Flexible space-filling designs for complex system simulations

Abstract

In order to better understand the complex nature of a system, analysts need efficient experimental designs that can explore high-dimensional simulation models with multiple outputs. These simulation models are critical to the early phases of system design and involve complicated outputs with a wide variety of linear and nonlinear response surface forms. The most common response surface form for analyzing complex systems is the second-order model. Traditional designs that fit second-order response surface models do not effectively explore the interior of the experimental region and cannot fit higher-order models. We present a genetic algorithm that constructs space-filling designs with minimal correlations between all second-order terms for a mix of continuous and discrete factor types. These designs are specifically suited to fit the second-order model with excellent space-filling properties and are flexible enough to fit higher-order models for a modest number of factors; these high-order terms are what characterize the system complexities. We demonstrate the utility of these designs with a Model-Based Systems Engineering application that integrates multiple simulation outputs to form a trade-off environment for a system design. This research enables the simulation analysis and system design community to better understand the complex nature of multiple simulation outputs.