Quantile Stein Variational Gradient Descent for Batch Bayesian Optimization
Authors: Chengyue Gong, Jian Peng, Qiang Liu
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on a number of real-world applications, and show that it consistently outperforms other recent state-of-the-art batch Bayesian optimization methods. |
| Researcher Affiliation | Academia | 1Department of Computer Science, UT Austin 2University of Illinois at Urbana Champaign, IL. |
| Pseudocode | Yes | Algorithm 1 Quantile Stein Variational Gradient Descent Algorithm 2 Quantile Stein Bayesian Optimization (QSBO) |
| Open Source Code | No | The paper mentions and links to open-source implementations of *baselines* (MACE, BUCB, LP-UCB) and a related work (SVGD by Liu & Wang), but does not provide a link or explicit statement about the availability of their *own* code (QSBO). |
| Open Datasets | Yes | We test our method and the baselines in ten benchmark problems from Bingham (2019)... We focus on pre-breeding genomic prediction using rice dataset... (Zhao et al., 2011) 4. ... The dataset contains 43k genotypes whose traits have been pre-evaluated by researchers. For our simulation, we find the nearest neighbors of the query points as the true evaluate points at each iteration of BO; this is equivalent to running standard BO on a piece-wise constant function constructed by a nearest neighbor regression on the dataset. ... The VAE is trained on the ZINC dataset5. |
| Dataset Splits | No | The paper describes an online optimization process where points are evaluated and added to the dataset iteratively. It does not define or use traditional training, validation, and test splits for a static dataset in the context of model evaluation. |
| Hardware Specification | No | The paper mentions "We would like to acknowledge Google Cloud for their support." but does not provide any specific details about the hardware used (e.g., CPU, GPU models, memory, or specific Google Cloud instance types). |
| Software Dependencies | No | The paper mentions that an RBF kernel is used and Ada Delta is the optimizer, but it does not list specific software dependencies with version numbers (e.g., Python version, TensorFlow/PyTorch version, specific library versions). |
| Experiment Setup | Yes | The hyperparameter η is chosen by following the setting in Srinivas et al. (2009), with η = p log td/2+2π2/3/δ, where t is the iteration step and δ is set as 0.05. We set the number of quantile SVGD steps at each iteration to be 600, with a constant learning rate of 0.1. The kernel in quantile SVGD is the standard RBF kernel and the optimizer is Ada Delta as suggested in the open-source code by Liu & Wang (2016)3. Empirically, we find it is useful to turn off the entropy regularization (setting τ = 0) at the last 10% iterations of quantile SVGD (when the different particles have already been pushed into different modes), so that we can find the exact local optima without the impact from the entropy regularization term. We investigated the choice of the entropy regularization coefficient τ and risk aversion coefficient λ, and found that the results are not sensitive to these parameters once they are the right region. We set τ = 5 10 2 and λ = 1 in all the experiments unless specified otherwise, and recommend it as a general default setting when using UCB-based acquisition. |