Almost Minimax Optimal Best Arm Identification in Piecewise Stationary Linear Bandits

Authors: Yunlong Hou, Vincent Tan, Zixin Zhong

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compare PSεBAI+ to baseline algorithms using numerical experiments which demonstrate its efficiency.
Researcher Affiliation Academia Yunlong Hou Department of Mathematics National University of Singapore yhou@u.nus.edu Vincent Y. F. Tan Department of Mathematics Department of Electrical and Computer Engineering National University of Singapore vtan@nus.edu.sg Zixin Zhong Data Science and Analytics Thrust Hong Kong University of Science and Technology (Guangzhou) zixinzhong@hkust-gz.edu.cn
Pseudocode Yes Algorithm 1 PIECEWISE-STATIONARY ε-BEST ARM IDENTIFICATION (PSεBAI)
Open Source Code Yes All the code to reproduce our experiments can be found at https://github.com/Y-Hou/BAI-in-PSLB.git.
Open Datasets No We utilize the instance defined in Example 1 with d = 2, ϕ = π/8, We generate a changepoint sequence C such that cl+1 = cl + Ll with Lmin = 3 10^4, Lmax = 5 10^4, P[Ll = Lmin] = 0.8, P[Ll = Lmax] = 0.2, and fix it throughout the whole set of experiments.
Dataset Splits No The paper uses a synthetic instance and describes how it is generated, but does not specify explicit training/test/validation dataset splits (e.g., percentages or sample counts).
Hardware Specification Yes All experiments are conducted via MATLAB R2021b on a Mac Book Pro with Apple M1 Pro chip and 16 GB memory.
Software Dependencies Yes All experiments are conducted via MATLAB R2021b
Experiment Setup Yes We set the confidence parameter δ = 0.05 and vary the slackness parameter ε from 0.04 to 0.6 (i.e., ε = 0.03 1.35k for k [12]). We set γ = 6, the window size w = Lmin/(3γ) and compute b via (3.5) in Assumption 1.