|
Abstract Aiming at the prediction shortcomings of the traditional nonlinear modeling methods，including partial least-squares regression(PLS)，artificial neural network(ANN) and support vector machines(SVM) using probability-integral method，a novel method named multi-scale kernel partial least-squares regression(multi-scale KPLS) is proposed to identify parameters of the probability-integral method. Firstly，an admissible multi-scale Gaussian kernel function is constructed. Secondly，fuzzy clustering is applied to determine the optimal number of categories，which is regarded as scale parameter，and then，all kernel widths are optimized by 10 times 10-fold cross-validation and grid search method. Finally，the modeling process was discussed detailedly. Contrasting the prediction results of multi-scale KPLS with other methods of PLS，RBF neural network(RBF-NN) and SVM respectively，it shows that the former?s prediction accuracy is obviously better than the others? because of considering the characteristic of multi-scale in modeling samples；multi-scale KPLS has a stronger robustness and efficiently overcomes the multicollinearity between factors effecting on prediction results disadvantageously；multi-scale KPLS is suitable for parameters identification of probability-integral method with several induced variables versus several independent variables and its parameters could be determined by self-adaptive，so by terms of modeling efficiency，it is better than RBF-NN and SVM.
|
|
Received: 28 October 2010
|
|
|
|
2011, Vol. 30 
