Initial sampling methods in metamodel-assisted optimization

The modern engineering design process often relies on numerical analysis codes to evaluate candidate designs, a setup which formulates an optimization problem which involves a computationally expensive black-box function. Such problems are often solved using a algorithm in which a metamodel approximates the true objective function and provides predicted objective values at a lower computational cost. The metamodel is trained using an initial sample of vectors, and this implies that the procedure by which the initial sample is generated can impact the overall effectiveness of the optimization search. Approaches for generating the initial sample include the statistically based design of experiments, and the more recent search-driven sampling which generates the sample vectors with a direct-search optimizer. This study compares these two approaches in terms of their overall impact on the optimization search and formulates guidelines in which scenario is each approach preferable.