Influence-Guided Agent Adaptation for Dynamic Human–AI Collaboration
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
To facilitate effective human–agent collaboration, agents must not only understand the intent of their human partners and adapt to it reactively, but also account for how their own behavior influences their partners. Rather than treating this influence as a passive consequence of interaction, we propose actively predicting and optimizing it as part of the agent's decision-making process. In doing so, the agent explicitly models how its actions are likely to shape the subsequent behavior of its human partner. To this end, we introduce i-STAR, a novel framework that not only optimizes an agent's own actions but also reasons over and jointly optimizes the anticipated response of the human partner to those actions. Our framework employs a multi-layer architecture comprising three core components: (1) a team response model, which predicts how the human partner will respond to candidate agent actions; (2) a joint utility model, which estimates the overall utility of the agent's selected action together with the anticipated partner response; and (3) an action policy whose advantage function is shaped by the maximum-utility joint action. We find that, using this framework, our agent outperforms current state-of-the-art agents across all layouts of the challenging Overcooked benchmark. Furthermore, in a human-subject study, i-STAR is perceived as significantly more fluent than the strongest agent–agent baseline.