Intelligent Batching for Bayesian Optimization

Although most BO implementations are sequential, it is possible to model the interactions between batches of configurations in order to distribute objective function evaluations on a cluster. This is a major feature which allows Bayesian Optimization to be put into production in a trading context. Batched Bayesian Optimization is particularly interesting when: - Your objective function is expensive to compute - You have multiple workers available to evaluate the recommended configurations The performance gains are illustrated in the attached figure. Increasing the size of the batch significantly improves the Rate of Convergence and the Accuracy of the optimization process. The figure on the left represents the difference between recommended and true global minimum and the figure on the right represents the distance between recommended and true global minimizers. You can find out how we implemented it in this article https://www.linkedin.com/pulse/batched-bayesian-optimization-charles-brecque/