다중 응답 시스템의 변수 선별을 위한 개선된 이웃성분 특징 선택

초록

In optimization problem, as the number of design variables increases, the convergence rate of optimization tends to decrease. Therefore, the number of function calls increases and time and cost increase. Variable screening techniques such as analysis of variance (ANOVA) are often used to select variables which have large influence on responses. However, ANOVA could take extra time and cost by additional analyses or experiments because the data must satisfy orthogonality. Therefore, previous study implemented neighborhood component feature selection (NCFS) which can select significant design variables regardless of the orthogonality of data. However, NCFS may have inconsistent results depending on the hyperparameter in the formulation. In addition, it may be difficult to make a fair comparison in calculating the variable screening measure of multiple response system because the orders of the weight for each response are different. This study aims to solve the problems addressed before by modifying the formulation of previous NCFS. Mathematical example is used to compare the results of previous method with proposed method to verify its effectiveness.

키워드