Density Ratio Estimation-based Bayesian Optimization with Semi-Supervised Learning

Authors: Jungtaek Kim

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we show the empirical results of our methods and several baseline methods in two distinct scenarios with unlabeled point sampling and a fixed-size pool, and analyze the validity of our methods in diverse experiments.
Researcher Affiliation Academia 1University of Wisconsin Madison, Madison, WI 53706, USA. Correspondence to: Jungtaek Kim <EMAIL>.
Pseudocode Yes Algorithm 1 DRE-BO-SSL Algorithm 2 Labeling Unlabeled Data
Open Source Code No The paper mentions that its methods are implemented using various libraries, including Bayes O (Kim & Choi, 2023), which is an open-source framework for Bayesian optimization. However, it does not explicitly state that the *specific implementation code for the methodology described in this paper* (DRE-BO-SSL) is released or provide a direct link to it.
Open Datasets Yes We run several synthetic functions for our methods and the baseline methods. Tabular Benchmarks (Klein & Hutter, 2019) NATS-Bench (Dong et al., 2021) 64D minimum multi-digit MNIST search... in the MNIST dataset (Le Cun et al., 1998).
Dataset Splits Yes To train and test the model fairly, we create a training dataset of 440,000 three-digit images, a validation dataset of 40,000 three-digit images, and a test dataset of 80,000 three-digit images using a training dataset of 55,000 single-digit images, a validation dataset of 5,000 single-digit images, and a test dataset of 10,000 single-digit images in the original MNIST dataset (Le Cun et al., 1998).
Hardware Specification Yes To carry out the experiments in our work, we use dozens of commercial Intel and AMD CPUs such as Intel Xeon Gold 6126 and AMD EPYC 7302. For the experiments on minimum multi-digit MNIST search, the NVIDIA GeForce RTX 3090 GPU is used.
Software Dependencies No Our proposed methods and baseline methods are implemented with scikit-learn (Pedregosa et al., 2011), PyTorch (Paszke et al., 2019), NumPy (Harris et al., 2020), SciPy (Virtanen et al., 2020), XGBoost (Chen & Guestrin, 2016), and Bayes O (Kim & Choi, 2023). While these software components are listed with citations, specific version numbers for the libraries/packages used in the experiments are not provided.
Experiment Setup Yes All experiments are repeated 20 times with 20 fixed random seeds, where 5 initial points are given to each experiment. By following the previous work by Tiao et al. (2021); Song et al. (2022), we set a threshold ratio as ζ = 0.33 for all experiments. To solve (6), we use L-BFGS-B (Byrd et al., 1995) with 1,000 different initializations. MLP, BORE and LFBO: These methods are built with two-layer fully-connected networks; the detail of the multi-layer perceptron is described as follows. The architecture of the multi-layer perceptron is set as the following: (i) first layer: fully-connected, input dimensionality d, output dimensionality 32, ReLU; (ii) second layer: fully-connected, input dimensionality 32, output dimensionality 1, Logistic. The Adam optimizer (Kingma & Ba, 2015) with learning rate 1 × 10−3 is used to train the network for 100 epochs.