BlackjackNN - Deep Q-Learning Blackjack Agent

A neural network-based Blackjack AI that learns optimal playing strategies through deep Q-learning. The project includes both training capabilities and interactive GUI interfaces for playing against or getting advice from the trained agent.

Features

• Deep Q-Learning implementation for Blackjack strategy learning
• Interactive Blackjack GUI for playing games
• Decision Helper GUI for getting real-time advice
• Detailed training metrics and logging
• Basic and advanced reward shaping based on optimal Blackjack strategy

Project Structure

• main.py - Core Blackjack game environment and logic
• blackjack_agent.py - Deep Q-Learning agent implementation
• train_agent.py - Training script with metrics logging
• blackjack_gui.py - Interactive Blackjack game interface
• decision_helper_gui.py - Real-time strategy advice interface

Installation

1. Clone the repository:

git clone https://github.com/jonathanung/blackjackNN.git
cd blackjackNN

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate

3. Install required packages:

pip install -r requirements.txt

Usage

Training the Agent

python ./agent/train_agent.py [num_episodes]

This will:

• Train the agent for the specified number of episodes (default is 5000)
• Generate training metrics plots
• Save the trained model as 'blackjack_agent.pkl'
• Create detailed training logs

If you want to train the agent on a GPU or MPS, you can uncomment the relevant lines in blackjack_agent.py under the __init__ method.

Playing Blackjack

python ./gui/blackjack_gui.py

Features:

• Play Blackjack with a graphical interface
• Get real-time suggestions from the trained agent
• View agent's confidence in its suggestions
• Color-coded advice based on confidence levels

Getting Strategy Advice

python ./gui/decision_helper_gui.py

This will open a GUI interface for getting real-time strategy advice from the trained agent.

Input your:

• Current hand value
• Number of cards
• Dealer's visible card

Get:

• Recommended action
• Confidence levels
• Q-values for all possible actions

Training Metrics

The training process generates:

• Win rate over time
• Average reward progression
• Detailed state-action logs
• Agent's exploration rate (epsilon) decay

Acknowledgments

• Built using PyTorch for deep learning
• Implements standard Blackjack rules and optimal strategy rewards
• Uses epsilon-greedy exploration strategy

Future Features

• More sophisticated reward shaping
• Train with multiple agents in parallel to simulate multiplayer Blackjack
• Implement agents in multiplayer Blackjack game, allowing players to choose their position on the table

Data over 10M episodes in 750 size batches