The study of hurricanes
Posted: Fri Nov 05, 2004 7:35 am
Dear all,
there is an article in the 16 September edition of the Economist about the
study of hurricanes (“Shelter from the storm“). In this article the author
writes: “…The difficulty of forecasting hurricane intensity is such that
one group has simply given up trying to model precisely what is happening.
Instead, Michael McGauley of the University of Miami uses so-called neural
networks to predict the evolution of intensity. A neural network is a
computer program that can be trained to look for correlations between
inputs (in this case the same set of meteorological data that goes into
detailed models of hurricanes) and outputs (in this case wind speeds). But
it is not really a model of what is going on, because the process of
correlation does not attempt to mimic the actual interactions inside a
storm. Nevertheless, Mr McGauley has found the technique effective. It
significantly outperforms conventional models in its predictions of
hurricane intensity both 24 and 48 hours into the future, although it is no
better if asked to look 72 hours ahead…”
Now, is this an area of application where the causality approach does not
(and cannot) provide satisfactory answers? Or haven’t they tried hard
enough yet? Or is it about “prediction” where System Dynamics, for example,
has never said it was meant for?
Thanks,
Regards,
Thomas Beck
From: Thomas_Beck@swissre.com
Swiss Re
there is an article in the 16 September edition of the Economist about the
study of hurricanes (“Shelter from the storm“). In this article the author
writes: “…The difficulty of forecasting hurricane intensity is such that
one group has simply given up trying to model precisely what is happening.
Instead, Michael McGauley of the University of Miami uses so-called neural
networks to predict the evolution of intensity. A neural network is a
computer program that can be trained to look for correlations between
inputs (in this case the same set of meteorological data that goes into
detailed models of hurricanes) and outputs (in this case wind speeds). But
it is not really a model of what is going on, because the process of
correlation does not attempt to mimic the actual interactions inside a
storm. Nevertheless, Mr McGauley has found the technique effective. It
significantly outperforms conventional models in its predictions of
hurricane intensity both 24 and 48 hours into the future, although it is no
better if asked to look 72 hours ahead…”
Now, is this an area of application where the causality approach does not
(and cannot) provide satisfactory answers? Or haven’t they tried hard
enough yet? Or is it about “prediction” where System Dynamics, for example,
has never said it was meant for?
Thanks,
Regards,
Thomas Beck
From: Thomas_Beck@swissre.com
Swiss Re