Near-Optimal Collaborative Learning in Bandits

This repository is the official implementation of [Near-Optimal Collaborative Learning in Bandits] (to appear in the proceedings of NeurIPS'22).

Requirements

To install requirements (Python 3.8.5, using Conda to set up a virtual environment, and pip to install packages):

conda create -n near_optimal_federated python=3.8.5
conda activate near_optimal_federated
python3 -m pip install -r requirements.txt
conda deactivate

Getting started

Acquire a personal academic license for MOSEK (free)

Go to https://www.mosek.com/products/academic-licenses/ to acquire a personal academic license, and follow instructions.

Reproduce results from the paper

• Run command

bash commands.sh

Run

Parameters can be changed in params.py, and are overridden by the parameter values stored in a JSON file at the path provided by params_path.

Compare the collaborative setting versus independently solving the bandit instances per population

To compute the corresponding complexity constants from the lower bound, run

python3 test_useful.py

Add new elements of code

• Add a new bandit by creating a new instance of class FederatedAI in file algorithms.py
• Add a new dataset by adding a few lines of code to file data.py
• Add new types of rewards by creating a new instance of class problem in file problems.py

Results

Please refer to the paper.

Contributing

All of the code is under MIT license. Everyone is most welcome to submit pull requests.