Evolving Genetic Algorithm Applied To Variable Selection

Variable selection, which is used in finding effective factors for modeling from a great number of factors, has always been an important research topic. In variable selection, there are lots of methods with their own advantage and disadvantage, which are suitable in different problems.The algorithm we studied is evolving genetic algorithm (EGA). Combined with partial least squares (PLS), EGA is used for modeling and prediction in tobacco model. Good predicted results are received and controllability of tobacco model is enhanced. Results show that EGA is effective in dealing with such discrete and nonlinear model. EGA is further used in linear model. In order to validate the stability, reliability and correctness of EGA, spectrums simulated by Gauss algorithm are in use. When orthogonal degree is between 0.3 and 1.0, EGA can select decades of effective variables from model with about two hundred variables. Compared with the unprocessed model, better prediction result and higher efficiency are received.In order to compare with EGA, stepwise regression (SR) and partial correlation coefficient matrix (PCCM) are utilized in tobacco model. The average relative errors of EGA are respectively: nicotinamide 5.87%, tar 4.44%, CO 5.35%. The average relative errors of SR are respectively: nicotinamide 5.19%, tar 4.96%, CO 6.95%. The average relative errors of PCCM are respectively: nicotinamide 5.83%, tar 4.76%, CO 8.94%. Predicted results demonstrate that EGA is a reliable and effective method for variable selection.Besides, Kernel-PLS algorithm is realized in programming language java for chemometrics platform. The application of Kernel-PLS in quality control of wheat grain is also introduced in this paper.