README.md

metadata

title: QP-RNN Interactive Demo
emoji: 🎮
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.35.0
app_file: app.py
pinned: false
license: mit

QP-RNN: Quadratic Programming Recurrent Neural Network

This is an interactive demo for the paper "MPC-Inspired Reinforcement Learning for Verifiable Model-Free Control" (L4DC 2024).

What is QP-RNN?

QP-RNN is a novel neural network architecture that combines:

• 🎯 Structure of Model Predictive Control (MPC)
• 🧠 Learning capabilities of Deep Reinforcement Learning
• ✅ Verifiable properties (stability, constraint satisfaction)

At each time step, QP-RNN solves a parameterized Quadratic Program:

min   0.5 * y'Py + q(x)'y
s.t.  -1 ≤ Hy + b(x) ≤ 1

Where the parameters (P, H, q, b) are learned through RL instead of derived from a model.

Demo Features

This interactive demo lets you:

• 🎮 Control a double integrator system with QP-RNN
• 🔧 Adjust controller parameters in real-time
• 📊 Visualize system response and phase portraits
• 📈 See performance metrics and constraint satisfaction

Key Advantages

1. Interpretable: QP structure provides clear understanding
2. Verifiable: Enables formal stability and safety analysis
3. Efficient: Fixed-iteration solver suitable for real-time control
4. Robust: Handles constraints and disturbances naturally

Links

• 📄 Paper
• 💻 GitHub Repository
• 🤖 Full Training Code

Citation

@InProceedings{lu2024mpc,
  title={MPC-Inspired Reinforcement Learning for Verifiable Model-Free Control},
  author={Lu, Yiwen and Li, Zishuo and Zhou, Yihan and Li, Na and Mo, Yilin},
  booktitle={Proceedings of the 6th Conference on Learning for Dynamics and Control},
  year={2024}
}