Play Atari Pong with REINFORCE and Deep Q-Learning

Simple implementations of fundamental reinforcement learning algorithms to learn to play Pong. Instead of using Pixels, the agents perceive the environment using simple hand-crafted features.

Limitations:

• The implementations can only learn to play Pong, but could be adapted to other similar games.
• No optimization has been done - both algorithms could be massively parallelized.

Train a neural network with the REINFORCE algorithm.

To watch a pre-trained agent in action, just type:

python pgm_evaluate.py

To start the training from scratch, just type:

python pgm_train.py

Result after 5000 batches of 10 episodes each: the agent beats the computer-controlled player most of the times.

Train a Deep Q-Network.

To watch a pre-trained agent in action, just type:

python dqn_evaluate.py

To start the training from scratch, just type:

papermill --log-output dqn_train.ipynb dqn_results.ipynb

Result after 10000 episodes: the agent wins most games with nearly perfect score.