Simulation metamodelling with Bayesian networks

Abstract

This paper introduces the use of Bayesian networks (BNs) as an exploratory metamodelling tool for supporting simulation studies conducted with stochastic simulation models containing multiple inputs and outputs. BN metamodels combine simulation data with available expert knowledge into a non-parametric description of the joint probability distribution of discrete random variables representing simulation inputs and outputs. The distributions of the inputs are determined based on expert knowledge and/or a real-world data source while the conditional distributions of the outputs are estimated from the simulation data. The exploratory use of the BN metamodels is an iterative process including the construction and validation of the BNs and allowing various analyses dealing with the dependencies among the inputs and the outputs, input uncertainty, and inverse reasoning. The results of these analyses are applied to guide and aid the utilization and interpretation of the simulation model under consideration. In addition, the analyses are used for studying the behaviour of the simulated system. The exploratory use is illustrated with an example involving a simulated queue.