Optimal Multi-Fidelity Best-Arm Identification

Authors: Riccardo Poiani, Rémy Degenne, Emilie Kaufmann, Alberto Maria Metelli, Marcello Restelli

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

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, in Section 5, we present the results of numerical experiments which demonstrate the good empirical performance of our new algorithm compared to prior work. We conclude this work by presenting numerical simulations whose goal is to show the empirical benefits of our approach. We compare MF-GRAD against IISE [25], and the gradient approach of [22] that simply does BAI using samples collected at fidelity M.
Researcher Affiliation Academia Riccardo Poiani DEIB, Politecnico di Milano, Milan, Italy riccardo.poiani@polimi.it Rémy Degenne Univ. Lille, Inria, CNRS, Centrale Lille, UMR 9189-CRISt AL, F-59000 Lille, France remy.degenne@inria.fr Emilie Kaufmann Univ. Lille, CNRS, Inria, Centrale Lille, UMR 9189-CRISt AL, F-59000 Lille, France emilie.kaufmann@univ-lille.fr Alberto Maria Metelli DEIB, Politecnico di Milano, Milan, Italy albertomaria.metelli@polimi.it Marcello Restelli DEIB, Politecnico di Milano, Milan, Italy marcello.restelli@polimi.it
Pseudocode Yes We present our solution for solving MF-BAI problems, an algorithm called Multi-Fidelity Sub Gradient Ascent (MF-GRAD). Its pseudocode can be found in Algorithm 1.
Open Source Code Yes We provide the codebase with instructions on how to reproduce the results.
Open Datasets No Given this setup, first, we test all methods on a 4 5 multi-fidelity bandit with Gaussian arms that have been randomly generated, using a risk parameter δ = 0.01.
Dataset Splits No We test all methods on a 4 5 multi-fidelity bandit with Gaussian arms that have been randomly generated, using a risk parameter δ = 0.01.
Hardware Specification Yes For the experiments we relied on a server with 100 Intel(R) Xeon(R) Gold 6238R CPU @ 2.20GHz cpus and 256GB of RAM.
Software Dependencies No No specific software dependencies with version numbers are mentioned in the paper.
Experiment Setup Yes using a risk parameter δ = 0.01. We also define αt = 1/t and γt = 1/(4 * t). βt,δ = log(K/δ) +M log(log(t) +1)