This code has been run and developed solely on Ubuntu 18.04. We recommend using the same operating system in order to replicate our results.

You will need the following system dependencies to run our code:

• Conda
• GNU Parallel
• Bash

Run the following commands to install the environment in the requirements.yaml and activate it.

conda env create -f requirements.yaml
conda activate DDT

Make the tests directory your working directory and execute run_tests.sh

cd ./tests
./run_tests.sh

If the tests run the environment should have been installed correctly and you can start running the experiments.

There are scripts to run experiments in parallel located in the exp_cmds directory. All the scripts inside that directory assume exp_cmds is the current directory.

1. Supervised Learning Experiments: Run the following command:

cd ./exp_cmds
./supervised_learning_command.sh

As your output you should see A cancer, ttt and cesarean directories generated under the plots directory. Each of these corresponds to one of the UCI datasets used for comparison. For each UCI dataset you should see a tree and ddt directory showing the results for a regular decision tree and a differentiable decision tree respectively.

2. Reinforcement Learning Experiments:

The following commands will run the experiments used to collect the MLP baselines as well as the differentiable decision tree experiments used in the paper. Notice that this command may take a long time to execute.

cd ./exp_cmds
./collect_baselines.sh
./command_runner.sh

The output of these commands should be a series of folders in the tensorboards folder corresponding to each experiment. To view the results run the following command from the toplevel directory:

tensorboard --logdir tensorboards

This should allow you to view the training curves and relevant data for each of the experiments.

To replicate the figures in the paper please run the following command:

cd exp_cmds
./exp_plotter

Notice that for this command to you must arrange the tensorboard results for each experiment such that your directory structure looks like this: tensorboards/final_exp/$$EXPERIMENT_NAME$$/$$MODEL$$/$$SEED_FOLDER$$. As provided the exp_plotter script expects the follwing experiment names: LunarLander, CartPole and Acrobot. A sample directory structure could look like this:

tensorboards/final_exp/Acrobot/ddt4/env.Acrobot-v1-m.ddt.4-reg.b0.0001.m1.01859.tleaf.mdnever-ceps.1.0-alpha.b1.0.m1.0.inc0.0-lr.0.03-alg.ppo-init.uniform-s.1

Please note that we use lightly modified version of ikostrikov's implementation of PPO, ACKTR and A2C.

1. pg_pipeline.py

This file contains the reinforcement learning pipeline we use.

2. sl_pipeline.py

This file contains the supervised learning pipeline.

We provide a Linear Policy, an MLP and two implementations of the differentiable decision trees. DDTN uses a recursive implementation which is quite slow. For all of our experiments we use FDDTN which takes advantage of a full tree structure to parallelize tree execution in the GPU. The ActorCritic model is simply a wrapper to be able to use ikostrikov's PPO/ACKTR/A2C implementation.

1. The data folder contains the UCI cesarean datset.
2. Plot.py generates the learning curves with error shades shown in the paper.