Gradient Based Criteria for Sequential Experiment Design

Our work will study the use and effectiveness of gradient based criteria for developing metamodels that can be used for nearly instantaneous estimation of a complex computer simulation output at any set of inputs. The goal of the criteria is to have the metamodel be most accurate in the most important areas in the input space of the simulation. Gradient methods focus on areas where the output of the surface is changing quickly and these are often the areas of most interest to decision makers since small changes in the inputs can result in dramatically different outputs. Guided by a criterion based on the gradient, the sequential experiment will build up the accuracy of the metamodel in the most relevant areas allowing for decision makers to get results for the areas of interest more quickly and more accurately.