Cost-aware Adaptive Sampling for Global Metamodeling Using Voronoi Tessellation

Metamodeling has become a common approach to replace costly and time-consuming physical experiments or computer experiments (e.g., numerical simulation, training of AI models) by an easy-to-evaluate metamodel, which is trained on samples of the experiment. Since the choice of sample points significantly impacts model accuracy, these are in many cases determined using adaptive sampling methods. In addition, the cost of conducting an experiment often depends decisively on the choice of its parameters. However, only few strategies for selecting the sample points have been proposed, that take into account parameter-dependent costs. In this work, we introduce a novel Voronoi-based cost-aware adaptive sampling algorithm for global metamodeling that is independent of the choice of sampling strategy and metamodel. The method is evaluated on a variety of randomly generated black-box and cost functions, where it has shown to vastly outperform existing sampling strategies.