Generative Training Scenario Development Toolset (GenScen)

Stottler Henke proposes to develop a Generative Training Scenario Development toolset (GenScen) applying generative AI methods to automate the creation of high-fidelity air defense scenarios with adaptive red threat behaviors learned from past pilot-in-the-loop data, while also supporting communications analysis during exercises for performance assessment purposes. With the inclusion of rapid generative AI technology into the Next Generation Threat System (NGTS), its scope and utility could expand well beyond its current state. These capabilities will not only increase instructor efficiency with learning objective (LO) based informed joint and allied force virtual asset generation, but also increase the scope and scale of NGTS mission sets with AI-generated threat behavior and decision-making. Scenario generation methods apply techniques from Unsupervised Environment Design (UED) simultaneously with techniques from human-AI interaction studies (specifically, the Boltzmann Policy Distribution). GenScen communications analysis methods integrate advanced natural language processing (NLP) and machine learning to assess Naval aviation communications, utilizing cross encoders, Large Language Models, and semantic embeddings. Prototype development in Phase I is facilitated by our team’s combination of subject matter expertise and experience implementing a wide range of AI applications, natural language analytics, and training debrief tools.