Selective ensemble simulate metamodeling approach based on latent features extraction and kernel learning

Simulate metamodeling technique based on complex physical model is one of the key methods to improve simulation effectiveness and assist decision maker to cognize behavior of complex system. There are strong collinearity among input variables of the simulate metamodel. Moreover, number of the training samples is also limited. Popularly used back propagation neural network (BPNN) model suffers from lower learning speed and over-fitting problems. Although random vector functional-link (RVFL) networks have faster learning speed, its prediction performance stability isn't satisfied for modeling little sample data. Thus, a new selective ensemble simulates meta-modeling approach based on latent features extraction and kernel learning is proposed. Partial least square (PLS) is used to extract the latent features from original input variables, which can eliminate collinearity among inputs and simplify structure of the simulate metamodel. Kernel based RVFL (KRVFL) networks, selective ensemble learning and global optimization viewpoint based modeling parameters selection algorithms are used to improve prediction accuracy and satiability of the simulate metamodel. Simulate results show that the proposed approach is effective.