Reinforcement Learning and Wavelet Adapted Vortex Methods for Simulations of Self-propelled Swimmers

We present a numerical method for the simulation of collective hydrodynamics in self-propelled swimmers. Swimmers in a viscous incompressible flow are simulated with a remeshed vortex method coupled with Brinkman penalization and projection approach. The remeshed vortex methods are enhanced via wavelet based adaptivity in space and time. The method is validated on benchmark swimming problems. Furthermore the flow solver is integrated with a reinforcement learning algorithm, such that swimmers can learn to adapt their motion so as to optimally achieve a specified goal, such as fish schooling. The computational efficiency of the wavelet adapted remeshed vortex method is a key aspect for the effective coupling with learning algorithms. The suitability of this approach for the identification of swimming behaviors is assessed on a set of learning tasks.