This codebase contains implementation of experiments for the "Causal Bandits without Graph Learning" paper.
The raps package was developed using Python 3.9 and compatibility with other
Python versions is not guaranteed. First, clone the repository with the
submodules using git clone --recurse-submodules command. After that, to
install simply run
pip install -e raps
Installing the package automatically adds the raps script to the $PATH
variable. This script could be run to obtain the results of experiments
measuring the regret, for example:
raps --logdir logdir/tree-p3d2n20.01 --num-parents 3 --domain-size 2 \
--num-nodes 20 --nruns 10
This adds 10 tasks to the task spooler. Task spooler can be installed using
brew install task-spooler or sudo apt install task-spooler. If you're
using a Linux system, then be sure to pass --laucher tsp argument to the
script as on Linux the command to launch task spooler is different. You can
control the number of tasks run at the same, for example, to launch all 10 runs
at the same time use ts -S 10 or tsp -S 10 if on Linux. The progress bars
of the runs could be watched by running ./watch-runs first last where first
and last are the first and last indices of the tasks in task spooler.
Alternatively, you can use SLURM to manage the runs, for this pass --launcher slurm argument to the script.
Running the script generates pickle files with matplotlib figure and experiment objects. Later figure objects are used in the corresponding jupyter notebook in the directory notebooks to obtain the final figure aggregating the result from multiple runs of the same experiment.
For the experiments that test our theoretical findings regarding
the number of interventions performed by RAPS see
notebooks/num-interventions.ipynb jupyter notebook.