Generalization Evaluation of a Nonlinear Auto-Regressive Neural Network for PON Technologies

Building robust models with high generalization capability that can efficiently solve problems in a wide range of situations is essential to promote the development of enabling technologies. In this context, this paper evaluates the generalization capacity of a nonlinear autoregressive with external input neural network (NARXNET) for a passive optical network (PON). We adopted this type of neural network, given its applicability for nonlinear filtering processes, in which the target output is a noise-free version of the input signal. The built NARXNET was assessed with eight extra distinct experimental data sets demonstrating average RMSE and R2 values of 0.295 and 0.961, respectively, with a high generalization capability. Furthermore, NARXNET showed a fast training time of 240 seconds and an improved eye diagram for on-off keying (OOK) modulation format on a PON system.