How the AI works
Flappy Bird is solved with NEAT (NeuroEvolution of Augmenting Topologies). Instead of training one network with backpropagation, a population of 50 small neural networks plays simultaneously, and evolution — not gradient descent — improves them.
Inputs → outputs
Each bird's network reads a few normalized inputs: its vertical position and velocity, plus the horizontal and vertical distance to the next pipe gap. It outputs one value; above a threshold, the bird flaps.
Evolution loop
- Fitness: birds that survive longer and pass more pipes score higher.
- Selection + crossover: the best networks are recombined.
- Mutation: weights change and new nodes/connections are added only when useful, so networks stay compact.
What you see on screen
The live network panel shows neurons firing and weighted connections in real time, so you can watch the flying policy emerge generation after generation.