Bayesian Optimization over Hybrid Spaces

Authors: Aryan Deshwal, Syrine Belakaria, Janardhan Rao Doppa

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on synthetic and six diverse realworld benchmarks show that Hy BO significantly outperforms the state-of-the-art methods.
Researcher Affiliation Academia 1School of EECS, Washington State University, Pullman, USA.
Pseudocode Yes Algorithm 1 Hy BO Approach
Open Source Code Yes The code and data are available on the Git Hub repository https://github.com/ aryandeshwal/Hy BO.
Open Datasets Yes bbox-mixint is a challenging mixed-integer blackbox optimization benchmark suite (Tuˇsar et al., 2019) that contains problems of varying difficulty. This benchmark suite is available via COCO platform1. ... We employ six diverse realworld domains. The complete details (function definition, bounds for input variables etc.) are in the Appendix. ... We consider hyper-parameter tuning of a neural network model on a diverse set of benchmarks (Gijsbers et al., 2019)
Dataset Splits No The paper discusses 'training data' and 'testing dataset' but does not explicitly provide details about dataset splits for training, validation, or testing, such as percentages or sample counts.
Hardware Specification Yes All experiments were run on a AMD EPYC 7451 24-Core machine.
Software Dependencies No The paper mentions using specific open-source Python implementations and libraries (e.g., CMA-ES, CoCaBO, SMAC, TPE) but does not provide their specific version numbers required for reproducibility.
Experiment Setup Yes We configure Hy BO as follows. We employ uniform prior for the length scale hyperparameter (σ) of the RBF kernel. Horse-shoe prior is used for β hyper-parameter of the discrete diffusion kernel (Equation 4.8) and hyper-parameters θ of the additive diffusion kernel (Equation 4.9). We employ expected improvement (Mockus et al., 1978) as the acquisition function. For acquisition function optimization, we perform iterative search over continuous and discrete sub-spaces as shown in Algorithm 1. For optimizing discrete subspace, we run local search with 20 restarts. We normalize each continuous variable to be in the range [ 1, 1] and employed CMA-ES algorithm 3 for optimizing the continuous subspace. We found that the results obtained by CMA-ES were not sensitive to its hyper-parameters. Specifically, we fixed the population size to 50 and initial standard deviation to 0.1 in all our experiments.