Investigation of a two-phased strategy for simulation optimization

Computer simulation is a widely used analytical tool that permits the study of complex systems that cannot be modeled by other mathematical and statistical methods. Simulation optimization can determine the state of controllable inputs to a system that will cause system outputs to be at their most favorable or optimal condition which in turn will maximize system performance.; This research compares several nonlinear optimization methods for simulation optimization and investigates the utility of a two-phased optimization strategy. Optimization methods considered in this research are scatter search, genetic algorithms, evolution strategies, evolutionary programming, the pattern search method of Hooke and Jeeves, the simplex method of Nelder and Mead, and tabu search. A two-phased strategy is developed where globally oriented search methods are used for exploration to identify the region of the true optimum followed by a locally oriented search method for exploitation to better estimate the optimal solution. The utility of the two-phased strategy is compared to a single-phased strategy.; The Simulation Optimization Application Resource (SOAR) conducts both single and two-phased simulation optimization experiments to determine the best optimization strategies with respect to the miss distance from the observed optimum to the true global optimum. The research investigated simulation optimization over 48 experimental conditions that addressed variable type, number of available simulation evaluations, dimensionality of the solution vector, and level of noise present in the simulated system.; Recommendations based on experimental data suggest appropriate search methods and strategies for each of the 48 experimental conditions. Notable results include the recommendation of scatter search and tabu search as the preferred exploration and exploitation search methods for all high-dimensional problems. A genetic algorithm is the recommended exploratory search method for low-dimensional problems if higher numbers of simulation evaluations are available. Recommendations concerning preferred search methods are provided for the other experimental conditions. The use of a two-phased strategy was effective for half of the 48 experimental conditions to include all low-dimensional problems with low noise. Recommendations concerning preferred strategies are provided for the other experimental conditions.; Knowledge gained through this research has direct application for the development of tools to conduct simulation optimization.