Versatile and Interpretable System Explanations for AI in Multi-Domain Operations (VISION)

As the Army pivots to larger combat operations, the increased scale and complexity of multi-domain operations (MDO) has heightened the difficulty of synchronizing collection and operations across domains. Commanders conducting planning are limited by “human bandwidth,” which can degrade the quality of plans and desynchronize operations. Unfortunately, the policies learned from many current AI approaches are not yet interpretable, meaning planners have no effective way of understanding how decision recommendations are informed by the current situation, future predicted situations, or AI training data. To address this need and improve military planning in MDO using the power of AI systems, Charles River Analytics, in collaboration with Dr. Paulo Shakarian at Arizona State University, propose to design and demonstrate the feasibility of Versatile and Interpretable System Explanations for AI in Multi-Domain Operations (VISION). VISION is an XAI middle layer that provides a suite of techniques to extract comprehensible data representations of the AI system decision space. VISION will realize the full potential of AI systems to aid military planning and MDO by enabling Commanders to better understand the decision-making process behind AI recommendations and how it is predicated on future outcome predictions and past training.