Design of Experiments and Metamodeling for Stochastic Simulation

Computer simulation of complex systems is a widely accepted method to obtain highly accurate estimates of performance measures. However, detailed modeling of the real system and experimenting at a large number of scenarios typically take a long time. In this thesis, we design support tools for decision-making via simulation to produce accurate predictors with reasonable simulation investment. These predictors provide results with the fidelity of a detailed simulation model and the "what if" capability of a simple static model or a queueing approximation. The first part of the thesis presents a methodology called simulation on demand that uses well--known models from queueing theory as the basis of the predictors for the mean steady-state cycle time of products in a manufacturing system, created from a limited but adaptively selected set of simulation runs. In the second part of the thesis, we build predictors using a new modeling procedure for simulation on demand called stochastic kriging, that is applicable where the form of the response surface model is unknown. We propose design strategies that guide sequential algorithms with and without adaptation to the data to make allocation and stopping decisions such that a prespecified relative precision is realized with some confidence. The third part describes how simulation can be used in the concept--development phase of a new production line. We illustrate that when little information is available about the real system, and details are subject to frequent change creating predictors is not the appropriate analysis technique.