abstract
- The performance of coherent optical systems is influenced by the combination of additive noise and nonlinear impairments, motivating the development of efficient equalization strategies. This paper explores decision tree (DT) classifiers as a lightweight detection method for a 112 Gbps dual-polarization 16-ary quadrature amplitude modulation (16QAM) optical interconnects. A power-diversity training approach is introduced, in which classifiers are trained on grouped launched power ranges (low, medium, high), avoiding the harmful overfitting. Simulation results show that this method improves bit error rate (BER) within the range between 1 and 12 dBm compared to per-power training and maximum likelihood detection. Furthermore, joint polarization processing provides additional gains at medium and high powers by addressing cross-polarization nonlinearities.