Explainable Reinforcement Learning for Broad-XAI: A Conceptual Framework and Survey

Reinforcement Learning systems currently operate as "black boxes" — the paper identifies that RL algorithms, despite their success in solving complex decision-making problems, obscure how they actually make choices through "complex arrays of values and functions," making it difficult for users to understand why an AI agent took a particular action.

A new framework (CXF) aims to extract and explain what AI agents are actually "thinking" — the Causal XRL Framework proposes methods to reveal an agent's perception of its environment, its motivations, goals, and decision-making values, moving beyond explaining single decisions to providing integrated explanations of overall behavior.

Broad-XAI represents a shift toward holistic explanations rather than isolated ones — instead of explaining individual AI decisions in isolation, this approach seeks to integrate explanations from multiple algorithms into coherent narratives about an agent's behavior that match what different users actually need to understand.

The framework is designed to be adaptable and communicative — CXF incorporates external knowledge systems (ontologies) and communication tools, enabling AI agents to answer questions and adjust to their environments, suggesting a path toward AI systems that can actively explain themselves to users.