BOIDS: High-Dimensional Bayesian Optimization via Incumbent-Guided Direction Lines and Subspace Embeddings
Authors: Lam Ngo, Huong Ha, Jeffrey Chan, Hongyu Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive experimental results show that BOIDS outperforms state-of-the-art baselines on various synthetic and real-world benchmark problems. We demonstrate that BOIDS empirically outperforms state-of-the-art baselines on a comprehensive set of synthetic and real-world benchmark problems. We additionally conduct ablation study to understand each component of the proposed method. |
| Researcher Affiliation | Academia | 1RMIT University, Australia 2Chongqing University, China EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Overall, the BOIDS algorithm operates as follows (pseudocode in Alg. 1 and illustration in Fig. 1). Algorithm 1: The BOIDS Algorithm |
| Open Source Code | Yes | Code https://github.com/Lam Ngo1/boids |
| Open Datasets | Yes | For synthetic problems, we use Ackley-100D, Branin-500D and Hartmann-500D, which are widely-used in BO research works (Eriksson et al. 2019; Song et al. 2022; Ngo et al. 2024). For real-world benchmark problems, we use Mopta-124D, Lasso DNA-180D and Half Cheetah-102D. Mopta-124D is a vehicle design task aiming to minimize the vehicle s weight, with code implementation from Eriksson and Jankowiak, 2021. Lasso DNA-180D is a hyperparameter tuning task involving a microbiology DNA dataset, with code implementation from ˇSehi c et al., 2022. Half Cheetah-102D is a reinforcement learning task aiming to maximize a cumulative reward, with code implementation from Song et al., 2022. |
| Dataset Splits | No | The paper discusses benchmark problems like Ackley-100D, Branin-500D, Hartmann-500D, Mopta-124D, Lasso DNA-180D, and Half Cheetah-102D. However, it does not explicitly provide information regarding dataset splits (e.g., training/test/validation percentages or counts) for these problems in the main text. It mentions 'Most methods are run with 1000 iterations on all problems', which refers to the number of function evaluations in the Bayesian Optimization process, not data splitting. |
| Hardware Specification | No | The first author (L.N.) would like to thank the School of Computing Technologies, RMIT University, Australia for providing computing resources for this project. Additionally, this project was undertaken with the assistance of computing resources from RACE (RMIT AWS Cloud Supercomputing Hub). This statement mentions computing resources and a specific hub (RMIT AWS Cloud Supercomputing Hub) but lacks specific hardware details like CPU/GPU models, memory, or instance types. |
| Software Dependencies | No | For the surrogate model, we use a GP with Mat ern 5/2 kernel function. For acquisition function, we use Thompson Sampling and employ NSGA-II algorithm (Deb et al. 2002) as the MO optimizer for Eq. (4). While specific algorithms and functions are named, there are no explicit version numbers provided for any programming languages, libraries, or frameworks (e.g., Python version, PyTorch/TensorFlow version, or specific library versions for NSGA-II or the GP implementation). |
| Experiment Setup | Yes | The settings of BOIDS are as follows. For the incumbent-guided direction lines, we use a common configuration as with PSO, m = 20, w = 0.729, c1 = c2 = 2.05w (Shi and Eberhart 1998; Regis 2014; Li et al. 2020). For the surrogate model, we use a GP with Mat ern 5/2 kernel function. For acquisition function, we use Thompson Sampling and employ NSGA-II algorithm (Deb et al. 2002) as the MO optimizer for Eq. (4). |