New acquisition speeds Bayesian optimization on the unit cube
Original title: Acquisition for Spherical-Linear Bayesian Optimization on the Cube
The proposed acquisition is globally solvable faster than multi-start quasi-Newton and matches Gaussian-process accuracy per evaluation on cheaper black boxes.
The paper studies the complexity of globally solving Bayesian optimization acquisition subproblems when the decision region is the unit cube. Spherical domains already allow rapid exact acquisition, but the cube usually relies on multi-start quasi-Newton. The authors propose a new acquisition they say takes significantly less time to optimize globally while maintaining or improving performance, and is competitive with Gaussian-process BO on accuracy per evaluation.
Unconfirmed: Speed and accuracy claims are the authors' reported comparisons, not an independent benchmark.
- The target domain is the unit cube, the common BO setting.
- It is positioned against multi-start quasi-Newton acquisition solving.
- Claimed wall-time gains apply to low-to-moderate-cost black boxes.
From the original article
1 min read at arxiv.orgRecent work demonstrated that executing Bayesian Optimization (BO) using simply a linear surrogate with a clever transformation is enough to approximate the performance of the standard Gaussian-Process (GP) based version on high-dimensional test functions. When the decision region is a sphere, this strategy admits rapid exact acquisition.
By Nathan Wycoff · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.