|
Titel |
Consensus on Long-Range Prediction by Adaptive Synchronization of Models |
VerfasserIn |
G. Duane, J. Tribbia, B. Kirtman |
Konferenz |
EGU General Assembly 2009
|
Medientyp |
Artikel
|
Sprache |
Englisch
|
Digitales Dokument |
PDF |
Erschienen |
In: GRA - Volume 11 (2009) |
Datensatznummer |
250030979
|
|
|
|
Zusammenfassung |
A collection of climate models typically gives divergent results for reasons that are not
obvious a priori. The differences seem buried in detailed dynamical choices, none clearly
superior, with effects that amplify over the course of a simulation. The chaotic dynamics of
the climate system is a well known obstacle to short-term weather prediction and to the
validation of forecast models. In predicting long-term climate change, one hopes that
different models will produce similar attractors defining overall climate, but the divergent
results of different climate models suggest similar difficulties. If one were able to combine
the different models - in a way better than by simply averaging their outputs - some advantage
might result.
A theoretical paradigm that has been applied to describe order in the climate, and seems
appropriate for model fusion, is that of the synchronization of loosely coupled chaotic
systems. Two or more chaotic systems, loosely coupled through only a few of many degrees
of freedom, fall into synchronized motion along their strange attractors under a surprisingly
wide variety of conditons, despite sensitivity to differences in initial conditions. The
phenomenon has been used to establish a new theoretical framework for data assimilation as
the synchronization of two systems, corresponding to “truth” and “model”, respectively.
Here, we suggest that a collection of models, synchronized by limited exchange of
information as they run, could form an “intelligent” consensus in regard to long-range
predictions, improving on the simple output-averaging procedures that have been used
previously.
The synchronous coupling approach succeeds because detailed
coupling coefficients can be chosen adaptively, as an instance of a general scheme for model
learning in the synchronization context. That scheme adjusts model parameters so as to
reduce synchronization error, with the possible addition of noise to escape local optima.
Here, the parameters to be adjusted are weights attached to each of many model
variables in each pairwise comparison with a corresponding variable of another
model. By training on 20th century data, weights will be selected so that the best
features of each model will maximally influence the collective behavior of the entire
suite.
We show that a small collection of Lorenz systems with different parameters can be fused in
this manner, so as to represent yet another “real” Lorenz system with a different
parameter set than that of any in the collection. The error in the representation is much
less than the error with any weighted combination of the Lorenz “models”, run
independently. Further, the Lorenz fused “model” continues to reproduce the behavior of the
real system after the adaptation process is turned off. We explore the limits of this
procedure as the parameters of the “real” system are changed after the training
period.
Then we show that the procedure can be extended so as to adaptively fuse two
quasigeostrophic channel models with different forcing terms so as to represent a third “real”
system. Results suggest that the procedure can naturally be extended to large climate
models.
The procedure can indeed be extended to form a multi-model of any time series produced by
a real physical system. As with the synchronization approach to data assimilation, it is argued
that the fusion approach is justified by contemporary theories of the role of synchronization in
neurobiological processing. |
|
|
|
|
|