Model abstraction for discrete event systems using neural networks and sensitivity information

Simulation is one of the most powerful tools for modeling and evaluating the performance of complex systems, however, it is computationally slow. One approach to overcome this limitation is to develop a "metamodel". In other words, generate a "surrogate" model of the original system that accurately captures the relationships between input and output, yet is computationally more efficient than simulation. Neural networks (NN) are known to be good function approximators and thus make good metamodel candidates. During training, a NN is presented with several input/output pairs, and is expected to learn the functional relationship between inputs and outputs of the simulation model. So, a trained net can predict the output for inputs other than the ones presented during training. This ability of NNs to generalize depends on the number of training pairs used. In general, a large number of such pairs is required and, since they are obtained through simulation, the metamodel development is slow. In DES simulation it is often possible to use perturbation analysis to also obtain sensitivity information with respect to various input parameters. We investigate the use of sensitivity information to reduce the simulation effort required for training a NN metamodel.