Optimal Top-Two Method for Best Arm Identification and Fluid Analysis
Authors: Agniv Bandyopadhyay, Sandeep Juneja, Shubhada Agrawal
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 6, we describe the numerical experiments. Detailed proof of all results are in the appendix. ... In Section 6, we numerically demonstrate the dynamics followed by the algorithm AT2, and also compare its performance against the β-EB-TCB algorithm of [16] for different values of β, and TCB algorithm of [22]. ... In Appendix J, we demonstrate by several examples that both the AT2 and IAT2 algorithms significantly outperform the β-EB-TCB and β-EB-ITCB of [16] when β is chosen different from the optimal β. |
| Researcher Affiliation | Academia | Agniv Bandyopadhyay TIFR Mumbai, India agniv.bandyopadhyay@tifr.res.in Sandeep Juneja Ashoka University, India sandeep.juneja2010@gmail.com Shubhada Agrawal Georgia Institute of Technology, USA sagrawal362@gatech.edu |
| Pseudocode | Yes | In the appendix, we provide pseudo-codes of AT2 and IAT2 in Algorithms 1 and 2, respectively. ... Algorithm 1: Anchored Top-Two (AT2) Algorithm ... Algorithm 2: Improved Anchored Top-Two (IAT2) Algorithm |
| Open Source Code | Yes | Reproducibility: Our code is implemented in Julia 1.7.1, and the plots are generated with the Plots.jl package. Other dependencies are listed in the Readme.md file, which also includes instructions to reproduce the figures and tables presented here. We build upon the publicly available code for [16]. |
| Open Datasets | No | The paper describes using '4 armed Gaussian bandit with unit variance and mean vector µ = [10, 8, 7, 6.5]' and 'Bernoulli bandit with means µ = [0.91, 0.73, 0.64, 0.59]'. While these are standard distribution types, the specific instances (parameter values) are created for the simulations and no link or citation to a public, downloadable dataset is provided. |
| Dataset Splits | No | The paper describes various experimental setups, including the parameters of the bandit instances and confidence levels. However, it does not explicitly mention the use of 'training/test/validation dataset splits' or their specific percentages/methodology for reproducibility. |
| Hardware Specification | Yes | Our experiments are conducted on an institutional cluster computing facility having an Intel Xeon Gold 6130 2.1GHz CPU with 32 cores. |
| Software Dependencies | Yes | Reproducibility: Our code is implemented in Julia 1.7.1, and the plots are generated with the Plots.jl package. Other dependencies are listed in the Readme.md file, which also includes instructions to reproduce the figures and tables presented here. |
| Experiment Setup | Yes | The AT2 algorithm takes in confidence parameter δ > 0 and exploration parameter α (0, 1) as inputs... We consider a 4 armed Gaussian bandit with unit variance and mean vector µ = [10, 8, 7, 6.5]... The error probability δ in both these experiments is set to 0.001. |