Asymptotically Optimal and Computationally Efficient Average Treatment Effect Estimation in A/B testing

Authors: Vikas Deep, Achal Bassamboo, Sandeep Kumar Juneja

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical comparisons demonstrate that both policies perform similarly across practical values of ϵ and δ, offering efficient solutions for A/B testing.
Researcher Affiliation Academia 1Kellogg School of Management, Northwestern University, Evanston, IL 60201 2Ashoka University, Sonipat, Haryana, India.
Pseudocode No The paper describes the 'Assignment rule', 'Estimation rule', and 'Stopping rule' for Policy P1 and P2 in paragraph form, but it does not present these as a structured pseudocode block or a clearly labeled algorithm.
Open Source Code No The paper does not provide any statement about releasing source code for the described methodology, nor does it include a link to a code repository.
Open Datasets No Specifically, we model the outcomes for treatments A and B as exponentially distributed with means µA = 10 and µB = 0.1, respectively. We select ϵ = 0.5 and explore different values of δ, including 10%, 5%, and 1%. For each δ setting, we generate 2000 sample paths and calculate the average outcomes. The paper describes modeling and generating synthetic data for its experiments, rather than using a publicly available dataset with concrete access information.
Dataset Splits No The paper describes modeling and generating synthetic data for its experiments, and mentions using '2000 sample paths,' but it does not specify any training/validation/test dataset splits.
Hardware Specification No The paper does not provide any specific details about the hardware used to run the experiments, such as CPU or GPU models, memory, or cloud resources.
Software Dependencies No The paper mentions using specific methods and rules from other research (e.g., 'D-tracking rule introduced in Garivier & Kaufmann (2016)', 'Kaufmann & Koolen (2021) for the choice of β(n, δ)'), but it does not list any general software dependencies with specific version numbers (e.g., Python, PyTorch).
Experiment Setup Yes We select ϵ = 0.5 and explore different values of δ, including 10%, 5%, and 1%. For each δ setting, we generate 2000 sample paths and calculate the average outcomes.