RESPONSE SURFACE AND SENSITIVITY-BASED OPTIMIZATION IN LS-OPT: A BENCHMARK STUDY

This paper evaluates the robustness of LS-OPT for response surface and design sensitivity-based optimization. The methodology uses linear response surfaces constructed in a subregion of the design space. These are constructed using either a design of experiments approach with a D-optimal experimental design or the available analytical or numerical gradient. The approach utilizes a domain reduction scheme to converge to an optimum. The scheme requires only one user-defined parameter, namely the size of the initial subregion. To test its robustness, the results using the method are compared to SQP results of a selection of the well-known Hock and Schittkowski problems. Although convergence to a small tolerance is predictably slow when compared to SQP, LS-OPT does remarkably well for these, sometimes pathological, analytical problems.