|
Items in cart [0] |
Questions? Call 888-624-8373 |
| Copyright © 2009. National Academy of Sciences. All rights reserved. Terms of Use and Privacy Statement |
Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter.
Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 113
EXPERT SYSTEMS: AEPITCAT~ONS IN SPACE
Bruce G. Buchanan
~TE~rXJOl'lON
Artificial intelligent- is one of the most important trends in
computing becau ~ making cc mputers behave intelligently is at least as
important as manipulating data efficiently. Opportunities for using
intelligent programs in NASA space station environments are numerous
and abvicus. But many of those opportunities require substantial
research in artificial intelligence before they can be realized.
This
_
paper looks at the technology of artificial intelligence, especially
expert systems, to define "from the inside Out" what capabilities exist
that are relevant for applications and environments In the space
station, and what research needs to be promoted in order to achieve
systems better able to interact symbiotically with a variety of persons
for long times in space.
Anderson and Chambers (1985) mention a number of characteristics of
systems in a human-centered space station. These include:
symbiosis with humans: human and machine capabilities may
complement one another
autonomous,
continuing operation for a period up to 20 years,
operating in an information-rich environment,
consequent ~ of interactions with humans not entirely
predictable,
maturation of system implies flexibility to accommodate
opera;~ion21 growth and minor up graces,
evolution of system implies flexibility to acoom=odate new and
enhanced functionality,
system may include electrons c crew belters (ECHO),
humans may have to learn new skills to interact productively
with computers,
computers may learn from humans,
autonomous agents may serve a variety of roles with varying
degrees of decision making power and authority.
These are some of the relevant considerations in a top-5cwn design
of systems for the space station. Each of these points implies a
113
OCR for page 114
114
research and development program of some intensity. This paper takes a
bottc m-up view of the same considerations--i.e., starts with what
exists today and asks how we can achieve these design goals. By doing
so, I hope to Introduce some relevant details into the design of
Systems and the pi q of research.
Expert systems are now being used in many decision-ma-ding situations
of direct relevance to N~SA's mission, spanning manufacturing,
engineering, medic me, and science. At present, they are used more as
"intelligent assistants" than as replacements for technicians or
experts. That is, they help people think through difficult problems
and may provide suggestions about what to do, without taking over every
aspect of the Cask.
Cc mar programs that reason autoncmously are also of extreme
importance in space, but they, too, must be integrated into an
environment that is centered around people. They are extensions of
present technology along several dimensions discussed here, that
involve all of the same principles of design as the Intelligent
assistant prc grams.
One primary consideration is why intelligent systems are necessary
in space. Although there are many reasons to build an expert system,
they are all based on the premise: "Expertise is a scarce resource."
The corollary (by Mhrphy's Law) is: "Even when there is enough
expertise, it is never close enough to those who need it in a hurry."
Because this is true--almost by definition of the term
' expertise '--constructing expert systems that reason at the level of
N~SA's, or their contractors', specialists may have several benefits.
These we summarized in Table 1.
WHAT IS AN EXPERI SYSTEMn
The general nature of expert systems is familiar to everyone within
NASA. A reiteration of the four major characteristic is provided
below to help define the most important dimensions for research and
development efforts.
An expert system is a computer program with expert-level problem
solving abilities, which also fits some other criteria: it is a
IABIE 1 Some Perceived Benefits of Expert Systems: Responses of 86
Users of Knowledge Engineering Tools
1. Replicate expertise
2. Preserve expertise
3. Increase productivity and cost savings
4. Free human experts for more demanding problems
5. Provide expert consultations to inexperience staff
SOURCE: Bauman (1984)
OCR for page 115
115
symbolic reasoning program that uses heuristics, its reasoning and
knowledge base are understandable, and--most importantly--it is
flexible. These characteristics are discussed below. All are
important for applications in the space station, and all define
research topics that will enhance current capabilities.
Example
One well-known expert system that has become a classic, although not
actively used, is MYCIN. It was developed at Stanford by E. H.
Shortliffe and others ~ the m~d-1970s. Its task is two-fold: (a)
diagnose the cause ts) of infection in a patient and (b) recommend
appropriate drug therapy. From a medical perspective, MYCIN's
knowledge base is now dated; from the perspective of expert systems it
represents much of the kind of reasoning that is captured in today's
systems. MYCIN's conclusions were demonstrated to be equal in quality
to those of infectious disease specialists at Stanford Medical Center.
The sample typescript shown in Appendix A illustrates MYCIN's
requesting information about a case and reason mg to conclusions about
the best treatment.
Performance
Naturally we want computer programs to solve problems without error.
But that is not always possible--in fact, outside of mathematics and
logic we don't have flawless methods we can put into programs.
Specialists in engineering, science, education, the m~litary--and every
ares outside of pure logic--must solve problems with less than perfect
methods. How do they do it? Mostly by building up specialized
knowledge through extra years of training and experience and by
reasoning carefully with that knowledge in situations they have learned
to recognize. They are not infallible, though. Specialists' decisions
are challenged frequently--most noti~=hly in the courts. So it is
also unreasonable to expect computer programs to reason infallibly in
all of these areas. Occasionally new methods are disocvered that
provide much better results than the established methods of the old
practitioners. But these improvements can then be put into prc grams,
thus raising the overall standard of performance while still keeping
the same relative standard of comparison with the best specialists.
Reasoning
When we say that expert systems are reasoning--and not just calculating
with numters--we are saying that they belong to a class of programs
using the methods of artificial intelligence thereafter AI). In the
1940s, computers were used almost exclusively for large mathematical
problems. At Los Alamos, for instance, scientists had to solve complex
mathematical equations in order to calculate elements in the design of
OCR for page 116
116
the atomic bomb. These applications are usually referred to as
large-scale scientific computation, or "number crunching" for short.
In the 1950s, IBM and other computer manufacturers, realized the
enormous value m helping b~-=iness solve problems of record keeping,
payroll and the like. These applications extended the concept of
comput~r-as-ca1culator to ccmputer-as-data-manager.
In both of these classes of applications, the method of computation
is error-free. There is no question that the result is correct,
providing of course that the computer has been programmed correctly. A
mathematical equation is solve] correctly; an employee roster is sorted
correctly--if the Beth ~ are followed precisely. And computers are
better able to follow complex instructions than people are. In
computer science, logic an] mathematics we call these procedures
algorithms. They a ~ procedures that can be guaranteed to provide a
correct answer in a finite time, if there is one, and otherwise will
provide a statement that the problem is not solvable.
Some algorithms are too expensive to use, however, even in
camputers. A classic example is finding the shortest route a
Averting salesman can tat to visit many cities once and end up at
home. With more than a hard of cities, aigorit~nic me~chc~s will not
finish in time to be useful. For this reason, alternative methods have
been developed.
Art the mid l950s and early 1960s an alternative side of
Sting ~ to be romanized as i~por~cant. Instead of always using
aigori~ns, a Outer may case heuristics rules of thumb that aid In
finding plausible answers quirky without guaranteeing the correctness
of the results. Sc~t;~s these rules of throb are introduced into
large primal simulations in order ~ get the situations ~ Card
out anshrers more quickly. Or approximate methods may be sup tituted
for more precise ones for the same reason. The assumptions may not all
be correct; thus the results of the simulation may not be correct.
When heuristic (non-algorithmic) methods are combined with symbolic
(non-numeric) data, we are dealing with that part of computer science
known as artificial intelligence.
Understandability
When someone truly knows something, he or she can "give an account" of
what he knows. In our terms, good performance is not enough to call a
person (or pro gram) an expert he/she (it) should also be able to
explain why the solution is plausible, what features of the situation
were noted to be important, what knowledge and problem solving methods
were used. Otherwise we label a person ~~ "consistently (but
unacocNnt~h~y) lucky", or maybe "psychic". Each field has its own
standards of what a reasonable explanation is. A surgeon who
recommends amputation of a leg generally talks about the process of
disease or extent of injury and what will happen if it is not
amputated. A broker So aWised liquidation of one's stock portfolio
may explain the arrive with resent to technical Carts, historical
trends, or sad ec~onc~nic principles that point to a stock market
OCR for page 117
117
collapse. In their awn c~nities, both ache surgeon and the broker
can usually justif~r--in hurt if necessary--the advice they give. Arxl
we rears them as experts party because they have he ~a~riedge ached
lets them do this.
Flexibility
We e; ~ experts to be flexible In their Inlet ~ . Arx] we regard
persons as amateurs, not experts, when we encounter opinions that are
rigid, locked-in ways of dealing with problems, or an inability to deal
with new situations.
In particular, there are two situations in which we want expert
systems to be flexible:
At advicergiving tamp we want the program (or a person) to
provide good advice about situations that have never been
encountered before. Novices with good memories may be able to
provide the "textbook" answers for classic situations. Experts
however, should, in addition, be able to reason about novel
situations.
2. At the time a program is being constructed or modified (or a
person is rearm ng), we want it to be flexible enough to
assimilate new bodies of information. There should be a
capacity for growth of knowledge, not a rigidity that freezes
either the depth or breadth of the program's knowledge.
SC ME APPIICAIIONS
Some of the types of problems for which expert systems have been
constructed are shown in Table 2. Many of these, such as small
troubleshooting assistance programs, are relatively straightforward.
Although the state of the art is difficult to quantify, the programs In
the Tahoe represent the kinds of commercially robust systems that can
be built for NASA today, provided adequate resources and an app striate
problem. We don't have an adequate taxonomy of problem types. Many of
these overlap, ~ being different forms of data interpretation, for
example. Even this brief characterization, however, provides a
reasonably good idea of what expert systems can do.
In general, expert systems can reduce costs or increase quality of
goods and services--in a single phrase, they can increase productivity
in an organization. If you believe either that there is not enough
expertise in the world, or that it is not well distributed, then you
will be willing to entertain the idea that putting human expertise into
an -~ily-replicated form may answer some productivity problems. Or,
at least expert systems may provide a partial answer. Consider medical
OCR for page 118
OCR for page 120
OCR for page 121
OCR for page 122
OCR for page 123
OCR for page 131
OCR for page 132
OCR for page 133
OCR for page 134
OCR for page 135
OCR for page 136
OCR for page 137
OCR for page 138
OCR for page 139
OCR for page 140
OCR for page 141
Representative terms from entire chapter:
please enter
118
HE 2 Some Emblems for which Exam: Systems are ~vidi~
Solutions
RISK ASSESSMENT
E~I~ DAIS
BEEN
ADA
En BEG
It ~1']!;~'~
SO
q~ ~Nr
_~ . _
Ax . ~ : A;
P~L10 ~ =
~U~
AURA S"1~;=
CROP ~
#
St. Paul Once Co.—assess risk of
insuring large camme~ial clients
Genera ~tors~determi~ causes of
vibration retie= and reccm=end repairs
Digits Equipment Corp. translate
customers' orders for computer sys ~ c into
Hi Fin orders
U. S. Army--design loading plan of cargo and
equipment into aircraft of different types
U. S. Air Force--plan an aircraft's route
from base to target and back to avoid
detection and threats when possible
Schlumberger--~nterpret down-hole data from
oil well bore holes to assist in prospecting
Wasting house--plan manufacturing steps in a
plant so as to avoid bottlenecks and delays
Stanford Medical Center--~cist in managing
malti-step chemotherapy for cancer patients
IBM- monitor operations of MA operas mg
sat
U. S. Enr~ ion
Agency--determine which requests for
information fall under the exceptions to the
Freedom of Information Act
First Financial Planning Systems (Travelers
Inch- analyze an indivi~sl's
financial situation and ~ types of
investments
Hewlett Packard~iagnose causes of
problems in potolith~ra~hy steps of wafer
fabricatica,
Virginia Polyphonic Institute -
119
diagnosis. Specialists at university medical centers generally see more of the
unusual disorders than a rural practitioner and thus stand a better chance of
diagnosing them correctly. Putting some of that expertise more directly at the
service of the rural practitioner could al low more effective treatment, and
save patients the time and trouble of travel to the medical center.
Or consider troubleshooting a complex piece of equipment. Persons with the
most field experience are often the ones promoted to desk jobs in the central
office. When subtle combinations of causes keep a less experienced field
service technician from fixing a mechanic=] failure, someone with more
expertise is needed. On earth, depending on travel times and =~he criticality
of the work flow in the central office, calling the experienced specialist out
may be a very expensive repair procedure.
The following situations are all cases where it may make good sense to build
an expert system:
too few specialists for the number of problems;
specialists not at the sites of problems when they occur;
long training time for a specialist;
high turnover among technicians;
combination of complex equipment and poorly trained technicians;
organization's best (or only) specialist in an area is nearing
retirement;
too many factors for a person to think through carefully in the time
available.
KEY CONCEPTS
The four goals that characterize expert systems can be achieved with a few key
methodological ideas. In this section, the key ideas will be introduced; in
successive sections they will be elaborated on so as to explain a little how
they work. The main organizational principle of expert systems is to keep
specialized knowledge separate frog the logical and heuristic inference methods
that use it. m is is -any to say but difficult to follow, for reasons that
will be described later.
Keep Dc main-Specific Knowledge
Separate frog General Reasoning Methods
-- FEY IDEA #1 --
Another key concept, which is imported from principled design of software
generally, is modularity. (The first key idea is an instance of this, but that
instance has taken on more importance than all the other instances of the
general concept.) Modularity at the level of knowledge about the problem area
implies conceptual separation of elements in the knowledge base. For example,
medical knowledge about penicillin, although not totally independent, can often
be separated from knowledge of other drugs. It can be codified in major ways,
or deleted, without altering the program's knowledge of other drugs. So, this
is to say that the concepts used to talk about objects ~ the Clan should be
chosen so as to allow talking separately about an individual object, a single
120
properly of an object, or a single relation of one type of object with
another. Modularity at the level of programming constructs implies that the
program's 1ntern21 representation of knowledge elements (e.g., objects,
properties, relations) is similarly "clean".
Keep independent pieces of knowledge independent.
Keep the rest as nPa'ly-independent as possible.
--KEY IDES #2 --
A third key concept is uniformity of conceptualization and representation of
knowledge. The underlying intuition is that it is easier for a person or a
program to build, understand, and codify a body of knowledge if it doesn't mix
and merge a variety of different types of things. This is as true at the
knowledge level as at the programming level. For instance, one of the most
compelling aspects of Newbonts laws is that all physical bodies ore treated as
quantities with mass. He didn't need one set of laws for planets and another
for apples. So it is desirable to build an expert system with a "conceptually
clean", well-organized, simple collection of concepts. And it is important to
use a simple, well-organized collection of programming constructs as well.
Otherwise there are too many different kinds of things to keep track of and
reason with.
There is more dispute among AI specialists about this principle. There are
good reasons to violate it, as we shall see, ~ the interest of being able to
say more about the objects and relations of interest than can conveniently be
said in a single language. We are frequently told by bi-lingual friends, for
instance, that there are some concepts that just can't be expressed fully in
English. m e same is true for programming constructs, but the basic principle
for constructing expert systems is to try to ma Attain uniformity as much as
Possible.
Strive for uniformity of language
and programming constructs
KEY IDES #3
A fourth principle is to design the expert system to mirror the ways experts
think about problems in their dc mains. That means using the same terms and the
same rules of reasoning as the experts use. One reason for this is that
building and debugging a knowledge base depends necessarily on the expert, and
using 1=== fa ~ liar terminology or methods will introduce confusion and error
before the knowledge base is completed. Also, aft or it is completed it needs
to be comprehensible and unambiguous to the practitioners using the system or
else confusion and error will result.
Note that we are assuming here that the expert designing the system knows
how to make it understandable to users. Great care must be taken when building
a system, however, to insure that this assumption is true.
There are times when this principle will be, or should be, violated. For
example, when efficient computer algorithms can solve part of a problem, it
doesn't often make good sense to use anything else for that part, even if the
experts don't think about it in that way.
121
As huh as possible, he the sane vocabulary and
methods In the program as the experts and
practitioners use.
-- KEY IDEA #4 --
These key ideas help us achieve all of our four goals in the following ways.
· HENCE In problems whose solution methods are not airea~r well
formalized, which are considerable, Huh of the effort In building a
knowledge base frees an expert system lies in building the conceptual
frank. ~i~hpr~ies and relations of objects to describe is
often not well specified at the beginning. So the Repledge base is
built Duncan - lay, Here experience with one kna~riedge base guides
future modifications, extensions, or reformations.
· RFASON~-~en the solution methods are not well characterize, it is
important to encode heuristics that experts say they use. Storing these
separately and In a simple form allows them to be changed easily. Since
it is nearly impossible for an expert to articulate a complete and
consistent set of heuristics at one sitting, it must be easy to add,
remove, or modify the heuristics that determine the reasoning.
· UNDERSTAND~BILPTY--with modularity, individual elements of the knowledge
base can be displayed meaningfully in isolation. Moreover, with the
separation of knowledge base and inference prcce~ures it is possible to
peruse the knowledge base in order to find just those elements that were
used to reason about a new case. And with uniformity of data
structures, it is possible to build one set of procedures that produce
explanations.
· FLEX~BILITY--when the elements of the knowledge base are in separate
data structures, and not intents med with code for inference procedures,
we can add more knowledge with considerably more ease. When the
indivi*u31 items in the knowledge base are nearly separate, we have
fewer interactions to worry about when we change an items. And when the
representation is homogeneous, we can more easily write other programs
that act as "editing assistants" or explainers that help us insure
correctness of new items and help us understand what is in the knowledge
bit.
~RMANCE ISSUES
Expert systems constitute one class of computer programs. As such, they work
the same way as every other program: they process input data to produce output
data. But the nature of the processing is different from most conventional
programs. The key ideas motioned earlier are the key differences In the
design am] irritation of expert systems.
In order to design a rezoning program, we need to provide knowI~ge to
radon with ark] reasoned methods to use. Both are needed. A
122
powerful thinker needs something to think about, and a body of facts
without methods for using them is sterile. Over the last few decades,
research in AI has elucidated programmed y methods for making inferences
and storing knowledge. We briefly characterize these topics below,
although with some reservations about cversimplifying, in order to
highlight research issues relevant to increasing the performance of
expert systems. In addition to research on inference methods and
representation of knowledge, several other issues are mentioned briefly
as needing more research in order to improve the performance of expert
systems.
Inf~ren~- Methods
Aristotle's Theory of the syllogism defined a~r~ptahie inference
methods cutside of mathematics for about 2000 years. His theory has
been extended in this century by Russell & Whitehe,ad, and others, in a
formal theory that includes methods of reason ng with several
statements and several variables in an argument.
Formal logic defines several inference rules which are guaranteed to
create true conclusions if the premises of the argument are true. The
chain rule (mcdus ponens) is the single most important inference rule
in expert systems. It allows us to chain together a string of
inferences=
If A then B
If B then C
If C then D
A
Many of the inferences we make In our lives are not guaranteed by
the rules of logic, however, nor do we have certain knowledge about the
truth of our premises. Whenever we argue that the future will be like
the past, as in stock market predictions, we have to be prepared for
exceptions. These inferences, labeled "plausible inferences" by George
Polya, are the ones of most interest In AT.
One set of p~r~ methods were In AI for making plausible
inferences is to assert the facts categorically--as if they wee hood
to be true with certa~n~r and then reason about exceptions that might
force revisions to he conclusion.
Anther set of naptha= deals explicitly with the degree- of
unc~ca~n~r In the facts and In he associations. MYC]N (see Premix
A) uses this style of neasonir~. Usually the degrees of uncertainty
implied by words like "often" and ' ~ y" are expressed as numbers. And
often these numbers are interpreted as probabilities.
A third, and most powerful, set of methods is to introduce heuristic
rules, or rules of plausible inference, into the reasoning. These are
facts or relationships that are not guaranteed to produce correct
conclusions, but will often do so. Moreover, they often produce
answers more quickly than their algorithmic counterparts. In the
123
traveling salesman problem, for example, the problem is to plan a route
for visiting each cider in a set exactly once and ~ at the hcme cider.
Ihis is an NP~nplete problem, that is, the algorithm for solving it
Woes times that is exponential with ache number of cities. One
heuristic we Tray ~n~uce is to go to the nearest city Scat has rot
yet been visited. This certainly is up the computation of the
route, but may (and probably will) miss the route that is shortest
overall. Some mules of plausible inference used, with caution, In some
exit systems are Shown below:
Satisficing: If it will be expensive to find the very best
solution to a problem, then stop with the first solution that
satisfies easier criteria of being good enough.
Inheritance: (Some specified) properties of a whole are shared
by all its parts. E.g., An ice cube is cold and hard. Pieces
of an ice cube are cold and hard. LBut other properties, like
'~eight", do not behave the same.]
Single Fault: If a piece of equipment (or any organized system)
is malfunctioning, and one hypothesis explains the problem, then
there prod ably is only a single cause of the problem.
Compelling Evidence: If you have gathered a lot of evidence in
favor of hypothesis Hi, and very little evidence against it, and
you have gathered little positive evidence for alternative
hypotheses, then HI is a plausible hypothesis.
Dcccmposability: If there are many parts to a problem that are
nearly independent, assume they can be solved independently.
Then adjust the composed solution to take account of known
interactions.
Parsimony of Design: Designs or plans with fewer elements are
preferred to those with more.
In pr mciple, the rules of inference (both logical and plausible)
may be applied again and again to a situation description, in any
order, and the resulting conclusions will be the same. This is not
always possible in practice, however. There may not be enough time to
reason exhaustively about all possibilities and contingencies. For
that reason AI researchers talk about controlling the inferences as
being a more important, and more difficult, problem than making the
inflect
Controlling inferences breaks down into two subtasks: (a) deciding
which rules to apply now, at this stage of the problem-solving process,
and (b) deciding which part of the problem to work on now. Since we
believe these sNbtasks require some intelligence, all of the principles
for building knowledge-based systems also apply at this level of
reason m g On participle, ~t is desirable to make this control
knowledge explicit and separate from the inference methods
131
Real Tim Monitoring
As Herb systems beeline faster, it will be Tier to build systems
ached monitor other devices or processes with rapid change.
Conceptually a difficult problem is managing time~ependlent relations
efficiently, which is one of the necessary ~ onents of a monitoring
system. The large amounts of data received and the speed with which
they are received are also critical issues. Integrating AI methods of
reasoning about the data with numerical methods for digitizing and
filtering is essential.
Richer Ihput/Output
No one likes to interact with computers by typing. Considerable work
on interactive graphics has reduced the need for typing. But it will
be even easier when we can communicate with programs by giving voice
commands and receiving spoken English output in return.
Models of Users and Situations
No single style of interaction is best for all users at all times.
Specialists do not need explanations of the meanings of terms, for
example, while less experienced users used considerable help
understanding the context of the problem. Also, the criticality of the
situation may demand taking shortcuts in data acquisition or reasoning
to reduce the risk immediately before taking a more systematic,
detailed look at the problem. Expert systems must be sensitive to
models of both the user and the situation in order to request
appropriate input, reason at an appropriate level of detail, and
present conclusions and suggestions in an appropriate way.
CONCLUDING OBSERVATIONS
Expert systems already are saving organizations millions of dollars and
performing tasks routinely that ordinarily require human expertise.
The number of applications of to~ay's technology is nearly
boundless--consider, for example, the number of pieces of equipment in
a space station that we don't readily know how to fix. The first
commercial shells on the market are robust enough to be used
effectively. Integrating ~nt=1ligent systems with conventional
computer programs and with persons ~ the space station involve= new
research in many dimensions. The single biggest advantage of AI
programs, amply demonstrate in expert systems, is their flexibility.
This matches precisely the single biggest design requirement on
software in the space station.
What we see now is just the beginning of ~ wave of intelligent
software that can have as great an effect as business data prccessing
software. It is impossible in any area of technology to make accurate
132
predictions. However,
there are many parallels between the growth of
expert systems and or computing hardware, with about a 25-30 year lag.
When electronic computers became available commercially, businessmen
began to ask about applications that would make a difference to them.
In 1955, several of these innovators assembled at Harvard to discuss
they e~ien~-C. She of the conclusions they Mew falcon their party
experience are summarized belay (Sheehan, 1955~:
.
1.
"The initial c~v~n~chusiasm, which inevitably Zanies a
project of this scope, can and does make the job harder. Too
many people had the impression that this was the answer to all
Problems. Plans it is. but we haven't bed sprat enough to
. .
develop all of them...
, —
2. "Some of cur original thinking has been partly confirmed in that
the greatest benefits to be derived from a computer will
probably consist of information impossible to obtain
previously...
3. "Our experience has shown that the compute' is more adaptable to
some projects than others...
4. "Programmers should be recruited within your own company...It is
easier to teach men the required computer and program techniques
than to acquaint them properly with the complex pracedurcs and
routines of mcdern-day industry...
5. "I doubt if it is possible to overemphasize the desirability of
providing for convenient corrections or deletion of errors in
data. . .
6. 'the maximum justifiable amount of flexibility for extending or
integrating applications must be included in the initial
programming..."
The parallels with expert systems are clear: we have just begun to
explore the first set of applications an] have only a limited view of
possibilities in the future. Over the next 5-10 years, however, we do
know what extensions in capabilities will be possible because they are
the ones that are active research areas in the present.
Future developments will make the technology more broadly applicable
and easier to use. Every professional activi~y--from medicine to
manufacturing--contains many problem solving tasks that are not
mathematical. The develcpment of artificial intelligence methods that
can solve many of these kinds of problems will change the nature of
applications of computers in cur society as much as the develcpment of
transistors changed the a architecture and applicability of computers
themselves.
133
Appendix: Annotates!
Typescript of a Sample
MYCIN Consultation
Instructions for Users at the Start of an
A. 1 Infectious Disease Consultation (Optional)
[Here, as elsewhere. tore responses hyped by the user follow a double-asterisk prompt.
MYCIN 3-Jun-78 . . .
Special options (type ? for help):
·.
Instructions? (y or N)
YES
This is a computer program named MYCIN that has been designed to advise you regarding an appropriate
choice of infectious disease therapy. it is understood that you have a patient who may have an infection. Please
answer the following questions, terminating each response with RETURN. To correct typing errors. use the
DELETE key to delete single characters, ~ctrl>W to delete a word, and Q to delete the whole line.
If you are not certain of your answer, you may modify the response by inserting a certainty factor (a number
from 1 to 10) in parentheses after your response. Absolute certainty (10) is assumed for every unmodified answer
It is likely that some of the following questions cannot be answered with certainty.
You may change an answer to a previous question in two ways. If the program is waiting for a response from
you (that is, has typed ), enter CHANGE followed by the number(s) of the question(s) whose answers will be
altered. You may also change a previous answer at any time (even when the program is not waiting for a
response from you) by typing cctr1>F (Fix), which will cause the program to interrupt its computation and ask
what you want to change. (If the response to F is not immediate. try typing the RETURN key in addition.)
Try to avoid going back because the process requires reconsidering the patient from the beginning and therefore
may be slow.
Note that you may also enter UNK (for UNKnown) if you do not know the answer to a question, ? if you wish to
see a more precise definition of the question or some examples of recognized responses. ?? if you want to see all
recognized responses, the word RULE if you would like to see the decision rule which has generated the question
being asked. the word WHY if you would like to see a more detailed explanation of the question. or the letters QA
if you would like to interrupt the consultation in order to ask questions regarding the decisions made so far In the
consultation. If you are ever puzzled about what options are available to you during a consultation. enter the word
HELP and a list of options will be listed for you.
Sample Response [user input follows the 1
Does the patient have a risk factor for tuberculosis?
134
Appendix: Annotated Typescript of a Sample MYCIN Consultation
One or more of the following are considered risk factors for TB: a) positive PPD (5TU). by history of close contact
with a person having active TB. c) household member with a past history of active TB. d) chest X-rav showing
apical scarring. e) granulomas seen on biopsy of any organ tissue.
expected responses are: YES NO
Ente! HELP for user options.
·- YES
SUMMARY:
(type ctrl-O to abort printout)
UNK - Answer not known.
? - Rephrases the question and gives examples of recognized responses.
?9 - Prints a list of all recognized responses.
Prints the current decision rule.
Program enters question-answering mode.
Go back and re-request answer to question number #.
Enter comments about any aspect of the system s performance. Your comments will be
forwarded to those in charge of the MYCIN program.
Gives high-level explanation of the current reasoning chain that provoked this question
Explains HOW the system wii achieve a goal referred to by number # in a previous
explanation.
Provides a more detailed explanation of a previous answer given by a WHY command
Resets the explanation of the reasoning chain back to lowest level. as if you never used the
WHY EXPLAIN commands.
Halts the program without completing the consultation.
Halts the program saving the current patient on a disk file retrievable at a later date.
Prints this list.
RULE -
QA -
CHANGE #
COMM ENT
WHY -
EXPLAIN
FORGET -
STOP -
SAVE
HELP
A.2 Infectious Disease Consultation Session
[This is a typescript of a MYCIN consu/tabon on an actua/ patient treated at Stanford MYCIN was no' used
to suggest treatment. Spaces have been added for readability. and comments (in italics) have been
inserted Otherwise it is a verbatim transcript. Users responses follow double asterisks
----PATIENT-538--------
1 ) Pat~ent s name:
2'
l
.,
PT538
Age:
34 YEARS
3) Sex:
MAEL
= MALE
4) Race:
WHY
[i.e. WHY is it important to determine the race of PATIENT-538?]
[Name age. sex. and race are always asked because they are INlTlALDA`-A j
[MYCIN corrects misspellings
[The user asks for an explanation In this Instance it is trivia!
[1.0] The race of a patient is one of the 5 standard parameters that identify
a patient.
This may be useful later in the consultation. For example. it Is
relevant In determining the organisms (other than those seen on cultures
or smears) which might be causing the Infection.
[back to question 4...l
CAUCASIAN
Please supply one answer under each heading. If your response Is not
fans is the answer to the quest/on about race
135
Appendix: Annotated Typescript of a Sample MYCIN Consultation
16) Please give the date on which clinical evidence (symptoms. signs. or
laboratory tests) of the meningitis first appeared.
29-Jan-77 22:15
17j Does Pt538 have an abnormal chest X-ray?
NO
18) Does Pt538 have one or more of the following risk factors for TB: a)
positive intermediate strength PPD b) history of close contact with a
person with active TB. c) a household member with a past history of
active TB d) apical scarring on chest X-ray. e) history of granulomas
NO
19) Does Pt538 have any cutaneous lesions or rash on physical examination?
NO
20) Does Pt538 have evidence of ocular nerve dysfunction?
NO
21) is Pt538 a compromised host (e.g. alcoholic. sickle-cell-disease.
immunosuppressed)?
YES
22) Is Pt538 ~mmunosuppressed (e 9 corticosteroid therapy. cytotoxic drug
therapy. radiation therapy. leukemia. Iymphoma)?
N<:)
23) For how many days has PtS38 had abnormal necrologic signs?
On biopsy of liver lymph node. or other organs?
7.2 HOURS
24) For how many days has Pt538 had necrologic symptoms?
~ DAYS
25) Has PtS38 had recent evidence of tender or enlarged salivary glands?
'' NC
26) Has Pt538 been exposed to any contagious disease recently (e.g.
men~ngococcal disease. mumps) ?
NO
iNote that the answer was given in hours.
MYCIN has to be flexible enough to convert un~ts.J
Please enter CSf findings In the following table
CSF Percent Protein CSF Peripheral
WBC PMN Glucose Glucose
27) 100 99 280 5 UNKNOWN
1Note that some information is missJng.j
28) Do you know the results of a recent CBC of PtS38?
·. YES
Please enter CE3C results in the following table
WBC Percent Percent
In 1000s PMN BANDS
29) 1.9 UNKNOWN UNKNOWN
30) Has Pt538 had any InJury or insult to. or defect in the central-
nervous-system (excluding neurosurgery)?
NO
31) Has PtS38 ever undergone any type of surgery?
- YES
32) Does PtS38 have a history of NEUROSURGERY?
' M0
33) Does PtS38 I've in a crowded environments
NO
34! Do you suspect recent alcoholic history In Pt538?
NO
35) is meningitis a hospital-acqu~red Infection?
WHY
[This is a more interesting question. requ/rmg presentation of the rule that caused question 35 to be
asked. Note that the first five clauses of the rule have been ~nvesngated and question 35 IS motivated
by the last clause. MYCIN first rephrases the question tO give the user its ,nterpretanon of the WHY|
36
Infectious Disease Consultation Session
(i e. WHY is it important to determine whether the Infection was acquired
while the patient was hospitalized?]
[2.01 This will aid in determining the organisms (other than those seen on
cultures or smears! which rn~gnt be causing the Infection
It has already been established that
Author: Yu
[RULE5453
[back to question 35.
HOW 2.3
[2 1 J the infection which requires therapy Is meningitis and
[2 21 organisms were not seen on the stain of this pending CSF culture.
and
[2 3] the type of the infection is bacterial. and
[2.41 neurosurgery is not one of the relevant items from the history
of Pt538. and
[2.5] the age of Pt538 is greater than 10 days
Therefore. if
[2.6] the infection was acquired while the patient was hospitalized
then
there is evidence that the organisms (other than those seen on
cultures or smears) which might be causing the Infention is e cold
( 75) staphylococcus-coag-pos ( 3) pseudomonas-aerug!nosa ( 3)
klebsiella-pneumoniae (.5)
{MYCIN now requests the answer to question 35 again.
Instead of answering. the user asks how clause 2.3 was confirmed.]
[i.e. HOW was it established that the type of the infection is bacterial?]
[Ramer than printing the contents of the rules. MYCIN shows the rule
numbers and the strength of positive and negative evidence from each rule.|
The following rules concluded about:
the likelihood that the type of the infection is bacterial
Rule Yes No
[2.71 RULE148 (.19)
[2.8] RULE500 (.51)
[2.91 RULE501 (.95)
[3.0] RULE502 (.97)
[3.1] RULE526 ( 98)
(3.2] RULE504 (.97)
[3.31 RULE524 ( 98)
Thus it has been established that the type of the infection is bacterial.
back to question 35...
HOW 2.5
[i.e. HOW was it established that the age of pt538 is greater than 10 days?
In answer to question 2 you said that the age of pt538 is 34 years
Thus it has been established that the age of pt538 is greater than 10 days.
[back to question 35. . .]
~~ y
36) Is Pt538 s clinical history consistent with EPIGLOTTITiS?
NO
37) Is Pt538s clinical history consistent with OTITIS-MEDIA?
~ NO
38) Has Pt538 ever undergone splenectomy?
NO
39) Is Pt538 a burn patient?
WHY
[Another explanation is reglJested.
137
Appendix: Annotated Typescript of ~ Sample MYCIN Consultation
[i.e. WHY is it important to determine whether PATIENT-538 has been seriously
bumed?l
4.01 This will aid in determining the organisms (other than those seen on
cultures or smears) which might be causing the infection.
It has already been established that
[4.1] the infection which requires therapy is meningitis and
[4.21 organisms were not seen on the stain of this pending CSF culture.
and
[4.3] the type of the infection is bacterial
Therefore. if
then
14.41 pt538 has been seriously burned
there is suggestive evidence (.5) that pseudomonas-aeruginosa is one
of the organisms (other than those seen on cultures or smears)
which might be causing the infection
Author: Yu
[RULE578]
[back to question 39. . .]
WHY [In this case the user asks another WHY question. backing up the subgoal tree.J
[i.e. WHY is it important to determine the organisms (other than those seen
on cultures or smears) which might be causing the infection?]
[5.0] ... in order to determine the therapeutic regimen of PATIENT-538.
If
[5.1 ] information has been gathered about organisms isolated from the
patient organisms noted on smears of the patient negative
cultures of the patient. suspected infections without
microbiological evidence. current drugs of the patient and prior drugs of the patient and
5.2] an attempt has been made to deduce the organisms which require
therapy and
[5.3] You have given consideration to organisms (other than those
seen in cultures or smears) which might be present and found
at least one organism which requires therapy.
then
determine the best therapy recommendation from among the drugs likely
to be effective against the significant organisms
tRULE092]
[back to question 39. . .]
~~ y
1MYCIN now prints its intermediate conclusions about the likely causes of the infection.]
I have found evidence (RULE545 RULE557) that treatment should cover for E.coli
causing meningitis.
I have found evidence (RULE545 RULE557 RULE578) that treatment should cover
for Pseudomonas-aerug~nosa causing meningitis.
I have found evidence (RULES45 RULE557) that treatment should cover for
Klebsielia-pneumoniae causing meningitis.
I have found evidence (RULE507) that treatment should cover for Diplococcus-
pneumoniae causing meningitis.
[Now MYCIN begins to ask questions related to appropriate theraPyJ
40) Does PtS38 have focal neurological signs?
UNKNOWN
Some questions are motivated by the organism under consideration. Thus
the following lines sometimes provide useful contextua/ information to a users
(Determining which drugs are desirable for use against the E.coli. . .]
[Determining which drugs are desirable for use against the Pseudomonas-
aerug~nosa. .
138
[Determining which drugs are desirable for use against the F
139
understood. you will be prompted by the standard questioning method.
(enter ? for normal prompting, for questions that are not applicable )
Please enter the results of positive cultures using the following table.
site culture date dated
number collected positive
5) DONE
Please enter the results of pending cultures in the following table.
site culture date stain
number collected examined
6) ' CSF 1234 30-Jan-77 YES
Infectious Disease Consultation Session
method
of collection
[urine,sputum only
[The next question asks for severe/ items in tabular form
since they are found together in the hospital record.
method ofl
collection
[urine,sputum only|
stain of morph! fMYClN follows up on the stain of
smear the smear and the morphology.!
[type DONE if no organism was seen on the smear
7)--DONE
[No information about stain or morphology is available. MYCIN now
follows up question 6, asking about other positive cuJtures.
site cul # date coil. smear coil. ~nothodl
examined?!
8)''DONE [There are no more positive cu/tures.1
Please enter negative cultures in the following table.
site culture dated method of
number collected collection
[urine,sputum
only]
9)--DONE
10) Do you suspect Pt538 may have an infection at a site from which you
have not obtained culture specimens?
NO
Please enter information on current antimicrobials.
drug route of dater
name administration started
11). CEPHALOTHIN IV 23-Jan-77
1 2) 'DONE
Please enter information on prior antimicrobials.
drug route of
name
1 3)--DONE
dater
administration started
Considering organisms (other than those isolated from cultures) which might
be print. . .]
14) Has Pt538 recently had symptoms of persistent headache or other
abnormal necrologic symptoms (dizziness, lethargy, etc.)?
YES
15) Has Pt538 recently had objective evidence of abnormal neurologic signs
(nuchal rigidity, coma, seizures, etc.) documented by physician
observation or examination?
~ YES
The CSF cultures will be considered to be associated with MENINGITIS.
[There are no negative cuitures.J
dater
stopped
140
Appendix: Annotated Typescript of a Sample MYCIN Consultation
RECOMMENDATION-2 is equivalent to flECOMMENDATION.1.
Would you like to see details of this comparison?
YES
Looking at the coverage of the 'most likely' items, we see that:
RECOMMENDATION-1 consists of 1 first choice drug (AMPICILLIN for ITEM-1
and ITEM - ) and 1 third choice drug (GENTAMICIN for ITEM-1, ITEM-2 and
ITEM-3)
RECOMMENDATION-2 consists of 1 first choice drug (PENICILLIN for ITEM-4)
and 1 third choice drug (GENTAMICIN for ITEM-2. ITEM-1 and ITEM-3)
ORGANISMS RECOMMENDATION-1 RECOMMENDATION-2
most likely' Drug -- Choice Ranking Drug -- Choice Ranking
ITEM-1 AMPICILLIN-AND-GENTAMICIN--1 st GENTAMICIN--3rd
ITEM-2 GENTAMICIN--2nd GENTAMICIN--2nd
ITEM-3 GENTAMICIN--3rd GENTAMICIN--3rd
ITEM-4 AMPICILLIN--1st PENICILLIN--1st
(The desirability of a drug is defined to be its lowest ranking for the items
it covers.)
Do you wish to see the next choice therapy?
' NO
141
RENGF~
Arxierson, J. L. and A. Cl~.
1985 ~nan~nter~ Space Station Design. Paper presented at
Conference on Cc - uters ~ Aerospace, Lord Beadh, CA,
October 21-23.
Bagman, B.
1984 MOP Thesis, University of Texas, Or.
Buchanan, B. G.
1986 Expert systems: working systems and the Church
literature. Expert Systems 3 (1) :32-51. Also appear as
S`canford REL Report 85-37 and SON CS-85-1075.
Han, Paul.
1986 ~ventor)~ and analysis of existing expert system. Expert
S:rst~; Strategies 2 (8): 1-16.
Miller, R. A., Pople, H. E. and M\rers, J. D.
1982 ~NTERNTST-1, Art Experiments ~ter-Based Diagnostic
Consultant for General Thternal Mectic~ne. N=M
307(8) :468-476. peso ahead; in 'Readings In Medical
Artificial Intelligence: The First Decade', Clingy. W. J
and Shortliffe, E. H., Ad., Wading, Ha, A~ison-Wesley,
Pp. 190-209.
Sheehan, G. M
1955 An application to payroll.
Data processing Conference.
Cambridge, He, September.
.
Feedings of the Autistic
Heard University ess,
