Snake AI DQN

A Python implementation of the classic Snake game powered by Deep Q-Learning Network (DQN). This project combines PyGame for visualization with PyTorch for deep reinforcement learning, creating an AI agent that learns to play Snake through experience.

SnakeDQN.mp4

Features

• Classic Snake game implementation using PyGame
• Deep Q-Learning Network (DQN) with PyTorch
• Experience replay for stable training
• Model save/load functionality
• Real-time visualization of training
• Configurable hyperparameters
• Support for both CPU and CUDA training

Requirements

pygame
numpy
torch

Installation

1. Clone the repository:

git clone https://github.com/Abhigyan126/Snake-DQN
cd Snake-DQN

2. Install dependencies:

pip install pygame numpy torch

Usage

Run the main script to start training:

python train.py

You'll be presented with three options:

1. Start new training
2. Continue training existing model
3. Exit

The training process will display:

• Real-time game visualization
• Episode progress
• Total rewards
• Current exploration rate (epsilon)

Technical Details

Training pipeline

flowchart LR
    subgraph Game["Snake Game Environment"]
        State["State\n(12 inputs)"] --> Action
        Action --> Reward
        Action --> NextState["Next State"]
    end

    subgraph Agent["DQN Agent"]
        Policy["Policy Network\n(5 layers)"]
        Target["Target Network\n(5 layers)"]
        Memory["Replay Buffer\n(10000 capacity)"]
        
        Policy --> |"Select Action"| Action
        State --> Policy
        Memory --> |"Batch (64)"| Policy
        Policy --> |"Update"| Target
    end

    State --> Memory
    Action --> Memory
    Reward --> Memory
    NextState --> Memory

Loading

Neural Network Architecture

• Input layer: 12 nodes (state space)
• Hidden layers: 64 → 64 → 128 → 64 nodes
• Output layer: 4 nodes (action space)
• Activation function: Leaky ReLU

State Representation

The game state consists of 12 binary values:

• Danger detection (4 directions)
• Current direction (4 possibilities)
• Food location relative to snake (4 directions)

Training Parameters

• Replay buffer size: 10,000
• Batch size: 64
• Learning rate: 0.001
• Discount factor (gamma): 0.99
• Initial epsilon: 1.0
• Minimum epsilon: 0.01
• Epsilon decay: 0.995

Project Structure

snake-ai-dqn/
├── train.py              # Main game and training logic
├── test.py               # Test scrip to run the model
├── README.md             # Project documentation
└── snake_dqn_model.pth   # Saved model checkpoints