Click a button to begin
Original behavior - mutate top performers
Mario AI - Neuroevolution
đ§Ŧ What is Neuroevolution?
Neuroevolution trains neural networks using genetic algorithms instead of traditional backpropagation. Networks "evolve" through mutation and selection - just like biological evolution!
- Create random neural network weights
- Evaluate by playing the game (fitness = distance traveled)
- Select the best performers
- Mutate/Breed to create new generation
- Repeat until it beats the level!
đ Training Modes
- Simple: Mutation only - copy and mutate top performers
- SBX Crossover: Two-parent breeding using Simulated Binary Crossover. Creates offspring near parents, good for fine-tuning.
- Uniform Crossover: Random gene mixing from two parents. More exploration, can combine diverse strategies.
- Optimize: Start from pre-trained reference weights and continue training. Good for pushing past Level 1-1!
đī¸ Network Architecture
Input: 80 values (7Ã10 tile vision + 10 row encoding)
Hidden: 9 neurons with ReLU activation
Output: 6 buttons (LEFT, RIGHT, A, B + filtered UP/DOWN)
⥠Training Pipeline
Foreground: Shows best performer playing
Background: Runs population evaluation (headless, fast)
Python: Handles neural network evolution via PyScript
Previous