Deep reinforcement learning for the control of CPU-intensive environments

Jonathan Viquerat

The 2010’s decade has seen the fast-paced development of deep reinforcement learning (DRL) as a powerful approach for active control in many domains, including robotics or games. Yet, as the state-of-the-art methods now present a sufficient level of maturity, the application of such techniques in the context of numerically-resolved PDEs is still lagging behind. Here, we propose a short introduction to DRL in general, and a focus on the challenges that arise when applying such techniques to CPU-intensive environments, such as numerical PDE solvers. Then, illustrations from recent works on flow control are proposed.

J. Viquerat defended his PhD on discontinuous Galerkin time-domain methods for nanophotonics at INRIA Sophia-Antipolis in 2015, after what he spent 3 years there as a research engineer, mostly writing the core of the Diogenes time-domain solver.>He is now a research engineer in the CFL team at CEMEF-Mines ParisTech, working on DRL-based control for fluid flows, as well as on the development of the in-house CFD solver.