Thompson Sampling via Local Uncertainty

This repo corresponds to the Thompson Sampling via Local Uncertainty paper, published in ICML 2020. We propose a new probabilistic modeling framework for Thompson sampling, where local latent variable uncertainty is used to sample the mean reward. We implemented our algorithms LU-Gauss/SIVI based on the Benchmark proposed by Carlos Riquelme.

Please, use the following when citing the code or the paper:

@article{zhendong2020luts, 
title={Thompson Sampling via Local Uncertainty},
author={Wang, Zhendong and Zhou, Mingyuan},
journal={International Conference on Machine Learning, ICML.}, 
year={2020}
}

Dependencies

This code is based on Python 2.7, with the main dependencies being TensorFlow==1.14 and other dependencies stated in the Benchmark

Dataset

You can download the required dataset from the benckmark.

Training

You can run comparison of all algorithms from run_bandit.py file, with one argument for a specific dataset. It will save the output results in one .npz file.

    python run_bandit.py -dataset='mushroom'

Acknowledgement

We thank greatly Riquelme et al. for making their code public.