Blackjack AI

简体中文 | English

An AI project for Blackjack based on deep reinforcement learning, using Deep Q-Networks (DQN) to train agents that learn optimal game strategies.

Key Features

• 🧠 AI Agent: Intelligent agent trained using Deep Q-Learning algorithm to master complex Blackjack strategies
• 🎮 Human vs. AI: Play against the trained AI opponent
• 📊 Card Analysis: AI analyzes cards that have appeared to adjust decision strategies
• 🔄 Adaptive Learning: AI continuously improves its gameplay through extensive training

Simplified Version Note

Note: The current implementation is a simplified version of Blackjack, focusing on basic hit and stand decisions, excluding the following advanced features:

• Splitting pairs
• Double Down
• Insurance
• Surrender

This simplified design allows the AI to focus on learning core strategies. Future versions plan to gradually add these advanced features to expand game complexity and strategic depth.

Project Structure

blackjack-ai/
│
├── blackjack_env.py   # Blackjack game environment implementation
├── dqn_agent.py       # Deep Q-Network agent implementation
├── play_with_ai.py    # Human vs. AI interface
├── train.py           # AI training script
├── models/            # Saved trained models
└── requirements.txt   # Project dependencies

Installation Guide

1. Clone the repository:

   git clone https://github.com/korman/blackjack-ai.git
   cd blackjack-ai

2. Install dependencies:

   pip install -r requirements.txt

Usage Instructions

Training the AI Model

python train.py

After training, the model will be saved in the models/ directory.

Playing Against the AI

python play_with_ai.py

Follow the on-screen prompts to play:

• h: Hit (request another card)
• s: Stand (end your turn)

Technical Details

State Representation (15-dimensional vector)

The AI understands the game state through a 15-dimensional vector:

• 2 player state features: hand value and availability of usable Ace
• 13 card counting features: tracking the distribution of known cards

Agent Architecture

• Network Structure: Multi-layer Perceptron (MLP) with 15 neurons in the input layer, hidden layers, and 2 neurons in the output layer (corresponding to hit and stand actions)
• Learning Mechanism: Experience Replay and Double Network Architecture
• Decision Strategy: ε-greedy policy, exploring during training and selecting optimal actions during gameplay

Game Rules

Standard Blackjack rules:

• The goal is to get a hand value as close to 21 as possible without exceeding it
• Number cards are worth their face value, J/Q/K are worth 10 points, and A can be worth 1 or 11 points
• Players can choose to hit (draw a card) or stand (end their turn)
• Busting (exceeding 21 points) results in automatic loss
• When the dealer busts, all non-busted players win

License

MIT License

---

This project was created for educational purposes to learn about deep reinforcement learning and AI applications.