happywu / a3c

MXNET + OpenAI Gym implementation of A3C from "Asynchronous Methods for Deep Reinforcement Learning"

Still in progress.

A3C

This is a MXNET implementation of A3C as described in "Asynchronous Methods for Deep Reinforcement Learning.

Requirement

• openai gym
• mxnet

Flappy Bird

Game source from Using Deep Q-Network to Learn How To Play Flappy Bird.

If you don't want to run FlappyBird, you can ignore this.

To run experiment:

python a3c.py --game-source=flappybird --num-threads=16 --save-model-prefix=a3c-flappybird --save-every=1000

To eval, I have upload a checkpoint of mine, you could try your own parameters.

python a3c.py --test --model-prefix=a3ce-8 --load-epoch=305000 --game-source=flappybird

Notice

If you train on computer without GPUS, please change "devs = gpu(1)" to "devs = cpu()"