A data-driven QoT decision approach for multicast connections in metro optical networks

Abstract

A data-driven technique for analyzing Quality-of-Transmission (QoT) data of previously established connections is proposed for accurately deciding the QoT of the newly arriving multicast requests in metro optical networks. The proposed approach is self-adaptive, it is a function of data that are independent from the physical layer impairment (PLIs) and thus does not require specific measurement equipment, and it does not assume the existence of a system with extensive processing and storage capabilities. It is also fast in processing new data, and fast in finding a near-accurate QoT model provided that such a model exists. The proposed technique can replace the existing Q-factor models that are not self-adaptive, they are a function of the PLIs, and their evaluation requires timeconsuming simulations, lab experiments, specific measurement equipment, and considerable human effort. The proposed datadriven QoT approach is based on the utilization of a feed-forward neural network that is trained on a dataset previously generated from a known Q-factor model. The dataset fed to the neural network is represented in a way that specifically describes the QoT of the multicast connections requesting to be established in the network but it is independent from the PLIs. The validity of the proposed approach is examined for two distinct networks, exhibiting a high accuracy when compared to the results of the Q-factor model utilized for generating the QoT data.

An Overview on Application of Machine Learning Techniques in Optical Networks

“…Similarly, in the context of multicast transmission in optical network, a NN is trained in [43], [44], [46], [47] using as features the lightpath total length, the number of traversed EDFAs, the maximum link length, the degree of destination node and the channel wavelength used for transmission of candidate lightpaths, to predict whether the Q-factor will exceed a given system threshold. The NN is trained online with data mini-batches, according to the network evolution, to allow for sequential updates of the prediction model.…”

Learning quantile QoT models to address uncertainty over unseen lightpaths

“…To achieve this objective, several ML models have been exploited, such as DNNs, k-nearest neighbors, support vector machines, and Gaussian Processes [23,31], with DNNs validated both in the field [32,33] and experimentally [2,34] with the use of real datasets. In general, the model of choice, is trained either as a binary classifier [29,[35][36][37] or as a regressor [5,6,22,27] by means of supervised learning, given a labeled dataset of previously observed lightpaths.…”

Representing Uncertainty in Deep QoT Models

“…In spite, however, of the underlying model of choice, ML-aided QoT models are usually trained either as binary classifiers [15], [20]- [22] or as regressors [8], [9], [13], [16]. Classifiers are trained to estimate the class of an unestablished lightpath; that is, the feasible or the infeasible class, determined according to a predefined QoT threshold.…”