Information Directed Sampling for Sparse Linear Bandits

Authors: Botao Hao, Tor Lattimore, Wei Deng

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results demonstrate significant regret reductions by sparse IDS relative to several baselines. Through several experiments, we justify the great empirical performance of sparse IDS with an efficient implementation. We plot the empirical cumulative Bayesian regret.
Researcher Affiliation Collaboration Botao Hao Deepmind haobotao000@gmail.com Tor Lattimore Deepmind lattimore@google.com Wei Deng Department of Mathematics Purdue University weideng056@gmail.com
Pseudocode Yes Algorithm 1 Empirical Bayesian sparse sampling and Algorithm 2 Sparse IDS are present.
Open Source Code No In the 'Ethics Statement' section, under 'If you ran experiments', it is stated: '(a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No]'
Open Datasets No The paper describes generating synthetic data for its experiments: 'We set d = 10, s = 3, n = 100 and the actions are drawn i.i.d from a multivariate normal distribution N(0, Σ) with Σij = 0.6|i j|.' and 'We consider a more general case where each action is generated from multivariate normal distribution N(0, Σ) with Σij = 0.6|i j|.' It does not refer to a publicly available dataset with concrete access information.
Dataset Splits No The paper does not explicitly provide training/validation/test dataset splits. It describes generating data for experiments but does not specify how this data is partitioned for different phases of model evaluation.
Hardware Specification No In the 'Ethics Statement' section, under 'If you ran experiments', it is stated: '(d) Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [N/A]'. Therefore, specific hardware details are not provided.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers) needed to replicate the experiment.
Experiment Setup Yes Setting All the true parameters are randomly generated from a multivariate normal distribution, truncated to be sparse and normalized to have square norm 1. The noise variance is fixed to be 2 and we replicate the experiments over 200 trials. Each Bayesian algorithm will take 10000 posterior samples. We use the TS without blow-up factor for the variance and tune the length of confidence interval of Lin UCB over a candidate set.