Simulation Metamodeling Using Dynamic Bayesian Networks with Multiple Time Scales

Abstract

The utilization of dynamic Bayesian networks (DBNs) in simulation metamodeling enables the investigation of the time evolution of state variables of a simulation model. DBN metamodels have previously described the changes in the probability distribution of the simulation state by using a time slice structure in which the state variables are described at common time instants. In this paper, the novel approach to the determination of the time slice structure is introduced. It enables the selection of time instants of the DBN separately for each state variable. In this way, a more accurate metamodel representing multiple time scales of the variables is achieved. Furthermore, the construction is streamlined by presenting a dynamic programming algorithm for determining the key time instants for individual variables. The construction and use of the DBN metamodels are illustrated by an example problem dealing with the simulated operation of an air base.