Representing Uncertainty in Deep QoT Models

Quality-of-transmission (QoT) estimation of unestablished lightpaths has been extensively studied in the literature through the development of linear physical layer models (PLMs) and more recently through the application of machine learning (ML). ML has proven its superiority over PLMs on more accurately modeling the non-linear nature of physical layer impairments (PLIs), leading to margin reduction and consequently to network capacity savings. Currently, in optical networks, margins compensating for ML model uncertainties are estimated through empirical myopic approaches, largely ignoring the fact that diverse input lightpaths are subject to different levels of uncertainty. Clearly, to further exploit the advantages of ML, it is necessary that uncertainty is appropriately represented through margins that capture uncertainty over each unestablished lightpath. In this work, we present and compare two uncertainty representation frameworks for deep QoT estimation models. The first framework is based on deep quantile regression and the second on the Monte Carlo dropout technique. Both frameworks achieve further margin reductions and improve classification accuracy of unestablished lightpaths, as opposed to myopic margins traditionally considered. Further, it is shown that the Monte Carlo dropout technique results in better representation of the model uncertainty compared to the deep quantile framework. Importantly, the implementation strengths and limitations of each framework are for the first time revealed and discussed.