Fixed Confidence Best Arm Identification in the Bayesian Setting

Authors: Kyoungseok Jang, Junpei Komiyama, Kazutoshi Yamazaki

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

Reproducibility Variable Result LLM Response
Research Type Experimental Simulations verify the theoretical results.We conduct two experiments to demonstrate that the expected stopping times of frequentist δ-correct algorithms diverge in a Bayesian setting and that the elimination process in Algorithm 1 is necessary for more efficient sampling. In Tables 1 and 2, each column Avg , Max , and Error represents the average stopping time, maximum stopping time, and the ratio of the misidentification, respectively.
Researcher Affiliation Academia Kyoungseok Jang Universitá degli Studi di Milano ksajks@gmail.com Junpei Komiyama New York University / RIKEN AIP junpei@komiyama.info Kazutoshi Yamazaki The University of Queensland k.yamazaki@uq.edu.au
Pseudocode Yes Algorithm 1 Successive Elimination with Early-Stopping. Algorithm 2 No Elimination (No Elim) Algorithm.
Open Source Code Yes The code used in the experiments for this paper can be found at the following link: https://github.com/jajajang/FC_BAI_Bayes.
Open Datasets No The paper describes generating data based on a specified Gaussian prior distribution for simulations (e.g., "k = 2 arms with standard Gaussian prior distribution, which means mi = 0, ξi = 1 for all i [k]" and random numbers for prior means/variances in additional experiments). It does not use a pre-existing publicly available dataset with a concrete link, DOI, or formal citation.
Dataset Splits No The paper describes a simulation-based experimental setup (e.g., "ran N = 1000 Bayesian FC-BAI simulations") rather than traditional train/validation/test splits from a fixed dataset.
Hardware Specification Yes Hardware We used Python 3.7 as our programming language and Macbook Pro M2 16 inch as our hardware.
Software Dependencies Yes Hardware We used Python 3.7 as our programming language and Macbook Pro M2 16 inch as our hardware.
Experiment Setup Yes We design an experiment setup that has k = 2 arms with standard Gaussian prior distribution, which means mi = 0, ξi = 1 for all i [k]. We set δ = 0.1 and ran N = 1000 Bayesian FC-BAI simulations to estimate the expected stopping time and success rate.