Robust Outlier Arm Identification

Authors: Yinglun Zhu, Sumeet Katariya, Robert Nowak

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our algorithms are both robust and about 5x sample efficient compared to state-of-the-art.
Researcher Affiliation Collaboration 1University of Wisconsin-Madison 2Amazon. Correspondence to: Yinglun Zhu <yinglun@cs.wisc.edu>.
Pseudocode Yes Algorithm 1 Construction of Confidence Intervals; Algorithm 2 ROAIElim; Algorithm 3 ROAILUCB
Open Source Code Yes Our code is publicly available (Zhu et al., 2020). URL https: //github.com/yinglunz/ROAI_ICML2020.
Open Datasets Yes We also compare the performance of all algorithms on the real-world Wine Quality dataset (Sathe & Aggarwal, 2016), which is widely used to compare outlier detection algorithms.
Dataset Splits No The paper describes generating synthetic data and simulating rewards from means of a real-world dataset but does not explicitly provide training, validation, and test splits for reproducibility in the traditional sense of fixed datasets. The multi-armed bandit setting involves adaptive sampling.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, processor types, or memory used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9).
Experiment Setup Yes We choose the arm configuration in Fig. 1(a) containing 15 normal arms (in blue) with fixed means equally distributed from 0 to 2, an outlier threshold θ ≈ 2.837, and 2 outlier arms (in orange) above the outlier threshold. The distance between the outlier arms is fixed at 0.2. We decrease θ from 0.6 to 0.2, and this changes the theoretical sample complexity. Note that the threshold does not change. The reward of each arm is normally distributed with standard deviation 0.5. (...) We generate 100 normal arm means from N(0.3, 0.075^2) and 5 outlier means from Unif(0.8, 1). We draw rewards of each arm from a Bernoulli distribution with respect to its mean.