Intermittency in predicting the behavior of stochastic systems using reservoir computing

A new behavior type of reservoir computing model for predicting dynamics of stochastic systems has been observed. It has been shown that when the control parameters of the predicted stochastic system and the reservoir computing model are turned, we observe intermittent behavior, i.e., close to the threshold parameter value the reservoir computing model demonstrates the accurate prediction most of the time, but there are time intervals during which the accurate prediction is interrupted by intervals characterized by the lack of prediction. The characteristics of the intermittency in predicting the behavior of the stochastic system correspond to the well-known on-off intermittency. The concept of the effective noise to describe the quality of prediction is proposed, and the technique of its amplitude value estimation is developed.