By Jessica Keyes
Over the last several years, more than 80 percent of the Fortune 500 companies have
explored "smart system" techniques. While there are a preponderance of smart
system technologies including fuzzy logic, neural networks and natural languages, this
article will concentrate on and explain the one with the most buzz - expert systems.
The usual selling point of expert systems is that they encode the knowledge and reasoning
skills of experts. It permits users to conduct intelligent dialogues with automated
systems, providing an enormous boost to productivity, and dramatically extending the power
of the computer.
In 1985, E.I. du Pont Nemours & Co., realizing the strategic value of this type of
technology, made a wholesale effort to train anyone on the staff who was interested in
expert systems to build his or her own. Today, there are more than 600 expert systems
installed in du Pont's business units. What's even more interesting is that this effort
has saved du Pont some $100 million.
Along with making a great impact on the P&L statement, the goal of more than one of
those 600 expert system was to get du Pont a greater percentage of market share - or even
to break into a market that they were never in before. An example of this last strategy is
a system named the Packaging Advisor. This expert system is used for designing rigid
plastic food containers and helped du Pont break into the very competitive barrier resin
market.
Digital Equipment Corp., another pioneer in using artificial intelligence techniques, has
some 50 expert systems in place, which has led to a $200 million dollars savings to their
bottom line. The expert system they are best known for, now approximately a decade old, is
called XCON, short for expert configurator. XCON automatically writes the technical
specification for a mid-ranged computer configuration. Because a Digital computer can have
from 200 to 8,000 parts, the human technical writer, whom XCON replaced, made more than a
few errors. One of the largest expert systems on record, XCON, with a rulebase of over
10,000 rules, was one of the key elements responsible for making Digital a strong
competitor to IBM during the 1980s.
But what are expert systems? In a human resources knowledge base, for example, there are
no facts and figures, only stored knowledge about how the human resource department works.
It's this expertise that proves so valuable.
Expert systems are found in a myriad of areas such as geology, information management,
law, manufacturing, medicine, meteorology, the military and space. In the world of
agriculture, an expert system predicts damage to come due to cutworm. Another expert
system manages apple orchards. In the world of chemistry, the expert system
"Dendral" can determine the molecular structure of unknown compounds from mass
spectral and nuclear magnetic response data. In the area of engineering, the
"Reactor" expert system can help operators in the diagnosis and treatment of
nuclear reactor accidents. The secret behind an expert system, therefore, is its
knowledge.
Expert system technology has been available for several decades, although it's only the
last that has seen its popularity grow in the commercial sector. From the outset, AI
tools, particularly expert systems, have required the power of a heavy-duty workstation.
But both a refinement of the software and the increasing power of the PC has seen a surge
of applications downsized to the Intel environment.
Still, the workstation has its share of proponents in the AI arena, particularly in the
areas of high finance and manufacturing where compute power is often the difference
between success and failure.
When the name Campbell is mentioned, one is likely to think of soup, because Campbell has
cornered this market for well over 40 years. Campbell also holds the number one or two
brands in its frozen, baked goods, beverage and grocery business units. Over the last
decade or so, Campbell has seen its market share eroding as more and ever more aggressive
rivals add soup cans to supermarket shelves. Given the problem of narrow margins and
severe price competition, Campbell needed to become more competitive - and more quality
oriented. This they did with an expert system named Cooker.
It was the job of one human being to diagnose problems with what is known as a hydrostatic
sterilizer or cooker. A cooker is a unit about 30 feet square, over 70 feet high and able
to process tens of thousands of cans per hour. The problem with these 70-foot cookers was
that it was not unusual to incur significant lost production time on each cooker
malfunction.
When the human expert decided it was time to retire - after 44 years on the job - Campbell
decided to try to capture his knowledge in an expert system. The system had two primary
goals. The first was to be able to replace as much of the expert's diagnostic judgment as
possible. The second goal strived for a final product that would be a useful training tool
for production and maintenance engineers. The end result was an expert system that
contained 44 years worth of experience and was able to give advice about the operation as
well as the start-up and shutdown of the hydrostatic system.
Another early adopter of the AI technology that also saw its image gain a new luster as a
result of an early adoption of an advanced technology is American Express.
AMEX is one of the largest and most profitable of American companies, running a host of
diverse business enterprises, all of which heavily depend on technology.
AMEX's first, and best known, project was a system known as the Authorizer's Assistant.
The goal was a difficult one. AMEX wanted to use advanced technology to solve the problem
of increasing bad debt and fraud in the use of credit cards. Since conventional computer
systems didn't make much of a dent in reducing this problem, the idea of an expert system
was appealing.
Authorizer's Assistant assists operators in granting credit to card holders based on a
review of the customer's records. Because there is no preset credit limit this process can
be a bit tricky.
The original system took four and a half months to prototype and consisted of 520 decision
rules running on a workstation. Ultimately expanded to more than 800 rules, it was found
that usage of the system provided a 76 percent reduction in bad credit authorizations. The
system was found to be accurate 96.5 percent of the time as compared with the human rate
of only 85 percent.
Both American Express and Campbell utilized workstation and workstation-based AI toolsets
to deliver, at least, the prototypes of their expert systems. There are a wide variety of
expert system toolsets available across the heterogeneous platforms that are common in
today's organizations. In the workstation arena both Inference and IntelliCorp remain the
leaders, but there is no dearth of competitors such as Palo Alto-based Neuron Data.
Today's expert system tools are very much object-oriented, although they still retain
their decision tree or rule flavor which usually gets coded somewhat similarly to the
example below:
If ?X is in class REACTORS and
The PRESSURE of
The PRIMARY.COOLING.SYSTEM of ?X
is DECREASING and
The STATUS of
The HIGH.PRESSURE.INJECTION.SYSTEM of ?X is ON
Then The INTEGRITY of
THE PRIMARY.COOLING.SYSTEM OF ?X
is CHALLENGED
Do ACTIVIATE.ALARM
GET.VALUE X 'PRIMARY.COOLING.SYSTEM
The rules and objects of an expert system are processed by what is known as an inference
engine, which is the real heart of the system. In expert systems, unlike conventional
systems, rules do not execute in any particular order. Rules execute according to a
control strategy that prescribes the methodology used to search through a knowledge base.
There are two major categories of searching: forward and backing chaining.
To simplify a very complicated subject, in forward chaining we move from a set of
assertions or facts to one or more possible outcomes in a sort of top-down reasoning
approach. Here the system searches for a value where the conditions in the IF part of the
rule are deemed to be true. Backward chaining is forward chaining's opposite. This
strategy operates from the perspective that you already possess an outcome and are
searching for the conditions or circumstances that would lead to that result.
In both cases, the inference engine maintains a truth table that permits the process to
stop, back up and go in a different direction, depending upon the inputs of the user of
the system. Additionally, since truth is a matter of degree, most expert system tools have
the ability to factor in a certain degree of variability in the form (most often) of
confidence factors.
Thus the outcome of a particular expert system consultation might provide a set of answers
of the ilk: a) There is a 60 percent possibility that Mr. X is the guilty party; b) There
is a 99 percent chance that Mr. Y is the guilty party. This is something the trial of the
century in L.A. could have used to everyone's advantage.
Jessica Keyes is president of New Art Communications, a consulting firm based in New York
City. She is the author of seven books including the McGraw-Hill "Multimedia
Handbook" and several books on artificial intelligence. She can be reached at
212-362-0559.