Multiple meta-models based design space differentiation method for expensive problems

Meta-models and meta-models based global optimization methods have been commonly used in design optimizations of expensive problems. In this work, a multiple meta-models based design space differentiation (MDSD) method is proposed. In the proposed method, an important region will be constructed using the expensive points inside the whole design space. Then, quadratic function (QF) will be employed in the search of the constructed important region. To avoid the local optima, kriging is employed in the search of the whole design space simultaneously. The MDSD method employs different meta-models in the different design space instead of space reduction, which preserves the advantages of high efficiency of the space reduction methods and avoids their shortcomings of removing the global optimum by mistake in theory. Through extensive test and comparison with three meta-model based algorithms, efficient global optimization (EGO), Mode-pursuing sampling method (MPS) and hybrid and adaptive meta-modeling method (HAM) using several benchmark math functions and an engineering problem involving finite element analysis (FEA), the proposed method shows excellent performance in search efficiency and accuracy.