Considering User Preferences In Autonomous Agent Tasks

Autonomous agents working in an environment shared with humans must respond to a variety of changes. Human users have a variety of needs and preferences, and agents should be able to follow those preferences across a wide range of users. This thesis advances human-robot interaction by adapting autonomous agent tasks to user preferences from feedback. Our existing task architecture is behavior-based and encodes temporal constraints for the task. The architecture selects actions at runtime for efficiency, but users may have personal constraints that should also be encoded. This thesis proposes an approach that accounts for diverse user preferences by adjusting the architecture's action selection mechanism as soft suggestions, or by altering the task description entirely as hard changes to handle user preferences. In addition, we describe a large language model-driven preference elicitation system that bridges unstructured user feedback to our internal task representation. This research aims to produce an adaptable and interpretable task architecture for use with human users overseeing autonomous agents. We demonstrate our work in a simulated cooking environment, programmatically introducing preferences to test our approaches, and demonstrate the preference elicitation system as a proof-of-concept.

Level thesis:degree_level

Master’s Degree

Year dc:date.issued

2026

Author dc:creator

• Peterson, William Stanislaw

Advisor dc:contributor.advisor

• Nicolescu, Monica

Committee members dc:contributor.committeemember

• Feil-Seifer, David
• Panorska, Ania

• Autonomous Agents
• Human-Robot Interaction
• Preferences
• Task Trees

Language dc:language.iso

en_US, English

Repository record dc:identifier.uri

https://scholarwolf.unr.edu/handle/11714/11833

OAI identifier oai:identifier

oai:scholarwolf.unr.edu:11714/11833