相关操作；
邮箱：

文档浏览记录