A GMDH network using backpropagation and its application to a controller design

When designing control systems, it is necessary to make mathematical models which describe the controlled objects in detail. However, since the controlled objects generally have nonlinearities and uncertainties, it is difficult to obtain exact models. As an artificial network model, the group method of data handling (GMDH) scheme has been proposed. The GMDH network has a feature that the nonlinear dynamics are clearly expressed as a mathematical model. Therefore, it is relatively easy to obtain the system properties in detail. A new design scheme is proposed, which adjusts all weighting coefficients based on backpropagation. The newly proposed scheme enables us to obtain a mathematical model with superior approximation ability for nonlinear systems. Furthermore, the generalized minimum variance control (GMVC) system is constructed by using the proposed GMDH network. A numerical example for a process system is demonstrated to illustrate the effectiveness of the proposed scheme