Traffic Prediction

A fourth important problem is traffic prediction. This problem is of extreme importance because traffic prediction is crucial for successful congestion control, congestion avoidance, dynamic bandwidth allocation, dynamic contract negotiation, traffic shaping and engineering, etc. More precise congestion avoidance mechanisms can be implemented by considering both traffic prediction and the rate control previously presented. We have also explored this issue for the following reasons: $\bullet$

• because of its direct relation with control techniques,
• because neural network tools work quite well for the case of rate control and the quality assessment problems.

However, there are some difficulties that so far prevented providing good traffic predictors that can work in real situations, mainly related to the fact that network traffic is, in general, a complex, time-variant, nonlinear, and nonstationary process. Furthermore, this process is governed by parameters and variables that are very difficult to measure. Sometimes, there are completely nonpredictable portions of this time-variant process (the so-called ``spiky'' fragments, see Section 9.3.4). Therefore, a precise model of this process becomes difficult as its complexity increases. Despite these problems, traffic prediction is possible because, as the measurements have shown, there coexist both long-range and short-range dependencies in network traffics. For example, the amount of traffic differs from the weekend to that in the weekdays. However, it is statistically similar for all the weekends, also during the same day, in the morning, in the nights, at some specific parts of the day, etc. There are many proposals in the literature for traffic prediction [85,27,121,120,156,39]. These proposals concentrate on the short-range dependencies, somehow neglecting the long-range ones. When testing these models on real traces lasting for several minutes or even hours, they give good performance. However, when employing them for predicting the traffic for days or weeks, their performance degrades significantly. Our aim in this part of our dissertation is to propose a new model for traffic prediction that takes into account both long-range and short-range dependencies.