Local Anti-Concentration Class: Logarithmic Regret for Greedy Linear Contextual Bandit

Authors: Seok-Jin Kim, Min-hwan Oh

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conducted numerical experiments to evaluate the performance of the greedy algorithm and compare it with existing bandit algorithms, Lin UCB from Abbasi-Yadkori et al. [1] and Lin TS from Agrawal and Goyal [4].
Researcher Affiliation Academia Seok-Jin Kim Columbia University New York, NY, USA seok-jin.kim@columbia.edu Min-hwan Oh Seoul National Univeristy Seoul, South Korea minoh@snu.ac.kr
Pseudocode Yes Algorithm 1 Lin Greedy: Greedy Linear Contextual Bandit
Open Source Code Yes We provide code in supplementary material.
Open Datasets No We conducted numerical experiments to evaluate the performance of the greedy algorithm and compare it with existing bandit algorithms, Lin UCB from Abbasi-Yadkori et al. [1] and Lin TS from Agrawal and Goyal [4]. We conducted experiments for three cases with varying parameters: d = 20, K = 20, T 1000, d = 100, K = 20, T 1000, and d = 20, K = 100, T 1000, and five different distributions of contexts: Uniform in a ball, truncated Student s t, Laplace, Gaussian, and exponential.
Dataset Splits No The paper describes generating contexts from various distributions and running the bandit algorithms for a fixed number of rounds (T), but it does not specify explicit train/validation/test dataset splits.
Hardware Specification No The paper describes the experimental settings (d, K, T, and context distributions) but does not specify any particular hardware used for running the experiments (e.g., GPU/CPU models, memory).
Software Dependencies No We conducted numerical experiments to evaluate the performance of the greedy algorithm and compare it with existing bandit algorithms, Lin UCB from Abbasi-Yadkori et al. [1] and Lin TS from Agrawal and Goyal [4]. No specific software versions (e.g., Python version, library versions) are mentioned.
Experiment Setup Yes We conducted experiments for three cases with varying parameters: d = 20, K = 20, T 1000, d = 100, K = 20, T 1000, and d = 20, K = 100, T 1000, and five different distributions of contexts: Uniform in a ball, truncated Student s t, Laplace, Gaussian, and exponential.