Selective ensemble simulation meta-modeling approach based on multiple kernel features extraction and fuzzy inference

To construction effective simulation meta-models for complex physical simulation system, the “curse of dimension” and the “uncertain and imprecise information” problems have to be addressed firstly. Although simulation meta-models based on neural networks can obtain well performance, the fuzzy inference mechanism of domain expert for practical application problems cannot be simulated. Thus, some prediction results may be contradicted with that of the actual physical systems. Aim at these problems, a novelty selective ensemble (SEN) simulation meta-modeling approach based on multiple kernel feature extraction and fuzzy inference modeling is proposed in this paper. A new ensemble construction method based on multiple candidate kernel latent features is used to produce candidate training sub-samples. Candidate sub-models based on Mamdani's fuzzy inference are constructed with these training sub-samples. Brand and bound (BB) and adaptive weighting fusion (AWF) algorithms are used to select and combine the ensemble sub-models to obtain final SEN simulation meta-model. Simulate results based on synthetic data show that the proposed approach is effective.