Neural networks associated with the "black box" models of non-linear dynamic systems

The approximation of non-linear dynamic system operators by way of describing the input-output relationship with the help of mathematical models is considered. A neural network is one of famous mathematical models. The types of neural networks are represented as the universal approximators of non-linear operators. The classification of recurrent neural networks according to the feedback location is described. The recurrent Hammerstein network is used as a mathematical model of a non-linear compensator for digital communication channels. The source of non-linear signal distortion in communication channels is a power amplifier. It is found that the model of a non-linear compensator in the form of the recurrent Hammerstein network exceeds the two-layer perceptron network in the accuracy of signal processing and the Volterra polynomial in the simplicity of hardware implementation.