-lE93 FROM GUIDON TO NEOMYCIN AND IdERACLES IN TWENTY SINORT 1/i- LESSONS(U) STANFORD UNIV CA DEPT OF COMPUTER SCIENCE

Wd J CLANCEY JUL "E STAN-CS-87-ii72 NSORI4-95-K-GM8

UNCLASSIFIED F/C 5'9 U

EooliEmomhohiE

MICROCop'r RiSOLUTK3W TEST CHART!4AfTON AI J Of ST ATDAROT 19b3 A

DIC FILE COIJvI CARY Report No. STAN-CS4-172Also numbered KSL-86-11

In

From GUIDON to NEOMYCIN and HERACLESin Twenty Short Lessons

by

William J. Clancey

Department of Computer Science

Stanford UniversityStanford, CA 94305

DTIC,LCTE

*~~p~JUNJ~~ '~DECl1 4198~

GA H.5.

S Approved for pulic reI.w0*Distributon u.-amited

12~ 1

SECURITY CLASSIFICATION OF THIS PAGE |AM /4% 953REPORT DOCUMENTATION PAGE

Ia. REPORT SECURITY CLASSIFICATION lb RESTRICTIVE MARKINGS

UNCLASSIFIED _,2a. SECURITY CLASSIFICATION AUTHORITY 3. DISTRIBUTION /AVAILABILITY OF REPORT

APPROVED FOR PUBLIC RELEASE,2b. DECLASSIFICATION DOWNGRADING SCHEDULE DISTRIBUTION UNLIMITED

4, PERFORMING ORGANIZATION REPORT NUMBER(S) S. MONITORING ORGANIZATION REPORT NUMBER(S)

STAN-CS-87-1172 or KSL-86-11 ONR TECHNICAL REPORT #20

6& NAME OF PERFORMING ORGANIZATION 6b. OFFICE SYMBOL 7a. NAME OF MONITORING ORGANIZATIONSTANFORD KNOWLEDGE SYSTEMS (if applicable) PERSONNEL AND TRAINING RESEARCH PROGRAMS

LABORATORY I6c. ADDRESS (City, State, and ZIPCode) 7b. ADDRESS (City, State, and ZIPCode)

COMPUTER SCIENCE DEPARTMENT OFFICE OF NAVAL RESEARCH (CODE 1142PT)701 WELCH ROAD, BUILDING C 800 NORTH QUINCY STREETPALO ALTO, CA 94304 ARLINGTON, VA 22217-5000

Ba. NAME OF FUNDING/SPONSORING Sb. OFFICE SYMBOL 9. PROCUREMENT INSTRUMENT IDENTIFICATION NUMBERORGANIZATION (If applicable)

I _N00014-85K-03058c. ADDRESS (City, State, and ZIP Code) 10. SOURCE OF FUNDING NUMBERS

PROGRAM PROJECT TASK WORK UNITELEMENT NO. NO- NO. ACCESSION NO61153N RR04206 OC NR702-003

1 TITLE (Include Security Classiication)FROM GUIDON TO NEOMYCIN AND HERACLES IN TWENTY SHORT LESSONS

12 PERSONAL AUTHOR(S)WILLIAM J. CLANCEY

13a TYPE OF REPORT 13b. TIME COVERED 14. DATE OF REPORT (Year Month, Day) S. PAGE COUNTFINAL FROM TO JULY 1986 32

16 SUPPLEMENTARY NOTATIONKSL 86-11; APPEARED IN THE AI MAGAZINE, 7(3):40-60, CONFERENCE EDITION 1986

17 COSATI CODES IS SUBJECT TERMS (Continue on reverse if necessary and identify by block number)FIELD GROUP SUB-GROUP GUIDON, TUTORING, DIAGNOSIS, STRATEGY, RULE-BASED EXPERT05 09 SYSTEM, KNOWLEDGE ENGINEERING, MEDICAL INSTRUCTION,

I IKNOWLEDGE REPRESENTATION.19 ABSTRACT (Continue on reverse if necessary and identify by block number)

I review the research leading from the GUIDON rule-basedttlforinq system, i ncluding the reconfiguration of MYCIN into NEOMYCINand NEOMYCIN's generalization in the heuristic classification shell,IIERACLES. The presentation is organized chronologically aroundpictures and dialogues that represent conceptual turning points andcrystallize the basic ideas. My purpose is to collect the importantreslilts in one place, so they can be easily grasped. In theconclusion, I make some observations about our research methodology.

20 DISTRIBUTION/AVAILABILITY OF ABSTRACT 21 ABSTRACT SECURITY CLASSIFICATIONrUNCLASSIFIED/UNLIMITED § SAME AS RPT. Q oTIC USERS UNCLASSIFIED

22& NAME OF RESPONSIBLE INDIVIDUAL 22b. TELEPHONE (Include Area Code) 22c. OFFICE SYMBOLDR. SUSAN CHIPMAN (202) 696-4318 ONR 142PT

DD FORM 1473,84 MAR 83 APR edition may be used untI exhausted. -SECURITY CLASSIFICATION OF THIS PAGEAll other editions are obsolete.

" --' ~ ~ ~ ~ ~ ~ ~~~~~~~~~~~F ... ..."....... M'-' : "" I '

From GUIDON to NEOMYCIN and HERACLES

in Twenty Short Lessons

byWilliam J. aancey

Stanford Knowledge Systems LaboratoryDepartment of Computer Science

701 Welch Road, Building CPalo Alto, CA 9430

The studies reported here were supported (in part) by:

The Office of Naval Research \ ,Personnel and Training Research ProgramsPsychological Sciences DivisionContract No. N00014-85K-0305

AcoOssiOf orThe Josiah Macy, Jr. Foundation -

Grant No. B852005 NTIS GlRA&I

New York City D'iC TAB :0Unannounceded e C]__',,tt _justi f iear. iOL

The views and conclusions contained in this document are the authors'and should not be interpreted as necessarily representing the officialpolicies, either expressed or implied of the Office of Naval Research By

or the U.S. Government. - -

whol orin prt s ~~rmied or ay prpos oftheAvellabilltT CodeSApproved for public release: distribution unlimited. Reproduction in -- AVail and/owhole or in part is permited for any purpose of the United States 101st speolasl

Government. Ais

I I11I

i I

From GUIDON to NEOMYCIN and HERACLES in Twenty Short Lessons

William J. Clancey

Origins it knows about, but it is unrealistic for medical diagnosis

The idea of developing a tutoring program from the in general. Using NEOMYCIN, we can convey the moreMYCIN knowledge base was first described by Ted Short- complex processes of forming hypotheses, grouping com-

lifle (1974). In fact, it was the mixed-initiative dialogue of petitors discriminating among competitors and reasoningthe SCHOLAR teaching program (Carbonell, 1970) that causally. A second important idea in NEOMYCIN, distin-inspired Shortliffe to produce the consultation dialogue of guishing it from other expert systems, is that the inference

MYCIN. He conceived of it as a question-answer program procedure for diagnosis is represented in a well-structuredin SCHOLAR's style, using a semantic network of disease language, separate from the medical knowledge. This fa-knowledge. Shortly after I joined the MYCIN project in cilitates explanation and student modeling.early 1975, Bruce Buchanan and I decided that developing HERACLES is the generalization of NEOMYCIN,a tutoring program would be my thesis project standing for Heuristic Classification Shell (Clancey 1985a).

The GUIDON program was operational in early 1979 By analogy with EMYCIN, we might say that HERACLESThisevew G dONe pgra w oe iona in eDONay 1 . is "NEOMYCIN without the knowledge," but there is a big

aThis review describes the key ideas in GUIDON and the difference. We retain NEOMYCIN's diagnostic procedure;important developments of the following six years as re- itfereued We r edin ne w iagns.search continued under funding from the Office of Naval it is reused and adapted in new applications.Research (ON R), the Defense Advanced Research Projects GUIDON2 is a set of tutoring systems that work forAgeny (ARPA, ad th Ary Rsearh Istitte.The any HERACLES knowledge base; it is currently being de-Agency (DARPA), and the Army Research Institute. The veloped with NEOMYCIN. In the GUIDON2 family offirst three years were covered briefly in an earlier report veoged w NeoCin t oN familytof(Clancey & Buchanan 1982). In general, only publications programs, we are exploring different forms of student andfrom this project are cited; many other references appear teacher initiative (Clancey 1984a).in the cited publications. ;ep, - ', ,- -_, e.oJ -- A A- Guidon: "Transfer of Expertise"

Overview: Introduction to the Programs In GUIDON (See Figure 2), we held the MYCIN knowl-

Figure 1 shows the relationship between programs we edge base constant and considered the additional knowl-have constructed in the past six years, including MYCIN edge about teaching that would provide a good tutoringand EMYCIN, which served as the foundation. system. We were especially interested in teaching from dif-

The medical consultation system, MYCIN, was gener- ferent knowledge bases using one program. This excitingalized to EMYCIN (van Melle, 1979). The tutoring sys-tem, GUIDON, was designed to work with any EMYCINknowledge base (Clancey 1979a, Clancey & Letsinger 1984,Clancey 1982a). Abstract

NEOMYCIN, another medical diagnosis program, ex- I review the research leading from the GUIDON rule-basedpands MYCIN's disease knowledge to include competing tutoring system, including the reconfiguration of MYCIN intoalternatives, for example, diseases that might be confused NEOMYCIN and NEOMYCIN's generalization in the heuristicwith meningitis. This provides an opportunity for teaching classification shell, HERACLES. The presentation is organizeddiagnostic strategy. MYCIN's strategy of exhaustive, top- chronologically around pictures and dialogues that representdown refinement is sufficient for the small set of diseases conceptual turning points and crystallize the basic ideas. My

purpose is to collect the important results in one place, so theyBill Clancey is a senior research associate in the Stanford Knowledge can be easily grasped. In the conclusion, I make some obeerva-Systems Laboratory, 701 Welch Road, Building C, Palo Alto, CA tions about our research methodology.94304.

40 THE AI MAGAZINE August, 1986

Lam

Kflowlsdg (and wmf SysGUION

ym5,.'- Mvcs

NEOMYCIN 30 HERACLES )I GUIDON2

CASTER WATCH MANAGE

MYCIN - EMYCIN - GUIDON um St

PUF SACON

Figure 1. A Map Showing the Evolution of Research. Figure 2. Guidon Teaches from MYCIN's Knowl-NEOMYCIN, a reconfiguration of MYCIN, is also a medical edge Base. The MYCIN knowledge base, combined withconsultation program. Research from both programs follows an interpreter for applying rules and interacting with a user,a parallel path: generalizing the system into a knowledge sys- forms a consultation program. The same knowledge base is in-tem shell, applying the shell to develop knowledge systems in terpreted by teaching rules for interacting with a student in aother domains (PUFF, SACON, CASTER), and developing an case-method dialogue, constituting the GUIDON instructionalinstructional system compatible with any knowledge system de- program. MYCIN's rules are ranked, relating them to years ofveloped from the shell. This paper describes the key ideas in medical experience (for modeling the student and selecting newGUIDON and the upper path of research. material for the student to learn). Additional annotations indi-

cate subtype sad causal relations among rule clauses and relatethe rules to a general description of the infectious process which

idea was motivated by the EMYCIN design that allows is used by GUIDON to provide more concise explanations of

putting in a different knowledge base and carrying on a MYCIN's reasoning. Within a few limits, GUIDON can discuss

consultation in a different domain. GUIDON's teaching any case that MYCIN or any EMYCIN program can solve. Inke , it conventional computer-assisted instruction, a new program isknowledge is separate from the medical knowledge; so t written for each case.

is reusable and adaptable to new applications. This is a

significant advance over traditional computer-assisted in-struction methodology that requires writing a new pro-gram for each case to be discussed. We now have twokinds of generality: First, this tutor can discuss any case toring or t-rules, numbering about 200. We built the sys-that MYCIN can solve. Second, we can swap in a different tem very much like a traditional expert system, runningknowledge base and discuss cases in another domain. The cases and incrementally modifying the t-rules. When theseparation of the knowledge base from the procedures that progrpm said something inappropriate, we modified t-ruleinterpret it is the important idea. conditions to change when that kind of remark would oc-

cur. Similarly, GUIDON sometimes missed an opportunityDiscourse Procedures: to say something interesting. For example, if a fact can be

Alternative Dialogues and Transitions. inferred by definition, there is no need to go through a

Another successful aspect of GUIDON's design is the rep- long dialogue, gathering data and forming hypotheses and

resentation ofthe tutoring knowledge. This knowledge can so on; so we added t-rules to deal with this case, leading

be shown as a transition diagram, where each node rep- the program to give MYCIN's conclusion or to ask for the

resents a situation within a tutoring dialogue (See Figure student's conclusion, depending on the model of what the

3). The program has a list of rules for reasoning about student knows and the goals for the dialogue.

what to do at each step. For example, when GUIDON de- This works rather well, though it lacks a theoreticaltects that a goal under consideration has been determined foundation. Arbitrary strategies are encoded in the tutor-(from MYCIN's point of view), it selects from three al- ing rules. Building on the t-rule idea, Beverly Woolf hasternative transitions: presenting a conclusion, presenting added a hierarchical structure to the alternative dialogues,a summary, and asking the student to make a hypothesis couched in the terminology of discourse analysis. This rep-(Clancey 1979b). resents in a more principled way the choices the program

Each of these transitions is encoded by rules called tu- is making. (See Woolf & McDonald 1984.)

THE AI MAGAZINE August, 1986 41

RULE GOAL OAL: €VffFOf

recent - success GL DATAr t bock

oone-ts I\\ ; a pua

all-rules-fail 0 f i L / -- "

completed

present, values summary HYPOTHESIS, Oisaseve

/0 "present-values pseiA0 -ow

knw-b-efntinadvice NewsomAknown- by-definition Figure 4. A Student Model Constructed by the

E Overlay Technique. The student states that the organismsGOAL causing the infection might be Diplococcus, Psuedomonas,

present-values RULE or Neisseria. MYCIN rules that conclude about the organ-0{X. isms causing the infection are shown with associated patient

I related-rules data. For example, rule 507 states that if the patient is be-tween 15 and 55 years old, then Diplococcus and Neisseriaare organisms that therapy should cover. Circled values are

missing from the student's hypothesis (for example, E.cols) orwrongly stated (for example, Neisseria). Dotted lines lead

Figure 3. Dialogue Transition Diagram. Each node from rules the student probably did not use. m = evidencestands for a situation in a case-method dialogue, represented link that the tutor believes is unknown to the student. R andin GUIDON as a stylized procedure of ordered rules or a rule W = links to the right and wrong values the tutor believes areset, totalling 200 rules in 40 situations. For example, in pur- known by the student; ! = a unique explanation; the tutor

suing a goal explicitly agreed upon by student and program, knows of no other evidence at this time. ? = questionable, thethe student can request more case data. GUIDON can recog- tutor is uncertain about which evidence was considered by thenize the data as relevant to a subgoal, provide it as a set of student. For example, R? means that the student stated thisrelated information (block), or determine that there is no need value, it is correct, and more than one MYCIN rule suppliesto ask (perhaps because the requested data can be inferred evidence for it.from known information). Arrows that loop back indicate thata situation may occur iteratively or recursively. For example,several related rules might be presented after a given rule is dis- If MYCIN's rules argue for and against this hypothesis,cussed. The italicized labels indicate the basis for a transition: then the student might know the positive evidence, butEconomy, Domain logic, and Tutoring goals. not the negative evidence. GUIDON concludes in a simi-

lar way that Psesdomonas is believed by the student be-

cause the patient is burned-but not because of the whiteOverlay Model: Evaluating a Student Hypothesis blood count (WBC) or because the infection occurred in

Perhaps the most interesting reasoning in GUIDON in- the hospital (nosocomial) -given that the student didn't

volves evaluating a student's partial solution (See Figure mention the two other diagnoses associated with this ev-4) (Clancey 1979c) In this example, the student says idence (E.cohi and Klebsiella). It's a straightforward, log-

that the organisms causing the infection could be Diplo- for aindeing c e nlue and udin aupro

coccus, Pseudomonas, or Neissera. The program looks at for indexing how facts are concluded and used in a pro-

MYCIN's rules and sets up a consistent mapping. It uses gram.

double evidence, a history of interaction with the student, The Sacon Tutorial:and a measure of rule difficulty to construct a consistent Experimenting with other Knowledge Basesmodel.

For example, suppose the student mentions Neisserta. Figure 5 shows an excerpt from a dialogue with GUIDON

42 THE Al MAGAZINE August, 1986

I

using the SACON knowledge base. It graphically demon- indicates that the the age of the patient, 34, is not ev-strates the value of keeping GUIDON's tutoring knowl- idence for Neisseria. Yet, a rule in the knowledge baseedge general: The same tutoring program can interpret says, "If the age of the patient is between 15 and 55, thenSACON's rules and discuss a structural analysis problem Neisseria is one of the organisms." I was rather surprised.with a student. This rule is consistent with the student's hypothesis and

justification.We must now determine whether the shape of the floor sec- The problem is that some of the information in thetion (SUB-STRUCTURE-i) is one of: the surface shapes. premise of this rule is still unknown, so MYCIN can't ap-Here is some relevant data you could have asked for: the ply the rule. Specifically, there is no indication that the agegeometry of the floor section is planar, the modelling dimen- of the patient is causally related to Neisseria and that thesionality of the floor section is 2, and continuum is one of the age would be sufficient in itself to suggest this conclusion.constructions of the floor section. GUIDON has no way of knowing that one of the clauses

is more directly associated with the conclusion than anyWhat does this tell you about the shape of the floor section? other clause. To make this clear, consider another rule: "If** HELP the age of the patient is greater than 17 and the patientCan you conclude that the shape of the floor section is (select is an alcoholic, then Diplococcus might be causing the in-from the following): fection." Considering this rule and knowing only that the(1) beam-suggestive evidence (.62) patient is 34 would not make you think of Diplococcus.(2) semimonococque-suggestive evidence (.62) Here the age clause controls the application of the rule,(3) not shell--strongly suggestive evidence (-.95) preventing the program from asking whether a child is an(4) plate-with certainty (1.0) alcoholic. The causal relation is between alcoholism and** 4 Diplococcus.

Yes, that's right (see RULE100). The student's knowledge and the procedure being usedare very different from MYCIN. The student has probably

Figure 5: Excerpt of GUIDON Tutorial Using SACON formed a hypothesis just hearing the age of the patient andKnowledge Base. GUIDON encounters an incomplete sub- some tentative information (not shown in the excerpt) thatgoal in a rule that it's trying to discuss with the student. A suggests meningitis. MYCIN will only conclude Neisseriat-rule in the procedure for discussing an incomplete subgoal when, from its point of view, it has exhaustively consid-finds that the subgoal can be inferred by a definitional rule ered the evidence for meningitis and considered whetherand then invokes the procedure for discussing definitional rules, it is bacterial and so on. MYCIN does a top-down searchGUIDON gives the student new information (the geometry, di- through the set of diseases, but the student has "triggered"mensionality, and construction of the floor section), and then meningitis from just partial information, with no direct ev-asks him if he can now infer the shape of the floor section. The menit fr just p atial infm tion t nde vstudent asks for help, and GUIDON converts the question into idence for an infection or bacterial infection at all.

' ~ a multiple choice. Reasoning about the current problem state, To properly respond to the student, we would havetext generation, and quiz construction and evaluation are all to represent the association between age and Neisseriaaccomplished by general t-rules that were originally developed explicitly and separate out the search procedure. How-in the context of a medical diagnosis dialogue, ever, to recognize what strategy the student is following,

we'd have to encode a different strategy, expressing whyit makes sense to think about Neisseria just knowing the

This interaction plausibly captures some of the behav- age and some tentative evidence for meningitis. The very

ior we'd like to see in a teaching program. It was produced idea of a hypothesis is foreign to MYCIN.

entirely by t-rules that were written for medical examples The Idea of Strategy: From "tracing a parameter"and then just plugged into SACON. It took about an hour to "focusing on a hypothesis"to make it all work, with a few modifications to cope withsyntactic variations in SACON's rules. For further dis- Figure 7 and Figure 8 illustrate that at a certain levelcussion and an example from PUFF, see (Clancey 1979a, MYCIN's reasoning is arbitrary, lacking the focus on hy-Clancey 1982a). potheses we find in people. People group their questions

* logically; they don't jump around without reason. How-Inadequacy of Mycin: ever, MYCIN does not focus on a particular hypothesis as

Implicit, Nonpsychological Strategy it goes down through its (implicit) tree of diseases. Whenit considers types of meningitis or organisms, the types

We now consider the analysis that led to NEOMYCIN. are considered arbitrarily, based on the order in whichWhat problems arise in using MYCIN for teaching? rules were entered into the program. The program pro-

Figure 6 shows an excerpt from an experiment with ceeds systematically from infection to meningitis to bacte-GUIDON; this was a pivotal example for me. GUIDON

THE Al MAGAZINE August, 1986 43

What are the organisms which might be causing the infec- in Davis's original conception of metarules, is inade-tion? quate. The problem is that there is an implicit, undisci-

Splined mapping between medical knowledge and MYCIN's>>-ISRAMNNIII parameter-value language. For example, if MYCIN's dis-What facts about this case tell you that Neisseria-meningitidis paaeteralueplane. Fo ex ame s dis-is one of the organisms which might be causing the infection? eae wr a reral parameter (rtherthan by a general parameter called "coverfor" with or-

>> AGE ganisms as values), then the normal back-chaining process>> would make the reasoning focused. Thus, a strategy can- > The fact that the age of Pt538 is 34 years is not evidence be implicitly encoded in the relation between parametersfor the conclusion that Neisseria-meningitidis is one of the and their values. Recalling the age-alcoholism example, aorganisms which might be causing the infection, strategy is already implicitly coded in the ordering of ruleDo you have a question? clauses. Before metarules can be written to systematically

control domain rules according to a hypothesize-and-teststrategy, conventions must be established for distinguish-

Figure 6: Problematic Excerpt of GUIDON Tutorial ing between data and hypothesis parameters and consis-Using MYCIN Knowledge Base. The student offers to tently encoding causal and subtype relationships amongstate a partial solution (student input follows ">>"). The pro- them.gram responds by rephrasing the current topic as a question,"What are the organisms which might be causing the infec- The Tetracycline Rule:tion?" When the student says Neisseria, GUIDON checks and Structure, Strategy, and Supportfinds that MYCIN has made no conclusion at all up to thispoint. A t-rule prompts the student to justify his hypothesis. This brings us to about 1980 when I studied MYCIN's 400The student says that he is considering the age of the patient. rules to determine how they might be reconfigured for useIt's a lost cause for the student, however; whatever he says in teaching. Up to this point, in constructing GUIDON,next GUIDON will reply, "No, that's not sufficient," because only limited annotations had been added to the originalMYCIN has made no conclusion. At the final prompt, the stu- rule set. Now any change at all would be allowed.dent can review the available data and MYCIN's reasoning, if Early on, I developed a framework that turned out todesired. In fact, the student's hypothesis is reasonable, but be very useful in protocol analysis. In this framework, ex-GUIDON would need to know how MYCIN's rules are con- planations are analyzed according to knowledge roles, howstructed and a different model of reasoning to understand why knowledge is used in relation to other knowledge (Clanceythe student did something different. 1983a) (See Figure 9):

9 The heuristic rule: A relation between data and diag-noses or therapies

rial meningitis to organism, but the process is unorderedat each level of refinement with regard to children. This a tue bpingnoses, and therapiesis because the goals that MYCIN pursues are always moregeneral than the conclusions in the rules being applied. In * Strategy: The procedure for applying rules

order to teach a procedure to a student and to recognize e Support: The justification for ruleswhat the student is doing, we need a program that will As an example of structural knowledge, you mightdeliberately focus on particular diseases and that will be think of "SSS" when trying to remember this framework.able to articulate its focusing principle. To understand the different kinds of knowledge here,

This analysis of MYCIN was directly inspired by a suppose that you have a patient who is four years old.* study of strategy by Brown (Brown, Collins, & Harris You would probably say, "Well, I'm not going to prescribe

1977). He points out that a problem solver does not apply tetracycline." However, that's not a very good model ofalgebraic operators randomly when simplifying an equa- how a physician reasons; this would be a strange conclu-tion; there is some logic behind each choice, describing a sion to make right at the beginning. Just as there is aline of reasoning. Applying this analysis to MYCIN, I un- logic for requesting data and for focusing on hypotheses,derstood for the first time how a strategy reasons about there is a logic for making assertions given available infor-operators or problem-solving methods, focusing their ap- mation. All together, we call this logic the inference proce-plication. In MYCIN, a rule corresponds to an operator, dure. (In the literature, it is also called control knowledgeand problem solving involves some strategy for selecting or, more specifically, diagnostic strategy. To emphasize thewhich rule to apply. Specifically, diagnostic reasoning is reasoning about control alternatives, the term managerialusefully controlled by focusing data requests and hypoth- strategy or metastrategy is also used, particularly in theesis testing. education literature [Schoenfeld 1981]).

From this perspective, it can be seen that describ- By strategy, I mean the general goal that leads theing strategy only in terms of domain rule ordering, as physician to remember a heuristic rule. For example, when

44 THE Al MAGAZINE August, 1986

mahul

32) Does J. Smith have a history of NEUROSURGERY? STRATEGIC KNOWLEDGE

** NO33) Does J. Smith live in a crowded environment? eorp in.cin t

consider Conraidicatons&

* NO dosage sdkistmena.34) Do you suspect recent alcoholic history in J. Smith?** NO35) Is the meningitis a hospital-acquired infection?** YES36) Is J. Smith's clinical history consistent withEPIGLOTTITIS?

/ NO conraindications STRUCTURAL37) Is J. Smith's clinical history consistent with KNOWLEDGE

OTITIS-MEDIA?paltient infection -ecifiC

* NO factors

38) Has J. Smith ever undergone splenectomy?** NO39) Is J. Smith a burn patient? HEURISTIC RULE

** YES AGE< 7 -o NO TETRACYCLINE

Figure 7: Sequence of Data Requests from MYCINConsultation.

SUPPORTING KNOWLEGE

Chetlon mechanism: discoloraton

of growing bones A teem.

GOAL I10omM MM OUETON

E.COU - (RulS11) - 032 NEROSURGERY Figure 0. Analysis of Knowledge Relating to- NESSERA - (Atis.) - o3 CRowo MYCIN's Tetracycline Rule. The rule states, "If the

patient is less than seven years old, then remove tetracycline(Rueess-) - 34 ALcoouc from the list of drugs under consideration." Relation of age to

o0 4EUm. _E other contraindication factors (such as whether the patient isL_. (Rul9ss) - 038 SPLENECTOMY pregnant), justification for the rule, and time when it would be

covE5o) -- considered are relevant to explaining this rule, but are not rep-

__L ( -s64) - 036 OO resented in MYCIN. Making explicit this structural, support,H U (Rusas) 3- 04 EOLOrMmS and strategic knowledge enhances our ability to understand and

- 037 OlTlS-MEDA modify MYCIN.P EWo. - ( Ru~) - 039 BUR is it important to remember not to prescribe Tetracycline?

Obviously the physician must take this into accountFigure S. Relating MYCIN's Data Requests to Or- when prescribing therapy.

ganism Hypotheses. MYCIN's questions, shown in Figure By structural knowledge, I mean the relations by7, have been reordered according to the hypotheses that moti- which heuristic rules are indexed and subsequently con-vate them. For example, question 33 about living in a crowded trolled. In general, this involves categorizing the facts theenvironment is asked in order to apply rule 533, which con- rules use (for example, patient factors) and the facts theycludes Neisseria. All of the questions pertain to the samegoal - determining what organisms therapy should cover - conclude about (for example, therapies).but the rules conclude about different organisms. Neither By support knowledge, I mean the justification for the

[['S the sequence of rule applications nor the questions are sorted rule. Why wouldn't you prescribe tetracycline to someoneby organism. Questions 34 and 38 pertain to Diplococcus- who is less than seven years old? Here we have a chem-pneumoniae, with three intervening questions pertaining to ical process, a chelation mechanism, that results in theHemophilus-influenzae. Thus, in pursuing a goal, MYCIN's molecule binding to the growing teeth and bones, and areasoning is unfocused at the level of possible values for the social consideration that attests people don't want to havegoal, in this case organisms that might be causing the infec- discolored teeth. This is a very interesting justification be-tion. cause it shows that giving tetracycline might save

THE AI MAGAZINE August, 1986 45

II

the patient's life, even though it might have an undesir-'le We're talking at the top of infections, but beforeside effect. That's important to know if tetracycline is the we go down infections, are there any other thingsonly drug available. It's a nice example of why it's useful you can think of? The mistake you don't want toto know the justification of a rule-so you can violate the make is leaving out the important things on top.rule and know what the consequences will be. We repeatedly heard these general statements-move

Generally, this figure suggests a framework for under- down different diagnoses, ask general questions, don't leavestanding an expert's explanations. When I ask a physician out important things on top. These were the strategicwho is solving a problem, "Why did you ask that ques- gems-better than I could have expected-that would al-tion?" I classify the answer into one of these categories, low us to construct NEOMYCIN. Essentially, I saw theIf the physician tells me, "Well I'm not going to prescribe opportunity here for a program that would talk procedu-tetracycline because the age is less than seven," I am be- rally about these operations: Moving down different diag-ing told what assertions were made from given information noses, asking general questions, not leaving out important(that is, a heuristic rule). If I ask the physician why and things at the top. This procedure is separate from theI get an explanation having to do with chelation, then medical knowledge, describing how the medical knowledgeI'm being given the justification for the assertion (that is, is searched. That is, the statement of strategy does notsupport). If the physician says "This is just one of the con- directly mention domain terms; it is abstract.traindications I'm going to consider," then I'm being told In Beckett's explanations, we see regular switchingabout the organization of his knowledge, the categories back and forth between the concrete situation and a gen-used for focusing (that is, structure). Finally, if the physi- eralization:cian tells me when contraindications are considered andhow each type is considered, then I'm getting the infer- Ask it very generally, like "Have you had any ma-ence procedure (that is, strategy). I tried to consistently jor medical problems, or are you on any medica-apply this analysis when working with physicians, particu- tion?" Those types of general questions are impor-

jlarly to focus their explanations on strategy and avoid the hao s ry can ocause own. alytelo

bottomless pit of support explanations. hwso o a ou on

NEOMYCIN research focuses on representing strategy You have to think of some of the common things,and structure because this is the deficiency of GUIDON we but at the same time you have to think of somemost want to improve. We also sense that structure and of the serious things that may not be common.

staey r tthe top of a pyramid of knowledge and are What is a serious infection that can get in your

more limited in nature. A research effort focused on themtho?is attractive because this knowledge conceivably might be This last example shows most clearly my model of in-carefully and exhaustively explored. ference in NEOMYCIN.

The Beckett Tapes: An Articulate Teacher Refining the diagnosis and thinking of same of thecommon things, the physician looks into the domain model

In 1980, Reed Letsinger and I worked with Tim Beckett, and asks, "What is a serious infection that can get in theMN.D., who was recommended by Ted Shortliffe and who throat?" and "What are some of the common things thatturned out to be a rather fortunate choice. Beckett was could cause it?" This is how the metarules in NEOMYCINknown at Stanford for being a good teacher. He could work.articulate general principles for reasoning very well. He As confirmation of the potential effectiveness of Beck-didn't just say what it is you should ask about or what your ett's approach, we analyzed his best student's reasoning.conclusions should be-he was able to speak in general The student obviously followed the procedure Beckett ar-

*terms about how you should think. ticulated in class. Of course, not all students would nec-We taped interviews and classroom interactions, and essarily find Beckett's teaching approach to be useful, but

transcribed and studied them (Clancey 1984b). In one we had an existence proof and clear statements of at leastinteraction, Beckett interrupts a student who is examining one diagnostic procedure, so we wrote the approach down.a patient played by another student: About this time, we also had the first glimmer of how

When you ask these questions about whether gar- an explicit procedure could help a student learn relevant0~ .gling makes it better or worse, or whether it's bet- medical knowledge. When I had Beckett present problems

ter certain times of the day, are you thinking about to me, I often lacked the medical knowledge to carry outhow that's going to help you move down different the procedure. However, knowing the procedure, I founddiagnoses? ... ask a couple of general questions that I could ask reasonably intelligent questions: "I knowmaybe that could lead you into other areas to fol- I should be thinking about some of the serious and cam-low up on, rather than zeroing in. mon causes of this disease, but I don't know what they

Note the absence of medical terms in his strategic ad- are." This has evolved into our version of explanation-vice. Again: based learning (see The Situation-Specific Model: From

46 THE Al MAGAZINE August, 1986

a Diagnosis to an Exploration). We also applied the pro- this relation in the knowledge base. In this sense, the in-cedure to an analysis of Beckett's interruptions of students: ference procedure is interpreting the domain model. If weGiven this modt. of his reasoning, could we use it to infer compiled the procedure- instantiating and composing ithis strategy for interrupting students and providing assis- with respect to a particular knowledge base- we wouldtance? The most telling example, occurring just before get something very similar to MYCIN's rules In makingBeckett asks the question about sore throats shown here, this abstraction, stating these general rules. I'm not claim-is analyzed in (Clancey 1984c). ing that people reason through general statements every

time or even realize that these patterns exist. In partic-Neomycin: Separating the Medical Knowledge ular, reasoning categorically probably involves automatic

from the Diagnostic Procedure processes of memory. Some distinctions, such as consider-

Figure 10 shows the architecture of NEOMYCIN, illus- ing causal prerequisites of diseases before effects, might be

trating the idea of separating the diagnostic inference regularities that the physician does not consciously realize

procedure (control knowledge) from the medical knowl- (Clancey 1984c).

edge. Crucially, both are represented in well-structured I now believe that these domain relations are in largelanguages so that they can be reasoned about by the expla- part what we want to teach students, as generalizations,to help them learn about new diseases. In describing hownation, knowledge-acquisition, and student-modeling pro- to focus reasoning, we are indirectly saying how knowl-grams (Clancey 1983b). edge should be practically organized. For example, we

say, "You should think in terms of common causes and se-(FOLLOW-UP-OUAS'noN headache SM4G) rious causes." That is much more informative than saying,

"You should form a hypothesis" or "You should reasonforward." We hypothesize that the procedure is automatic

nce you have the knowledge. A medical student mightDmas macanot have to be told to refine hypotheses, but he' has to be

NW&,od n a databes) / Pmced" taught the subtypes of fungal meningitis.

The Disease Taxonomy:Searching an Abnormal Process Classification

(CAUSED-BY diniocia SH-YPOTHESIS) I

There are several dimensions for describing NEOMYCIN'sFigure 10. Architecture of NEOMYCZN. An infer- reasoning: psychological aspects of memory and attention,

4% ence procedure queries the knowledge base, relating findings Al representation and control techniques, and aspects ofand hypotheses to one another in order to make a diagnosis. medical causal reasoning. Figure 11 provides one perspec-For example, given that the patient has diplopia (double vi- tive in which these dimensions come together.sion), the program asks the knowledge base what could cause The main part of the knowledge base is a taxonomyit. One or more hypotheses might be returned, which the in- of diseases or, more generally, a classification of abnormalference procedure will proceed to discriminate, test, and refine, processes. Each disease describes a process, somethingmaking further inquiries about disease and symptom relations. that has happened to the patient in the past, accounting

for the set of observed manifestations. In general, therecan be many different taxonomies, orthogonal and tangled.

Davis's conception of metarules for expressing strategy How do we know that a given taxonomy is complete?inspired this design. However, TEIRESIAS's metarules This important question did not explicitly arise in MYCIN

1 compose domain facts with procedure, just like MYCIN's research because we didn't isolate the disease taxonomy asrules (Clancey 1983a). NEOMYCIN's metarules mention a separate object of study. We now hypothesize that theno domain terms. Moreover, they constitute a coherent physician's diagnostic classification, particularly its level ofprocedure that completely controls every data request and specificity, depends on how it will be used. The physicianevery inference; so there is no back chaining of rules at all. is not involved in scientific research here; what goes into

As is apparent in Beckett's generalizations, we can the taxonomy is based on distinctions useful for selecting9'" think of this procedure as "asking questions of the domain therapy. For example, NEOMYCIN makes no attempt to

model." The language of relations used in metarules corre- determine precisely which type of viral meningitis the pa-sponds to the propositions in the knowledge base. These tient has. The reason is that they're all treated the same-relations impose a classification on domain terms. This with a lot of aspirin and orange juice-and it is irrelevant

% is what I called structural knowledge in the tetracycline to resolve the cause any further. Thus, NEOMYCIN'sanalysis.

Given a hypothesis, the program asks, "What is a com- I Masculine expressions in this article are used as generic terms. No

mon cause of this disorder"" The program then looks up bias is intended.

TlE Al MAGAZINE August, 1986 4704

a program that can understand arbitrary LISP code. Too" 'much of the design is implicit and not available for expla-

nation. Therefore, we devised a highly structured repre-.. sentation, organized around the idea of rule sets, with ev-

ery "loop" encoded as a separate task (subprocedure) andi, a Ct1, aT,, the controi of rules stated declaratively (simple vs. itera-

.. tive, try-all versus stop-on-success). Each task has a typed-".. - "focus (argument.), local variables, and an explicit "end con-

dition" (equivalent to the "while" or "until" condition of aloop). Making every program statement a rule facilitates

.... interpreted control, annotation, and record keeping.

The overall design is similar to LOOPS, which*, evolved at the same time as NEOMYCIN. However,

NEOMYCIN's metarules use variables, rather than do-.. . main terms. Also, the end condition, inherited by task

E ...... /a W"F * invocation, enables a procedure anywhere on the currentstack to regain control, either because its goal is completed

/- or there is reason to reconsider how its subgoals are beingaccomplished. Figure 13 shows the flow of control in terms

A of focus changes.Figure 11. Looking Up and Looking Down in Di- In writing down the diagnostic procedure as rules, we

agnostic Search. Disease knowledge is represented as a are following the same methodology used in developing. taxonomy of processes. At the highest level are internal aber- MYCIN and GUIDON. With the knowledge expressed in a

rations in structure building or maintenance (for example, con- disciplined way, it now becomes possible to study patternsgenital diseases) and processes involving environmental interac- and to consider how the knowledge could be derived. Suchtion (for example, infection, trauma). Processes are specializedhere by location, temporal extent, and specific agent. The tax-onomy is overprinted to show hypothetically how it might be interested reader will find the metarules listed in (Clanceysearched. Initial information--chief complaints-triggers some 1984c), with a discussion of the procedure in terms of op-hypothesis, shown arbitrarily here in the middle of the dis- erators and the cognitive, social, mathematical, and caseease taxonomy. Two operations follow: (1) looking up, think- population constraints implicit in the rules. The next sec-ing of the high-level categories and discriminating among them tion considers the procedure as a grammar.(GROUP-AND-DIFFERENTIATE) and (2) looking down torefine hypotheses when distinctions are important for selectingtherapy (EXPLORE-AND-REFINE). This is to be contrasted Image and Odysseus:with an exhaustive, top-down search, which a large knowledge Parsing the diagnostic processbase makes impractical. Given the abstract nature of the tasks and metarules, they

can be viewed as a kind of grammar for parsing a problemsolver's sequence of requests for data. Such an analysis is

disease taxonomy deliberately remains a partial model of shown in Figure 14, the picture I had in the back of myabnormal processes within this area of medicine, mind in about 1980 when I wanted some way for GUIDON

SAnother part of the knowledge base, the causal to reason about what a student was doing. An interpreta-network, is discussed in the context of CASTER (See tion of a student's partial solution provides a good basis forCASTER: From Disease to Abnormal Substances and Pro- assisting him when he doesn't know what do next. Suchcesses.) an interpretation is also a source of information for relat-

The Diagnostic Procedure: ing a student's explicitly stated diagnosis to a model of hisSearchpeanosanroeura s domain knowledge. As a contextual analysis, it potentially

Search Operators and Constraints shortens the interactive dialogue that might be necessary

The overall diagnostic strategy or inference procedure is a to confirm the student's understanding.program consisting of a set of subprocedures as shown in Bob London, David Wilkins, and I have been devel-Figure 12. oping student-modeling programs with the common goal

Each procedure is represented as a set of ordered and of using NEOMYCIN's diagnostic procedure to interpretcontrolled conditional statements called metarules. Rules a sequence of student requests for data. London has fol-provide a uniform, well-structured language. Although ex- lowed a top-down approach in the IMAGE program (Lon-perienced programmers can read a LISP encoding of the don & Clancey, 1982); Wilkins's ODYSSEUS program usesdiagnostic procedure easily enough, it is difficult to write exhaustive, bottom-up reasoning (Wilkins, Buchanan, &

48 TIlE Al MAGAZINE August, 1986

ASI -ASERAN- -U$I

~Figure 12. Invocation of Diagnostic Tasks, Shown as a Lattice. Each task is represented in NEOMYCIN as a stylized

procedure, shown here as a node, with the subprocedures it calls below it. For example, pursuing a hypothesis involves testingand refining the hypothesis. To relate new findings and hypotheses, all tasks eventually call FORWARD-REASON, which invokesadditional tasks not shown here. GENERATE-QUESTIONS is invoked when there is insufficient information to proceed; it chasesdown leads in different ways, thus explaining its central position. Note also that FINDOUT calls TEST-HYPOTHESIS so thatdomain rules will be selected deliberately, replacing the back chaining of EMYCIN. Using this representation for explanationand student modeling requires additional knowledge about task preconditions and postconditions and how metarues controllingtask invocation are ordered.

Clancey 1984). Evaluation of these alternative approaches is available prosaically (by asking WHY) or through theis in progress, task stack.

Figure 14 shows a parse of reasoning produced by theODYSSEUS program. We're testing this program with"synthetic" students, systematically varying NEOMYCIN Although our WHY/HOW system goes up the goaland comparing ODYSSEUS's interpretation to the known stack in a way similar to MYCIN's explanation program,variations in the knowledge base. Another application is this nwpormtakes avngeof the structured rp

newpogrm avanagerep

-: to give ODYSSEUS a sequence of data requests and to resentation to be more selective about what it says. Inhave it determine what knowledge base changes would be particular, it looks at the focus of a task to determinerequired to produce this sequence, consistent with the in- whether to mention the task as it goes up the stack. Aference procedure. We believe that the simple classifica- focus can be one of three basic terms- a finding, hy-tion nature of the inference procedure makes this approach pothesis, or a domain rule--or a list of these. If the focusplausible. We're developing this capability for a tutoring is a rule or list of rules, the explanation program skips overprogram called GUIDON-DEBUG (Clancey, el ai. 1986). the task (for example, APPLYRULES). The task is men-The same program could be used for knowledge acquisi- tioned if its metarule establishes a new focus, such as going

n tion. from a list of hypotheses to a single hypothesis (GROUP-AND-DIFFERENTIATE) or from a hypothesis to a rule

"yt Neoxpl: Strategic Explanation (TEST-HYPOTHESIS). This turns out to be a goodUsing NEOMYCIN's well-structured representation, Di- planation heuristic. A new explanation system under de-

arane Hasling, Glenn kennels, and I (1983) reformulated velopment uses the propositional encoding of the metarulesMYCIN's WHY/HOW explanations in terms of metarules (described later) to select particular rule-premise clausesand tasks. Figure 15 shows how procedural information to mention.

r cTHE Al MAGAZINE August, 1986 49

Th0aepormcudb sdfrkoldeaqii indi t eaueetbihsanwfcs uha on

(Clancey & Bock 1982). Unfortunately, recoding the inter-owse-o. preter slowed down the program by an order of magnitude

E. .. -O- and made the procedure too obscure to read or maintain.&GI.Ff wowIn the current version of the program, we retain the origi-

S1 ") nal interpreter and use a variant of MRS as a specification> .How."" cw-o(V,' or ~.e' language for metarule premises, which are compiled into

7VUoNIn-u'.es Lisp. This provides the well-structured, uniform languageour modeling and explanation programs require withoutsacrificing runtime efficiency. Figure 16 illustrates howMRS is used in the metarules and definitional rules forrelations.

FocvFmW Primitive relations are compiled as direct LISP oper-Mations, using explicit declarations about how propositions

are represented in the LISP-encoded knowledge base. For

Figure 13. Predominant Focus Shifts in Diagno- example, a TYPE proposition is represented as a propertysis. This diagram simplifies the dynamic flow of control be- list structure, so the compiler substitutes a GETPROP, antween tasks, revealing how findings and hypotheses are related. ASSOC, or more complex loop construction, depending onNew findings suggest new hypotheses and support existing hy- what terms are known when the proposition is encoun-potheses (FORWARD-REASON); a decision is made to focus tered in the metarule. In encoding propositions in stan-on a particular HYPOTHESIS (ESTABLISH-HYPOTHESIS- dardized LISP structures, distinguishing between the Ian-SPACE); a decision is made to focus on a particular finding guage for expressing knowledge and how it is stored in the(TEST-HYPOTHESIS); the implications of the new informa- computer, we are exploiting the multiple representationtion are considered, and so on. In contrast, MYCIN does not aspect of MRS, which is one interpretation of its name.change its goals on the basis of new data or deliberately or- A number of elegant patterns in the metarules made theder the goals and data it will pursue. Accomplishing this by compiler easy to write (Clancey, forthcoming). Figure 17abstract metarules (not specifying domain terms) requires ex- summarizes how rules, tasks, and relations are encodedplicitly representing relations between findings and hypotheses,on the basis of which they will be selectively considered, as EMYCIN rules and parameters and how these entities

are related. Our success in building HERACLES on topTasks (appearing in bold italics) can be related to Figure of EMYCIN demonstrates the generality of the original12, which shows the subtasks they invoke. In practice, parameter-rule representation language. It is closer to aESTABLISH-HYPOTHESIS-SPACE is only invoked if there typical frame language than is commonly realized.is reason to stop pursuing the current HYPOTHESIS. Criteria The most exciting result of this reformulation is whatfor applying domain rules in FORWARD-REASON are com- it reveals about the relational nature of the knowledgeplex. For example, new findings are related to hypotheses "infocus"; if a new HYPOTHESIS "explains" the known findings base. It is now evident that the metarules are selecting fociat least as well as existing hypotheses, it is considered; new (findings, hypotheses, domain rules) on the basis of howhypotheses are related to previously known findings, etc. The they are related to one another. These relations can beprogram stops when its differential, the list of most-specific hy- either static (for example, red-flag finding, one that needspotheses under consideration, has been discriminated, tested, to be explained) or dynamic (for example, hypothesis inand refined. focus). The knowledge base can be viewed as a database,

defined in terms of these three primitive terms and rela-tions among them. Writing a new metarule tends to re-

MRS/Neomycin: quire defining a new preference relation for discriminatingFrom Findings and Hypotheses to Relations among findings, hypotheses, and domain rules. That is,

Student modeling, debugging, and explanation require each new relation further classifies the primitive terms inthat our programs reason about the premises of metarules, a way useful for controlling reasoning. For example, theparticularly to determine which domain facts matched metarule shown in Figure 16 required the new relation,ariuy ru lesf ie Ori l mare prmises "a finding that needs to be explained." As this exampleSw ta n d w h y ru les fa iled . O rig in a lly , m eta ru le p rem ises sh w , t e m a i g o a r l t o n s t ed o h w t e e &were encoded in LISP. In a hybrid system called shows, the meaning of a relation is tied to how the rela-MRS/NEOMYCIN, Conrad Bock and I rerepresented tion is used. This is particularly clear for relations such asmetarule premises in MRS, a logic-programming language follow-up question and trigger rule.that provides a framework for multiple representations of A detailed analysis shows that the metarules are col-knowledge and control of reasoning (Genesereth & Smith lecting, sorting, and filtering domain terms and rules on1982). Bock also recoded the interpreter in MRS rules, the basis of their applicability as operands (foci) for theand placed a simple deliberation-action loop at the top operators (subtasks) that will accomplish the current task.

50 TIE Al MAGAZINE August, 1986

S

.com-~t(mSu-aa&-imnmn-h~omL306TI11WCHhCT-flMff[CuVUJ-h~qffiH3U)

,qI 'A54fIEF3qI-1,PWiTR0aIMAL A"tptU239

THHWWAATERl W T5JIDM[(4

TnqWAV-ELrCn]-Anqf7?si 4["F4B9JRAPA1A

lr llyl l l011 - ll~lAlll f1? # J- N(iT RIaI ASLO T / !81URE ]-J--AR1(N( 3 33IJ

tTHME5IJ-i[O~ 45 TF I"A RA"-#fin4MX - hrI!(RUa H303A]UTA hA&FKUIMIO~A -AI33)

T~O~f-cFrInqOHR~cws-Cfr4su3OI/-.1AlqE0

Figure 14. ODYSSEUS's Parse of a Student's Data Requests. Given a sequence of requests for patient data,Q listed on the right side of the figure (Q5, Q6, Q7), the program indicates all of the alternative justifications for why a question~might have been asked. For example, the student's query about seizures (FINDOUT/Seizures, Q6) might have been asked to

determine whether the disease is caused by an Intracranial Mass Lesion, Subarachnoid Hemorrhage, and so on. The programindicates in inverse video, combining its bottom-up analysis with a top-down parse, that this question relates to meningitis

*(TEST-HYPOTHESIS/Meningitis), as part ofteprocess ofdiscriminating hypotheses (GROUP-AND-DIFFERENTIATE).

OO[O~ -KOM ofTN-0F0A EM

w Thus, the problem state (hypotheses under consideration) arid the tasks interact to explain finding requests in terms of a logic forfocusing on hypotheses and findings. Note that by the same anaysis the question about a fever (Febrile, Q5) has three consistent

~interpretations. This kind of analysis is not possible using MYCIN because, first, its reasoning did not involve "looking up"from "triggered hypotheses" and, second, its inference procedure is represented behaviorally as specific productions. A functionalrepresentation, as diagnostic tasks, relates surface behavior to abstract goals, which can be accomplished in multiple ways.

SFor example, metarules for TEST-HYPOTHtESIS col- Guidon-Watch: Reifying the Processlect, sort, and filter potential findings to support a hypoth- The availability of graphics has changed how we can illus-

esis. trate reasoning and is shaping our ideas of what we'd likeRefining a hypothesis means collecting, sorting, and to show. As a first step toward implementing a new in-

• wt filtering its causes and subtypes (for example, distin- structional program on top of NEOMYCIN, Mark Richer: g. guishing between common and serious causes). Gener- and I (1985) used the Interlisp-D window and menu fea--_ ally, the domain relations classify NEOMYCIN's experien- tures to construct a complex interactive system for brows-

tilknowledge ofpredefined dies oes(see Hacs: ing the knowledge base and watching reasoning. ThisinFrom Diseases to Stereotypes) according to how they are cludes the dynamic task tree (similar to Figure 14) and thetriggered, tested, discriminated, and refined by operators task stack (see Figure 15). Our work has been directly in-(tasks) for constructing a problem-specific, historical ac- spired by Brown's emphasis on reifying or making concrete

. counting of the disease process (see The Situation Specific the reasoning process (Brown 1983).Model: From a Diagnosis to an Explanation). Figure 18 shows how the disease taxonomy is over-

NEOMYCIN has about 170 relations in its control printed to reveal the pattern of NEOMYCIN 's reasoning.vocabulary. They appear in the 75 metarules, grouped In GUIDON-DEBUG, now under development, it is possi-into 29 tasks. In HERACLES, the generalization of ble to roll back the consultation display to show any win-NEOMYCIN, the knowledge engineer can modify these dow at the time any given question was asked. This is ametarules, defining new relations for describing his do- debugging facility we could hardly have imagined even five

- main. years ago.

THE AI MAGAZINE August, 1986 51

0, Pr" AA tIP'

" '

7) What is Susanne's temperature (in Fahrenheit)? TO TEnd Eht 0

** 185.8 fChrenhelt

RULE62

RULE384

COLILECT-SF [ L[' i.e PHY are we asking nhether Susanne ht a experitenced seizures

recent L Y RULE062.e are. ryig to decide whether Ssane hr .... E8Tboigi-YnTtERI-PA [

Recent seizures is strongly associated path meningitisRULE585

RULE400TrEST4M-OTHESIS [ MEMYWGI1 74

RULE411

APPLYAT8 [ RULE RLE.60Figue 1. Mutipl mViw of cthe iagnot e RUE3 e

S saiZURES de W E Csed ab seizures ( s ) tnd,TNE-FO TANvL iULtE6 6 rl 2) The pen ingnts asr focursiv ESSaSREOFLAG-CNFin men ingti in c fonte t a bno e lalfy tacaiNEatM rue E424 see or buing inForeampe te ead e v t i then he ompont s cestved APPLYRULEI44 R.WheEn heGusslectdFEXPLAItm heme nEad424 acn tvdence (.) that the

pseo d ie h kw gbspate in has meningitiss .RULE11 7 -ae

vac RULEI -,. FIN UT SE2LURES]CSFPROTE |N RULE11 ? 7 g

SRULE169

Figure 15. Multiple Views of the Diagnostic Process: Question, Evidence Relation, Task Stack, Metarulesoand Prosaic Condensation. When NEOMYCIN asked about seizures (question 8), the user selected a subiten in the KBWINDOWS menu, which caused the task stack-the current line of reaoning-to be displayed. The rule above a task is themetarule that invoked it; thus, rule 400 selected meningitis a focus, invoking TEST-HYPOTHESIS with it as a argument.Selecting meningiti s window caused the table in the lower left to be displayed. Here, boldface type indicates positivefindins aId sucessusy applied tules. Greyed ateas correspond to negative findings ad failed rules. Thus, the patient is not aneonute; rule 424 succeeded. Arows preceding finding indicate that the finding is in a triggering relation with the hypothesis.For example, the headache volunteered in the chief complaint caused the program to try to apply rule 424. When the userselected EXPLAIN in the menu adjacent to the consultation typescript, the program summarized the line of reasoning, skipping

over Tuninteresting tasks.

Heracles: From Diseases to Stereotypes NEOMYCIN approach to Al students, I found that it waspossible to redescribe other knowledge bases in its terms.

In late 1983 1 began to consider how NEOMYCIN might be For example, in terms of the mapping between models ofgeneralized. What kinds of problems can be conveniently situation descriptions and selected solutions, "people are

solved by an architecture consisting of a classification net- to diseases" as "meals are to wines." I had also recentlywork and a separate, abstract control strategy? In partic- reread Rich's work on user modeling (Rich 1979), intend-ular, to what problems can the same diagnostic strategy ing to apply this to our explanation program. I recognizedbe applied? It was obvious from the start that the proce- that it fit the same pattern-models of people related toadure had nothing specifically to do with medicine; was it taxonomy of books. Finally, I recalled that Rubin (1975)

more general than diagnosis? In attempting to teach the and Aikins (1983) emphasized that diseases are described

52 THE AI MAGAZINE August, 1986

0N

TASK: PROCESS-FINDINGFOCUS: $FINDING EMC Vew by

PAI~

Metarule

IF: (AND (OR (FINDINGTYPE SFINDING REDFLAG)(NOT (DIFF.EXPLAINED $FINDING)))

(MAKESET (TRIGGERS? SFINDING $RULE)RULELST)) ombn dM610

THEN:(TASK APPLYRULES RULELST) ows P"N-r

If the finding must always be explained or - "

it is not currently explained by the differential,then trigger hypotheses that explain it.

Definitional relation rules = b

IF: (AND (DIFFERENTIAL SHYP)(EXPLAINEDBY SFINDING SHYP))

THEN:(DIFF.EXPLAINED SFINDING)

A finding is explained by the differentialif it's explained by some hypothesis in the differential.

IF: (OR (CAUSED-BY SF SH) MAN

(AND (TYPE SH $PARENT)(EXPLAINEDBY SF SPARENT))) I

THEN: (EXPLAINEDBY SF SH) Figure 17. How Control Knowledge Is Encoded in

HERACLES. HERACLES is implemented as a specializationA finding is explained by a hypothesis of EMYCIN. Above is shown the original conception of domainif it is caused by the hypothesis or parameters and rules. In HERACLES parameters are special-by some more general category. ized as domain relations, control tasks, and domain terms, con-

ditionally inferred and invoked by rules. We use "relation" inthe mathematical sense to refer to both predicates and func-tions. "Finding" and "hypothesis' formally classify the domainterms (for example, meningitis), and informally are used to re-

Figure 18: Propositional Representation of a Metarule. fer to propositions in a situation-specific model; so, we say thatThis is one of six metarules for accomplishing the task 'the patient has meningitis' is a hypothesis.PROCESS-FINDING, which is invoked whenever a new find- Tasks are accomplished by an interpreter that appliesins becomes known. The metarule detects that this finding metarles. Propositions used by metarule premises (such asis serious and has to be explained (a red-flag finding), or it's (EXPLAINED-BY SF SH) appearing in Figure 16), can be in-something that's not currently explained by the set of possibil- ferred definitionally by rules or can be inferred by proceduralities under consideration. The program gathers up the trigger attachment (for example, accessing Lisp structures). Theserules-automatic inferences-and tries to apply them. The propositions are both static and dynamic. They classify do-idea is that if the finding doesn't always have to be explained main propositions and domain rules, as well as characterizingand it's explained by hypotheses that were already triggered, the problem-solving state (such as whether a hypothesis is inyou shouldn't trigger a new hypothesis. For example, if the the differential or whether a task has been done yet). Addi-patient has a headache, and other evidence suggests menin- tional relations that classify tasks are used by the task inter-gitis, which would explain the headache, there's no need to preter (not shown here). Metarule actions apply domain rules,consider other explanations of the headache. Intermediate re- request (from the user) or conclude domain propositions, orlations, such as EXPLAINEDBY, are defined by other rules invoke other tasks. In particular, the task FINDOUT uses all(simplified here). All pattern variables in these rules are instan- of these methods to infer domain propositions. In HERACLEStiated as domain rules or terms. All expressions are implicitly all domain rules are applied directly by metarules rather thanuniversally quantified. by back chaining. Only domain rules mention domain terms

directly; other rules use variables.

THE AI MAGAZINE August, 1986 53

I

ilW~m ontkle Ti1 0 R~de

, "I 1 h IS the p oteir, value (mg'loocc) in the (Cg7

Ha s a T a,, be en r un on Susanie'

1Kv_-1LI tw 4ewol

_ - _ -_.-C,, -_ _ _ _ . . _ _= = . _

!1/ .- , .Al

FINoING ftOLE(S) WXACF *IlICr F INING vat m C H THESIS cF aCmrIILI 1 LL:. 950 AGE 44 yeari IWCIU-PROCESS 00 991

I £;-. oeAvf RULEM54 900 FACE Necas'.u SENINGITIS 850 90",Aao "Ca,' CULE We5 HEADACHE A(iJTE -MENIN G1 AL -1U 0 0P APILEENA RtULE219 900 P 0OO A /AcuTE -1ACTERIAL-ENINGI1IS 80

CILC tULE373 80 PA'HE' INCRlEASED- INTRACRANIAL-PflESSURF 200HE AO*CHf -CIONIC* RULE23 700 FEBiILE INIRA ANIAL -NASS-LESHI0N 50L;T N RULE23: lHE ADACHE-PUFATION h UrIS S5nO i mO MSSEIZIRES 1.EZ62 290 CN -F INDI 1C:T ;FF-NECI,- IGN' B4rA IN -ANEWI" 100

t FF- NE C 0 -LE-,Irl"

Hm. I GM _i,= r,,. ,ADE -FE .....II

Figure 18. Overprinting a Classification Network to Show How it is Searched. Nodes blink and are boxed tomake visible the "looking up" and "looking down" process of diagnosis. Numbers indicate the relative certainty of conclusions;the cumulative certainty factor (CUMCF) includes hierarchical propagation. Hteavy-bordered boxes indicate the program'sdifferental -the most specific cut through the taxonomy and causal network. The differential is printed in the lower rightwindow with indenting to show specialization by process subtype and cause. When a hypothesis is selected, the evidence windowcan be displayed, indicating which findings and rules have been considered and the outcome of each consideration Dozens of

0 other windows are available, including different views of causal networks and the history of task invocation.

(in knowledge bases) as stereotypes. The general model of eases as classes is inadequate given what is required inheuristic classification fell into place: Some problems can general and what is evident in other programs (for exam-be solved by selection, heuristically relating a classifica- pie, allowing for multiple inheritance). We call the recon-tion of problem data to a classification of known solutions ceptualized framework HERACLES. It is not a completed(Clancey 1985a). tool but an idea that, continues to evolve

o To my chagrin, this new model required a reconceptu- Figure 19 illustrates the heuristic classification analy-V alization of parts of NEOMYCIN. We began to consider sis of SACON, a program that many of us knew about andthe diseases as stereotypes, we introduced qualitative ab- talked about for five or six years, but that few understoodstraction of numeric data where it had been omitted in until its knowledge base was portrayed in this way. 'IheMYCIN, and we realized that our representation of dis- purpose of SACON is to select a configuration of programs

54 THE Al MAGAZINE August, 1986

0U, " " " " , . . h '" " ' *

in a structural analysis software package developed by the assembly flaw (congenital), environmental influence (infec-Marc Corporation. These programs can analyze an object tion, toxicity, trauma, psychological load), or degenerationfor structural failure in many ways, some of which are un- (vascular disorder, immunoresponse, muscular disorder).necessarily accurate and time consuming. An expert can In both of these physical systems, externally observabletell you which of the programs should be run to analyze a manifestations are explained in terms of internal systemparticular structure, and that is SACON's task. Imposing behavior, tracked back to faulty structures and malfunc-a type classification on SACON's concepts, and labeling tions of subsystems. These are in turn explained by theinferences as abstractions, heuristics, and refinements, we etiologies, processes in which the system interacted with itsfind a previously hidden secondary structure which helps environment, bringing it to its current state. In medicine,us to understand what SACON does. these etiologies include congenital problems (caused by the

Studying and generalizing knowledge-based programs, mother's lifestyle or her environment), psychogenic prob-we can go quite a bit further. First, we can realize that as lems (emotional overload), trauma (structurally damagingstereotypes the classifications are models of systems: spec- the body), toxic environment, and so on. In the humanifications or descriptions of systems and plans for assembly body, internal systems generate new subsystem structures,or modification of systems. Second, the classification se- so developmental and degenerative processes are also im-quences, relating one model to another, are regular and portant etiologies. We believe that this analysis can belimited in nature, constituting tasks. A model of system generalized to cover all physical systems.being monitored is related to a plan for controlling its be- A second interesting result is the set of heuristics wehavior. A diagnostic model of a faulty system is related discovered for constructing a well-formed causal networkto a repair plan. A specification is related to a design (Clancey 1984d). These heuristics include asking the ex-and then to an assembly plan. Finally, the idea of sys- pert about categories of states; asking about unobserv-tems, tasks, and common sequences is independent of how able states that track back to different etiologies; distin-the solutions are computed each step along the way. Ei- guishing clearly between substances and processes, par-ther heuristic classification or some constructive method ticularly, never causally linking substances directly; and(perhaps involving nonmonotonic reasoning, hypothetical working backward from repairs to causes. This last pointworlds, and so on) might be used. It is important to re- emphasizes that the purpose of the causal-associationalmember that this inference structure shows the pattern of and etiologic taxonotny is to make choices about repair, ainferences that map given information to final solutions, point I emphasized in The Disease Taxonomy: Searchingand makes no claims about the process or order in which an Abnormal Process Clarification. Uncertainty in diag-the inferences are made. Further examples and extensive nostic reasoning need only be resolved to the extent thatdiscussion appear in (Clancey 1985a, 1986). it makes a difference in distinguishing among repairs.

Our heuristics can be viewed as criteria for critiquingCaster: From Diseases to a behavioral causal model. Can we formalize these con-

Abnormal Substances and Processes straints so that they can be taught to a student? Viewing

In addition to the disorder taxonomy (Figure 11), a knowl- a diagnosis as a model is the first step.edge base for diagnostic problems constructed in HERA- The Situation-Specific Model:CLES might include a causal-associational network. Dis- From a Diagnosis to an Explanationorders in this network are descriptions of internal states inthe system being diagnosed. Figure 20 shows such a net- This lesson might be the most important It is the ideawork for CASTER, a knowledge system for sand-casting that a diagnosis is not the name of a disease but an argu-diagnosis. ment which causally relates the manifestations which need

* Tim Thompson and I (1986) developed this program to be explained (because they are abnormal) to the pro-in order to better understand the distinction between the ceases that brought them about (See Figure 21). A numberpathophysiological states of the causal net and the etiolo- of ideas come together here:gies, or final causes, of the disorder taxonomy. This dis- 9 Diseases are processes (see The Disease Taxonomy:tinction was emphasized in the CASNET program (Weiss Searching an Abnormal Process Classification andet al. 1978); our interest was to apply the ideas to a non- Caster: From Diseases to Abnormal Substances andmedical problem. Processes.) Thus, a diagnosis is a network causally

What did we learn from the CASTER experiment? linking manifestations and states to processes.First, for diagnosing malfunctions in some manufactur- e A causal explanation applies the general concepts anding process, it is useful to organize the disorder taxon- links in a knowledge b. i , construct a case-specificomy according to each stage in the overall process (pat- model (Patil, Szolovits. k, Shwartz 1981). Thus, thetern design, melting, and so on). In contrast, the top level network linking manifestations and diseases is a modelof NEOMYCIN's taxonomy corresponds to defects in the of a particular sequence of events in the world (alsoneurological system, viewing it as an object, not a process: called a situation-specific model)

TIE Al MAGAZINE August. 1986 55

@4

Ofw"e .. P*4 "Rd *e W, . - -r it

Szolovits, & Schwartz 1981).Auaveft ProUwfl

e A causal explanation has the structure of a geome-try proof: It must account for all of the findings and

o, ..um. ~ must be coherent and consistent. Thus, the situation-A k §n.W, specific model must be a connected graph with one

BI T process at the root (assuming a single fault).

O u ,-c oon The evolution of these ideas is intriguing, revealingOf S AM1 WW 00MOW how our computational tools and the use of the computer

oE'rO'u as a modeling medium changes how we think. Sometime inHE4MSTIC MATCHN

- 1985 it occurred to me that we could extend the windowsDATA offered by GUIDON-WATCH to include a graph show-

ABsmRCON ing how the final diagnosis related to the known findings.s ,ucw o.0sao When I saw the way Anderson replaced a linear geome-

try proof by a graph (using the same Interlisp-D graph-_P, -.wo ics package), the analogy between a causal explanation

- and a proof became concrete (Anderson, Boyle, & Yost/ 0 1985). Thus, the example from another domain showedS -9d L o.V' eyc how Patil's idea of a patient-specific model could be useful

osfiecr om. in teaching, and the availability of the graphics package

4 UAJTA1W encouraged us to create the picture to see what it wouldI ,look like.

It is astounding to realize how many hundreds of ex-pert systems are cranking out diagnoses with neither the

SnFNTICN"L programs nor their designers ever explicitly considering a

see" sp.Ie.PW . diagnosis as a coherent causal model. They don't even4' check to see if all of the findings are covered by the final

m6nT. diagnosis. Our language is too loose: The program prints

0WdRW out the name of a disorder, and we say, "The programhas made a diagnosis." However, where is the explanationargument?

Figure 19. Inference Structure of SACON. An ab- For the purpose of teaching, this graph could perhapsstract description of inference chains is shown above a par- be the best way to reify the process of diagnosis. For sev-ticular sequence of associations. SACON abstracts the given eral years, inspired by Brown's emphasis on "process ver-structure and relates this abstraction to half a dozen rules of sus product" (possibly derived from Dewey [1964]), I'vethumb that make a quantitative prediction of the structure's been searching for some written notation that we couldbehavior under stress. Specifically, the fact that the structure use, something analogous to algebra, to make visible whatis a beam is combined with information about its size, support, the operators of diagnosis (NEOMYCIN's tasks) are do-and load distribution in order to select a numeric equation, ing. The analogy with geometry turns out to be strongerwhich computes stress and deflection. These predictions are than the analogy with algebra because each inference itselfabstracted, and definitionaly related to the structural analysis relies on a proof, analogous to the causal arguments be-program. Specifically, the characterization of stress, combined hind each link of the situation-specific model. In algebrawith information about loading and error tolerance, is classi- hinec l es ae alluatios.fied as a particular kind of *analysis" for which the program the inference rules are all axioms.is specialized. The SACON program selects from about 30 dif- Giving this window to the student, we might have himferent program combinations. This corresponds to the number carry out the diagnosis by posting his hypotheses and link-of organisms in MYCIN, and is probably good to remember ing them to the known findings. Each step along the way,when considering whether the heuristic classification method is there are visible problems to be solved. The student canappropriate for solving a problem. see that he is trying to construct a logically consistent

network. Behind each request for data is an operation formaking the network hang together-explaining the find-

* Diagnostic operators examine and modify the differ- ings that need to be explained and refining the hypothesesential (most specific diseases under consideration), that need to be made more specific. An instructional pro-linking and refining them. Thus, HERACLES tasks gram is now being developed based on this idea. Calledare operators for constructing a situation-specific GUIDON-MANAGE, it has a student "manage" the diag-model (similar to ABEL's diagnostic operators (Patil, nosis by explicitly applying strategic operators.

56 THE Al MAGAZINE August, 1986

0

,#EOF 0-84TOOMLL EXCEISPIMY'uAR ETsUS

/ E -TOO IMLL ECE@@1IV11Y-. 4iNGATE

F -IEAVY -THW-THIN PATTEM-EBIMFEE-SlT -FF ISTTOO-8100AL. Was-PATTERm

i -WAWALL TT E4fUl-t@DWPRT-WA F Y-FITTED-JACET\ /METAl.-TOO-HOT

SlWIOA(E-GAVITES --KADEQU)ATE -FEEDOG BFlpNivp4A*OETL TO-OB/KKEN-M(D NOT "4NOSJG4EAS '~

fl HOT -.O'EFIMATIN4(---5KS1 -NqAhG

-/ OT .EFOA TION -MATERIALSIO-METAL -TO-FEWD

10101T 400..hW± /FLASKTO4 MM04LL\CORE -FLASK -TOO -

Figure 20. CASTER's Causal-Associational Network for Shrinking Defects in Cast Iron. This simplified networkrelates structural failures (for example, mold wail movement) to functional failures (for example, inadequate mold support).These are all internal to the system and often can't be observed directly. Reasoning proceeds as follows. Given some surfacefault, such as shrinkage cavities in the cast iron, we reason backwards to possible causes: (1) feed of metal shut off, (2) a brokenmold (leak), and (3) absence of metal to feed. Gates, risers, and fillets refer to structures for shunting metal and venting gases.Terminal nodes, on the right side, track the problem back to some problem in the iron-casting process (pattern design, moldformation, metal melLing, and so on), thus relating system behaviors to external causes (the designer's assumptions, previoustreatment of the sand, contamination of the metal supply, and so on). We believe that analyzing such networks, relating them tothe well-defined structure and function of the sand-casting system, will help us to redefine in a principled way the causal relationsgiven to us by experts in other domains, such as medicine. Working in multiple domains proliferates metaphors and helps us todevelop more general theories about expert knowledge.

This is an amazing change. Ten years ago I thought e Reimplementing the explanation program to use theI was trying to teach parameters and rules, and now I'm logic encoding of the metarules (stating this programsaying that I want to teach the student to be an efficient in the same task-metarule language so that it mightmodel builder. What can we tell the student that will reason about its own explanations)help him critique the model that he's constructing? For o Generalizing our graphics package using object-example, we'll say, "All the important findings need to be oriented techniquesexplained." Observing that he has failed to do something e Applying the student-modeling program, ODYSSEUS,that needs to be done, we'll tell him about the opera- t Appldge st iontors, so he can step back and say, "Well, what knowledge to knowledge acquisitionmight I be missing that prevented me from carrying out * Preparing HERACLES for use by other peoplethat task?" So debugging by explanation of failure-pro- I'm going to jump up a level here to consider someceeding from model constraints to operators to knowledge methodological lessons we can draw from this research.relations-is the approach we're following. This leads to Figure 1 provides a simplified summary of how thean interesting model of learning (Clancey et al. 1986). various programs and research ideas are connected. We

observe two examples of a specific expert system beingMethodological Lessons generalized, with the resulting shell used to construct other

Tspecific systems and a tutoring shell. Is there any logic inTo summarize ongoing projects mentioned or alluded to this sequence that might reveal something about learninghere, we are currently doing the following: in general or at least about how we learn by constructing

* Developing instructional programs based on NEOMY- programs?CIN In the section names in this article, I indicated the

* Studying learning in the setting of debugging a knowl- sequence of terminological changes ("from ... to ... ") thatedge base seem to mark each major change in my understanding.

THE Al MAGAZINE August, 1986 57

aim

IGO

-PULER1 AULI RULI,2800111

SO0 000 HIQH~~ ~~-IM E CH S-,1NDItO -DURIATION q u .

50 0W(1201ors

,IWCCT1OU$ .PIC9SS $i ULL[ 2,

700 Oro

RULE423 HEADACIE ST -IP. E-ONI- LEXON

Figure 21. Partial Diagnostic Model in NEOMYCIN. The process of diagnosis is the construction of a proof tree,relating the findings and disorders that could have caused these findings. At some intermediate state when solving the problem,the network is disconnected and partial. The patient has seizures; what could have caused that? There is some support forAcute Bacterial Meningitis and Increased Intracranial Pressure, but these two hypotheses haven't been related. Is there someunderlying cause (process) that could account for all of the manifestations? Diagnostic operators can be viewed as graphconstruction operators, focusing on particular nodes and trying to grow the graph down to support possible explanations orrefining it upward to more specific explanations. A final situation-specific model is a connected network, with some root processthat we say explains the internal states (such as Increased Intracranial Pressure), which, in turn, explain the observed findings.This graph, as an argument having the structure of a proof, is the diagnosis, not the term Acute Bacterial Meningitis.

The renaming that occurred in moving from "clinical are some clear patterns:parameter" to "model" is dramatic. None of the interme- * Abstracting or generalizing terminology to incorporatediate concepts (hypothesis, relation, process, and so on) another specific domain (for example, moving fromis new, but it is interesting to note how they are retained disease to disorder process)and how they build upon one another as the knowledge * Separating a domain model (what is "true") fromstructures are reinterpreted from different perspectives, the inference process (what to do) by identifying and

Thus, in HERACLES today, we have parameters, justifying procedural sequences (for example, defin-terms, hypotheses, diseases, processes, stereotypes, and ing relations for ordering MYCIN's rule clauses and,models. All of these remain true descriptions of what's later, defining relations for ordering NEOMYCIN'sin our program. The perspective changes, broadening metarules)from language terminology (parameters, terms) to reason-ing phenomenoogy (hypotheses), domain ontology (disease * Justifying domain relations in terms of underlying con-

process taxonomy); and, finally, epistemological distinc- straints and patterns (for example, a theory for gener-

ions (stereotypes, models). With the heuristic classifica- ating appropriate follow-up questions or trigger rules

tion perspective at the top-couched in terms of systems, or a theory for generating a causal network in terms

tasks, and models (Clancey 1986)-previous terminology of faulty structures and malfunctions)

is retained for describing the program at different levels. Figure 22 summarizes the overall pattern. The pointLooking closely at the sequence of research itself, there of the analysis phase is to detect patterns that we want to

58 THE Al MAGAZINE August, 1986

model explicitly and that have been mapped into the lan-guage in an implicit and perhaps undisciplined way. Thus,findings and hypotheses, causality and subtype, and dis- COMPUTATIONAL REPRESENTATIONSease knowledge and procedures are not distinguished inMYCIN. Findings and hypotheses are both representedas parameters. Cause and subtype are represented by se-quences of clauses in rules, or in the relation between aparameter and its values (for example, parameter- "thekind ot Meningitis"; value--"Bacterial"). Focusing proce- PATTERNSdures are also encoded by rule clause ordering.

There is apparently no end to this criticism; the samegame can be played with NEOMYCIN. For example, inattempting to improve the explanation program we find DSCRIB

that the use of terms in NEOMYCIN's original metarulesis ludicrously undisciplined; they are used like arbitraryprogram variables, with no apparent connection betweenSHYP and $CURFOCUS. Interpreting this representation GENERATIVEfor diagnosis causes no difficulties, but the explanation JUSTIFCATIONSprogram needs to know that the metarules refer to thesame kind of entity, a hypothesis.

This analysis suggests that detecting patterns of state-ments in some language, articulating a new classificationmodel, and defining a new procedure by which the state-ments are to be interpreted are intricately related. Recall-ing the analysis of metarules (MRS/NEOMYCIN: FromFindings and Hypotheses to Relations), we observe that LANGUAGEeach new purpose for interpreting a representation re-quires new distinctions-new relations-to classify exist- _I

ing domain terms, rules, and relations among them. Thus, Figure 22. Methodology for Improving Computa-the compiler needs to know which domain relations are tional Models. In the process of knowledge representation,predicates and which are functions (in the mathematical we write statements in some language; we organize what wesense). ODYSSEUS needs to know when metarules can have written down, describing and classifying patterns; we ex-be reordered. The teaching program needs to know why plain the patterns in terms of primitive relations; and we definemetarules are ordered a certain way. In classifying rela- a new language that enables us to explicitly state these primi-tions and terms, we are constantly asking, "Which things tive relations and generate the original patterns. For example,can be procedurally operated upon in the same way?" clause correlations in in MYCIN's rules are now reformulated in

Winograd reached the same conclusion in his analysis tasks, metarules, and domain relations. Another cycle occursowhen we study these metarules and articulate the constraintsof how language arises. The need to take action reorients behind their design. Similarly, patterns in NEOMYCIN's dis-us to the world, forcing us to make new distinctions. The ease taxonomy and CASTER's causal network are articulatedrelevant properties attributed to an object are determined by characterizing diseases as processes and states as abnor-by the role the object plays in an action: "This grounding mal structures and malfunctions. These new perspectives-of description in action pervades all attempts to formalize the search for patterns and their articulation in a new lan-the world into a linguistic structure of objects, properties, guage-all rise in an attempt to formulate some generativeand events" (Winograd & Flores 1986). Indeed, by this rationale for constructing similar structures in new domains asanalysis the world and its objects exist only in language, well as to evaluate existing networks for consistency and com-mediated by action. pleteness. A generative theory of a representation facilitates

The expert system methodology of writing down teaching people how to use the representation, reformulatingknowledgeein smeystruture d m o y so taiti wn it for efficiency, and constructing explanation programs and, knowledge in some structured way so that it can later knowledge-acquisition tools.

be studied and better formalized is a remarkable, excit-

ing turning point in epistemological practice. We try tounderstand why a relation holds by abstracting it and thentrying to find similar relations in the knowledge base. If a the same kind of concepts, leaving out the same kind ofpattern holds, we restate everything more abstractly. Why details? Do all links in the network connect structures tois it correct to say that "broken mold" causes "inadequate functions? Is there any reason why they should?feeding"? What other causal links in the network connect Having written a model down, the most powerful tools

THE Al MAGAZINE August, 1986 59

of language come into play: References* It is possible to reflect on what was said, to ask whyIt is trusle, to eeo wbtter u rstadg tor thry Aikins, J. S. 1983. "Prototypical Knowledge for Expert Systems."it is true, to develop a better understanding or theory. Artificial Intelligence 20(2): 163-210." An incremental critique and transformation process Anderson, J. R., C. F. Boyle, & G. Yost. 1985. "The Geometry

becomes possible-the best way to build anything so Tutor." In Proceedings of the International Joint Conference onArtificial Intelligence 9, vol. 1. Los Altos, California: William

that it is reliable and useful (Petroski 1985). Kaufmann, Inc., 1-7.Computational languages provide a way of writing Brown, J. S. 1983. "Process Versus Product-A Perspective on Tools

for Communal and Informal Electronic Learning." In Educationthings down so that the model is executable, the very in the Electronic Age. New York: Educational Broadcasting Cor-thing we need for modeling processes. Al research is ex- poration, WNET.ploring how to model physical, inferential, communicative, Brown, J. S., A. Collins, & G. Harris, 1977. "Artificial Intelligencemotoric, and perceptual processes using qualitative (prin- and Learning Strategies." In O'Neill (Ed.), Learning Strategies.

New York: Academic Press.cipally nonnumeric, relational) representations (Clancey Carbonell, J. R. 1970. Mixed-Initiatie Man-Computer Instructional1986). Graphics provide a medium for visualizing pro- Dialogues. Tech. Rep. 1971, Bolt Beranek and Newman.

cesses, so we can understand the complexity of the sys- Clancey, W. J. 1979. "Transfer of Rule-Based Expertise Through atems we construct (Hollan, Hutchins, & Weitzman 1984; Tutorial Dialogue." Ph.D diss., Stanford University.

Clancey, W. J. 1979. "Dialogue Management for Rule-Based Tu-Clancey 1983c, Richer & Clancey 1985) and even start to torials." In Proceedings of the International Joint Conferenceask new questions as icons and graphs become part of our on Artificial Intelligence 6, Tokyo, Japan. Los Altos: Williamlanguage for stating theories. The marriage of qualitative Kaufmann, Inc., 155-161.modeling and graphics in the 1980s, made available on Clancey, W. J. 1979. "Tutoring Rules for Guiding a Case Method

Dialogue." The International Journal of Man-Machine Studies.cheaper, more powerful machines, provides a sharp stim- 11: 25-49. (Also in Sleeman& Brown, (Eds.), Intelligent Titoringulus to Al research and a good reason to be optimistic Systens. New York: Academic Press).about the progress to come. Clancey, W. J. 1982. "GUIDON." In A. Barr & E. A. Feigenbaum

(Eds.), The Handbook of Artificial Intelligence. Los Altos, CA:William Kaufmann, Inc.

Clancey, W. J. 1983. "The Epistemology of a Rule-Based ExpertSystem: A Framework for Explanation." Artificial Intelligence20(3), 215-251.

Clancey, W. J. 1983. "The Advantages of Abstract Control Knowl-edge in Expert System Design." In Proceedings of the National

Acknowledgements Conference on Artificial Intelligence, Washington, D.C. MenloPark, California: American Association for Artificial Intelligence,

This article was prepared as a final ONR technical report for the 74-78.period 1979-1985. The GUIDON project was first funded jointly byMarshall Farr at ONR and Dexter Fletcher at DARPA. Other tutor- Clancey, W. J. 1983. "Communication, Simulation, and Intelligenting programs in the ONR/ARPA family at that time included WHY, Agents: Implications of Personal Intelligent Machines for MedicalWEST, and SOPHIE. Avron Barr, John Seely Brown, Ira Gold- Education." In Proceedings of AAMSI-8. 556-560.stein, and Keith Wescourt helped us to define the initial proposal, Clancey, W. J. 1984. "Teaching Classification Problem Solving (Ab-conveying the problems and methods then current in the research stract)." In Proceedings of the Sixth Annual Conference of thecommunity. Bruce Buchanan was principal investigator, shaping the Cognitive Science Society, Boulder, Colorado. New York: Cog-clarity and style of the first proposal and continuing today to serve nitive Science Society, 44-46.as advisor on the project. Clancey, W. J. 1984. "Methodology for Building an Intelligent Tutor-

I sincerely thank the students and programmers who have made ing System. " In Kintsch, Miller, & Poison (Eds.), Method andso many significant contributions to the NEOMYCIN and HER- Tactics in Cognitive Science. Hillsdale, New Jersey: LawrenceACLES programs (chronologically): Reed Letsinger, Bob London, Erlbaum Associates, 51-83.Conrad Bock, Diane Warner Hasling, David Wilkins, Glenn Ren- Clancey, W. J. 1984. "Acquiring, Representing, and Evaluating anels, Mark Richer, Tim Thompson, Steve Barnhouse, Arif Merchant, Competence Model of Diagnosis." HPP Memo 84-2, StanfordDavid Leserman, and Naomi Rodolitz. Victor Yu, M.D. helped me University, February. (To appear in M. Chi, R. Glaser, and M.understand MYCIN in the early years. I am especially grateful to Farr (Eds.), Contributions to the Nature of Expertise.)the late Tim Beckett, M.D. for showing me what a good physican Clancey, W. J. 1984. "Knowledge Acquisition for Classification Ex-teacher can do. Curt Kapsner, M.D., John Macias, and Bevan Yueh pert Systems." In Proceedings of ACM Annual Conference. Newhave enthusiastically collaborated in our recent work. York: Association for Computing Machinery, 11-14.

The project was funded jointly by ONR and the Army Research Clancey, W. J. 1985. "Heuristic Classification." Artificial Intelli.-Institute during the period 1981-1985 (ONR contract N00014-79C- lence 27(December), 289-350.0302). Computational resources have been provided by the Sumex- gnce Fotcoming. 289-350.Aim National Resource (NIH grant RR00785). Organization and Clancey, W. 3. Forthcoming. "Representing Control Knowledge asmaintenance of computer resources in the Knowledge Systems Lab- Abstract Tasks and Metsrules." In M. Y. Coombe & L. Boc.oratory is managed by Tom Rindfleisch. Continuing research is sup- (Eds.), Comiiter Expert Spstemn. New york: Springer-Verlag.ported in part by ONR contract N00014-85-K-0305 and by a grant Clancey, W. .. 1986. The Science and Engineering of Qulitativefrom the Josiah Macy, Jr., Foundation. Models. KSL Report 86-27, Stanford University.

This paper is based on a talk I gave at the Computer Science Clancey, W. J. & C. Bock. 1982. "MRS/NEOMYCIN: RepresentingColloquium at Carnegie-Mellon University on 19 November 1985. I Metscontrol." Is Predicete Calculs. HPP Memo 82-31, Stan-thank David Evans for arranging to have the talk taped and Rene ford University.Bornstein for her accurate transcription. Many useful revisions were Clancey, W. J. & B. C. Buchanan. 1982. Esploration of Problem-suggested by Steve Barnhouse, Ted Crovello, Bob Engelmore, and Soling and Fstoria# Strategies: 197D-1989. Tech. Rep. STAN-Mark Richer. Special thanks to Susan King for TEXing the final CS-82-910, HPP 82-8, Stanford University.copy. Continued on Page 187.

60 THE Al MAGAZINE August, 1986

0

a

Clancey, continued from page 60

Clancey, W. J. & R. Letainger. 1984. "NEOMYCIN: Reconfiguringa Rule-Based Expert System for Application to Teaching." In W.J. Clancey, & E. H. Shortliffe (Eds.), Readings in Medical Arti-ficial Intelligence: The First Decade. Reading, Massachusetts:Addison-Wesley, 361-381.

Clancey, W. J., M. Richer, D. Wilkins, S. Barnhouse, C. Kapsner, D.Leserman, J. Macias, A. Merchant, & N. Rodolitz. 1986. Guidon-Debnu: The Stndent as Knowledge Engineer. KSL Report 86-34,Stanford University.

Dewey, J. 1964. "The Process and Product of Reflective Activity:Psychological Process and Logical Form." In R. D. Archambault(Ed.), John Dewey on Education: Selected Writings. New York:Random House, 243-259.

Genesereth, MR. & D.E. Smith. 1982. Meta-Level Architecture.HPP Report 81-6, Stanford University.

Hasling, D. W., W. J. Clancey, G. R. Rennels. 1983. "Strategic Ex-planations in Consultation." The International JoLrnal of Man-Machine Studies 20(1), 3-19. Also in M. J. Coombs, (Ed.), De-velopment in Expert Systems. London: Academic Press.

Hollan, J. D., E. L. Hutchins, & Weitzman, L. 1984. "STEAMER:An Interactive Inspectable Simulation-Based Training System."Al Magazine 5(2), 15-27.

London, B. & W. J. Clancey. 1982. "Plan Recognition Strategiesin Student Modeling: Prediction and Description." In Proceed-ings of the Second National Conference on Artificial Intelligence,Pittsburgh, PA. Menlo Park, California: American Associationfor Artificial Intelligence, 335-338.

Patil, R. S., P. Szolovits, & W. B. Schwartz. 1981. "Causal Under-standing of Patient Illness in Medical Diagnosis." In Proceedingsof the Seventh International Joint Conference on Artificial Intel-ligence. Los Altos, California: William Kaufmann, Inc., 893-899.

Petroski, H. 1985. To Engineer is Human: The Role of Failure inAD Space Successful Design. New York: St. Martin's Press.

Rich, E. 1979. "User Modeling Via Stereotypes." Cognitive Science3, 355-366.

Richer, M. H. & W. J. Clancey. 1985. "GUIDON-WATCH:A Graphic Interface for Viewing a Knowledge-Based System."IEEE Computer Graphics and Applications 5(11), 51-64.

Rubin, A. D. 1975. Hypothesis Formation and Evaluation in Med.ical Diagnosis. Tech. Rep. AI-TR-316, Artificial IntelligenceLaboratory, Massachusetts Institute of Technology.

Schoenfeld, A. L. 1981. Episodes and Executive Decisions In Math-ematical Problem Solving. Tech. Rep., Hamilton College, Math-ematics Department.

Shortliffe, E. H. 1976. "MYCIN: A Rule-Based Computer Programfor Advising Physicians Regarding Antimicrobial Therapy Selec-tion." Ph.D diss., Stanford University.

Thompson, T. & W. J. Clancey. 1986. "A Qualitative ModelingShell for Process Diagnosis." IEEE Software. March. 3(2), 6-15.

van Melle, W. 1979. "A Domain-Independent Production Rule Sys-tem for Consultation Programs." In Proceedings of the Sixth In-ternational Joint Conference on Artificial Intelligence. Los Al.tos, California: William Kaufmann, Inc., 923-925.

Weiss, S. M., C. A. Kulikowski, S. Amarel, & A. A. Safir. 1978."A Model-Based Method for Computer-Aided Medical DecisionMaking." Artificial Intelligence 11, 145-172.

Wilkins, D., B. G. Buchanan, & W. J. Clancey. 1984. "Inferring* an Expert's Reasoning by Watching." In Proceedings of the 1984

Conference on Intelligent Systems and Machines.Winograd, T. & C. F. Flores. 1986. Understanding Computers and

Cognition: A New Foundation for Design. Norwood, New Jersey:Ablex Publishing Company.

Woolf, B. & D. D. McDonald. 1984. "Context-Dependent Tran-sitions in Tutoring Discourse." In Proceedings of the NationalConference on Artificial Intelligence, Austin, Texas. MenloPark,California: American Association for Artificial Intelligence, 355-361.

THE AI MAGAZINE August, 1986 1870

Personnel Analysis Division, Dr. Gordon H. BowerAT/NPXA Dr. Eva L. Baker Department of Psychology

5C360, The Pentagon UCLA Center for the Study Stanford UniversityWashington, DC 20330 of Evaluation Stanford, CA 94306

145 Moore HallAir Force Human Resources Lab University of California Dr. Richard BrabyAFHRL/MPD LOS Angeles, CA 90024 NTSC Code 10Brooks AFB, TX 78235 Orlando, FL 32751AFOSR, Director of Automation Dr. Robert Breaux

Life Sciences Directorate Allen Corporation of America Code N-095RBolling Air Force Base 401 Wythe Street Naval Training Systems CenteWashington, DC 20332 Alexandria, VA 22314 Orlando, FL 32813

Dr. Robert Ahlers Dr. Meryl S. Baker Dr. John S. BrownCode N711 Navy Personnel R&D Center XEROX Palo Alto ResearchHuman Factors Laboratory San Diego, CA 92152-6800 CenterNaval Training Systems Center 3333 Coyote RoadOrlando, FL 32813 Dr. Donald E. Bamber Palo Alto, CA 94304

Code 71Dr. Ed Aiken Navy Personnel R & D Center Dr. Bruce BuchananNavy Personnel R&D Center San Diego, CA 92152-6800 Computer Science DepartmentSan Diego, CA 92152-6800 Mr. J. Barber Stanford University

Dr. William E. Alley HQS, Department of the Army Stanford, CA 94305AFHRL/MOT DAPE-ZBRBrooks AFB, TX 78235 Washington, DC 20310 Dr. Howard Burrows

National Institute of HealthCapt. J. Jean Belanger Bldg. 10, Room 5C-106Dr. Earl A. Alluisi Training Development Division Bethesda, MD 20892HO, AFHRL (AFSC) Canadian Forces Training System Dr. Patricia A. ButlerBrooks AFB, TX 78235 CFTSHO, CFB Trenton

Astra, Ontario, KOK OERIDr. John R. Anderson CANADA 555 New Jersey Ave., NWDepartment of Psychology Washington, DC 20208Carnegie-Mellon University CDR Robert J. Biersner, USNPittsburgh, PA 15213 Naval Biodynamics Laboratory Dept. of Psychology

Dr. Nancy . AndersonP. 0. Box 29407DetofPyhlgDr. Nancy S. Anderson Orleans, LA 7019 University of South Carolina

Department of Psychology Columbia, SC 29208University of Maryland Dr. Menucha Birenbaum Dr. Robert CalfeeCollege Park, MD 20742 School of Education School of Education

Tel Aviv University Stanford UniversityDr. Steve Andriole Tel Aviv, Ramat Aviv 69978 Stanford, CA 94305George Mason University ISRAELSchool of InformationTechnology & Engineering4400 University Drive Dr. Werner P. Birke Joanne CapperFarfax, VA 22030 Personalstammamt der Bundeswehr Center for Research into Practice

Kolner Strasse 262 1718 Connecticut Ave., N.W.Dr D-5000 Koeln 90 Washington, DC 20009Dr. John Annett FEDERAL REPUBLIC OF GERMANY Dr. Jaime CarbonellSUniversity of WarwickDrJae ronlDepartment of Psychology Dr. Gautam Biswas Carnegie-Mellon UniversityCoventry CV4 7AJ Department of Computer Science Department of PsychologyENGLAND University of South Carolina Pittsburgh, PA 15213

Dr. Phipps Arabie Columbia, SC 29208 Mr. James W. CareyUniversity of Illinois Dr. John Black Commandant (G-PTE)Department of Psychology College, Columbia Univ. U.S. Coast Guard603 E. Daniel St. 525 West 121st Street 2100 Second Street, S.W.Champaign, IL 61820 New York, NY 10027 Washington, DC 20593

Technical Director, ARI Dr. Arthur S. Blaiwes Dr. Susan Carey5001 Eisenhower Avenue Code N711 Harvard Graduate School of

i Alexandria, VA 22333 Naval Training Systems Center EducationSpecial Assistant for ProJec Orlando, FL 32813 337 Gutman LibrarySpeialAssstatafr Paje

OASN(M&RA) Dr. Robert Blanchard Appian way5D800, The Pentagon Navy Personnel R&D Center Cambridge, MA 02138Washington, DC 20350 San Diego, CA 92152-6800 Dr. Pat Carpenter

Dr. Michael Atwood Dr. Jeff Sonar Carnegie-Mellon UniversityITT - Programming Learning R&D Center Department of Psychology1000 Oronoque Lane University of Pittsburgh Pittsburgh, PA 15213Stratford, CT 06497 Pittsburgh, PA 15260 Prof. John W. Carr III

Department of Computer ScienceUniversity of Colorad Dr. J. C. Boudreaux Moore School ofUnrsity of Pcholor g Center for Manufacturing Electrical EngineeringDepartmnt of Psychology eneri University of Pennsylvanialox 345 National Bureau of Standards Philadelphia, PA 19104* Slder, CO 80309 Gaithersburg, ND 20899

Dr. Robert Carroll Dr. Stanley Collyer Goery DelacoteOP 0187 Office of Naval Technology Directeur de L'informatiqueWashington, DC 20370 Code 222 Scientifique et Technique

800 N. Quincy Street CNRSLCDR Robert Carter Arlington, VA 22217-5000 15, Qual Anatole FranceOffice of the Chief 75700 Paris FRANCE

of Naval Operations Dr. Lynn A. CooperOP-OlB Learning R&D Center Mr. Robert DentonPentagon University of Pittsburgh AFMPC/MPCYPRWashington, DC 20350-2000 3939 O'Hara Street Randolph AFB, TX 78150

Pittsburgh, PA 15213Dr. Fred ChangNavy Personnel R&D Center Dr. Meredith P. Crawford Mr. Paul DiRenzoCode 51 American Psychological Association Commandant of the Marine CorpsSan Diego, CA 92152-6800 Office of Educational Affairs Code LBC-4

1200 17th Street, N.W. Washington, DC 20380Dr. Alphonse Chapanis Washington, DC 200368415 Bellona Lane Dr. R. K. DismukesSuite 210 Dr. Hans Crombag Associate Director for Life ScielBuxton Towers University of Leyden AFOSRBaltimore, MD 21204 Education Research Center Boilling AFB

Boerhaavelaan 2 Washington, DC 20332Dr. Davida Charney 2334 EN LeydenDepartment of Psychology The NETHERLANDS Defense TechnicalCarnegie-Mellon University Information CenterSchenley Park LT Judy Crookshanks Cameron Station, Bldg 5Pittsburgh, PA 15213 Chief of Naval Operations Alexandria, VA 22314

OP-112G5 Attn: TCDr. Eugene Charniak Washington, DC 20370-2000 (12 Copies)Brown UniversityComputer Science Department Dr. Kenneth B. Cross Dr. Thomas M. DuffyProvidence, RI 02912 Anacapa Sciences, Inc. Communications Design Center

P.O. Drawer Q Carnegie-Mellon UniversityDr. Paul R. Chatelier Santa Barbara, CA 93102 Schenley ParkOUSDRE Pittsburgh, PA 15213PentagonWashington, DC 20350-2000 Dr. Mary Cross Barbara Cason

Department of Education Military Educator'sDr. Michelene Chi Adult Literacy Initiative Resource NetworkLearning R & D Center Room 4145 InterAmerica Research AssociatesiUniversity of Pittsburgh 400 Maryland Avenue, SW 1555 Wilson Blvd3939 O'Hara Street Washington. DC 20202 Arlington, VA 22209Pittsburgh, PA 15213 • Edward 3. Eddowes4r. Raymond E. Christal CT/McGraw-Hill Library CmATRA 1301AFHRL/MOE 2500 Garden Road Naval Air StationBrooks AFD, TX 78235 Monterey, CA 93940 Corpus Christi, TX 78419

Dr. Yee-Yeen Chu Dr. John EllisPerceptronics, Inc. CAPT P. Michael Curran Navy Personnel R&D Center21111 Erwin Street Office of Naval Research San Diego, CA 922522oodland Hills, CA 91367-371 800 N. Quincy St.Code 125 Dr. Richard Elster)r. William Clancey Arlington, VA 22217-5000 Deputy Assistant Secretarytanford University of the Navy (Manpower)<nowledge Systems Laboratory Dr. Cary Czichon OASN (M&RA)'01 Welch Road, Bldg. C Intelligent Instructional Sy Department of the Navylalo Alto, CA 94304 Texas Instruments Al Lab Washington, DC 20350-1000

P.O. Box 660245Thief of Naval Education Dallas, TX 75266

and Training Dr. Susan s retsonLiaison Office Bryan Dalleen University of KansasAir Force Human Resource Lab AFHRL/LRT Psychology DepartmentOperations Training Division Lowry AFS, CO 80230 426 FraserWilliams AFS, AZ 85224 LT John Deaton Lawrence, KS 66045

Assistant Chief of Staff ONR Code 125for Research, Development, 800 N. Quincy Street Dr. Randy EnqleTest, and Evaluation Arlington, VA 22217-5000 Department of PsychologyT aaval Education and University of South CarolinaTraining Con and (N-5) Dr. Natalie Dehn Columbia, SC 29206A Pensacola, FL 3250) Department of Computer andInformation Science Lt. Col Rich ntlich

Dr. Allan M. Collins University of Oregon NO, Department of the ArmyBolt Beranek & Newman, Inc. We, 06 97403 OCSA(DACS-D 2)50 Moulton Street Dr.Wshngton, DC 20310Camridge, MA 02136 Artificial IntelligeNe Grou Dr. Susan Ipstein

Dr. John J. Collins Coordinated Science Laoto Hunter CollegeDirector, Field Research University of Illinois 144 S. ountain AveMe0 Office, Orlando Urbana, IL 61601 nontclair, l43 07042NP0C Liaison OfficerWTSC Orlando, FL 32813

Dr. William Epstein Dr. Carl H. Frederiksen Dr. Joseph GoguenUniversity of Wisconsin McGill UniversityW. J. Brogden psychology Bld 3700 McTavish Street Computer Science Laboratory1202 W. Johnson Street Montreal, Quebec H3A 1Y2 SRI International

Madison, WI 53706 CANADA 333 Ravenswood AvenueMenlo Park, CA 94025

ERIC Facility-Acquisitions Dr. John R. Frederiksen4833 Rugby Avenue Bolt Beranek a Nemn Dr. Sherrie GottBethesda, MD 20014 50 Moulton Street AFHRL/MODJ

Cambridge, MA 02138 Brooks AFB, TX 78235

Dr. K. Anders EricssonUniversity of Colorado Dr. Norman Frederiksen Jordan Grafman, Ph.D.Department of Psychology Educational Testing Service 2021 Lyttonsville Road

Boulder, CO 80309 Princeton, NJ 08541 Silver Spring, MD 20910

Dr. Edward Esty Dr. Alfred R. Fregly Dr. Richard H. GrangerDepartment of Education, OEr AFOSR/NL Department of Computer Science

Room 717D Bolling AFB, DC 20332 University of California, Irvine

1200 19th St., NW Irvine, CA 92717Washington, DC 20208 Dr. Bob Frey

Commandant (G-P-l/2) Dr. Wayne GrayDr. Beatrice J. Farr USCG HO Army Research InstituteArmy Research Institute Washington, DC 20593 5001 Eisenhower Avenue5001 Eisenhower Avenue Alexandria, VA 22333Alexandria, VA 22333 Dr. Alinda Friedman

Department of Psychology Dr. James G. GreenoDr. Marshall J. Farr University of Alberta University of California2520 North Vernon Street Edmonton, Alberta Berkeley, CA 94720Arlington, VA 22207 CANADA T6G 2E9

H. William GreenupDr. Pat Federico Education Advisor (E031)CDe 511 eInformation Technology Education Center, MCDECCode 511 Applications Division Quantico, VA 22134nD92152-6800 Admiralty Research EstablishSan Diego, CA Portsdown, Portsmouth P06 4A Dipl. Pad. Michael w. Habon

Dr. Jerome A. Feldman UNITED KINGDOM Universitat DusseldorfUniversity of Rochester ErziehungswissenschaftlichesComputer Science Department Dr. R. Edward Geiselman Universitatsstr. 1Rochester, NY 14627 Department of Psychology D-4000 Dusseldorf 1

University of California WEST GERMANY

Dr. Paul Feltovich Los Angeles, CA 90024Southern Illinois University Prof. Edward HaertelSchool of Medicine Dr. Michael Genesereth School of EducationMedcal Education Department Stanford University Stanford UniversityP.O. Eox 3926 Computer Science Department Stanford, CA 94305

Springfield, IL 62708 Stanford, CA 9.4305Mr. Wallace Feurzeig Dr. Dedre Gentner Haff Resources, Inc.

Educational Technology University of Illinois 4918 33rd Road, NorthBolt Beranek & Newman Department of Psychology Arlington, VA 2220710 Moulton St. 603 E. Daniel St.Cambridge, MA 02238 Champaign, IL 61820

Dr. Craig I. Fields Dr. Robert Glaser Prof. of Education & PsARPA Learning Research University of Massachusetts1400 Wilson Blvd. & Development Center at Amherst

Arlington, VA 22209 University of Pittsburgh Hills House3939 O'Hara Street Amherst, MA 01003

Dr. Gerhard Fischer Pittsburgh, PA 15260University of Colorado Dr. Cheryl HamelDepartment of Computer Science Dr. Arthur M. G;Lenberg NTSCBoulder, CO 80309 University ot Wisconsin Orlando, FL 32813

W. J. Brogden Psychology Bldg.J. D. Fletcher 1202 W. Johnson Street Dr. Bruce W. Hamill9931 CotsLca Street Madison, WI 53706 Johns Hopkins UniversityVienna ';A 22180 Applied Physics Laboratory

Dr. Marvin D. Glock Johns Hopkins Road

Dr. Linda Flower 13 Stone Hall Laurel, MD 20707

Carnegie-Mellon University Cornell UniversityDepartment of English Ithaca, MY 14853 Dr. Ray HannapelPittsburgh, PA 15213 Scientific and Engineering

Dr. Gene L. Gloye personnel and EducationDr. Kenneth D. ForbuS Office of Naval Research National Science FoundationUniversity of Illinois Detachment Washington, DC 20550Department of Computer Science 1030 Z. Green Street1304 est Springfield Avenue Pasadena, CA 91106-2485Urbana, IL 61801 Stevan Harnad

Dr. Sam Glucksberg Editor, The Behavioral andDr. Marbara A. rox Department of Psychology Brain SciencesUniversity of Colorafo Princeton University 20 Nassau Street, Suite 240Deparment of Lingu stics Princeton, W4 08540 Princeton, 14 08540BOide, CO 80309

Dr. James Howard Dr. Richard JohnsonMr. William Hartung Dept. of Psychology Boise State UniversityPEAM Product Manager Human Performance Laboratory Simplot/MicronArmy Research Institute Catholic University of Technology Center5001 Eisenhower Avenue America Boise, ID 83725Alexandria, VA 22333 Washington, DC 20064

CDR Tom JonesDr. Wayne Harvey Dr. Steven Hunka ONR Code 125SRI International Department of Education 800 N. Quincy Street333 Ravenswood Ave. University of Alberta Arlington, VA 22217-5000Room B-S324 Edmonton, AlbertaMenlo Park, CA 94025 CANADA D.ol AJn

~Dr. Douglas A. JonesDr. Reid Hastie P.O. Box 6640Northwestern University Dr. Earl Hunt LawrencevilleDepartment of Psychology Department of Psychology NJ 08648Evanston, IL 60201 University of Washington

Seattle, WA 98105 Col. Dominique Jouslin de NoProf. John R. Hayes Etat-Major de l'Armee de TerCarnegie-Mellon University Dr. Ed Hutchins Centre de Relations HumainesDepartment of Psychology Intelligent Systems Group 3 Avenue Octave GreardSchenley Park Institute for 75007 ParisPittsburgh, PA 15213 Cognitive Science (C-015) FRANCE

UCSDDr. Barbara Hayes-Roth La Jolla, CA 92093 Dr. Marcel JustDepartment of Computer Science Carnegie-Mellon UniversityStanford University Dr. Dillon Inouye Department of PsychologyStanford, CA 95305 WICAT Education Institute Schenley Park

Provo, UT 84057 sburgh, Pa 1 3

Dr. Frederick Hayes-Roth Pittsburgh, PA 15213Teknowledge Dr. Alice Isen525 University Ave. Der-irtment of Psychology Dr. Milton S. KatzPalo Alto, CA 94301 Univ"ersity of Maryland Army Research Institute

Catonsville, MD 21228 5001 Eisenhower Avenue

Dr. Joan I. Heller Alexandria, VA 22333505 Haddon Road Dr. Zachary JacobsonOakland, CA 94606 Bureau of Management Consulting Dr. Scott Kelso

365 Laurier Avenue West Haskins Laboratories,Dr. James Hendler Ottawa, Ontario KlA 05 270 Crown StreetDept. of Computer Science CANADA New Haven, CT 06510University of MarylandCollege Park, MD 20742 Dr. Robert Jannarone Dr. Dennis Kibler

Department of Psychology University of CaliforniaUniversity of South Carolina Department of Information

Dr. Jim Hollan Columbia, SC 29208 and Computer ScienceIntelligent Systems Group Irvine, CA 92717Institute for Dr. Claude Janvier

Cognitive Science (C-015) Directeur, CIRADE Dr. David KierasUCSD Universite' du Quebec a Montreal University of MichiganLa Jolla, CA 92093 Montreal, Quebec H3C 3P8 Technical Communication

CANADA College of EngineeringDr. Melissa Holland 1223 E. Engineering BuildingArmy Research Institute for COL Dennis W. Jarvi Ann Arbor, MI 48109

Behavioral and Social Sci Commander5001 Eisenhower Avenue AFHRL Dr. Peter KincaidAlexandria, VA 22333 Brooks AFB, TX 78235-5601 Training Analysis

& Evaluation GroupDr. Robert W. Holt Dr. Robin Jeffries Department of the NavyDepartment of psychology Hewlett-Packard Laboratories Orlando, FL 32813George Mason University P.O. Box 104904400 University Drive Palo Alto, CA 94303-0971 Dr. Walter KintschFairfax, VA 22030 Department of Psychology

Dr. Robert Jernigan University of ColoradoDr. Keith Holyosk Decision Resource Systems Campus Box 345University of Michigan 5595 Vantage Point Road Boulder, CO 80302Human Performance Center Columbia, MD 21044330 Packard Road Dr. Paula KirkAnn Arbor, MI 48109 Oakridge Associated UniversitiesMargaret Jerome University Programs DivisionProf. Lutz F. Hornke c/o Dr. Peter Chandler P.O. Box 117Institut fur Psychologie 83, The Drive Oakridge, TN 37831-0117RWTH Aachen HoveJaegerstrasse 17/19 Sussex Dr. David KlahrD-5100 Aachen UNITED KINGDOM Carnegie-Mellon UniversityWEST GERMANY Department of Psychology

Dr. Joseph E. Johnson Schenley ParkMr. Dick Hoshaw Assistant Dean forOP-135 Graduate Studies Pittsbugh, PA 15213Arlington Annex College of Science and Math*Room 2834 University of South CarolinaWashington, DC 20350 Columbia, SC 29208

Dr. Mazie Knerr Dr. Jim Levinprogram Manager University of California Dr. Richard E. MayerTraining Research Division Laboratory for Comparative Department of PsychologyHumRRO Human Cognition University of California1100 S. Washington D003A Santa Barbara, CA 93106Alexandria, VA 22314 La Jolla, CA 92093

Dr. Janet L. xolodner Dr. John LevineGeorgia Institute of Technology Learning R&D Center Dr. James McBride

School of Information University of Pittsburgh Psychological Corporation

& Computer Science Pittsburgh, PA 15260 c/o Harcourt, Brace,

Atlanta, GA 30332 Javanovich Inc.1250 West 6th Street

Dr. Stephen Kosslyn Dr. Michael Levine San Diego, CA 92101Harvard University Educational Psychology1236 William James Hall 210 Education Bldg. Dr. David J. McGuinness33 Kirkland St. University of Illinois Gallaudet CollegeCambridge, MA 02138 Champaign, IL 61801 800 Florida Avenue, N.E.washington, DC 20002

Dr. Kenneth Kotovsky Dr. Charles Lewis

Department of Psychology Faculteit Sociale Wetenschappen Dr. Kathleen McKeownCommunity College of Rijksuniversiteit Groningen Columbia UniversityAllegheny County Oude Boteringestraat 23 Department of Computer Science800 Allegheny Avenue 9712GC Groningen New York, NY 10027

Pittsburgh, PA 15233 The NETHERLANDSDr. Joe McLachlan

Dr. David H. Krantz Dr. Clayton Lewis Navy Personnel R&D Center2 washington Square village University of Colorado San Diego, CA 92152-6800Apt. # 15J Department of Computer ScienceNew York, NY 10012 Campus Box 430 Dr. James McMichael

Boulder, CO 80309 Assistant for MPT Research,Dr. Benjamin Kuipers Development, and Studies

University of Texas at Austi Library OP 017

Department of Computer Scien Naval War College O 1B7

T.S. Painter Hall 3.28 Newport, RI 02940 Washington, DC 20370

Austin, Texas 78712 Dr. Barbara MeansLibrary Human ResourcesDr. David R. Lambert Naval Training Systems Cente Research organization

Naval Ocean Systems Center Orlando, FL 32813 1100 South WashingtonCode 441T Alexandria, VA 22314271 Catalina Boulevard Dr. Charlotte LindeSan Diego, CA 92152-6800 SRI International Dr. Douglas L. Medin

333 Ravenswood Avenue Department of Psychology

Dr. Pat Langley Menlo Park, CA 94025 University of IllinoisUniversity of California 603 E. Daniel StreetDepartment of Information Dr. Marcia C. Linn Champaign, IL 61820

and Computer Science Lawrence Hall of ScienceIrvine, CA 92717 University of California Dr. Arthur Melmed

M. Diane Langston Berkeley, CA 94720 U. S. Department of Education

communications Design Centei Dr Robert Linn 724gBrown

Carnegie-Mellon University College of Education Washington, DC 20208

Schenley Park University of Illinois Dr. Jose Mestrepittsburgh, PA 15213 Urbana, IL 61801 Department of Physics

Dr. Jill Larkin DHasbrouck Laboratory

Carnegie-Mellon University Dr. Don Lyon University of MassachusettsCarege-Mlln nivrstyP.O. Box 44 Amherst, MA 01003

Department of Psychology Higley, AZ 85236Pittsburgh, PA 15213

Dr. Jane Malin Dr. Al MeyrowitzDr. Robert Lawler MaOl Code SR 111 office of Naval ResearchInformation Sciences, FRL NASA Johnson Space Center Code 1133GTE Laboratories, Inc. Houston, TX 77058 800 N. Quincy40 Sylvan Road Arlington, VA 22217-5000Waltham, MA 02254 Dr. William L. Maloy r. Ryszard S. Michalski

Dr. Paul E. Lehner Chief of Naval Education University of IllinoisPAR Technology Corp. and Training Department of Computer Scien7926 Jones Branch Drive Naval Air Station 1304 West Springfield AvenueSuite 170 Pensacola, FL 32508 Urbana, IL 61801McLean, VA 22102

Dr. Sandra P. Marshall Prof. D. MichieDr. Alan M. Lesgold Dept. of Psychology The Turing InstituteLearning R&D Center San Diego State University 36 North Hanover StreetUniversity of Pittsburgh San Diego, CA 92182 Glasgow Gl 2AD, Scotlandpittsburgh, PA 15260 UNITED KINGDOMDr. Mantofi K. Matthews UIE IGO

Dr. Alan Leshner Department of Computer ScienDeputy Division Director University of South Carolina Dr. George A. Miller

9havioral and Neural Sciences Columbia, SC 29208 Department of Psychology

mational Science Foundation Princeton University

1800 G street Princeton, iv 08540

VaShington, DC 20550

Leadership Management Educat Office of Naval Research,Dr. James R. Miller and Training Project Offi code 1133

MCC Naval medical Comand 800 N. Quincy Street

9430 Research Blvd. Code 05C Arlington, VA 22217-5000Echelon Building #1, Suite 2 washington, DC 20372Austin, TX 78759 Office of Naval Research,

Dr. mark Miller Technical Director,

Code 1142

Dr.Makiter*Th ht Navy Health Research Ctr. 800 N. Quincy St.

Computer*Thought Corporation P.O. Box 85122 Arlington, VA 22217-5000

1721 West Plano Parkway San Diego, CA 92138Plano, TX 75075 Office of Naval Research,

Dr. Andrew R. Molnar Dr. T. Niblett

Code 1142EP

Scientific and Eineering The Turing Institute 800 N. Quincy Street

ennel and Egeerin 36 North Hanover Street Arlington, VA 22217-5000

Personnel and Education Glasgow GI 2AD, Scotland

National Science Foundation UNITED KINGDOM office of Naval Research,Washington, DC 20550 Cod* 1142PT

Dr. William Montague Dr. Richard E. Nisbett 800 N. Quincy StreetD. Wiv t Arlington, VA 22217-5000NPRDC Code 13 University Of Michigan ( ois

San Diego, CA 92152-6800 Institute for Social Researc (6 copies)

Dr. Tom Moran 48109 Director, Technology Progra

Xerox PARC Ann Arbor, M0 Office of Naval Research

3333 Coyote Hill Road Dr. Mary Jo Nissen Code 12

Palo Alto, CA 94304 Univetsity of Minnesota 800 North Quincy Street

N218 Elliott Hall Arlington, VA 22217-5000

Dr. Melvyn Moy Minneapolis, MN 55455Navy Personnel R & D Center

Special Assistant for MaritSan Diego, CA 92152-6800 Dr. Donald A. Norman cor Ma

Institute for Cognitive Scie Corps Matters,

D l nUniversity of California ONR Code 00NC

eaiDr. Allen Munro La Jolla, CA 92093 800 N. Quincy St.Behavioral Technology Arlington, VA 22217-5000

Laboratories - usC Director, Training Laborator

1845 S. Elena Ave., 4th Floor NPRDCo 0n) Assistant for Long Range

Redondo Beach, CA 90277 San Diego, CA 92152-6800 Requirements,

Director, CNO Executive Panel

Drecion Director, Manpower and perso OP 00KDecision Support LabOratory, 2000 North Beauregard Stre

Systems Division, NNPC NPRDC (Code 06) Alexandria, VA 22311

Naval Military Personnel Command San Diego, CA 92152-6800

N-164 Assistant for MPT Research

Washington, DC 20370 DirectOr, Human Factors Development and Studies

D t& Organizational Systems Op 01B7

Director, NPRDC (Code 07) Washington, DC 20370Distribution Department, NPC San Diego, CA 92152-600

N-4 Dr. Judith Orasanu

Washington, DC 20370 Fleet Support Office, Army Research Institute

NpRDC (Code 301) 5001 Eisenhower Avenue

Director, San Diego, CA 92152-6800 Alexandria, VA 22333Overseas Duty Support

Program, NMPC Dr. Jesse Orlansky

N-62 Library, NPRDC Institute for Defense Anal

Washington, DC 20370 Code P201L 1801 N. eauregard St.

Head, HRM Operations Branch, San Diego, CA 92152-6800

Alexandria, VA 22311

NMPC Commanding officer, Prof. Seymour Papert

N-62F Naval Research Laboratory 20C-109Washington, DC 20370 Code 2627 Massachusetts Institute

Washington, DC 20390 of Technology

Assistant for Evaluation, acambridge, MA 02139

Analysis, and MIS, NMPC Dr. Harold F. O'Neil, Jr.N-6C School of Education - WPM e01 CDR R. T. Parlette

Washington, DC 20370 Department of Educational Chief of Naval Operations

SePsycholoiy & Technology OP-112GSpec. Asst. for Research, Exper- university of Southern California Washington, DC 20370-26,

mental & Academic Programs, LoS Angeles, CA 90089-0031NTTC (Code 016) Dr. James Paulson

NAS Memphis (75) Dr. Stellan Ohlsson Department of PsychologyMillington, TN 38054 Learning R & D Center Portland State UniversLtt

university of Pittsburgh P.O. Box 751Director, 3939 O'nara Street Portland, OR 97207

Research & Analysis Div., Pittbrgh, PA 15213NAVCRUITCON Code 22 Dr. Douglas Pearse

4015 Wilson Blvd. Director, Research Programs, DCIUtArlington, VA 22203 Office of Naval Research Box 2000

$00 North Quincy street Dwsiw nai

Technical Director Arlington, VA 222175000Navy Malth Research CenterP.O. Box 85122San Diego, CA 92136

Dr. Virginia S. Pendergrass Dr. Mary C. Potter Dr. Donald RubinCode 711 Department of Psychology Statistics DepartmentNaval Training System Center MIT (E-10-032) Science Center, Room 60SOrlando, FL 32613-7100 Cambr 1g/e, MA 02139 1 Oxford Street

Harvard UniversityDr. Joseph Psotka Cambridge, MA 02138

Dr. Robert Penn ATTN: PERI-ICNPRDC Army Research Institute Dr. David RumelhartSan Diego, CA 92152-6800 5001 Eisenhower Ave. Center for Human

Alexandria, VA 22333 Information Processing

Dr. Nancy Pennington Univ. of California

University of Chicago Dr. Roy Rada La Jolla, CA 92093

Graduate School of Business National Library of Medicine1101 E. 58th St. Bethesda, M0 20894 Ms. Riitta Ruotsalainen

Chicago, IL 60637 General Headquarters

military Assistant for Training and Dr. Lynne Reder Training SectionDepartment of Psychology Military Psychology Office

Personnel Technology, Carnegie-Mellon University PL 919OUSD (R A E) Schenley Park SF-00101 Helsinki 10, FINLAN

Room 3D129, The Pentagon Pittsburgh, PA 15213Washington, DC 20301-3080 Dr. Michael J. Samet

Perceptronics, IncDr. Ray Perez Brian Raiser 6271 Variel AvenueARI (PERI-II) Department of Psychology woodland Hills, CA 913645001 Eisenhower Avenue Princeton UniversityAlexandria, VA 2233 Princeton, NJ 08544 Dr. Roger Schank

Yale UniversityDr. David N. Perkins Computer Science DepartmentEducational Technology Center Dr. James A. Reggla P.O. Box 215837University of Mary3.arn P.O. Co 6520

337 Gutman Library School of Medicine New Haven, C? 06520Appian Way Department of Neurology Dr. W. L. ScherlisCambridge, MA 02138 22 South Greene Street Dr.t. L.mpuerie

LCDR Frank C. Petho, MSC, USN Baltimore, MD 21201 Dept. of Computer Science

CNATA Coe N3, Blg. 1Carnegie-Mellon UniversityCNATRA Code N36, Bldg. I Schenley ParkNAS Pittsburgh, PA 15213

Corpus Christi, TX 78419 Naval School of Health SciencesNational Naval Medical Center Mrs. irgitte Schneidelbach

Administrative Sciences Department, Bldg. 141 Frsertte forn e dercNalPsrautScolWashington, DC 208114 Forsvarets Center for Leders

Naval Postgraduate School WChristianshavns Voldgede 8Monterey, CA 93940 Dr. Fred Reif 1424 Kobenhavn K

Department of Computer Science, Physics Department DENMARK

Naval Postgraduate School University of California Dr. Walter SchneiderMonterey, CA 93940 Berkeley, CA 94720 Learning R&D Center

Department of operations Research, Dr. Lauren Resnick University of Pittsburgh

Naval Postgraduate School Learning R a D Center 3939 O'Hara Street

Monterey, CA 93940 University of Pittsburgh Pittsburgh, PA 15260

3939 O'Hara StreetPittsburgh, PA 15213 Dr. Alan H. Schoenfeld

Dr. Tjeerd Plomp University of California

Twente University of Technol Dr. Mary S. Riley Department of EducationDepartment of Education Program in Cognitive Science Berkeley, CA 94720P.O. Box 217 Center for Human Information7500 AE ENSCHEDE Processing Dr. Janet Schofield

University of California Learning R&D CenterSLLa Jolla, CA 92093 University of Pittsburgh

S THA William Rizzo Pittsburgh, PA 15260

Dr. Martha Polson Code 712 Dr. Marc SebrechtsDepartment of Psychology Naval Training Systems Center Department of PsychologyCampus Box 346 Orlando, FL 32813 Wesleyan UniversityUniversity of Colorado Miiddletom, CT 06475Boulder, CO 80309 Dr. Linda G. Roberts

cience, Education, and Dr. Judith SegalDr. Peter Polson Transportation Program OuRxUniversity of Colorado Office of Technology Assessment 555 New Jersey Ave., NWDepartment of Psychology Congress of the United States Washington, DC 20208Boulder, CO 80309 Washington, DC 20510

Dr. Robert J. SeidelDr. Steven E. Poltrock Dr. Andrew 1. Rose US Amy Research InstituteMCC American Institutes 5001 Eisenhower Ave.

9430 Research Blvd. for Research Alexandria, VA 22333Echelon Bldg *l 1055 Thomas Jefferson St., NWAustin, TX 78759-6509 Washington, DC 20007 Dr. Ramsay W. Selden

Assessamet CenterDr. Harry E. Pople Dr. William B. Rouse CCSSOUniversity of Pittsburgh Search Technology, Inc. Suite 379Decision Systms Laboratory 25-b Technology Park/Atlanta 400 f. Capitol, NW1360 Scaife Hall Norcross, GA 30092 Washington, DC 20001Pittsburgh, PA 15261

Dr. W. Steve Sellman Dr. Linda B. SmithOASD(MRA&L) Department of Psychology Navy Personnel R&D CenterOD(P&)Indiana UniversityNayPronlRD etr

29269 The Pentagon Indintoniersity San Diego, CA 92152-6800

Washington, DC 20301 Bloomington, IN 47405

Dr. Alfred F. Smode Dr. John TangneyDr. Sylvia A. S. Shafto Senior Scientist AFOSR/NL

OERI Code 07A Bolling AFB, DC 20332

555 New Jersey Ave., NW Naval Training Systems CenterWashington, DC 20208 Orlando, FL 32813 Dr. Kikumi Tatsuoka

CERL

Dr. T. B. Sheridan Dr. Richard E. Snow 252 Engineering Research

Dept. of Mechanical Engineering Department of Psychology LaboratoryMIT Stanford University Urbana, IL 61801

Cambridge, MA 02139 Stanford, CA 94306~Dr. Martin M. Taylor

Dr. Ben Shneiderman Dr. Elliot Soloway DcrEm

Dept. of Computer Science Yale University Box 2000

University of Maryland Computer Science Department DOwAsview, OntarioCollege Park, MD 20742 P.O. Box 2158

Dr. Ted Shortliffe New Haven, CT 06520 Dr. Perry W. ThorndykeComputer Science Department Dr. Richard Sorensen FC CorporationStanford University Navy Personnel R&D Center Central Engineering Labs

Stanford, CA 94305 San Diego, CA 92152-6800 Santa Clara, CA 95052

Dr. Lee Siulman Dr. Kathryn T. SpoehrStanford University Brown University Major Jack Thorpe

1040 Cathcart Way Department of Psychology DARPA

Stanford, CA 94305 Providence, RI 02912 1400 Wilson Blvd.Arlington, VA 22209

Dr. Randall Shumaker James J. Staszewski Dr. Douglas TowneNaval Research Laboratory Research Associate Behavioral Technology Labs

Code 7510 Carnegie-Mellon University 1845 . Elena Ave.

4555 Overlook Avenue, S.W. Department of Psychology Redondo Beach, CA 90277

Washington, DC 20375-5000 Schenley Park

Dr.Pittsburgh, PA 15213 Dr. Amos Tversky

Carnegie-Mellon University Dr. Marian Stearns Stanford University

Department of Psychology SRI International Dt.nof Psychology

Schenley Park 333 Iaventoav Stanford, CA 94305

Pittsburgh, PA 15213 Room B-S324 Dr. James TweeddaleMenlo Park, CA 94025 Technical DirectorDr. Herbert A. Simon Navy Personnel R&D Center

Department of Psychology Dr. Robert Sternberg San Diego, CA 92152-6800

Carnegie-Mellon University Department of Psychology

Schenley Park Yale UniversityPittsburgh, PA 15213 YaeUiest Dr. Paul TwohigBox IdA, Yale Station Army Research Institute

New Haven, CT 06520 5001 Eisenhower AvenueLTCOL Robert Simpson AeadlV 23

Defense Advanced Research Dr. Albert StevensProjects Administration Bolt Beranek & Newman, Inc. Headquarters, U. S. Marine Corp

1400 Wilson Blvd. 10 Moulton St. Code MPI-20

Arlington, VA 22209 Cambridge, MA 02238 Washington, DC 20380

Dr. Zita M Simutis Dr. Paul J. Stichainstructional Technology Senior Staff Scientist Dr. Kurt Van Lehn

Systems Area Training Research Division Department of PsychologyARI HumRRO Carnegie-Mellon University5001 Eisenhower Avenue 1100 S. Washington Schenley ParkAlexandria, VA 22333 Alexandria, VA 22314 Pittsburgh, PA 15213

Dr. H. Wallace Sinaiko Dr. Thomas Sticht Dr. Beth WarrenManpower Research Navy Personnel R&D Center Bolt Beranek & Newman, Inc.

and Advisory Services San Diego, CA 92152-6800 50 Moulton StreetSmithsonian Institution Cambridge, MA 02138801 North Pitt Street Dr. David StoneAlexandria, VA 22314 KAJ Software, Inc. Dr. David J. Weiss

3420 East Shea Blvd. N660 Elliott Hall

Dr. Derek Sleeman Suite 161 University of MinnesotaStanford University Phoenix, AZ 85028 75 E. River RoadSchool of Education Minneapolis, MN 55455Stanford, CA 94305 Cdr Michael Suman, PD 303

Naval Training Systems Cento Roger Weissinqer-Daylon

Dr. Charles F. Smith Code N51, Comptroller Department of AdministrativeNorth Carolina State Univers Orlando, FL 32813 SciencesDepartment of Statistics Naval Postgraduate School

Raleigh, NC 27695-8703 Dr. Hariharan Swaminathan Monterey, CA 93940Dr. Edward 9. Smith Laboratory of Psychometric a D. D WeitzmanDr. ead a Sm Evaluation Research MITRDolt lranek a Newman, Inc. School of Education 1 o M nS

50 Moulton Street University of Massachusetts 1020 Dolly Madison Blvd.Carids, A 02136 Amherst, MA 01003 MacLen, Va 22102

Dr. Keith T. Wescourt Dr. Joseph L. YoungFMC Corporation Memory G CognitiveCentral Engineering Labs Processes1185 Coleman Ave., Box 580 National Science FoundationSanta Clara, CA 95052 Washington, DC 20550

Dr. Douglas Wetzel Dr. Steven ZornetzerCode 12 Office of Naval ResearchNavy Personnel R&D Center Code 1140San Diego, CA 92152-6800 800 N. Quincy St.

Arlington, VA 22217-5000Dr. Barbara WhiteBolt Beranek & Newman, Inc. Dr. Michael J. Zyda10 Moulton Street Naval Postgraduate SchoolCambridge, MA 02238 Code 52CK

Monterey, CA 93943-5100LCDR Cory deGroot WhiteheadChief of Naval OperationsOP-112G1Washington, DC 20370-2000

Dr. Michael WilliamsIntelliCorp1975 El Camino Real WestMountain View, CA 94040-2216

Dr. Hilda WingArmy Research Institute5001 Eisenhower Ave.Alexandria, VA 22333

A. E. WinterbauerResearch AssociateElectronics DivisionDenver Research InstituteUniversity ParkDenver, CO 80208-0454

Dr. Robert A. WisherU.S. Army Institute for the

Behavioral and Social Sciences5001 Eisenhower AvenueAlexandria, VA 22333

Dr. Martin F. WiskoffNavy Personnel R & D CenterSan Diego, CA 92152-6800

Dr. Merlin C. WittrockGraduate School of EducationUCLALos Angeles, CA 90024

Mr. John H. WolfeNavy Personnel R&D CenterSan Diego, CA 92152-6800

Dr. Wallace Wulfeck, IIINavy Personnel R&D CenterSan Diego, CA 92152-6800

Dr. Joe YasatukeAFHRL/LRTLowry AFB, CO 80230

Dr. Masoud YazdaniDept. of Computer ScienceUniversity of ExeterExeter EX4 40LDevon, ENGLAND

Major Frank Yohannan, USMCHeadquarters, Marine Corps(Code PI-20)Washington, DC 20380

Kr. Carl YorkSystem Development Foundation161 Lytton AvenueSuite 210Palo Alto, CA 94301

a'C