Bayesian error isolation for models of large-scale systems

A methodology is presented for use in isolating sources of misspecification in system models that are known to be invalid. The methodology relies on a technique based on stochastic approximation in the context of a Bayesian formulation. This approach has significant advantages in computational efficiency, relative to a straightforward Bayesian analysis, for large-scale systems. Moreover, it applies to arbitrary model forms (e.g. state-space, regression, etc.) and applies when the probability distribution for the system output is not necessarily Gaussian.<
>