Effect of pseudo datasets for the classification-based engineering design

Machine learning classification techniques have been used widely to recognize the feasible design domain and discover hidden patterns in engineering design. An accurate classification model needs a large dataset; however, generating a large dataset is costly for complex simulation-based problems. After training by a small dataset, surrogate models can generate a large pseudo dataset efficiently. Errors, however, may be introduced by surrogate modeling. This paper investigates the mutual effect of a large pseudo dataset and surrogate modeling uncertainty. Four widely used methods, i.e., Naive Bayes classifier, support vector machine, random forest regression, and artificial neural network for classification, are studied on four benchmark problems. Kriging is used as the basic surrogate model method. The results show that a large pseudo dataset improves the classification accuracy, which depends on both design problems and classification algorithms. Except for the Naive Bayes, the other three methods are recommended for using pseudo data to improve classification performance. Also, a wind turbine design problem is used to illustrate the effect of the pseudo dataset on feasible subspace recognition. The large pseudo dataset improves the recognized subspace bound greatly, which can be reproduced by classification models well except for the Naive Bayes classifier. Under the uncertainty of surrogate modeling, the random forest presents high robustness to recognize the feasible design domain bound, while the artificial neural network demonstrates a high sensitivity to uncertainty with the recognized bound deteriorated.