System Identification by Adaptive Boosting

Abstract: In the field of machine learning, the algorithm Adaptive Boosting has beensuccessfully applied to a wide range of regression and classification problems.Still, there is no known method to use the algorithm to estimate dynamical systems.In this thesis, the relationship between Adaptive Boosting and systemidentification is explored. A new identification method, inspired by AdaptiveBoosting, called TM-Boost is introduced. It fits a dynamical model byiteratively adding orthonormal basis functions. An interesting feature of themethod is that there is no need to specify a model order. It is also proven mathematicallyand verified in a series of identification experiments that TM-Boost,under reasonable conditions, converges to the true underlying system.