Dresden 2026 – scientific programme

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 5: Poster

AKPIK 5.2: Poster

Thursday, March 12, 2026, 15:00–16:30, P5

Physics-based Reinforcement Learning for Balancing the Cart-Pole — •Igor Polonskiy, Atreya Majumdar, and Karin Everschor-Sitte — Faculty of Physics and Center for Nanointegration Duisburg-Essen (CENIDE), University of Duisburg-Essen, 47057 Duisburg, Germany

Balancing a pole on a moving cart by applying lateral forces is a standard benchmark problem in reinforcement learning. Deep Q-Networks [1], which integrate reinforcement learning with neural networks, have been highly effective in solving this problem. Training the multiple hidden layers of Deep Q-Networks, however, is computationally expensive and thereby energy-demanding. Replacing these hidden layers with an Echo State Network reduces training costs while maintaining performance [2]. Echo State Networks have been shown to be replaceable by physical systems [3]. We explore the potential of solving the Cart-Pole problem with a physics-based Echo State Network.

[1] V. Mnih et al., Nature 518, 529 (2015)

[2] I. Polonskiy, Bachelor Thesis, University of Duisburg-Essen (11/2024)

[3] K. Everschor-Sitte et al., Nature Reviews Physics 6, 455 (2024)

Keywords: Machine Learning; Reinforcement Learning; Unconventional Computing; Physical Reservoir Computing