Simulation optimization methods that combine multiple comparisons and genetic algorithms with applications in design for computer and supersaturated experiments
暂无分享,去创建一个
The central issue in this dissertation is the developing relationship between the studies of simulation optimization and of experimental design. We begin by using simulation to evaluate and compare experimental design and analysis techniques for computer experiment meta-modeling. We used two test functions to compare all combinations of four experimental design classes with either second-order response surface (RS) or kriging modeling methods. The findings included for the first test function, any conclusion could be reached about which combination achieved the best prediction accuracy based on the experimental design used to represent each class. Also, we tentatively concluded that, for cases in which the number of runs is comparable to the number of terms in a quadratic polynomial, similar prediction errors can be expected from both kriging and regression modeling procedures as long as regression is used in combination with designs generated to address bias errors.
Next, we propose heuristics for simulation optimization based on combining genetic algorithms with elitist reproduction and comparisons based on confidence intervals. We use eight test functions to compare the proposed method with alternatives from the literature and conclude that, among general-purpose stochastic solvers, the proposed class of algorithms is promising. We also define stochastic equivalence to relate cases in which only a finite number of Monte Carlo simulations are feasible with imagined cases in which an infinite number of simulation are possible. We apply results from statistical selection and ranking and multiple comparison with the best to estimate the sample size requirements for stochastic equivalence to be achieved.
The main application of the proposed heuristic is the development of so-called “supersaturated” experimental designs. These designs are used for identifying the important inputs or factors in situations in which the number of factors is larger than the number of runs. Using simulation optimization we are able to evaluate and derive experimental designs from more realistic assumptions and more relevant objectives than was possible previously. We use hierarchical prior assumptions from the design literature that include the possibility of interactions between factors. Also, we directly optimize the probability of success in selecting important factors, which is new.