Learn2Slither

A reinforcement learning project where a snake learns to navigate and survive in a dynamic environment through Q-learning.

Project Overview

Learn2Slither implements a classic snake game controlled by an AI agent that learns optimal strategies through reinforcement learning. Using a neural network built with Keras/TensorFlow, the snake progressively improves its performance through multiple training sessions.

Features

Game Environment

• Flexible Board Size: Default 10x10 grid, configurable to any size (width 3-24, height 3-13)
• Dynamic Elements: Green apples (increase length), red apples (decrease length)
• Snake: Starts with length of 3, randomly positioned

AI Implementation

• Learning Framework: Q-learning with neural networks (Keras/TensorFlow)
• State Representation:
  The snake's "vision" is encoded as 12 neurons (3 features × 4 directions: UP, DOWN, LEFT, RIGHT):
  • 0/1: Is it an obstacle?
  • 0/1: Is it green food?
  • 0-1: Normalized distance to that object
• Action Space: 4 possible movements
• Reward System: Simple yet effective (+1 for eating apple, -5 for death)

Performance

• Average Length: 28 cells
• Maximum Length: 50 cells
• Multiple Trained Models: Various training configurations available

Technical Features

• Model Management: Save and load trained models
• Visualization: Graphical interface with adjustable speed
• Step-by-Step Mode: For detailed analysis of agent decisions
• Training Configuration: Command line parameters for sessions, learning mode, map size, etc.

Installation

git clone https://github.com/ivotints/Learn2Slither.git
cd Learn2Slither
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Running the Snake

# Start with default settings
python3 start.py

# Start with a pre-trained model
python3 start.py --load_model models/Gorynych_v1.5_4/snake_model_190_20250410_15:20.keras

# Train a new model with 10000 episodes
python3 start.py --episodes 10000 --name my_model

# Run without graphics
python3 start.py --no_graphics

# Run in evaluation (no training) mode
python3 start.py --evaluation_mode --load_model models/Gorynych_v1.5_4/snake_model_190_20250410_15:20.keras

# Show agent vision in terminal
python3 start.py --show_vision

# Use a custom map size (e.g. 20x13)
python3 start.py --map_width 20 --map_height 13

# Use graphical lobby
python3 start.py -ul

Command Line Arguments

Argument	Short	Type	Default	Description
--no_graphics	-ng	flag	False	Turn off graphics
--evaluation_mode	-e	flag	False	Run in evaluation mode (no training)
--load_model	-lm	str	None	Path to model to load
--name	-n	str	model	Name of model folder
--first_layer	-fl	int	32	Number of neurons in the first layer
--second_layer	-sl	int	16	Number of neurons in the second layer
--episodes	-ep	int	10000	Number of training episodes
--show_vision	-sv	flag	False	Show agent vision in terminal
--use_lobby	-ul	flag	False	Show configuration lobby
--map_width	-mw	int	10	Width of the map (3-24)
--map_height	-mh	int	10	Height of the map (3-13)
--show_history	-sh	flag	False	Show training history plot after training

Training Process

The agent learns through trial and error, with the neural network adjusting its weights based on the rewards received. The learning process involves:

1. Exploration: Initially taking random actions to discover the environment
2. Exploitation: Gradually favoring actions that lead to higher rewards
3. Iterative Learning: Refining decision-making over thousands of game sessions

Results

After extensive training, the snake AI demonstrates impressive capabilities:

• Efficiently navigates to food sources
• Avoids walls and self-collisions
• Achieves an average length of 28 cells
• Maximum recorded length of 50 cells