Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization

Authors: Samuel Daulton, Maximilian Balandat, Eytan Bakshy

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluation demonstrates that q EHVI is computationally tractable in many practical scenarios and outperforms state-of-the-art multi-objective BO algorithms at a fraction of their wall time.
Researcher Affiliation Industry Samuel Daulton Facebook sdaulton@fb.com Maximilian Balandat Facebook balandat@fb.com Eytan Bakshy Facebook ebakshy@fb.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Acquisition functions are available as part of the open-source library Bo Torch [5]. Code is available at https://github.com/pytorch/botorch.
Open Datasets Yes For synthetic problems, we consider the Branin-Currin problem (d = 2, M = 2, convex Pareto front) [6] and the C2-DTLZ2 (d = 12, M = 2, V = 1, concave Pareto front), which is a standard constrained benchmark from the MO literature [16]
Dataset Splits No The paper does not specify training, validation, and test dataset splits in the conventional sense for supervised learning tasks. It describes evaluation budgets and number of trials for optimization problems.
Hardware Specification Yes Table 1: Acquisition Optimization wall time in seconds on a CPU (2x Intel Xeon E5-2680 v4 @ 2.40GHz) and a GPU (Tesla V100-SXM2-16GB).
Software Dependencies No While BoTorch is mentioned as an open-source library used, specific version numbers for software dependencies (e.g., BoTorch, PyTorch) are not provided in the paper's main text.
Experiment Setup Yes Both plots show optimization performance on a DTLZ2 problem (d = 6, M = 2) with a budget of 100 evaluations (plus the initial quasi-random design). We report means and 2 standard errors across 20 trials.