Flappy Bird AI (NeuroEvolution of Augmenting Topologies)

An autonomous agent trained to play Flappy Bird using NeuroEvolution of Augmenting Topologies (NEAT) and Pygame.

Overview

This project implements a self-learning agent to play Flappy Bird. Instead of hardcoding jumping thresholds, the agent uses a feedforward neural network evolved over multiple generations. Once the network reaches a score threshold of 50, the best genome is serialized and saved.

How it Works

1. Neural Network Input Features

The feedforward neural network accepts three input parameters:

• Bird Y-Coordinate: The vertical position of the bird on the screen.
• Distance to Top Pipe: The absolute distance from the bird to the bottom edge of the upper pipe.
• Distance to Bottom Pipe: The absolute distance from the bird to the top edge of the lower pipe.

2. Genetic Algorithm & Fitness Function

• Base Fitness: The bird accumulates a fitness score of +0.1 for every frame it remains alive.
• Obstacle Clearance: The bird is rewarded with +5 fitness points for each pipe successfully passed.
• Failure Penalty: A penalty of -1 is subtracted from its fitness if the bird collides with a pipe or boundary.
• Generational Mutation: Genomes mutate, perform crossover, and split into species across generations based on configuration parameters.

File Structure

flappy_bird/
├── Font/                               # Font assets for display text
│   ├── PixelifySans-Bold.ttf
│   ├── PixelifySans-Medium.ttf
│   └── PixelifySans-VariableFont_wght.ttf
├── imgs/                               # Game sprites (bird, pipes, background)
│   ├── base.png
│   ├── bg.png
│   ├── bird1.png
│   ├── bird2.png
│   ├── bird3.png
│   └── pipe.png
├── applied_model.py                    # Plays game autonomously using the trained model
├── config-feedforward.txt              # NEAT genetic algorithm configuration
├── flappy.pickle                       # Serialized trained model
├── game.py                             # Manual playable version of the game
└── main.py                             # NEAT training loop script

Requirements

Install dependencies using pip:

pip install pygame neat-python

Running the Code

Train the AI:

python main.py

Run the Trained Agent:

python applied_model.py

Play the Game Manually:

python game.py