Estimating Throughput in Optical Backbone Networks Using Deep Neural Networks

Abstract

Optical backbone networks, which use optical fibres as the transmission medium, form the core infrastructure used by network operators to deliver services to users, as well as by Internet companies to route traffic between data centres. The network throughput is a key parameter in the analysis of the networks' performance. However, its determination can be a complex process that involves long computation times, since aspects related to both the physical and network layers need to be accounted for. To face this challenge, we propose a machine learning solution: a deep neural network (DNN) model, that has the goal of estimating the values of the network throughput and of a closely related parameter, average channel capacity, accurately and with short computation times. The simulation results indicate that the DNN model accurately predicts both outputs, with mean relative errors of 6.17% for the network throughput and 2.84% for the average channel capacity. These predictions are made in just a few milliseconds, providing a significant advantage over the heuristic routing algorithms, which can take up to tens of seconds in larger networks.