Bayesian Optimization of High-dimensional Outputs with Human Feedback

We consider optimizing the inputs to a black-box function that produces high-dimensional outputs such as natural language, images, or robot trajectories. A human decision maker (DM) has a utility function over these outputs. We may learn about the DM's utility by presenting a small set of outputs and asking which one they prefer. We may learn about the black-box function by evaluating it at adaptively chosen inputs. Given a limited number of such learning opportunities, our goal is to find the input to the black box that maximizes the DM's utility for the output generated. Previously proposed methods for this and related tasks either do not scale to high dimensional outputs or are statistically inefficient because they ignore information in the outputs. Our proposed approach overcomes these challenges using Bayesian optimization and a novel embedding of high-dimensional outputs into a low-dimensional latent space customized for this task. This embedding is designed to both minimize error when reconstructing high-dimensional outputs and support accurate prediction of human judgments. We demonstrate that this approach significantly improves over baseline methods.