A1086
Title: Dimension reduction for Gaussian process models via convex combination of kernels
Authors: Lulu Kang - Illinois Institute of Technology (United States) [presenting]
Abstract: Many computer simulation models in engineering and scientific domains involve a high number of input variables, which can result in high computational costs and reduced prediction accuracy for the Gaussian process (GP) model. However, some simulation models may be influenced by only a small subset of the input variables, referred to as the active variables. Identifying these active variables can help researchers overcome the GP model's limitations and better understand the simulated system. To address this issue, a new approach for identifying the effective input dimensions of the GP model is proposed. Specifically, the covariance kernel function of the original GP model is approximated using a convex combination of kernels from lower-dimensional input dimensions. An iterative algorithm based on the Fedorov-Wynn algorithm from the optimal design literature was developed to determine the best approximation. The effect heredity principle while selecting the active input variables, which ensures that the subset of variables identified is sparse, is also incorporated. We demonstrate the effectiveness of the proposed method through several examples, showing that it outperforms some alternative approaches in correctly identifying the active input variables.
