ChenyuEcho/texas-holdem-ppo

🎰 Texas Hold'em PPO Agent

Play heads-up Texas Hold'em poker against AI agents trained with Proximal Policy Optimization (PPO)!

🎮 Features

• —Multiple AI Opponents: Choose from CNN or MLP-based agents at different training stages
• —Real-time Game Info: Track pot size and bet amounts
• —Professional UI: Casino-themed dark interface
• —Fair Play: Watch agent actions and see cards in real poker order

🤖 Available Models

• —Stage 1: Early training agents (vs rule-based opponents)
• —Stage 2: Improved agents (fine-tuned vs frozen agents)
• —Stage 3: Final agents (advanced fine-tuning)

Each stage available in both:

• —CNN: Card-aware convolutional network (~357k params)
• —MLP: Pure multi-layer perceptron (~120k params)

🎯 How to Play

1. 1.Select an opponent from the dropdown
2. 2.Set your starting chips
3. 3.Click "🎲 New Match" to begin
4. 4.Use action buttons: Call, Raise, Fold, or Check
5. 5.Try different opponents to test various strategies!

🧠 Architecture

Trained using PPO with a three-stage curriculum:

• —Stage 1: Multiple agents vs rule-based opponents
• —Stage 2: Fine-tuning with replay buffer
• —Stage 3: Advanced refinement

See the GitHub repository for training details and source code.

📊 Model Comparison

Compare CNN vs MLP architectures to see which performs better:

• —CNN uses card-structure inductive bias
• —MLP processes features directly
• —Both trained with identical hyperparameters

Built with PettingZoo, PyTorch, and Gradio