基于深度学习的机会网络拓扑预测机制研究

发布时间：2018-01-19 16:08

本文关键词： 机会网络深度信念网络拓扑预测相似性指标最小二乘支持向量回归机　出处：《南昌航空大学》2016年硕士论文　论文类型：学位论文

【摘要】：机会网络(Opportunity Network,ON)是一种不需要源节点和目标节点之间存在完整连通路径,利用节点移动带来的相遇机会实现通信的移动自组织网络。机会网络中节点移动频繁,节点之间间歇性连接,致使其网络拓扑结构随时间频繁地发生改变,这给机会网络研究带来了诸多困难,主要包括路由转发机制、网络负载与效率、网络服务质量、网络行为预测等。本文来源国家自然科学基金项目,研究机会网络中网络行为预测的拓扑预测问题,主要内容如下:(1)相似性指标的建立;(2)深度信念网络(Deep Belief Network,DBN)模型的建立;(3)支持向量回归机的建立。针对机会网络的时变性,基于时间序列理论和方法,在综合考虑节点之间权值、局部路径和节点强度三个方面的基础上,构建了一种能够反映机会网络拓扑结构随时间动态变化的相似性指标;基于信息熵理论、自适应学习率构建作为特征提取器的DBN模型,其中基于信息熵理论自动计算得到受限玻尔兹曼机(Restricted Boltzmann Machine,RBM)隐含层神经元数量,采用自适应学习率使RBM的重构误差快速达到平稳,缩短网络收敛时间;采用高斯核函数、K折交叉验证等方法构造基于最小二乘支持向量回归机(Least Squares Support Vector Regression Machine,LS-SVR)的回归机模型(DBN-LS-SVR)。本文采用命中率HITR_和受试者工作特征曲线(Receiver Operating Characteristic Curve,ROC)中的Precision、Accuracy指标来评价拓扑预测结果,并且在INFOCOM05(INF’05)、MIT数据集上设计了多组对比实验验证DBN-LS-SVR模型。实验结果表明信息熵方法可以根据输入数据找出RBM隐含层神经元数量的合适值,自适应学习率可以加快RBM网络的收敛速度,在一程度上提高了DBN网络的计算效率;与LS-SVR模型相比,DBN-LS-SVR模型的建模能力和拟合输入数据的能力更强,能够获得更好的预测效果。
[Abstract]:Opportunity Network (ON) is one that does not require a complete connectivity path between the source node and the target node. Mobile ad hoc networks are implemented by using the encounter opportunities brought by node mobility. In opportunistic networks, nodes move frequently and internodes connect intermittently, which results in frequent changes in network topology over time. This brings many difficulties to the research of opportunity network, including routing and forwarding mechanism, network load and efficiency, network quality of service, network behavior prediction and so on. The topology prediction problem of network behavior prediction in opportunistic networks is studied. The main contents are as follows: (1) Establishment of similarity index; (2) Establishment of Deep Belief Network (DBN) model; 3) the establishment of support vector regression machine. Aiming at the time-varying of opportunity network, based on the theory and method of time series, considering the weight between nodes, the local path and the strength of nodes. A similarity index which can reflect the dynamic change of opportunity network topology with time is constructed. Based on the information entropy theory, the adaptive learning rate is used to construct the DBN model as a feature extractor. Based on the information entropy theory, the number of neurons in the hidden layer of restricted Boltzmann machine is calculated automatically. The adaptive learning rate is used to make the reconstruction error of RBM stable quickly, and the convergence time of the network is shortened. Gao Si kernel function is used. K fold cross validation and other methods based on least squares support vector regression (LS-SVM). Least Squares Support Vector Regression Machine. The regression model of LS-SVR is DBN-LS-SVR.The hit ratio HITR_ and the operating characteristic curve of the subjects are used in this paper. Receiver Operating Characteristic Curve. The PrecisionAccuracy indicator in ROC) is used to evaluate the topology prediction results, and is found in INFOCOM05 / INF5). The experimental results show that the information entropy method can find the appropriate number of neurons in the RBM hidden layer according to the input data. Adaptive learning rate can accelerate the convergence speed of RBM network and improve the computational efficiency of DBN network to a certain extent. Compared with LS-SVR model, the model of DBN-LS-SVR has better modeling ability and ability of fitting input data, and can obtain better prediction effect.
【学位授予单位】：南昌航空大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TN929.5

【相似文献】