Reinforcement Learning using Deep Q-Network for FlappyBird

pygame.window.2022-07-28.22-35-29_Trim.mp4

How to Run?

git clone https://github.com/mewbot97/FlappyBird_DQN.git
cd FlappyBird_DQN_master/src/
python dqn_play.py

Train new model/ retrain existing model

PS:delete the existing model('b2d.pt') in '/src/model/' folder to train a new model

cd FlappyBird_DQN_master/src/
python dqn_train.py

NN architecture can be changed in '/src/dqn_model_class.py'

The states used are

horizontal distance to the next pipe;
difference between the player's y position and the next hole's y position.
refer here for more info

The agent receives +1 reward for each timestep alive and -100 when it crashes