A HETEROSCEDASTIC T-PROCESS SIMULATION METAMODELING APPROACH AND ITS APPLICATION IN INVENTORY CONTROL AND OPTIMIZATION

In this paper we develop a heteroscedastic t-process metamodeling approach (TP) for approximating the mean response surface implied by a stochastic simulation and performing metamodel-based optimization. We provide details on how to construct a TP metamodel, make inference and perform prediction based on TP. We show that TP can retain the attractive properties of approaches that rely on Gaussian processes, but it also enjoys enhanced flexibility, at no additional computational cost. We further provide a closedform expression for the TP-based expected improvement to perform metamodel-based optimization. We compare the predictive performance of TP and stochastic kriging (SK) via an M/M/1 inspired example, and demonstrate the performance of TP and SK-based algorithms for optimizing a simple periodic review (s, S) inventory system. The preliminary numerical evaluations indicate that TP can serve as a promising simulation analytics tool for solving problems encountered in production planning and supply chain management.