Deep Q Network

Reproduce performance (rewards) of the following deep reinforcement learning methods using TensorFlow:

• Deep Q Network (DQN): Human-level Control Through Deep Reinforcement Learning

• Dueling DQN: Dueling Network Architectures for Deep Reinforcement Learning

• Double DQN: Deep Reinforcement Learning with Double Q-Learning

This is an easy to understand and modify the DQN structure as well as memory efficiency implementation that can store 1M transitions using ~8GB memory. The model architecture used in this code is not quite similar to the one described in the original DQN paper. I have tested with Pong, Breakout, and MsPacman so far.

It took ~25 hours of training to reach its first 400 points reward on Breakout evaluation using 1 GTX 1080.

Environment I used

• Python 3.6
• TensorFlow 1.10
• OpenAI Gym 0.10.5
• OpenCV 3.4.2
• mpi4py 3.0.0

How to run

Set up hyper-parameters in config.py. To run the program:

python train.py