Newsgroups: comp.risks
X-issue: 3.61
Date: Sat, 20 Sep 86 16:11:42 edt
From: harvard!wanginst!infinet!rhorn@seismo.CSS.GOV (Rob Horn)
To: risks-request@CSL.SRI.COM
Subject: Simulation risk

One kind of risk that I have not seen discussed here is the problems posed
by using computer simulation models that are not adequate.  In particular I
am refering to situations where due to either insufficient computer
resources, or insufficient mathematical analysis, the really accurate model
results are not available.  Usually more primitive, inaccurate model results
are available and being used by the ideologues on both sides of an issue.
This places the responsible scientists and engineers in a difficult
situation.  How do you say ``I don't know yet'' and how do you deal with
making recommendations in the absence of adequate information.

I can think of two such situations that have major public decision-making
impact.

The first is the ``nuclear winter'' situation.  I remember many years ago
reading the sensitivity analysis of the one-dimensional and two-dimensional
climate models to solar input.  They were hyper-sensitive, with variations
on the order of measurement error causing massive climate change.  It was
not until recently (1982) that the vectorized Climate Model was analyzed and
shown to be reasonably well behaved.  And even it has some contentious
approximations.  This model requires 15 hours on a CRAY-1 to analyze one
situation for one season.

When the nuclear winter stories came out I had my doubts.  Where did these
people find a solid month (12 seasons x 4(?) test cases) of CRAY time?  Had
they used one of the hyper-sensitive 1 or 2-dimensional models.  What would
the accurate models find?  And how should I respond when I knew that it
would probably be a year or more before that much CRAY time and post-
simulation analysis could be finished?  (Fortunately I only had to handle
party conversation with people who knew that I had worked in that field.)

The same kind of problem occured in the ozone layer issues during the mid
70's.  The more accurate model had two extremely severe problems: 1) it was
unconditionally unstable when phrased as a finite difference problem or
exceedingly temperamental when phrased as an implicit differencing problem.
2) It involved solving extremely stiff differential equations.  In this case
the official answer given was ``we don't know.  It will take several years
of mathematical research effort to make this problem tractable.  The real
answer is anyone's guess.  The published model answers are meaningless.''  A
truthful answer but of little value to decision makers.  (There was a brute
force throw-computers-at-it solution.  Estimated run-time on a CRAY was
about 1,000 years per simulated year.  Easier to wait and see what
happened.)
 
How often are we placed in a situation where the inaccurate computer
simulation is available, but the accurate simulation unavailable?  
What is an appropriate way to deal with this problem? 
 
				Rob  Horn
	UUCP:	...{decvax, seismo!harvard}!wanginst!infinet!rhorn
	Snail:	Infinet,  40 High St., North Andover, MA