HR: 0800h
AN: IN41A-0309 [Abstracts]
TI: Ensemble Smoothing and Markov Chain Monte Carlo for Data Assimilation in Highly Nonlinear
Systems
AU: * Turmon, M
EM: turmon@jpl.nasa.gov
AF: Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, CA 91109
United States
AU: Chin, T M
EM: tmchin@jpl.nasa.gov
AF: Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, CA 91109
United States
AU: Jewell, J B
EM: jewell@jpl.nasa.gov
AF: Jet Propulsion Laboratory, 4800 Oak Grove Drive, Pasadena, CA 91109
United States
AU: Ghil, M
EM: ghil@atmos.ucla.edu
AF: Dept. of Atmos. and Oceanic Physics, University of California, Los Angeles, Los Angeles, CA 90095
United States
AB:
Current methods for atmosphere and ocean data assimilation propagate
Gaussian distributions for gridded state variables forward in time.
Powerful as these methods are, they do not handle outliers well and cannot
simultaneously entertain multiple hypotheses about system state. The
alternative of propagating the system's full probability distribution is
burdensome, and ensemble methods have been introduced into data
assimilation for nonlinear systems to get around this problem. By
propagating an ensemble of representative states, algorithms like the
Ensemble Kalman Filter (EnKF) and the Resampled Particle Filter (RPF) rely
on existing modeling infrastructure and capture the weights to be assigned
to the data based on the evolution of this ensemble. We present an
ensemble-based smoother that is applicable to Monte Carlo filtering schemes
like the EnKF and the RPF. At the minor cost of retrospectively updating a
set of weights for ensemble members, this smoother provides superior state
tracking for two simple nonlinear problems, the double-well potential and
the trivariate Lorenz system. The algorithm does not require retrospective
adaptation of the ensemble members themselves, and is thus suited to a
streaming operational mode. The accuracy of the proposed backward-update
scheme in estimating non-Gaussian distributions is evaluated by comparison
of its posterior distributions with ground truth provided by a Markov chain Monte Carlo algorithm.
DE: 0555 Neural networks, fuzzy logic, machine learning
DE: 3245 Probabilistic forecasting (3238)
DE: 3315 Data assimilation
DE: 4445 Nonlinear differential equations
SC: Earth and Space Science Informatics [IN]
MN: Fall Meeting 2005