Multi-Kernel Partial Least Squares Regression based on Adaptive Genetic Algorithm

Kernel learning has been a focus of machine learning domain recently. Kernel partial least squares (KPLS) algorithm can construct nonlinear model using extract latent variables from the input and output data space simultaneously. However, generalization performance of KPLS model relies mostly on kernel types and kernel parameters, which are difference to modeling of different applicable background. Intelligent optimization algorithm can be used to search these parameters. Thus, a new multi-kernel partial least squares regression approach based on linear multi-kernel construction method and adaptive genetic algorithm (AGA) is proposed in this paper. Normally used global and local kernels are weighed to obtain the mixed multi-kernel of KPLS algorithm. These kernel's parameters and weighting coefficients are selected using AGA optimization algorithm. The experimental results based on Benchmark data set show that the proposed approach has better prediction performance than that of single kernel based modeling method.