University of Minnesota
Analyzing Benchmarks and Methods for Contextual Bayesian Optimization Towards Personalized Healthcare Decisions
Abstract
dc:description.abstractIn low-budget personalization problems, an optimizer must choose actions for a new target task using only contextualinformation and a small number of online evaluations. Transfer from related tasks is therefore a natural strategy. Existing contextual and transfer optimization methods provide useful guideposts in this regime, but they are rarely studied under a shared offline-to-online benchmark that isolates the tradeoff between source-task transfer and target-task adaptation. This thesis introduces a benchmark suite for low-budget offline-to-online contextual optimization and a modular framework that decomposes methods into a fixed transfer component and a target-task adaptation mechanism. This framing enables direct comparison between warm-start methods and persistent transfer strategies across real, simulator-backed, and semisynthetic benchmarks. Across the benchmark suite, transfer is often beneficial, especially at initialization, but persistent transfer is not uniformly reliable and can suppress target-task adaptation when the transfer model is misaligned. Across the full online budget, warm-start methods are among the strongest overall performers, suggesting that in this regime transfer is often most valuable as an initializer rather than as a mechanism used throughout optimization.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Sortland, Eric
Subjects
dc:subject × 3Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Sortland, Eric. (2026). Analyzing Benchmarks and Methods for Contextual Bayesian Optimization Towards Personalized Healthcare Decisions. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/281697.