martindoff / dc-rnn

Learn dynamical systems as a difference of convex functions (DC) using a Recurrent Neural Network (RNN) architecture with DC structure. Example on the coupled tank problem.

Recurrent Neural Network (RNN) model with difference-of-convex-functions (DC) structure.

Table of Contents

1. About The Project
   • Built With
2. Getting Started
   • Prerequisites
   • Running the code
   • Options
3. Application
4. Contact

About The Project

Learn dynamical systems from data using Recurrent Neural Network (RNN) models with a special difference-of-convex-functions (DC) structure. The so-called DC-RNN model approximates the system trajectories by a DC input-output map between the control input sequence and state trajectory. The DC-RNN model is leveraged in this example to predict the trajectory of the
coupled tank dynamical system.

Built With

• Python 3
• Keras

Getting Started

Prerequisites

You need to install the following:

• numpy
• matplotlib
• tensorflow / keras

Run the following command to install all modules at once

pip3 install numpy matplotlib tensorflow

Running the code

1. Clone the repository

   git clone https://github.com/martindoff/DC-RNN.git

2. Go to directory

   cd DC-RNN

3. (Optional) Generate data from the dynamical system

   python3 generate_data.py

   The data are stored in two files: input.npy and output.npy

4. Run the program

   python3 DC_RNN_model.py

Options

1. To load an existing model, set the load variable in DC_RNN_model.py to True

   load = True

Set the variable to False if the model has to be (re)trained.

Application

Such model have applications, e.g. in the framework of robust tube MPC for systems representable as a difference of convex functions (see paper on DC-TMPC) The DC-RNN model allows one to learn the dynamics in DC form and then apply the DC-TMPC algorithm.

Contact

Martin Doff-Sotta - [email protected]

Linkedin: https://www.linkedin.com/in/mdoffsotta/