Toward Explainable Reinforcement Learning and a Custom RL Benchmark for Strategic Decision Making
Open Access
Author:
Panda, Sourav
Graduate Program:
Informatics
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
March 19, 2025
Committee Members:
Jonathan Dodge, Thesis Advisor/Co-Advisor Aron Laszka, Committee Member Carleen Maitland, Program Head/Chair C Lee Giles, Committee Member Abhinav Verma, Committee Member
Keywords:
Reinforcement Learning Explainable AI Explainable RL Sequential Domain Decision Making
Abstract:
Reinforcement Learning (RL) has achieved remarkable success in complex domains such as real-time strategy (RTS) games. However, the opaque decision-making of RL agents poses challenges for human-AI collaboration and strategic planning. This thesis presents a framework for Explainable Reinforcement Learning (XRL), integrating explainability techniques with RL-based decision-making in StarCraft II, a representative RTS environment. The work introduces MIXTAPE (Middleware for Interactive XAI with Tree-based AI Performance Evaluation), a system designed to enhance transparency in RL-driven strategy games. Additionally, we developed a custom RL benchmark to bridge the gap between simplistic mini-games and full-game complexity, enabling progressive learning and evaluation of AI decision-making strategies. By incorporating reward decomposition, visualization modules, and an initial user study design, this research paves the way toward making RL agents more interpretable, ultimately aiming to foster better human-AI collaboration in complex, real-world settings.
