Deep Reinforcement Learning Lab (DRL Lab)

English | 简体中文

"Bridging theory and practice: An exploration from Markov Decision Processes to Rainbow DQN."

drl_lab is a modular, extensible Deep Reinforcement Learning (DRL) experimental platform designed to provide out-of-the-box implementations of classic control tasks with a modern debugging experience.

✨ Features

• Modern Architecture: Built on PyTorch and Gymnasium, featuring a modular design for Agent, Task, and Trainer.
• Rainbow DQN Integration:
  • ✅ Double DQN (DDQN): Mitigates Q-value overestimation.
  • ✅ Dueling Networks: Separates value and advantage streams for faster convergence.
  • ✅ Huber Loss: Gradient clipping for stability.
• Interactive TUI: Terminal User Interface powered by Textual, enabling real-time visualization of training and inference (Braille animations, status dashboards, live logs).
• Developer Friendly: Lifecycle hooks (pre_training, on_step, etc.) and a standardized BaseTask interface.
• Practical Optimizations: Includes Reward Shaping for sparse reward tasks like CartPole.

🚀 Quickstart

Installation

This project uses uv for package management:

# Clone the repository
git clone https://github.com/1shin-7/rl_lab.git
cd rl_lab

# Sync dependencies
uv sync

Training

Start training the CartPole task with TUI visualization enabled:

uv run rlab train cartpole --visual --episodes 500

Inference

Run inference using a trained model:

uv run rlab infer cartpole --visual --weight outputs/cartpole.pth

Clean

Remove generated models and plots for a task:

uv run rlab clean cartpole

📚 Documentation

• Development Overview: Architecture, Tasks, and Hooks.
• TUI Guide: Visualization and UI design.
• CLI Reference: Detailed documentation for all CLI commands.

🤝 Credits

• PyTorch: The backbone of the entire project, enabling flexible and efficient deep learning.
• Gymnasium: Standardized RL environment API.
• UI/UX: Textual, Loguru, and Rich were used to provide a convenient debugging interface.
• Gemini: For assisting in problem-solving throughout the internship.