Multi-task Representation Learning for Pure Exploration in Bilinear Bandits

Authors: Subhojyoti Mukherjee, Qiaomin Xie, Josiah Hanna, Robert Nowak

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conduct proof-of-concept experiments on both single and multi-task bilinear bandits. In the single-task experiment, we compare against the state-of-the-art RAGE algorithm (Fiez et al., 2019). We show in Figure 1 (left) that GOBLIN requires fewer samples than the RAGE with an increasing number of arms. In the multi-task experiment, we compare against the state-of-the-art Dou Exp Des algorithm (Du et al., 2023). We show in Figure 1 (right) that GOBLIN requires fewer samples than Dou Exp Des with an increasing number of tasks. As experiments are not a central contribution, we defer a fuller description of the experimental set-up to Appendix A.8.
Researcher Affiliation Academia Subhojyoti Mukherjee ECE Department UW-Madison Wisconsin, Madison smukherjee27@wisc.edu Qiaomin Xie ISy E Department UW-Madison Wisconsin, Madison Josiah P. Hanna CS Department UW-Madison Wisconsin, Madison Robert Nowak ECE Department UW-Madison Wisconsin, Madison
Pseudocode Yes Algorithm 1 G-Optimal Design for Bilinear Bandits (GOBLIN) for single-task setting... Algorithm 2 G-Optimal Design for Bilinear Bandits (GOBLIN) for multi-task setting
Open Source Code No No statement or link providing access to the source code for the described methodology was found.
Open Datasets No The paper does not provide information about specific datasets used in the experiments nor their public availability. It only refers to a 'number of arms' and 'number of tasks' in its experimental section.
Dataset Splits No The paper mentions experiments but does not provide details on dataset splits (training, validation, test) or cross-validation setup.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory) used for running experiments were mentioned in the paper.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) were mentioned in the paper.
Experiment Setup No As experiments are not a central contribution, we defer a fuller description of the experimental set-up to Appendix A.8. The main body of the paper does not contain specific hyperparameters or training details.