‘Structural variance’, randomisation strategies and the design of experiments using combat models

This paper is concerned with a feature of many deterministic combat models, commonly referred to as ‘structural variance’. Due mainly to the effects of differential tasking of battlefield elements, the output may appear as an irregular and non-monotonic function of any input variables. This can make it difficult to generalise from any particular set of results, complicating any decision-making process based on them. It is generally agreed that the addition of stochastic features to such models can assist in alleviating this problem. By building a dedicated experimental combat model, allied to a very extensive experimental programme, this paper gives provisional guidance on: the varieties of ‘structural variance’ and their likely resistance to remedial measures; the choice of parameters for stochastic treatment and the range over which they should be varied; the effectiveness of different randomisation strategies; the impact of different sampling distributions; and ‘variance reduction’ methods which may increase the precision and information-value of combat modelling experiments.