Reservoir Computing Networks by Using Gain Saturation in a Semiconductor Optical Amplifier (SOA) Applied to Nonlinear Channel Equalization
Xiyong Liu, Hanwen Gao, Huiwen Luo, Feng Wen · 2024
Reservoir computing (RC) is an efficient information processing model inspired by biology, which makes the training simpler and more efficient through focusing on the training behavior only in the output layer. As the neurons, the nodes in the reservoir layer are connected with the random weights, and stores the state information of the system from the previous time. In such network, the RC is particularly suitable for processing the time related tasks, such as the time series prediction, the nonlinear channel equalization, etc. In the paper, we use a single nonlinear node and the delayed feedback as the reservoir, where the power saturation response from the semiconductor optical amplifier (SOA) performed as the activation function. We experimentally measured the gain curves of SOA under different injection currents, and investigated the optimization of the nonlinear channel equalization (NCEQ) task through these activation functions. Based on the numerical simulation, the symbol error rate (SER) of recovered signals is 0, and the average mean-squared-error (MSE) is only 10−4. Moreover, we also tested the training-length optimization through tuning the injection current of SOA, which could further reveal the dependency of NCEQ on the proposed SOA-RC scheme.