Alternatives to neural networks for inferential measurement

Abstract Neural networks have attracted much attention as a means of modelling nonlinear phenomena, for example for inferential measurement in process control. A neural network is a nonlinear multivariable function whose main potential advantage: the ability to represent highly nonlinear input—output relationships, is in fact not essential in many potential process engineering applications. This advantage is in any case frequently outweighed by its major disadvantage: the intractability of the parameter estimation procedure resulting from the highly nonlinear form of its parameters. In this work we show how moderately nonlinear functions, with easily estimated parameters may be used in certain inferential measurement applications for which neural networks have been proposed. These functions are as effective in representing input—output relationships and their parameters can be fitted far more rapidly than can the “weights” of a neural network. Furthermore, we show that the performance of both these functions and neural networks, being arbitrary representations having no physical basis, may almost invariably be improved upon by the use of even very simple approximate models based on proper physical understanding.