abstract
- We investigate multi-objective hyperparameter optimization of machine-learning-based nonlinear equalizers for coherent optical links. A sequentialized (S) MLP, also denominated time-delay neural network (TDNN) is adopted as a baseline model, and its hyperparameters (spanning discrete, categorical, and continuous variables) are jointly optimized using the Nondominated Sorting Genetic Algorithm II (NSGA-II). The objectives are the bit-error ratio (BER) after equalization and the computational cost expressed in floating-point operations. Results for a 400ZR-like DP-16QAM transmission show that NSGA-II efficiently constructs well-distributed Pareto fronts and identifies configurations that significantly reduce complexity while preserving near-optimal BER. Although the study focuses on MLPs due to their simplicity and interpretability, the proposed NSGA-II framework is directly extensible to more advanced architectures, such as CNNs and RNNs, where the hyperparameter space becomes even more complex. These results confirm the suitability of evolutionary multi-objective methods for designing practical nonlinear equalizers in coherent receivers.