Pure exploration in Kernel and Neural Bandits

This repository is the official implementation of Pure exploration in Kernel and Neural Bandits

Requirements

To install requirements:

Packages used in this folder include: numpy, functools, scipy, sklearn, math, sys, logging, torch, itertools, pickle, gzip.

Training and Evaluation

This paper includes results on two sythentic datasets and two real datasets.

To run the model(s) in the paper, run this command:

bash run.sh

MNIST dataset

mnist.pkl contains the raw data of the MNIST dataset.

method_list = [neural_elim, kernel_elim, linear_elim]

For method in method_list:

 python run_minst.py method seed

For example, to run Alg.2 NeuralEmbedding, we use:

python run_mnist.py neural_elim 43

Linear dataset

- python run_linear_data.py method

Results

Our model achieves the following performance on Mnist Dataset and [Yahoo Dataset](https://webscope.sandbox.yahoo.com/?guccounter=1 style="zoom:50%;" />)

Sample complexity

• MNIST dataset:

• Yahoo dataset:

Success rate

	Neural Elimination	Kernel Embedding	Linear Embedding	RAGE	Action Elimination
MNIST Dataset	98%	100%	100%	100%	100%
Yahoo Dataset	100%	98%	88%	90%	100%

Acknowledgement

We would like to thank Sequential Experimental Design for Transductive Linear Bandits for open-source code.

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follow.

@article{zhu2021pure,
  title={Pure Exploration in Kernel and Neural Bandits},
  author={Zhu, Yinglun and Zhou, Dongruo and Jiang, Ruoxi and Gu, Quanquan and Willett, Rebecca and Nowak, Robert},
  journal={arXiv preprint arXiv:2106.12034},
  year={2021}
}