Berlin 2024 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

BP: Fachverband Biologische Physik

BP 20: Poster IIIa

BP 20.14: Poster

Wednesday, March 20, 2024, 11:00–14:30, Poster B

Reinforcement Learning in Active Colloidal Reservoir Computing — •Jonas Scheunemann, Samuel Tovey, and Christian Holm — Institute for Computational Physics, Stuttgart, Germany

The capacity to process information through a physical system can be exploited and further understood by using the recently introduced framework of physical reservoir computing. The concept involves utilizing the dynamics of nonlinear physical systems for time series forecasting, speech recognition, or classification tasks. The characteristics of an effective reservoir are still under discussion, and we use reinforcement learning to delve deeper into this question. Our reservoir substrate consists of a swarm system of active matter colloids, which has recently been demonstrated to work using a modified Reynolds boids model. We train the swarm by rewarding the colloidal agents through a concentration field approach, inspired by the behaviour of E. coli, and tested by forecasting a chaotic time series with a Lorenz attractor input. As the swarm reservoir's memory depends on the colloids' correlation time, we employ the Langevin equation to set up the system at diverse temperatures. We identify a potential connection between temperature and prediction accuracy, opening up research on the advantages of temperature-induced noise in the reservoir.

Keywords: Reinforcement Learning; Active Matter; Reservoir Computing