Computing confidence intervals for stochastic simulation using neural network metamodels

Abstract

This paper discusses the use of supervised neural networks as a metamodeling technique for discrete-event, stochastic simulation. An (s, S) inventory simulation from the literature is translated into a metamodel through development of parallel neural networks, one estimating expected total cost and one estimating variance of expected total cost. These neural network estimates are used to form confidence intervals, which are compared for coverage to those formed directly by simulation. It is shown that the neural network metamodel is quite competitive in accuracy when compared to the simulation itself and, once trained, can operate in nearly real-time. A comparison of metamodel performance under interpolative versus extrapolative predictions is made.