On Worst-case Regret of Linear Thompson Sampling

This repository contains the code that reproduces the simulations in our paper.

Installation

This project requires python 3.7 or higher. The list of the package dependencies can be found in requirements.txt. The following command installs the required packages.

conda install --file requirements.txt

Reproducing the results

By running the following commands, the plots in the paper can be reproduced.

PYTHONPATH=src python -m example_1
PYTHONPATH=src python -m example_2 --change mu --mu 1.0 --dim 2000
PYTHONPATH=src python -m example_2 --change rho --rho 1.0 --dim 2000
PYTHONPATH=src python -m example_2 --change dim --mu 0.1 --n-value 18
PYTHONPATH=src python -m experiments

Each script also accepts a set of parameters used in the experiment which can be found using --help argument. For example, running

PYTHONPATH=src python -m example_1 --help

yields

usage: example_1.py [-h] [--n-iter N_ITER] [--sigma SIGMA] [--tau TAU]
                    [--seed SEED]

Simulate the first example for TS failure.

optional arguments:
  -h, --help       show this help message and exit
  --n-iter N_ITER  number of iterations
  --sigma SIGMA    prior sd
  --tau TAU        noise sd
  --seed SEED      initial random seed