Regensburg 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 39: Machine Learning in Dynamics and Statistical Physics II

DY 39.5: Vortrag

Donnerstag, 20. März 2025, 16:00–16:15, H47

Reinforcement learning for autonomous navigation of active particles in complex flow fields — •Diptabrata Paul and Frank Cichos — Peter Debye Institute for Soft Matter Physics, Universität Leipzig, 04103 Leipzig, Germany

Sensing and feedback on environmental stimuli are integral to regulating diverse functions in living systems, ranging from sub-cellular processes to evolution of navigation strategies such as chemotaxis and phototaxis. Unlike living systems, noisy artificial microswimmers have limited ability to adapt to various stationary and dynamic environmental perturbations to yield optimized behaviour for a given task. Consequently, reacting to such environmental cues becomes indispensable for achieving effective navigation and control in complex and noisy settings. In this context, we explore incorporation of machine learning algorithm for autonomous decision making for navigation of an active microswimmer within noisy environments. While naive navigation policies yield inefficient and ineffective solutions under changing conditions, employing actor-critic reinforcement learning (RL) framework trained in experiments leads us to quasi-optimal policies that are capable of navigating, even in presence of complex flow fields. Our study exhibits that a model trained under noisy conditions successfully learns effective navigation policies and are robust with respect to environmental perturbations such as hydrodynamic flow fields as well as varying initial conditions. This work paves the way for development of online RL for modelling adaptive behaviour and navigation of active microswimmers in complex fluidic scenarios.

Keywords: Reinforcement Learning; Microswimmers