Multi-armed-Bandits-with-Local-Differential-Privacy

This repo is for our work "Ren, W., Zhou, X., Liu, J., & Shroff, N. B. (2020). Multi-armed bandits with local differential privacy. arXiv preprint arXiv:2007.03121."

Any entity can use this repo as long as the above work is properly cited.

The structure of .py files is as follows.

DPclass.py: define all the necessary classes and functions

-- ucb_bandit: basic class for MAB Bern reward under UCB with LDP	

-- ucb_bandit_gaussian: child class of ucb_bandit, defined for gaussian reward

-- ucb_bandit_heter: child class of ucb_bandit, defined for mixed reward

-- experiment_with_algorithms: fix a LDP parameter (i.e., epsilon), test for different algorithms, defined by type_list (e.g., Non-private, LDP-UCB-L, LDP-UCB-B)

-- experiment_with_epsilon: fix a algorithm (e.g., LDP-UCB-B), test for different LDP parameters, defined by epsilon_list (i.e., )

-- plot_algorithms: plot the results obtained by experiment_with_algorithms

-- plot_epsilon: plot the results obtained by experiment_with_epsilon

main.py: run the experiment and plot the figures.

-- reward_type: Bern, Mixed or Gaussian as considered in the paper

-- iters: number of time-slots, i.e., T

-- episodes: number of runs

-- mu_array: define the problem instance