Heavy-Tailed Linear Bandits: Huber Regression with One-Pass Update

Authors: Jing Wang, Yu-Jie Zhang, Peng Zhao, Zhi-Hua Zhou

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we evaluate the empirical performance and time efficiency of our proposed Hvt-UCB algorithm. We present two experiments under heavy-tailed noise (Student s t) and light-tailed noise (Gaussian).
Researcher Affiliation Academia 1National Key Laboratory for Novel Software Technology, Nanjing University, China 2School of Artificial Intelligence, Nanjing University, China 3Center for Advanced Intelligence Project, RIKEN, Japan.
Pseudocode Yes Algorithm 1 Hvt-UCB
Open Source Code No The paper does not provide any explicit statement or link indicating that the source code for the methodology described in this paper is publicly available.
Open Datasets No We consider the linear model rt = X t θ + ηt... We conduct two synthetic experiments with different distributions for noise ηt: (a) Student s t-distribution with degree of freedom df = 2.1 to represent heavy-tailed noise; and (b) Gaussian noise sampled from N(0, 1) to represent light-tailed noise.
Dataset Splits No The paper uses synthetic data and thus does not involve traditional training/test/validation dataset splits. It specifies the number of rounds T = 18000 and the number of independent trials (10 or 5) for evaluation.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No We implement all three distributions using Python s scipy.stats. (No version numbers are provided for Python or scipy.)
Experiment Setup Yes Settings. We consider the linear model rt = X t θ + ηt, where the dimension d = 2, the number of rounds T = 18000, and the number of arms n = 50. Each dimension of the feature vectors for the arms is uniformly sampled from [ 1, 1] and subsequently rescaled to satisfy L = 1. Similarly, θ is sampled in the same way and rescaled to satisfy S = 1. We conduct two synthetic experiments with different distributions for noise ηt: (a) Student s t-distribution with degree of freedom df = 2.1 to represent heavy-tailed noise; and (b) Gaussian noise sampled from N(0, 1) to represent light-tailed noise. For the heavy-tailed experiment, we set ε = 0.99 and ν = 1.31, while for the light-tailed experiment, we set ε = 1. (From Section 5) Additionally, from Theorem 1: By setting σt, τt, τ0, α as in Lemma 1, and let λ = d, σmin = 1 T , δ = 1 8T