Dissertation/Thesis Abstract

Adaptive control and learning using multiple models
by Wang, Yu, Ph.D., Yale University, 2017, 229; 10783473
Abstract (Summary)

Adaptation can have different objectives. Compared to a learning behavior, which is mainly to optimize the rewards/experience obtained through the learning process, adaptive control is a type of adaptation that follows a specific target guided by a controller. Although the targets may be different, the two types of adaption share common research interests.

One of the popular research techniques for studying adaptation is the use of multiple models, where the system will utilize information from multiple environment observers instead of one to improve the adaptation behavior in terms of stability, speed and accuracy. In this thesis, applications of multiple models for two types of adaptation, adaptive control and learning, will be investigated separately. For adaptive control, the research focuses on second-level adaptation, which is a new multiple-model-based approach; for learning, the multiple model concept is designed and embedded into a type of reinforcement scheme: learning automata.

The stability, robustness and performance of second-level adaptation will be first investigated in the context of various environments, including time-varying plants and noisy disturbances. Then, a new design of second-level adaptation for general systems and input-output accessible systems will be discussed. The reasons for the improved performance using second-level adaptation are analyzed theoretically. The second part of the thesis contributes to a new method of learning automata using multiple models. The method is first applied to a two-state (binary) reward environment in the simplest case, and it is later extended to the feed-forward case when multiple states or actions are presented. Finally, general reinforcement learning automata for network cases will be discussed. In all cases, simulation studies are given, wherever appropriate, to demonstrate the improvement in performance compared to conventional approaches.

Indexing (document details)
Advisor: Narendra, Kumapti S.
School: Yale University
School Location: United States -- Connecticut
Source: DAI-B 79/05(E), Dissertation Abstracts International
Subjects: Electrical engineering, Systems science, Artificial intelligence
Keywords: Adaptive Control, Artificial Intelligence, Learning Automata, Machine Learning, Multiple Models, Reinforcement Learning
Publication Number: 10783473
ISBN: 978-0-355-70939-1
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy