An Artificial Neural Network Based Metamodel for Analysing a Stochastic Combat Simulation

In this paper, we discuss the use of multi-layer perceptron (MLP) neural networks to predict the outputs from a stochastic combat simulation model. The use of a faster, auxiliary model to approximate the input-output relationships of a complex simulation model is known in the simulation community as metamodelling. Since many simulation models contain a large number of input parameters, however it is necessary to reduce the set considered and determine the most important ones to include in a metamodel given a particular modelling context. This paper employs a two-stage experimental design — first, Morris’ randomised one-at-a-time (OAT) design is used as a factor screening method, and then a Latin Hypercube design is employed to select the set of input configurations which defines the required runs of the simulation model. The set of input configurations together with the associated simulation outputs provide the training data for the development of the MLP networks. The approach is illustrated with reference to a stochastic combat simulation called SIMBAT. The paper then investigates a number of aspects relating to the development of neural network-based metamodels of stochastic combat simulations. It shows that using the outputs from each replication of a stochastic simulation is generally better than only using the mean output.