Bayesian optimization improves by using cheap paired auxiliary data
Gaussian Process Bandits with Paired Auxiliary Observations
Machine Learning
Summary
Optimizing complex systems can take a long time because each test is noisy and expensive. The authors propose a method that uses cheaper related data collected beforehand to make better guesses during optimization. Their approach learns how to correct this auxiliary data based on paired test results, which speeds up learning without needing the cheap data to closely match the real objective. They show mathematically and experimentally that their method reduces mistakes and performs better than traditional approaches.
What this means in practice
- •For machine learning engineers: Improve tuning of expensive models by integrating prior cheap data with online experiments for faster convergence.
- •For industrial process control teams: Use low-cost sensor data to speed up optimization of costly system parameters with fewer expensive tests.
Authors
Weibo Huang, Qing Feng, Ruihao Zhu, Cheng Hua
Abstract
Bayesian optimization (BO) is widely used for expensive black-box optimization, yet observation noise makes each evaluation less informative and slows posterior learning. Standard BO learns from costly outcomes gathered during optimization. In many applications, however, low-cost auxiliary data are available beforehand. We introduce Auxiliary-Informed Gaussian Process Bandit (AI-GPB), a variance-reduction method that combines an offline estimate of the auxiliary mean with a correction coefficient learned from paired online observations. Notably, it exploits correlation between paired observation errors without requiring the auxiliary mean to approximate the objective. We establish finite-sample GP confidence bounds and regret guarantees, including explicit cumulative and simple regret rates for Matérn kernels. Numerical experiments across continuous BO benchmarks with varied landscapes and dimensions show consistent reductions in cumulative and simple regret relative to vanilla GP-UCB, even when within-pair correlation is moderate and the auxiliary mean is estimated from noisy offline data.