Parts | Days | Selection | Search | Updates | Downloads | Help
BP: Fachverband Biologische Physik
BP 35: Active Matter V (joint session BP/DY)
BP 35.5: Talk
Friday, March 22, 2024, 10:30–10:45, H 1028
SwarmRL: Lowering the entry barrier to reinforcement learning for active matter research — •Samuel Tovey, Christoph Lohrmann, and Christian Holm — Institute for Computational Physics, University of Stuttgart, Stuttgart, Germany
As scientists learn to better design and control devices at a microscopic scale, so too must the tools used to control these devices develop. Multi-agent reinforcement learning (MARL) is a powerful machine learning paradigm for learning control strategies in agents at all scales. Recent work has applied MARL to controlling microscopic agents, whether in learning chemo-taxis behaviour, object manipulation, or swarming.
This talk introduces SwarmRL, a powerful open-source library for applying MARL to microscopic environments. We demonstrate how SwarmRL is used in our group to control micro-scale agents in simulation and experiments and how to interpret the learned policies. The talk introduces the library broadly before looking into results from our recent work using SwarmRL, including a better understanding of the role of temperature on learned strategy and the emergence of chemotactic behaviour in unstable regimes. Finally, we discuss our vision for the future of the library and its integration into experiments and simulations.
Keywords: Active matter control; Deep multi agent reinforcement learning; Microrobotics; Bio-inspired robotics; Emergent collective behaviour