Implicit Estimation of Ecological Model Parameters

• DOI: 10.1007/s11538-012-9801-6
• Publication Date: 2013-01-04T16:45:56Z
• Authors: Brad Weir; Robert N. Miller; Yvette H. Spitz
• Keywords: 0301 basic medicine; Stochastic Processes; 03 medical and health sciences; Models, Biological; Monte Carlo Method; Algorithms; Ecosystem

We introduce an implicit method for state and parameter estimation and apply it to a stochastic ecological model. The method uses an ensemble of particles to approximate the distribution of model solutions and parameters conditioned on noisy observations of the state. For each particle, it first determines likely values based on the observations, then samples around those values. This approach has a strong theoretical foundation, applies to nonlinear models and non-Gaussian distributions, and can estimate any number of model parameters, initial conditions, and model error covariances. The method is called implicit because it updates the particles without forming a predictive distribution of forward model integrations. As a point of comparison for different assimilation techniques, we consider examples in which one or more bifurcations separate the true parameter from its initial approximation. The implicit estimator is asymptotically unbiased, has a root-mean-squared error comparable to or less than the other methods, and is accurate even with small ensemble sizes.