Contrastive Explanations for Explainable Artificial Intelligence Planning

Explainable Planning is an important area of research which focuses on providing justifications and explanations for planner decisions. Much effort has been made in this field to provide explanations for users. However, there has not been as much effort in determining what users want explained. We observe that the need for plan explanation is driven by the fact that a human and a planning agent may have different models of the planning problem and different computational abilities. We hypothesise that users prompt explanations through the use of local contrastive questions rather than global "how" or "why" questions. To test this hypothesis we conducted an empirical study to find what questions users have when faced with a plan. We present our analysis of the results of the study, categorising our findings to create a taxonomy of questions.
We find that the majority of user questions are contrastive questions of the form "Why A rather than B?". These kinds of questions can be answered with a contrastive explanation that compares properties of the original plan containing A against the contrastive plan containing B. We introduce the plan negotiation problem where there is some disagreement between the proposed plan and the plan expected by the user. We propose a solution to this problem through an iterative process of question asking and answering. We propose a domain-independent approach of compiling these questions into constraints. These constraints are added to the planning model, so that a solution to the new model represents the contrastive plan which we use to produce an explanation. We introduce a formal description of the compilation from user question to constraints in a temporal and numeric PDDL2.1 planning setting. We created a framework implementing the above.
We then evaluated the framework through a user study. The results of the user study show that the explanations we provide are majoritively satisfactory to the user.