Problem of Extracting the Knowledge of Experts Fkom the Perspective of Experimental Psychology

AI Magazine Volume 8 Number 2 (1987) (© AAAI)

The ‘Problem of Extracting the Knowledge of Experts fkom the Perspective of Experimental Psychology

RobertR.Hoffman

or perceptual and conceptual and represent their special knowledge The first step in the development of an problems requiring the skills of . . . [It] may take several months of the expert system is the extraction and charac- an expert, expertise is rare, the expert’s time and even more of the terization of the knowledge and skills of an expert’s knowledge is extremely system builder’s” (p. 264). Three years expert. This step is widely regarded as the detailed and interconnected, and our later, Duda and Shortliffe (1983) major bottleneck in the system develop- scientific understanding of the echoed this lament: “The identifica- ment process To assist knowledge engi- expert’s perceptual and conceptual tion and encoding of knowledge is one neers and others who might be interested in the development of an expert system, I processes is limited. Research on the of the most complex and arduous offer (1) a working classification of meth- skills of experts in any domain affords tasks encountered in the construction ods for extracting an expert’s knowledge, an excellent opportunity for both of an expert system” (p. 265). (2) some ideas about the types of data that basic and practical experimentation. Some common phrases that occur the methods yield, and (3) a set of criteria My investigations fall on the experi- in the literature are “knowledge acqui- by which the methods can be compared mental psychology side of expert sys- sition is the time-critical component” relative to the needs of the system develop- tem engineering, specifically the prob- (Freiling et al. 1985), and “extracting er The discussion highlights certain issues, lem of generating methods for extract- and articulating the knowledge is the including the contrast between the empiri- ing the knowledge of experts. I do not most important phase” (Gevarter cal approach taken by experimental psy- review the relevant literature on the I984), and “the process of extracting chologists and the formalism-oriented approach that is generally taken by cognition of experts.l I want to share knowledge can be laborious” (Quinlan cognitive scientists a few ideas about research methods 1984). I encountered passages of this that I found worthwhile as I worked type in every major review article or with expert interpreters of aerial pho- textbook on expert systems that has tographs and other remotely sensed come to my attention (for example, data [Hoffman 1984) and on a project Bramer 1982; Denning 1986; Hayes- involving expert planners of airlift Roth, Waterman, and Lenat 1983; Mit- operations (Hoffman 1986). These tal and Dym 1985; Raulefs 1985; ideas should be useful to knowledge Weiss and Kulikowski 1984). Such engineers and others who might be articles dutifully assert that “the first interested in developing an expert sys- step is to get the knowledge” (Freiling tem. et al. 1985, p. 152). What follows, however, is typically a discussion of TheBottleneck abstract inference strategies, details of In generating expert systems, one Lisp code, and descriptions of system architecture rather than an answer to must begin by characterizing the the question of exactly how to get the knowledge of an expert. As far as I can knowledge. tell from the literature on expert sys- Some papers have titles that explic- tems, getting this characterization is itly suggest they deal with the nuts the significant bottleneck in the sys- and bolts of how to extract an expert’s tem development process. For exam- knowledge (for example, “Acquisition ple, from their experience in the of Knowledge from Domain Experts,” development of PROSPECTOR, Duda Friedland 1981; see also Nii and Aiello and Gasbnig (1981) concluded that ‘I... 1979; Politakis and Weiss 1984). How- something must be done to shorten ever, such papers deal mostly with the the time needed to interview experts

SUMMER 1987 53 nosing (interpretation of data), planning, designing, and explaining. Psychologically, the tasks that an expert typically performs involve at least the following: (1) the analysis of complex stimuli into relevant features or cues based on a process psychologists call “perceptual learning,” (2) the analysis of conceptual categories in terms of the relevant features (the perception of similarities and differences), (3) the analysis of the features and the categories in terms of relevant underlying causal laws (involving “concept- formation processes”), and (4) abilities to infer and test hypotheses. Although these (and probably other) psychological factors are involved, the products that result from familiar tasks bear on these tasks and might not actually be very informative about the expert’s reasoning. For example, the standard method of aerial-photo interpretation (terrain analysis) pro- vides lots of information about the land that appears in an aerial photo but says precious little about how the expert arrived at the description (cf. Mintzer and Messmore 1984; Way Table 1. Types of Methods That Can Be Used to Extract the 1978). Such is also probably true for Knowledge of an Expert. the kinds of diagnostic notations made by radiologists when they interpret X representation of knowledge. III short, type of method involves studying the rays (Feltovich 1981). Nevertheless, an apparently little or no systematic expert’s performance on the “tough analysis of familiar tasks can be very research has been conducted on the cases” that sometimes occur. The cate- beneficial because it can give the question of how to elicit an expert’s gorization shown in table 1 forms the knowledge engineer a feel for the kinds knowledge and inference strategies topical organization of this paper. Fol- of knowledge and skill involved in the (Duda and Shortliffe 1983; Hartley lowing a brief discussion of each of domain. 1981).2 How can one find good tasks these method types is an examination An analysis of familiar tasks (includ- for extracting the knowledge of an of some ways in which the data they ing an analysis of available texts and expert? How can various tasks be yield can be analyzed and the various technical manuals) can be used to gen- compared? Can tasks be tailored to methods compared. erate a “first pass” at a data base. What extract specific subdomains of knowl- the expert knows is represented as a edge within the expert’s broader The Method of Familiar Tasks categorized listing of statements cast domains of knowledge? My research in some sort of formal language (such The method of familiar tasks involves as propositions) using terms and cate- has addressed these questions. studying the expert while he or she is Methods for extracting expert gories that are meaningful and related engaged in the kinds of tasks that are to the domain at hand (Freiling et al. knowledge seem to fall neatly into a usually or typically engaged in. Look- handful of categories, as shown in 1985). Such propositions can express ing across a set of experts’ specific tac- observation statements or facts as well table 1. One obvious methodological tics and procedures, one should see category involves observing the perfor- as implications or potential if-then commonalities in terms of goals, the rules of inference.3 Table 2 presents mance of the expert at the kinds of information the experts like to have tasks the expert is familiar with or example excerpts from the data base of available, and the data or records that the aerial-photo-interpretation project. usually engages in. A second category are produced (Mittal and Dym 1985). of methods is the interview. Artificial In a number of reviews (for example, The Unstructured Interview tasks can be devised that depart from Duda and Shortliffe 1983; Stefik et al. what the expert usually does by limit- 1982), the various tasks that experts As far as I can tell, the development of ing the information that is available to engage in have been analyzed and cate- most existing expert systems started the expert or by constraining the prob- gorized into basic types, such as diag- with unstructured interviews of the lem the expert is to work on. Another

54 AIMAGAZINE experts. Indeed, most system developers have apparently relied exclusively on the unstructured interview method. ROCK FORMS #3--DOME Some developers have even apparently Raised circular, linear or ellipsoid rock taken it for granted that an unstruc- Can be small, compound, or clustered tured interview is the only way to Can have radiating fractures extract expert knowledge (for example, Can be salt, gypsum, or intrusive bedrock Weiss and Kulikowski 1984, p. 105). Radial drainage pattern, annular pattern at base of the slope In an unstructured interview, the knowledge engineer asks more-or-less ROCK TYPE # 1 --FLAT SHALE spontaneous questions of the expert Gently rolling, irregular plain while the expert is performing (or talk- Symmetrical finger ridges ing about) a familiar task [see Freiling Branching rounded hills with saddle ridges et al. 1985). For instance, the inter- Scalloped hill bases viewer might ask the expert a question V- and U-shaped stream gullies such as “How do you know that?” Uniform slope gradients imply homogeneous rock whenever the expert seems to tap into Compound slope gradients imply thick bedding knowledge or make an inference, or Escarpments, very sharp ridges, steep slopes, and steep “Do you want a cup of coffee?” when- pinnacles imply sandy soils ever they get tired. HUMID CLIMATE Presumably, the knowledge engi- Implies rounded hills neer’s prior analysis of the familiar Implies fine dendritic drainage pattern tasks has turned the engineer from a ARID CLIMATE novice into an apprentice (or even a ImP lies steep, rounded hills and ridges with asymmetrical slopes master). Thus, the engineer is trained Implies intermittent drainage ’ - in what to look for during the inter- Implies barren land or shrub land view. Should the training or prepara- Implies light or mottled soil tones tion proceed so far as to turn the knowledge engineer into an expert? I FLUVIAL LANDFORMS # 17--PLAYAS do not know if there can be any princi- Dry, very low relief lakebeds in arid regions pled answer to this question. It has Can include beach ridges been claimed (Davis 1986) that the Few drainage features novice interviewer might be better Irrisiation and intense cultivation able to ask questions about ideas or Scrabbled surface implies alkaline deposits procedures which an expert tends to leave implicit or take for granted (see SOIL TYPE #Z--SILT also Hartley 198 1). Light tones Table 3 presents an example excerpt U-shaped stream gullies from an unstructured interview conducted during the airlift-planning pro- DRAINAGE PATTERNS #5--THERMOKARST ject (Hoffman 1986). This excerpt is Stream gullies form polygons and hexagons, linked by meandering prototypical in that there are inquiries from the interviewer, speech pauses and hesitations, ungrammatical segments, and unintelligible segments. The expert’s monologues in this excerpt are brief. Occasionally, they can be much longer. Knowledge engineers sometimes make an audiotape of the expert’s ruminations; this recording is called a verbal protocol (Ericcson and Simon 1984). The trick is to train the expert into thinking out loud, which requires practice because the act of speaking Table 2. An Example Set of Excerpts from the Aerial-Photo-Interpretation one’s thoughts can actually interfere Data Base (Hoffman 1984). Within each category and subcategory there can with or inhibit trains of thought for appear potential if-then rules (“X implies Y”), as well as facts (observation state- some individuals. Other individuals, ments). Stream gully shapes refer to their cross sections. those who habitually think out loud, might not actually leave a very good

SUMMER 1987 55 record of their reasoning. Audiotapes are necessarily a partial representation of the key information. An expert’s facial expressions and ges- tures can also reveal inference-making processes (McNeil1 and Levy 1982). For instance, in the analysis of 10 hours of unstructured interviews (Hoffman 1986), I found that a majori- ty of the 957 questions asked by the interviewer were not marked by clear- ly rising intonation. Typically, the interviewer was repeating some bit of knowledge that had just been asserted by the expert. The intention was that the expert would either confirm, deny, or qualify the interviewer’s statement. The pragmatics of such utterances as “ah” and “urn” (and their gestural counterparts) are only4 captured as ephemera on audiotape. The moral here is that the interviewer should always take copious notes and not rely passively on audiotapes. Furthermore, it is prudent to distinguish between interviews, which are one-on-one, and discussion groups. In general, discussion group sessions need not be recorded. An important exception to this rule is the recording of the discussions of teams of experts who are working on tough cases (to be discussed later). Table 3. Example Excerpt from an Unstructured Interview. ber of effects on the data base. It can during the interpretation of photos. In The Structured Interview lead to (I) the addition or deletion of the limited-information task, such entries, (2) the qualification of entries, contextual information can be with- A powerful and yet conceptually sim- (3) the reorganization of the hierarchi- held, forcing the expert to rely heavily ple alternative to unstructured inter- cal or categorical structure of the data upon (and hence provide additional views is the structured interview. In a base, or (4) the addition or deletion of evidence about) their knowledge and sense, the structured interview com- categories. The result is a second pass reasoning skills. bines an analysis of familiar tasks at the data base. In general, experts do not like it with an unstructured interview. In when you limit the information that is order to add structure to an interview, Limited- available to them. It is commonly the knowledge engineer initially assumed in familiar tasks that all the makes a first pass at a data base by Information Tasks relevant information is available and analyzing the available texts and tech- Limited-information tasks represent that the expert’s judgments are cor- nical manuals, or by conducting an the application of basic scientific rect. Some experts would rather give unstructured interview. The expert method: to understand how something no judgment at all than give one that then goes over the first-pass data base works in nature, we tinker with it. is based on incomplete information. It one entry at a time, making com- Limited-information tasks are similar is important when instructing the ments on each one. Recording this to the familiar tasks, but the amount expert to drive home the point that the process is not necessary because the or kind of information that is available limited-information task is not a chal- knowledge engineer can write changes to the expert is somehow restricted. lenge of their ego or of their expertise: and notes on a copy of the printout of For example, an expert radiologist The goal of the task is not to deter- the first-pass data base. might like to have available all sorts of mine how “clever” the expert is. A structured interview in one way information about a patient’s medical Once adapted to the task, experts or another forces the expert to system- history before interpreting an X ray. can provide a wealth of information. atically go back over the knowledge. Many expert aerial-photo interpreters The limited-information task is espe- Any given comment can have a num- like to have all sorts of maps available cially useful for revealing an expert’s

56 AI MAGAZINE strategies (as opposed to factual weather forecasters (Moninger and judgments, the limited or constrained knowledge). The incompleteness of Stewart 1987). The experts have to do tasks can be good sources of informa- the information affords the formula- a forecast but in displaced real time. In tion about the expert’s reasoning tion of hypotheses (rather than final this simulated familiar task, the clock strategies. judgments), strategic thinking (What if can be stopped at any point and the . . . ?), and the use of heuristics [rules expert queried with regard to reason- of thumb). ing strategies or subdomains of knowl- The Method of The limited-information task can be edge as incoming data are monitored used to provide information about sub- and predictions are made. Tough Cases domains of the expert’s knowledge, to Research on an expert’s reasoning fill in any gaps in the data base or a set encounters special difficulty during its of inference rules. For example, in the The Method of Scenarios later phases when some of the expert’s aerial-photo-interpretation project, the knowledge and methods of reasoning structured interview with one particu- While analyzing cases, experts often have already been described. The task lar expert did not yield much informa- draw analogies to previously encoun- that now confronts the knowledge tion about certain geological forms. tered situations or cases. A given case engineer is to get evidence about the Hence, other experts could be run is explored in terms of any relevant or subtle or refined aspects of the expert’s through a limited-information task salient similarities and differences (at reasoning. One needs to know some- that involved photos of precisely these either perceptual or conceptual levels) thing about the expert’s reasoning that geological forms. relative to a previously encountered is not already known. case. For example, Klein (1987) exam- Subtle or refined aspects of an ined how expert mechanical and elec- expert’s reasoning are often manifest- Constmined- trical engineers design new compo- ed when an expert encounters a tough nents by analogy to other components case, a case with unusual, unfamiliar, Processing Tasks that perform similar functions. or challenging features. It is usually Constrained-processing tasks are like Moninger and Stewart (1987) docu- intuitively obvious to an expert when limited-information tasks in that both mented how expert weather forecast- a case is a tough one. Almost by defi- involve tinkering with the familiar ers make predictions about the growth nition, such cases are rare. As a conse- task. Constrained-processing tasks of storms based on analogies to previ- quence, the knowledge engineer must involve deliberate attempts to con- ously experienced storms6 adopt special methods of study. The strain or alter the reasoning strategies One type of constrained-processing limited-information and constrained- that the expert uses. One simple way task involves having the interviewer processing tasks work by making rou- to achieve this goal is to limit the deliberately encourage the use of sce- tine cases challenging; they do not amount of time that the expert has in narios during the performance of a turn routine cases into tough cases. which to absorb information or make familiar task. Such a practice should Methodologically, the method of judgments. For example, the interpre- evoke evidence about the expert’s rea- tough cases is quite simple. The tation of aerial photos typically takes soning for the kinds of scenarios knowledge engineer is not likely to be hours, but in a constrained-processing involved in the data at hand. present when an expert encounters a task the expert might be allowed, say, tough case; so, the expert is equipped two minutes to inspect a photo and with a small tape recorder and is fiv; minutes to make judgments about Combined Constraints instructed how to make a verbal proto- it. A task can involve combining limited- col of personal ruminations when Another way to constrain the pro- information constraints with process- encountering a tough case. cessing is to ask the expert a specific ing constraints. For example, the It seems to me that for all important question rather than to require the full expert aerial-photo interpreter could domains of expertise--those which analysis that is conducted during the be asked to interpret a photograph take years of practice--all experts and familiar task. Two subtypes of a con- without the benefit of maps and with teams of experts should routinely strained-processing task that involve only two minutes in which to view make recordings of their tough cases. this single-question processing are the photo [Hoffman 1984). Expertise should never be wasted. what can be called the method of sim- If an expert does not like limited- Although the analysis of protocols can ulated familiar tasks and the method information tasks, chances are the be time consuming (more on this of scenarios. opinion will be the same about con- point later), benefits will no doubt strained-processing tasks. In general, accrue from recording the knowledge The Method of the more constrained the task--or the that took someone a lifetime to Simulated Familiar Tasks more altered it is relative to the famil- acquire. Ultimately, recording of iar tasks--the more uncomfortable the knowledge might be the crowning Here, a familiar task is performed expert is doing the task. However, in achievement of the modem work on using archival data. The best example my experience, once an expert begins expert systems. As Doyle (1984) put it, of this method I’ve seen is used in to open up and becomes less hesitant “It may be fruitful to separate training probing the knowledge of expert about giving uncertain or qualified in articulate apprenticeship from

SUMMER 1987 57 Table 4. Some Salient Advantages and Disadvantages of the Various Methods for Extracting an Expert’s Knowledge. training in current computer systems, Task and Material Simplicity mation and constrained-processing for the’former will be useful today and tasks can be intensive and should be The method of familiar tasks and the tomorrow, while the latter will con- designed to take somewhere between method of tough cases are essentially tinually become obsolete” (p. 61). 15 and 45 minutes. Although they are equal in terms of the simplicity of the not very time consuming, these tasks tasks and the materials, as are the lim- do require time for transcribing and ited-information and constrained-pro- coding the audiotapes (more on this cessing tasks. For all these methods, point later). Comparison of the Methods the instructions involved are about a For an independent example of how all page or so long, and the materials will these various methods crop up in work be one or a few pages of information, on expert systems, see the section displays, graphs, and so on. The struc- Task Flexibility “Step by Step” in a review article by tured interview stands out because it Ideally, the task should be flexible: it Patton (1985). Looking through her requires a first-pass data base, which should work with different sets of discussion, one can ferret out an can itself consist of a relatively large materials and with variations in example of a structured interview, an and complex set of materials. instructions. For some experts, abun- analysis of tough cases, a simulated dant information about reasoning can familiar task, and a constrained-pro- be evoked by the simplest of ques- cessing task. Given all these various Task Brevity tions. For other experts, the verbaliza- tasks, how can they be compared? One tions might be less discursive. For Ideally, one wants to disclose the way to compare them is in terms of some experts, a tape recorder might be expert’s reasoning as quickly as possi- their relative advantages and disadvan- absolutely necessary; for others, short- ble. Familiar tasks can take anywhere tages, some of which I have already hand notes might suffice. alluded to for the various tasks. Table from a few minutes (X-ray diagnosis) 4 summarizes these points. to an hour [weather forecasting] to an Apart from the qualitative advantages entire day (aerial-photo interpretation). and disadvantages, the experimentalist Interviews take on the order of days or Task Artificiality would like to have some quantifiable even weeks. Although a time-consum- The task should not depart too much criteria for use in a comparative analy- ing process, interviews can yield a from the familiar tasks. After all, one sis. Some possible criteria are present- great deal of information, especially does not want to build a model of rea- ed in table 5. when used in the initial phase of soning based on research evidence developing a data base. Limited-infor- gained from tasks that never occur in

58 AIMAGAZINE the expert’s experience. The further the task departs from the usual problem-solving situation, the less it tells the knowledge engineer about the usual sequences of mental operations and judgments. However, deliberate violations of the constraints that are involved in the familiar tasks (that is, tinkering with the familiar tasks) can be used to systematically expose the expert’s reasoning. None of the tasks that I have described in this paper depart radically from the familiar tasks. They do differ in the degree to which they relax the constraints that are involved in the familiar tasks in order to evoke instances of reasoning.

Data Format Ideally, the data that result from any method should be in a format ready to be input into the data base. Only the structured interview has this charac- teristic. The other methods can result in verbal protocol data, which need to be transcribed and analyzed. ’

Data Validity The data that result from any method should contain valid information, valid in that it discloses truths about the expert’s perceptions and observa- tions, the expert’s knowledge and reasoning, and the methods of testing Table 5. Some Criteria by Which Methods Can Be Analyzed and Compared. hypotheses. In addition to validity in the domain of expertise being studied. from earlier studies of some other the sense of truth value, one would However, just because a given state- expert). This redundancy can be taken like to feel that a data base is comment (or datum) is relevant does not as evidence of the validity of the data, plete--that it covers all the relevant mean that it is at all important. validity in the sense of agreement or subdomains of knowledge. One would Validity in the sense of the relative reliability across different experts. also like to know that different importance of a given fact or rule can How many experts is enough? I am experts agree about the facts. This be assessed in a number of ways. One not sure there can be a principled agreement is validity in the sense of way is the direct approach--ask the answer to this question. Some reliability or consensus across experts. experts some more-or-less structured domains have only one expert, leaving Finally, one would like to have some questions to get their judgments of the one with little choice. At the other assurance that a set of facts includes relative importance of a data-base extreme is the “shotgun” approach of the most important ones. entry. Another way to assess impor- interviewing as many different experts Without such valid information, an tance is to see how early or how often as possible (for example, Mittal and expert system cannot be built. HOW a given fact crops up in the data from Dym 1985) in order to get a complete can one be sure that a data base repre- various experts. data base. sents valid information? Validity: Reliability or Consensus The Validity: The Completeness of the Validity: The Importance of the Data running of additional experts in a Data Base Having more than one Some of an expert’s statements will knowledge-extraction task presumably expert perform various tasks can also obviously be irrelevant (Do you want a has as its major effect the generation of help assure validity--in the sense of the cup of coffee?). Excepting these state- lots of information which is redundant completeness of the data base. Indeed, ments, the knowledge engineer gener- in that it repeats data which are a good rule of thumb is to “pick the ally takes it for granted that a given already in the data base (having come brains” of more than one expert for any statement by the expert is relevant to

SUMMER 1987 59 photo interpreters sometimes work in teams of research groups. For cases such as these team efforts, the experts’ discussions, debates, and exchanges of hypotheses can be very informative, especially when the team encounters a tough case. Whether disagreements among experts are pervasive, are due to uncertainty in the application of relevant knowledge, or are due to uncertainty about specific hypotheses, it is ulti- Table 6. A Matrix for Classifying the Data Generated by a Task That Is Intend- mately the responsibility of the knowledge engineer to resolve any unreliable ed to Extract the Knowledge of an Expert. The frequency of propositions of each of two types (facts or observation statements, and inferences or potential if-then or contradictory rules or facts. Suppose that each of a set of expert aerial-photo rules) can be used to assess the relative difficulty of extracting the two types of interpreters produces a rule about the knowledge. “old” or “new” relative to the current data base. identification of limestone sinkholes domain in which different experts that is wholly idiosyncratic at best and but that the rules are somehow mutu- have widely different degrees of expe- based on trivial knowledge at worst. In ally incompatible. The knowledge rience with different subdomains. In any event, it could be that the domain engineer really has only two choices: addition to their idiosyncratic knowl- at hand is not really well-suited to the either combine the rules into a new edge, they might have idiosyncratic application of expert system tools. The logically coherent rule, or select one of strategies (Mittal and Dym 1985). advice of trying to avoid contradictions them for use. In any event, only one Thus, having more than one expert go seems to the experimental psycholo- rule for identifying limestone sink- through the various knowledge-extrac- gist to miss the point of expert system holes can appear in the expert system tion tasks should have the effect of work. Disagreements should be used [that is, there can be no contradictions “pushing” the data base toward com- as clues about the basic research that in its logic). Formally, the machine can pleteness by filling in the gaps, which might be needed to fill in the knowl- only model one expert at a time. occurred in the aerial-photo-interpre- edge gaps, perhaps even before any tation project. One expert knew a lot expert system work can be done. about desert forms, another expert Less extreme than the assumption Method Efficiency knew a lot about glacial and arctic of pervasive uncertainty and disagree- Perhaps the most important criterion forms. ment is the assumption that experts in in a practical sense is overall efficien- a given domain agree about most of cy. For the sake of people out there “in Validity: The Truth Value of the Data the basic facts and relevant causal the trenches,” I thought I’d try to apply The question of how many experts is laws, but might disagree about how to some efficiency metrics to the various enough cuts two ways. The study of go about applying them. An example is methods. In the case of knowledge more than one expert is bound to gen- the project on the effects of sunspots acquisition, efficiency involves the erate some disagreements, and dis- on the weather (McIntosh 1986). number of propositions generated per agreements are indigestible contradic- Although the relevant laws are known unit of time. Time is expressed as total tions as far as any logic-based machine (that is, electromagnetism, particle task time, that is, the time it takes the system is concerned. Nevertheless, it physics, and so on], the dynamics of knowledge engineer to prepare to run might become inevitable that contra- the sun’s atmosphere are only partially the expert at the task plus the time it dictions arise as one refines a data understood. Thus, the solar weather takes to run the task plus the time it base, adds details, and fleshes out the experts sometimes disagree on how to takes to analyze the data. experience-dependent subdomains of go about predicting solar phenomena. One can assess the production rates knowledge. Another realistic possibility is the of propositions of different types. One It has been proposed that expert sys- experts agreeing on the relevant laws seeks not only facts [predications or tem development projects should only and about how to go about applying observation statements), one also use one expert precisely in order to these laws but disagreeing about seeks information about potential if- avoid contradictions (for example, Pre- hypotheses for some specific cases. then rules (the expert’s inference rau 1985). This strategy assumes that Such occurrences routinely happen in strategies and heuristics). Further- disagreements are pervasive, but it knowledge-extraction tasks that more, one seeks not just propositions also seems to assume an underlying involve more than one expert at a of these types but new ones, informa- reason for the pervasiveness--it is the time. For example, weather forecasters tive ones because they are not already expert knowledge itself that is plagued can be paired in the simulated familiar in the data base. Table 6 presents this by uncertainty. If this statement is task (Moninger and Stewart 1987) and classification of propositional data. true in a given domain, one obviously they often enter into debates as Table 7 presents some efficiency runs the risk of generating a system hypotheses are hashed about. Aerial-

60 AI MAGAZINE analyses for the projects I’ve been involved with. I present these analyses not to imply that knowledge engineers should routinely perform such analyses, but hopefully to save them the trouble. I suspect that the rate statistics presented here are representative of what people can expect in work on expert systems. Let me focus on unstructured interviews because this knowledge-acquisition method is the most commonly used. For unstructured interviews, one can assess efficiency in a number of ways. Total time can be divided into the number of propositions obtained (for both observation statements and potential if-then rules). Also, the number of propositions obtained can be divided into the number of questions asked by the interviewer (or the number of pages of transcript]. For the project on expert airlift planners (Hoffman 1986), the analysis of the 10 hours of taped unstructured interviews yielded a total task time of 100 hours, a total of 957 questions asked by the interviewer, and a total of 753 propositions (including both Table 7. A Comparison of the Results for Four Methods in Terms of Some Mea- observation statements and potential sures of Efficiency The unstructured interview results are from the project on if-then rules). Thus, the unstructured expert airlift planners (Hoffman 1986). The other results are from the aerial interview generated only approximate- photo interpretation project (Hoffman 1984) The third task listed here is one ly 0.13 propositions per task minute that combined limited information with processing constraints. All rate com- and approximately 0.8 propositions for putations are based on total task time. each question asked by the knowledge engineer. The unstructured interview pro- somewhere. If the rate is closer to tasks should be conducted and a first- duced lots of propositions in the first three per minute, the knowledge engi- pass data base gleaned from this analy- hours of the interview (over 5 proposi- neer is golden. sis. Once a first-pass data base has tions per page of transcript or over 200 been produced, then unstructured propositions per hour), but the rate interviews (using note taking rather trailed off rapidly (to less than 1 propo- than tape recording) can be used to sition per page of transcript or about Recommendations determine user needs. If the knowl- 40 propositions per hour). In part, this edge engineer wishes to build a refined result was due to the fact that in the Overall, the unstructured interview is or second-pass data base, structured later sessions, the interviewer and not too terribly efficient. This finding interviews are more efficient. In fact, expert spent time sitting at a comput- deserves emphasis because most they can be many times more efficient er terminal and informally discussed expert systems have apparently been overall, according to my results. user needs and system design. In just developed using unstructured inter- As data bases become very large, it the first 5 of the 10 hours, the unstruc- views Preliminary unstructured inter- becomes necessary to use special tured interview yielded approximately views can be critical for obtaining tasks. Limited-information or con- 1.6 propositions per task minute. information on three key questions: (1) strained-processing tasks or the In general, if a knowledge engineer Who are the experts? (2) Is this prob- method of tough cases can be used in finds that knowledge is being extract- lem well-suited to the expert system order to focus on specific subdomains ed at a rate of about two new proposi- approach? and (3) What are the needs of knowledge or to disclose aspects of tions per task minute, the engineer of the people who will use the system? reasoning that the data base might can be assured of probably being on However, if the knowledge engineer lack. Otherwise, the knowledge engi- the right track. If the rate is closer to needs to become familiar with the neer might spend hours working at an one proposition per task minute, domain, then an analysis of familiar inefficient task, inefficient in that it chances are there is some inefficiency

SUMMER 1987 61 produces lots of propositions which MYCIN (Davis, Buchanan, and Short- happen to already reside in the data liffe I977), a system for diagnosing Some Lingering Issues base. infectious diseases, took many years. All experts differ, and all domains of The development of INTERNIST, None of the methodological ideas that expertise differ; so, too, will all expert another medical-diagnosis system, are presented here is intended to be system-development projects differ. took 10 years with the help of a full- logically airtight. My categorization of Some methods for extracting an time specialist in internal medicine methods is intended to be pragmatic expert’s knowledge will fit some pro- (Buchanan 1982). Rl, which configures rather than complete. The criteria for jects and not others. For example, an the VAX computer (McDermott 1980), analyzing methods are probably insuf- analysis of familiar tasks might not be took two man-years to develop by a ficient. I welcome reactions and sug- very fruitful for domains in which team of about a dozen researchers and gestions and hope that these ideas are there are no texts or in which the is still being refined. In general, it helpful to others. available technical manuals provide takes one to two years to develop a To those who have already devel- little information about what the prototype and about five years to oped an expert system, many of the experts actually do. Such was the case develop a full-scale system (Duda and research ideas I have referred to might for the airlift-planning project (Hoff- Shortliffe 1983; Gevarter 1984). seem detailed (if not trivial). I wanted man 1986). It made sense in that pro- In contrast, the PUFF system for the to spill out the gory details for the ject to begin with an unstructured diagnosis of pulmonary disorders was sake of those who have not yet interview to learn not only about user reported to have been developed in attempted to generate an expert sys- needs but to simultaneously get some less than 10 weeks (Bramer 1982). The tem. As I said at the outset, I address very basic information about what air- likely reason for this brevity was that the expert system engineering process lift planners actually do in their plan- most of the rules were easily gleaned from a nuts-and-bolts perspective. To ning tasks (that is, the familiar task). from archived data (Feigenbaum 1977), those who are not familiar with exper- Despite such domain specificity, I and only one week was spent inter- imental psychology, it might be sur- can propose a generalized series of viewing the experts (Bramer 1982). prising that experimentalists deal with steps for extracting the knowledge of Apparently, if one is relying on minutiae such as better ways of tape experts, as shown in table 8. Situa- unstructured interviews, is interview- recording or the variety of ways to tions probably exist for which my rec- ing a large number of experts, or is assess the validity of a data set. How- ommended series of steps is not quite working on a system that requires ever, such minutiae are precisely the fitting. However, I’d wager that one or hundreds or thousands of rules and sorts of things that experimental psy- another slight variation on this theme facts in its data base (as in the chologists like to attend to. Thus, my would fit. INTERNIST system), then many years discussion focused on methods and The figures in table 8 for “Effort of effort are required to develop the methodology (the analysis of meth- Required” are approximate ranges system to the prototype stage. ods). Hidden in the discussion are a based on my experience. According to In this light, a minimum of three few broader issues that merit explicit my estimates, it can take at least three months to build a second-pass data consideration and a wee bit of feather months to get to the point of having a base seems like an underestimate. ruffling. refined or second-pass data base. How However, it cannot be a gross does this time frame compare with the underestimate. The second-nass data experiences of other artificial intelli- base for the aerial-photo-interpretation The Social gence [AI) researchers? Actually, it is project (Hoffman 1984) contained over PsychologyofExpertise impossible to tell from reading the 1400 rules and facts, and it took me expert system literature exactly how about 45 days to develop it by means Should the knowledge engineer be the much effort went into building the of structured interviews and special expert? In other words, should the data base in any expert system project. tasks. expert become a knowledge engineer? (It is for this reason that I feel com- AI researchers will not be eager to I do not know if there can be a princi- pelled to present my own “ballpark” use all the methods and carry out all pled answer to this question. Some estimates.) Typically, authors do not the efficiency computations presented researchers claim that the knowledge even state how the knowledge was here, but I do not propose that they engineer should be the expert (Fried- obtained--sometimes from texts and should. I do not want to shoehorn the land 1981), that “expert systems are manuals, sometimes from archived “tricks of the trade” of building expert best conceived by the experts them- data, and usually from unstructured systems into the rigid statistics-orient- selves” (Taylor 1985, p. 62). Some interviews. Typically, research reports ed structure of a psychological experi- researchers claim that the situation in jump right into a discussion of system ment. However, I do want to strongly which experts interview other experts architecture. recommend that developers of expert can be disastrous. Other researchers Some reports on expert system pro- systems routinely report the methods say the more the expert knows about jects state how long it took to develop used to extract experts’ knowledge and AI, the more likely it is that the expert a prototype, but one can only guess the efficiency of the methods [that is, has biases (for example, about which how long it took just to acquire (or the amount of effort taken to con- knowledge representation format to extract) the data. The development of struct the data base). use) (McIntosh 1986).

62 AI MAGAZINE Table 8 Some Steps for Extracting and Characterizing the Knowledge of an Expert Prior to the Construction of an Expert System The analysis of familiar tasks includes the analysis of texts and technical manuals. The activity des- ignated “Special tasks” includes limited information tasks, constrained processing tasks and analyses of “tough cases ” All figures for “effort required” are approximations based on the author’s experience.

The answer to this question proba- elusion or inference._ . Typically, the what degree does a given expert sys- bly lies in the social-psychological system returns with what is essential- tern need to represent conceptual dynamics of particular working ly a list or a printout of the rules knowledge? The jury is still out. We groups. Certainly, the expert should involved. Such an explanation is not cannot be certain the purely rule-based know enough about AI to be skeptical very adequate. Until expert systems approach will always effectively or and enough about expert systems to deal with reasoning and knowledge at efficiently solve the various problems appreciate the ways in which the sys- a semantic or conceptual level, their that people would like their expert tern can aid, rather than replace them. artificiality will remain painfully obvi- systems to solve. ous, and they might not even merit Most concepts do not exist in Are Expert Systems AI, or Are being called intelligent. human cognition in any pristine form. Concepts are typically defined in AI Experts, unlike expert systems, do not They Cognitive Simulation? using logically necessary and sufficient always reason in terms of a logical set Some expert systems have an “expla- conditions. However, can an expert of rules (Dreyfus and Dreyfus 1986). nation” component; that is, one can system work (or work well) without What is clear is that in many domains, query the system about a given con- also being a “cognitive simulation?” To experts reason in terms of their percep-

SUMMER 1987 63 tual and conceptual understanding, er program! Davis and Lenat (1982, p. tationally elegant. Such hasty episte- such as their “mental models” of 348) assert that a system which has mological pronouncements will never causal laws (Gentner and Stevens knotiledge about its own representa- do.9 1983; Sridharan 1985). tions can allow for “knowledge acqui- If AI is to solve the really tough When the expert aerial-photo inter- sition using a high-level dialog.” But problems, it would do well to put a bit preter looks at a photo, the complex can’t humans use dialog to extract less effort into short-term solutions geo-biological events that led to the knowledge, or did I miss something? [for example, reliance on naked if-then formation of the terrain are perceived AI really does like to put the logical rules) ma premature applications (for (for example, mountain building, cart before the empirical horse. example, expert systems that rapidly glaciation, and so on). When the expert In the literature on expert systems, become too complex to be rapidly cardiologist listens to heart sounds I’ve found numerous assertions, such computable) and a bit more effort into over a stethoscope, the complex as the claim that experts are better systematic research (for example, how biomechanical events that occur in able to build good data bases than experts really reason]. In the past few the cardiovascular system are per- computer scientists. Such assertions years, federal funding for basic ceived (Jenkins 1985; Johnson et al. might seem solid and might indeed be research on cognition and graduate 1982). Such perceptual and conceptual correct, but they are rarely empirically training in experimental psychology understanding, the result of perceptu- grounded, which causes the experi- have suffered from much more than al-learning ma concept-formation pro- mental psychologist to yank out their fair share of cuts. Rather than cesses, might be direct (Gibson 1979); clumps of hair. just bemoaning the lack of trained that is, it might not be mediated by The broader literature on AI and knowledge engineers, anyone who is any analysis of naked if-then rules, cognitive science is characterized by interested in making a real contribu- serial or otherwise. its penchant for philosophical specula- tion to our needs in the field of AI Recently, attempts have been made tion (Hoffman and Ned 1983). The might consider investing some to incorporate conceptual knowledge typical issue of a cognitive science resources in the support of basic into AI systems (Chandrasekaran and technical journal contains evermore research on human cognition and per- Mittal 1983; Kunz 1983; Pople 1982) clever theories, theories that combine ception. and to define rule systems that behave “schemas,” “scripts,” “inference nets,” like the human reasoner (Sridharan and “labeled relations” in all sorts of 1985). As instances of AI systems, clever computational ways. However, The Benefits of Collaboration expert systems might not need to pre- there is rarely a clear statement about The collaboration of AI researchers cisely simulate cognition in order to whether the theory has an evidential and experimental psychologists can be get the job clone. However, we do need base. Sometimes, no statement even of benefit to work on expert systems. I to learn more about experts’ reasoning exists about whether the theory is sup- hope my article has made this point. through basic research if our expert posed to have anything at all to do Benefits can also accrue for psycholo- systems are ever to get the job clone with heads or whether it is just sup- gy, however. For example, a major the- well, which brings me to the last issue posed to be about computers, which ory of memory asserts that forgetting I address. causes the experimental psychologist is caused by the interference of new to yank out even larger clumps of hair. learning with old (Reder 19873 Smith, While remembering that the roots Adams, and Schon 1978). This hypoth- Does AI Have Its Foundations of AI are in the philosophy of the mind esis generates what is called the para- in Philosophy or Research? (cf. McCarthy and Hayes 19691, AI dox of expertise: how can experts deal practitioners seem to have forgotten The cognitive-simulation work hints so adroitly with so many remembered about the roots in experimental psy- facts? Hopefully, research on the cog- at the fact that some AI workers are chology. It is the rare paper that really closet psychologists. They rely nition of experts can help clarify this acknowledges the need for basic paradox. on psychological terms and mentalis- research on cognition, that is not put tic metaphors, and they work on psy- Expert system work also bears on off by the experimental psychologist’s theories of perceptual learning. chological problems. More to the stubborn concern with methodological point, the problems they work on Although it is obvious experts have details, or that is sensitive to the dif- special knowledge and concepts, it often beg for an empirical answer, an ference between AI and cognitive sim- answer based on research findings. should be kept in mid that they also ulation.8 have special perceptual skills However, AI workers tend to turn to If one is at all interested in cogni- computing theories and abstract con- (Shanteau 1984). For example, the tion or cognitive simulation, then expert radiologist “sees” X rays differ- cepts to produce solutions. experiments are necessary to ensure No better example of this tendency ently than the first-year medical stu- that the model works like the mind dent (Feltovich 1981). Another, more exists than that provided by the field does, or to test hypotheses about an of expert systems itself. Having recog- commonplace example is the expert expert’s strategies. One cannot simply sports commentator who “sees” some- nized that the knowledge-acquisition assert, for instance, that human or process is a bottleneck, what solution thing that we novices can only pick up computer memories must rely on when shown the slow-motion replay. do AI workers seek? Let’s build anoth- schemas because schemas are compu-

64 AI MAGAZINE Lifetimes of good research could be rized to reproduce and distribute reprints ing and Expert Systems. IEEE Expert l(2): carried out to distinguish experts from for governmental purposes not withstand- 17-28 novices in terms of their perceptual ing any copyright notation herein. Ericcson, K. A, and Simon, H A. 1984 Pro- processes, learning processes, and rea- tocol Analysis: Verbal Reports as Data. soning strategies. What information do References Cambridge, Mass : MIT Press experts focus on? What reasoning bias- Bierre, P 1985 The Professor’s Challenge Feigenbaum, E A 1977 The Art of AI. In es do they have? How can experts be AI Magazine 514):60-70 Proceedings of the Fifth International Joint identified? Boose, J. H. 1984 Personal Construct Theo- Conference on Artificial Intelligence Inter- What I find particularly appealing ry and the Transfer of Human Expertise. In national Joint Conference on Artificial about such questions is that the Proceedings of the Third National Confer- Intelligence research they engender is practical and ence on Artificial Intelligence, 27-33. Feltovich, P. J. 1981 Knowledge-Based basic at the same time--practical Menlo Park, Calif : American Association Components of Expertise in Medical Diag- because it contributes to the solving of of Artificial Intelligence. nosis, Technical Report, PDS-2, Learning important problems, basic because it Bramer, M A 1982 A Survey and Critical Research and Development Center, Univ. of contributes to the body of knowledge. Review of Expert Systems Research. In Pittsburgh. The collaboration of experimental psy- Introductory Readings in Expert Systems, Freiling, M.; Alexander, J.; Messick, S.; chologists and AI researchers can be of ed D. Michie, 3-29. London: Gordon and Rehfuss, S.; and Shulman, S 1985. Starting great mutual benefit. Breach a Knowledge Engineering Project: A Step- Buchanan, B. G 1982 New Research on by-Step Approach AI Magazine 613): 150- Expert Systems In Machine Intelligence, 163. Acknowledgments vol 10, eds J. Hayes, D Michie, and Y-H Friedland, P. 1981. Acquisition of Procedu- Most of the ideas reported here were first Pao New York: Wiley ral Knowledge from Domain Experts In developed under the aegis of the U S. Army Chandrasekaran, B., and Mittal, S 1983 Proceedings of the Seventh International Summer Faculty Research and Engineering Deep versus Compiled Knowledge Joint Conference on Artificial Intelligence, Program (Delivery Order No 0550). The Approaches to Diagnostic Problem Solving. 856-861 International Joint Conferences on work was conducted in the summer of 1983 International lournal of Man-Machine Artificial Intelligence at the Engineer Topographic Laboratories Studies 19:425-436 Gentner, D., and Stevens, A, eds 1983 (ETL), U.S. Army Corps of Engineers, Ft Mental Models Hillsdale, N.J : Lawrence Belvoir, Virginia The author would like to Cooke, N. M 1985 Modeling Human thank Olin Mintzer, master aerial-photo Expertise in Expert Systems, Technical Erlbaum interpreter, for his help and guidance The Report, MCCS-85-12, Computing Research Gevarter, W B. 1984 Artificial Intelligence author would also like to thank all the peo- Laboratory, New Mexico State Univ. Computer Vision and Natural Language ple at the Center for Physical Science and Davis, J 1986 Summary of the Workshop Processing Park Ridge, N. J : Noyes the Center for Artificial Intelligence at ETL on Knowledge Acquisition Paper presented Gibson, J G. 1979. T he Ecological for their many stimulating discussions at Artificial Intelligence Research in Envi- Approach to Visual Perception Boston: Clarifications and amplifications arose in ronmental Science Conference. Boulder, Houghton-Mifflin discussions at the Artificial Intelligence Co10 : NOAA Gorfein, D., and Hoffman, R, eds. 1987 Research in Environmental Science Confer- Davis, R ; Buchanan, B G ; and Shortliffe, Memory and Learning. The Ebbinghaus ence held at the Environmental Research E H 1977 Production Systems as a Repre- Centennial Conference. Hillsdale, N J.: Laboratories of the National Oceanic and sentation for a Knowledge-Based Consulta- Lawrence Erlbaum. Forthcoming. Atmospheric Administration (NOAA), tion Program. Artificial Intelligence 8:15- Boulder, Colorado, 28-29 May 1986 The Hartley, R 1981 How Expert Should Expert 45. author would especially like to thank Systems Be? In Proceedings of the Seventh William Moninger of NOAA for his encour- Davis, R and Lenat, B. 1982 Knowledge- International Joint Conference on Artificial agement Based Systems in AI. New York: McGraw- Intelligence International Joint Confer- Many others deserve thanks for their Hill ences on Artificial Intelligence encouragement and support of the research Denning, P J 1986. Towards a Science of Hayes-Roth, F ; Waterman, P. A ; and reported here: Gary Grann, Lieutenant Expert Systems IEEE Expert l(2): 80-83. Lenat, D., eds. 1983 Building Expert Sys- Colonel John Whitcomb, Captain Eric Doyle, J 1984 Expert Systems without tems. Reading, Mass : Addison-Wesley. Goepper, Captain Thomas Moran, Captain Computers or Theory and Trust in Artifi- Hoffman, R. 1986 Procedures for Efficient- Jeffrey Valiton, 2d Lieutenant Joseph Bessel- cial Intelligence AI Magazine 5(2): 59-63. ly Extracting the Knowledge of Experts, man, and Lieutenant Ray Harris of the Technical Report, U.S. Air Force Summer Strategic Planning Directorate, Electronics Dreyfus, H , and Dreyfus, S. 1986 Why Expert Systems Do Not Exhibit Expertise. Faculty Research Program, Office of Scien- System Division, Hanscom Air Force Base, tific Research, Bolling Air Force Base. Massachusetts; Murray Daniels of MITRE IEEE Expert l(2): 86-90. Corporation; Rodney Darrah, Susan Espy Duda, R. 0, and Gashnig, J. G. 1981 Hoffman, R. R. 1985. Some Implications of and the people at Universal Energy Sys- Knowledge-Based Expert Systems Come of Metaphor for the Philosophy and Psycholo- tems, Inc ; Karen Thompson and Pat Carey Age. Byte September:238-281. gy of Science. In The Ubiquity of Metaphor, eds W. Paprotte and R Dirven, at Adelphi University. Duda, R O., and Shortliffe, E. H 1983. Preparation of this paper was supported 327-380. Amsterdam, Netherlands: John Expert Systems Research. Science 220:261- Benjamin by a grant from Summer Faculty Research 276. Program (Contract No F49620-85-C-0013) Hoffman, R. R. 1984. Methodological Pre- Einhorn, H J 1974. Expert Judgment /our- of the U.S. Air Force Office of Scientific liminaries to the Development of an Expert nal of Applied Psychology 59:562-571. Research. The U.S. Government is autho- System for Aerial-Photo Interpretation, Eliot, L. B. 1986. Analogical Problem Solv- Technical Report, ETL-0342, The Engineer

SUMMER 1987 65 Topographic Laboratories Mintzer, 0, and Messmore, J 1984 Terrain IEEE Conference on Systems, Man, and Hoffman, R R 1980 Metaphor in Science Analysis Procedural Guide for Surface Con- Cybernetics, ed M El Hawaray. New York: In Cognition and Figurative Language, eds figuration, Technical Report, ETL-0352, Institute of Electrical and Electronics Engi- R Honeck and R Hoffman, 393-423 Hills- Engineer Topographic Laboratories neers, Inc. dale, NJ: Lawrence Erlbaum Associates Mittal, S , and Dym, C L 1985. Knowledge Shortliffe, E H 1980 Consultation Sys- Hoffman, R R, and Kemper, S 1987 What Acquisition from Multiple Experts Al Mag- terns for Physicians In Proceedings of the Could Reaction-Time Studies Be Telling Us azine 612) 32-36 Canadian Society for Computational Stud- about Metaphor Comprehension? In Moninger, W, and Stewart, T 1987 A Pro- ies of Intelligence. Victoria, B C : Canadian Metaphor and Symbolic Activity vol 2 posed Study of Human Information Pro- Society for Computational Studies of Intel- Forthcoming cessing in Weather Forecasting. In Bulletin ligence Hoffman, R R., and Nead, J M 1983 Gen- of the American Meteorological Associa- Smith, E E.; Adams, N ; and Schon, D. era1 Contextualism, Ecological Science, and tion Forthcoming 1978 Fact Retrieval and the Paradox of Cognitive Research The fournal of Mind Nii, H P , and Aiello, N 1979 AGE Interference Cognitive Psychology 10:438- and Behavior 4 507-560 (Attempt to Generalize]: A Knowledge- 464 Jenkins, J J 1985 Acoustic Information for Based Program for Building Knowledge- Sridharan, N S. 1985 Evolving Systems of Objects, Places, and Events In Persistence Based Programs. In Proceedings of the Sixth Knowledge AI Magazine 6(3): 108-120 International Joint Conference on Artificial and Change Proceedings of the First Inter- Stefik, M; Aikins, A.; Balzer, R.; Benoit, J; Intelligence, 645-655 International Joint national Conference on Event Perception, Birnbaum, L ; Hayes-Roth, F.; and Conferences on Artificial Intelligence eds W Warren and R Shaw, 115-138 Hills- Sacerdoti, E 1982 The Organization of dale, N.J Lawrence Erlbaum Pachella, R G 1974 The Interpretation of Expert Systems: A Tutorial Artificial Intel- Johnson, P E.; Hassebrock, F ; Duran, A S.; Reaction Time in Information-Processing ljgence I8:I35.I73, and Moller, J H. 1982 Multimethod Study Research In Human Information Process- ing: Tutorials in Performance and Cogni- Stemberg, R J 1977 Intelligence, Informa- of Clinical Judgment. Organizational tion Processing, and Analogicai Behavior and Human Performance 30 201- tion, ed B Kantowitz Hillsdale, N.J : Lawrence Erlbaum. Reasoning Hillsdale, N J : Lawrence Erl- 230 baum Patton, C 1985 Knowledge Engineering: Kelly, G 1955 The Psychology of Personal Taylor, E C 1985 Developing a Knowledge Constructs New York: Norton Tapping the Experts Electronic Design, May:93- 100 Engineering Capability in the TRW Defense Kintsch, W 1972 The Representation of Systems Group AI Magazine 6(2): 58-63 Politakis, P, and Weiss, S. 1984 Using Meaning in Memory Hillsdale, N J : Townsend, J T, and Ashby, F G 1983 Lawrence Erlbaum Empirical Analysis to Refine Expert Sys- terns Knowledge Bases Artificial Intelli- Stochastic Modeling of Elementary Psycho Klein, G. A 1987 Applications of Analogic gence 22:23-48 logical Processes. Cambridge, Mass Cam Reasoning Metaphor and Symbolic Activi- bridge University Press ty 2(3): 201-218 Pople, H E 1982. Heuristic Methods for Imposing Structure on Ill-Structured Prob- Way, D. S 1978 Terrain Analysis Strouds Kunz, J 1983 Analysis of Physiological lems In Artificial Intelligence in berg, Pa.: Dowden, Hutchinson, and Ross Behavior Using a Causal Model Based on Medicine, ed P Szolovits Boulder, Co10 : Weiss, S., and Kulikowski, C. 1984 A Prac First Principles In Proceedings of the Sec- Westview. tical Guide to Designing Expert Systems ond National Conference on Artificial Totowa, N J.: Rowman and Allanheld. Intelligence. Menlo Park, Calif.: American Posner, M I 1978 Chronometric Explo- Association for Artificial Intelligence rations of Mind Hillsdale, N.J Lawrence Woods, D. D. 1985. Cognitive Technolo Erlbaum. gies: The Design of Joint Human-Machinc McCarthy, J 1983. President’s Quarterly Cognitive Systems. AI Magazine 6(4): 86 Message: AI Needs More Emphasis on Basic Prerau, D S. 1985. Selection of an Appro- Research AI Magazine 4[4): 5 priate Domain for an Expert System. Al 92 Magazine 6(2): 26-30 McCarthy, J , and Hayes, P 1969 Some Philosophical Problems from the Stand- Purves, W K 1985 A Biologist Looks at Notes Cognitive AI AI Magazine 6(2) 38-43. point of AI In Machine Intelligence, vol 4, 1 For discussions of the psychology o eds B Meltzer and D Michie. Edinburgh, Quinlan, J. R. 1984 Fundamentals of the expertise, see Cooke (1985), Einhom [ 19741 Scotland: Edinburgh University Press Knowledge Engineering Problem In Intro- Feltovich (1981), and Shanteau (1984) ductory Readings in Expert Systems, ed. D McDermott, J 1980 Rl: A Rule-Based Con- 2 One exception is the work of BOOSI Michie, 33-46. London: Gordon and Breach figurer of Computer Systems, Technical (19841, who has adopted ideas about inter Report, Computer Science Dept., Camegie- Raulefs, P 1985 Knowledge Processing viewing from psychologist George Kell Mellon Univ Expert Systems In AI’ Towards Practical (1955) ~~~~~~~ cOncern has been wit] McIntosh, P. S 1986 Knowledge Acquisi- Applications, eds T. Bernold and G Albers building a computer system for interview tion for THEO: An Expert System for Solar- Amsterdam, Netherlands: North-Holland ing experts Flare Forecasting Paper presented at Artifi- Reder, L M 1987 Beyond Associationism: 3. A proposition is generally regarded as al cial Intelligence Research in Environmental Strategic Components in Memory “atomic fact”--an assertion of the existent Science Conference Boulder, Co10 : NOAA Retrieval In Memory and Learning: The of some entity and the predication of som McNeill, D , and Levy, E 1982 Conceptual Ebbinghaus Centennial Conference, eds D property or relation of this entity Treat Representations in Language Activity and Gorfein and R Hoffman. Hillsdale, N J ments that are relevant to protocol analysi Gesture In Speech, Place, and Action, eds Lawrence Erlbaum Forthcoming can be found in Kintsch (1972, chapter 2 .R Jarvella and W. Klein, 271-295 New Shanteau, J 1984. Some Unasked Ques- and Ericcson and Simon (1984, chapters York: Wiley tions about the Psychology of Expert Deci- and 6). sion Makers In Proceedings of the 1984

66 AI MAGAZINE 4. Tape-recording tips: [ 1J conduct the inter- solving, see Eliot (1986), Gentner and transcript. The coding should involve at view in a small, quiet room; (2) use batter- Stevens [1983), Hoffman (1980, 1985), and least two independent coders (or judges) ies liberally so you always receive a clear Sternberg (1977) until evidence indicates that their respec- signal; and (3) preface your recording by 7. Some detailed comments are in order tive codings show agreement which is high identifying each of the participants. about the transcription process for the sake enough to warrant dropping the extra 5. Constrained-processing tasks can also of those who might choose to go this route. coders. involve the measurement of reaction time As I have already implied, the transcription 8 To be fair, some AI authors are sensitive or decision latency Such measures are process takes time Without any doubt, the to the difference, for example, Bierre (19851, important for the purposes of cognitive most time-consuming aspect of the tran- McCarthy (19831, Purves (1985,) and Woods simulation and cognitive psychology, scription process is the time spent pausing (1985). Furthermore, the major AI conven- because the results can be informative and backing up the tape to interpret and tions include sessions on cognitive about specific sequences of mental opera- write down the segments where more than research, and the AI technical journals tions (such as memory encoding and one person speaks at a time. The moral: include occasional reviews of recent cogni- retrieval] For a discussion of the logic of The examiner should consciously try to tive research reaction-time experiments, see Hoffman withhold comments or questions while the 9 I am hesitant to attribute this particular and Kemper (1987), Pachella (1974), Posner expert is talking. The examiner should use claim about schemas to any one person I’ve [ 1978), or Townsend and Ashby [ 1983). a notepad to jot down questions and return read so many claims about the representa- 6. What is salient here is that the analogic to them when the expert’s monologue is tion of knowledge [templates, prototypes, reasoning is in terms of a comparison to over The goal of any interview, whether networks, semantic features, and so on), previously encountered scenarios Because structured or unstructured, is not to edify and I’ve yanked out so much hair, that I’d reasoning by analogy pervades all problem the knowledge engineer, but to let the rather ruffle everyone’s feathers I also find solving, the analogy component of the expert talk and get the knowledge out so it worrisome when claims are made about method of scenarios does not seem as that it can be formalized (hence my earlier computer or mental representations on the salient for present purposes as the reliance distinction between interviews and discus- basis of the argument that there are “stor- on scenarios to form the analogies For sions] It also takes time to code the tran- age” limitations. Indeed, “storage” might be reviews of the abundant literature on the script for propositional content, anywhere the wrong metaphor altogether. For further role of analogy and metaphor in problem from one to five minutes for each page of discussion, see Gorfein and Hoffman (1987)

m Halbrecht Associates,Inc. Executive Search Consultants

SERVING THE ARTIFICIAL INTELLIGENCE COMMUNITY Halbrecht Associates, founded in 1957, was the first search firm in the United States to specialize in Znformation Systems Management, Operations Research, Management Sciences, Telecom- munications, Computer Sciences and Decision Support Systems. We are now also in the forefront of the Artificial Intelligence field. Because we have been so extensively involved in this emerging technology, our portfolio of people includes extensive numbers of AZ professionals at all levels, from the United States and abroad. Our client list includes companies coast to coast, large and small, including start-ups with equity opportunities. Accordingly, depending on your needs, we are in a unique position to assist you rapidly with any of your requirements. Our current assignments range downward from several Chief Scientists, or those who can provide the technical leadership to head up major AZ programs, to Senior LISP Programmer. Zf you would like to be considered to be part of our AZ data base, please send us your credentials. This will enable us to give you first notice of appropriate opportunities. All inquiries are in strictest confidence. Contact: Daryl Furno, Senior Associaie

1200 Summer Street l Stamford, CT 06905 l 203-327-5630

For free information, circle no. 193

SUMMER 1987 67