Title: Temporality in Sequential Explainable AI
Amal Alabdulkarim
Ph.D. Student in Computer Science
School of Interactive Computing
Georgia Institute of Technology
Date: April 30, 2025
Time: 10 AM ET
Location: Coda C1115 Druid Hills, or join online on Microsoft Teams (link)
teams.microsoft.com
Committee
Mark Riedl (Advisor) - School of Interactive Computing, Georgia Institute of Technology
Sonia Chernova - School of Interactive Computing, Georgia Institute of Technology
Matthew Gombolay - School of Interactive Computing, Georgia Institute of Technology
Kartik Goyal - School of Interactive Computing, Georgia Institute of Technology
Sarath Sreedharan - Department of Computer Science, Colorado State University
Lara Martin - Department of Computer Science and Electrical Engineering, University of Maryland, Baltimore County
Abstract
Explainable Artificial Intelligence (XAI) is a growing area of research that focuses on making AI systems easier for people to understand and trust. Explanations are critical in sensitive areas that require human-AI teaming, where people must understand and trust the AI’s decisions. However, making AI models explainable becomes much harder in sequential decision-making, like reinforcement learning agents. These agents make decisions based on a series of past and future actions, so an actionable explanation has to look at the bigger picture, not just what happened right before a choice was made.
In this thesis, I present algorithms and representations for reasoning about how future expectations and past events play into an agent's decision-making to generate actionable explanations. I develop an interactive pipeline that personalizes system explanations and meets variable user explainability goals, leveraging the affordances of retrospectives. I present my work across three domains: single-agent reinforcement learning, multi-agent systems, and narrative understanding.
]]>