Selective ensemble kernel partial least squares method based on dual layer genetic algorithm optimization with its application

摘要

Selective ensemble (SEN) learning algorithm can improve generalization performance of regression model. How to select SEN models' leaning parameters is an important issue. In this paper, based on kernel partial least squares (KPLS) constructed candidate sub-models, a new SENKPLS modeling method is proposed using dual layer genetic algorithm (GA) optimization. At first, adaptive GA (AGA) is used to produce the initial solutions of the local SEN models' learning parameters. Then, bootstrap production strategy-based training sub-samples are used to construct candidate sub-models with the initial learning parameters for each population. Thirdly, the ensemble sub-models are selected and combined with GA optimization toolbox (GAOT) and adaptive weighting algorithm (AWF). Thus, the local SEN models of all AGA populations are built. Finally, prediction performance of these local SEN models are used as fitness, and the selection, crossover and mutation process are repeated until satisfy the pre-set stopping criterion. Simulations based on the benchmark dataset of concrete compressive strength show that the proposed method is effective in terms of prediction accuracy.