Explainable AI for Mission Planning and Execution with Interpretable Courses of Action
Abstract
Navy mission planning and execution requires the ability to consider a large number of potential courses of action (COAs) and ensure optimal decision-making in uncertain environments. However, current methods can only explore a limited number of COAs, do not accurately model commander biases, and struggle when facing intelligent adversaries. Recent advances in artificialintelligence (AI) have the potential to address these limitations, thus creating a paradigm shift in how the Navy operates. Yet several critical challenges limit our ability to fully realize this shift, many of which relate to a lack of explainability in deep learning-based AI systems. This proposal aims to address these challenges through three primary research thrusts. Each thrust containsseveral open research questions that need to be resolved before AI systems can support mission planning and execution in an effective and trustworthy manner.Research thrust one will develop methods to classify decision sequences from AI systems into clusters representing potential COAs using generative models. We will then leverage temporalepistemic logic to translate those clusters into coherent and interpretable COAs. Research thrust two will develop Learning from Demonstration methods to discover commander biases based onobservations of their decision sequences. We will then develop adapt-to-learn techniques to incorporate desired biases into trained AI decision policies. Research thrust three will develop a methodology to evaluate hierarchical RL architectures using an explainability taxonomy and assessments of policy adaptation, in addition to mission performance. We will then develop methodsto optimize hierarchical architectures using unsupervised learning methods and confidence-based training. The expected outcomes of this research will be algorithms and systems that improve explainability in AI systems for future use in Navy mission planning and execution. Methods developed in thrust one will allow human planners to utilize AI systems to generate coherent COAs that canbe implemented in a mission of interest. Methods developed in thrust two will enable accurate representation of friendly and enemy commanders when training AI systems for mission planning or execution. Methods developed in thrust three will provide architectures that can be used to support mission execution in contested environments, for example by providing AI decisionpolicies for autonomous systems that leverage explainability benefits of hierarchical systems. The proposed research is fundamental in nature, but will performed in a manner that emphasizes its value towards the Navy by focusing on a Navy-relevant concept of operations. We will demonstrate developed methods in theory, simulation, and hardware to ease their transition into higherTechnology Readiness Levels in future efforts.Explainable AI for Mission Planning and Execution with Interpretable Courses of Action
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 29, 2020
- Source ID
- N000142012249
Entities
People
- Huy Tran
Organizations
- Office of Naval Research
- United States Navy
- University of Illinois Urbana–Champaign