flappy-rl 0.2.3

FlappyBird game with Reinforcement Learning AI

🏆 FlappyAlpha: Reinforcement Learning vs Human

🎮 Overview

FlappyAlpha is a reinforcement learning (RL) project inspired by AlphaGo, where an RL agent learns to play a custom Flappy Bird game environment with realistic physics (gravity, collision, etc.). The agent (FlappyAlpha) is trained using Q-Learning to maximize its score by passing obstacles and avoiding crashes. The project also features human vs. agent battles to showcase the agent's learning progress.

🎮 Features

• Custom Flappy Bird environment built with Pygame
• Q-Learning agent with reward/punishment system
• RL training for "Beginner" and "Hard" agents
• Human vs Agent battle playground with score visualization
• Performance interpretation and result plots
• Easy to use and play for normal users (a PyPI package is prepared)

🎮 Tech Stack

• Game & UI: Python, Pygame
• Reinforcement Learning: Q-Learning (NumPy)
• Data Analysis & Result Visualization: NumPy, Matplotlib

🎮 Simple Demo

📂 Project Structure

The project uses a standard src layout to separate the installable package from developer scripts.

zephyra/
│
├── .github/workflows/      # CI/CD pipeline for PyPI publishing
├── src/
│   └── flappy_rl/          # Core package source code
│       ├── game/           # Game environment logic
│       ├── models/         # Pre-trained Q-table models (.npy)
│       ├── assets/         # Game assets
│       ├── __init__.py
│       └── main.py         # Entry point for PyPI package to run
│
├── scripts/                # Scripts for developers (training, testing)
│   ├── train_qLearningBEGINNER.py
│   ├── train_qLearningHARD.py
│   ├── playground.py
│   └── playgroundHARD.py
│
├── results/                # Output for training & playing result graphs
├── LICENSE
├── pyproject.toml          # Main project configuration for packaging
├── README.md
└── requirements.txt        # Dependencies

---

🚀 Quick Start (For Users that are just want to play only)

Get started in under a minute! Play against the AI directly from your terminal. I've make it as an offical python package.

1. Install from PyPI

pip install -U flappy-rl

2. Challenge the Agent

• Play against the Beginner agent:

  flappy-alpha --mode beginner
  

• Play against the Hard agent:

  flappy-alpha --mode hard
  

Example

---

🛠️ For Developers & Researchers

This section is for those who want to dive deeper, retrain the models, or experiment with the code.

1. Clone the Repository

git clone https://github.com/kaifeng-cmd/zephyra.git
cd zephyra

2. Install Dependencies

It's recommended to use a virtual environment.

# Create and activate a virtual environment (optional but recommended)
python -m venv venv

# For Windows
venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

3. Train the RL Agent

Pre-trained models are included, but you can start your own training sessions.

• Train the Beginner agent (2500 episodes):

  python scripts/train_qLearningBEGINNER.py
  

• Train the Hard agent (3000 episodes):

  python scripts/train_qLearningHARD.py
  

  Training progress and results will be saved as .png plots in the results/ folder.

4. Run the Playground

Test the agents or play against them using the original developer scripts.

• Play against the Beginner agent:

  python scripts/playground.py
  

• Play against the Hard agent:

  python scripts/playgroundHARD.py
  

---

🎮 Highlights & Insights

• The RL agent receives positive rewards for passing pipes/blocks and negative penalties for crashing.
• As training episodes increase, the agent's performance improves dramatically:
  • After 2500 episodes (beginner mode): Best score 17.
  • After 3000 episodes (hard mode): Best score 75 (441% increase).
  • After 4000 episodes: Best score 3693 (near-perfect play).

  U can modify the training parameter, ex. EPISODES = 4000 if u want.

• Human players can easily beat the Beginner agent (2500 episodes), but the Hard agent (3000 episodes) is a tough opponent, often outperforming humans in average score, and I'm sure for > 4000 episodes training, human can't beat RL agent as it plays near perfect.

🎮 Acknowledgements

Inspired by AlphaGo and the power of reinforcement learning.

---

💎 Feel free to fork, do experiment, and challenge FlappyAlpha.