Regensburg 2022 – wissenschaftliches Programm

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

DY: Fachverband Dynamik und Statistische Physik

DY 51: Machine Learning in Dynamics and Statistical Physics (joint session DY/SOE)

DY 51.2: Vortrag

Freitag, 9. September 2022, 10:15–10:30, H19

Deep reinforcement learning for chemotactic active particles — •Edwin Loran, Mahdi Nasiri, and Benno Liebchen — Institute of Condensed Matter Physics, Technische Universität Darmstadt, D-64289 Darmstadt, Germany

Throughout evolution, microorganisms have developed efficient strategies for locating nutrients and avoiding toxins in complex environments. Understanding their adaptive policies can provide new key insights for the development of smart artificial active particles. Here, we use a machine learning approach, namely deep reinforcement learning, to develop smart foraging strategies for chemotactic active particles which consume nutrients for their survival. Our method is able to devise efficient chemotactic navigation strategies guaranteeing "survival" inside unknown and complex landscapes while only having access to local sensory data. Our approach is based on deep Q-learning and uses the particle's observation of its surrounding chemical (nutrient) concentration as the input. The presented method highlights the extent of the capabilities of reinforcement learning approaches in mimicking (and going beyond) the evolutionary strategies learned by microorganisms.