Latent Space Bayesian Optimization

Latent space Bayesian optimization (LSBO) aims to efficiently optimize complex, black-box functions by leveraging generative models, often variational autoencoders (VAEs), to map high-dimensional input spaces into lower-dimensional latent spaces where Bayesian optimization techniques can be more effectively applied. Current research focuses on improving exploration within these latent spaces, addressing the mismatch between generative model and optimization objectives, and developing efficient surrogate models for high-dimensional latent spaces, including the use of trust regions and in-context learning. LSBO shows promise for accelerating optimization in diverse fields, such as materials science, drug discovery, and hyperparameter tuning for machine learning models, by significantly reducing the computational cost of exploring vast search spaces.