Are We On The Same Page? Hierarchical Explanation Generation for\n Planning Tasks in Human-Robot Teaming using Reinforcement Learning

Abstract

Providing explanations is considered an imperative ability for an AI agent in\na human-robot teaming framework. The right explanation provides the rationale\nbehind an AI agent's decision-making. However, to maintain the human teammate's\ncognitive demand to comprehend the provided explanations, prior works have\nfocused on providing explanations in a specific order or intertwining the\nexplanation generation with plan execution. Moreover, these approaches do not\nconsider the degree of details required to share throughout the provided\nexplanations. In this work, we argue that the agent-generated explanations,\nespecially the complex ones, should be abstracted to be aligned with the level\nof details the human teammate desires to maintain the recipient's cognitive\nload. Therefore, learning a hierarchical explanations model is a challenging\ntask. Moreover, the agent needs to follow a consistent high-level policy to\ntransfer the learned teammate preferences to a new scenario while lower-level\ndetailed plans are different. Our evaluation confirmed the process of\nunderstanding an explanation, especially a complex and detailed explanation, is\nhierarchical. The human preference that reflected this aspect corresponded\nexactly to creating and employing abstraction for knowledge assimilation hidden\ndeeper in our cognitive process. We showed that hierarchical explanations\nachieved better task performance and behavior interpretability while reduced\ncognitive load. These results shed light on designing explainable agents\nutilizing reinforcement learning and planning across various domains.\n