On Universally Optimal Algorithms for A/B Testing

Authors: Po-An Wang, Kaito Ariu, Alexandre Proutiere

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

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the performance of the ETT algorithm with α = 1/4 and different thresholds ε, and compare it to that of the uniform sampling algorithm and to that of an Oracle algorithm that selects arms using optimal exploration rate x (µ) = argmaxx g(x, µ). ... The error probabilities are derived from 40000 trials for each setting and algorithm.
Researcher Affiliation Collaboration 1EECS and Digital Futures, KTH, Stockholm, Sweden 2Cyber Agent, Tokyo, Japan.
Pseudocode Yes Algorithm 1 Successive Rejects (SR) ... Algorithm 2 Estimate and Thresholded Tracking (ETT) ... Algorithm 3 Randomized TCSF ... Algorithm 4 De-randomized TCSF
Open Source Code No The paper does not provide an explicit statement or link indicating that the source code for the described methodology is publicly available.
Open Datasets No We consider the instance: µ = (0.0005, 0.0001). ... The error probabilities are derived from 40000 trials for each setting and algorithm. The paper discusses theoretical properties of bandit instances, not public datasets.
Dataset Splits No The paper describes simulation trials for specific instances of bandit problems, not experiments on datasets with explicit train/validation/test splits.
Hardware Specification No The paper mentions numerical experiments but does not provide specific details on the hardware used, such as GPU/CPU models or memory.
Software Dependencies No The paper does not provide specific software dependencies with version numbers for reproducibility of the experiments.
Experiment Setup Yes We illustrate the performance of the ETT algorithm with α = 1/4 and different thresholds ε, and compare it to that of the uniform sampling algorithm and to that of an Oracle algorithm that selects arms using optimal exploration rate x (µ) = argmaxx g(x, µ). ... Figure 6 displays the error probability with a fixed budget of T = 20000 and varying ε from 0 to 0.0008.