Mid-Range Approximations in Sub-Spaces for MDO Problems with Disparate Discipline Attributes
Jonathan Ollar, VASSILI V. TOROPOV, Royston Jones · 2014
In the metamodel-based design optimization the dimensionality of the space in which a metamodel is built directly affects the quality of the metamodel achieved for a given computational budget, i.e. a certain number of data points. Because of such a consequence of the “curse of dimensionality”, design variable ranking is often used to only include variables that affect the responses of interest most. In a multi-disciplinary problem, various disciplines may depend on different subsets of the design variable space. In this paper an approach to building metamodels in sub-spaces while running the optimization problem in the full space is proposed. This technique is implemented using the Moving Least Squares metamodels within a trust region based optimization scheme known as the Multipoint Approximation Method and demonstrated by an analytical example and a small-scale FE test example.