Teaching neural networks to flap via natural selection

What it does FlappyLearning is a live browser experiment where a population of neural networks attempts to master Flappy Bird through neuroevolution. The project includes Neuroevolution.js, a standalone library that evolves network weights using genetic algorithms, selecting and mutating birds based purely on survival score.

The interesting bit The library treats weights as genes: each generation breeds, mutates, and culls networks based solely on how long a bird survives. It is a compact, fully client-side illustration of evolutionary computation.

Key highlights

• Pure JavaScript implementation running entirely in the browser
• Configurable perceptron structure, population size, elitism rate, and mutation parameters
• Neuroevolution.js is decoupled from the game logic for reuse in other toy problems
• Live demo visualizing each generation’s incremental progress (and frequent crashes)
• 3,993 stars, suggesting it has long served as an accessible entry point into ML

Caveats

• The README is minimal: it documents API options but offers no theory, benchmarks, or maintenance history.
• Code snippets rely on older JavaScript conventions (var, non-modular scope), so expect a pre-ES6 dialect.
• It is strictly a teaching demo, not a general-purpose reinforcement-learning framework.

Verdict Grab it if you want a visual, low-friction introduction to neuroevolution or a seed for browser-based genetic algorithms. Pass if you need production-grade ML tooling or rigorous algorithmic documentation.