Simulation-driven metamodeling of complex systems using neural networks

Simulation of large complex systems for the purpose of evaluating performance and exploring alternatives is a computationally slow process, currently still out of the domain of real-time applications. To overcome this limitation, one approach is to obtain a 'metamodel' of the system, i.e., a 'surrogate' model which is computationally much faster than the simulator and yet is just as accurate. We describe the use of Neural Networks (NN) as metamodeling devices which may be trained to mimic the input-output behavior of a simulation model. In the case of discrete event system (DES) models, the process of collecting the simulation data needed to obtain a metamodel can also be significantly enhanced through concurrent estimation techniques which enable the extraction of information from a single simulation that would otherwise require multiple repeated simulations. We will present applications of two benchmark problems in the C3I domain: A tactical electronic ground-based radar sites; and an aircraft refueling and maintenance system as a component of a typical Air Tasking Order. A comparative analysis with alternative metamodeling approaches indicates that a NN captures significant nonlinearities in the behavior of complex systems that may otherwise not be accurately modeled.