Normalized neighborhood component feature selection and feasible-improved weight allocation for input variable selection

In design optimization, as the number of input variables increases, the convergence rate of optimization tends to decrease, and the number of function calls and design change costs tend to increase. Neighborhood component feature selection (NCFS) was adopted to select significant input variables. However, the parameter determination process of the NCFS incurs a high computational cost and weakens robustness. Therefore, this study proposes a normalized NCFS (nNCFS) by normalizing scales between mean loss and regularization terms via the initial dataset Additionally, in the case of a multi-response system, complex decision-making processes that involve the allocation of weights for multiple responses are required. It is possible to allocate weights by using conventional methods such as the analytic hierarchy process and entropy methods. However, the analytic hierarchy process method is highly influenced by the designer's subjectivity, and the entropy method is unable to consider a design optimization problem. Accordingly, the feasible-improved weight allocation (FIWA) method is now proposed by considering a design optimization problem objectively. Comparing the NCFS with the nNCFS through mathematical examples, we found that the nNCFS significantly improved the computational cost and robustness. Moreover, the FIWA method selected significant input variables that yielded feasible and improved designs. Then, the nNCFS and the FIWA methods were applied to the design of the body-in-white of a vehicle. The significance of input variables was analyzed using the nNCFS, and feasible and improved designs were obtained on the basis of the significant input variables selected using the FIWA method. (c) 2021 Elsevier B.V. All rights reserved.

Automated summary

The study proposed the normalized NCFS (nNCFS) and feasible-improved weight allocation (FIWA) methods, which significantly improved the computational cost and robustness of design optimization, selected significant input variables, and obtained feasible and improved designs.