Strong and weak prediction of stochastic dynamics using reservoir computing

We propose an approach to replicate a stochastic system and forecast its dynamics using a reservoir computing (RC). We show that such machine learning models enable the prediction of the behavior of stochastic systems in a wide range of control parameters. However, the quality of forecasting depends significantly on the training approach used for the RC. Specifically, we distinguish two types of prediction-weak and strong predictions. We get what is called a strong prediction when the testing parameters are close to the training parameters, and almost a true replica of the system trajectory is obtained, which is determined by noise and initial conditions. On the contrary, we call the prediction weak if we can only predict probabilistic characteristics of a stochastic process, which happens if there exists a mismatch between training and testing parameters. The efficiency of our approach is demonstrated with the models of single and coupled stochastic FitzHugh-Nagumo oscillators and the model of an erbium-doped fiber laser with noisy diode pumping. With the help of a RC, we predict the system dynamics for a wide range of noise parameters. In addition, we find a particular regime when the model exhibits switches between strong and weak prediction types, resembling probabilistic properties of on-off intermittency.