In a collaboration with Paul Monderkamp, Fabian Schwarzendahl, and Hartmut Löwen, we trained a microswimmer, in simulations, to navigate in random motility fields.
Paul A. Monderkamp, Fabian Jan Schwarzendahl, Michael A. Klatt, and Hartmut Löwen, Active particles using reinforcement learning to navigate in complex motility landscapes, Mach. Learn.: Sci. Technol. 3, 045024 (2022).
The “microswimmer” was an overdamped active Brownian particle with an additional ability to influence its orientation based on only limited, local information. The choices of the swimmer were trained using Q-learning, that is, a reinforcement learning technique, where the aim was to swim as quickly as possible across the observation window.
An important ingredient to success was to train the swimmer in a stealthy hyperuniform model. The homogeneity of this model allowed the swimmer to gradually improve its performance. Once the training was complete, the strategy immediately generalized to much more difficult models.
Michael Andreas Klatt 