Ewerton Rocha Vieira, Rutgers University

Abstract

Models for evolutionary processes like physical systems are conceptualized via continuous  dynamical systems. However, in general, the model is unknown and only finite data is observed, hence, it is a significant challenge to learn a continuous function that describes the desired dynamical system robustly. To address this, data-driven dynamics are being increasingly employed in order to identify the underlying dynamics based on finite data.

To overcome the gap between the complexity of a continuous system and the  description based on finite data, we use a Gaussian process as a surrogate model together with combinatorial dynamics to capture the global behavior of the underlying function that generates the dynamics. After introducing the main ideas, I will show some applications to robotic systems.

Video