PULSE SHAPING OPTIMIZATION VIA MACHINE LEARNING FOR OPTICAL FIBER TRANSMISSION

In this work, we propose a non-conventional framework based on machine learning (ML) able to provide equalization capabilities in a communication channel with impairments. Traditional Pulse Shaper/Matched Filter blocks will be replaced by trainable FIRs aimed at the optimization of the pulse shaping in order to compensate for channel impairments in absence of a traditional equalization block. The specific application case that motivated this work is the short-reach optical link; on this scenario we will base the literature review and the problem statement. Then, in the simulations, the impaired channel we tested is a band-limited AWGN channel whose perfect knowledge allowed us to provide a benchmark for the obtained results, achieving the goal of validating the proposed ML-based method. In the view of a more widespread adoption of ML in communications, this work can be considered as a methodological evaluation of a non-conventional ML-based equalization method, aimed at exploring the opportunities that ML can enable on the Pulse Shaper/Matched Filter blocks and, in principle, can be extended to more complex channels where the optimal solutions have not been found yet - like the optical channel - providing effective improvements in terms of performance and complexity.