We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Mathematics > Dynamical Systems

Title: Linearly-Recurrent Autoencoder Networks for Learning Dynamics

Abstract: This paper describes a method for learning low-dimensional approximations of nonlinear dynamical systems, based on neural-network approximations of the underlying Koopman operator. Extended Dynamic Mode Decomposition (EDMD) provides a useful data-driven approximation of the Koopman operator for analyzing dynamical systems. This paper addresses a fundamental problem associated with EDMD: a trade-off between representational capacity of the dictionary and over-fitting due to insufficient data. A new neural network architecture combining an autoencoder with linear recurrent dynamics in the encoded state is used to learn a low-dimensional and highly informative Koopman-invariant subspace of observables. A method is also presented for balanced model reduction of over-specified EDMD systems in feature space. Nonlinear reconstruction using partially linear multi-kernel regression aims to improve reconstruction accuracy from the low-dimensional state when the data has complex but intrinsically low-dimensional structure. The techniques demonstrate the ability to identify Koopman eigenfunctions of the unforced Duffing equation, create accurate low-dimensional models of an unstable cylinder wake flow, and make short-time predictions of the chaotic Kuramoto-Sivashinsky equation.
Comments: 37 pages, 16 figures
Subjects: Dynamical Systems (math.DS); Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1712.01378 [math.DS]
  (or arXiv:1712.01378v2 [math.DS] for this version)

Submission history

From: Samuel Otto [view email]
[v1] Mon, 4 Dec 2017 21:29:43 GMT (2101kb,D)
[v2] Tue, 15 Jan 2019 19:15:19 GMT (2527kb,D)

Link back to: arXiv, form interface, contact.