Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (CiBO)

CiBO introduces a novel framework that combines flow-based models with posterior inference in latent space to solve high-dimensional constrained black-box optimization problems. By iteratively training flow-based models to capture data distributions and surrogate models to predict function values and constraint violations, our method effectively handles the challenges of high-dimensional constrained optimization. The key innovation lies in performing posterior inference in the latent space of flow-based models, which helps overcome the multi-modality and plateau issues in the posterior distribution, especially when dealing with binary constraint feedback. Our approach demonstrates superior performance across various synthetic and real-world tasks, offering a scalable solution to constrained black-box optimization problems.

For a visual explanation of the CiBO framework, please see the image below:

You can find the workshop paper on OpenReview, the latest preprint on arXiv, and the code on GitHub.

Recommended citation: Om, K.*, Sim, K.*, Yun, T.*, Kang, H., & Park, J. (2025). "Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization." NeurIPS Workshop on Structured Probabilistic Inference & Generative Modeling (Oral).
Download Paper