Explainable Polices for Emergent Raid Tactics (ExPERT)

Toyon Research Corporation proposes to continue its research to develop explainable artificial intelligence (AI) agents for salvo management against an exo-atmospheric raid using multi-agent Deep Reinforcement Learning. In Phase I, Toyon and the University of Colorado Boulder developed models and explainability tools using Advanced Framework for Simulation, Integration, and Modeling (AFSIM) to model simple intercept scenarios. The Phase I development included both policy visualization and distillation algorithms. In Phase II we will extend this work significantly by training Graph Convolutional Neural Networks (GNNs) in a large-scale league that includes both trained and heuristic agents. This effort includes improving the realism of the AFSIM-based training environment, and performing red side raid management agent training. The most performant agents will then be analyzed and distilled into decision trees to generate new tactics, techniques and procedures (TTPs) for US Missile Defense System (MDS) operator training. Approved for Public Release | 24-MDA-11673 (8 Jan 24)