Finding Human-Aligned Abstractions Efficiently for Explaining Plan Quality Differences

Abstract

Explaining differences in the quality of plans is an effective method for helping users understand the decisions made by planners. When a user suggests an alternative plan that is lower in quality, these explanations can clarify why the original plan was preferable, or reveal hidden preferences and trade-offs in the model. Recent advances in plan explanation have used abstraction to generate explanations for differences in plan quality. Details of planning problems are abstracted until the quality of plans they produce are equal in cost. These abstractions are then used as causal factors in explanations. We build on this work in two ways. Firstly, we introduce a number of pruning techniques based on properties of abstractions, extracted from the human-subject study, to improve the efficiency of finding human-aligned abstractions. Secondly, we formalise and implement a number of heuristics, derived from a human-subject study capturing how people explain differences in plan quality.