Decision making under uncertainty: randomization, online learning, and applications
File(s)
Author(s)
Wang, Zhengchao
Type
Thesis
Abstract
This dissertation develops data-driven decision-making tools for operational problems under uncertainty, with applications in assortment planning and non-profit fundraising. It addresses key challenges in predictive and prescriptive analytics, such as sparse data and the amplification of estimation errors through optimization.
By building on and extending foundational paradigms including robust optimization and online learning, this work contributes both new methodologies and practical insights.
The first part of the dissertation studies randomized robust assortment optimization.
We show that in many practical settings, offering a single deterministic assortment is suboptimal under model uncertainty.
We introduce the notion of randomization-receptiveness and analyze its implications across three widely used customer choice models.
We further provide tractable algorithms for computing optimal randomized strategies and demonstrate their practical value using both synthetic and real-world datasets.
The second part of the dissertation
focuses on donor targeting in annually recurring fundraising campaigns.
In collaboration with a global non-profit organization, we develop multi-armed bandit algorithms that incorporate clustering to identify donors' latent campaign preferences.
Due to the short lifetime of donors, learning their preferences from limited interaction histories presents a fundamental bias-variance trade-off.
To address this, we introduce a clustering approach that enables more effective generalization across donors.
We propose tailored algorithms for settings where donor types are either known in advance or must be inferred, and demonstrate both theoretical guarantees and strong empirical performance on real-world data from nearly one million donors.
Together, these projects advance our understanding of robust and adaptive decision-making under uncertainty, offering theoretically grounded solutions that are immediately relevant to high-impact operational contexts.
By building on and extending foundational paradigms including robust optimization and online learning, this work contributes both new methodologies and practical insights.
The first part of the dissertation studies randomized robust assortment optimization.
We show that in many practical settings, offering a single deterministic assortment is suboptimal under model uncertainty.
We introduce the notion of randomization-receptiveness and analyze its implications across three widely used customer choice models.
We further provide tractable algorithms for computing optimal randomized strategies and demonstrate their practical value using both synthetic and real-world datasets.
The second part of the dissertation
focuses on donor targeting in annually recurring fundraising campaigns.
In collaboration with a global non-profit organization, we develop multi-armed bandit algorithms that incorporate clustering to identify donors' latent campaign preferences.
Due to the short lifetime of donors, learning their preferences from limited interaction histories presents a fundamental bias-variance trade-off.
To address this, we introduce a clustering approach that enables more effective generalization across donors.
We propose tailored algorithms for settings where donor types are either known in advance or must be inferred, and demonstrate both theoretical guarantees and strong empirical performance on real-world data from nearly one million donors.
Together, these projects advance our understanding of robust and adaptive decision-making under uncertainty, offering theoretically grounded solutions that are immediately relevant to high-impact operational contexts.
Version
Open Access
Date Issued
2025-07-14
Date Awarded
01/09/2025
License URL
Advisor
Wiesemann, Wolfram
Peura, Heikki
Sponsor
Imperial College London
Engineering and Physical Sciences Research Council
Grant Number
EP/R045518/1
EP/T024712/1
EP/W003317/1
Publisher Department
Business School
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)