The generalization ability of artificial neural networks in forecasting TCP/IP network traffic trends.

Artificial Neural Networks (ANNs) have been used in many fields for a variety of applications, and proved to be reliable. They have proved to be one of the most powerful tools in the domain of forecasting and analysis of various time series. The forecasting of TCP/IP network traffic is an important issue receiving growing attention from the computer networks. By improving upon this task, efficient network traffic engineering and anomaly detection tools can be created, resulting in economic gains from better resource management. The use of ANNs requires some critical decisions on the part of the user. These decisions, which are mainly concerned with the determinations of the components of the network structure and the parameters defined for the learning algorithm, can significantly affect the ability of the ANN to generalize, i.e. to have the outputs of the ANN approximate target values given inputs that are not in the training set. This has an impact on the quality of forecasts produced by the ANN. Although there are some discussions in the literature regarding the issues that affect network generalization ability, there is no standard method or approach that is universally accepted to determine the optimum values of these parameters for a particular problem. This research examined the impact a selection of key design features has on the generalization ability of ANNs. We examined how the size and composition of the network architecture, the size of the training samples, the choice of learning algorithm, the training schedule and the size of the learning rate both individually and collectively affect the ability of an ANN to learn the training data and to generalize well to novel data. To investigate this matter, we empirically conducted several experiments in forecasting a real world TCP/IP network traffic time series and the network performance validated using an independent test set. MATLAB version 7.4.0.287’s Neural Network toolbox version 5.0.2 (R2007a) was used for our experiments. The results are found to be promising in terms of ease of design and use of ANNs. Our results indicate that in contrast to Occam’s razor principle for a single hidden layer an increase in number of hidden neurons produces a corresponding increase in generalization ability of ANNs, however larger networks do not always improve the generalization ability of ANNs even though an increase in number of hidden neurons results in a concomitant rise in network generalization. Also, contradicting commonly accepted guidelines, networks trained with a larger representation of the data, exhibit better generalization than networks trained on smaller representations, even though the larger networks have a significantly greater capacity. Furthermore, the results obtained indicate that the learning rate, momentum, training schedule and choice of learning algorithm have as much a significant effect on ANN generalization ability. A number of conclusions were drawn from the results and later used to generate a comprehensive set of guidelines that will facilitate the process of design and use of ANNs in TCP/IP network traffic forecasting. The main contribution of this research lies in the identification of optimal strategies for the use of ANNs in forecasting TCP/IP network traffic trends. Although the information obtained from the tests carried out in this research is specific to the problem considered, it provides users of back-propagation networks with a valuable guide on the behaviour of networks under a wide range of operating conditions. It is important to note that the guidelines accrued from this research are of an assistive and not necessarily restrictive nature to potential ANN modellers