Exploring the Fidelity of Synthetic Data for Decision Support Systems in Military Applications

Military decision support tools rely on models and simulations as no historical examples exist for the battle to come. Techniques such as reinforcement learning are used for generating synthetic data by the simulator. This raises a fundamental question of how the quality of synthetic data influences the quality of the decision support as a whole. This work seeks to explore the relation between the realism of the simulator and the predictive quality of a decision support system. The concept of fidelity is introduced as a means to assess the quality of synthetic data generated by these applications. To address this research question, this study employs the referent-abstract model concept. In essence, the fidelity of an abstract model is evaluated by comparing it to a referent model, which may also contain realworld data. We specifically investigate the manoeuvre fidelity of soldiers that are tasked to clear a village from enemy forces. The findings illustrate a direct correlation between the quality of generated synthetic data and the predictive performance of the decision support application. Additional results highlight how various parameters affect the motion fidelity as well as the performance of the decision support application.
^1
1
This paper was originally presented at the NATO Science and Technology Organization Symposium (ICMCIS) organized by the Information Systems Technology (IST) Panel, IST-205-RSY - the ICMCIS, held in Koblenz, Germany, 23–24 April 2024.