You are on page 1of 7

AI Magazine Volume 8 Number 3 (1987) (© AAAI)

A Graduate-Level
Expert Systems Course
David C. Brown

t seems that these days everyone introduction to expert systems, and


This article presents an approach to a
graduate-level course in expert, knowl-
I wants to take a course in expert
systems! The problem is how to teach
discuss some advanced topics. Most
systems are presented by the students,
edge-based, problem-solving systems. The it in order to make it anything more with a final lecture by me about my
core of the course, and this article, is a set than a look at a few applications of a research. The choice of systems is
of questions called a profile, that can be few AI techniques. intended to provide a view of some of
used to characterize and compare each
The Computer Science Department the most well-known, successful, and
system studied.
at Worcester Polytechnic Institute influential systems, yet it allows for a
offers a course called Expert, Knowl- contrast between different types of
edge-Based, Problem-Solving Systems problems and different approaches to
to graduate students who have already their solution. The systems also cover
taken the graduate-level introduction a wide range of claims about how
to AI. The course size is limited to 20. much they resemble human problem
It is run as a 14-week course, with solving.
one 3-hour class per week. There is some disagreement about
One goal of the course is to exam- which systems are in fact expert sys-
ine a number of expert, knowledge- tems. For example, some lists include
based, problem-solving systems, look- speech-understanding systems.
ing at each system in some depth. Although humans could be called
Another important goal is to make expert speakers, I don’t think that we
comparisons across systems in a would call them experts in the speech
domain-independent way. An attempt domain. I have tried to only include
is made to relate systems by their systems where the domain is such
problem-solving capabilities rather that it is learned as an adult and
than merely by the AI techniques where we would be inclined to
used. acknowledge that someone could be
Each student is set the goal of called an expert.
understanding at least one system in
detail. The student becomes the Teaching Method
expert on, and the representative of,
the system. The goal of improving the At the start of the term, the students
student’s presentation skills is satis- are given a package of readings that
fied by having each student make a provide background material for the
presentation about a particular sys- whole course as well as specific mate-
rial for the first two lectures. This
tem. Another goal is to have each
material includes an introduction to
member of the class contribute during
how the course is to be run and a
class discussions.
description of its goals. It also
includes surveys, comments on the
Course Content field, and technical papers.
The content of the course as currently The assignment of students to sys-
taught is given in table 1. The refer- tems is done during the first meeting.
ences are merely pointers to the liter- Wherever possible, it is done accord-
ature and are not intended to be com- ing to student’s requests because
plete. some students have appropriate
The first two lectures are used to domain knowledge, and others (part-
organize the course, present a general time students) find some link with

FALL 1987 33
BeProfile, MDX/PATREC
lomain. Medicine; especially a subset of liver PATREC also contains some anatomical global calculus and was in general found to be
diseases: CHOLESTASIS knowledge, temporal reasoning logic, and comparable (although it performed less well
tain General Function. Input and organization flexible input parsing knowledge than MYCIN when both were denied signifi-
of a patient’s medical data (PATREC), and Characterization of Output. MDX currently cant laboratory findings about the patient’s
diagnoses based on these data to determine produces as output a list of possible diseases, condition)
the cause(s) of cholestasis each of which may account for a particular set Phases. The first phase of operation is data
ystem Name. Diagnostic component: MDX of cholestasis-related symptoms The list is entry via PATREC The patient’s symptoms,
(Medical Diagnostic Expert); database compo- accompanied by the set of symptoms that history, and so forth are entered and organized
nent: PATREC (Patient Record] support each disease PATREC produces two in PATREC’s patient record Initial inferences
types of output: specific values and trends regarding the data may then be drawn Search
bates. Approximately 19781983; some ongoing
(such as white blood cell count), and logical is then initiated in MDX, starting with the
development root node Successor nodes are called in a
responses to queries, which can have values
.esearchers. B Chandrasekaran, F Gomez, S true, false, or unknown recursive establish/refine process in order tc
Mittal, J Smith, J Sticklen, T Bylander form a list of plausible diseases with support-
Characterization of Data. Input data is via
ocations. Laboratory for Artificial Intelligence PATREC, and consists of a set of medical data ing evidence Finally, this list is checked tc
Research, Ohio State University, Columbus, pertinent to the diagnosis of cholestasis (clini- see if all the abnormal data are explained
Ohio cal findings, lab tests, medical history, etc ), Subfunctions. MDX does not recommend thera-
anguage. MDX is written in LISP; PATREC is organized by major temporal episodes (e g “at py or perform other sub-functions outside 01
written in LISP and FRL An implementation admission”] These data may be incomplete or diagnosis
language (CSRL] for developing MDX-type unreliable in a variety of ways: (a] test results Use of Simulation or Analysis. No numerical
expert systems is also available may be lacking; (b) test results may be of simulation or analysis is performed
Iachine. MDX was initially implemented on a questionable validity; [c) the actual relation-
System/Ctonrol Implementation Architecture.
DEC-10, PATREC on a DEC-20 ship between symptoms and the disease(s)
MDX is based on a hierarchy of specialist:
which cause them may be unclear
#rief Summary. This system consists of two (active agents), each of which organizes a set
major components: MDX, the diagnostic com- Generic Tasks. PATREC uses knowledge-direct- of rules In principle these active agents coulc
ponent, and PATREC, the database compo- ed information passing; MDX’s main task is be processed in parallel and communicate vis
nent Each component is responsible for all classification, but hypothesis matching and a blackboard The current implementation is
knowledge and reasoning in its own area For abductive assembly play a major role in the serial
both MDX and PATREC, knowledge is dis- classification process
Characterization of the Structure Knowledge
tributed throughout a hierarchy of major con- Theoretical Commitment and Reality. Three Knowledge in MDX is grouped into a hierar
cepts relevant to either diagnosis or medical major claims are made about the theoretical thy of concepts that represent disease cate
data This hierarchy is called the conceptual underpinnings of MDX The first is that its gories at various levels of abstraction Each
structure In MDX, the concepts represent dis- conceptual structure represents one valid concept has local, highly specific knowledge
eases or disease categories The diagnostic model of how experts view knowledge in the about how the concept establishes itself ant
task is also distributed among these concepts domain, and that the hierarchical nature of how it refines its successor concepts Both tht
Each one solves a small part of the diagnostic this structure promotes focused problem solv- diagnostic knowledge base and the inference
problem by establishing itself (if possible) as a ing Several sources of information (textbooks engine are distributed among these concepts
partial explanation of the symptoms, and then on diagnosis, consulting human experts) sup- A similar approach is taken in PATREC
by refining itself with calls to its sub-concepts port this claim The second claim is that a where patient medical data is organized into :
to achieve a more specific diagnosis for the modular, local uncertainty calculus is more hierarchy of frames with attached inferencc
symptoms At many points during the diagno- realistic than a global one, since the same procedures
sis, MDX queries PATREC for information piece of knowledge may have different mean-
Characterization of Process Knowledge. The
PATREC answers by following links in its ing and/or diagnostic value in different con-
main effect of a concept’s knowledge is t(
own hierarchical organization of patient med- texts MDX is most definitely an attempt to
establish whether or not that concept is a rel
ical data to retrieve specific values or make simulate the method of human
evant component of the final diagnosis Thus
inferences from them as necessary diagnosticians, rather than simply to repro-
active knowledge serves either to confirm o
Mated Systems. RED is an expert system used duce their diagnoses A final claim is that
to exclude certain concepts (hypotheses], am
for red blood cell antibody identification in symbolic uncertainty measures represent a
to recommend which sub-concepts should bf
blood banks Its architecture and distributed logically valid method for combining compo-
tried next The net result is a classification
approach to problem solving are based on nents of evidence, a method which closely
<subset of symptoms> --> <partial explana
MDX, but RED has an abductive assembler resembles the reasoning of the experts them-
tion> The partial explanation would include
that produces a single best identification selves evidence for or against a given concept
MDX-MYCIN is a version of MDX written for Completeness. MDX currently performs diag-
Deep or Surface?. The knowledge in MDX i:
a sub-domain of MYCIN (meningitis), and it noses in the sub-class of liver diseases charac-
highly compiled, surface knowledge Thl
has been used to compare the performance of terized by cholestasis, and within this snb-
claim is made, however, that the organizatiot
MYCIN’s global uncertainty calculus to class it is fairly complete The possibility of
of knowledge into a conceptual structure cap
MDX’s local approach AUTO-MECH is a implementing diagnoses of other liver dis-
tures the result of all “relevant” deep know1
small expert system that tests the MDX eases exists In the current implementation,
edge, and that compiled knowledge is there
approach in a non-medical domain, and also MDX produces a list of likely diseases but
fore adequate for all diagnostic tasks withii
examines the usefulness of CSRL for imple- does not choose the “best” diagnosis from
the domain
menting expert systems of the MDX type among them Work has been done on an
abductive assembler that could perform such Search Space. The search space in PATREC i
lharacterization of Givens. MDX contains composed of all possible paths through th
knowledge about the logical relations a task One such assembler has been imple-
mented for RED (see “related systems”) MEDATA hierarchy from general data catega
between its classificatory concepts (partial ry to very specific data values The MD:
hypotheses) Each concept contains local Use and Performance. The system has not been search space consists of all possible path
knowledge about conditions under which it tested in an actual clinical or diagnostic set-
through the cholestasis hierarchy from gener
may establish itself and how to refine this ting It has been run on test cases drawn from al disease category to very specific hypothese
hypothesis by calling sub-concepts PATREC the medical literature, and it produced reason-
about the disease The search states ar
contains knowledge about the logical rela- able diagnoses in these cases The perfor-
explicit in the hierarchical knowledge repre
tions between its data frames and their mance of MDX’s local uncertainty calculus
sentation Search states represent successiv
attribute slots (inheritance, inference, etc ) was measured in comparison to MYCIN’s
refinements of a diagnosis

34 AI MAGAZINE
SpaceTraversal. In MDX, after a concept estab- hypothesis at some level of confidence, that accounts for some subset of symptoms with
lishes itself, it uses local knowledge to refine concept is rejected along with all its succes- an explicit degree of certainty To get the best
itself: its children are called to try to establish sors, resulting in a significant pruning effect diagnosis from this list of candidates, some
themselves, and this establish/refine process If the data are insufficient to establish or sort of abductive assembly is required MDX
continues until the diagnosis cannot be pur- reject a hypothesis, the concept suspends currently lacks this capability, although it has
sued further If a concept cannot establish itself If a complete line of reasoning fails, been implemented in RED (see “related sys-
itself, all its successors are ruled out, resulting control returns to the highest established con- tems”)
in considerable pruning of the search space In cept along this line, which then tries other Interaction. While the system is being devel-
PATREC, a global query processor analyzes possibilities If a normal search cannot oped, if any concepts are invoked which have
queries and directs search to one or more account for all the data, an exhaustive search not yet been implemented in detail, an inter-
patient record frames Local query knowledge through all likely possibilities may be tried active human expert can be “plugged in” to
associated with each frame may infer values Uncertainty. Both data and knowledge may be play the role of the missing concept MDX
by inheritance through the procedural trigger- uncertain, and as a result solutions to sub- and PATREC also use some interesting acqui-
ing of related searches, or sometimes from problems may also be uncertain sition interactions [see below)
default values Management of Uncertainty. MDX uses a fixed Data Collection and Acquisition. Typically all
Search Control Strategy. The control strategy is range of 7 certainty values, represented by the patient data are made available to PATREC at
called Establish-Refine This strategy may be integer range -3 [very unlikely) to +3 (very the outset of a diagnosis These data are made
expressed as follows: Consider a general dis- likely) These represent an ordered set of like- available to MDX on an as-needed basis, and
ease category that at least partially accounts lihoods; no numerical calculations are per- MDX can query the user for relevant data that
for the symptoms Then consider successively formed A local method of combining certain- are lacking in PATREC PATREC guides the
more specific diseases as causes, accounting ty values is used, and each concept uses its user in data acquisition by prompting for cate-
for more of the symptoms, until the most spe- own specific situation-dependent knowledge- gories of data (e g lab tests), and by requesting
cific cause(s) are reached This represents the based approach Certainty values are com- that all data from the same temporal episode
collective diagnosis (in medical jargon, the bined using expert specified tables, that asso- be entered at the same time The input mod-
“differential”] ciate a resulting certainty with each meaning- ule prompts for data in an order compatible
Standard Search Strategies. MDX can be consid- ful combination of the component values with the hierarchical organization of the
ered to use a form of the generate-and-test Not all combinations are represented Howev- patient data base
search strategy, except that the available er, it is likely that many of those not repre- Data Format. Data are input in the form of a
choices are always given in the concept hier- sented at a given node may have been ruled patient record The actual input uses a flexi-
archy The abductive assembly process imple- out in prior processing ble LISP-like syntax For example, the datum
mented in the related RED medical expert Management of Time. Much of the patient data “patient exhibits mild pain in the epigastrum”
system uses an explicit form of means-end is time-dependent: when certain symptoms can be input as (PAIN EPIGASTRUM MILD)
analysis began, what their trend has been, etc The or as [PAIN MILD EPIGASTRUM]
Search Control Characterization. Top-down, actual temporal dependencies are inferred by Learning. None
best first search driven by local knowledge In PATREC, which clusters temporal data into Explanation. When a concept is established, it
some cases, an established concept might con- “episodes” such as at-admission and at-
maintains a list of the symptoms that sup-
sider all its children if information about surgery PATREC then orders these episodes ported it The final diagnosis contains, for
which is most likely is not present In this as linearly as possible Imprecision of time each plausible disease, a list of all the symp-
case, the best-first search degenerates to depth estimates is also considered: three weeks may toms that supported all the concepts that led
first be represented as ([years O)(weeks 3)jdays to the inclusion of the disease The system
Subproblems. Concepts represent distributed unknown]), since 3 weeks might be anything can also supply a trace of its goals during exe-
problem solvers Each concept forms a partial from 18 to 24 days cution Because of the organization inherent
solution to the overall diagnostic problem Knowledge Representation Method. Mainly in the conceptual structure, this trace can
Concepts establish or reject themselves, so hierarchically organized clusters of local give the appearance of a causal explanation
each “knows” if it is on the right track or not rules, with some look-up tables, and with Strengths. Conceptual structure has a strong
Sub-concepts are largely independent of their local procedural knowledge to handle message
intuitive appeal as a valid model of how
siblings, but not of their parents This would passing between concepts experts organize domain knowledge. The hier-
not necessarily be true of any expanded imple- Knowledge Representation Generality. archical nature of this structure allows
mentation designed to deal with multiple dis- Although MDX was written in LISP, there is focused problem solving with significant
ease hierarchies now a special language, CSRL, for building pruning It permits local control of reasoning
Search Control Representation. Search control classification systems CSRL facilitates the and local uncertainty handling It facilitates
representation is largely explicit Some con- construction of a hierarchy of local specialists modular implementation and local enhance-
trol knowledge is explicit in the conceptual whose knowledge is represented in one or ment, and it can also lend itself readily to dis-
hierarchies of MDX and PATREC Some con- more “Knowledge Groups” (KG) A KG con- tributed processing techniques
trol knowledge is explicit in a concept’s local tains a set of rules or a table One group is
Weaknesses. Distributed knowledge and con-
production rules or in a frame’s procedural devoted to summarizing the other groups con- trol structures can be costly to implement
attachments Other control knowledge in clusions without tools such as CSRL, and, even with
PATREC’s global data acquisition and query Knowledge Structuring. Knowledge is organized such tools, global modifications may be diffi-
modules is more implicit as a hierarchy of concepts, each of which con- cult MDX’s use of compiled tables for
Search Control Strength. Methods are strong: tains specific, local knowledge about itself hypothesis matching or for combining uncer-
they are domain dependent (except for some This hierarchy is called the conceptual struc- tainties does not facilitate in-depth justifica-
of PATREC’s global query processing module), ture The claim is made that the hierarchy of tions of the inference process Systems which
and knowledge rich concepts should be similar or identical to the rely on hierarchical representation methods
Failure Method. PATREC rejects input data for way an expert would organize his knowledge may be forced to incorporate complex control
which it cannot find an appropriate frame or of the domain This hierarchy helps to focus mechanisms in order to handle some realistic
slot; it also responds “unknown” to a query if the problem solving process problems such as multiple diseases
the data are insufficient to infer an answer Alternative Structuring. The system doesn’t cur- Other. This is quite enough!
“Unknown” must be clearly distinguished rently use alternative representations for data
from “false” or default values In MDX, local This example profile was witten by Bill Sloan and
Alternative Solution Methods. None
knowledge is used to establish or reject indi- Doug Greenfor CS 525, Topicsin Computer Science.
Optimization and Multiple Results. MDX pro- Expert,Knowledge-Based,Problem-Solving Systems
vidual concepts (hypotheses) If data exclude a duces a list of diseases, each of which

FALL 1987 35
Waterman, and Lenat (1983). Howev-
er, it primarily comes from discus-
sions, seminars, and courses in which
I have been involved (see “Acknowl-
edgments”)
For several systems, some profile
questions are easy to answer; for other
systems, they are hard. Despite my
best intentions, some of the questions
overlap. Because the class discussion
contributes to the understanding of
each system, I think it is better that
the profile be completed after the class
in which the system is presented The
completed profile is submitted to the
instructor prior to the next class so
that copies can be made, to be dis-
tributed the following week. The col-
lection of completed profiles is a very
important benefit of this class.
The following section contains the
format for the profiles the students are
to provide. A sample profile of the
Table 1. Content of the Expert, Knowledge-Based, Problem Solving Systems Course. MDX/PATREC system, taken from
their work. Verbal ability is an impor- readings and is collected at the start of the course, follows at the end of this
tant factor in the course because the lecture in which the AIR- article.
every student has to make a presenta- CYL/DSPL system is presented. All
tion. Because it is often the case that other profiles are done with the bene- The Profile
some of the students are not native fit of having heard class discussion as The profile is divided into specific sec-
speakers, every attempt is made to well as the instructor’s comments and tions. Each section has a number of
pair these individuals with a compe- questions. headings that represent questions to
tent speaker. Each three-hour presen- Because much of the literature be asked The questions in each sec-
tation is made by one or two students. about expert, knowledge-based, prob- tion are related in some way Com-
After each presentation, the presen- lem-solving systems is hard to find, a ments or sample answers follow the
ter( s) providei s) a written profile of the collection of major papers is provided headings.
system. This completed profile is for each presenting person or group.
copied and then distributed at the Two weeks in advance of their presen- Domain. Chemistry, electronics,
start of the following class. The out- tation, the students are given a folder computers, and medicine (plus spe-
line of the profile (see the next sec- containing about 5 to 10 papers relat- cific subarea)
tion) is provided, thus ensuring a con- ing to their system. By limiting the
Main General Function. Diagnosis,
sistent format for the summaries. The access to material in this way, an
prediction, planning, and data inter-
concepts used in the profile make it attempt is made to equalize prepara-
pretation
essential that every student has tion time across the class.
already had a course in AI. System Name. AM, DENDRAL, and
The week before each presentation, Using a Profile MYCIN
a paper about this system is distribut- After the presentation, the specific Dates. Rough period of development
ed to the class as a reading assign- person or group is expected to provide Researchers. Main people responsible
ment. This preparation is essential if the class with a completed profile of for research and development (R&D)
the class is to fully absorb the details the system. The purpose of providing Location. Where R&D took place
of each system. In addition, it raises this profile is to delineate a substan- [can be more than one)
the quality of the in-class discussions. tial set of attributes by which systems
Preparation can be tested and, to a can be categorized and subsequently Main References. The clearest, as
certain extent, enforced by directly compared. This profile is also intend- well as the most comprehensive
asking students for their comments. ed to guide the form of each presenta- Language. Lisp, OPS5, KEE, Loops,
As an additional test of the stu- tion to the class, although this struc- and ART (can be more than one)
dents’ ability to evaluate systems, a ture does not need to be strictly Machine. Lisp Machine, DEC20, and
final profile about my research is adhered to. The profile was developed VAX
required from each class member. with the help of sources such as
Brief Summary. A short paragraph
This report is produced only from the Gevarter (1982) and Hayes-Roth,
describing the system

36 AI MAGAZINE
Related Systems. Systems from tion (for example, diagnosis], does it Space Traversal. How is the space
which this system was developed, use other types of problem solving traversed? What does it mean in
systems developed from this sys- as part of the system (see the terms of the problem to move from
tem, and systems by others using answers to the last question for one part of the space to another, for
same methods (clones) clues to whether this answer might example, subproblem decomposi-
be yes)? What kind(s)? For example, tion, gradual refinement, moving
diagnostic systems sometimes down a predetermined hierarchy,
Characterization of Givens. What is select therapy, too; this can be done and instantiation?
the information given and built into by synthesis or selection (which is
the system (expressed as abstractly Search-Control Strategy. Does the
often a kind of classification; that system use a strategy that is
as possible)? is, can this disease situation be clas- expressible in terms of the problem,
Characterization of Output. What is sified as one which this drug can or does it appear just to be an AI
the information produced by the help?). technique that happens to fit?
system (expressed as abstractly as Use of Simulation or Analysis. Does
possible)? Standard Search Strategies. Does the
the system use a numeric simula- system explicitly use generate and
Characterization of Data. Is the data tion or analysis, either done by test (G&T) or means-ends analysis
reliable? Is the data complete? itself or some package, during its (Be careful, almost everything is a
Generic Tasks. Which generic tasks operation? form of G&T!)? HOW would you
are obviously included, explicitly or System-Control Implementation characterize the search control, for
implicitly, for example, classifica- Architecture. What is the overall example, depth first, best first,
tion, state abstraction, knowledge- architecture? For example, breadth first, knowledge based, and
directed information passing, object metarules + rules, blackboard + random!
synthesis, hypothesis matching, knowledge sources, production Subproblems. Is evaluation of partial
abductive assembly, and simple rules, active agents, activation nets, solutions possible (that is, can the
selection? and so on. system know when it is on the right
Theoretical Commitment. Does the track?)? Are the subproblems inde-
system have any theoretical under- Characterization of the Structure pendent? Totally? Partially?
pinning? Is it claiming to show that Knowledge. IS it grouped into Search-Control Representation.
some theory of its type of problem types? What types, for example, Where is the search-control knowl-
solving is correct? IS the method component knowledge, chemical edge? How is it expressed? Explicit-
used claimed to work for other sim- knowledge, functional knowledge, ly?
ilar domains? and causal knowledge? What are the Search-Control Strength. Is it based
Reality. Is there any psychological types used for? Do they correspond on a very domain independent and
validity to the method used-the to phases of the system? knowledge-free method (a weak
structure of the knowledge, the con- Characterization of the Process method), or is it very domain depen-
trol mechanisms? Is it a system that Knowledge. For active knowledge, dent and knowledge full (a strong
is merely a simulation of result, or how would you characterize the method]?
is it in any way a simulation of effect of this knowledge, for exam-
ple, <partial situation descrip-
Failure Method. When part of the
Completeness. Has the system been tion>--> <classification> ?
system fails or reaches an incorrect
fully implemented? Has all the Deep or Surface. Is the system using conclusion, how does the system
domain been included? deep knowledge, or is it only using attempt recovery? What type of
Use. Has the system been used with surface knowledge? How would you knowledge does it use? Is it global
real users from outside the original characterize it? Are there levels of or local? IS there a significant varia-
development situation? Has the sys- representation or reasoning? For tion from the normal flow of con-
tem been used with real users in the example, if there is a qualitative trol? If it uses backtracking, what
user’s own working environment? simulation, it could be argued that kind does it use?
Are there any perfor- it is using deep knowledge.
Performance. Uncertainty. What is it that is uncer-
mance measures available? How tain? A piece of data? A piece of
was the system evaluated? How did Search Space. What space or spaces knowledge? A solution to a sub-
it fare? does the system search through? Is problem? If all are used, how do
the search explicit? Are the states they interact?
represented explicitly? How? What Management of Uncertainty. Does
Phases. IS the system organized into
distinct phases of different activity? do the states represent, for example, the system use probabilities, scor-
complete alternative solutions, ing values, a fixed range of certainty
Distinct subtasks? What are they?
solution refinements, and plans? values? What does a value mean?
Subfunctions. Despite the fact that How big is the space? What is the method of combina-
the system has a single main func-

FALL 1987 37
tion? IS it a global method, or does database or user) depending on the per week on the average, it actually
it vary (that is, a local method)? Are problem being solved? seems quite reasonable.
there any apparent problems? Data Format. In what form is the Because different people have differ-
data given, for example, patient ent presentation skills, some talks are
Management of Time. Is there any
records, Lisp, or natural language? much harder to understand than oth-
time-dependent data? How does it
ers. The reading distributed the week
affect the problem solving? Acquisition. Does the system have before helps to ensure that everyone
any way of acquiring knowledge understands the material. In addition,
Knowledge Representation Method. from the expert user? Does it guide the instructor should summarize the
What method(s) does it use to repre- the user? Does it have a way of vali- major points every so often during
sent knowledge? For example, rules, dating the knowledge? each presentation. Discussion can be
procedures, tables, semantic nets, Learning. Does the system learn prompted by asking students for sum-
and logic. from its own performance? If so, maries. In addition, comparison
Knowledge Representation Generali- how? What does it learn? between systems is facilitated by call-
Explanation. Does the system have ing on one of the student representa-
ty* Is there a special language for
this system, or does it use a general the ability to explain where its tives of some other system.
method? Is the method provided by result came from? Is this obtained If the presentation is poor as a result
some expert system building tool? from a trace of the goals and sub- of language problems (that is, a nonna-
goals formed during system execu- tive speaker), the instructor needs to
Knowledge Structuring. Is the sys- play an active role. However, presenta-
tem based on a hierarchy, a net- tion? If not, how is it formed?
tions can be improved by insisting
work, or some other structure? that the students prepare overheads of
Does this structure correspond to Strengths. What do you think are the major points and diagrams prior to the
the domain in some way? Does it strong points of the system? class.
correspond to the problem solving Weaknesses. What do you think are A problem that might occur in the
being carried out? the weak points of the system? future is students cheating by using
Other. Any other comments that profiles produced for a previous
Alternative Representations. Does help characterize this system. course. Because previously submitted
the system use alternative represen- profiles are not included in the folders
tations for the same piece of knowl- of technical papers, the students
edge in order to allow for alterna- would have to seek out the prior pre-
tive solution methods? Conclusions senters of a system. Cheating would
Alternative Solution Methods. Does require the cooperation of another
the system use alternative methods
to reach the same solution(s)?
T he students and I have found the
course to be very successful.
However, it is not without problems.
advanced graduate student, which is
possible but not likely. My view is
that students get out of a course an
Optimization. Does the system pro- One major criticism is that not amount proportional to what they put
duce the best answer? Sometimes? every student has the same amount of into it. If they cheat in the course,
Always? How? Why does it need to2 work to do. Those working by them- they cheat themselves, too. However,
Multiple Results. Does the system selves have to prepare a three-hour one way to combat the problem is to
produce more than one result? If it presentation, but others only prepare a gradually change the list of systems to
obtains several, does it try to order one-and-a-half-hour presentation. _This
. be examined.
them by evaluating them? Does it situation can be compensated for The course places some burden on
try to combine them? If so, how? somewhat by insisting that each the instructor. Of course, the instruc-
member of a group submit a separate tor should be able to answer all the
profile. Another factor, however, is profile questions for every system-if
Interaction. Is there anything inter- that some systems are harder to not before each presentation, at least
esting about the way the system understand than others, perhaps afterward! The only weekly work is to
interacts with the user (for both because of the complexity of the sys- read and grade the profiles submitted,
input and output), for example, by tem or the quality of a researcher’s
do the reading prior to class, and work
using menus, pictures, diagrams, writing. Systems that are perceived to out some key questions (the moral) in
picking, color, or sound? be more difficult than others can be advance. The major effort occurs
Data Collection. Does the system assigned a larger group. before the first class when you identi-
require all its data (that is, details of Because there are no exams, some fy and collect all the major papers on
this particular problem) before exe- students feel at the start of the class the chosen systems. Save a whole
cution, or does it allow or require that the course doesn’t entail much drawer of your filing cabinet!
incremental addition? If incremen- work. However, one student com- The benefit obtained by the stu-
tal, does it vary the type, number, plained that he spent 45 hours prepar- dents depends highly on the quality of
or order of data-gathering actions ing his talk and profile. Because this discussion, which in part depends on
(for example, questions of a amount works out to only three hours the quality of the students. However,

38 AI MAGAZINE
the instructor must play an active Expert Systems in the Micro-Electronic
part in stimulating and controlling Age, ed D. Michie, 3-25. Edinburgh: Edin-
discussion and must be prepared. burgh University Press.
Apart from the few problems men-
tioned, the comments from students
Friedland, P 1979. Knowledge-Based Exper-
iment Design in Molecular Genetics. Ph.D Spang Robinson
on their teaching evaluation forms diss., TR-CS-79-771, Dept. of Computer
Science, Stanford Univ The Spang Robinson Report
have been extremely positive. The
Gevarter, W B 1982 An Overview of
on AI
presentation approach using a profile
is an excellent way to provide an in- Expert Systems, Technical Report, NBSIR
depth comparison of expert systems 82-2505, National Bureau of Standards
for a graduate-level course. Hart, P E. 1982. Directions for AI in the The Spang Robinson
Eighties SIGARTNewsletter 79:11-16 Commercial AI Database
Hayes-Roth, F 1984 Knowledge-Based
Acknowledgments Expert Systems IEEE Computer 17(10): ARE PLEASED TO ANNOUNCE TBE
The author would like to acknowledge the 263-273 PUBLICATION OF AN IMPORTANT
intellectual stimulation that has been pro- Hayes-Roth, F; Waterman, D. A.; and ADVANCED COMPUTING BUSINESS
vided by members of the Laboratory for AI Lenat, D B. 1983 Building Expert NEWSLE’ITER:
Research at Ohio State University and Systems Reading, Mass : Addison-Wesley
members of the AI Research Group at Kulikowski, C., and Weiss, S 1982 Repre- THE SPANGROBINSON
Worcester Polytechnic Institute I would sentation of Expert Knowledge for Consul- REPORTON
also like to acknowledge Drs. Doug Green tation: The CASNET and EXPERT Pro-
and Bill Sloan for allowing me to use their
profile of MDX as an example
jects In AI in Medicine, ed P Szolovitz, Supercomputing and
21-55. Boulder, Co10 : Westview
Lenat, D., and Brown, J S 1983 Why AM Parallel Processing
References and Eurisko Appear to Work In Proceed-
ings of the Second National Conference on EDITOR-IN-CHIEF: Richard H.
Brown, D C., and Chandrasekaran, B Artificial Intelligence, 236 Menlo Park, Hill (512) 280-6706 -- formerly
1986. Knowledge and Control for a Calif.: American Association for Artificial with Honeywell and the Advanc-
Mechanical Design Expert System IEEE Intelligence. ed Computer Architecture Pro-
Computer 19(7):92-100
Lindsay, R.; Buchanan, B ; Feigenbaum, E. gram at the Microelectronics and
Buchanan, B , and Shortliffe, E H 1984 Rule A; and Lederberg, E 1980 Applications of Computer Technology Corpora-
Based Expert Systems Reading, Mass : Artificial Intelligence for Organic Chem- tion (MCC).
Addison-Wesley istry: The DENDRAL Project New York:
Chandrasekaran, B 1986 Generic Tasks in McGraw-Hill. COVERING: All significant
Knowledge-Based Reasoning: High-Level Marr, D. 1981 AI: A Personal View In applications in supercomputing
Building Blocks for Expert System Design Mind Design, ed. J. Haugeland, 129-142 and parallel processing: the
IEEE Expert 1(3):23-30 Cambridge, Mass : MIT Press. architectures, the theories,
Chandrasekaran, B., and Mittal, S 1983a McDermott, J 1982 RI: A Rule-Based product comparisons, the
Deep versus Compiled Knowledge Configurer of Computer Systems Artificial markets, the financial impli-
Approaches to Diagnostic Problem-Solv- Intelligence fournal 19( 1):39-88. cations, the companies, the
ing. International lournal of Man-Machine personalities, the potential, the
Nii, H. P; Feigenbaum, E A; Anton, J J.;
Studies 19:425-436
and Rockmore, A J 1982 Signal-to-Sym- limitations.
Chandrasekaran, B, and Mittal, S. 19831, bol Transformation: HASP/SIAP Case
Conceptual Representation of Medical Study AI Magazine 3(2):23-35 VOL. 1, NO. 1: September 1,
Knowledge for Diagnosis by Computer: 1987
Pople, H 1982. Heuristic Methods for
MDX and Related Systems. In Advances in
Imposing Structure on Ill-Structured Prob-
Computers, vol 22, ed M. Yovits, 217.
293. New York: Academic.
lems: The Structuring of Medical Diagno- FREQUENCY: Monthly
sis In AI in Medicine, ed. P Szolovitz, 119.
Davis, R 1982. Expert Systems: Where Are 190 Boulder, Co10 : Westview
We? And Where Do We Go From Here? AI INTRODUCTORYOFFER: Until
Stefik, M 1981 MOLGEN Artificial Intel- September 1, 1987, subscriptions
Magazine 3(2):3-22
ligence Iournal 16(2): 11 l-169.
Davis, R , and Lenat, D 1982 Knowledge- will be $200, a discount of $75
Based Systems in AI New York: Academ- off of the regular annual sub-
ic. scription rate of $275.
Duda, R 0, and Reboh, R. 1984 AI and
Decision Making: The PROSPECTOR To SUBSCRIBE: Contact Sly;;
Experience In AI Applications for Busi- Robinson, P.O. Box
ness, ed. W. Reitman, 111-147 Norwood, Manchester, MA 01944 USA
NJ : Ablex
(617) 526-4820
Feigenbaum, E A 1979. Themes and Case
Studies of Knowledge Engineering In
For free information, circle no. 33

FALL 1987 3S

You might also like