Robust minimax design from costly simulations.

Abstract : Design of complex physical systems most often relies on numerical simulations that may be extremely costly. In this paper, design is formalized as the optimization of a performance index with respect to control variables. Uncertainty is modeled via a vector of environmental variables that can take any value in a known compact set and may have an adverse effect on performance. In this context, the determination of a robust design requires the continuous minimax optimization of black-box functions. An algorithm combining Kriging-based optimization with relaxation is presented, which makes it possible to find approximate solutions to such problems on a limited computational budget. The design of a vibration absorber is presented as an illustrative example.