Machine Learning algorithms for optimal signal generation in a Quasi-Coherent system with cost-effective devices

Abstract

The applications for Machine Learning (ML) are growing every day thanks to the algorithms that can solve really complex problems with little information of the systems. Deep Reinforcement Learning (a branch of ML) has been involved with control tasks, and in this project, we study the introduction of learning algorithms to control hardware optical devices (considering parameters like bias voltage, signal distortion, Q factor) in complex environments such as the QuasiCoherent receiver and different modulations as OOK and PAM-4 to get the best performance in terms of BER. Additionally, the possibility of using cost-effective devices (like EMLs or VCSELs) to mount the QuasiCoherent receiver and get similar performances than using a TLS was studied, that is why some PCB boards had to be designed and fabricated.