An AI that plays Atari 2600 Pong. Trained using reinforcement learning using OpenAI Gym and Keras

You might need to install Swig before being able to use the python libraries.

• Windows
• Arch: sudo pacman -S swig
• Debian/Ubuntu: sudo apt install swig

You need to have OpenAI Gym and Keras installed, and wheel is needed as a dependency.

pip install wheel gym[all] Keras

Run python train_agent.py weightfile.h5 to train an agent. If weightfile.h5 already exists, the training will continue from the last point, otherwise a new training run will be started and the progress saved to a new weights file.

A logfile with the training progress will be written to weightfile.h5.log.

If you do not want to show the game screen (to speed up training), set the render variable in the script to False

The repository also provides start_weights.h5 containing the randomly initialized weights that were used to train the trained_weights.h5. Although I did not have any problems with the current iteration of the training procedure, the method can be very sensitive to the initial weights and it might happen that your agent does not learn. For those cases make a copy of start_weights.h5 to have the same starting point I had during training.

You can run a game in real time with a weights file using python use_agent.py weightsfile.h5 To run the agent with the trained weights provided run python use_agent.py trained_weights.h5. If use_agent.py is given a second argument it will save each frame as a .png file to the media directory using the second argument as a prefix for the filename.

Untrained	Trained

The code is in part based on https://github.com/mkturkcan/Keras-Pong