Date of Award
Doctor of Philosophy (PhD)
Department of Operational Sciences
Jeffrey D. Weir, PhD.
Many Air Force studies require a design and analysis process that can accommodate for the computational challenges associated with complex systems, simulations, and real-world decisions. For systems with large decision spaces and a mixture of continuous, discrete, and categorical factors, nearly orthogonal-and-balanced (NOAB) designs can be used as efficient, representative subsets of all possible design points for system evaluation, where meta-models are then fitted to act as surrogates to system outputs. The mixed-integer linear programming (MILP) formulations used to construct first-order NOAB designs are extended to solve for low correlation between second-order model terms (i.e., two-way interactions and quadratics). The resulting second-order approaches are shown to improve design performance measures for second-order model parameter estimation and prediction variance as well as for protection from bias due to model misspecification with respect to second-order terms. Further extensions are developed to construct batch sequential NOAB designs, giving experimenters more flexibility by creating multiple stages of design points using different NOAB approaches, where simultaneous construction of stages is shown to outperform design augmentation overall. To reduce cost and add analytical rigor, meta-learning frameworks are developed for accurate and efficient selection of first-order NOAB designs as well as of meta-models that approximate mixed-factor systems.
DTIC Accession Number
Little, Zachary C., "Experimental Designs, Meta-Modeling, and Meta-learning for Mixed-Factor Systems with Large Decision Spaces" (2018). Theses and Dissertations. 1848.