Abstract
In this survey, we analyze the newest machine learning (ML) techniques applied in modern optical orthogonal frequency division multiplexing (O-OFDM) systems for access, core networks, and multi-channel transmission. From rudimentary to more advanced approaches, ML is proven to be a gold standard technique for signal quality improvement when low transmitter modulation extinction ratio dominates in coherent O-OFDM, and when stochastic-induced nonlinearities are present such as parametric noise amplification in long-haul transmission and the interplay between polarization-mode dispersion and the Kerr-induced nonlinearity. In addition, ML algorithms can effectively tackle determinist nonlinear distortions in O-OFDM networks, as well as inter-subcarrier nonlinear effects (ie, inter-subcarrier four-wave mixing and cross-phase modulation). In essence, ML techniques could be potentially beneficial for any multi-carrier approach (eg, filter bank modulation). The survey illustrates an extensive ML taxonomy for O-OFDM based networks, covering supervised, reinforcement learning and unsupervised ML categories. The transmission performance of various ML-assisted O-OFDM systems is presented taking into account the ML computational complexity toward real-time implementation. We also highlight the strict operating conditions for such systems under which a ML algorithm should perform classification, regression or clustering. Finally, the survey opens research issues and future directions toward ML implementation in radio-over-fiber (RoF) and 5G new radio (NR) systems.