CDC2023: Nonlinear Youla-REN

This repository contains the code for our paper Learning Over Contracting and Lipschitz Closed-Loops for Partially-Observed Nonlinear Systems accepted at CDC 2023. The code is built on the RobustNeuralNetworks.jl package which implements the REN models.

This code has been tested with juilia v1.7.3 and RobustNeuralNetworks.jl version v0.1.0.

Installation

Clone this git repository and start the Julia REPL within the project root directory.

git clone https://github.com/nic-barbara/CDC2023-YoulaREN.git
cd CDC2023-YoulaREN

Start a Julia session, then activate the repository and install dependencies

using Pkg
Pkg.activate(".")
Pkg.instantiate()

This project depends on a number of larger packages (eg: Flux.jl) and an older version of Julia (v1.7.3) so installation may take a few minutes.

Usage

The main scripts used to log experimental data are:

• src/MagLev/mag_experiment.jl to train models on magnetic suspension
• src/QubeServo/qube_experiment.jl to train models on the rotary-arm pendulum
• Within src/Robustness/, run both mag_adversarial.jl and mag_ecrit_save.jl to generate robustness results on magnetic suspension
• Within src/Robustness/, run both qube_adversarial.jl and qube_ecrit_save.jl to generate robustness results on the rotary-arm pendulum

The main scripts used to visualise results are:

• src/MagLev/mag_plot_results.jl to reproduce Fig. 3a in the paper
• src/QubeServo/qube_plot_results.jl to reproduce Fig. 3b in the paper
• src/Robustness/mag_plot_robustness.jl to reproduce Fig. 4a in the paper
• src/Robustness/mag_plot_robustness.jl to reproduce Fig. 4b in the paper

Contact

For any questions, please contact Nicholas Barbara (nicholas.barbara@sydney.edu.au)