Jean Rabault - Recent advances in Deep Reinforcement Learning applied to Active Flow Control

Title: Recent advances in Deep Reinforcement Learning applied to Active Flow Control, by Jean Rabault

Abstract: Active Flow Control (AFC) is one of the many illustrations of the
challenges presented by 'complex', i.e., non-linear, high-dimensional,
time-dependent systems. While AFC is a topic of both industrial and
academic interest, and has been the focus of considerable research
efforts over several decades, few (if any) real-world or laboratory
achievements have been attained. One reason for this apparent lack of
tangible results lies in the difficulty of applying classical
optimization methods, which are often local and / or rely on
regularity properties of the underlying problem, to tackle 'complex'
systems. Indeed, if a system is 'complex' enough, any method based on
an analysis of its local properties may have very much difficulties to
understand its behavior away from the initial position in the phase
space. In the most extreme 'complexity' scenario, the only method for
truly discovering a complex system, and, therefore, understanding and
controlling it, may be to explicitly explore (and simultaneously map /
learn) its phase space through a trial-and-error approach.

Fortunately, considerable progress has recently taken place in the
Machine Learning (ML) field regarding the development of methods
aiming at learning the dynamics of a system through a trial-and-error
approach. The most striking illustration of these ML advances was
arguably the defeat of the best human player at the game of Go in
2016, a feat that was widely thought to be at least decades away.
Following this success, Deep Reinforcement Learning (DRL) has
attracted considerable interest from research communities in robotics,
optimal control, and, recently, Fluid Mechanics (FM), among others.

In this presentation, I will present a quick reminder of the main
principles and algorithms within DRL, before discussing a series of
recent results this method achieved in Fluid Mechanics, including the
control of laminar, and turbulent, 2D flows, as well as the control of
the Lorentz attractor, and of the Rayleigh-Benard instability.
Finally, I will discuss what results can be expected to be reached in
the next months and years, and suggest a few promising real-world
applications.
 

Publisert 10. juni 2020 13:35 - Sist endret 19. juni 2020 11:20