Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 6: Machine Learning in Dynamics and Statistical Physics I
DY 6.2: Vortrag
Montag, 9. März 2026, 09:45–10:00, HÜL/S186
From Phase-Space Fluctuations to Predictive Power: Entropy Production as a Metric for Swarm Reservoir Computing — •Patrick Egenlauf1,2 and Miriam Klopotek1,3 — 1University of Stuttgart, Stuttgart Center for Simulation Science, SimTech Cluster of Excellence EXC 2075, Stuttgart, Germany — 2University of Stuttgart, Interchange Forum for Reflecting on Intelligent Systems, IRIS3D, Stuttgart, Germany — 3Heidelberger Akademie der Wissenschaften, WIN-Kolleg, Heidelberg, Germany
In reservoir computing, a time-varying input is projected onto a high-dimensional state space, allowing a simple linear readout to retrieve task-relevant features. Physical substrates such as active-matter swarms promise efficient, low-energy computation, but a quantitative selection criterion, that reliably indicates a good reservoir, is missing. We simulated an interacting swarm subjected to an external driver and evaluated two entropy measures: system entropy, quantifying phase-space density fluctuations, and environment entropy, representing heat dissipation. For each parameter set of the swarm interactions, we computed the relative differences for the system and environment entropy between undriven and driven cases and measured the driver work performed on the system. Both relative differences display robust linear correlations with forecast accuracy, while the driver work matches the performance curve almost perfectly, indicating that driver-induced entropy production dominates the reservoir’s information-processing capacity. Consequently, entropy production offers a quantitative metric for tuning swarm-based reservoirs toward optimal performance.
Keywords: Machine learning; Reservoir computing; Non-equilibrium; Dynamical systems; Entropy production