mario-neural-network

The code from my video where I try to make an AI from scratch. A neural network, trained by a genetic algorithm, learns to play Super Mario Bros.

What's here

• mario_nn/ — the actual Mario AI
  • main.py — the neural network, breeding/crossover/mutation, and the training loop
  • gym_mario.py — emulator wrapper; reads the NES RAM and builds the 16×13 tile grid the network sees
• Everything else is the practice exercises I did while learning:
  • simple_perceptron.py — a single perceptron by hand
  • perceptrons.py — perceptron that learns to classify points above/below a line
  • neural_network.py — first layers + ReLU with numpy
  • string_evolution.py — genetic algorithm evolving a target string
  • rockets.py — smart rockets evolving toward a target
  • notes.md — notes I took while learning

Running the Mario AI

pip install numpy opencv-python gym-super-mario-bros nes-py
python mario_nn/main.py

main.py currently resumes from a saved population (last_population.npz, not included). To train from scratch, swap elements = load_population() for the commented-out n = 100 / elements = create_population(n) lines.

The exercises additionally use matplotlib, nnfs, and pygame.