Optimal and Adaptive Off-policy Evaluation in Contextual Bandits
Authors: Yu-Xiang Wang, Alekh Agarwal, Miroslav Dudı́k
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically evaluate the SWITCH estimators against a number of strong baselines from prior work, using a previously used experimental setup to simulate contextual bandit problems on real-world multiclass classification data. |
| Researcher Affiliation | Collaboration | 1Carnegie Mellon University, Pittsburgh, PA 2Microsoft Research, New York, NY. |
| Pseudocode | No | The paper does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any statement about providing open-source code for the described methodology. |
| Open Datasets | Yes | We next empirically evaluate the proposed SWITCH estimators on the 10 UCI data sets previously used for off-policy evaluation (Dudík et al., 2011). |
| Dataset Splits | Yes | Following Dudík et al. (2011), DR is constructed by randomly splitting the contextual bandit data into two folds, estimating ˆr on one fold, and then evaluating π on the other fold and vice versa, obtaining two estimates. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU, GPU models, or memory) used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., library or solver names and versions). |
| Experiment Setup | Yes | The target policy π is the deterministic decision of a logistic regression classifier learned on the multi-class data, while the logging policy µ samples according to the probability estimates of a logistic model learned on a covariate-shifted version of the data. ... Then, in the simulator, we randomly draw i.i.d. data sets of size 100, 200, 500, 1000, 2000, 5000, 10000, . . . until reaching n, with 500 different repetitions of each size. ... In all these approaches we optimize among 21 possible thresholds, from an exponential grid between the smallest and the largest importance weight observed in the data, considering all actions in each observed context. |