Topology Prediction Mechanism For Opportunistic Network Based On Deep Belief Network

Opportunity network(ON) is a kind of mobile self-organizing network, who does not need a complete communication route between source node and destination node.It makes use of the encounter chance bring by the node mobility to communicate.ON has the features of node mobility and intermittent connection between nodes,which lead to the network topology changed frequently over time and bring some research challenges, including routing forwarding mechanism, network load, quality of network service, network behavior prediction, and etc.The thesis is supported by the National Natural Science Foundation, it will study on topology prediction issues of network behavior prediction for ON. It focus on, 1)building of similarity metrics; 2) building of deep belief network(DBN); and 3)building of support vector regression machine(SVR). According to temporal variation of ON, in consideration of the link weight, the node strength and local path between the nodes, the similarity metrics that can reflect the topology dynamic change over time in ON will be built based on time series theory. The DBN model is constructed in terms of information entropy and self-adapting learning rate for feature extraction. The number of hidden layer neurons of restricted boltzmann machine(RBM) will be computed automatically through information entropy. And through tuning learning rate self-adaptively, reconstruction error of RBM goes to stable rapidly, so that convergence time is shortened. The regression model will be built in terms of Gaussian core function and k-fold cross verification based on least squares support vector regression machine(LS-SVR).Evaluating indicator HITR_, Precision and Accuracy were used for evaluating the results of topology prediction. And groups of contrast experiment on real data from INFOCOM05(INF’05) set and MIT set are designed for verifying of models.The experiment results show that information entropy can be used to find out the appropriate value of the number of hidden layer neurons of RBM according to the input data, and adaptive learning rate can speed up the convergence rate of RBM.Compared with LS-SVR, DBN-LS-SVR has better modeling ability and the ability to fit the input data, and can get better prediction results.