Comparison of Neural Network and Kriging Method for Creating Simulation-Optimization Metamodels

Abstract

The intent of this study is to provide an initial exploration of the metamodeling capabilities of two methods, i.e. neural network (NN) and Kriging approximation, in the context of simulation optimization. A total of four performance measures are adopted, and they describe different kinds of metamodel performance, such as ability to provide good starting points for gradient-based search, accuracy of placing optima in the correct location and so on. With the help of the four measures, the performance of the two metamodeling methods is evaluated through the examination of two 2-D test functions. Both test functions have multiple local optima over the design space, and they are representative of the modeling challenges typically encountered in realistic simulation optimization problems. In the process of performance comparison, different empirical formulas are used to set the number of neurons in the hidden layer, while diverse correlation functions are adopted to create different kinds of Kriging metamodels. Preliminary research results reveal that Kriging approximation is in general likely to be preferred.