A portfolio approach to massively parallel Bayesian optimization
Abstract: One way to reduce the time of conducting optimization studies is to evaluate designs in parallel rather than just one-at-a-time. For expensive-to-evaluate black-boxes, batch versions of Bayesian optimization have been proposed. They work by building a surrogate model of the black-box to simultaneously select multiple designs via an infill criterion. Still, despite the increased availability of computing resources that enable large-scale parallelism, the strategies that work for selecting a few tens of parallel designs for evaluations become limiting due to the complexity of selecting more designs. It is even more crucial when the black-box is noisy, necessitating more evaluations as well as repeating experiments. Here we propose a scalable strategy that can keep up with massive batching natively, focused on the exploration/exploitation trade-off and a portfolio allocation. We compare the approach with related methods on noisy functions, for mono and multi-objective optimization tasks. These experiments show orders of magnitude speed improvements over existing methods with similar or better performance.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: We thank the reviewers for their helpful and detailed comments. We appreciated the positive comments on the originality of tackling large batch sizes. We are not aware of other works enabling such large batches while handling the exploration-exploitation-replication trade-off.
We carefully modified the paper to address the issues raised by the reviewers.
In particular:
- We have streamlined the Introduction and moved most of the related works to the background Section 2 to improve the exposition.
- We have largely modified Section 3 to clarify the adaptation of the HSRI portfolio selection strategy to the Bayesian Optimization context and now provide further detail in Algorithm 2 (previously Algorithm 1).
- New experiments are provided to show the scaling of qHSRI with larger $q$ on the synthetic benchmarks. As in the two realistic test cases, the only available method to compare with in reasonable time is random search.
We further reply to the specific comments and questions of the reviewers from their individual reviews.
Assigned Action Editor: ~Michael_U._Gutmann1
Submission Number: 800
Loading