7.    Results – Genetic Algorithm Framework

In the previous chapter, it was demonstrated that neural networks show the best calibration for the refrigerant data set due to the nonlinearities present in the data. It was also shown that a compression of the input variables by a simple combination of a PCA and neural networks shows results comparable with the neural networks using all variables. Therefore, it is expected that a more sophisticated selection of the input variables can improve the generaliza­tion ability and thus the calibration quality. Hence, a genetic algorithm for the variable selection is combined with neural networks for the calibration in this chapter. As single applications of this combination neither show superior calibrations nor repro­ducible variable selections, a framework is setup, which uses many parallel runs of the genetic algorithm for different data subsets resulting in improved calibrations and a high reproducibility.