Integrating neural networks with special purpose simulation

Traditional methods of dealing with variability in simulation input data are mainly stochastic. This is most often the best method to use if the factors affecting the variation or the nature of the relationships between the factors and the outputs cannot be easily identified. Artificial neural networks have the ability to learn complex relationships between inputs and outputs. Their use can greatly enhance simulation models and allow for more accurate representations of real life scenarios. The paper proposes a generic approach for integrating external processes such as neural networks with simulation models. The object oriented method is used to `expose´ the properties of the simulation models to external processes, and allow for users to define relationships at run time. This approach was tested by integrating a neural network model for predicting the productivity of an excavator with an earth moving simulation process. This proved to be of extreme benefit because the defined neural network parameters depend on certain factors which varied during the simulation