Counterfactual Off-Policy Evaluation with Gumbel-Max SCMs

Overview

This repository contains all the code required to replicate the figures in the ICML 2019 paper. To do so from scratch, you will need to run the following steps:

• To generate the MDP parameters, run learn_mdp_parameters.ipynb, which will save the learned parameters in the data folder; This takes ~2 hours.
• Alternatively, you can use the parameters that are already learned - To do so, unzip data/diab_txr_mats-replication.zip locally (e.g., using unzip diab_txr_mats-replication.zip)
• To re-create the plots in the main paper, run plots-main-paper.ipynb; This assumes that you have data/diab_txr_mats-replication.pkl, by one of the methods above
• To re-create the plots in the appendix, see the corresponding notebooks

Dependencies

This code was run using Python 3.7 in a conda environment: Running the following commands should cover all the dependencies of the code (e.g., installing pandas will install numpy, and so on)

conda install jupyter
conda install pandas
conda install seaborn
conda install tqdm
pip install pymdptoolbox

Updated Simulator

As we receive suggestions for improving the realism of the sepsis simulator, we will collect them in the sim-v2 branch of this repository, in case it is useful for others. The master branch will remain unchanged to facilitate reproduction of the original paper.

Acknowledgements

First, we would like to thank Christina X Ji and Fredrik D. Johansson for their work on an earlier version of the sepsis simulator we use in this paper.

Second, for some of the code used in the posterior inference over Gumbel variables, we borrowed from Chris Maddison's blog post here

Finally, in this repository (in pymdptoolbox/) we have the source code for the pymdptoolbox package from sawcordwell/pymdptoolbox, which is in turn based the toolset described in Chades I, Chapron G, Cros M-J, Garcia F & Sabbadin R (2014) 'MDPtoolbox: a multi-platform toolbox to solve stochastic dynamic programming problems', Ecography, vol. 37, no. 9, pp. 916–920, doi 10.1111/ecog.00888. We reproduce it here because we needed to make a slight modification to the mdp class to bypass certain checks; in particular, it checks for whether or not the rows of the transition matrix sum to one, but can fail due to floating-point inaccuraries - we replace this check in the main code with an assertion using np.allclose instead of checking for strict equality.