Conditions to Retrain the NN

As previously mentioned, the performance of the model may degrade when there are changes in the system. In this case, the NN should take this into account. This can be done either by on-line training or off-line training (in this case, retraining the working model when the MSE increases over some threshold). For our model, trained by the first week, we tested it using samples of the next complete week as given in Section 9.3.4 which gave for the MSE=0.0046. However, when we used the same trained NN to predict the sixth week, we got MSE=0.012. This may be noticed from Figures 9.5 and 9.8, where the accuracy at the end is not the same as at the beginning. That means that the performance degraded and the NN should be retrained. Retraining the trained NN blindly will not improve the performance too much. This process is a critical one, so, it should be carried out carefully. The retraining samples should be selected as follows. Removing all the spiky samples is a must to get better performance. We have to choose some samples that give good prediction of the last tested period. We have also to choose some samples that are mis-predicted. The choice of the mis-predicted samples is to let the NN learn the changes of the traffic process. The choice of the well-predicted samples is important to keep the NN informed about the past history of the traffic process and to avoid spoiling the trained NN. We have carried out two experiments to retrain the NN. In the first one, we trained it blindly with a database containing spikes. We got MSE=0.009 on the testing database. Once we removed the spikes from the training database in the second experiment, the MSE improved to 0.0054 on the testing database. It is important to mention that in [56], the authors argued that neither the on-line nor the off-line training improves the performance. But this is due to the fact that the spiky samples do not help. As we have shown, by removing them and choosing a fair number of samples from the well- and mis-predicted samples, the performance improved.