Experimental Reinforcement Learning for Control of Aerodynamic Systems

Systems such as unmanned aerial vehicles (UAVs) and wind turbines can be severely damaged by forces resulting from large atmospheric flow disturbances (e.g. gusts). Machine learning (ML) methods present a possible solution for mitigating these potentially disastrous situations.  Informed by non-intrusive flow measurements in real-time, ML may be used to estimate incoming flow conditions and identify appropriate control strategies for these applications. Reinforcement learning (RL) is a type of ML in which an “agent” may be trained through trial-and-error goal-directed search to identify superior control strategies of dynamical system. A large-scale experimental application of RL is presented, in which a system has been trained to control aerodynamic forces resulting from an unsteady flow. The performance of these methods is analyzed in the context of real-world flow control applications.