Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback

Authors: Chicheng Zhang, Alekh Agarwal, Hal Daumé Iii, John Langford, Sahand Negahban

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we evaluate some of these algorithms on a large selection of datasets, showing that our approach is both feasible, and helpful in practice.
Researcher Affiliation Collaboration 1Microsoft Research 2University of Maryland 3Yale University.
Pseudocode Yes Algorithm 1: Adaptive Reweighting for Robustly Warmstarting Contextual Bandits (ARROW-CB)
Open Source Code No No explicit statement or link providing concrete access to the source code for the methodology described in this paper was found.
Open Datasets Yes We compare these approaches on 524 binary and multiclass classification datasets from Bietti et al. (2018), which in turn are from openml.org.
Dataset Splits Yes Partition S to E+1 equally sized sets Str, Sval 1 , . . . , Sval E . ... where a separate validation set is used to pick λ.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, processor types, memory amounts, or detailed computer specifications) used for running its experiments were provided.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment were provided.
Experiment Setup Yes All the algorithms (other than SUP-ONLY and MAJORITY, which do not explore) use ϵ-greedy exploration, with most of the results presented using ϵ = 0.0125. We additionally present the results for ϵ = 0.1 and ϵ = 0.0625 in Appendix J. ... We vary the number of warm-start examples ns in {0.005n, 0.01n, 0.02n, 0.04n}, and the number of CB examples nb in {0.92n, 0.46n, 0.23n, 0.115n}.