A Comparison of Machine Learning Techniques for Fiber Non-Linearity Compensation: Multilayer Perceptron vs. Learned Digital BackPropagation

In this paper, we analyze the performances of a Multilayer Perceptron (MLP) and the physics-based model called Learned Digital Backpropagation (LDBP) for compensating for fiber non-linear impairments in coherent transmission systems. We applied both methods numerically to a single-channel single polarization 200Gb/s 16QAM system along 5x100km and 10x100km of standard single-mode fiber. We evaluated both techniques in terms of bit error rate (BER) and computational complexity (CC). Furthermore, we identified three study cases for the MLP depending on its position in the receiver. In our setup, the results showed that the LDBP improves BER much better than the MLP, even at 1 step/span. We also showed that the CC of the LDBP is lower than the MLP one concerning real multiplications per symbol. However, in the real life, the LDBP, being a sequential technique, might take more processing time than the MLP which is easily parallelizable.