Скачать или смотреть A Survey of Explainable Reinforcement Learning : Targets, Methods and Needs

The paper, "A Survey of Explainable Reinforcement Learning: Targets, Methods and Needs," focuses on *Explainable Reinforcement Learning (XRL)**. This field is dedicated to understanding the actions of **Reinforcement Learning (RL) agents**. These agents learn to make sequential actions to maximize cumulative reward within an environment, often through **deep neural networks* that make their internal mechanisms opaque. The authors propose a *new taxonomy* for XRL methods, which is structured around two key questions: "What" the method aims to explain (e.g., the agent's overall **policy**, a specific **sequence of interactions**, or a particular **action**) and "How" the explanation is provided. This extensive survey reviews over 250 papers. It also discusses related domains that warrant further research, such as **Relational RL (RRL)**, and highlights several **needs for the XRL field**, including the importance of comparing methods, establishing unified metrics, conducting more user studies, and developing dedicated user interfaces.

https://arxiv.org/pdf/2507.12599