Flappy Bird AI with Scratch Neural Network

Project Overview

In this solo project, I recreated Flappy Bird entirely from scratch using Pygame, then implemented a custom neural network library (no external ML frameworks) that learns to play the game.

Instead of gradient descent or backpropagation, I used an evolutionary approach with mutation. Networks are evaluated based on fitness (game score), the best network is selected, and all others mutate its weights and biases to form the next generation. Over time, networks learn effective flight strategies.

AI bird learning to play via generations.

Technical Design

Game built in Pygame with custom sprites, physics, and collision detection.
Neural network library coded from scratch, including:
- Dense layers
- ReLU, Tanh, and Softmax activations
- Forward pass functionality
Input features for the NN:
- Bird Y position
- Bird vertical velocity
- Horizontal distance to next pipe
- Vertical distance to pipe gap
Output: a single value (-1 → don’t flap, +1 → flap) via Tanh activation.
Fitness = score (pipes passed + survival time).
Mutation: applies random Gaussian deltas to weights/biases of the best network each generation.

Results

Over generations, performance improved significantly: early generations would crash instantly, while later generations consistently survived dozens of pipes. By logging best, worst, and average fitness as well as weight/bias statistics, I visualized learning progress.

Fitness over generations

Weight statistics during evolution

Future Improvements

Add backpropagation training for direct supervised learning (hybrid evolution + gradient descent).
Experiment with different mutation strategies (e.g., crossover).
Improve visualization: plot live learning curves alongside gameplay.
Package as a standalone executable for others to try.

Explore the Code

You can view the full source code, training loop, and game implementation on GitHub.

View on GitHub