Improved Algorithms for Nash Welfare in Linear Bandits

Overview

Official Python implementation for the work by Sarkar et al. on improved algorithms for Nash/p-mean regret minimisation in linear bandits. [arXiv]

Repository Structure

FairLinBandit/
├── algorithms/       # Contains code for all bandit algorithms used
├── cached/           # Stores regret data from first-time runs
├── plots/            # Final figures
├── main.py           # Master script: runs all experiments sequentially
├── experiments.py    # Functions for running and plotting individual experiments
├── data_loader.py    # Functions for loading datasets
└── requirements.txt  # Python dependencies

cached/

Stores serialized regret data generated during the first run of each experiment which can be directly loaded for subsequent runs.

plots/

Holds the final plot images (e.g., .png files) for each experiment. These plots compare algorithmic performance under different settings.

Installation

1. Clone the repository:

   git clone https://github.com/NP-Hardest/FairLinBandit.git
   cd FairLinBandit
   

2. Create a conda environment (optional but recommended):

   conda create -n FairLinBandit 
   conda activate FairLinBandit
   

   One can also use venv instead of conda.

3. Install dependencies:

   pip install -r requirements.txt
   

Usage

Run All Experiments

Execute the master script to run all experiments in sequence. Use np.random.seed(42) for reproducing results from the paper:

python main.py

Tweak Parameters

Adjust experiment parameters in main.py such as:

-   Horizon length (T)
    
-   Number of trials
    
-   Algorithm-specific settings

-   $p$ and $d$ values.

While running the experiments for the first time, the regret data is stored in cached/ directory. For subsequent runs, this data can be directly loaded using np.load().

Results & Visualization

• After running the experiments, plots will be saved in the plots/ folder.

• Use the functions in experiments.py to customize plot styles or export formats.

Contributing

Feel free to fork the repository to reproduce our results and other suggestions for modifications.