Nearly Optimal Catoni’s M-estimator for Infinite Variance

Authors: Sujay Bhatt, Guanhua Fang, Ping Li, Gennady Samorodnitsky

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate the performance improvement in using our proposed algorithm based on Catoni’s estimator for ε ∈ (0, 1] by comparing it with the state of the art best arm identification algorithms proposed in Yu et al. (2018). (...) A significant improvement in performance using our algorithms is observed in Figure 1, more so in case of smaller ε, indicating the tightness for smaller ε. Table 1 shows the performance of Algorithm 3 for K from 10 to 1000.
Researcher Affiliation Collaboration Sujay Bhatt, Guanhua Fang, Ping Li Gennady Samorodnitsky Cognitive Computing Lab School of ORIE Baidu Research Cornell University 10900 NE 8th St. Bellevue, WA 98004, USA 220 Frank T Rhodes Hall, Ithaca, NY 14853, USA {sujaybhatt.hr, fanggh2018, pingli98}@gmail.com gs18@cornell.edu
Pseudocode Yes Algorithm 1 Lepskii Moment Adaptation (LMA) (...) Algorithm 2 Iterative Elimination with Catoni (...) Algorithm 3 Phase-based Iterative Elimination with Catoni (...) Algorithm 4 Adaptive Elimination with Catoni (...) Algorithm 5 Restricted Differential Privacy for Heavy-tailed Data
Open Source Code No The paper does not contain an explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets No The experimental setup is as follows: the rewards for all arms in all experiments are generated from the heavy tailed Student’s-t distribution. The shifted mean of each arm is given using the following rule: µ(i) = 2 − (i − 1/K)0.6 for i = 2, ..., K and µ(1) = 2. The number of degrees of freedom for the first two figures from the left is ν = 3 which corresponds to ε = 1 and υε = 3. The third figure uses ν = 2 which has infinite variance and for ε = 0.85, an upper bound is υε = 50.
Dataset Splits No The paper describes experiments in a multi-armed bandit setting with sequentially generated data, rather than using predefined training/validation/test splits from a fixed dataset.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., CPU, GPU models, memory) used to run the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9) required to replicate the experiments.
Experiment Setup Yes The experimental setup is as follows: the rewards for all arms in all experiments are generated from the heavy tailed Student’s-t distribution. The shifted mean of each arm is given using the following rule: µ(i) = 2 − (i − 1/K)0.6 for i = 2, ..., K and µ(1) = 2. The number of degrees of freedom for the first two figures from the left is ν = 3 which corresponds to ε = 1 and υε = 3. The third figure uses ν = 2 which has infinite variance and for ε = 0.85, an upper bound is υε = 50. Smaller values for τ and larger values for h will have sharper bounds, and can be chosen depending on the tolerance to the cost of initial exploration. One choice parameters used is τ = 0.05 and h = 0.7. Further, γ(> 1) in Algorithm 3 is set to 1.1; this can be tuned further to improve performance.