Predicting ADSL Lines Data Rate Using Neural Network

Asymmetric Digital Subscriber Line - ADSL - technology is the preferred high-speed access to the Internet, offering up to 24Mb/s on downlink channels. Exploiting the copper media already deployed by Telecom Operators, physical link quality often limits the maximum ADSL data rate, so that the actual bitrate of a line is often unpredictable. In this work, we analyze a large set of end-users' ADSL links, and try to correlate the effective bitrate with the physical measurements exposed by network devices. By exploiting a Neural Network (NN) predictor, we acquire knowledge about the behavior of ADSL lines from the huge amount of measured data. We show that NNs are good tools to automatically predict end-users' ADSL available bandwidth. However some ingenuity is required to guide the learning phase and to avoid misbehaving lines to fool the NN prediction.