Atari Reinforcement Learning

Aim of this project is to implement simple learning mechanism for Atari 2600 games on ALE Platform (Arcade Learning Environment). As a reinforcement learning technique we use NFQ Learning (see next).

Main Principles of learning:

• First of all, input image of a game is read by ALE interface (lib/ImgProc)
• The input image is encoded using Autoencoder in lib/Autoencoder/Encoder.py
• The extracted features from encoder (and also reward, action selected.. = transition data) are pushed into the NFQ module which saves these transition as a training data for Neural Network learning
• When enough training data are acquired, the NFQ trains the Q-function, which is used as a predictor of the most suitable move

Find out more about NFQ at link.

Dependencies

• Pip
• Numpy
• Theano
• Keras
• ALE

Quick Start

make init

which installs all dependencies

cd example/
python PongAgent.py

Pong game example

To run a simple game example, you first need to get a game binary (you can find one in example/ folder). The next step is to train the encoder lib/Autoencoder/Encoder.py for your game using train method. When the encoder is train well enough, you can use this pre-trained NN within initialization of AleAgent after the first required parameter, which is the Game ROM. AleAgent contains 2 more methods, which are train() and play() which first train the Q-function for prediction and then, can play the game using trained network.