On direct gradient enhanced simulation metamodels

Traditional metamodel-based optimization methods assume experiment data collected consist of performance measurements only. However, in many settings found in stochastic simulation, direct gradient estimates are available. We investigate techniques that augment existing regression and stochastic kriging models to incorporate additional gradient information. The augmented models are shown to be compelling compared to existing models, in the sense of improved accuracy or reducing simulation cost. Numerical results also indicate that the augmented models can capture trends that standard models miss.