Treffer: Towards explainable agent behaviour

Agents are a special kind of AI-based software in that they interact in complex environments and have increased potential for emergent behaviour, even in isolation. Explaining such behaviour is key to deploying trustworthy AI, but the increasing complexity and opaqueness of agents makes this hard. Beyond narrow-task and instant-based goals, agents may exhibit durative behaviour and be required to have planning or deliberative capabilities, or even to reason over other's behaviours. This precludes machine learning explainability -i.e. explanations over single predictions or actions-from giving complete and useful explanations. There is a need for extending explainability tools. We split the capabilities of agents into several levels, each more abstract, and produce explanations by climbing these levels: from actions, tellic (ends), deliberation, and more. The first two have been solved through frequentist models (Policy-Graphs), and the third is work in progress. We intend to extend this work by adding components for explaining epistemology, agent-agent interaction, norms and values. ; Peer Reviewed ; Postprint (published version)