Multiple Multivariate Regression And Global Optimization in a Large Scale Thermodynamical Application

We describe a large scale real-world application of neural networks for the modelization of heat radiation emitted by a source and observed through the atmosphere. For this problem, thousands of regressors need to be trained and incorporated into a single model of the process. On such large scale applications, standard techniques for the control of complexity are impossible to implement. We investigate the interest of i) integrating several regressors into a single neural network, and ii) refining the learned functions by optimizing simultaneously all regressors over a global function. The two approaches described offer a solution to these problems, and were crucial for the development of a fast and accurate model of radiation intensity.