ASYMMETRIC KRIGING EMULATOR FOR STOCHASTIC SIMULATION

In many situations, e.g., simulation optimization and input uncertainty quantification, we need to assess the system performance at a large number of alternative inputs. Since each simulation run could be computationally expensive, statistical emulator could efficiently use the simulation budget to estimate the system performance. This paper proposes a new emulator for stochastic simulation, called asymmetric kriging (AK), which can be used to emulate the distribution of simulation outputs at each input point. Different from existing methods in the simulation literature, our approach does not require strong assumptions on either the functional form of the response surface or the normal distribution of the simulation estimation error. Numerical studies indicate the efficacy of our approach compared to alternative methods in the literature.