Approximation Theory Based Methods for RKHS Bandits

Authors: Sho Takemori, Masahiro Sato

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In synthetic environments, we empirically show that APG-UCB has almost the same cumulative regret as that of IGP-UCB and its running time is much shorter.
Researcher Affiliation Industry 1 FUJIFILM Business Innovation, Kanagawa, Japan.
Pseudocode Yes Algorithm 1 Construction of Newton basis with P-greedy algorithm (c.f. Pazouki & Schaback (2011))
Open Source Code No The paper does not explicitly state the release of source code or provide a link to a repository for the described methodology.
Open Datasets No The paper describes generating synthetic environments and reward functions for experiments, but does not refer to a publicly available dataset with concrete access information (link, DOI, or specific citation to an established dataset).
Dataset Splits No The paper describes synthetic environments and online learning simulations, but does not specify training/validation/test dataset splits for reproducibility.
Hardware Specification Yes Computation is done by Intel Xeon E5-2630 v4 processor with 128 GB RAM.
Software Dependencies No No specific software dependencies with version numbers are mentioned in the paper.
Experiment Setup Yes We take l = 0.3 for the RQ kernel and l = 0.2 for the SE kernel, because the diameter of the d-dimensional cube is d. For each kernel, we generate 10 reward functions as above and evaluate our proposed method and the existing algorithm for time interval T = 5000 in terms of mean cumulative regret and total running time.