Making (selected) model parameters time-dependent was proposed long ago
as a means for supporting the identification of causes of model
structure deficits. I will present recent results of implementing this
idea in a continuous-time framework that is based on describing the
time-dependent parameter(s) by a mean-reverting Ornstein-Uhlenbeck
process. I will present this approach in a notation that emphasizes the
similarty of underlying ideas with the approach we were discussing over
the past weeks at SAMSI (e.g. Bayarri et al. 2005). In some aspects,
our approach is a special case of that approach as it focuses on time
series output. On the other hand, it makes a step further in trying to
track back the cause of the bias to mechanisms in the model and/or
input error. This is intended to support model improvement and bias
reduction (instead of just bias description). I will try to present
preliminary results of the application of this approach to a simple
hydrologic model described in Kuczera et al. 2006.
My talk is based on Tomassini, Reichert, Kuensch, Buser and Borsuk 2006
(with a new application) which extends earlier work by Brun (and
Kuensch and Reichert) and Buser (and Kuensch).
The Papers by Bayarri
et al. and Tomassini
et al. are available on the home page of the methodology group. The
one by Kuczera can be provided on request (we only have the journal
version of the paper in press which we are not allowed to put on the
web).