Efficient explaining policies in reinforcement learning and applications
Year: 2026
Type: article
Abstract: Although explainable reinforcement learning (XRL) has made significant progress, policy explanation still lacks a unified classification framework and standardized evaluation criteria. Based on a structured review of XRL works related to policy explanation, this paper identifies this gap and proposes a comprehensive taxonomy for policy explanation methods, organized along primary axes of decision-making mechanisms and explanation scope, supplemented by secondary axes and five main method families. To elucidate the inter-connections among components in the explanation process, we introduce an integrated pipeline: Decompose → Score → Select → Explain → Validate , highlighting the interaction and information flow between stages. In addition, a standardized evaluation framework with quantitative metrics is proposed to assess explanation quality. Three case studies were conducted to validate these contributions; results show that the taxonomy facilitates the positioning of existing methods, the pipeline demonstrates a transparent process for generating and validating explanations, and the evaluation framework enables objective assessment of key attributes, thereby supporting the production of transparent and systematically assessable policy explanations.
Source: Applied Soft Computing
Authors: Dac Hoang Nguyen, Nhu-Tai Do, Huy Quoc Nguyen
Institution: Saigon University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,730
OpenAlex ID: W7222808619