High-Dimensional Dueling Optimization with Preference Embedding

In many scenarios of black-box optimization, evaluating the objective function values solutions is expensive, while comparing a pair relatively cheap, which yields dueling optimization. The side effect optimization that it doubles dimension solution space and exacerbates dimensionality scalability issue e.g., Bayesian To address this issue, existing methods fix one when throughout process, but may reduce their efficacy. Fortunately, has been observed that, in recommendation systems, results are mainly determined by latent human preferences. paper, we abstract phenomenon as preferential intrinsic inject into resulting embedding (PE-DBO). PE-DBO decouples pairwise comparison via matrix. Optimization performed subspace with much lower dimensionality, completed original space. Theoretically, disclose preference can be approximately preserved lower-dimensional subspace. Experiment verify on molecule discovery web page tasks, exists superior compared state-of-the-art (SOTA) methods.