Discipline-Specific Trust-Region Sizes for Variable-Fidelity Multidisciplinary Optimization

摘要

Trust-region model-management methods (TRMM) are an efficient, provably convergent technique for solving design optimization problems. They use multiple fidelities, and converge to a local minimum of the high-fidelity space while performing the majority of the analysis using the low-fidelity problem, thus allowing significant computational savings. Until now, TRMMs have treated multidisciplinary problems as a single, monolithic problem and used a single trust region size. This paper presents an extension of the traditional trust region algorithm to allow a multidisciplinary trust-region algorithm to scale the dimensions of the trusted region independently for the variables associated with each discipline. The algorithm is tested on two synthetic multidisciplinary problems and compared to an active set algorithm. The results show that the new method performs at least as well as the state-of-the-art methods and is therefore a promising direction of research for multidisciplinary engineering optimization problems.