Optimization strategy using dynamic radial basis function metamodel

Since global approximation accuracy of traditional static metamodel is difficult to be ensured and its computation efficiency is relatively low for flight vehicle multidisciplinary design optimization, the optimization strategy using dynamic radial basis function metamodel is proposed to overcome such defects above. The radial basis function metamodel is constructed with initial sampling points generated by Maximin Latin hypercube design method. Global optimization algorithm is employed to optimize current metamodel to find the potential global optimum of the true optimization problem, and then, the significant sampling space based on current known information can be identified. During optimization process, the new sampling points in the significant sampling space are added, and the metamodel is updated for the purpose of improving the approximation accuracy around the global optimum until the potential global optimum is satisfied the convergence conditions. The optimization strategy is validated by using a benchmark numerical test problem and the NASA speed reducer optimization design. As the optimization results shown, the global optimal solutions are obtained by using the dynamic metamodel optimization strategy. Compared with directly using genetic algorithm, 95% reduction on the number of function evaluations using dynamic metamodel is got. Compared with traditional static radial basis function metamodel, the number of function evaluations using dynamic metamodel is reduced by 50%.