TY - GEN

T1 - A novel Gaussian sum filter method for accurate solution to the nonlinear filtering problem

AU - Terejanu, Gabriel

AU - Singla, Puneet

AU - Singh, Tarunraj

AU - Scott, Peter D.

N1 - Copyright:
Copyright 2008 Elsevier B.V., All rights reserved.

PY - 2008

Y1 - 2008

N2 - The paper presents two methods of updating the weights of a Gaussian mixture to account for the density propagation within a data assimilation setting. The evolution of the first two moments of the Gaussian components is given by the linearized model of the system. When observations are available, both the moments and the weights are updated to obtain a better approximation to the a posteriori probability density function. This can be done through a classical Gaussian Sum Filter. When the measurement model offers little or no information in updating the states of the system, better estimates may be obtained by updating the weights of the mixands to account for the propagation effect on the probability density function. The update of the forecast weights proves to be important in pure forecast settings, when the frequency of the measurements is low, when the uncertainty of the measurements is large or the measurement model is ambiguous making the system unobservable. Updating the weights not only provides us with better estimates but also with a more accurate probability density function. The numerical results show that updating the weights in the propagation step not only gives better estimates between the observations but also gives superior performance for systems when the measurements are ambiguous.

AB - The paper presents two methods of updating the weights of a Gaussian mixture to account for the density propagation within a data assimilation setting. The evolution of the first two moments of the Gaussian components is given by the linearized model of the system. When observations are available, both the moments and the weights are updated to obtain a better approximation to the a posteriori probability density function. This can be done through a classical Gaussian Sum Filter. When the measurement model offers little or no information in updating the states of the system, better estimates may be obtained by updating the weights of the mixands to account for the propagation effect on the probability density function. The update of the forecast weights proves to be important in pure forecast settings, when the frequency of the measurements is low, when the uncertainty of the measurements is large or the measurement model is ambiguous making the system unobservable. Updating the weights not only provides us with better estimates but also with a more accurate probability density function. The numerical results show that updating the weights in the propagation step not only gives better estimates between the observations but also gives superior performance for systems when the measurements are ambiguous.

UR - http://www.scopus.com/inward/record.url?scp=56749095344&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=56749095344&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:56749095344

SN - 9783000248832

T3 - Proceedings of the 11th International Conference on Information Fusion, FUSION 2008

BT - Proceedings of the 11th International Conference on Information Fusion, FUSION 2008

T2 - 11th International Conference on Information Fusion, FUSION 2008

Y2 - 30 June 2008 through 3 July 2008

ER -