Fuzzy local linear approximation-based sequential design

Abstract

When approximating complex high-fidelity black box simulators with surrogate models, the experimental design is often created sequentially. LOLA-Voronoi, a powerful state of the art method for sequential design combines an Exploitation and Exploration algorithm and adapts the sampling distribution to provide extra samples in non-linear regions. The LOLA algorithm estimates gradients to identify interesting regions, but has a bad complexity which results in long computation time when simulators are high-dimensional. In this paper, a new gradient estimation approach for the LOLA algorithm is proposed based on Fuzzy Logic. Experiments show the new method is a lot faster and results in experimental designs of comparable quality.