Explainable Reinforcement Learning Survey

Léo Saulières’ latest Journal of Artificial Intelligence Research paper, released on October 22, 2026, offers an extensive review of eXplainable Reinforcement Learning (XRL) methods ¹. The study focuses specifically on explaining the actions of agents trained via reinforcement learning, distinguishing it from broader explainable AI research.

The author introduces a clear two‑axis taxonomy based on the questions ‘What’ and ‘How’. The first axis identifies what is being explained—such as policy decisions, reward structures, or state transitions—while the second axis describes how explanations are provided. This framework organizes over 250 papers, giving researchers a systematic way to classify existing techniques and identify gaps in target–method coverage.

Regulatory context frames part of the motivation for this work. EU legislation effective January 1st, 2027 will mandate explainability for high‑risk applications such as credit scoring and autonomous vehicles ¹. Saulières notes that while the law sets broad requirements, concrete guidance remains sparse, underscoring the need for rigorous, domain‑specific explanation methods.

The paper also references earlier classification attempts, notably Guidotti et al.’s 2019 effort focused primarily on image‑classification models. While those works laid groundwork for taxonomy design, they did not address reinforcement learning’s unique sequential decision structure; Saulières’ contribution fills that gap.

Despite covering over 250 studies, the author acknowledges that the field still lacks comprehensive coverage of emerging counterfactual explanations—a technique gaining traction within academia for its ability to illustrate “what if” scenarios in policy execution. Future surveys may therefore need to incorporate such approaches to fully capture XRL’s evolving landscape.

In summary, Saulières provides a structured lens through which researchers can navigate the growing body of XRL literature, align their work with regulatory demands, and spot opportunities where novel explanation mechanisms could make real‑world AI systems more transparent and trustworthy. The survey is an essential reference point for anyone looking to understand or contribute to explainable reinforcement learning today.