Date of Award
Spring 5-12-2025
Degree Name
Master of Science (MS)
Degree Type
Thesis
Abstract
Reinforcement learning algorithms can enable autonomous systems to learn the control skills needed to accomplish a task specified by a linear temporal logic formula. However, they cannot be transferred to a new task, even when the two are very similar. For each new task, the policy must be redesigned from scratch, which is a common limitation of existing reinforcement learning methods for temporal logic tasks. A proposed solution to this problem leverages the similarity between past and new tasks to reuse already learned control skills to accomplish the new task, with minimal or no retraining.
Rather than learning a single policy that satisfies a task specified by a linear temporal logic formula, this approach decomposes the formula into reach-avoid subtasks, where policies are instead learned for each reach-avoid subtask and stored in a library of reusable control skills. Similarity metrics between two formulas are defined and algorithms for their computation are proposed. These are essential for determining if the library of reusable control skills contains the necessary policies to accomplish the new task. Finally, a simulation of an autonomous system that was trained to satisfy a temporal logic task is presented completing a new, unseen temporal logic task without any retraining.
Language
English (en)
Chair
Dr. Yiannis Kantaros
Committee Members
Dr. Sankalp Bhan Dr. Andrew Clark