Adaptive Sampling for Global Meta Modeling Using a Gaussian Process Variance Measure

Adaptive sampling methods have been widely used for meta modeling, when dealing with expensive-to-evaluate experiments in practical design and optimization tasks to approximate their performance measure. Existing methods try to minimize the global model error in various ways. However, there is still no general method for the denser sampling of regions with high performance. In this work, we introduce a new adaptive sampling approach that samples regions of high performance more densely, while also exploring unseen regions. A Gaussian process is used as meta model and a variance-based measure is defined for computing the adaptive sample points. Furthermore, Voronoi tessellation is used to reduce the complexity for application in high-dimensional design spaces. The proposed approach allows for higher model accuracy in regions of high performance by efficiently placing the available samples.