Bayesian optimization is a standard technique for expensive black-box problems, but observation noise can make each evaluation less informative and slow down posterior learning. A new preprint on arXiv, titled Gaussian Process Bandits with Paired Auxiliary Observations, addresses this issue directly.
The authors propose leveraging paired auxiliary observations—additional data collected alongside the main costly outcome—to improve the learning process. While the abstract is truncated, the stated motivation is clear: standard BO relies only on the costly outcomes, and the new method aims to use auxiliary signals to mitigate noise and accelerate optimization.
Because the full paper is not yet available in the abstract, the specific mechanism and experimental results remain unspecified. The work appears to target settings where evaluations are expensive and noisy, making auxiliary information a potentially valuable resource for more efficient black-box optimization.