Data driven discovery of systems of ordinary differential equations using nonconvex multitask learning - Archive ouverte HAL Accéder directement au contenu
Article Dans Une Revue Machine Learning Année : 2023

Data driven discovery of systems of ordinary differential equations using nonconvex multitask learning

Résumé

In physical sciences, dynamic systems are modeled using their parameters within governing equations that often form a system of ordinary differential equations (SODE). This system consists of multiple equations, each of which relates the time derivative of a single parameter to several parameters. A parameter can appear in multiple equations, and this parameter potentially links the equations to each other. Although in certain cases the SODE can be written by domain experts, it is often unknown. With advances in sensor technology, large quantities of data can be sampled from dynamic systems, thus enabling the data-driven discovery of closed-form SODEs. State-of-the-art approaches are based on sparse single-task learning, which means that each equation from the SODE is learned independently. Omitting the coupling features of equations leads to SODEs that weakly identify the dynamic system. Furthermore, the convexity of the sparse penalty included in the learning criterion gives an SODE that is biased with respect to the true SODE. To reduce such a bias, we propose a multitask learning (MTL) based penalty which can learn the closed-form SODE with unbiasedness. The purpose of each task is to discover a single equation. But discovering an SODE is nontrivial, as dynamic systems are often nonlinear and the available data are noisy. Our proposal improves SODE identification by harnessing a nonconvex sparse matrix-structured penalty which takes into account the coupling feature as well as addresses the bias issue. Experimental results, based on noisy data simulated from known SODEs, confirm that, compared to single-task learning, MTL is more effective for recovering the closed-form SODE, and the proposed nonconvexity ensures that it can be estimated with unbiasedness. We also show the benefits of our approach on a realworld public dataset sampled from a laboratory-based ecological experiment. Keywords Sparse estimator • Unbiased sparsity • Multitask learning • Data-driven identification of dynamics • Multivariate time series Editors: Krzysztof Dembczynski and Emilie Devijver.
Fichier principal
Vignette du fichier
ECML_JournalTrack_Lejeune_authorversion.pdf (1.93 Mo) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)
Licence : CC BY ND - Paternité - Pas de modifications

Dates et versions

hal-04142117 , version 1 (11-07-2023)

Identifiants

Citer

Clément Lejeune, Josiane Mothe, Adil Soubki, Olivier Teste. Data driven discovery of systems of ordinary differential equations using nonconvex multitask learning. Machine Learning, 2023, 112 (5: Special Issue of the ECML PKDD 2022 Journal Track), pp.1523-1549. ⟨10.1007/s10994-023-06315-y⟩. ⟨hal-04142117⟩
60 Consultations
53 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More