Dynamic Bayesian Networks as a Probabilistic Metamodel for Combat Simulation

Simulation modeling is used in many situations. Simulation meta-modeling is used to estimate a simulation model result by representing the space of simulation model responses. Metamodeling methods are particularly useful when the simulation model is not particularly suited to real-time or mean real-time use. Most metamodeling methods provide expected value responses while some situations need probabilistic responses. This research establishes the viability of Dynamic Bayesian Networks for simulation metamodeling, those situations needing probabilistic responses. A bootstrapping method is introduced to reduce simulation data requirement for a DBN, and experimental design is shown to benefit a DBN used to represent a multi-dimensional response space. An improved interpolation method is developed and shown beneficial to DBN metamodeling applications. These contributions are employed in a military modeling case study to fully demonstrate the viability of DBN metamodeling for Defense analysis application.