École Nationale Supérieure de
Techniques Avancées

Abstract

Estimation of chaotic non-linear dynamical systems embedded in highly dimensional vector space lead to huge computational cost and large model errors if treated naively. The main objective of the present work is to present filters run in low dimension subspaces determined by different reduction techniques with different forward model during the predictive phase. Moreover, the underlying dynamics is not always known. The results show that efficient reductions are possible. The quality of the filter logically depends on the amount of information known about the "truth dynamics"