Dr. Clifford Mass
summarizes the challenge to improve forecaster
workstation tools as follows:
|
"We are entering an era in which high-resolution models realistically
simulate regional mesoscale circulations, when mesoscale ensembles produce
calibrated regional-scale probabilistic guidance, and forecasters can no
longer consistently 'beat' objectively produced forecasts. The frequent
use of human beings to laboriously alter deterministic forecasts for a
week into the future would be a serious mistake that would lessen
forecasters' time for more productive work.
IFPS [the
Interactive Forecast Preparation System]
must be altered to facilitate short-term forecasting
and the communication of such forecasts.
It should also evolve into a more probability-capable platform that can
make use of ensemble information and can disseminate probabilistic
information. Last, human forecasters must be left with sufficient time to
fill an irreplaceable role--to explain and interpret the probabilistic and
deterministic forecasts created by automated numerical weather prediction
systems and to provide critical advisories, watches, and warnings."
Source: Mass, Clifford F., 2003: IFPS and the Future of the National Weather Service. Weather and Forecasting: Vol. 18, No. 1, pp. 75-79. |
Using current tools,
expert forecasters cannot make full use of
available knowledge and data for nowcasting.
1
There is an increasing amount of remotely sensed data,
a
"data deluge",
more than humans can intelligently and consistently
and quickly process during rapidly changing weather situations.
2
In many non-meteorological professions,
experts systems have been developed to
help experts multiply their effectiveness
by doing much of the most repetitive, systematic work
that was traditionally done by experts.
A commonly used definition of
artificial intelligence (AI) is:
"The application of computers to areas normally regarded as requiring human intelligence."
Because the development of meteorological expert systems has not kept up with the increases in data and the increases in demand for forecast products, forecasters are increasingly challenged to attend to the details that experts are expected to attend to, and in so doing, are increasingly losing sight of the "big picture". As Henry David Thoreau might have put it: "Men have become the tools of their tools." The "What" to "How" spectrum of computing In the "man-machine mix", AI can mediate between perception and decision making and computation. According to Edward Feigenbaum, AI enables people to spend more time controlling "what" computers do and less time controlling "how" computers do. |
"The 'What' to 'How' spectrum of computing. Off to the left, the user wishes to express goals and needs in his
or her own terms, not in terms of the computer's step-by-step instructions. The history of the computer
software is a series of steps that allow the user more of that freedom. AI software aspires to be at the point
closest to the user's desires and problems."
Edward A. Feigenbaum, How the "What" Becomes the "How", Communications of the ACM, Volume 39, No. 12 (Dec. 1996). |
Currently, nowcasting tools use very little AI.
To produce nowcasts,
forecasters have to:
1) use their heads, to perceive, think, and decide;
and
2) use their hands, to specify details, step-by-step, in computers' terms.
When nowcasting tools use more AI,
forecasters will spend less time working with their hands,
to specify
how
forecasts are constructed,
and will spend more time working with their heads,
applying their meteorological expertise,
to determine
what
forecasts say.
Nowcasting tools, for both regular and severe weather, should be designed to enable synergistic use of the complementary abilities of expert forecasters and AI. For this, the different needs and abilities of forecasters and AI should be initial considerations in design of nowcasting tools. National Meteorological Services need to develop and use intelligent nowcasting tools to enable intelligent integration of predictive, heterogeneous fields -- such as observations, model based guidance, and projected radar and satellite images -- to enable such data to be intelligently and consistently and quickly processed into better nowcasts.
Three types of intelligent nowcasting tools Three types of intelligent nowcasting tools -- heuristic, non-fuzzy, and fuzzy rule base -- are described here, along with their relative advantages (+), disadvantages (-), and a few examples: 1. Heuristic: experience-based, useful nuggets of wisdom gleaned from experience, perception, observation, and analysis. Basis of rules of thumb, expertise, insight, and intuition. (+) Sound results, accord with knowledge and experience; (-) Inconsistent and "expensive", in development and in application; (-) Context dependent; (+/-) Refer to abstract (fuzzy) features, such as "high to the north", which people process well, computers not so well. Examples:
2. Non-fuzzy: basically, a decision tree with answers at terminal nodes. Often implemented as "if-then-else ladders" or as conditional "equations". Discrete thresholds between predictors/conditions. (+) Time-tested "work horse", effective in many situations; (+) All meteorologists and developers of meteorological systems are familiar with decision trees; (-) Lack explanation capability. Usually a "black box".
(-) Lack facility for intelligent data fusion. For example, given fields of current observations, model output, and radar and satellite projections, how would a system "intelligently" compose a projection of a field of interest (e.g., severe thunderstorm at CYYZ at t+1 hr); (-) ... hence, non-fuzzy systems are poor for nowcasting and "brittle" if "misapplied" (e.g., recent radar should "outweigh" older MOS). Examples:
3. Fuzzy rule base: answers based on a weighted sum of all rules "firing" simultaneously, each rule activated to the degree to which its conditions are valid. Predictors/conditions specified by fuzzy sets. Rules expressed in fuzzy language (hence fuzzy logic has been described as "Computing with Words"). For example, the method used in a classic example of a fuzzy rule base for controlling an industrial furnace has been adapted directly for forecasting radiation fog. (+) Numerous: www.rap.ucar.edu/technology/iws/advantages.htm; (+) Rules are clear to developers and forecasters, easy to develop, tune, and maintain (www.nrlmry.navy.mil/~medex/tutorial/toc/toc.html); (+) Can incorporate heuristics and non-fuzzy decision trees and other forecast models and heterogeneous data with intelligent integration. According to Richard Wagoner, Deputy Director/Program Development Manager at the Research Applications Program, National Center for Atmospheric Research (RAP/NCAR): Over 90% of systems developed at RAP use fuzzy logic as the intelligence integrator. (+) Operates intelligently on heterogeneous data (e.g., observations, model output, and radar and satellite projections) to help nowcasting, as explained by the Research Applications Program (RAP) in an article about Intelligent Weather Systems design. (+) ... hence fuzzy rule based systems are good for nowcasting, more capable than heuristics and non-fuzzy decision trees by themselves, and less brittle; (-) Most meteorologists and developers of meteorological systems are [still] not familiar with fuzzy logic. Examples:
I've only roughly outlined three basic and distinct intelligent nowcasting approaches. Of course, in practice, forecasting experts, researchers, and developers combine the approaches. For example:
Bjarne Hansen 1. This webpage describes how to improve forecasting tools through intelligent integration of frequently updated predictive fields, for example, regular observations, and current and projected weather radar images. Such data is very valuable for forecasting a few hours forwards, but its value drops off after a few hours. I refer to such forecasting as "nowcasting" to emphasize the value of current data. However, all three of the following types of forecasts could benefit from the development of tools for intelligent integration of data.
2. If weather forecasting is like flying an airplane, then nowcasting is like landing an airplane. To help to explain why human forecaster expertise is needed in the forecasting process, an autopilot analogy is often used. For example, as Richard Verret recently put it:
Flying an airplane consists of three stages: takeoff, flight, and landing. Pilots generally fly on autopilot during flight, when events requiring rapid and intelligent intervention are relatively rare. However, they still fly manually during takeoffs and landings, when events requiring rapid and intelligent intervention are relatively common. A forecast "takes off" when, for example, a forecaster decides whether to choose model A or model B. A forecast "flies" when, for example, a forecaster performing weather watch perceives no reasons to revise a forecast. A forecast "lands", for example, when a forecaster decides whether or not an existing thunderstorm is likely to track near a target, such as a major city or an airport.
Last updated 28 June 2003. Opinions expressed here are solely those of the author except where indicated as otherwise. |