search

kevin-fang / reinforced-flappy-bird

Using deep reinforcement learning and policy gradients to train a neural network to play flappy bird
https://www.youtube.com/watch?v=PNE6KFNn1yY
GNU General Public License v3.0
3 stars 0 forks source link
flappy-bird machine-learning neural-network policy-gradient reinforcement-learning

readme

Playing Flappy Bird with Reinforcement Learning

Demo video available at https://www.youtube.com/watch?v=PNE6KFNn1yY

Running

After installing the requirements (NumPy, PyGame, and TensorFlow), run python train_nn.py -n to generate a new model. It will play 20 games of Flappy Bird, train, and do that again, retraining each round. If you have a saved model, run python train_nn.py -l and it will retrain on the model already available in the folder ./models.

Once the model has been trained, it can be tested with python test_flappy_results.py <model path> and it will play the game with the existing model.

Some pretrained models exist in ./saved_models

Game Animation

Requirements:

Anaconda (https://anaconda.org/anaconda/python), NumPy (conda install numpy), PyGame (pip install pygame), TensorFlow (conda install -c conda-forge tensorflow-gpu/conda install -c conda-forge tensorflow), Python 3.

Info

Flappy bird implementation written in Python with PyGame from https://github.com/Max00355/FlappyBird, significantly modified to work with machine learning code.

python train_nn.py -n will generate a new model with random initial weights and keep training in iterations of 20 games. It will save a model in a new folder models/.

python test_flappy_results.py <model_name> will play Flappy Bird games with an initially trained model.

neural_jumper.py is a wrapper for the neural network to choose a jump.

tf_graph.py contains the neural network architecture.

config.py contains some configuration variables, such as whether to save images and directories.

global_vars.py contains some global variables - it should not be modified.