Overfitting Characteristics of Reservoir-Computing-Based Nonlinear Equalizer Trained on PRBS and Repeated Random Bit Sequence

We investigated the overfitting characteristics of a reservoir-computing (RC)-based nonlinear equalizer, which is used to compensate for optical nonlinear waveform distortion in optical fiber communications systems. The RC-based nonlinear equalizer is gaining attention due to its high-speed training capability and feasibility for hardware implementation. However, its overfitting characteristics have not been thoroughly explored yet. We evaluated the overfitting of the RC-based nonlinear equalizer by comparing it with that of a well-studied three-layer artificial neural network (ANN)-based nonlinear equalizer. The number of reservoir units was 50, 100, and 200. The spectral radius of the RC was also varied to control the performance of the RC-based nonlinear equalizer. We employed pseudorandom binary sequences (PRBSs) and repeated random bit sequences (RRBSs) to train both equalizers and evaluated the overfitting on the training signals. Our results showed that when the equalizers were trained on PRBSs, the ANN exhibited stronger overfitting compared to the RC. This is because the ANN is more capable of efficiently learning the generation rules of the PRBSs and predicting the received PRBSs more easily than the RC. In contrast, when RRBSs were used to train the equalizers, the overfitting was weaker compared to the case with PRBSs in both the ANN and RC.