Prediction Based On Neural Network Self-similar Traffic

Predicting the behavior of network traffic is very important for admission management and congestion control in the communication network. With the rapid development of modern communication technology, networks have been expanding and the network traffic appears diversity. The increasingly complex behavioral characteristic of the network has brought new challenges to the network traffic modeling and forecasting. In recent years, a great number of studies have shown the common existence of self-similarity in network traffic. The self-similar traffic has negligible effects to network performance. On the other hand, the feature of self-similarity indicates that there exists a predictive structure of the network traffic. The traditional methods of network traffic prediction are to build mathematical models using its statistical characteristics. With the deepening of research in neural network, some scholars also utilize neural network for traffic modeling and forecasting.According to the self-similar characteristic of network traffic, neural networks are applied to self-similar traffic prediction and its practical application. After exploratory researches for the structure and learning algorithm of neural network, an algorithm based on adaptive gain coefficient is presented. The main research work of this dissertation is shown as follows.1. The characteristic of self-similar traffic, as well as the causes of self-similar traffic and its impact on network performance is summarized. After analyzing the prediction researches of self-similar traffic using neural network, two kinds of feed forward neural network, which are BP and FIR neural network, are used as prediction model in this paper.2. The structure and information processing of neuron model are studied. After analyzing the standard algorithm of BP and FIR neural network, a new method based on adaptive gain coefficient is presented. Utilizing improved algorithm in BP and FIR neural networks, the update forms of parameters are deduced. Under the condition of same neural network structure and initial parameters, the influences of proposed algorithm on the prediction results and the convergence speed of mean square error are analyzed.3. The BP and FIR neural networks are applied for self-similar traffic prediction. For the same samples of real network traffic, the best structures of BP and FIR neural networks for prediction are determined by comparing the prediction results of different structures. After comparing the prediction results of BP, FIR and Wiener filter for self-similar traffic, the optimal model is used for prediction tool in dynamic buffer allocation.4. According to the prediction results, the buffer space is allocated dynamically. In several buffer allocation schemes, the analysis of self-similarity on performance of queuing systems provides evidence for the dynamic buffer allocation. Based on the analysis of Complete Partitioning (CP) scheme and Real-Time Dynamic Sharing (RTDS) scheme, a Prediction-Based Dynamic Sharing (PBDS) scheme is proposed. The influences of three different schemes on the packet loss rate of a queuing system are compared in the simulation. In the same time, the fairness of these schemes to each traffic source is analyzed.