The construction and implementation of metamodels

Abstract

This paper suggests a computationally economical alternative to the common method of using simulation for sensitivity studies, i.e., by making repeated simulation runs with incrementally changed inputs and then observing the magnitudes of the resulting changes in the outputs of interest. Although no theory now exists for constructing the requisite "metamodels" for direct sensitivity analysis, it is possible to construct them, as is illustrated by three examples in the field of operations research.