Bayesian network meta-models from combat simulation for Defence decision analysis
Abida Shahzad, Steven Mascaro, Thang Cao, Kevin B. Korb · 2019
Defence Science and Technology Group (DST) is investigating the characteristics of Land Combat Vehicles (LCVs) with a view to improving purchasing decisions between different LCV options. Using a combination of closed-loop combat simulation (i.e., one run without any human interventions) and subject matter experts (SMEs), these studies seek to understand the close combat capability factors for a future LCV. These factors are being investigated for different kinds of physical environment and different levels of enemy lethality, with each LCV's lethality, survivability, signature (the ease with which the enemy identifies friendly forces) and knowledge acquisition (friendly forces identifying enemies) being measured, yielding a multi-dimensional view of the different LCVs and their performance. Our study describes a new approach, developing a Bayesian network (BN) meta-model (i.e., a model based upon the combat simulation model; see Fig 1) that combines those multiple dimensions in a single multicriterial decision model. Figure 1. A Bayesian network meta-model (from the PC algorithm). Both deterministic and stochastic methods have been developed for multicriteria decision analysis (MCDA). Limitations to prior approaches include treating utilities as independent of each other and also very limited ability to deal with uncertainty about preferences. Bayesian networks offer advantages in these regards, so here we develop an alternative approach using a BN meta-model for decision making and evaluating the operational impact of different LCV choices. BNs support both the prediction of operational results and their causal explanation. These BNs were developed both from expert elicitation and causal discovery (data mining) from combat simulation data, and enable causal visualization, sensitivity analysis and optimization. We evaluate the performance of some of these BNs against alternatives, using common accuracy-based evaluation metrics. Please note that the example and data presented in this paper are not representative of the real LCV options and testing scenario, they are used for methodology illustration only.