adr: Advantage Doubly Robust (ADR) Estimator for Learning When-to-Treat Policies

This repository implements ADR for learning when-to-treat policies, as proposed by Nie, Brunskill and Wager (2019).

Authors

This package is written and maintained by Xinkun Nie (xinkun@stanford.edu).

Reproducibility:

The code has the option to run on a multi-core cluster or on an individual machine. You can set the flag cluster=T if the former in main.R. There are also a few additional flags in main.R you can set for running with different options, with comments/documentations in the script.

Here we include instructions on how to reproduce results in the paper.

First, make new directories for experiments.

mkdir results
mkdir plots
mkdir logs

The following two commands start running two bashscripts that write to disk Monte-Carlo evaluated oracle values for all policies specified in a policy class, for the two simulation setups respectively. start_oracle_1.sh is a bash script runs the oracle simulations for the multiple-treatment setup specified in the paper, and start_oracle_2.sh is a bash script runs the oracle simulations for the binary-treatment setup specified in the paper

./start_oracle_1.sh
./start_oracle_2.sh

Running the above two commands will write to the directory results_oracle. Note that the results of the runs are included in the folder results_oracle so you can skip these two commands if you choose to.

Next run the following two commands to actually start the experiments, which read oracle values from the results_oracle directory. results_oracle

./start_1.sh
./start_2.sh

Finally, to produce tables and make plots, run the following files in the following order combine_results.R, parse_results.R, make_plot.R. You might need to make small modifications to adjust for which setup's results is being compiled and plotted, etc.

To produce plots that compare Q-Opt decision boundaries with ADR, turn log_trajs=T in main.R, run a few repeats, and then run plot_policy.R.

References

Xinkun Nie, Emma Brunskill and Stefan Wager. Learning When-to-Treat Policies. 2019. [arxiv]