Abstract
The scope of this paper is to present a general purpose method for approximating an arbitrary continuous function on a compact set from a given set of observations. The method consists of constructing a model based on a feedforward multilayer network, embedding both classical data analysis techniques and connectionist or neural network techniques. The model construction can be divided into three steps: (1) principal components analysis is first applied to reduce the number of input variables and uncorrelate them, (2) multiple regression analysis is used to derive the best linear estimator and (3) multilayer neural networks are trained to extract the non-linear components of the function. A review of these techniques is presented and an application to the prediction of apple quality from near infra-red spectra is discussed. Apple quality can be estimated by the sugar content of the fruit for which a non-destructive measure is available through the use of infra-red spectrometry. A prototype was built for this purpose and several experiments were conducted. The results of the experiments were used to derive a model for predicting apple quality from near infra-red spectra. The percentage variation explained ( R 2 ) of the linear model was equal to 0·82 and the remaining error was reduced by 5% using multilayer neural networks.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have