Explainable Reinforcement Learning with Counterfactuals and Causal Reasoning

Abstract

Explainable Reinforcement Learning (XRL) is an emerging field of research that attempts to provide human-understandable explanations for the decisions made by RL agents. Such agents often demonstrate complex behaviours, learned through many interactions with the environment, that their human counterparts may have not considered before. Being able to explain such complex behaviour in real-world crucial before they can responsibly be deployed on real-world problems. Reliable explanations can develop trust in the previously black-box policies of RL agents and enable effective interrogation of their behaviour in case of potential failure. By addressing the �why� and �why not�, �what if�, and �how� questions under human guidance, in this basic research project, we hope to provide better explanations from XRL and bring new opportunities for the use of models in various applications such as Unmanned Aerial Vehicle (UAV) route planning for C2 mission and intelligent driving.

Document Details

Document Type
DoD Grant Award
Publication Date
Feb 16, 2024
Source ID
FA23862314066

Entities

People

  • Flora Salim

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force
  • University of New South Wales

Tags

Fields of Study

  • Computer science

Readers

  • Agent-Based Social Robotics and Mobile-Assisted Learning in Virtual Environments.
  • Theoretical Analysis.

Technology Areas

  • AI & ML
  • AI & ML - Autonomous Systems
  • Autonomy
  • Autonomy - Human-Robot Interaction