Explainability and Terrain Reasoning for Autonomy (EXTRA)

Artificial Intelligence (AI) based systems are showing great promise for application in the DoD. In particular, Reinforcement Learning based autonomous agents are appealing for their ability to make high-speed decisions in complex environments. However, their decision-making process is inherently black-box, and their true intent is opaque. EXTRA is a framework for providing: explainablity; terrain reasoning; and intent inference to sophisticated RL agents. This framework benefits provides commanders with rich, accurate, and actionable explanations on how their AIs, and adversaries are formulating and executing strategies. The explanations are unpacked according to features about the environmental features are critical in decision making, what strategy the blue (friendly) AI is perceiving in the red (adversary) AI, and future valuable states deemed valuable by the AI. The explanations are delivered via an intuitive Explanation User Interface for rapid and straightforward comprehension.