Metodi statistici e di machine learning per reti ottiche a minimo margine = Statistical and Machine learning method for minimal margin deployment in optical networks.

In this thesis we consider the problem of the uncertainty on the GeneralizedSignal to Noise Ratio (GSNR) value over fiber optics. We work in an opticalnetwork of Optical Line Systems (OLSs) where all the capacity comes from thetransmission layer. Our project is based on a Python implementation. We studythe GSNR distribution in realistic scenario applying Monte Carlo and machinelearning (ML) techniques.We started by considering a trivial example in which we consider a periodic OLS.Then, we generalize the problem by randomly changing the physical parametersregarding fibers and EDFAs(length, dispersion, connector losses, gain ripples) inorder to reproduce realistic uncertainties over these parameters.We study the variation of the GSNR distribution due to uncertainties over theconsidered parameters assuming Gaussian fluctuations around reference values.We create a dataset of GSNR variations using a python abstraction of the OLSsthat simulates the signal propagation with different parameter values. In orderto study the GSNR distributions we analyse separately the ASE and NLI noisesintroduced during the signal propagation. This enables a deeper insight on launchpower dependency of the GSNR, which is a critical aspect when different fibertypes are considered.After a comprehensive description of the statistical features of the GSNR distribu-tion over a periodic OLS, we extend these results over an entire network.In this scenario, we investigate the total GSNR inaccuracy over generic lightpathsdue to physical parameters uncertainty.As a matter of facts, despite an accurate GSNR prediction can be produced by va-rious quality of transmission estimators(QoT-Es), the latter requires highly preciseknowledge of the physical parameter. As this condition is not guaranteed in reali case scenarios, the study of the GSNR fluctuation is a variable tool in fixing theproper system margins.This investigated framework represents an ideal scenario for the application ofML on top of the statistics of GSNR fluctuations.This may be achieved by collecting a dataset of the OLS responses to variousspectral loads in order to train a ML algorithm, allowing a QoT-E to be calculatedfor both untested spectral load configurations and LPs which have not yet beenexplored.