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

DY: Fachverband Dynamik und Statistische Physik

DY 23: Stochastic Thermodynamics

DY 23.8: Vortrag

Mittwoch, 20. März 2024, 11:30–11:45, BH-N 128

Thermodynamically efficient agents — •Paul C. Barth, Lukas J. Fiderer, Isaac D. Smith, Marius Krumm, Fulvio Flamini, and Hans J. Briegel — Institute for Theoretical Physics, University of Innsbruck, 6020 Innsbruck, Austria

Landauer’s bound and its generalizations such as the information-processing second law provide energetic limits not only on information erasure but also on pattern manipulation and computations in general. In this work, we generalize these bounds further to account for agents which interact with an environment via a percept-action loop. Within our framework, we then design and analyze toy environments and thereby demonstrate that efficient agents, which maximize work production, do not always adhere to zero entropy production. Furthermore, we introduce a thermodynamic framework for learning by leveraging similarities between maximum work extraction and reward maximization in reinforcement learning. We apply this framework to an existing reinforcement learning scheme, namely projective simulation. We also discuss a possible quantization of our framework. This line of research promises new insights into the energetic aspects of adaptive behavior in natural and artificial system.

Keywords: Landauer's bound; second law; computational mechanics; reinforcement learning; projective simulation