Neural networks for the metamodeling of simulation models with online decision making

Abstract

Abstract We present a methodology for an artificial neural network (ANN) based metamodeling of simulation models in the special case when online decision making routines are invoked repetitively by the simulation model throughout the simulation run. For a practitioner, the benefit of such a simulation metamodel lies in the possibility to compare different decision making routines (operational control strategies) without excessive computational time for running multiple simulation configurations with different control strategies. Contrasting to the conventional setting of ANN based simulation metamodeling, in this paper ANNs have to take as input not only numerical parameters, but also different control strategies. The methodology is finally put into practice in a case study of an order picking system. Results show that on average the relative error of the ANN metamodel is fairly acceptable and allows for a first assessment of system parameters and control strategies. However, also large outliers in the relative error are encountered. Hence, for the analysis of different parameters and control strategies in applications, the use of ANN-based simulation metamodels requires an accompanying statistical assessment of ANN-based performance prediction accuracies.