A comparison of modeling nonlinear systems with artificial neural networks and partial least squares

Artificial neural networks (ANN) can be used to model nonlinear and noisy calibration systems. Models of such systems can also be made by partial least squares (PLS) regression after linearization of the data. These different models and their predictive properties have been tested. The data used are measurements of inorganic and organic air pollutants, solar light intensity, temperature, and corresponding ozone (O3) concentrations. The total data set sizes are: 710×57 and 710×10 for X and 710×1 for y. The large number of objects permits splitting the data into calibration and test sets. The orthogonality properties of the derived linear and nonlinear functional basis sets are investigated. This investigation shows that certain aspects of latent variable based linear modeling can be transferred to the ANN models. Nonlinear neurons can be linearized after the training iterations have been completed. The use of this mixed approach permits the development of additional understanding of the nature of the basis set expansion that is used in the typical neural network (NN). This approach also avoids overfitting and appreciably improves the predicted results.