An investigation of the behavior of simulation response surfaces

This paper is part of a research stream whose purpose is to study the effect of simulation response surface behavior on the choice of appropriate simulation optimization search technique. This paper's research lays some groundwork by examining the behavior of simulation response surfaces themselves. The point here is not to criticize existing simulation-optimization techniques (such as Response Surface Methodology (RSM)). Rather, one point is to emphasize the care and precision that must be used to invoke extant procedures properly, while another is to demonstrate the need for additional methods such as nonparametric approaches. In particular, this paper examines a simple, inventory-simulation model under various experimental conditions, including some factors under a user's control, and some not. Both point and region estimates of surface characteristics are determined and graphed while such factors as number of replications, simulation run length, and demand and lead-time variances are varied. It is found, for example, that even for this simple surface such optimization techniques as first-order RSM can be inappropriate over 21–98% of the feasible region, depending on the case. Four implications are noted from the research: the care that should be exercised with existing simulation-optimization techniques; the need for a simulation-optimization starter; the importance of examining global, nonparametric-metamodeling approaches to simulation optimization; and the desirability of investigating a multi-strategy approach to optimization. The paper concludes with a call for further research investigating these suggestions.