Snake RL

A reinforcement learning implementation of the classic Snake game using Deep Q-Network (DQN). The agent learns to play Snake through self-play using PyTorch.

Features

• Dueling DQN Architecture: Advanced neural network with value and advantage streams
• Prioritized Experience Replay: Focuses learning on important experiences
• N-step Returns: Better value estimation through multi-step returns
• Optimized Environment: Fast game engine with collision detection and visual rendering
• Multiple Training Modes: Train, play, evaluate, and benchmark the agent

Performance

Average score: ~43 | Best: 80+ | Success rate: 100%

Installation

1. Clone and enter directory:

git clone <https://github.com/thanhtulearncode/Snake>
cd Snake

2. Create virtual environment:

python -m venv venv
venv\Scripts\activate  # Windows
source venv/bin/activate  # Linux/Mac

3. Install dependencies:

pip install -r requirements.txt

Usage

# Train agent
python main.py --mode train --episodes 5000

# Watch agent play
python main.py --mode play --games 5

# Evaluate performance
python main.py --mode eval --episodes 500

# Benchmark speed
python main.py --mode benchmark

Arguments:

• --mode: train, play, eval, or benchmark
• --episodes: number of episodes (default: 5000)
• --render: show game during training
• --games: number of games in play mode (default: 5)

Technical Details

Algorithm: Dueling DQN with n-step returns, prioritized replay, and double Q-learning

State: 24 features (danger detection, direction, food position, space analysis)

Actions: 3 (turn left, straight, turn right)

Rewards: Food consumption (+10 + length bonus), approaching food (+0.8), collisions (-10)