The neural network to identify an object by a sequential training mode

Abstract

The problem of building a real-time model for the nonlinear dynamic object is considered. The parameters of the object are unknown in the General case can be treated as time-varying. For control, it is necessary to build an adequate model of this object. The model is built using a recursive neural network. The identification quality criterion is the average loss. Based on observations of the input and output of the object and the configurable model, the identification algorithm changes the parameters so that the average loss reaches a minimum with growth. The results of the analytical construction of the model and the results of modeling based on specially developed software are presented.