Small Total-Cost Constraints in Contextual Bandits with Knapsacks, with Application to Fairness
Authors: Evgenii Chzhen, Christophe Giraud, Zhen LI, Gilles Stoltz
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Section 5: Brief overview of the numerical experiments performed. Appendix G: Numerical experiments: full description. These sections detail simulations and present performance figures (Figure 1) and tables (Table 1), indicating empirical evaluation. |
| Researcher Affiliation | Collaboration | Evgenii Chzhen Christophe Giraud Université Paris-Saclay, CNRS, Laboratoire de mathématiques d Orsay, 91405, Orsay, France; Zhen Li BNP Paribas Corporate and Institutional Banking, 20 boulevard des Italiens, 75009 Paris, France; Gilles Stoltz Université Paris-Saclay, CNRS, Laboratoire de mathématiques d Orsay, 91405, Orsay, France HEC Paris, 78351 Jouy-en-Josas, France. |
| Pseudocode | Yes | BOX B: PROJECTED GRADIENT DESCENT FOR CBWK WITH FIXED STEP SIZE; BOX C: PROJECTED GRADIENT DESCENT FOR CBWK WITH ADAPTIVE STEP SIZE; Algorithm 1: Pseudo-code for the Box C strategy (Appendix C). |
| Open Source Code | No | The paper mentions a GitHub repository (https://github.com/stanford-policylab/learning-to-be-fair) in Section 5 and Appendix G, but explicitly states it is from Chohlas-Wood et al. [2021] and describes it as an existing "experimental setting" used for inspiration, not their own source code for the methodology presented in this paper. |
| Open Datasets | Yes | We follow strictly the experimental setting of Chohlas-Wood et al. [2021] as provided in the public repository https://github.com/stanford-policylab/learning-to-be-fair. |
| Dataset Splits | No | The paper describes a sequential online learning setup for contextual bandits with simulated data over T rounds (T=10,000 individuals). It mentions an "initial 50 rounds as a warm start" but does not specify formal train/validation/test dataset splits, percentages, or predefined partitions typically used for reproducibility in static dataset experiments. |
| Hardware Specification | Yes | Appendix G.4: Our experiments were ran on the following hardware environment: no GPU was required, CPU is 3.3 GHz 8 Cores with total of 16 threads, and RAM is 32 GB 4800 MHz DDR5. |
| Software Dependencies | No | The paper describes the mathematical models (e.g., logistic regression for rewards) and estimation procedures, but it does not specify the software libraries, frameworks, or their versions (e.g., Python, PyTorch, TensorFlow, or specific solvers with version numbers) used for implementation. |
| Experiment Setup | Yes | Appendix G.3: We take T = 10,000 individuals... and set initial 50 rounds as a warm start for strategies. Appendix G.1: In our simulations, we tested a range of values and picked (in hindsight) the well-performing values Cδ = 0.025 and λlogistic = 0. Appendix G.2: We run all these strategies not with the total-cost constraints B = (0.05, 0.2, τ, τ, τ, τ) , where τ {10 7, 0.025} , but take a margin b = 0.005 and use B = B (b, b, 0, 0, 0, 0) instead. |