Uploaded May 2024 | Updated September 2026, 2 hours ago
Microsoft Future Leaders in Robotics and AI Seminar Series: Enhancing Transparency in Human-Robot Teams: A Machine Teaching Approach to Communicate Robot Decision-Making to Diverse Human Teammates
Online Seminar
Suresh Kumaar Jayaraman
Postdoctoral Fellow
Carnegie Mellon University
For transparent and effective collaboration between an agent and a human group, the challenge arises in teaching a diverse group of individuals about the agent’s decision-making process in a potentially time-sensitive and resource-limited environment. To address this challenge, we employ machine teaching through demonstrations for teaching groups of human learners modeling them as inverse reinforcement learners and using counterfactual reasoning to generate personalized informative demonstrations. Differing individual knowledge introduces challenges in personalization which we address by using aggregated team knowledge representations and developing models of team beliefs using particle filters. We present several group teaching strategies based on individual and aggregated team knowledge and conducted a simulation study comparing these different group teaching strategies with a baseline method of teaching each group member individually. We ran this study for various groups with varying combinations of learning abilities of its members. We discuss the results on multiple metrics such as learning resource utilized, learning rate, and final knowledge gained and discuss the implications of this work for human-agent teaming.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
Microsoft Future Leaders in Robotics and AI Seminar Series: Enhancing Transparency in Human-Robot Teams: A Machine Teaching Approach to Communicate Robot Decision-Making to Diverse Human Teammates
Online Seminar
Suresh Kumaar Jayaraman
Postdoctoral Fellow
Carnegie Mellon University
For transparent and effective collaboration between an agent and a human group, the challenge arises in teaching a diverse group of individuals about the agent’s decision-making process in a potentially time-sensitive and resource-limited environment. To address this challenge, we employ machine teaching through demonstrations for teaching groups of human learners modeling them as inverse reinforcement learners and using counterfactual reasoning to generate personalized informative demonstrations. Differing individual knowledge introduces challenges in personalization which we address by using aggregated team knowledge representations and developing models of team beliefs using particle filters. We present several group teaching strategies based on individual and aggregated team knowledge and conducted a simulation study comparing these different group teaching strategies with a baseline method of teaching each group member individually. We ran this study for various groups with varying combinations of learning abilities of its members. We discuss the results on multiple metrics such as learning resource utilized, learning rate, and final knowledge gained and discuss the implications of this work for human-agent teaming.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
