Project intro

    This project uses the main framework of pytorch+pygame+opencv,Papers on Deepmind《Human-level control through deep Reinforcement learning》,Choosing Flappy Bird as a game instance to reproduce its core algorithm.As we finish trainging,this brid can beat most of people players.

                                                  《Human-level control through deep Reinforcement learning》

     AS we can see,DQN(blue)'s performance is better than traditional machine learning algorithms.

Project video and image

Project structure

Reinforcement Learning five Elements

Deep Q network

Explanation of file or folder

final_log_file.txt :this txt file records the detail step 's state of the bird;
score.txt this txt :file records the scores of the flappy bird every 1000 steps;
score.txt this txt :records the socres of the flappy bird afters 5390000 iterations;
assets :this foler records the source files of the Game Flappy Bird;
saving_nets2:this folder records the last few tiemes models;

Data Analysze

the avg scores of 2900000 iterations.

the avg scores of 5300000 iterations.as we can see ,there is a obvious score growth about 3000000 iterations.Clearly,there is no doubt that the score will continuely grow if we add the training iterations.

At point ot 5300000 ,the avg scores and the peak score of the bird.