Computationally Efficient Adaptive Design of Experiments for Global Metamodeling through Integrated Error Approximation and Multicriteria Search Strategies

Abstract

Gaussian processes (GPs) are a popular technique for global metamodeling applications. Their objective in such settings is to establish an efficient and globally accurate approximation of the response surface of computationally expensive simulation models. When developing such GPs, the design of (simulation) experiments (DoE) plays an important role in reducing the required number of model runs for obtaining accurate approximations. Sequential (adaptive) selection of experiments can provide significant advantages, especially when the response surface is characterized by localized nonlinearities. Such adaptive DoE strategies for global metamodeling applications typically focus on minimizing the predictive GP variance, representing an exploration strategy, while recent developments have additionally considered the reduction of the GP bias obtained through cross validation, representing an exploitation strategy. While significant focus has been placed on the definition of appropriate adaptive DoE criteria, computational challenges still exist that limit the widespread adoption of adaptive DoE techniques—for example, related to the additional computational demand for identifying the optimal new experiment(s) or the necessity to establish proper schemes to combine exploration and exploitation strategies. To address these specific challenges, this research investigates two new adaptive DoE formulations. The first one focuses on the approximation of the popular integrated mean square error (IMSE) DoE criterion. The computationally demanding GP predictive variance update (after addition of each candidate experiment), required in the original IMSE formulation, is replaced by an approximation based on the current predictive variance and the domain of influence that surrounds each new experiment. The approximation is established through a parametric formulation that leverages the GP kernel to describe the aforementioned domain, with characteristics that are progressively calibrated across the GP training stages, to minimize the discrepancy between the actual and the approximated IMSE. The second formulation establishes a multicriteria search for simultaneously identifying multiple Pareto optimal experiments that balance exploration and exploitation objectives, replacing conventional strategies that establish a weighted combination of these objectives to promote a single DoE selection criterion.