Assessment of Domain Adaptation Approaches for QoT Estimation in Optical Networks

Predicting the Quality of Transmission (QoT) of a candidate lightpath prior to its establishment plays a pivotal role for an effective design and management of optical networks. In the last few years, supervised Machine Learning (ML) techniques have been advocated as promising approaches for QoT estimation, but to ensure the effectiveness of their training phase, a large amount of samples (training set) must be provided to the learning algorithm. Unfortunately, the collection of training samples is often hindered by practical issues (e.g., lack of dedicated telemetry equipment in every network node) or is too costly to permit the acquisition of large datasets. However, it is sometimes possible to rely on large training datasets from a different network (source domain) than the one on which the ML model operates (target domain). In such a scenario, we wish to exploit at best the data from the source domain to tailor a good model to the target domain. This approach is known in ML research as Domain Adaptation (DA). Note that most of the existing DA techniques require to complement the dataset from the source domain with a few samples from the target domain: quantifying the amount of samples extracted from the target domain needed to achieve satisfactory predictive performance of the adopted learning model is a crucial issue to determine the practical applicability of DA techniques in real-world scenarios. In this thesis, we evaluate the effectiveness of two existing DA approaches (i.e., feature augmentation and domain adaptation) for ML-based QoT estimation of candidate lightpaths, for a fixed/variable number of available training samples from the source/target domain. As source/target domains, we consider several network topologies with varying transmission equipment characteristics. Results show that, when the number of samples from the target domain is very limited (e.g., in the order of tens), DA approaches consistently outperform standard supervised ML techniques.