Work

Reservoir computing with the minimum description length principle

Year: 2025

Type: article

Abstract: We use the minimum description length (MDL) principle, which is an information-theoretic criterion for model selection, to determine echo-state network readout subsets. We find that this method of MDL subset selection improves accuracy when forecasting the Lorenz, Rössler, and Thomas attractors. It also improves the performance benefit that occurs when higher-order terms are included in the readout layer. We provide an explanation for these improvements in terms of decreased linear dependence and improved consistency.

Source: Chaos An Interdisciplinary Journal of Nonlinear Science

DOI: 10.1063/5.0252938

Authors: Antony Mizzi, Michael Small, David M. Walker

Institutions: The University of Western Australia, Commonwealth Scientific and Industrial Research Organisation

Language: English

Topic: Neural Networks and Reservoir Computing

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funder: Australian Research Council

Awards: ARC Training Centre for Transforming Maintenance through Data Science, TSuNAMi: Time Series Network Animal Modelling