<![CDATA[The problems in prediction of calibration model are multicolinearity and the number of variables is larger than the number of observations. Principal Component Analysis-Artificial Neural Network-Genetic Algorithm (PCA-ANN-GA) models were applied for the relationship between sample of concentration which is limited and transmittance data which is in large dimensions. A large number of variables were compressed into principal components (PC’s). From these PC’s, the ANN was employed for prediction of concentration. The principal components computed by PCA were applied as inputs to a backpropagation neural network with one hidden layer. The models was evaluated using GA for the best network structure on hidden layer. Root Mean Square Error (RMSE) for 80% training set and 20% testing set are 0.0314 and 0.5225, respectively. Distribution of data according to the percentage of training data and testing data were also very influential to obtain the best network structure with the smallest RMSE achievement. The best model for these methods is two layers Neural Network with eight neuron in the hidden layer.]]>
Copyrights © 2011
