Code for "Non-Stationary Latent Autoregressive Bandits" paper.
Set experiment parameters such as max_seed or num_time_steps in the global_params.py file.
ucb_agents.py contains the code for all agents. It contains the parent class which implements standard UCB / OFUL for stochastic linear bandits as well as child classes such as standard UCB for multi-armed bandits and our algorithm presented in the paper.
simulations.py contains the skeleton code for running an experiment with an agent and an environment variant. There is also code for calculating the ground-truth optimal actions and mean rewards for each action.
To recreate the results in the main experiments section of our paper, run python3 run_simulations.py. We vary the true AR order k and the noise of the latent process. This script will save all results in the experiment_results folder.
To recreate the ablation results in section considering different mis-specified values of AR order k, run run_sims_wrong_k.py. This script will save all results in the experiment_results folder.
To generate the figures in our paper, run python3 stats_and_figs/make_plots.py and python3 stats_and_figs/make_plots_wrong_k.py. These scripts generate figures using results saved in experiment_results with correct k and mis-specified k, respectively.