chenxingqiang / dqn-benchmark

Pytorch realize DQN,Duel DQN and Double DQN

[PYTORCH] Deep Q-learning Pytorch with flappy bird

Introduction

Here is my python source code for training an agent to play flappy bird. It could be seen as a very basic example of Reinforcement Learning's application.

demo

How to use my code

With my code, you can:

• Train your model from scratch by running bash train.sh
• Test your trained model by running bash test.sh

Trained models

train with python

python ./tool/train.py --model DuelDQN --train_model DDQN --clip_grad True --Fire train
#python ./tool/train.py --model DQN --train_model DDQN --clip_grad True --Fire train

--model means network structure
--train_model means using DQN train-policy or double DQN train-policy

train with shell

You could find DDQN model at model/DDQN
You could find DQN model at model/DQN

NetWork

You could find onestream at demo/img_net--->DQN
You could find duel-stream at demo/img_net--->DuelDQN

Result

Methods	Game	Mean score(second)
DQN	flappy bird	61.2s
Double DQN	flappy bird	>160s
Duel DQN	flappy bird	> 150S
Duel Double DQN	flappy bird	>300s

Requirements

• python 3.6
• pygame
• cv2
• pytorch-1.0
• numpy

Reference

• DQN2013
• DQN2015
• Double DQN
• Duel DQN