Near-Optimal Collaborative Learning in Bandits

Authors: Clémence Réda, Sattar Vakili, Emilie Kaufmann

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We propose in Appendix E an empirical evaluation of W-CPE-BAI for the weight matrix wm,n = α1(n = m) + 1 α M which corresponds to the setting studied by [30]. Our experiments on a synthetic instance show that W-CPE-BAI and PF-UCB-BAI have similar performances in terms of exploration cost and that W-CPE-BAI becomes better when the level of personalization α is smaller than 0.5.
Researcher Affiliation Collaboration Clémence Réda Université Paris Cité, Inserm, Neuro Diderot, F-75019 Paris, France clemence.reda@inria.fr Sattar Vakili Media Tek Research, Cambourne Business Park, CB23 6DW, United Kingdom sattar.vakili@mtkresearch.com Emilie Kaufmann Université de Lille, CNRS, Inria, Centrale Lille, UMR 9189 CRISt AL, F-59000 Lille, France emilie.kaufmann@univ-lille.fr
Pseudocode Yes Algorithm 1 Weighted Collaborative Phased Elimination for Best Arm Identification (W-CPE-BAI)
Open Source Code Yes Our code and run traces are available in an open-source repository. 5https://github.com/clreda/near-optimal-federated
Open Datasets No The paper states: 'We evaluated W-CPE-BAI against PF-UCB-BAI on a synthetic instance. We generated M = 10 agents and K = 10 arms, with fixed means (µk,m)k,m .' However, it does not provide concrete access information (link, DOI, citation) to this synthetic instance or the method to reproduce it.
Dataset Splits No The paper mentions generating a 'synthetic instance' and averaging results over '100 random seeds', but it does not describe explicit training, validation, or test dataset splits.
Hardware Specification Yes All the numerical experiments were performed on a personal computer (Intel(R) Core(TM) i5-3210M CPU @ 2.50GHz x 4, 8GB RAM) using Python 3.8.10.
Software Dependencies Yes All the numerical experiments were performed on a personal computer (Intel(R) Core(TM) i5-3210M CPU @ 2.50GHz x 4, 8GB RAM) using Python 3.8.10. The solvers used for the optimization problems are CVXPY 1.2.0 [9] and MOSEK 9.3.13 [27] (with SciPy 1.7.3 [33] as linear algebra backend).
Experiment Setup No The paper specifies the generation of the synthetic instance (e.g., M=10 agents, K=10 arms, α=0.5, 100 random seeds, fixed random seed for mean matrix generation). However, it does not provide specific hyperparameters like learning rates, batch sizes, or other training configuration details for the W-CPE-BAI or PF-UCB-BAI algorithms themselves.