Machine Learning Assisted Optimization of Dynamic Crosstalk-Aware Spectrally-Spatially Flexible Optical Networks

We focus on optimization of dynamic spectrally-spatially flexible optical networks (SS-FONs), in which distance-adaptive, spectral super-channel (SCh) transmission is realized over weakly-coupled multi-core fibers (MCFs). In such networks, the inter-core crosstalk (XT) effect in MCFs impairs the quality of transmission (QoT) of optical signals, which has a negative impact on overall network performance. In this context, a key issue is the selection of a modulation format (MF) for a particular lightpath, since each MF offers a different tradeoff between spectral efficiency (number of occupied frequency slices) and transmission reach, and both these elements are in a close relation with the experienced XT. In fact, it is difficult to quantify the impact of the MFs used in a given network on the deterioration in QoT due to the XT effect and on overall network performance, as the levels of XT experienced by particular lightpaths depend directly on the ongoing transmissions in neighbor MCF cores, and those (i.e., established lightpaths) are subject to stochastic changes in dynamic SS-FONs. Therefore, in this article, we propose an approach that selects a MF for a lightpath based only on the transmission distance, i.e., length of the applied routing path. In particular, for each considered MF, we define a Modulation format Distance Limit (MDL) parameter that is used to select the MF for a lightpath by a simple comparison of the considered routing path length against the MDL parameter. To facilitate the very complex process of determining the best values of the MDL parameter, we propose a machine learning (ML) approach, namely, supervised learning regression method, with the overall goal to optimize network performance in terms of bandwidth blocking probability (BBP). Extensive simulation experiments performed in different network scenarios show that the proposed method ML allows to achieve performance gains when compared to various reference approaches, including a more computationally expensive method that is based on dynamic (flexible) selection of MFs.