Eventos Anais de eventos
ENCIT 2020
18th Brazilian Congress of Thermal Sciences and Engineering
ADDRESSING OVERFITTING ISSUES IN THE SPARSE IDENTIFICATION OF NONLINEAR DYNAMICAL SYSTEMS
Submission Author:
Leonardo Santos de Brito Alves , RJ , Brazil
Co-Authors:
Leonardo Santos de Brito Alves
Presenter: Leonardo Santos de Brito Alves
doi://10.26678/ABCM.ENCIT2020.CIT20-0646
Abstract
Over the past four years, the derivation of dynamical models using symbolic regression has become a fixture in machine learning due to the development of the SINDy tool for the Sparse Identification of Nonlinear Dynamical systems. As far as the author is aware, this tool has been applied assuming a polynomial representation of the unknown model that uses a monomial basis including only up to cubic nonlinearities. In the present paper, this issue is further explored. It turns out the library of candidate functions becomes ill-conditioned as the maximum nonlinearity order is increased, preventing the LASSO regularization in SINDy from eliminating unphysical terms due to increased error propagation.
Keywords
machine learning, symbolic regression, SINDy, Candidate Functions, Ill-Conditioned Library
DOWNLOAD PDF VIEW PRESENTATION
