Output details
11 - Computer Science and Informatics
University of Oxford
Bayesian Optimization in High Dimensions via Random Embeddings
<24>
Big data organizations typically have millions of users and complex systems implemented by different teams. Automation is essential to deliver personalized products and to configure hardware, software, interfaces and tests. Bayesian optimization is a powerful approach for increased automation in these domains, provided the number of parameters (dimensions) is small. Scaling to high dimensions has been the holy grail of Bayesian optimization for decades, but little progress was made due to the difficulty of the problem. Here, we introduce a novel approach that scales to high dimensions, and provide theoretical analyses. The paper received a prestigious IJCAI Distinguished Paper Award.