Active flow control of the flow past a cylinder under Reynolds number variation using deep reinforcement learning This study is published in German! In a 2D cylinder flow, a Kármán vortex street is created in the wake of the cylinder, which causes oscillating forces on the cylinder. An agent trained with a deep reinforcement learning algorithm is designed to reduce these forces. To achieve this, the agent rotates the cylinder based on the pressure values on the cylinder surface and thus actively regulates the flow. This algorithm uses a probability distribution, which is often a normal distribution, to select the rotation speed. However, other probability distributions can also be chosen. In this case, the difference in training speed and result between the normal distribution, a normal distribution to which a tangent hyperbolic function was applied, and the beta distribution was investigated. The beta distribution achieves the best results and can achieve the agent’s reward with the normal distribution after only 13 episodes, compared to 86 episodes. In addition, the beta distribution can increase the reduction of the average drag by 0.43 percentage points to 5.48% compared to the normal distribution. The agent using the tangent hyperbolic function was also able to achieve better results than the unmodified normal distribution, reducing the average drag by 0.1 percentage point to 5.15%. An agent with a beta distribution was then trained on 3 different but constant Reynolds numbers and applied to a time-varying flow. The maximum drag was reduced by 14.43 %. At times, the formation of the Kármán vortex street could even be completely prevented. Meanwhile, the average drag coefficient was reduced by 36.21 %.