Adaptive Training Scenario Generator (ATSG)

This proposal is to develop a new Adaptive Training Scenario Generator (ATSG).  The ATSG software tool will use artificial intelligence/machine learning (AI/ML) to support training in Live, Virtual, Constructive (LVC) environments by 1) rapidly generating scenarios to fit specified training objectives prior to exercises; and 2) recommend, or as an option, automatically implement adaptations to red team behavior to support achieving training objectives and in response to observed trainee performance in real time during exercises. We will focus our Phase I effort on the initial use case of fleet defense against air threats including coordinated enemy airstrikes on a Carrier Strike Group (CSG). Our technical approach includes the innovative integration of two complementary AI capabilities -- Large Language Models (such as ChatGPT) and Monte Carlo Tree Search (MCTS) -- into a powerful system for generating scenarios and adapting red team behavior during training.  In summary, ATSG uses LLMs to exploit domain knowledge to produce potential scenario modifications and AI-based MCTS planning algorithms to deeply evaluate those modifications. The result is a system that automatically and rapidly produces numerous varied scenario adaptations in real time during mission execution that are tailored to achieve effective training missions.