Learning Nonlinear Dynamics from Data Using Sparse Optimization and Compressed Sensing

点击翻译按钮获取中文摘要

Since the beginning of scientific revolution, scientists have always been interested in learning sophisticated underlying structures from experimental and observational data. In the past, this process was done manually by experts in related fields. Due to the huge amount of data as well as its complicated behaviours, there are great demands in designing efficient algorithms and analyzing theoretical aspects of that learning problem. This proposal is focused on answering those important questions for learning dynamical systems from time-dependent data. Specifically, we plan to develop innovative numerical methods for learning nonlinear dynamics by combining advanced techniques from optimization and compressed sensing theory. ******The motivation of the proposed method is based on two main observations. Firstly, the form of the governing equations is rarely known a priori; however, based on the sparsity-of-effect principle, one may assume that the number of potential functions needed to represent the dynamics is very small. In practice, sparsity is promoted through the addition of an L1 term (or related quantity) as a constraint or penalty in the optimization model. While sparse optimization techniques have demonstrated their success in image and signal processing, information sciences, and others, their applications in dynamical systems is still limited. On the other hand, compressed sensing theory has provided solid theoretical results in reconstruction guarantees for general data. For time-dependent data coming from dynamical flows with complicated behaviours and additional restrictions, current theory need to be extended and studied carefully. Using sparse-inducing methods and results from random sampling theory, this proposal aims to develop sparse models and sampling strategies to recover the governing equations of nonlinear dynamics from time-dependent data as well as understanding the reconstruction guarantees for the related minimization problems. Preliminary results by the PI and collaborators show that in physical spaces of dimension three, it is possible to identify the underlying equations exactly from possibly highly corrupted data as the solution of an L1 minimization problem, provided that the flow is sufficiently ergodic. Based on those initial results, the PI will investigate further the effective combination of sparse learning for dynamical systems and reconstruction guarantees from compressed sensing in studying the dynamics from a wide range of data such as high-dimensional data, noisy data, and data from bifurcation diagram. This research will provide new perspectives from sparse optimization and compressed sensing in learning data structures. It can be applied to problems in weather predictions and atmospheric models, controls for fluid flows, aircraft development, and disease control models.