MULTIPLE MULTIVARIATE REGRESSION AND GLOBAL SEQUENCE OPTIMIZATION - AN APPLICATION TO LARGE-SCALE MODELS OF RADIATION INTENSITY

We investigate the strengths and weaknesses of several neural network architectures applied to a large-scale thermodynamical application in which sequences of measurements from gas columns must be integrated to construct the columns' spectral radiation intensity profiles. This is a problem of interest for the aeronautical industry. The approaches p roposed for its solution can be applied to a wide range of signal prob lems. Physical models often make use of a number of fitted functions a s a simplified parametric base to approximate a high-dimensional nonli near (and usually computationally intractable) function. Realistically , models of radiation contain thousands of fitted functions. The use o f neural networks in applications of this scale are rare, and most eff ective techniques rely on cross-validation methods or involve other he avy computational overheads that are impracticable when a very large n umber of models need to be trained. We have employed here two differen t approaches: multiple multivariate regression, and global sequence mi nimization. The first approach shows that the integration of several n onlinear regression models into a single neural network may improve bo th generalization performance and speed of computation. For the former we propose a method of optimization by which we specialize our models globally, on typical sequences of input signals. We show how this doe s not degrade the overall accuracy but, rather, allows us to specializ e our models. (C) 1998 Elsevier Science B.V. All rights reserved.