On the generalization of cognitive optical networking applications using composable machine learning

Abstract

Model generalization characterizes the sustainability of machine learning (ML) designs applied to novel system states and therefore plays a vital role toward the realization of cognitive networking. In this paper, we present a composable ML framework (namely, CompML), aiming at generalizing ML-aided cognitive applications for optical networks. CompML makes use of three basic functional modules, i.e., the Loading, Recursion, and Readout modules, to model the loading/initialization processes (e.g., the launch of a signal), extract cumulative features by recursive operations, and produce model inferences, respectively. By the composition of the three modules and adoption of an end-to-end training mechanism, CompML allows for generalizing multiple tasks of the same domain [e.g., quality-of-transmission (QoT) estimation for different lightpaths]. We perform case studies of CompML on QoT estimation and nonlinearity compensation using both simulation and experimental data. Results show the superior generalization ability of CompML compared with the baselines, achieving mean absolute error (MAE) for generalized signal-to-noise ratio (GSNR) prediction error of below 1.06 dB for unseen lightpaths and up to 3 dB Q-factor improvement for nonlinearity compensation.