A repository for methods to learn to play Pong, using Deep Reinforcement Learning methods.
Install Docker and nvidia-docker, then build the image as
docker build -t deeprl-pong .
After the build you can open a terminal into the docker container as
nvidia-docker run --rm -it deeprl-pong
After getting a working Conda distribution, install the environment as
conda env create -f environment.yml
Then log into the environment
source activate deeprl-pong
And install the addons for Atari games
pip install gym[atari]
policygradientpytorch.py
A Pytorch implentation of the Policy Gradients method, maily based on Andrej Karpathy's blogpost on the topic.
policygradientkeras.py
Keras version of the Policy Gradients method. Way less efficient, as Keras does not allow to tinker with the gradients as easily as Pytorch.