Data-Driven Erbium-Doped Fiber Amplifier Gain Modeling Using Gaussian Process Regression

Abstract

We propose a data-driven erbium-doped fiber amplifier (EDFA) gain model utilizing Gaussian process regression (GPR). An additive Laplacian and radial-basis function kernel is proposed for the GPR and was found to outperform deep neural network (DNN) methods while additionally providing prediction uncertainty. Performance is measured using mean absolute error (MAE) averaged across five different EDFAs with three manufacturers. The GPR achieves an MAE of 0.1 dB using 30 training samples in contrast to the DNN that achieves an MAE of 0.25 dB using 3000 training samples. Additionally, we demonstrate that active learning can be used to improve robustness and repeatability of convergence.