A comparison of four methods for non-linear data modelling

Abstract This paper reports results from practical experiments with four different data modelling methods evaluated on five different real or simulated modelling problems. The methods are: partial least squares regression (PLS), back-propagation multilayer perceptron neural networks (MLP), radial basis function neural networks (RBF), and an adaptive B-spline modelling algorithm (ASMOD). The last four methods have the capability of identifying and representing general non-linear dependencies in the data without a priori specifying which specific non-linear dependencies to look for, whereas PLS needs active interaction by the data analyst to identify non-linear dependencies. The real-world modelling problems are: identifying the dynamic actuator characteristics of a hydraulic industrial robot manipulator, modelling carbon consumption as a function of other process variables in a metallurgic industrial process, and estimating water content in fish food products based on NIR spectrometry. In addition the methods have been applied for modelling data generated from a simulation of a chemical catalytic reactor, and to identify a 10-dimensional synthetic function from small samples of noisy data. Quantitative comparisons are made of the prediction accuracy in separate test data sets, and qualitative comparisons are made for some important performance characteristics, such as computation time and simplicity of use.