Learn2Slither is a 42 school project focused on implementing a reinforcement learning agent for the Snake game. The project involves creating an AI that learns to navigate through Q-learning implementation, enabling the snake to collect apples while avoiding collisions.
- Board size: 10x10 grid
- Two green apples randomly placed (increase snake length)
- One red apple randomly placed (decrease snake length)
- Initial snake length: 3 cells
- Game ending conditions: wall collision, self-collision, or zero length
The snake agent operates with information from four directions around its head, representing its local perception of the environment.
This restriction shapes the state space and decision-making process. The agent's actions are limited to four directional movements: UP, LEFT, DOWN, and RIGHT.
The agent processes information about the game state through a simple deep neural network (DNN) with multiple layers. The architecture begins with 20 input neurons that capture essential environmental data, processes this through two hidden layers, and produces action values through the output layer.
The input layer consists of two types of environmental sensors:
Each direction contains 4 normalized distance values representing:
- Distance to the nearest wall
- Distance to the nearest green apple
- Distance to the nearest red apple
- Distance to the nearest snake body segment
Four binary neurons (0 or 1) detect immediate collision threats in adjacent cells. These neurons activate when:
- A wall is directly adjacent in that direction
- A snake body segment is directly adjacent in that direction
Hidden Layers
The network processes this input through two hidden layers:
- First hidden layer contains 128 neurons, allowing for complex pattern recognition
- Second hidden layer contains 64 neurons, further refining these patterns
The final layer consists of 4 neurons, each corresponding to a possible movement direction (UP, DOWN, LEFT, RIGHT). These neurons output Q-values, representing the expected future reward for each action in the current state.
Bonus points are awarded for:
- Achieving snake lengths beyond 10 during sessions (15, 20, 25, 30, 35)
- Creating a polished visual interface
- Enabling the snake to generalize its learning across different grid sizes
- Python 3.9 or higher
- pip (Python package manager)
- git
# Clone repository
git clone https://github.com/Sleleu/AI_Snake.git
# Go to directory
cd AI_snake
# Create virtual environnement
python3 -m venv .venv
# Activate virtual environnement
# Linux/macOS:
source .venv/bin/activate
# Windows:
# .venv\Scripts\activate
# Install required packages in virtual environnement
pip install -r requirements.txtThe program can be run with various options to control its behavior:
python3 main.py [options]
-t, --train: Enable training mode for the AI agent-e EPISODE, --episode EPISODE: Specify the number of episodes to run-m MODEL, --model MODEL: Load a pre-trained model from a specified path
-v {on,off}, --visual {on,off}: Enable or disable the GUI (if you want to train your model faster, disable visualization to reduce computational overhead)-step-by-step: Enable step-by-step mode for detailed observation-plot OUTPUT_FILENAME: Save training statistics plots to a specified filename
Example of a training statistics plot:
-p, --player: Enable player mode to play the game manually
Train the AI for 1000 episodes with visualization:
python3 main.py -t -e 1000Load a pre-trained model and watch it play:
python3 main.py -m model/trained_model.pthTrain without visualization for faster processing:
python3 main.py -t -e 1000 -v offYou can modify basic parameters in settings.py like grid size, fruit population, initial snake length and game speed. Since a model can generalize its learning across different grid sizes, you can experiment with settings like:
# Default game size
GRID_SIZE = 30
CELL_SIZE = 15
...
# Default game settings
SNAKE_SIZE = 3
GREEN_FRUITS_NB = 300
RED_FRUITS_NB = 300
...
# Game speed
FPS = 150Here's what it looks like in action:
