#Deep Q Learning for ATARI using Tensorflow

(Version 1.0, Last updated :2016.03.17)

###1. Introduction

This is tensorflow implementation of 'Playing Atari with Deep Reinforcement Learning'.

This is renewal of (https://github.com/mrkulk/deepQN_tensorflow)

I used mrkulk's emulator interface and replay memory code and I made networks and main module

It needs 1~3 days to training.

I'm working on multiprocessing version for fast training.

You can also check A3C and batch-A3C implementations

A3C : https://github.com/gliese581gg/A3C_tensorflow

Batch-A3C : https://github.com/gliese581gg/batch-A3C_tensorflow

###2. Usage

python main_multithread.py (args)

where args :

-weight (checkpoint file) : for test trained network or continue training (default : None)
-network_type (nips or nature) : nature version is more complex, need more time for training but has better performance.(default : nips)
-visualize (y or n) : show opencv window for game screen or not (default : y)
-gpu_fraction (0.0~1.0) : fraction of gpu memory to use. Needs roughly 1~1.5 Gb. (default : 0.9)
-db_size (integer) : size of replay memory. Take 8Gb for size 1,000,000 (default : 1,000,000)
-only_eval (y or n) : doing only evaluation without training if set to y (default : n)

###3. Testing with pretrained networks

python main_multithread.py -network_type (nips or nature) -weight pretrained/(nips or nature)_pretrained -only_eval y

###4. Requirements:

• Tensorflow
• opencv2
• Arcade Learning Environment ( https://github.com/mgbellemare/Arcade-Learning-Environment )

###5. Video

https://www.youtube.com/watch?v=GACcbfUaHwc

###6. Changelog

-2016.03.17 : First upload!