Georgia Institute of Technology
Data-driven personalization techniques to account for heterogeneity in human-machine interaction
Abstract
dc:description.abstractAs robots and AI systems become more prevalent in every-day life, humans and machines will have to work closely together. Robotic devices will be used to support human health, service robots will operate alongside humans in homes, and autonomous vehicles will have to safely drive end-users to their destination. Yet, humans exhibit a high degree of heterogeneity which poses a challenge for robotic systems that are tasked with learning from and supporting humans. For example, in a medical setting, individual patients are likely to have different needs and varying biology that must be accounted for. Autonomous Vehicles (AVs) will have to learn about the differing preferences of end-users and adapt accordingly. Because of this human heterogeneity, one-size-fits-all algorithms will not suffice in many human-machine interaction scenarios. Instead, to effectively support humans, machines must be capable of recognizing individual desires, abilities, and characteristics and adapt to account for differences across individuals. This thesis focuses on the development of personalized algorithms that enable machines to better support and work with humans. Specifically, I aim to develop and research novel techniques for safely and efficiently supporting heterogeneous humans across various robotic domains. In this work, I develop data driven, personalized frameworks in healthcare, learning from demonstration, and autonomous driving domains to account for heterogeneity amongst end-users. In this thesis, I first investigate the question of how robots can best learn from human demonstrators who are suboptimal and heterogeneous in their suboptimality. I present my work on Mutual Information Driven Meta-Learning from Demonstration (MIND MELD), a framework enabling personalized robotic learning from heterogeneous human demonstrators via a learned, personalized embedding. I then extend this approach with Reciprocal MIND MELD and introduce a framework to provide personalized feedback to suboptimal human demonstrators to improve upon their ability to provide high quality demonstrations. Humans are not only heterogeneous in terms of their abilities when teaching machines; they also tend to differ in their preferences for various machine behaviors. To account for differing preferences amongst end-users, I draw upon our Reciprocal MIND MELD work and introduce Manipulating Autonomous Vehicle Embedding Region for Individuals’ Comfort (MAVERIC), an approach for personalizing driving styles of AVs to fit the preferences of end-users via personalized embeddings. In my final work, I consider personalization in safety critical domains such as healthcare. I introduce Safe Meta-Active Learning (Safe MetAL), an approach for determining the optimal, personalized parameter settings for a Deep Brain Stimulation (DBS) patient.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Mechanical Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Schrum, Mariah L.
- Advisor dc:contributor.advisor
-
- Gombolay, Matthew
- Committee members dc:contributor.committeemember
-
- Chernova, Sonia
- Feigh, Karen
- Howard, Ayanna
- Smart, Bill
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75143