DeePC-HUNT

Data-enabled predictive control hyperparameter tuning via differentiable optimization layers

DeePC-HUNT is a method for optimising over the hyperparameters of DeePC using analytical policy gradients and differentiable optimization layers. This method has been developed as part of my bachelor thesis, carried out at the Automatic Control Laboratory (IfA). Supervised by Alberto Padoan, Keith Moffat and Florian Dorfler.

Developed in a conda environment on Ubuntu 22.04 with Python 3.10.

Differentiable DeePC layer is inspired by Differentiable MPC and built using CvxpyLayers.

Installation

Clone the repo and install from source

cd DeePC-HUNT && pip install -e .

Extra packages necessary for running the example notebooks are in examples/requirements.txt if needed.

pip install -r examples/requirements.txt

DeePC-HUNT has the following dependencies.

Python3
PyTorch >= 1.0
CvxpyLayers >= 1.0

Usage

Data-enabled Predictive Control (DeePC) is a data-driven non-parametric algorithm for combined identification (learning) and control of dynamical systems. It leverages the solution of the following optimization problem in a receding horizon fashion.

DeePC can achieve performance that rivals MPC on non-linear and stochastic systems (see here) but is highly sensitive to the choice of regularization parameters $\theta_i$. DeePC-HUNT addresses this problem by automatically tuning these parameters. The performance of DeePC-HUNT has been validated on a rocket lander modelling the falcon 9, a noisy cartpole and a LTI system. To run these example notebooks, you can clone this directory and open it in a VS-Code environment with the Jupyter Notebook extension

Rocket - before training

Untrained-episode-0.mp4

Rocket - after training (episode 70)

After running DeePC-HUNT for 70 episodes, the controller now stabilizes the system.