Laws@SRI-AI.ARPA (01/10/84)
From: Ken Laws <Laws@SRI-AI.ARPA>
I have been talking with people interested in AI techniques for
weather prediction and meteorological analysis. I would appreciate
pointers to any literature or current work on this subject, especially
* knowledge representations for spatial/temporal reasoning;
* symbolic description of weather patterns;
* capture of forecasting expertise;
* inference methods for estimating meteorological variables
from (spatially and temporally) sparse data;
* methods of interfacing symbolic knowledge and heuristic
reasoning with numerical simulation models;
* any weather-related expert systems.
I am aware of some recent work by Gaffney and Racer (NBS Trends and
Applications, 1983) and by Taniguchi et al. (6th Pat. Rec., 1982),
but I have not been following this field. A bibliography or guide
to relevant literature would be welcome.
-- Ken Lawsrpw3@fortune.UUCP (01/12/84)
#R:sri-arpa:-1524700:fortune:21500007:000:1342
fortune!rpw3 Jan 11 15:35:00 1984
As far as the desirability to use AI on the weather, it seems a bit
out of place, when there is rumoured to be a fairly straightforward
(if INCREDIBLY cpu-hungry) thermodynamic relaxation calculation that
gives very good results for 24 hr prediction. It uses as input the
various temperature, wind, and pressure readings from all of the U.S.
weather stations, including the ones cleverly hidden away aboard most
domestic DC-10's and L-1011's. Starting with those values as boundary
conditions, an iterative relaxation is done to fill in the cells of
the continental atmospheric model.
The joke is of course (no joke!), it takes 26 hrs to run on a Illiac IV
(somebody from Ames or NOAS or somewhere correct me, please). The accuracy
goes up as the cell size in the model goes down, but the runtime goes up as
the cube! So you can look out the window, wait 2 hours, and say, "Yup,
the model was right."
My cynical prediction is that either (1) by the time we develop an
AI system that does as well, the deterministic systems will have
obsoleted it, or more likely (2) by the time we get an AI model with
the same accuracy, it will take 72 hours to run a 24 hour forecast!
Rob Warnock
UUCP: {sri-unix,amd70,hpda,harpo,ihnp4,allegra}!fortune!rpw3
DDD: (415)595-8444
USPS: Fortune Systems Corp, 101 Twin Dolphins Drive, Redwood City, CA 94065