A parameterized lower confidence bounding scheme for adaptive metamodel-based design optimization

Abstract

Purpose The purpose of this paper is to efficiently use as few sample points as possible to get a sufficiently explored design space and an accurate optimum for adaptive metamodel-based design optimization (AMBDO). Design/methodology/approach A parameterized lower confidence bounding (PLCB) scheme is proposed in which a cooling strategy is introduced to guarantee the balance between exploitation and exploration by varying weights of the predicting error and optimum of a metamodel. The proposed scheme is investigated by a set of test functions and a structural optimization problem, in which PLCB with four kinds of cooling control functions are studied. Moreover, other infill criteria (such as expected improvement and its extension versions) are taken into comparison. Findings Results show that the proposed PLCB (especially PLCB with the first cooling control function) based AMBDO method can find the optimum with fewer evaluations and maintain good accuracy, which means the proposed PLCB contributes to the excellent efficiency and accuracy in finding global optimal solutions. Originality/value The parameterized version of the lower confidence bound metric is proposed for AMBDO, typically used in the context of adaptive sampling in efficient global optimization.