Investigating the Creation of a Surrogate Model for Adaptive Control of Amplifier Operating Point Using Machine Learning

Abstract

Dynamic operation is one of the current challenges in optical communication and networks, and the adaptive control of optical amplifier (ACOP) is one of the problems in this challenge. The ACOP approaches aim to define the gains of the optical amplifiers dynamically to increase the quality of the transmission after a cascade of amplifiers. The most recent ACOP approach uses a multiobjective evolutionary optimization algorithm to define the gains of the amplifiers to maximize the optical signal to noise ratio (OSNR) and to minimize OSNR ripple. Despite the promising results regarding Quality of Transmission, it is not desirable to rely on an evolutionary algorithm to make decisions in real-time. In this work, we investigate the creation of a surrogate model that can obtain solutions as good as the multiobjective algorithm, but in real-time. We show the results for a machine learning (ML) regression technique, trained with the optimization algorithm solutions, can return configurations with OSNR less than 1 dB close to the best OSNR returned by the optimization algorithm. Moreover, the ML solution answers in milliseconds, whereas the optimization-based approach needs several minutes to find a proper configuration.