Comparison with Other Models

The majority of the existing works in this domain employ the 5-5-1 NN (as depicted in Figure 9.9) to predict the future traffic. To compare this architecture with our model, the traditional one has a ``five-now-window'' without day nor week windows, nor time nor date, and the NN has 5 hidden neurons. We run the same experiment as we did for our model but using the traditional 5-5-1 model, and we have found that the performance of our model is much better than the traditional one. With the latter, we have got for the MSE 0.0092 instead of 0.0046 with our model (a factor of 1.9 improvement when using our model). It is important to say that the traditional model were tested either on simulated traces or on traces that last only several hours sampled every second (Bellcore). It is not surprising to see in the literature that this simple model gives good results. However, once it is employed in long-term prediction, it does not give the same success rate as we do using our proposed methodology.

Figure: The traditional NN model that has been widely used to predict network traffic. This Figure is taken from [56, p. 115], where $z^{-1}$ represents a unit-step delay function.

From [56] and the work presented in [157], we show in Figure 9.10 which is taken from [56], that the NN can capture the traffic with very good performance. However, the authors obtained these results by generating the data to train and test the NN using the following equation.

$$X(t)=4\times X(t-1)\times\left(1-X(t-1)\right)$$ (91)

with $X(0)\in]0.0,1.0[$ except 0.5. It is clear that this traffic model is very simple and very easy to be learned by any classical NN architecture. However, this kind of data (claimed in [56] to simulate real network traffic) is far from being like real network traces as argued.

Figure: The actual against the NN prediction when training and testing it by data generated by Eqn. 9.1. This Figure is taken from [56].