9.2.6.   Conclusions

Similar to the previous chapters, it has been demonstrated that both, the genetic algorithm framework and the parallel growing neural network framework are superior to the common methods of multivariate calibration in terms of calibration quality. Both frameworks allow a reproducible variable selection resulting in smaller optimized neural network models with a better generalization ability compared with common neural networks. Both frameworks show reproducible and robust results whereby the parallel growing neural network framework seems to be more robust and shows a slightly better generalization ability, which has also been observed for the binary mixtures of the refrigerants. The classical chemometric methods like the PLS and INLR once more have problems to model the nonlinear relationships of the data similar to chapter 6. It was also shown that the PCA-NN could not deal with changes of the noise in the data resulting in significant biases of the predictions.