Abstract
dc:description.abstractRecently, Contextual Bandit algorithms with stage-wise constraints have attracted much attention due to their applicability in real-world contexts such as determining optimal dosages or ad policies. Nonetheless, a drawback is that they ensure high-probability guarantees only for the expected cost signal. In this work, we address this gap in literature, and propose a linear contextual bandit formulation with stage-wise constraints, accounting for both efficacy and toxicity in the optimal dosage finding problem. These metrics are modeled, respectively, by a linear reward and a linear cost function. For this model, we consider the problem of maximizing the expected total cumulative reward (i.e., efficacy), while ensuring that the toxicity constraints are satisfied with high probability. By using a UCB-like approach, we are able to guarantee a regret of \tilde{\mathcal{O}}(\sqrt{T}) in high probability, while guaranteeing that the toxicity levels do not exceed some prespecified values with high probability. We further generalize our results to non-linear functions and provide a regret bound in terms of the eluder dimension, a measure of function class complexity.
Degree
thesis:*- Name thesis:degree_name
- Master of Science
- Level thesis:degree_level
- masters
- Discipline thesis:degree_discipline
- Computer Science
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Dragazis, Spyridon Konstantinos
- Advisor dc:contributor.advisor
-
- Pacchiano, Aldo
Subjects
dc:subject × 5Rights
dc:rights- Statement dc:rights
-
- Attribution 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53411