Assisting human perception and control using theory of mind
Date:
October 22, 2020
Author:
Hrvoje Stojic
Abstract
In this talk, we will explore the idea that irrational human behavior can be modeled as rational with respect to incorrect internal beliefs about how the world works. We will discuss two recent papers that leverage this assumption to learn a user’s internal model from suboptimal demonstrations, then assist the user by making the real world feel more like their internal model. First, we will discuss how to assist a user with perception when their internal observation model is wrong: modify their observations to induce accurate state estimates when processed by their internal observation model. Second, we will discuss how to assist a user with control when their internal dynamics model is wrong: modify their actions to induce state transitions predicted by their internal dynamics model. These approaches to modeling and assisting human decision-making can be applied to problems in human-robot interaction, user interface design, accessibility, and education.
Notes
Assisted Perception: Optimizing Observations to Communicate State https://arxiv.org/abs/2008.02840
Where Do You Think You're Going?: Inferring Beliefs about Dynamics from Behavior https://arxiv.org/abs/1805.08010
Siddharth Reddy is a fourth-year computer science Ph.D. student at the Berkeley Artificial Intelligence Research Lab. He works on machine learning algorithms for augmenting human control in domains like robotics and education.
