Dissertation/Thesis Abstract

An End-to-End Framework for Control Synthesis from Demonstrations via Probabilistic Automata Learning
by Renninger, Nicholas, M.S., University of Colorado at Boulder, 2020, 111; 28087127
Abstract (Summary)

Formal Methods for Control Synthesis offer theoretically sound approaches to safety critical control of autonomous systems, but require specifying complex tasks via the difficult process of writing correct logical formulas, namely temporal logic formulas, to serve as task specifications. These formulas are typically translated to task specification automata and then composed with models of the autonomous systems to form new automata on which satisfying controllers are computed. Previous work has attempted to address the difficulty in temporal logic task specification by learning specification formulas as part of a new class of Learning from Demonstration (LfD) algorithms, where the learned outcome is a task specification. This thesis proposes bypassing the formula-to-automaton translation step by posing the task specification inference problem as a Grammatical Inference (GI) problem, allowing the use of existing GI algorithms to directly learn task specification automata from observations of the system captured during task demonstrations by an expert. Specifically, this thesis explores probabilistic automata learning using tools from the GI community, as well as the use of probabilistic automata as novel task specification models, which can encode expert task preferences. Then, an end-to-end framework combining probabilistic task specification automata learning and analysis with a novel control synthesis algorithm using these automata is presented. This framework is implemented in an open-source python library which has an interface to OpenAI's gym toolkit, allowing for automatic abstraction of existing, complex gridworld environments for easy experimentation with the proposed framework, as well as direct comparison with other LfD and reinforcement learning algorithms. Case studies on the performance characteristics of the proposed framework are presented to characterize and contextualize the approach in the broader control synthesis community.

Indexing (document details)
Advisor: Lahijanian, Morteza
Commitee: Ahmed, Nisar, Sankaranarayanan, Sriram
School: University of Colorado at Boulder
Department: Aerospace Engineering
School Location: United States -- Colorado
Source: MAI 82/3(E), Masters Abstracts International
Subjects: Computer science, Aerospace engineering, Artificial intelligence
Keywords: Control synthesis, Formal methods, Grammatical inference, Learning from demonstration, Probabilistic deterministic finite automaton (PDFA)), Task specification
Publication Number: 28087127
ISBN: 9798672128269
Copyright © 2021 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy