Data-Driven Modeling and Forecast of Noisy Nonlinear Dynamics

• Activity Number: 347 - Machine Learning and Applications in Complex Engineering Systems
• Type: Contributed
• Date/Time: Tuesday, July 31, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Kyongmin Yeo* and Youngdeok Hwang and Eun Kyung Lee
• Companies: IBM T.J. Watson Research Center and Sungkyunkwan University and IBM T.J. Watson Research Center
• Keywords: time series; nonlinear dynamical system; stochastic process; recurrent neural network; uncertainty quntification

Abstract

Data-driven modeling of a complex physical process is of current interest due to its direct relevance to cognitive manufacturing. Although domain knowledge on the underlying physical process has been playing a key role in understanding manufacturing processes, it is often impractical or impossible to build physics models from the first principles. Here, we propose a Recurrent Neural Network (RNN) based model for the nonlinear system identification and forecast of a stochastic process with an underlying physical process. The proposed RNN aims to directly estimate the probability distribution of the stochastic process by using a penalized maximum log likelihood method. It is shown that the RNN is essentially a state-space model, in which the underlying process is modeled by a set of linear dynamical systems. Also presented is a Monte Carlo procedure for a multiple-step prediction.