A methodology for the development of reverse simulation metamodels using neural networks

System design/redesign is a complex process in which models are used to make decisions on changes in existing or proposed systems. The goal of the design process is to design a system that meets or exceeds certain performance measures without violating any constraints. The objective of this research is to develop a methodology using simulation models and neural networks to build a reverse-simulation metamodel that can be used as a decision support tool when designing a new system or redesigning an existing one. This decision tool will suggest the system's design parameters when the required performance measures are specified.; The choice of the experimental design (D-optimal or orthogonal arrays) used to generate the training data and the type of learning algorithm used to train the neural network (Back-propagation or Genetic Algorithms) were investigated using 36 different simulation models with different degrees of complexity. Also the effects of model complexity on the performance of the neural network metamodel were studied. It was found that neural network metamodels trained using an orthogonal training set performed better than those trained using a D-optimal training set. It was also found that standard neural networks (Back-propagation) outperformed evolutionary neural networks in developing reverse simulation metamodels. It was also found that the saturation of the simulation model, the number of machines, and the number of processing levels had minimal impact on the performance of the back-propagation neural network, but the variability within the simulation model significantly affects the prediction capability of the neural network metamodel.