Posterior Sampling-Based Bayesian Optimization with Tighter Bayesian Regret Bounds

Authors: Shion Takeno, Yu Inatsu, Masayuki Karasuyama, Ichiro Takeuchi

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Furthermore, we demonstrate a wide range of experiments, focusing on the effectiveness of PIMS that mitigates the practical issues of GP-UCB and TS. Finally, we show broad experiments, particularly focusing on the practical effectiveness of PIMS compared with TS and GP-UCB-based methods.
Researcher Affiliation Academia 1Department of Engineering, Nagoya University, Aichi, Japan 2RIKEN AIP, Tokyo, Japan 3Department of Computer Science, Nagoya Institute of Technology, Aichi, Japan.
Pseudocode Yes Algorithm 1 TS
Open Source Code No The paper does not provide any explicit statement about releasing the source code for its described methodology, nor does it include a link to a code repository.
Open Datasets Yes We performed the experiments on synthetic functions generated from GP, benchmark functions, and emulators derived from real-world datasets. [...] benchmark functions called Ackley and Shekel functions in https://www.sfu.ca/~ssurjano/optimization.html. [...] The alkox dataset (Häse et al., 2021) is the measurement of alkoxylation reaction [...] The Fullerenes dataset (Walker et al., 2017) is the mole fraction of the desired products...
Dataset Splits No The paper states 'We generated 5 inputs for the initial dataset using the Latin hypercube sampling', but it does not specify explicit training, validation, and test dataset splits for its experiments. It refers to 'train' in the context of a 'Bayes neural network trained by each dataset' as used by others, not for its own BO experiment setup splits.
Hardware Specification No The paper does not specify the hardware (e.g., CPU, GPU models, or memory) used for running the experiments.
Software Dependencies No The paper mentions using 'random Fourier feature (RFF)' for posterior sampling but does not list any specific software dependencies or their version numbers, such as programming languages, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, scikit-learn).
Experiment Setup Yes We generated 5 inputs for the initial dataset using the Latin hypercube sampling(Loh, 1996). We fixed the noise variance σ2 = 10 6 and used the RBF kernel k(x, x ) = exp x x 2 2/(2ℓ2) , where ℓis hyperparameter. For this experiment, we used the theoretical confidence parameters, βt = 2 log |X|t2/ 2π for GP-UCB and ζt Exp (2 log (|X|/2) , 1/2) for IRGP-UCB. For MC estimation in MES and JES, we generated 10 MC samples. In this experiment, we employed the heuristic confidence parameters for GP-UCB and IRGP-UCB as βt = 0.2d log (2t) (Kandasamy et al., 2015) and ζt Exp (2/d, 1/2) (Takeno et al., 2023).