The AEROMATIC team has carried out a study on machine learning (ML) closed-loop flow control of the wake of a cylinder using Linear Genetic Programming Control (LGPC) and Deep Reinforcement Learning (DRL).
In this work, our team has compared LGPC and DRL in the limit of a few sensors for feedback, focusing on noise, sensitivity to initial conditions and interpretability of control actions. We have observed that DRL has higher robustness with respect to variable initial conditions and to noise contamination of the sensor data.
On the other hand, we have also observed that the LGPC is able to identify compact and interpretable control laws, which only use a subset of sensors, thus allowing reducing the system complexity with reasonably good results.
All this work, which is now accessible through Arxiv, has been possible thanks to the collaboration of our team with the group of Bernd Noack, and it is coauthored by Rodrigo Castellanos, Guy Yoslan Cornejo Maceda, Ignacio de la Fuente Fernández, Bernd Noack, Andrea Ianiro, and Stefano Discetti, PI of this project.