An approximation network for measurement systems

The problem of extraction of the measured value in optical measurement systems is addressed. It is required that the values of the calibration points be recreated exactly, while maintaining precise approximation between the points. A neural processing method is provided to solve the problem. A two-layer feed-forward network that offers a possibility of on-going insight into the approximation precision, the optimal selection of the successive training layouts, and the linear separability of the training inputs is constructed. Examples are given to illustrate the proposed method.<<ETX>>