Interpretable Policy Learning

Alihan Hüyük, Daniel Jarrett, Mihaela van der Schaar

International Conference on Learning Representations (ICLR) 2021

Code Author: Alihan Hüyük (ah2075@cam.ac.uk)

Overview

This repository contains the necessary code to replicate the main experimental results in the ICLR 2021 paper "Explaining by Imitating: Understanding Decision by Interpretable Policy Learning." Our proposed method, Interpole, is implemented in files adni/main-interpole.py and diag-bias/main-interpole.py for the decision environments considered in the paper, namely ADNI, DIAG, and BIAS.

Usage

First, install pomdp-solve v5.4 inside the empty directory pomdp/ by following the instructions on pomdp.org. Make sure the executable pomdp-solve is located at pomdp/src/pomdp-solve.

Install the required python packages as well by running:

python3 -m pip install -r requirements.txt

Then, the experiments in the paper can be replicated by running:

./adni/run.sh            # generates the results for ADNI given in Table 2
./diag-bias/run.sh       # generates the results for DIAG given in Table 3
./diag-bias/run-bias.sh  # generates the results for BIAS given in Table 4

Note that, in order to run the experiments for ADNI, you need to get access to the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Dependencies

The following packages are needed for running the model:

dill==0.3.2
jax==0.1.75
jaxlib==0.1.52
numpy==1.19
pandas==1.0.5
scikit-learn==0.23.1
scipy==1.5.1

Reference

If you use this code, please cite:

@article{huyuk2021explaining,
  title = {Explaining by Imitating: Understanding Decisions by Interpretable Policy Learning},
  author = {H{\"u}y{\"u}k, Alihan and Jarrett, Daniel and Tekin, Cem and van der Schaar, Mihaela},
  journal = {International Conference on Learning Representations},
  year = {2021},
}