Simulation metamodeling in continuous time using dynamic Bayesian networks

The application of dynamic Bayesian networks (DBNs) is a recently introduced approach to simulation metamodeling where the probability distribution of simulation state is represented as a function of time. The DBN metamodels reveal the time evolution of simulation and enable alternative what-if analyses unlike previous metamodels that imitate the simulation model as an input-output mapping. In earlier studies, the analysis of DBNs is restricted to discrete time instants selected beforehand in the construction phase of the metamodel. This paper introduces an extension to the framework of DBN metamodeling that employs multivariate interpolation and allows the analysis in continuous time. In practice, an approximation for the probability distribution of the simulation state is calculated by interpolating between conditional probabilities given by the DBN. The utilization of multivariate interpolation in the context of DBN metamodeling is illustrated by examples dealing with Poisson arrival processes and air combat simulation.