Collaborative Algorithms for Online Personalized Mean Estimation

This project is the implementation of the paper "Collaborative Algorithms for Online Personalized Mean Estimation" Accepted in TMLR: https://arxiv.org/abs/2208.11530 and https://openreview.net/forum?id=VipljNfZSZ

Please email llvllahsa @ gmail for questions and suggestions revolving around this code repository.

How to Run?

The code is written in python and you can access it via the folder named "CollaborativeCode" and to run the implementation, there are two steps:

1. Running the main file to run different algorithms and save the results in npFiles.npz:

   python3 main.py

2. Loading npFiles.npz and plotting the results using:

   python3 LoadFiles.py

How to modify and run different experiments?

You can go to the main.py file and main() function. To change the means, you can change "mus" variable. For instance:

mus = np.array([0.2, 0.3, 0.5, 0.7])

To change the maximum number of time steps for running each algorithm, you can modify:

horizon = 1000

To change the number of agents:

n_agents = 100

The list of algorithms implemented are: "soft-restricted-round-robin", "aggressive-restricted-round-robin", "restricted-round-robin", "round-robin", "oracle", "local" , "random" and you can add each of them to the list containing the algorithms to be run:

algorithm_set = ["soft-restricted-round-robin", "aggressive-restricted-round-robin"]

To change the distribution functions from Gaussian to Bernoulli, do:

bernoullichoice = True

To set PAC framework's epsilon and delta:

epsilon = 0.01

delta = 0.001

ConfidenceBound.py

Different concentration inequality bounds are implemented. However, they should be checked based on the analysis. We have used Lemma1 bound for our approach.