Parameter optimization of open-loop control of a circular cylinder by simplified reinforcement learning

Abstract

Open-loop control is commonly considered an efficient approach in flow control, in which the search for control parameters with excellent performance is mostly carried out by grid search, leading to an extremely tedious process of parameter optimization. With extensive applications of machine learning, reinforcement learning (RL) has emerged as a powerful tool to achieve optimal strategies, which constructively leads to the result that parameter optimization can be performed by RL. In this paper, we provide the concept of simplified RL formally and show the corresponding properties. In order to implement simplified RL for flow control, a high-order numerical approach is coupled with simplified RL to develop a new framework for parameter optimization and determination. In order to validate the performance of the framework, flows past a rotary oscillating circular cylinder at low Reynolds number Re = 200 (defined as Re=U∞D/ν, where U∞ is the free-stream velocity and ν is the kinematic viscosity) are investigated by varying the parameters of rotary amplitude and frequency individually or simultaneously. By numerical investigations, a satisfactory drag reduction effect is achieved, which demonstrates the capability of the framework to perform parameter optimization in terms of open-loop control.