Playing Retro Airstriker-Genesis game using Deep Q-learing with Experience replay

Introduction

I implement the simple DQN algorithm to play Openai-retro Airstriker-Genesis game to studing basic deep reinforment learning.

Algorithm

Here is pseudo-code of Deep Q-learnig with experience replay.

Networks

Here is the neuron network portion of DQN-v2013, it has three convolutional layers and two ful-connetion layers. The input is the original image from enviroment itself, and ouput is propability of each action of action space.

Requirements

• Python 3.6+
• openai retro
• tensorflow 1.5.0
• numpy 1.14.0

Usage

First, install all requirements.

To train a model:

$ python main.py --train True

To test your model via play the game:

$ python main.py --train False

Results

Currently, our best model can play the game warping to level 2.

Simple results

History graph of loss.

License

WTF (MIT)