The study of hurricanes

This forum contains all archives from the SD Mailing list (go to http://www.systemdynamics.org/forum/ for more information). This is here as a read-only resource, please post any SD related questions to the SD Discussion forum.
Locked
Thomas_Beck@swissre.com
Newbie
Posts: 1
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Thomas_Beck@swissre.com »

Alex leus
Junior Member
Posts: 6
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Alex leus »

Mr. Beck was asking about the application of neural networks to hurricanes. I would like to leverage off that thought and ask has anyone investigated Dr. Wolfram's claims in his book A New Kind of Science where basic simple rules will replace complex analysis, such as what is applied to understand hurricanes, and then dropped because of the complexity. When you read his book (ANKOS), he appears to be claiming that his approach is better than what is now available in the mathematical and science areas etc. In addition, are there basic simple rules for doing system dynamics? Thank you in advance.

Cheers,
Alex Leus
leusa@tds.net
Gary Hirsch
Junior Member
Posts: 2
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Gary Hirsch »

Thomas,

I don't know enough about meteorology and the physics of hurricanes to be
certain, but I'm not sure that an SD model of hurricanes would add anything
to the science of predicting their paths and intensity. It might be
possible to develop a good SD model of hurricanes for teaching students
about them, but I wouldn't try using such a model for real-world prediction.

On the other hand, I think SD is an excellent approach for modeling
VULNERABILITY to hurricanes which is really the important management issue.
A number of years ago, I was involved in a US-Canadian study of the
fluctuating levels of the Great Lakes and the economic, environmental, and
other consequences of these fluctuations. Here the fluctuations are quite
predictable if you look at a long enough historical record. At least you
can predict that there will be periods of high and low levels even if you
can't predict the exact timing. Yet vulnerability to fluctuations continued
to increase. When levels were low for a few years, people would build
closer to the water and then be ""surprised"" when levels predictably rose.
They would demand expensive and environmentally damaging mitigation measures
such as seawalls as well as insurance payouts for any damage done. When
levels were high for a while, they would build marinas and then be surprised
when water levels dropped and then demand expensive dredging to keep the
marinas viable. Local governments whose principal concerns seemed
short-term economic development (e.g., the jobs created by the new marina)
became co-conspirators in this process.

I suspect there are similar forces at work in creating vulnerability to
hurricanes. Short memories with regard to damage done by past hurricanes
allow people to build in places where the historical record tells us they
are vulnerable to future storms. While insurers such as Swiss Re cannot
manage the paths of hurricanes, they can certainly manage vulnerability to
them through premium rates for properties in different types of locations
and other rules associated with property and casualty policies.

Gary Hirsch
GBHirsch@comcast.net
Joel Rahn
Junior Member
Posts: 15
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Joel Rahn »

Wolfram's approach exploits the properties of cellular automata first
made famous (early 70s) by Conway's Game of Life in which local rules
for creating/destroying cells give rise to intriguing patterns on a
monitor when the cells are represented as pixels that are lit or dark
according to the state of the cell. Wolfram's claims have not exactly
taken the world of mathematical modelling by storm. The problem is one
of design: how to choose the rules to get a specific behaviour.
Agent-based modeling is essentially the same approach but less abstract.

>In addition, are there basic simple rules for doing system dynamics?

In answer to your last question: No, not if you want to _do_ significant
SD work, and Yes if you want to make toy models. The 'rules' or rather,
the guiding principles are described and illustrated in most SD software
packages like Vensim. A feel for the real effort needed to _do_ SD can
be picked up by working through Sterman's Business Dynamics.

Joel RAhn
jrahn@sympatico.ca
Joel Rahn
Junior Member
Posts: 15
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Joel Rahn »

Perhaps the question we should be asking ourselves on this list is
whether a neural network model of an SD model could tell us something
about the important causes of the model's behaviour, and if so, is it
more efficient to use a neural network instead of the dominant-loop or
dominant-mode approaches that have been developed over the last 15 years
or so?

Joel Rahn
jrahn@sympatico.ca
Alan Graham
Member
Posts: 27
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Alan Graham »

On Hurricanes and SD:

In the context of wondering which systems or problems System Dynamics is
good for, I'd characterize those systems as the systems in which we can
identify plausible aggregate quantities (those that will tend to behave as a
single group, and thus characterizable by one or a small vector of
quantities), and plausible causal relationships between them that will
determine overall behavior. In other words, aggregated, prior-rich and
perhaps data-poor. I know this is only a slightly more detailed version of
""We can model with SD systems that SD describes"".

When this isn't true, other methods are likely better. ""Black box"" methods
such as neural nets or statistical autocorrelation models don't deal with
internal causality, but just relations between inputs and outputs. There
has to be a major body of data for these techniques to be effective--in
other words, they're good for prior-poor and data-rich problems.

Hurricanes fall in a funny gray area: We do have good priors on how
elements interact, but the elements don't aggregate well. In lieu of a
small number of variables, meteorologists create very basic variables
(pressure, velocity, moisture content of a given cubic kilometer of air
space (or whatever), then repeat thousands or millions of times to cover a
goodly portion of the hemisphere. They know the causal relations among
these small parts, but are unsure of details (turbulence within the space),
but comparison of simulations with actual behavior tells us that we don't
yet have the combination of physics and computing power to be accurate in
predicting the aggregate.

So my answers to the questions posed would be: Have meteorologists tried
hard enough? Yes, the limits they're at right now are not theory or human
effort, but rather laying their hands on supercomputer time to run dynamic
models of adequately small granularity. Is this a problem causal modeling
is inappropriate for? We know from SD that predicting behavior is a great
deal harder than predicting direction of policy/strategy impact. But as to
what approach is ""right"", that may very well change--supercomputers will get
faster, and training data for neural nets will get better. Which will
happen faster? Hard to say. B.t.w., I'd expect some kind of combined
method to work better still--not using both methods would seem to throw away
data, which at least in theory reduces predictive power.

There are a couple items of good news for System Dynamics modeling, which
tends to focus on man-in-the-loop systems (market and institutional
dynamics), and sometimes living-thing-in-the-loop systems (ecology and
resource dynamics), as opposed to systems that were neither designed nor
evolved like hurricanes.

The first piece of good news is that many human institutions have
micro-level forces at work that tend to keep collections of things behaving
roughly the same over time--these sometimes go under the name of
""entrainment"". For example, people selling something too far different from
what others are selling for won't be the first to sell; they've got
incentive to move toward more typical market prices. (and this is true even
in markets like the stock market in which prices at any point in time are as
much a function of consensus as they are economic fundamentals.)

The second is that human institutions (and some naturual systems) often
follow a ""Peter Principle"" of complexifying only to the limit of their
controllability. Which means that things like markets and ecologies are
often nicely describable in terms of characteristics of aggregates (i.e.
variables) with causal relations among them. So we can teach SD as a
generally very useful method, even perhaps giving little mention to the more
exotic methods in , e.g., an MBA course.

(By contrast, you'd never design something like a hurricane to serve a
useful purpose--the formation depends on growth of tiny instabilities, as to
some extent do direction and velocity. When we have institutions with that
kind of goofy behavior, we dismantle them!)

Interesting questions, useful for articulating what the appropriate problem
space for System Dynamics, and injecting useful humility and guidance for
problems around the edges. Speaking of humility, Alex Leus' response in
this chain mentions Wolfram's A New Kind Of Science, which most reviewers
seem to find lacking in humility. I can't comment directly, having not read
it. However, we've certainly seen in our own field that when people first
discover the power of a new modeling method, the sensation is that it
applies to everything: new, old, useful, or useless. Hopefully in time we
learn some discernment.

cheers,

alan



Alan K. Graham, Ph.D.
Decision Science Practice
PA Consulting Group
Alan.Graham@PAConsulting.com
One Memorial Drive, Cambridge, Mass. 02142 USA
Joel Rahn
Junior Member
Posts: 15
Joined: Fri Mar 29, 2002 3:39 am

The study of hurricanes

Post by Joel Rahn »

In reply to George Simpson's REPLY: A large model is anything but
transparent. I was not thinking of using a neural network to replace an
SD model. I just thought that the parameters of the network model might
say something about where the important loops are i.e. the loops that
dominate a certain behaviour mode.

Joel Rahn
jrahn@sympatico.ca
Locked