ODYS Deep Learning

Learn prediction models from data

Model predictive control (MPC) requires a mathematical representation of the dynamics of the process to compute predictions. Physics-based nonlinear models are sometimes too complex to develop and maintain, and may lead to difficult nonlinear programming problems to solve online when evaluating the MPC controller. In alternative, data-driven models can be used to directly learn the prediction model (or parts of it) from experimental data.

Deep learning for MPC

ODYS Deep Learning is a software library, available in MATLAB® and in Python, conceived to help the development of black-box models from data for use with ODYS Embedded MPC. It offers the following main features:

  • Very efficient proprietary batch algorithms based on regularized nonlinear least-squares for learning deep neural networks, using most common activation functions, arbitrary differentiable output functions, and arbitrary differentiable loss functions, and recursive algorithms for online learning and adaptive MPC. Compared to most common deep-learning libraries, our algorithms are particularly efficient in learning small-scale neural networks, which are typically used in MPC, in terms of both quality of fit and computation time.
  • Interfaces with most common machine-learning libraries in Python (keras/tensorflow, sklearn, pytorch).
  • Automatic generation of optimized C code of neural network models and their Jacobian matrices for nonlinear MPC setup. The generated code is ready to be use as prediction model in ODYS Embedded MPC.
  • A black-box nonlinear system identification library to learn neural multivariable autoregressive models from experimental data.

Please contact us at info@odys.it for more information about ODYS Deep Learning.