Large-Batch, Iteration-Efficient Neural Bayesian Design Optimization
Navid Ansari, Alireza Javanmardi, Eyke H\"ullermeier, Hans-Peter, Seidel, Vahid Babaei
TL;DR
This paper introduces a scalable Bayesian optimization framework designed for large batch, data-intensive engineering design problems, emphasizing iteration efficiency over sample efficiency, and demonstrating superior performance on real-world applications.
Contribution
A novel, scalable, sample-based acquisition function for Bayesian optimization that effectively handles large batches and multiple objectives using Bayesian neural networks.
Findings
Effective in data-intensive environments with few iterations
Outperforms state-of-the-art multi-objective optimization methods
Validated on airfoil design and 3D printing problems
Abstract
Bayesian optimization (BO) provides a powerful framework for optimizing black-box, expensive-to-evaluate functions. It is therefore an attractive tool for engineering design problems, typically involving multiple objectives. Thanks to the rapid advances in fabrication and measurement methods as well as parallel computing infrastructure, querying many design problems can be heavily parallelized. This class of problems challenges BO with an unprecedented setup where it has to deal with very large batches, shifting its focus from sample efficiency to iteration efficiency. We present a novel Bayesian optimization framework specifically tailored to address these limitations. Our key contribution is a highly scalable, sample-based acquisition function that performs a non-dominated sorting of not only the objectives but also their associated uncertainty. We show that our acquisition function…
Peer Reviews
Decision·Submitted to ICLR 2024
- The paper tackles an important problem which is performing multi-objective BO with the large-batch setting. - The paper proposes an acquisition function for applying large-batch when performing BO while the current acquisition functions (qEHVI, qParEGO, qNEHVI) struggle, in terms of computation time. The concept of the proposed acquisition function is intuitive: it seems to further encourage explorative behavior, because it also maximizes the uncertainties in the surrogate model.
- Some technical details are not described clearly, making it sometimes hard to catch the main idea of the paper. For example, the formal problem statement is not described, the proposed method makes use of only epistemic uncertainty but the concept of epistemic uncertainty is not explained in the Background. The organization of the paper is sometimes a bit confused, for example, the overall process of BO should not be placed in method section. - The use of BNN as the surrogate model to enhance
- The paper considers an important problem relevant to real-world applications in engineering design. - I especially like the real world evaluation on two interesting benchmarks: airfoil design and 3D printing. It would be an interesting contribution to the BO community if they are released in the open-source code. - The idea is simple and works well on the benchmarks.
- Although I like the simplicity of the approach, the reasoning behind choosing this instantiation of multiobjective optimization is not entirely clear. Please considering some more analysis about the principles behind the proposed acquisition function. - Some relevant related work that can be useful to discuss in the paper: - A very similar idea utilizing multiobjective acquisition function with predicted mean and variance as objectives. [1] Gupta, S., Shilton, A., Rana, S., & Venkatesh,
1. The paper considers a novel black-box optimization setting where there are multiple objectives and the batch size can be very large. The authors empirically observe that contemporary multi-objective batch acquisition functions do not scale well with respect to the batch size. 2. To solve this issue, the authors propose a modified version of Deep Ensembles to approximate BNN and an acquisition function to maximize both predicted objectives and the uncertainty measure.
1. Contribution is not enough. The innovation of the paper can be summarized into a new predictive model with a minor modification on the original Deep Ensembles model and a new multi-objective batch acquisition function. For the predictive model, the reason for modifying the uncertainty measurement part from aleatory noise to epistemic noise is unclear. Also, the benefit from this change is not verified in the paper. Moreover, the novelty of 2$\textit{M}$D acquisition compared to the other acqu
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdvanced Neural Network Applications · Advanced Multi-Objective Optimization Algorithms · Machine Learning and Data Classification
MethodsFocus · Gaussian Process