Flappy Bird Genetic Evolution

Train a neural network using genetic evolution to play flappy bird! For a tutorial on how to get started with your own genetic algorithms, take a look here.

Final Results

Algorithm Details

• Starts out with a pool of 50 models
• Each iteration based on fitness scores of the models, perform crossover
• Crossover would swap the first layers (input -> hidden) for both the selected parents
• Random mutation ensures that the models are changed at every iteration

Progress Screenshots

Stage 1

Initially all the models would do the "same" wrong thing. So they would all die out quickly.

Stage 2

After some time however, they would start to show more variation, but still perform the wrong moves. This gives us a spread of flappy birds throughout the screen (lengthwise).

Stage 3

After a bit of training (~1hr) the spread decreases and is more concentrated at the height where there is a hole in the pipes. They start performing a lot better since they now understand when to flap and when not to.

Stage 4

The model with the maximum fitness can be considered as a trained model and it would perform much better than the average human.

Disclaimer

• Based on Flappy Bird clone in pygame, https://github.com/sourabhv/FlapPyBird