Experimental planning and sequential kriging optimization using variable fidelity data
Deng Yang Huang · OhioLink ETD Center (Ohio Library and Information Network) · 2005
Engineers in many industries routinely need to improve the product or process designs using data from the field, lab experiments, and computer experiments.Historically, designers have performed a calibration exercise to "fix" the lab system or computer model and then used an analysis method or optimization procedure that ignores the fact that systematic differences between products in the field and other environments necessarily exist.A new line of research is not based on the assumption that calibration is perfect and seeks to develop experimental planning and optimization schemes using data form multiple experimental sources.We use the term "fidelity" to refer to the extent to which a surrogate experimental system can reproduce results of the system of interest.For experimental planning, we present perhaps the first optimal designs for variable fidelity experimentation, using an extension of the Expected Integrated Mean Squared Error (EIMSE) criterion, where the Generalized Least Squares (GLS) method was used to generate the predictions.Numerical tests are used to compare the method performance with alternatives and to investigate the robustness to incorporated assumptions.The method is applied to automotive engine valve heat treatment process design in which real world data were mixed with data from two types of computer simulations.Dr. Theodore T. Allen, for their inspiration, encouragement and support which made this thesis possible, and for their patience in correcting both my scientific and stylistic