High-Dimensional Bayesian Optimisation with Variational Autoencoders and Deep Metric Learning

We introduce a method combining variational autoencoders (VAEs) and deep\nmetric learning to perform Bayesian optimisation (BO) over high-dimensional and\nstructured input spaces. By adapting ideas from deep metric learning, we use\nlabel guidance from the blackbox function to structure the VAE latent space,\nfacilitating the Gaussian process fit and yielding improved BO performance.\nImportantly for BO problem settings, our method operates in semi-supervised\nregimes where only few labelled data points are available. We run experiments\non three real-world tasks, achieving state-of-the-art results on the penalised\nlogP molecule generation benchmark using just 3% of the labelled data required\nby previous approaches. As a theoretical contribution, we present a proof of\nvanishing regret for VAE BO.\n