Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Feasible Action Search for Bandit Linear Programs via Thompson Sampling
Authors: Aditya Gangrade, Aldo Pacchiano, Clayton Scott, Venkatesh Saligrama
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our main simulations focus on constructing a practical methodology out of the theoretical study of FAST. These are presented in B.1 below. Additionally, we do auxiliary simulations that compare the behaviour of FAST and EOGT, which are presented in B.2. Figure 1. Estimates of the Optimism Rates under the Coupled Design Drawn According to Unif(γd Sd) as γ is varied. Top. Medians over 100 runs. Table 1. Comparison of the Quality of aout in the fixed d, varying M setting for Feasible Instances, with ε = 0.5. |
| Researcher Affiliation | Academia | 1Boston University 2Broad Institute of MIT and Harvard 3University of Michigan. Correspondence to: A. Gangrade <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Feasible Action Search via Thompson Sampling (FAST(µ, ε, δ)) |
| Open Source Code | No | The paper does not contain any explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | Throughout, we set Φ to be a certain 9 × 9 directed adjacency matrix, A, obtained from https://sparse.tamu.edu/van Heukelum/cage4 |
| Dataset Splits | No | The paper describes generating different instances by varying parameters (e.g., optimal safety margin M, constraint levels) for simulation purposes, but it does not specify conventional training, validation, and testing dataset splits as typically found in supervised machine learning experiments. |
| Hardware Specification | Yes | Indeed, with this change, these simulations took about 3 hours to execute on a midrange laptop computer running a 2022 Ryzen-5U CPU. |
| Software Dependencies | No | All instances are simulated 50 times using a MATLAB environment. The paper mentions a software environment (MATLAB) but does not provide a specific version number. |
| Experiment Setup | Yes | We execute FAST on the described instance with ε = 0.9, and ϕ chosen so that M = 1. As mentioned above, the noise laws investigated are the coupled design, with ζ ∼ Unif(γd Sd), for γ ∈ {3i : i ∈ [−4, 1]}. Note that since d = 9, the parameter γ = 3−2 boils down to choosing the uniform law on Sd. Throughout, we let the feedback noise have variance 1. We proceed by setting M ∈ {0.15, 0.2, . . . , 0.5}, and ε ∈ {0.5, 0.7, 0.9}, and simulating the behaviour of FAST on the described instances 100 times for each pair of parameters. We note that feedback noise is set to be independent Gaussian with standard deviation 0.1, and we use this value to adjust our confidence sets. |