AI learns to play Snake!

A neural network learns to play snakes in the terminal

It was built from scratch using Rust and Ratatui

Check out this for an older version of the AI that uses macroquad for GUI

Usage

• Clone the repo

git clone git@github.com:bones-ai/rust-snake-ai-ratatui.git
cd rust-snake-ai-ratatui

• Run the simulation

cargo run --release

Configs

• The project configs file is located at src/configs.rs
• Update IS_LOW_DETAIL_MODE for a minimal UI, it runs faster
• Set IS_SAVE_BEST_NET to train your own network. Networks are saved in data/net.json, once saved you can use the trained network by setting IS_LOAD_SAVED_NET
• Use data/net-100.json to watch the AI complete the game for 15x15 board size
• Set USE_GAME_CANVAS to true and update VIZ_GAME_SCALE to scale the game if needed.

Resources

Here are some excellent resources to learn more about genetic algorithms: - Video Series: Genetic Algorithm by The Coding Train. - Book: Nature of Code by Daniel Shiffman for those who prefer reading. - I highly recommend checking out Joshka's fork for more idiomatic Rust code.

Algorithm

1. Initialization:
2. The simulation begins at Generation 0.
3. A new population of snakes is created, each with a neural network initialized with random weights and biases.
4. Game Update:
5. Each step, every game is updated by passing vision inputs to the neural network to decide the snake's action.
6. A game is flagged as complete if:
   • The snake collides with walls or itself.
   • The snake fails to eat food within a certain number of steps, preventing indefinite looping.
7. Generation Completion:
8. The generation continues updating each game until all games are complete.
9. Fitness Evaluation:
10. At the end of each generation, snakes are ranked based on their performance.
11. Parent Selection:
12. Parents for the next generation are chosen based on rankings. Higher-ranked snakes have a higher probability of being selected as parents.
13. Reproduction:
14. Techniques such as roulette wheel selection, elitism, and other methods are used to generate children for the next generation.
15. New Generation:
16. A new population is created, and the process repeats from step 2 until the simulation is manually stopped.

This iterative process leads to snakes fine-tuning their strategies, resulting in longer snakes over time.

Extras

• I mostly post about my projects on - Twitter - @BonesaiDev
• Video of snake completing the game
• Check out my other projects on github https://github.com/bones-ai