Explaining Agent Intentions

Understanding why an agent acts in a certain way based purely on its observed behaviour is not an easy hill to climb, especially when explanations aspire to be reliable and interpretable. Despite the growing body of research on Explainable Agency (XAg), existing approaches to agent explainability often remain disconnected from causal and teleological theories of explanation or rely on the strong assumption that agents pursue a single goal or a set of independent goals. My thesis aims to develop methods to explain the reasoning behind when, how, and why agents pursue and prioritise multiple intentions, and how these intentions relate to one another. We first discuss what we mean by explainability, and we introduce a first step to extract the agent’s priorities from observed behaviour. We then outline future directions for using this information to support policy improvement and to infer inter-intention relationships, and open a discussion on the boundaries of agent explainability.

Authors

• Sara Montese Barcelona Supercomputing Center / Universitat Politecnica de Catalunya

Keywords

• Explainable AI
• Agent Explainability
• Intentions
• Causality