Survey Maps Explainable Reinforcement Learning

Léo Saulières has published a comprehensive survey on explainable reinforcement learning (XRL) in the latest issue of the Journal of Artificial Intelligence Research (JAIR). The work, dated 2026/10/22, systematically catalogs more than 250 papers that aim to clarify how agents trained by reinforcement learning make decisions, offering a clear taxonomy based on two axes: “What” – what is being explained – and “How” – the manner of explanation ¹.

The motivation behind XRL is growing regulatory pressure. Starting January 1st 2027, EU legislation requires high‑risk AI systems such as credit scoring or autonomous vehicles to provide explanations for their outputs ¹. Saulières notes that this regulation directly fuels interest in transparent RL agents, especially those deployed in safety‑critical domains ¹.

Saulières positions his survey within an existing lineage of explainability taxonomies. He cites earlier efforts like Guidotti et al., 2019, which focused primarily on image classification models, whereas his framework explicitly targets the sequential decision‑making nature of RL. By separating the target (“what”) from the delivery mechanism (“how”), the taxonomy helps researchers classify methods ranging from post‑hoc visualizations to counterfactual trajectory analyses ¹.

Beyond summarizing current techniques, the paper identifies several gaps. First, while many works address either policy interpretation or reward decomposition, few tackle both simultaneously. Second, most studies assume fully observable environments; real‑world deployments often involve partial observability and noisy sensors, raising questions about robustness of explanations under uncertainty ¹. The author also highlights a need for standardized benchmarks that evaluate not only predictive accuracy but also human interpretability metrics ¹.

The survey does not yet cover all available research. It acknowledges that future surveys may need to incorporate emerging approaches such as counterfactual explanations, which have gained traction recently within academia ¹. These methods generate “what‑if” scenarios by perturbing state trajectories to reveal causal relationships between actions and outcomes – a promising direction for XRL given its dynamic environment modeling ¹.

In conclusion, Saulières’ work provides an organized overview of over 250 XRL papers, introduces a practical two‑dimensional taxonomy, and outlines key needs for advancing explainable reinforcement learning in regulated contexts. Researchers can use this map to locate suitable explanation strategies for their agents and identify open problems where further theoretical and empirical investigation is warranted ¹.