Reproduce “Explainable Reinforcement Learning: A Survey” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Erika Puiutta, Eric MSP Veith
This survey provides a dedicated overview of Explainable Reinforcement Learning (XRL) methods, categorizing existing approaches based on time and scope of explanation. It reveals a prevalence of post-hoc methods and a neglect of human-centric design in current XRL research, while also proposing a uniform definition for "interpretability."
Key insights
- The majority of current XRL methods are post-hoc, indicating a tendency to simplify complex models for explanation rather than designing inherently interpretable ones.
- XRL research largely overlooks human-centric design principles, such as adapting explanations to user context or incorporating insights from social sciences on effective explanation forms.
- A proposed definition for "interpretability" emphasizes the need for explanations to be not only generated but also presented in a manner understandable by human non-experts, enabling behavior prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.