Distribution-free bootstrapping of the replicated responses of a given discrete-event simulation model gives bootstrapped Kriging (Gaussian process) metamodels; we require these metamodels to be either convex or monotonic. To illustrate monotonic Kriging, we use an M/M/1 queueing simulation with as output either the mean or the 90% quantile of the transient-state waiting times, and as input the traffic rate. In this example, monotonic bootstrapped Kriging enables better sensitivity analysis than classic Kriging; i.e., bootstrapping gives lower MSE and confidence intervals with higher coverage and the same length. To illustrate convex Kriging, we start with simulation-optimization of an (s, S) inventory model, but we next switch to a Monte Carlo experiment with a second-order polynomial inspired by this inventory simulation. We could not find truly convex Kriging metamodels, either classic or bootstrapped; nevertheless, our bootstrapped "nearly convex" Kriging does give a confidence interval for the optimal input combination.
INTRODUCTIONMany realistic simulation models have known characteristics such as convexity and monotonicity. For example, simulation models of supply chains consist of a sequence of submodels (building blocks, modules) for queues and inventories; higher traffic rates monotonically increase mean waiting time, and reorder levels and order quantities are often assumed to have a unique optimal combination because the cost function is convex (instead of having multiple local optima). However, in their classic textbook on convex optimization Boyd and Vandenberghe (2004) study problems with explicit functions, whereas simulation problems have implicit functions that are determined by the underlying simulation model. In this paper, we use a metamodel to approximate such an implicit function (also see Nesterov 2003, pp. 171-172).Metamodels (also called response surfaces, emulators, etc.) serve sensitivity analysis of the simulation models and optimization of the simulated systems. There are several types of metamodels, but the most popular types are linear regression analysis and Kriging (or Gaussian process) models; many references to various types of metamodels are given by Kleijnen 2008, p. 8. Well-known types of monotonic regression models are isotonic regression and "rank" regression; see Kleijnen 2008, pp. 98, 162. We, however, focus on Kriging. Monotonic Kriging metamodels are also examined by Kleijnen and van Beers (2011); we summarize and update that publication, and extend it to convexity.To estimate the Kriging metamodel, we simulate (say) n combinations (or points) x i of the k ≥ 1 simulation inputs; we replicate these combinations m i times (i = 1, ..., n) We assume that the simulation model is expensive; i.e., the simulation requires much computer time to obtain the outputs w i;r (r = 1, ..., m i ), so the set of input/output (I/O) data may be so small that "classic" Kriging does not preserve the assumed characteristic, and shows wiggling (erratic) behavior. We therefore derive bootstrapp...