[Paper Review] Order Matters: Generating Progressive Explanations for Planning Tasks in Human-Robot Teaming

Mehrdad Zakershahrak, Shashank Rao Marpally+3 more|arXiv (Cornell University)|Apr 16, 2020

AI-based Problem Solving and Planning32 references4 citations

TL;DR

This paper proposes a progressive explanation generation (PEG) framework for human-robot teaming that models human cognitive preferences for information order in explanations using a goal-based Markov Decision Process (MDP) with inverse reinforcement learning (IRL). The method learns a reward function from human preference data, demonstrating that progressive explanations—sequenced to align with cognitive assimilation—reduce cognitive load and improve task performance by 10.9% in accuracy and 18.5% in perceived effort (NASA TLX).

ABSTRACT

Prior work on generating explanations in a planning and decision-making context has focused on providing the rationale behind an AI agent's decision making. While these methods provide the right explanations from the explainer's perspective, they fail to heed the cognitive requirement of understanding an explanation from the explainee's (the human's) perspective. In this work, we set out to address this issue by first considering the influence of information order in an explanation, or the progressiveness of explanations. Intuitively, progression builds later concepts on previous ones and is known to contribute to better learning. In this work, we aim to investigate similar effects during explanation generation when an explanation is broken into multiple parts that are communicated sequentially. The challenge here lies in modeling the humans' preferences for information order in receiving such explanations to assist understanding. Given this sequential process, a formulation based on goal-based MDP for generating progressive explanations is presented. The reward function of this MDP is learned via inverse reinforcement learning based on explanations that are retrieved via human subject studies. We first evaluated our approach on a scavenger-hunt domain to demonstrate its effectively in capturing the humans' preferences. Upon analyzing the results, it revealed something more fundamental: the preferences arise strongly from both domain dependent and independence features. The correlation with domain independent features pushed us to verify this result further in an escape room domain. Results confirmed our hypothesis that the process of understanding an explanation was a dynamic process. The human preference that reflected this aspect corresponded exactly to the progression for knowledge assimilation hidden deeper in our cognitive process.

Motivation & Objective

• To address the gap in explanation generation that focuses on the explainer’s perspective rather than the explainee’s cognitive load and understanding process.
• To model human preferences for the order of information in sequential explanations, recognizing that progression enhances comprehension.
• To develop a generalizable framework that learns how to sequence explanation steps to minimize cognitive effort while maximizing task performance.
• To validate that domain-independent features (e.g., action distance, Levenshtein distance) and domain-dependent features (e.g., spatial position of contingencies) jointly shape human preferences for explanation order.
• To demonstrate that progressive explanations lead to measurable improvements in both objective task accuracy and subjective cognitive load metrics.

Proposed method

• Formulates progressive explanation generation as a goal-based Markov Decision Process (MDP), where each state represents a partial explanation and actions represent adding the next explanation step.
• Employs inverse reinforcement learning (IRL) to learn a reward function from human preference data collected via crowdsourced studies on MTurk.
• Uses a combination of domain-independent features (e.g., action distance, Levenshtein distance) and domain-dependent features (e.g., x_min, y_min, x_max, y_max) to represent the cognitive cost of explanation transitions.
• Trains the reward function on 21 human-annotated explanation sequences in a scavenger-hunt domain and tests on 87 subjects in an escape-room domain.
• Applies the learned reward function to generate progressive explanations that minimize perceived cognitive effort while maximizing task performance.
• Validates the method using both subjective (NASA TLX) and objective (task accuracy) metrics across two distinct domains.

Experimental results

Research questions

• RQ1 Do humans exhibit measurable preferences for the order of information in progressive explanations, and can these preferences be learned?
• RQ2 To what extent do domain-independent features (e.g., action distance, Levenshtein distance) and domain-dependent features (e.g., spatial position of contingencies) influence human preferences for explanation sequencing?
• RQ3 Does generating progressive explanations—sequenced to reflect cognitive assimilation—lead to improved task performance and reduced cognitive load compared to non-progressive baselines?
• RQ4 Is the cognitive process of understanding explanations dynamic, and is this reflected in the feature weights learned via IRL?
• RQ5 Can the proposed framework generalize across different domains (e.g., scavenger-hunt and escape-room) while maintaining consistent improvements in explanation quality?

Key findings

• The proposed PEG method significantly outperformed random, Manhattan, and baseline methods on all NASA TLX metrics, with a statistically significant improvement in the weighted TLX score (p < 0.05).
• Task accuracy improved from 85.4% (random) to 96.3% (PEG) in the escape-room domain, representing a 10.9 percentage point increase.
• The Levenshtein distance and action distance features received the highest normalized weights (0.46 and 0.44, respectively), indicating strong influence on cognitive effort perception.
• The plan cost distance feature had a low weight (0.04), suggesting that in the escape-room domain, increasing plan cost did not correlate with increased cognitive effort.
• The PEG-generated explanation sequences showed smoother action distance curves across steps, visually confirming the progressiveness of the explanation flow.
• The results confirmed that human preferences for explanation order are shaped by both domain-independent and domain-dependent features, validating the dynamic nature of cognitive assimilation during explanation comprehension.