Flappy AI · Flight Lab
Teach a whole flock of neural networks to fly.
I wanted to see whether a flock of simple neural networks could learn Flappy Bird. Each bird starts with random weights, and selection and mutation improve the population over successive generations. The Python app lets you watch training, inspect a bird's decisions, and replay a saved model through 100 increasingly difficult levels.
- Progressive levels
- 100
- Networks by default
- 160
- Recorded unseen clears
- 20/20
- The default network has 6 inputs, 12 hidden neurons, and 1 output. Training uses tournament selection, elitism, and Gaussian mutation.
- Train, pause, inspect individual birds, change playback speed, and resume reproducible checkpoints.
- Keep training, held-out validation, and independent completion evaluation separate.
- Included champion cleared 500 pipes across all 100 levels in each of 20 recorded unseen courses.
Observed on 20 independent seeds with the included policy and default physics; results vary with new seeds or settings.
