A Free Online VKI Seminar - Reliability of data-driven strategies in control and prediction of complex systems

A Free Online VKI Seminar
Reliability of data-driven strategies in control and prediction of complex systems.
Our Guest Speaker: Dr. Onofrio Semeraro, CNRS Research Associate at the Laboratoire Interdisciplinaire des Sciences du Numérique (LISN) - Universite Paris Saclay, Orsay (France)
Abstract
The breakthrough of machine-learning based applications is revolutionizing scientific computing; however, despite numerous successful examples -- mainly based on deep neural network -- the relative simplicity of applications of these tools is flawed by numerous issues that are too often overlooked such as generalizability of the models, lack of guarantees or case dependency. Increasing the size of the dataset of training, as well as relying on deeper and more expressive architectures do not necessarily guarantee a solution for these issues, while leading to longer training and thus to higher computational costs. The potential lack of robustness of these predictions is often mitigated by prior knowledge of physical constraints, when available. In this talk we will focus on two applications, by focusing on reliability and margins of guarantee.
First, we will consider Long-Short Term Memory neural networks and thoroughly investigate the impact of the training set, its structure and some issues relating the use of memory gates on the quality of the long-term prediction. Inspired by ergodic theory, we analyze the amount of data sufficient for a priori guaranteeing a faithful model of the physical system, relying on a proper dataset design. We show how an informed design of the training set, based on invariants of the system and the structure of the underlying attractor, significantly improves the resulting models, opening up avenues for research within the context of active learning. We show that this learning strategy corresponds at deploying a so-called curriculum, a technique often found in robotics.
In the second part of the seminar, we will focus on flow control, a subject that has attracted numerous research efforts in the last decades. Recently, Reinforcement Learning (RL), which encompasses a large variety of algorithms and strategies, has gained traction as an alternative protocol, thus circumventing the standard approaches involving a preliminary model reduction step. Indeed, RL algorithms do not require any a-priori knowledge of the equations governing the system to be controlled and solely rely on the local measurements of the flow, based on which a policy is learnt from the interaction of the agent with the environment. Despite successes, the first documented applications of RL for control in fluid dynamics often result in highly non-intuitive control policies, also when ``cheaper'' optimal solutions are available; we will discuss the impact of data selection on the convergence as well as an approach relying on optimistic policy iteration, where the margin of guarantees are explicitly found, with examples of fluid mechanics interest such as the Ginzburg-Landau system and the linearized boundary layer flows.
Biography
Onofrio Semeraro received his PhD in Mechanical Engineering at KTH-Stockholm (Sweden) in 2013. He served as postdoctoral researcher at Ecole-Polytechnique, Palaiseau (France) and Politecnico of Bari (Italy), and he is currently a CNRS Research Associate since 2017, at the Laboratoire Interdisciplinaire des Sciences du Numérique (LISN) - Universite Paris Saclay, Orsay (France). His studies focus mainly on control, data assimilation, modelling and data-driven techniques, ranging from system identification to deep learning for fluid mechanics. He is currently PI of an ANR project dedicated to optimal control and Reinforcement Learning for control of fluids, and contributors for projects at the intersection of machine learning, dynamical systems and fluid dynamics.
Event Information
Event Date | 09-02-2024 9:00 am |
Event End Date | 09-02-2024 10:00 am |
Cut off date | 08-02-2024 2:00 pm |
Location | von Karman Institute for Fluid Dynamics |
Venue Information - von Karman Institute for Fluid Dynamics
