Multi-step ahead prediction using neural networks

Multimedia applications transmitted over the Internet generate a major part of the Internet traffic. The bursty characteristics of the video traffic make it difficult to fulfill the requirements for Quality of Services (QoS) of such applications. Among other procedures for traffic and congestion control bandwidth allocation is also one of the options to guarantee the specified QoS. Dynamic bandwidth allocation can be successfully performed with the use of traffic prediction. Neural networks are a vastly used tool for prediction. The multi-step ahead prediction is more difficult approach then the single-step ahead prediction, but because video time series include long-range time dependencies, the multi-step ahead prediction seems as a better approach to video traffic prediction. In this paper we present three different approaches to multi-step ahead prediction and compare the results. First we describe basic principles of two types of neural networks, which we use for all the three approaches: the multilayer perceptron and the Nonlinear AutoRegressive model with eXogeneous inputs neural network (NARX). Then we briefly describe the composition of the video trace files. In the last section we present the results for multi-step ahead video prediction.