Stochastic Multi-Armed Bandits with Control Variates

Authors: Arun Verma, Manjesh Kumar Hanawal

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic problem instances validate performance guarantees of the proposed algorithms.
Researcher Affiliation Academia Arun Verma Department of Computer Science National University of Singapore arun@comp.nus.edu.sg Manjesh K. Hanawal Department of IEOR IIT Bombay, Mumbai, India mhanawal@iitb.ac.in
Pseudocode Yes UCB-CV UCB based Algorithm for MAB-CV problem 1: Input: K, Q, α > 1 2: Play each arm i [K] Q times 3: for t = QK + 1, QK + 2, . . . , do 4: i [K] : compute UCBt 1,i as given in Eq. (5) 5: Select It = argmax i [K] UCBt 1,i 6: Play arm It and observe Xt,It and associated control variates Wt,It. Increment the value of NIt(t) by one and re-estimate ˆβ NIt(t),It, ˆµc NIt(t),It and ˆνt,NIt(t) 7: end for
Open Source Code No No explicit statement about providing open-source code or a link to a code repository for the methodology was found.
Open Datasets No We empirically evaluate the performance of UCB-CV... on different synthetically generated problem instances. For all the instance we use we use K = 10, q = 1, and α = 2. All the experiments are repeated 100 times and cumulative regret with a 95% confidence interval (the vertical line on each curve shows the confidence interval) are shown. Details of each instance are as follows: Instance 1: The reward and associated CV of this instance have a multivariate normal distribution. The reward of each arm has two components. We treated one of the components as CV. In round t, the reward of arm i is given as follows: Xt,i = Vt,i + Wt,i, where Vt,i N(µv,i, σ2 v,i) and Wt,i N(µw,i, σ2 w,i).
Dataset Splits No The paper uses synthetically generated data and runs simulations over time. It does not describe traditional train/validation/test dataset splits as would be found in a supervised learning context.
Hardware Specification No No specific hardware details (e.g., GPU models, CPU types, or cloud resources) used for running experiments were provided in the paper.
Software Dependencies No The paper does not list specific software dependencies with version numbers for replicating the experiment. While plots mention Matlab, this is not a dependency of the experiment itself.
Experiment Setup Yes For all the instance we use we use K = 10, q = 1, and α = 2. All the experiments are repeated 100 times