FrankPSch / actor-critic

Implementation of A2C and ACKTR in TensorFlow.

Actor-Critic Reinforcement Learning

This project intends to provide a documented and extensible implementation of the A2C and ACKTR algorithms by OpenAI.

Based on the paper by Wu, Mansimov, Liao, Grosse, and Ba (2017): https://arxiv.org/pdf/1708.05144.pdf
Original implementation: https://github.com/openai/baselines

Documentation

The API documentation and a short guide can be found on Read the Docs.

Usage

Prerequisites

The following dependencies need to be installed besides TensorFlow and NumPy (click links for further details):

• OpenAI gym. Install with:

$ pip install gym

• KFAC for TensorFlow. You need the latest version (0.1.1), which currently is not hosted on PyPI. Install with:

$ pip install git+https://github.com/tensorflow/kfac

To use the Atari environments you need:

• OpenAI atari-py. Install with:

$ pip install atari-py

• OpenCV for Python. Install with:

$ pip install opencv-python

This project is only tested on Linux with Python 3.6.5.

Examples

Run the following to train an Atari model (see a2c_acktr.py for further details):

$ python -m actorcritic.examples.atari.a2c_acktr