Data-driven Learning and Control - DDLC

Using stored examples (templates) to solve complex dynamic tasks has appeared in many areas of engineering. This approach allows for fast inference without model reconstruction. However, a general theory of control, estimation and prediction from templates is still largely undeveloped. 

In this talk, we will focus on predicting the evolution of unknown dynamical systems from observed past data windows and templates. Specifically, we first construct a library of prototypical trajectories (templates), generated by different systems. We then combine these templates to build a new repertoire of sample trajectories that can successfully predict the future even when the observed past windows do not belong to any system in the original library. This demonstrates a form of generalization across dynamical behaviors.