Conclusion and Discussion

We have proposed in this chapter a new model for traffic prediction. The novelty of this model with respect to the existing literature, is that it makes use of both the long-range and short-range periodicity of the traffic process to provide a more accurate estimation. Good candidates for solving this kind of problems are neural networks, used here to learn from the past history to predict the future arrivals of the traffic process. As other models, for short-range periodicity, we use the information in the past few time steps (``now'' window). However, to make use of the long-range information, we added other inputs: the current time of the day, the current day of the week, past information about the traffic in the previous day at the same time, and past information about the traffic in the same day of the previous week. We validated our model by using real traffic traces collected from a large institution (the ENSTB) over a long period (6 months). We did not use traces generated by artificial simulations or mathematical equations (generally the case in the existing models) as they are easy to model by NN and do not characterize well real traffic processes. In addition, we did not use any special kind of traffic (like MPEG) which are also easy to learn by NN as they are periodical by nature. The existing models based on these kinds of traces work well for the special type of flow, but their performances degrade significantly when applied to traces corresponding to general traffic, as we show in this Chapter. We run some experiments to identify the best architecture of our model. The best candidate appears to be: 3, 2, and 2 for now, yesterday and last week window sizes respectively. This architecture may be the best for the ENSTB network, and should not be taken as a reference. We provided the guidelines to show how to identify the optimal values corresponding to a given network. We also derived a methodology to predict more than one step in the future and we evaluated it. We have found that when the retraining samples are selected carefully, the performance improves considerably. We showed through different experiments that our technique outperforms existing ones, and our explanation is that our method uses long-range information while classical attempts to solve the same problem don't. We discussed how different applications can take benefit from our traffic prediction. Dynamic bandwidth allocation, dynamic long-term contract negotiation, pricing, traffic shaping and engineering are some of them. ChapterChapter