snowkylin / async_rl

Tensorflow implementation of asyncronous 1-step Q learning in "Asynchronous Methods for Deep Reinforcement Learning" with improvement on weight update process (use minibatch) to speed up training.

Play Atari Games with TensorFlow and Asynchronous RL

This is a Tensorflow implementation of asyncronous 1-step Q learning with improvement on weight update process (use minibatch) to speed up training. Algorithm can be fount at Asynchronous Methods for Deep Reinforcement Learning

Demo

Dependencies

• Python
• TensorFlow
• gym (with atari environment)
• OpenCV-Python

Usage

Run play.py to play atari game (default is Breakout-v0) by trained network.

Run train.py to train the network on your computer.

You will get a comparatively good result (40+ score) when t is larger than 2000000. On my computer (i5-4590/16GB/GTX 1060 6GB), the training process need at least 2-3 hours.

Evaluation

You can find the eval at https://gym.openai.com/evaluations/eval_03aUUz45Sc6TBg0vifljwA , which takes 40 hours to train the network.

Credit

• coreylynch/async-rl
• yenchenlin/DeepLearningFlappyBird