Explainable AI for Mission Planning and Execution with Interpretable Courses of Action
Abstract
The overall goal of this project was to support Navy planning and execution by developing methods for improved explainability, interpretation, and performance in artificial intelligence (AI) systems. This project was organized into three primary research thrusts. Thrust one developed methods to identify coherent COAs from AI systems, including a heuristic temporal logic search algorithm and a neurosymbolic approach for learning general temporal logic formulae from agent trajectories. Thrust two developed methods to discover biases from decision sequencies and incorporate biases into trained decision policies, including a framework for an RL agent to adapt to biased teammates, methods for analyzing the impact of agent biases on team performance, and a survey of inverse reinforcement learning (RL) methods for discovering agent biases. Thrust three developed methods to improve the performance of explainable RL agents, including a method for training hierarchical RL agents that can handle adversaries, an RL method for training large teams of coordinated agents in adversarial environments, a framework for minimizing communications in multi-agent RL, and a method for deploying spiking neural network based RL policies on neuromorpohic hardware for low SWaP applications. Our results were disseminated to communities of interest through AI/robotics journals, conferences, workshops, and invited seminars.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 07, 2024
- Accession Number
- AD1227698
Entities
People
- Girish Chowdhary
- Heather Filippini
- Huy T. Tran
- R.S. Sreenivas
Organizations
- University of Illinois Urbana–Champaign