VIPER is a tool extracting decision tree policies from an oracle (e.g., a deep Q network) using imitation learning.

Table of Contents

0. Prerequisites
1. Running VIPER
2. Contact

Prerequisites

VIPER requires Python 3, and the Python packages numpy, tensorflow, scikit-learn, gym, and opencv-python (all of which can be installed using pip).

Running VIPER

We have included an example of how to run VIPER using the DQN oracle for the Atari Pong environment obtained from OpenAI baselines (available at https://github.com/openai/baselines/). To try this example, run

$ cd python
$ python -m viper.pong.main

Contact

For questions, feel free to contact obastani@csail.mit.edu.