Abstract
In this study, we introduce and consider a shape optimization process using deep reinforcement learning for the drag minimization of an object in an incompressible viscous fluid. Most previous studies on shape optimization have used the adjoint variable method. However, it has been demonstrated that shape optimization results are dependent on the initial shape and that the final shape is not necessarily the optimal shape1). Deep reinforcement learning is a method that learns only from the interactions between an agent and its environment. We believe that this is appropriate for optimization where the correct answer is unknown. Furthermore, if the shape design variables increase, the behavior patterns that an agent can consider will significantly increase. Therefore, we believe that deep learning is more suitable for this purpose.
