Robust parameter optimization strategies in computer simulation experiments

Abstract

An important consideration in computer simulation studies is the issue of model validity, the level of accuracy with which the simulation model represents the real world system under study. This dissertation addresses a major cause of model validity problems: the dissimilarity between the simulation model and the real system due to the dynamic nature of the real system that results from the presence of nonstationary stochastic processes within the system. This transitory characteristic of the system is typically not addressed in the search for an optimal solution.

In reliability and quality control studies, it is known that optimizing with respect to the variance of the response is as important a concern as optimizing with respect to average performance response. Genichi Taguchi has been instrumental in the advancement of this philosophy. His work has resulted in what is now popularly known as the Taguchi Methods for robust parameter design. Following Taguchi's philosophy, the goal of this research is to devise a framework for finding optimum operating levels for the controllable input factors in a stochastic system that are insensitive to internal sources of variation. Specifically, the model validity problem of nonstationary system behavior is viewed as a major internal cause of system variation.

In this research the typical application of response surface methodology (RSM) to the problem of simulation optimization is examined. Simplifying assumptions that enable the use of RSM techniques are examined. The relaxation of these assumptions to address model validity leads to a modification of the RSM approach to properly handle the problem of optimization in the presence of nonstationarity. Taguchi's strategy and methods are then adapted and applied to this problem. Finally, dual-response RSM extensions of the Taguchi approach separately modeling the process performance mean and variance are considered and suitably revised to address the same problem.

A second cause of model validity problems is also considered: the random behavior of the supposedly controllable input factors to the system. A resolution to this source of model invalidity is proposed based on the methodology described above.