Flappy Bird Neural Network Demo: Training AI to Play the Game Using Genetic Algorithms

This open-source educational demo uses native JavaScript, neural networks, and genetic algorithms to train AI to play Flappy Bird. Packaged in a single HTML file (no external dependencies), it visualizes the AI's learning process to help beginners understand abstract AI concepts intuitively.

Neural networks and genetic algorithms are core AI concepts but often hard for beginners to grasp due to abstract math and pseudocode. This project addresses this by visualizing the AI's evolution from random behavior to game mastery, lowering the understanding barrier.

The AI is controlled by a simple neural network. It takes game state inputs (bird position, pipe distance, pipe heights) and outputs a jump decision. This end-to-end learning mimics biological reflexes without explicit game rule programming.

Instead of backpropagation, the network's weights are evolved via genetic algorithms: 1. Initialize random weight populations; 2. Evaluate fitness (survival time/score);3. Select top performers as parents;4. Cross weights and add random mutations to generate new generations;5. Iterate to improve performance—mimicking natural selection.

Key visualizations include: population performance (best vs worst per generation), evolution trajectory (performance over generations), decision visualization (how inputs lead to jump decisions), and gene diversity (mutation's role in new behaviors). These help understand concepts like convergence and diversity maintenance.

Educational Advantages: Zero threshold (single HTML file, no setup), readable native JS code (transparent implementation), instant feedback (game success/failure). Technical Highlights: Canvas rendering for smooth gameplay, vector operations for neural network forward propagation, requestAnimationFrame for async loops, and state management for game/AI lifecycle.

This project can be extended: modify network topology to test effects; adjust GA parameters to study selection pressure; replace game environments to test generalization; add multi-agent competition for co-evolution.

This demo turns complex NN and GA concepts into intuitive visual experiences. Observing AI learn Flappy Bird from scratch builds learners' intuitive understanding, making it an ideal starting point for studying neuroevolution. Its concise implementation also aids in grasping core algorithms.