Mining and Validating Belief-Based Agent Explanations.

Agent explanation generation is the task of justifying the decisions of an agent after observing its behaviour. Much of the previous explanation generation approaches can theoretically do so, but assuming the availability of explanation generation modules, reliable observations, and deterministic execution of plans. However, in real-life settings, explanation generation modules are not readily available, unreliable observations are frequently encountered, and plans are non-deterministic. We seek in this work to address these challenges. This work presents a data-driven approach to mining and validating explanations (and specifically belief-based explanations) of agent actions. Our approach leverages the historical data associated with agent system execution, which describes action execution events and external events (represented as beliefs). We present an empirical evaluation, which suggests that our approach to mining and validating belief-based explanations can be practical.