Linear-Bandits-with-Memory

This repository contains the code for Linear Bandits with Memory (LBM), to reproduce the experiments presented in a paper currently under submission.

This repository contains the following files:

• exp.py contains the code to reproduce the performance of O3M, OFUL and Greedy on a rotting instance of LBM.
• model_selection_alpha.py contains the code to reproduce the performance of Bandit Combiner for the setting where the parameter $\gamma$ is misspecified.
• model_selection_m.py contains the code to reproduce the performance of Bandit Combiner for the setting where the parameter $m$ is misspecified.
• plot.py contains the code to plot the results obtained from exp.py, model_selection_alpha.py, model_selection_m.py,and showed in Fig. 1 (left pane).
• greedy_subopt_benchmark.py contains the code to reproduce the performance of O3M, Policy $\pi_2$, OFUL, and Greedy for the experiment in Fig. 1 (right pane).
• plot_greedy_subopt.py contains the code to reproduce the plot in Fig. 1 (right pane).
• regret_rising.py contains the code to reproduce the experiment where we compare the regret of O3M and OM-Block for several time horizons.
• regret_rising_plot.py contains the code to plot the results obtained from regret_rising.py for the experiment in Fig. 2 in the Supplementary Material.

In order to run the experiments presented in Fig. 1 (left pane), run (~2-3 hours):

$ python exp.py
$ python model_selection_alpha.py
$ python model_selection_m.py
$ python plot.py

In order to run the experiment related to Fig. 1 (right pane), run (~5 mins):

$ python greedy_subopt_benchmark.py
$ python plot_greedy_subopt.py

To run the additional experiment presented in the Supplementary Material, run the following command (~2 weeks):

$ python regret_rising.py
$ python regret_rising_plot.py