Optimization of simulation models using successive quadratic approximations

Abstract In this paper the development of a search procedure, successive quadratic approximations, and its application to the optimization of systems modeled through simulation are presented, where the measure of system performance is cost. Since random error is usually inherent in simulation, and since replication is one means of reducing the effect of random error, the number of replications required to estimate the measure of system effectiveness at each point evaluated in the search is specified in conjunction with statistical criteria which may be assigned by the user. The search is terminated when specified statistical tests indicate that continuation of the search is likely to improve the system less than the cost of computer time for the next iteration, that is, when a point of diminishing returns is reached.