Introduction
This repository contains code for "Pure Exploration in Asynchronous Federated Bandits" published at UAI2024. This project investigates the federated pure exploration problem within multi-armed and linear bandits settings, where multiple agents collaborate to identify the best arm under communication with a central server. Addressing common issues like latency and unavailability of clients in practical scenarios, we introduce the first federated asynchronous algorithms for multi-armed bandit (MAB) and linear bandit models aimed at pure exploration with fixed confidence. Our approach balances the exploration-exploitation trade-off efficiently, even in fully asynchronous environments, enhancing robustness against delays and agent unavailability while minimizing communication costs.
Main code
This repository contains implementation of the proposed algorithms UGapE, LinGapE, and FALinPE for comparison. For experiments on the synthetic dataset, run: SimLinear.py, SimTabular.py
Experiment results can be found in "./SimTabular/" and "./SimLinear/" folder, which contains:
- "SamConAndCommCost_dataset[number]_[startTime].png/pdf": plot of sample complexity / communication cost over time for each algorithm
- "SampleComplex_dataset[number]_[startTime].csv": sample complexty at each iteration for each algorithm
- "AccCommCost_dataset[number]_[startTime].csv": communication cost at each iteration for each algorithm
For experiments on realworld dataset (MovieLens):
- Available dataset and file scripts for data processing and feature vector generationare given in "./Dataset" folder.
- Experiment results can be found in "./MovieLinear/" folder run: plot_v.py
Reference
@inproceedings{
wang2024pure,
title={Pure Exploration in Asynchronous Federated Bandits},
author={Zichen Wang and Chuanhao Li and chenyu song and Lianghui Wang and Quanquan Gu and Huazheng Wang},
booktitle={The 40th Conference on Uncertainty in Artificial Intelligence},
year={2024},
url={https://openreview.net/forum?id=VqgArMVbZf}
}