7.4.   Genetic Algorithm Framework – Conclusions

In this chapter, a genetic algorithm was used for the optimization of neural networks by variable selection. It was shown that single runs of the GA are faced with irreproducible variable selections and with unstable predictions of the external validation data, which were even worse than the predictions by non-optimized neural networks. A genetic algorithm framework was suggested, which uses repeated runs of the GA. The predictions of the neural networks optimized by this framework were superior to the commonly used non-optimized neural networks and to all other calibration methods used before. It was also shown that the variable selection is reproducible and not subject to noise. Addi­tionally, it was demonstrated that the predictions are hardly influenced by interfering analytes rendering the combination of time-resolved measurements, the genetic algorithm framework for the variable selection and for the calibration open to the parallel quantification of more analytes using only one single sensor. The unique framework introduced in this works is not restricted to the calibration and variable selection of sensor signals, but can be used for the variable selection and multivariate calibration of virtually any data set as long as a sensible number of data is available.

table 3:      Comparison of the rel. RMSE of the calibration and validation data in % for the genetic algorithm approaches and the non-optimized neural networks used in section 6.8.