[Paper Review] Progressive Explanation Generation for Human-robot Teaming

Yu Zhang, Mehrdad Zakershahrak|arXiv (Cornell University)|Feb 2, 2019

AI-based Problem Solving and Planning23 references4 citations

TL;DR

This paper introduces Progressive Explanation Generation (PEG), a framework that minimizes cognitive effort for human teammates by sequencing explanation steps to improve comprehension. It formulates explanation as a sequential decision problem using model-distance heuristics, showing PEG explanations are nearly as concise as concise explanations but significantly reduce cumulative cognitive effort, especially under model mismatches between robot and human.

ABSTRACT

Generating explanation to explain its behavior is an essential capability for a robotic teammate. Explanations help human partners better understand the situation and maintain trust of their teammates. Prior work on robot generating explanations focuses on providing the reasoning behind its decision making. These approaches, however, fail to heed the cognitive requirement of understanding an explanation. In other words, while they provide the right explanations from the explainer's perspective, the explainee part of the equation is ignored. In this work, we address an important aspect along this direction that contributes to a better understanding of a given explanation, which we refer to as the progressiveness of explanations. A progressive explanation improves understanding by limiting the cognitive effort required at each step of making the explanation. As a result, such explanations are expected to be smoother and hence easier to understand. A general formulation of progressive explanation is presented. Algorithms are provided based on several alternative quantifications of cognitive effort as an explanation is being made, which are evaluated in a standard planning competition domain.

Motivation & Objective

• To address the gap in existing explanation generation methods that focus on the robot’s perspective while ignoring the human’s cognitive limitations.
• To improve human understanding in human-robot teaming by minimizing the cognitive effort required to interpret explanations.
• To model explanation as a sequential process where the order of information delivery affects comprehension efficiency.
• To develop a general formulation of progressive explanation generation that accounts for differences in cognitive models between robot and human partners.
• To evaluate the effectiveness of PEG in reducing cumulative cognitive effort compared to concise explanations in a standard planning domain.

Proposed method

• Formulates progressive explanation generation as a sequential decision-making problem where cognitive effort is minimized over time.
• Defines cognitive effort at each step using a model-plan distance metric between the robot’s model ($M^R$) and the human’s model ($M^H$).
• Employs A* search with heuristics based on feature mismatches (preconditions, add/delete effects) between $M^R$ and $M^H$, with a penalty for each feature change.
• Uses a missing probability for model features to simulate uncertainty in human understanding, with values ranging from 0.06 to 0.14 in evaluations.
• Compares PEG with concise explanations by measuring explanation size, computation time, and cumulative cognitive effort ($\sum \rho_i$).
• Evaluates the approach in the rover domain using Fast Downward as the underlying planner, with results based on two problem instances.

Experimental results

Research questions

• RQ1 How can explanation generation be improved by considering the cognitive effort required by the human explainee rather than just the robot’s perspective?
• RQ2 To what extent does sequencing explanation steps progressively reduce cumulative cognitive effort compared to non-sequential or concise explanations?
• RQ3 How does increasing model mismatch (i.e., missing features in the human’s model) affect the performance and computation time of progressive explanation generation?
• RQ4 Can progressive explanations achieve a favorable trade-off between conciseness and cognitive efficiency in human-robot teaming?
• RQ5 What is the impact of different distance metrics on the quality and efficiency of progressive explanation generation?

Key findings

• PEG explanations are nearly as concise as concise explanations, with only a small increase in size, while significantly reducing the cumulative cognitive effort ($\sum \rho_i$).
• In the first rover problem, PEG explanations achieved a lower $\sum \rho_i$ value than concise explanations, indicating reduced cognitive load, despite longer computation time.
• As the number of missing features in the human model increased, the computation time for PEG rose sharply, with a 1342.1-second search time for 11 missing features at a 0.11 missing probability.
• When no features were missing (0.14 probability), the PEG search time dropped to just 0.7 seconds, confirming that computation cost correlates with model mismatch.
• The size of the explanation had little impact on computation time except when the size was zero, where no search was needed.
• The results demonstrate that PEG effectively balances conciseness and cognitive efficiency, making it a viable approach for real-time human-robot teaming applications.