Synthesizing Machine-Learning Datasets from Parameterizable Agents Using Constrained Combinatorial Search

The tedious, often hand-modeled, activity of designing and implementing simulation scenarios can benefit from modern-day data-driven methods, i.e., machine-learning (ML). We envision a toolchain that exploits information obtained during live operations, such as the observed maneuvers, techniques, and procedures of all interacting players in live operational settings, that serves as input into an ML-based scenario authoring process. We present a mechanism, called the Parameter Diversifier (PD), that takes a base scenario structure and synthesizes the comprehensive datasets needed for the supervised machine-learning of a scenario authoring model. The design of the PD explores and exploits low-level agent state search space as it relates to it high-level implications at the scenario level. This work demonstrates an explicit sampling of the scenario parameter search space to build an implicit model for use in simulation scenario generation.