gisoo1989 / drts

Doubly Robust Thompson Sampling with Linear Payoffs

Python3 based implementation of the paper "Doubly Robust Thompson Sampling with linear payoffs". In this repository, you can generate Figures of the paper!

Directory tree

.
|-- experiment.py
|-- plot.py
|-- algorithms.py
|-- figure1.sh
|-- requirements.txt
|-- README.md

• algorithms.py contains the Thompson Sampling (TS), Balanced Linear Thompson Sampling (BLTS), and the proposed Doubly Robust Thompson Sampling (DRTS).
• experiment.py contains the simulation environments and evaluation of cumulative regrets and estimation error.
• plot.py plots the results generated by experiment.py
• figure1.sh contains a quick start that reproduces the result in the paper.
• requirements.txt contains the dependencies to run the codes.

Requirements

• python 3
• numpy
• scipy
• sobol_seq
• matplotlib
• tqdm

Quick start

First to install the dependencies,

pip install -r requirements.txt

To generate Figure 1 in the paper simply run,

sh figure1.sh

This code will generate the cumulative regrets and estimation error plots of TS, BLTS, and DRTS, when d=10, 30, and N=3, 10.

If you want to change the settings, use

python experiment.py -d 5 -N 7 -seed 1324 -T 10000

to evaluate performances of the three algorithms for T=10000 rounds when d=5, and N=7, with seed 1324.

After running experiment.py, the estimation error and cumulative regrets are saved in .txt format. Then run

python plot.py -d 5 -N 7

to plot the results of d=5, and N=7.

Example results

We introduce our example results in our paper.

Figure 1. Comparison of cumulative regrets with best hyperparameters.

Figure 2. Comparison of estimation error with best hyperparameters