Bayesian Optimization with BASS
Thesis research on spline surrogates and how experimental machinery affects Bayesian-optimization comparisons.
My master’s thesis explores Bayesian Adaptive Spline Surfaces (BASS) as a surrogate model for Bayesian optimization. The work investigates whether spline-based surrogates can serve as a flexible, well-calibrated alternative to Gaussian processes when navigating expensive, black-box objective functions.
My contribution
I develop and evaluate BASS-based surrogates alongside Gaussian-process baselines, with shared experimental code and recorded results. This is my MSc thesis work at ITAM, expected in September 2026 after completing the coursework in 2022.
The companion manuscript, The Machinery Confound, investigates how candidate generation, acquisition optimization, and repeated evaluations can affect comparisons attributed to the surrogate model. It is a research manuscript, not a claim that one surrogate is universally better.
Inspect the work
The repository contains the manuscripts, experiment code, results, and bo-audit, tooling for inspecting objective evaluations. The central question is how to make comparisons fair enough to support the conclusions we draw.
Check out the code on GitHub.