CobBO: Coordinate Backoff Bayesian Optimization with Two-Stage Kernels

Jian Tan, Niv Nayman, MengChang WANG, Rong Jin

Keywords: Bayesian Optimization

TL;DR: Bayesian optimization of high dimensional objective functions

Abstract:

Bayesian optimization is a popular method for optimizing expensive black-box functions. Yet it oftentimes struggles in high dimensions where a sufficient estimation of the global landscape requires more observations and the computation cost becomes prohibitively expensive.
We introduce Coordinate backoff Bayesian optimization (CobBO) with two-stage kernels to alleviate this problem.
In each iteration, a promising subset of coordinates is selected in the first stage, as past observed points in the full space are projected to the selected subspace adopting a simple kernel that sacrifices the approximation accuracy for computational efficiency.
Then in the second stage of the same iteration a more sophisticated kernel is applied for estimating the landscape in the selected low dimensional subspace where the computational cost becomes affordable. Effectively, this second stage kernel refines the approximation of the global landscape estimated by the first stage kernel through a sequence of observations in the local subspace. This refinement lasts until a stopping rule is met determining when to back off from a certain subspace and switch to another coordinate subset.
This decoupling significantly reduces the computational burden in high dimensions, while the two-stage kernels of the Gaussian process regressions fully leverage the observations in the whole space rather than only relying on observations in each coordinate subspace.
Extensive evaluations show that CobBO finds solutions comparable to or better than other state-of-the-art methods for dimensions ranging from tens to hundreds, while reducing the trial complexity and computational costs.

Code Of Conduct: I certify that all co-authors of this work have read and commit to adhering to the NeurIPS Statement on Ethics, Fairness, Inclusivity, and Code of Conduct.

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/cobbo-coordinate-backoff-bayesian/code)