Feature-Based Constrained Modeling

Abstract

Feature-based modeling exploits a particular structure of the system responses in order to further reduce the computational cost of setting up the surrogate. Representative examples are multiband antennas, whose input characteristics feature sharp and easily distinguished resonances. Reformulating the design problem in terms of the characteristic points of the system responses, such as the frequency and level allocations of the resonances for the mentioned case of multiband antennas, permits “flattening” the functional landscape to be handled. Constructing the model of the feature point coordinates rather than of the original responses (typically, frequency characteristics) requires a fewer number of training data samples. At the same time, the information carried by such a surrogate is sufficient to apply it for design purposes. This chapter demonstrates how the nested kriging method can be combined with the response feature approach to reduce further the computational cost of surrogate construction.