Anytime Model Selection in Linear Bandits
Authors: Parnian Kassraie, Nicolas Emmenegger, Andreas Krause, Aldo Pacchiano
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically we find that ALEXP consistently outperforms prior work across a range of environments. (Abstract) ... 6 Experiments |
| Researcher Affiliation | Academia | 1ETH Zurich 2Broad Institute of MIT and Harvard 3Boston University |
| Pseudocode | Yes | Algorithm 1 ALEXP (Page 5) |
| Open Source Code | Yes | The PYTHON code for reproducing the experiments is accessible on github.com/lasgroup/ALEXP. (Page 8) |
| Open Datasets | No | We create a synthetic dataset based on our data model (Section 3), and choose the domain to be 1-dimensional X = [ −1, 1]. (Experiment Setup) The paper describes how it generates synthetic data for experiments but does not provide a link to a publicly available dataset or a generated dataset for download. |
| Dataset Splits | No | For all experiments we set n = 100, and plot the cumulative regret R(n) averaged over 20 different random seeds, the shaded areas in all figures show the standard error across these runs. (Experiment Setup) The paper uses a synthetic data generation process and performs simulations over a fixed horizon (n=100). It does not mention explicit training, validation, or test dataset splits in the conventional sense of partitioning a static dataset. |
| Hardware Specification | No | The paper mentions software used ('PYTORCH', 'CELER') but does not specify any hardware components like GPU models, CPU types, or memory specifications for running the experiments. |
| Software Dependencies | No | We use PYTORCH [Paszke et al., 2017] for updating the upper confidence bounds... (Hyper-Parameter Tuning Results) ... To solve (1), we use CELER, a fast solver for the group Lasso [Massias et al., 2018]. (Experiment Setup) The paper mentions software tools but does not provide specific version numbers for reproducibility (e.g., PyTorch version, CELER version). |
| Experiment Setup | Yes | Experiment Setup. We create a synthetic dataset based on our data model (Section 3), and choose the domain to be 1-dimensional X = [ −1, 1]. For all experiments we set n = 100, and plot the cumulative regret R(n) averaged over 20 different random seeds... (Experiment Setup) ... To initialize ALEXP we set the rates of λt, γt and ηt according to Theorem 1, and perform a light hyper-parameter tuning to choose the scaling constants. (Experiment Setup) ... We set the rates for γt and ηt as prescribed by Theorem 1. For the scaling constants, we perform a hyper-parameter tuning experiment log-uniformly sampling 20 different configurations from γ0 [10−4, 10−1] and η0 [100, 102]. (Hyper-Parameter Tuning Results) ... For all the experiments, we set the exploration coefficient of UCB to βt = 2. (Hyper-Parameter Tuning Results) |