Theses and Dissertations

Date of Award

7-1-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mathematics

First Advisor

Oraby Tamer

Second Advisor

Zhuanzhuan Ma

Third Advisor

Hansapani Rodrigo

Abstract

Reinforcement Learning (RL) is a computational framework for sequential decision-making in which an agent learns optimal strategies through direct interaction with its environment. Unlike many traditional machine learning methods that rely primarily on datasets, RL emphasizes adaptive learning through trial-and-error and feedback from the environment. In recent years, RL has been treated as a powerful approach for solving complex dynamic problems in areas such as artificial intelligence, ecology, epidemiology, and public health. Common RL algorithms include Q-learning, Deep Q-Networks (DQN), REINFORCE, Actor-Critic methods, Soft Actor-Critic (SAC), Proximal Policy Optimization (PPO).

This dissertation focuses on the application of reinforcement learning and machine learning techniques to problems in vaccination behavior, ecological prediction, and epidemiological control. The first study develops a mathematical framework for vaccine acceptance dynamics by considering a binary decision process between vaccination and non-vaccination. A time-dependent vaccination quality function is introduced to represent adaptive learning through experience and information, inspired by reinforcement learning. The proposed model incorporates important societal factors, including vaccination cost, social norms, and bounded rationality, to examine their effects on disease dynamics. Analytical and numerical results indicate that reducing vaccination costs below a critical societal threshold and increasing initial vaccine acceptance significantly improve the likelihood of disease eradication. Furthermore, the study suggest that adaptive vaccination cost, together with strong social norms and lower levels of bounded rationality, can eliminate disease even when the disease has a relatively large basic reproduction number.

The second study focuses on an ecological application involving the prediction of the historical presence of Culex pipiens mosquitoes across the United States. The performance of several machine learning and reinforcement learning approaches is compared, including logistic regression, random forests, deep neural networks, Q-learning, DQN, REINFORCE, and Actor-Critic methods. Results show that traditional supervised learning models achieve higher predictive accuracy, while RL-based approaches exhibit better flexibility and avoid overfitting through continual learning from new data. These findings suggest that RL methods have strong potential for dynamic and real-time ecological prediction tasks.

The third study investigates epidemic intervention strategies for controlling chronic wasting disease (CWD) in deer populations under stochastic environmental conditions. A stochastic epidemiological model incorporating both environmental and demographic variability is developed to better capture uncertainties associated with disease transmission and wildlife behavior. PPO is applied to learn adaptive intervention policies involving population management and environmental decontamination strategies. Numerical simulations demonstrate that the learned policies remain robust under different epidemic intensities and stochastic fluctuations. The results demonstrate that this flexible RL framework for adaptive epidemic control can be extended to a broader class of infectious disease systems.

Comments

Copyright 2026 Wei Yin. All Rights Reserved. https://proquest.com/docview/3409129350

Share

COinS