Research on Model Order Reduction

My Ph.D. research at UC Berkeley focused on the problem of model order reduction (MOR), i.e., automated abstraction of “simpler” approximate models from high-dimensional models (usually derived from first principles).

If we view machine learning as learning a model from data, MOR is learning a (simple) model from a (complicated) model. While we can simulate the complicated model to obtain data and then apply machine learning to abstract a simple model, direct model abstraction can be faster and more accurate.

In particular, I focus on MOR of continuous-time continuous-value dynamical models, such as frac{d}{dt}x = f(x) + B u(t). But the ideas can be extended to discrete-time and/or discrete-value dynamical models. These models are widely used in modeling electronics, bio-chemical pathways, control systems, etc.

Practically, these abstract models can be used for fast simulation, validation and testing. The interpretable ones can also provide useful insights of how systems behave under input perturbation.


  • Chenjie Gu, Model Order Reduction of Nonlinear Dynamical Systems, University of California, Berkeley, 2011. [pdf] [bibtex]

Other Publications