This repository contains our pytorch implementation of Langevin Monte Carlo Thompson Sampling (LMC-TS), proposed in the paper Langevin Monte Carlo for Contextual Bandits

Abstract: Existing Thompson sampling-based algorithms need to construct a Laplace approximation of the posterior distribution, which is inefficient to sample in high dimensional applications for general covariance matrices. Moreover, the Gaussian approximation may not be a good surrogate for the posterior distribution for general reward generating functions. We propose an efficient posterior sampling algorithm, viz., Langevin Monte Carlo Thompson Sampling (LMC-TS), that uses Markov Chain Monte Carlo (MCMC) methods to directly sample from the posterior distribution in contextual bandits. Our method is computationally efficient since it only needs to perform noisy gradient descent updates without constructing the Laplace approximation of the posterior distribution.

To install the necessary packages, run

pip install -r requirements.txt

1. Find the data here to reproduce our results. https://hkzdata.s3.us-west-2.amazonaws.com/LMC-TS/gaussian50-20-1-1.pt
2. To generate synthetic data, run

python3 data_generator.py --config configs/data-linear.yaml

Run bandit algorithms on simulated bandit problems

python3 run_simulation.py --config_path configs/simulation/linear-LMCTS.yaml --repeat [number of experiments to repeat] 

You can add --log to turn on the wandb. Configuration file examples:

• Linear bandit: configs/simulation/linear-LMCTS.yaml
• Quadratic bandit: configs/simulation/quad-LMCTS.yaml
• Logistic bandit: configs/simulation/logistic-LMCTS.yaml

To run bandit algorithm on UCI datasets, use

python3 run_classifier.py --config_path configs/uci/shuttle-lmcts.yaml --repeat [number of experiments to repeat]

You can add --log to turn on the wandb. Configuration files are provided under folder configs/uci.

To run bandit algorithm on CIFAR10 datasets, use

python3 run_cifar.py --config_path configs/image/cifar10-lmcts.yaml

Configuration files are provided under configs/uci.

run_cifar.py, run_classifier.py, run_simulation.py are the main entries of the program, which contain the abstract code of bandit framework. These main entry files read configuration .yaml files from configs and parse them to run.

algo,train_utils,models realize different modules in the framework. More specifically,

1. algo implements different online bandit algorithms.
2. train_utils implements datasets, bandit instance, loss functions, and some helper functions.
3. models implements different model architectures.

@inproceedings{xu2022langevin,
  title={Langevin Monte Carlo for Contextual Bandits},
  author={Xu, Pan and Zheng, Hongkai and Mazumdar, Eric V and Azizzadenesheli, Kamyar and Anandkumar, Animashree},
  booktitle={International Conference on Machine Learning},
  pages={24830--24850},
  year={2022},
  organization={PMLR}
}