Machine learning-based QOT prediction for self-driven optical networks

Abstract

Nowadays, digital businesses with diverse deployment models such as cloud, mobile and edge devices for the internet of things will impact traffic, in both volume and dynamicity, at unprecedented rates. Moreover, due to the recent advances in optical networks and systems, the complexity of provisioning lightpaths is growing dramatically. Hence, optical network operators are forced to change their insight and move toward intent-based and self-driven networking, to cost-efficiently accommodate these challenging requirements. In this regard, knowledge-defined networking (KDN) promises to play a paramount role in realizing flexible and self-driven optical networks. In this work, we focus on one of the key aspects in this environment, i.e., prediction of quality of service for unestablished lightpaths. KDN is a solution that introduces machine learning techniques into the control plane of the network, to cope with inevitable complexities that arise in enabling network to operate autonomously and faster. For this, five machine learning models are evaluated for the classification and regression approaches. Multilayer perceptron, radial basis function and generalized regression neural network (GRNN) models are used for both of the regression and classification approaches, while support-vector machine and probabilistic neural network (PNN) models are used only for the classification scenario. Also, to discard the redundant features (among the considered experimental features) in the classification approach, input features are selected using the analysis of variance (ANOVA) test. The proposed models can accelerate and handle a significant part of operations in the closed-loop architecture of knowledge-defined optical networks, as a paradigm for designing self-driven optical networks. The best accuracies of quality of transmission prediction (classification approach) and optical signal-to-noise ratio estimation (regression approach) are achieved using PNN (with average accuracy of 99.6 ± 0.5%) and GRNN (with R-squared value of 0.957), respectively.