Echo-State Networks for Network Traffic Prediction

The ability to forecast future levels of network traffic is very important to network management and planning in enterprise, cloud, data center and WAN networks. A good prediction of future traffic levels makes it possible for network operators to plan ahead on how to achieve service delivery targets in the areas of resource allocation, congestion control, capacity planning, anomaly detection and quality of service. As a result, there has been significant research interest in finding good methods that can be used for traffic prediction.Various neural network based methods have been explored in research for prediction of network traffic. However, many of these approaches often require long processing times to create a model from the network data and predict traffic. In this paper, we present the results of our ongoing study on the use of echo state networks (ESN) for the prediction of network traffic. ESNs are a type of reservoir learning algorithm that have an internal state that is made up of a 'reservoir-like' pool of randomly connected neurons. In our experiments, we compare results obtained by using ESNs with the results obtained by using other techniques for network prediction on two datasets obtained from a well known public network trace repository. We show that ESNs achieve significantly faster training and prediction times for similar levels of accuracy when compared to the other methods.