Synthesis of a neural network controller for a linearized model of the plant

This article demonstrates the procedure for the synthesis of a neural network controller for a linearized model of the plant «two inverted pendulums on a cart». A feature of this plant can be considered its non-square matrix transfer function and the number of input effects is less than the number of output effects. The task of the control is to stabilize the angles of the inverted pendulums in a stable position and transfer the position of the cart to a set value. A distinctive feature of the demonstrated procedure for the synthesis of a neuroregulator is the deterministic choice of architecture and initialization of the weighting coefficients of the neural network. Data on the choice of architecture and on the values of initialized weighting coefficients for the neural network controller are obtained based on information about the transfer function of the controller obtained by the modal method using the polynomial matrix decomposition of the system. Recommendations for structural transformations of a neural network regulator containing recurrent connections are given. They are necessary for further training of a neural network controller with a deterministic approach to initialization of weight coefficients used in the demonstrated synthesis procedure. As a result of the complexity of the structure and further training of the neural network controller obtained, it is possible to increase the efficiency of the automatic control system compared with the system using the controller obtained by the modal method.

Authors: A. A. Voevoda, V. I. Shipagin

Direction: Informatics, Computer Technologies And Control

Keywords: neuroregulator, recurrent neural network, neural network training, initialization of weighting coefficients, structural transformations

View full article

Synthesis of a neural network controller for a linearized model of the plant – two inverted pendulums on a cart