SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization

Authors: Hong Qian, Yiyi Zhu, Xiang Shu, Shuo Liu, Yaolin Wen, Xin An, Huakang LU, Aimin Zhou, Ke Tang, Yang Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, baseline and state-of-the-art algorithms are tested and analyzed on SOO-Bench. Hopefully, SOO-Bench is expected to serve as a catalyst for the rapid developments of more novel and stable offline optimization methods.
Researcher Affiliation Collaboration Hong Qian, Yiyi Zhu, Xiang Shu, Shuo Liu, Yaolin Wen, Xin An, Huakang Lu, Aimin Zhou East China Normal University, China EMAIL, EMAIL Ke Tang Southern University of Science and Technology, China EMAIL Yang Yu Nanjing University, China Polixir Technologies, China EMAIL
Pseudocode Yes Algorithm 1 Offline Bayesian Optimization Trabucco et al. (2022) ... Algorithm 2 Covariance Matrix Adaptation Evolution Strategy (CMA-ES) Hansen (2006) ... Algorithm 3 Accumulative Risk Controlled Offline Optimization (ARCOO) Lu et al. (2023) ... Algorithm 4 Tri-mentoring for Offline Model-based Optimization Chen et al. (2023) ... Algorithm 5 Autofocused Model-based Optimization (Autofocusing Cb AS) Brookes et al. (2019) ... Algorithm 6 Tri-Training Data-Driven Evolutionary Algorithm (TTDDEA) Huang et al. (2021) ... Algorithm 7 Offline Data-Driven Optimization at Scale: A Cooperative Coevolutionary Approach (CCDDEA) Gong et al. (2023) ... Algorithm 8 Constrained Accumulative Risk Controlled Offline Optimization (CARCOO) ... Algorithm 9 Data-Driven Evolutionary Optimization with Penalty Function (DDEA-PF) Huang & Wang (2021b) ... Algorithm 10 Constrained Conservative Objective Models for Offline Optimization (CCOMs)
Open Source Code Yes The code is available at https: //github.com/zhuyiyi-123/SOO-Bench.
Open Datasets Yes GTOPX: Space Mission Optimization (Schlueter et al., 2021). ... The license for this dataset is GNU General Public License. ... CEC Task: Industrial and Design Optimization (Kumar et al., 2020). ... The license for this dataset is CC-BY 4.0 License. ... PROTEIN: DNA Sequence Optimization (Trabucco et al., 2022). ... You can access the data and code by visiting the URL https: //github.com/brandontrabucco/design-bench. ... The license for this dataset is MIT License.
Dataset Splits Yes In this paper, we select the middle 50% of the data (i.e., m% n% = 50%) to construct a simulated dataset as a reasonable baseline without leveraging any prior knowledge. Furthermore, when constructing the constraint dataset, we set the ratio of satisfying constraints to not satisfying constraints as 2 : 3.
Hardware Specification Yes The computing resources required for the research described in this paper are relatively modest, requiring only a single Nvidia Ge Force RTX 3090 GPU.
Software Dependencies No The paper mentions software like Bo Torch and various algorithm implementations from GitHub repositories but does not provide specific version numbers for these software dependencies. For example, 'We employ the quasi-expected Improvement (q EI) acquisition function within the Bo Torch framework (Balandat et al., 2019)' lacks a Bo Torch version.
Experiment Setup Yes We conduct a total of T = 150 optimization steps. ... CMA-ES: ... with the parameter σ set to 0.5. ... Autofocusing Cb AS: ... we have set this threshold to 0.9. ... ARCOO: ... initial momentum set at 0.2, ... sampling by 64 steps of Langevin dynamics. ... Tri-mentoring: ... number of neighborhood samples to 10. To maintain consistency, we standardized the number of optimization steps to 100 across all experiments.