Demonstrating the Impact of Modelling Decisions

Alan Lindsay, Ron Petrick

Keywords: Problem Modelling, Explainable Planning, Plan Visualisation, Explanations for Failure, Explanations for Suboptimality

TL;DR: Visualisations to justify to a user the modelling steps that have led to their plan being suboptimal or inconsistent in the agent's model

Abstract: The modelling problem involves making compromises between a variety of competing factors, including planning efficiency, plan language usefulness, and real-world optimisation goals. Optimal plans in the resulting model may appear suboptimal when executed in the world. As a consequence, a human observer might find it difficult to comprehend the apparent inefficient behaviours of the agent, which might impact on the human’s trust of the agent. In this work we consider modelling decisions, such as abstractions, and their impact on the resulting plans. Our aim is to build a general approach that can assist a user to better understand both the implications of a modelling step and provide justification to support the modelling step. As a start, we have extended an off-the-shelf plan visualisation tool to provide plan failure visualisations, to demonstrate the impact of these modelling decisions to the user.