Applications of Artificial Intelligence

Reports of panel

on

APPLICATIONS OF ARTIFICIAL INTELLIGENCE

Saul Amarel (Session Chairman), Rutgers University John Seely Brown, BBN Bruce Buchanan, Stanford University Peter Hart, SRI Casimir Kulikowski, Rutgers University William Martin, MIT Harry Pople, University of Pittsburgh

INTRODUCTORY. REMARKS. SUMMARY OF ISSUES AND OPEN

PROBLEMS IN AI APPLICATIONS . (S. Amarel)

In recent years there has been a growing amount of applications-oriented AI work. The distinction between an applied AI project and a 'basic' project is mainly determined by the prin­ cipal objectives of the effort. If the objective is to develop a high performance, expert, system for solving certain problems in a given domain, in a manner which is acceptable (and hopefully use­ ful) to users in the domain, then the project is in the applied category. What characterizes, in addition, work in several recent applied AI pro­ jects is the focus on natural domains, where the tasks involve reasoning with empirical data in the context of bodies of empirical knowledge. Prob­ lems in medicine, biochemistry and signal process­ ing are of this type. It is certainly possible to apply AI methods and tools to the development of expert systems in the context of formal task environments such as mathematics. Indeed, one of the most successful early applications of AI has been in this area - specifically, the MACSYMA project at MIT (Moses, 1971). It is a fact how­ ever that most of the efforts in AI applications over the last 5-10 years have concentrated on systems whose knowledge bases contain relatively large amounts of empirical information about pieces of the real world.

An important reason for this is the high potential usefulness (social value) of AI systems that could provide help to scientists and profes­ sionals in problems that involve the understanding and control of natural phenomena - especially in situations where available knowledge is complex, ill structured, and changing. Another essential reason is the expectation that work in real world problems will bring new challenges to AI and it will contribute significantly to the scientific development of the field.

At present, there are several applied AI pro­ jects in the country that have produced systems at an impressive level of expertise in limited real world task environments.

In analytical chemistry, the heuristic DENDRAL project at Stanford has been the pioneer effort in the area of computer- based expert assistants for problems in a scientific domain. The DENDRAL program can interpret mass spectra of organic molecules at a level of performance which equals, and in some cases surpasses, experts in the field (Buchanan et al, 1969; Smith and Carhart, 1976).

In medicine, the CASNET;tl LAUCOMA system at Rutgers, which provides consultation in diagnosis and therapy of glaucomas (Kulikowski and Weiss, 1971; Weisss 1974), and the MYCIN system at Stanford which assists in the treatment of infec­ tious diseases (Shortliffe et al, 1975) have now reached expert status. The INTERNIST system at the University of Pittsburgh, which provides clinical consultation in problems of internal medicine, (Pople et al, 1975; Pople 1975, 1977) is nearing expert status for diagnostic tasks in parts of its domain. Several other exper~ systems are now being developed in medicine, biochemistry, genetics, psychology, instruction, mineral explor­ ation, business management, language and speech processing, and computer programming.

Much experience has been accumulating in the course of developing these expert systems. Workers in the field found that many of the methods and techniques that grew from previous basic work in AI could be successfully adapted to the build­ ing of knowledge-based expert systems. Also, in their attempts to raise the levels of expertise and performance of systems in specific task environments, they identified several important AI problems on which more basic work is needed.

In general, we are still at the stage of learning how to build high-performance knowledge­ based systems. With the possible exception of DENDRAL, the expert systems that have been built to date for specific natural domains have not had as yet a significant impact on the users in these domains. The main value of these developments has been to demonstrate that expert systems of certain types are feasible within the present state of AI knowledge. Also, the exploratory work on realistic expert systems has forced us to look more closely into many aspects of knowledge handling that are of fundamental importance with­ in AI. Other basic topics whose study has been stimulated by these system development efforts include methods of plausible reasoning, strategies of planning under uncertainty and approaches to problems of interpretation and theory formation.

From the viewpoint of AI as a scientific discipline, work on AI applications is providing a valuable testing ground for existing AI schemes and tools, it is helping to identify their scope and limitations, and it is stimulating the development of new methodologies for the design and construction of AI systems. Furthermore, it is forcing us to direct more attention to scientific communications - both within the field and with experts in other fields. I expect that this will bring about a better understanding of the present state of knowledge in AI, and a better perspective of how this knowledge is related to other parts of computer science and to knowledge in other disciplines.

From the viewpoint of a discipline such as med­ icine i n whose domain AI systems are being developed, the effort to structure knowledge in the domain - which is an essential prerequisite for achieving an expert system - may be of a much more fundamental significance than was initially realized. The clar­ ification of basic concepts in the discipline, the identification of relationships between various pieces of knowledge, the assessment of the nature and status of underlying empirical knowledge, of theories and of beliefs, the recognition of gaps, and - most important - the providing of a conven­ ient computer-based environment which stimulates and supports these activities, are of great scientific value. In the long run, this may have a much strong­ er practical imoact on the disciPline than the im­ pact produced by the addition of a resource or a technological tool to assist users in a specific task area within the discipline.

Emphasis of panel. Relationship to AIM Workshop

The main emphasis of our panel will be on applications of AI to the development of knowledge­ based systems in medicine. In addition, appli­ cations in other domains (mineral exploration, education) will be considered, and certain method­ ological/technical problems that are shared by all these projects will be examined.

In particular, the panel is intended to pro­ vide a forum for an assessment of work being done in the AIM (Artificial Intelligence in Medicine) community. The AIM project is a national resource sharing activity which is supported by the Bio­ technology Resources Program (BRP) of DRR, NIH, and whose principal objective is to promote appli­ cations of AI to biological and medical problems. The main focus of AIM activity is at the Stanford SUMEX-AIM project, which provides computer shared resources to the AIM community via national net­ works. The Rutgers Research Resource on Computers in Biomedicine (Amarel, 1975) is one of the (BRP­ supported) projects in the AIM community; included among its functions is the organization of annual AIM Workshops. The objective of the Workshops is to strengthen scientific interactions within the national AIM community, and to disseminate AI-based methodologies, tools, and specific systems that are relevant to AIM. The 3rd AIM Workshop will be held at Rutgers on July 6 to 8, 1977. All the panel participants (except for Peter Hart who will be represented by Dick Duda) are expected to attend the Rutgers· Workshop. Points of view and con­ clusions developed at the AIM Workshop will be presented at the panel as part of the discussion of issues that are outlined in the present report.

The panel will provide an opportunity to summarize for a wide AI audience the highlights of certain key issues in the development of applied AI systems that will be discussed at the AIM Workshop.

Comments based on recent experience with AI applications

Experience with work in AI applications to date shows the following:

(a) The problem of acquiring specific knowledge in a domain, managing it in a AI system, modifying it, and using it appropriately is fundamental. The approach to most designs is incremental and re­ sponsive to the fact that the knowledge base in a domain is not stationary. Initially, a relatively low-performance AI system is created to provide the basis for subsequent stages of knowledge

acquisition and improvement, which eventually leads to a high-performance, expert, system. (b) Work on applications requires very close collaboration between AI experts and experts in the problem domain. Furthermore, special technical support facilities (e.g., computer net­ works) can play a significant role in establish­ ing these collaborations and in sustaining their effectiveness. This is a key point which has important implications on organizational and shared resource aspects of applied AI projects. (c) The development of an expert system within a reasonable time span requires more powerful technologies than those in use today - expecially when the knowledge bases will grow from the pre­ sent 102-103 'facts' to more realistic situations with 104 - · lOS 'facts'. So far,system development times (from conception to expert level in a re­ search environment) have been 4-8 years. To re­ duce this time span, or to keep it from growing too much as knowledge bases grow, we need more effective methods of knowledge acquisition and organization and more powerful program design environments. Related to this, we need better techniques for interfacing AI programs with experts and users. At a more basic level, we need better schemes for coordinating multiple knowledge bases and for handling information which is in­ consistent and/or uncertain.

Main issues and open problems in AI applications

There has been considerable progress in the development of schemes for representing knowledge in AI systems. Production systems, semantic nets and frame systems are among the major schemes used in the projects discussed in this panel. Experience is accumulating steadily with types of representations that are appropriate in different situations. Much more work is needed, however, on how to represent knowledge of different types (form, completeness, validity) for various problem solving tasks, and how to shift representations in a manner that facilitates problem solving. At present the choice of representational framework for a task (the set of basic concepts and their relationships, the grain of knowledge and the form of knowledge to be associated with specific types of processes) is the key decision made by the builder of a AI system. The success of a system depends critically on this decision. This funda­ mental problem of problem representation (or of conceptualization, as Buchanan puts it in his remarks later on in this report) has been with us for almost a decade (Amarel, 1970); it is now gaining added significance in the context of knowledge~based system design. Briefly, in a dynamically evolving system where the initial de­ sign is based on a given choice of problem representation (conceptualization), it is possible that no increase in performance - via accretion of the knowledge base - can be achieved beyond a certain point unless a change in representation (a re-conceptualization) can take place. This has strong implications on methods of system design, and it may determine the limits of performance of specifi'c applied AI system. Experience with this problem already exists in the MYCIN and CASNET projects - where knowledge-based consultation systems of high levels ot expertise have been achieved in limited domains of medicine. Until more is known about how to choose appropriate re­ presentations and how to handle shifts among re­ presentations, it would be reasonable to start new AI application projects only in domains with fairly well established conceptual frameworks.

Knowledge bases are built through processes of knowledge acquisition and assimilation. There has been some progress to date in the development of system aids for knowledge acquisition from experts. In his remarks in this report, Pople raises important conceptual issues in this area. He points out that the process of acquisition of an expert's knowledge in a given domain is a theory formation process. This is an intellectually demanding activity which involves the formation of hypotheses about the expert's knowledge structure, and the representation .of these hypotheses in a computer with the help of available AI methods and tools. A key role of the computer scientist who is working with the domain expert in the develop­ ment of an AI system is to perform this empirical theory formation process. The question arises whether AI methods can be used to assist the computer scientist in this task. Some work in this area is underway at Rutgers; it involves the development of the AIMDS representational frame­ work which is intended to facilitate the formula­ tion and testing of a psychologist's theories in the domain of belief systems (Schmidt et al, 1976). Much more work is needed in this area - both in problems of theory formation and in related issues of representation. Also, methods of automatic knowledge acquisition and assimilation - from a system's operating experience or from direct observation of empirical data - must await further progress in theory formation strategies. Thus, more work in formation problems is needed to strengthen the processes of incremental construc­ tion and improvement of complex knowledge-based systems.

Many applied AI projects face the problems of how to represent and use multiple knowledge sources. Each source may represent knowledge of different grain (level of resolution) or of different form. For example one body of knowledge in a medical consultation system may represent a qualitative model of pathophysiological- relation­ ships in a disease process, another may represent a detailed anatomical model underlying (parts of) the qualitative model of disease, and ·a third may express normative rules of action for treatment decisions. Approaches to the use of multiple knowledge sources are discussed in this report by Kulikowski and by Brown in the context of medical and instructional tasks respectively.

One of the key proces:;es in medical r~ason­ ing and in scientific inquiry is the interpreta­ tion of empirical data in the light of a given body of theoretical knowledge. Much of the work dis­ cussed in this panel is concerned with these pro­ cesses. Typically, a problem of interpretation involves reasoning about an individual case. Given evidence (data) about the case and a theory within which the evidence is to be understood/explained,

find a hypothesis which explains the case on the basis of the theory. There has been considerable progress in the development of strategies for sol­ ving interpretation problems. However, there remain several open problems: under what conditions the interpretation process should be controlled by the specific 'low level' features of the case under consideration, or by possible 'high level' hypo­ theses and by expectations derived from these hy­ potheses; how to best represent and keep track of information about a special case, of general know­ ledge in terms of which the case is interpreted, and of alternative interpretations of the case as the process evolves with time. These issues are discussed further in thi5 report by Buchanan, Mar­ tin and Hart.

Hart is also stressing the importance of me­ thods of plausible reasoning in diagnostic prob­ lem solving. In addition to work reported by Hart in the context of a system for mineral exploration, considerable work on plausible reasoning was done in recent years in the context of the medical AI projects. Also, work by LeFaivre at Rutgers (Le ~ivre, 1976) has been focusing on systems for re­ presenting, and experimenting with approximate reasoning methods. More controlled experimentation, as well as theoretical consolidation of existing methods, are needed in this area.

An important feature of current AI applica­ tions is the emphasis on explanation facilities. Explanation of the reasoning done by a knowledge­ based system in special cases is needed both for system development/debugging purposes and for in­ teraction between system and user . This point is stressed below by Buchanan, as well as by other panelists. The methodologies and techniques of explanation developed for applied AI systems are expected to have a broader impact on the problem of scientific communication in AI. We are still faced with a major problem of how to communicate effectively what is known in the AI field . This involves the difficult issue of communicating com­ plex programs, principles underlying their designs, accounts of their behavior, and their properties. The problem is general for computer science, but it can be seen most forcefully in the context of AI work. There is reason to hope that current work on exPlanation facilities in applied AI system will result in the development of general AI methods for effective scientific communication of complex programs.

Experience shows that we need powerful design methods and system tools for building knowledge­ based systems and for evolving them from the con­ cept stage to a user's environment. This is bring­ ing most AI applications-oriented efforts in clos­ er contact with many other areas of computer sci­ ence and technology. There is renewed concern with programming languages and other support soft­ ware, machine architectures, networking, communi­ cations and interfaces. This situation will prob­ ably result in an enrichment of system building technologies - both within AI and outside it. More fundamentally, it will clarify further the essen­ tial concerns and scientific content of AI and its many relations with other parts of computer sci­ ence.

Before concluding, iet me stress that any co~ plete assessmentof the state of AI applications must include two parts . The first should be con­ cerned with the performance achievements and the levels of expertise attained by specific programs in various application areas. The second, must fo­ cus on design issues that go across specific appli­ cations, and on the identification of new AI prob­ lems and approaches. The thrust of this panel is on the second part, because this part is more in­ trinsically relevant to the status of main techni­ cal issues within AI - and thus it is more appro­ priate for an AI Conference. However, this should not obscure the fact that in 1977 we are at a point where substantial achievements have been made in AI applications, and the prospects for the future of high performance AI expert systems are bright.

Let me now ~ummarize the areas of AI that I believe require more work in order to strengthen , the basis for the design of useful knowledge-based systems : guidelines for representing and using knowledge of different types; schemes for shifting representations; methods for solving formation problems; flexible strategies of interpretation; methods for acquiring and managing large bodies of knowledge ( 105 facts)from experts and from empir­ ical data; and programming techniques for AI sys­ tem development and for system transfer to various types of user environments. In addition, in order to increase the impact of AI on applications, we must find effective ways of organizing what is known in the field and of communicating this know­ ledge to others.

In the following sections, members of the panel - who represent major AI applications ef­ forts - will comment on several of the issues that I discussed above. Their comments are presented in the context of specific application projects with which they are associated. A list of refer­ ences concludes this panel report.

REPRESENTATION .AND USE OF EXPERT KNOWLEDGE IN MYCIN (B . Buchanan)

Successful applications of AI to. ~cience and

medicine require large amounts of specific know­ ledge. Yet this presents problems for the repre­ sentation, acquisitionand use of the knowledge by an AI program, as evidenced in the MYCIN program developed at Stanford (Shortliffe, 1974; Short­ liffe et al, 1975; Davis et al., 1977). This ten­ sion is not wholly resolved in MYCIN but we believe that extensions to the methodology will alleviate it.

A major goal of the MYCIN system was to pro­ vide a computer-based therapeutic tool designed to be clinically useful. This requires development of a system that has a medically sound knowledge base, and that displays a high level of clinical competence in its field.

Since many clinicians are not likely to ac­ cept the advice provided by a computer-based sys· tern unless they can understand why the recommended therapy has been selected, the system has to do more than just give advice dogmatically. It should have the ability to explain the reasoning behind its decisions, and should be able to do so in terms that suggest to a physician that the program ap­ proaches the problem in much the same way that he does. This permits the user to validate the pro­ gram's reasoning, and modify (or reject) the ad­ vice if he believes that some step in the decision process is not justified.

We have also attempted to provide the program with capabilities for experts in infectious disease therapy to augment or modify the knowledge base, in order to improve the validity of future consulta­ tions. The system therefore, includes a knowledge acquisition system which enables experts to update MYCIN's knowledge base, without requiring that the,r know how to program a computer. A principal fea­ ture of MYCIN, central to these goals, is the for­ mat in which its knowledge is encoded. Knowledge used by MYCIN is contained in diagnostic and thera­ peutic decision rules formulated during exten­ sive discussions of clinical case histories. The MYCIN knowledge base currently consists of approx­ imately 400 such rules. Each rule consists of a set of preconditions (the 'premise') which, if true, justifies the conclusion made in the "action" part of the rule. For example: If 1) the gram stain of the organism is gram negative, and 2) the morphol­ ogy of the organism is rod, and 3) the aerobicity of the organism is anaerobic, then there is sug­ gestive evidence (.6) that the identity of the or­ ganism is bacteroides.

Such rules form modular "chunks" of knowledge about the domain, represented in a form that is comprehensible to a clinician.

The consultation system uses its collection of rules to make conclusions about the patient. If, for instance, it is attempting to determine the identity of an organism responsible for a particu­ lar infection, it retrieves the entire list of rules which, like the one above, conclude about identity. It then attempts to ascertain whether the conclusion of the first rule is valid, by e­ valuating in turn each of the clauses of the pre­ mise. Thus, for the rule above, the first thing to find out is its gram stain. If this information is already available in the data base, the program retrieves it. If not, determination of gram stain becomes a new subgoal and the program retrieves all rules which conclude about it, and tries to use each of them to obtain the value of gram stain. If, after trying all the relevant rules, the an­ swer still has not been discovered, the program asks the user for the relevant clinical informa­ tion which will permit it to establish the vali~­ ity of the premise clause. Thus, the rules "un­ wind" to produce a succession of goals, and it is the attempt to achieve each goal that drives the consultation.

The use of a rule-based representation of · knowledge makes it possible for the system to ex­ plain the basis for its clinical recommendations. For example, if the clinician asks "How did you determine the identity of the organism?" the pro­ gram answers by displaying the rules which were actually used, and explaining, if requested, how each of the premises of the rules was established. This is something which the clinician can readily understand, and it provides a far more comprehen­ sible and acceptable explanation than would be possible if the program were to use a simple stat­ istical approach to diagnosis.

As work proceeds to expand the program's know­ ledge base, new "chunks" are added in much the same way that a clinician in training learns new pieces of knowledge about his field. This rule-based re­ presentation of knowledge means that the expert himself can offer new "chunks" of knowledge by expressing them in the same rule~based format. He can thus help make the program more competent, with­ out having to know anything about computer program­ ming. In addition, since the rules are independent of one another, and are used by the program as ne­ cessary in order to deal with the particular con­ sultation underway, the addition of a new rule or modification of an existing rule does not require altet· ation of other items in the knowledge base, as is often necessary with systems using the decj­ sion-tree methodology.

In order to represent the medical knowledge of infectious disease diagnosis and therapy for MYCIN, the designer of the rule base (the "expert") must first decide on the set of concepts that the rules will relate. In domains that are not fully codi­ fied already, this is a significant problem because there are many ways to think about the world and many ways to describe the organization of the des­ criptive terms. The whole methodology depends on the expert's choice of an adequate conceptualization.

The "grain size" for conceptualizing the domain will change a·s the expert gathers more experience with the performance of the system. That is, new · distinctions have to be made and, occasionally, se­ parate concepts need to be merged into a single one. The representation of rules can support this now as long as the descriptive terms are organized hier­ archically and the finer grain terms are added to the lower ends of the hierachy. Reorganizing the whole tree is more difficult than I would like.

We have found also that experts give the pro­ gram rules that are nice general principles, in a desire to codify the domain, but that these general rules have so many exceptions that high performance is impossible to achieve until the exceptions are also codified. This means that the expert has to be reminded to put in all the exceptions to the general principles, and has to spell out the excep­ tions in tedious detail. Moreover, it means that the representation for the rules mut tolerate many special ··case rules and exceptions and still be ef­ ficiently interpreted.

As the amount of detail required for good per­ formance grows, the amount of time the system needs for obtaining the information will also grow. This may show up either as increased I/0 time, to ask the user for the details required for making new inferences, or as increased computation time for reasoning through longer inference chains. We find that we are caught between the necessity of asking users for more and more data and the necessity of keeping the whole consultation short enought for users with little time and patience.

The formalism used by MYCIN can now include rules that ar~ triggered by events (in particular by new information), but it is not yet possible to set up expectations of future events that will con­ firm or disconfirm a current hypothesis. This is symptomatic of the difficulty we have in represent­ ing time dependent relations in static rules. For example, much of a physician's reasoning is of the form "If you see X today then look for Y tomorrow, and take appropriate action if you fail to observe Y (or if you do)."

These problems, among others, are topics of current research . Some extensions to the methc·d­ ology will be necessary, but we still believe ia the basic idea of goal-directed reasoning from an expert's set of rules.

ON THE KNOWLEDGE ACQUISITION PROCESS IN APPLIED A.I . SYSTEMS (H. Pople)

The process of knowledge acquisition, as this concept applies to the development of know­ ledge-based AI programs, can be considered from several different perspectives . There is, for ex­ ample, the ~nowledge acquisition process employed by the computer scientist who is called on, typ­ ically to develop a model of the expertise re~ quired to function productively in some problem domain. If we take such a model . as given, there is then the process of know­ ledge-base accretion, which expands the knowledge base in accordance with the fixed precepts of in­ formation structure and process underlying the gtven model. I would like to concentrate my re­ marks on the former of these two modes of know­ ledge acquisition, then discuss briefly what im­ plications this carries with respect to the gen­ eral concept of computer-based "knowledge engin­ eering".

What I consider to be the primary mode of knowledge acquisition for knowledge-based systems was brought into focus recently during an exposi­ tion of INTERNIST. This is the system we have de­ veloped at the University of Pittsburgh to pro­ vide cognitive support in the formation and solu­ tion of difficult clinical problems in internal medicine (Pople, et al. 1975; Pople 1975, 1977). Dr. Jack Myers revealed that when he and I embarked on the task of building a computer-based expert consultant for internal medicine some six or seven years ago, he knew nothing of computers and I knew even less of medicine. This prompted a noted philosopher of science in the audience to wonder aloud how we ever managed to get started. How were we able to ask useful questions of one another and make use of the answers provided?

On the basis of my experience in this mode of knowledge acquisition, I am inclined to character­ ize the modeling of expertise by a naive inquirer interacting directly with an expert as essentially an empirical process. One begins by acquiring a small number of facts, on the basis of which some hints may emerge as to the expert's structuring of knowledge. These structures in turn may provide cues as to the process by which this knowledge is accessed and used in the course of reasoning and problem solving. Further process cues may be ob­ tained from direct testimony of the expert, from a study of protocols and the expert reasoning a­ loud, and from the inquirer's own introspections concerning the structure and process of knowledge in other domains.

The next step is to fashJ~n from these empir­ ical findings a testable hypothesis; and here, for the computer scientist, there is a considerable ar­ mamentarium including but not limited to the models and methods of AI that can be brought into play in the construction of a working model. It should per­ haps be emphasized that what results from this stage is at best a model of the inquirer's concept, which may or may not bear much resemblance to that of the expert. Nonetheless, if reasonably faithful to the inquirer's emerging concept of the domain, the model can serve to guide and sharpen the furth­ er search for more subtle aspects of the expert's information structure and process.

There are a number of implications that flow from this characterization of the knowledge model­ ing process that I would like to put forth as top­ ics for consideration by this panel.

First, the term "applications" often used to characterize this sort of investigation is some­ thing of a misnomer. It suggests the existence of a general theory which is being instantiated by a "knowledge engineer", whereas experience suggests that the modeling of expertise is primarily an em­ pirical theory-formation activity.

A corollary of this proposition is that the models and methods of AI, which often prove useful at various stages of the modeling process, should be regarded merely as tools of the investi­ gator - not theories. One should not set about to fit a model of expertise to the models of AI. If would be better to devise new methods and techniques even at the risk of being called "ad hoc" if this is necessary in order to deal with the essential nature of the problem domain being investigated.

It also follows that one must be prepared to throw over any or all of a given model when further investigation reveals subtleties of expertise that cannot be represented within that framework. One must take whatever comfort there is in knowing that the model has served it's purpose if it has sharpened the investigator's awareness of the structures and processes that underlie expertise and then move on to the more sophisticated models and experiments enabled by this new level of understanding.

Tt is clear that there are certainaspects of knowledge acquisition currently amenable to com­ puter-based support. For example, at any stage of model development, there is typically the need to have encoded a sizable corpus of domain specific information so that realistic experiments may be carried out. Many techniques can be employed to facilitate this knowledge-base accretion process through direct interaction with experts in the do­ mains being_modeled 1 through induction on cases, or simply through batch entry of prestructured da­ ta. While thes.e methods of knowlege accretion are important, they constitute only a small part of the knowledge modeling process that we have here characterized as "empirical theory formation". Until we are able to model the role of naive in­ quirer and incorporate this critical aspect of the knowledge acquisition process in a computer­ based syste!Jl,i t seems . clear that the most signifi­ cant mode of knowledge exchange in the develop­ ment of knowledge based systems will continv.e to be person to person.

REPRESENTATION OF KNOWLEDGE FOR REASONING IN MED­

ICAL CONSULTATION SYSTEMS (C. Kulikowski)

Building a flexible and sophisticated knowledge bas.ed expert consultation system in medicine is a formidable task oecause of the complexity and het­ erogeneity of medical knowledge and our very limi­ ted understanding of clinical reasoning processes. During the past five years several artifical in­ telligence approaches have been taken in building descriptions of patients and diseases that combine knowledge from a variety of sources with a diver­ sity of structural representations, and experiments have been carried out with an equally varied array of inferential problem solving strategies. (Kuli­ kowski and Weiss, 1971; Amarel & Kulikowski, 1972; Shortliffe 1974, 1975; Pople et al, 1975; Rubin, 1975; Silverman,l974; Pauker, et al. 1976) . .

In the CASNET program which. we developed at Rutgers (Kulikowski & Weiss, 1971; Weiss, 1974), the knowledge of a group of clinical researchers in the domain of a disease is structured in the form of a causal-network representation which des­ cribes the mechanisms and evaolution of disease processes. The variability of individaul obser­ vations is accounted for by postulating a related associational structure of observations. The caus­ al network then stands as its underlying concep­ tual model.

The CASNET/GLAUCOMA system is an application of CASNET in the domain of the glaucomas. The system can be utilized in a variety of reasoning modes to provide diagnostic, prognostic, and ther­ apeutic recommendations, together with explanations and references to diverse, alternative expert opin­ ions.

A novel characteristic of the CASNET/GLAUCOMA

system is that it can, for a particular case, sim­ ul~neously present alternative opinions and reason­ ing derived from different consutlants. To provide the system with a variety of opinions we have es­ tablished a computer-based network of collaborating glaucoma expert consultants who share in the de­ velopment and testing of the programs. This oph­ thalmological network (ONET) includes glaucoma re­ searchers at the Mt. Sinai School of Medicine, Johns Hopkins University, Washington University, the University of Illinois at Chicago and the Un­ iversity of Miami.

By representing in the computer detailed pat­ terns of disease evolving with the passage of time, we are able to deal with multiple follow-up visits. Sequences of suggested therapies for the various types and stages of progression of glaucoma have ru­ so been incorporated in the system. Specific know­ ledge of disease is continually added by the ONET members. Tfiis is done by presenting the computer program with a variety of difficult clinical cases and weighing its performance against the judgment ·of the panel of consultants. Their suggestions are used to refine the diagnostic and therapeutic re­ commendations, improve the systems' assessment of signs and symptoms, and to perfect specific tech­ niques of acquiring and displaying the clinical da­ ta. Currently, the program runs and cases are stored on the SUMEX-AIM computer at Stanford, which is accessed via TYMNET from Rutgers and from · the ONET sites. The stored cases form a data base which serves as a source for clinical studies on prognostic indicators and treatment evaluation. Selected results from such studies can then be · used to improve the model of disease. The develop­ ment of the glaucoma system with the help of ONET has demonstrated the feasibility of collaboration between geographically remote medical investiga­ tors working on a common computer-based research project.

The CASNET/GLAUCOMA system was demonstrated and tested at the 1976 Meeting of the Anerican Academy of Ophthalmology and Otolaryngology. In addition, · the system was part of the symposiUm on glaucoma which was held at this meeting, where its recommendations were compared to those of a panel of experts. The results of this demonstration and evaluation were most successful. Most of the oph­ thalmologists who tested the glaucoma consultation system judged it to have reached a very competent­ to-expert status(95%). From the comments received at the meeting, it is clear that the scientific and health care significance of the system is widely re­ cognized by now.

More recently, we have explored several dif­ ferent representations for medical knowledge that extend the range of concepts and relationships be­ yond those defined in the CASNET formalism. Our mo­ tivation has been dual: a) to provide more modular descriptive capabilities so that explanations of reasoning can be more explicit than in CASNET; b) to broaden the scope of reasoning strategies as they apply to a greater variety of conceptual el­ ements. The first alternative representation that we used was the semantic network, which permitted us to define relationships other than causality at the process level (preconditions, complications, triggering effects, etc.). The strategies for de­ ~ision-making were formulated as normative rules for propagating information and confidence judg­ ments between nodes in the semantic net .in response to patient data. This provides a very general and flexible means of tracing the flow of reasoning for a particular case. A system called iRIS, embodying these ideas is now at an advanced state of imple­ mentation (Trigoboff & Kulikowski, 1977).

A more recent development involves the formu­ lation of a template-based description, which is well suited for building of the anatomical models that are associated with the existing process mo­ dels. We have most recently begun to develop fa­ cilities for aiding in knowledge acquisition from the expert using the semantic net and template ap­ proaches. A facility for building a model in neu­ ro-ophthalmology (visual pathways, and related an­ atomical structures together with the physiology of neural transmission) is being developed (Mathew, Kulikowski, and Kaplan, 1977).

In conclusion, we believe that the introduc­ tion of knowledge-based systems in medicine may help to reduce the gap between science and practice in this area. Medicine is primarily concerned with diagnostic decisions and with the design of courses of action for the treatment of patients. Knowledge of different types, some of which comes from re­ search in the life sciences, enters into the rea­ soning processes of the medical decision maker.By studying medical problem solving in a computer, in the context of models that embody physiological/ pathological knowledge of body processes, we ex­ pect to obtain a better understanding of how this knowledge can be used in the reasoning processes that lead to medical decisions. If we establish systematic ways of using such models in medical problem solving, then new knowledge about a body process, a basic mechanism of a disease, etc.,can be related to clinical observations and actions and thus it can be used and tested in medical prac­ tice. The contribution of the computer system in these situations is in handling the complexity of the relationship between a large and changing body of relevant general knowledge and decisions in special cases.

The work on applications of AI to the devel­ opment of medical consultation systems is part of our research in the Rutgers Research Resource on Computers in Biomedicine which is supported by the Biotechnology Resources Program of ORR, NIH.

REMARKS ON KNOWLEDGE-BASED PROGRAMS (W. Martin)

In the past few years we have worked on a num­ ber of knowledge-based programs at the M.I.T. Lab­ oratory for Computer Science . These include:

1. A present illness program focussing on kidney disease. (Pauker et al, 1976; Szolovitz and Pauker, 1976)
2. A digitalis therapy advisor (Silverman, 1975; Swartout, 1977).
3. An acid/base electrolyte disequilibrium advisor.
4. A program for the design of procurement systems. (Bosyj, 1976).

1. A program to assist in simple financial projec­ tions.
2. A program to write simple programs.

The work related to medicine has been funded by the Bureau of Health Manoower and involved Pro­ fessors G.A.Gorry (Now at Baylor Med.Center), and P. Szolovitz and Drs. W. Schwartz, S. Pauker and J. Kassirer at Tufts New England Medical Center. The work in nanagement and programming technology has been funded by ARPA and involved Prof. A. Hax and me.

One can divide such interactive systems, or functions of an eclectic system into three types:

1. ·operator based -- The system defines data ob­ jects and operators which may be applied to them. If is up to the user to determine a use­ fule sequence of operators. A high level lang­ uage 1 take to be a weak form of operator based system.
2. Model based The system provides one or more models which can be parameterized and run against the user data.
3. Knowledge based -- The system is operator and/ or model based, but it also 1) can exhibit self knowledge in the form of explantation, 2) is able to help the user formulate his problem be­ cause it has knowledge of situations which of­ ten can be cast in a form to which its opera­ tors · or models would apply.

There are several unresolved alternatives con­ cerning the best architecture for knowledge based systems. Consider first the problem of explanation . In the digitalis therapy advisor we have employed what might be called "semantic programming". All functions and variables have been called by the names they would have in English and the programs have been structured, with details supressed into subroutine calls. It is then possible to write a simple routine which translates a subroutine into English - high level subroutines give an overview of the processing and lower level ones give more details. By describing a routine or its execution in English the system can answer "How would you" or "How did you" questions. ·ane advantage of this ap­ proach is that we are guaranteedthat the explanation is not out of date with the code. There are two dis­ advantages. First, it is necessary that the user un­ derstand the terms and the model that is actually used in the calculations. Since there is common ter­ minology among users of digitalis this is not a problem with that program, but in experimenting with managers querying a data base we found a wide range of terminology for the same model concept. (Malhotra, 1975). A group at IBM Yorktown have op­ ted to build a separate knowledge base for inter­ preting the models to the users. This is obviously a difficult strategy to implement. Even the stra­ tegy employed in the digitalis advisor forces a careful structuring of the program for clarity of explanation. Those espousing production rules sometimes claim that even this can be avoided. The work of abstraction and organization turned over to the system. The best approach is not clear.

On a different but related dimension we have observed th~t our present illness program exhibits bad medical style. It often asks apparently un~eed­ ed questions; it sometimes runs on and on pursuing ever less plausible dead ends, and when it works correctly, it occasionally does things in a non­ standard order which is confusing and annoying.The standard style of operation of some of today's most sophisticated medical diagnostic programs(Pauker et al.,l976;Shortliffe,l974; Pople et al.,l975) in­ volves the application of a single complex cy­ clical processing algorithm to a uniformly struc­ tured database of medical knowledge. The processing algorithm repeatedly performs the following steps : 1) select a disease hypothesis, network node, or heuristic rule to invesitgate; 2) get some relevant fact from the database or program's user; 3) update the state of the database according to the inter­ pretation of the new fact; 4) re-iterate this process . In its responsiveness to new information this scheme is very attractive. However, it is very dif­ ficult to make a program structured in this way ex­ hibit a natural clinical style. Because the data­ base contains information only about diseases and their symptoms, and not about the state of the con­ sulting program, the responsibility for both correct medical conclusions and appropriate medical style rest on the purely medical knowledge and its use. The expert physician seems to have a great deal of specific experience which suggests to him definit­ ive ways of accomplishing a diagnostic goal . What we need is a way to add such strategies to our pro­ gram. One approach is to view these specific ex­ pert strategies as replacements for parts of the ~eneral diagnostic strategy, which replacements ap­ ply only in specific circumstances. Thus, before running a step of the general diagnostic engine, the system would first look to see if there were a more specific strategy pertinent. The recogni­ tion of when a strategy is applicable is a classic problem of AI. Here it will probably require that the data be characterized at different levels of abstraction. Currently, data are abstracted only to a small number of classes - disease, symptom, etc. - used by the single diagnostic cycle.

A further, nagging thought is the apparent orthogonality of the abstractions suggested by Schank to those required to cast all data into a particular scientific mode 1 or user point of view. While Schank seeks a uniform way to represent all knowledge which is problem independent, all useful svstems are prob­ lem dependent - their special model of the problem is at the heart of what they have to offer. To date, Schank's abstractions have had litle, if any, impact on knowledge based system- yet they do seem to capture important ideas about the re­ presentation of knowledge. What is going on here? Is it that Schank's techniques provide a level of generality to wn1~h we have not yet aspired, should not aspire, or will soon need to aspire? OBSERVATIONS ON THE DEVELOPMENT OF EXPERT KNOW7

.

LEDGE-BASED SYSTEMS (P. E. Hart)

The last five years have seen art accelerated interest in applications of artificial intelligence with what to my mind are beneficial effects both scientifically and politically. The most immedi­ ately practical of these applications use AI tech­ niques that are well-understood. For example, the commercial picture processing and word recognition systems available today are based directly on tech­ niques developed in the latesixties and early sev­ enties. Other applications, which may have equally_

practical goals, require more profound use of the methods of artificial intelligence. These applica­ tions are more interesting from a scientific point ofview, and I shall examine a few issues that they raise in the following.

Among the most prominent of these application themes is . the development of "expert knowledge­ based systems . " These systems -- DENDRAL (Buchan­ an, et al, 1969), MYCIN (Shortliffe, 1974), INTER­ NIST (Pople, et al., 1975) and others-- have forced AI technology ahead because of the need to r epresent and deploy substantial amounts of real­ world, or domain-specific, knowledge. Further, because the computer scientists involved in the design are usually inexperienced in the subject domain, there is a heightened need to develop me­ thodoldgies for extracting relevant knowledge fran those who are experienced. This is in sharp con­ trast to an earlier period of AI research, when the designer, programmer, and domain expert could oe one and the same person. ·

An expert system under development at SRI furnishes a number of instructive examples of how current applications efforts are advancing AI technology. We first briefly describe the sys­ tme, and then focus on some specific examples. PROSPECTOR (Duda et al., 1976; Duda et al, 1977) is a developmental system aimed at aiding geolo­ gists in their search for mineral resources. In some respects, the mineral exploration process is similar to the process of diagnosing diseases. A body of observations, often uncertain in nature, must be interpreted with the aid of a knowledge base that typically supports plausible reasoning but not strict logical inference. It is not surpris­ ing, therefore, that PROSPECTOR bears some resem­ blance to existing mediml diagnosis systems,parti­ cularly to MYCIN.

Domain specific knowledge is given to PROSPEC­ TOR as a set of inference rules of the form A IM­ PLIES B. Two numbers associated with the rule mea­ sure the degrees to which A is necessary and suf­ ficient for B. A prior probablility is also asso­ ciated with the conclusion B.

The inference rules are represented internally as a partitioned semantic network along the lines suggested by Hendrix. Such networks when used in their general form have all of the expressive power of predicate calculus, and additionally encourage the exploitation of subset-element relations, in­ verted indexing on terms, and a number of other such features. We are currently·using a pecialized form of semantic network with which we can easily represent both an important subset of predicate calculus formulas, and the inference rules themselves.

The top-level nodes of the network correspond to tqp~evel hypotheses about the presence of the various types of ore deposits. Lower-level nodes may correspond to directly observable geologic da­ ta, or to intermediate concepts that cannot be di­ rectly observed but that can be inferred from ob­ servables. A principal task of the system is to in­ fer probabilities for the top-level hypotheses on the basis of available observations.

The special characteristics of the mineral ex­ ploration process force the system designer to consider a number of issues that have not, to my know­ ledge, received much attention to date, but that could well have relevance to existing and future systems.

questions of Existence

The hallmark of exploration is the search for the existence of objects with specified properties . Thus, the quintessential question for the user (i. e., the geologist) is "Can you find an object X sat­ isfying P (X)?" The issue raised here by PROSPECTOR is in interesting contrast to medical diagnosis systems, for which the existence of a culture, a test, or indeed a patient is eldom in question. On philosophical grounds, it is always hard ·to rule out the possibility of finding something yet undis­ covered; on practical grounds, a committment to the absence of X sometimes needs to be made, leading to a need to handle the statement "There does not ex­ ist an X satisfying P(X)". PROSPECTOR currently handles this by denying the existence of any "real­ world binding" for a formal object satisfying P. However, although this statement is logically equi­ valent to the statement "For all X not P(X)" PROSPECTOR cannot handle the very similar statement "For all X Q(X)" except by setting Q equal to not P .· This stems from a repre­ sentation that highlights the possible existence of objects, rather than one that stresses logical com­ pleteness.

Problems of the existence of objects also lead to interesting choices about how to ask questions of the user. For example, a user may doubt (but not rule out) the existence of an X satisfying P(X). If we wa~t to ask him about other properties of X, we may ~l.nd ourselves in the position of asking "Does ~,wh1ch probably doesn't exist (but may), also sat­ J.sfy Q~X)?" PROSPECTOR handles this issue by simply resort1ng to a threshold probability for X below which X is deemed to be non-existent, but this ex­ pedient has little theoretical justification.

Antecedent reasoning

A typical consultation with a field geologist would begin by having the geologist tell the system about the significant geological features of the case, or "prospect" of interest. Ideally, the sys­ tem should be able to use this volunteered informa­ tion to help it focus on hypothesesof interest. It is therefore, important to be able to reason in the forward direction from volunteered features to the top-level hypothese that they suggest (even weakly) . The structure of PROSPECTOR supports antecedent rea­ soning through the use of an ·explicit network repre­ sentation of inference rules, and through the use of a taxonomy discussed below. It is interesting to note that Pople's INTERNIST system accomplishes the same aim in a very different system framework.

Taxonomies

A surprising amount of geologic reasoning is done thro~gh ~axonomies of objects. For example, chalcopyrJ.te l.S a copper-iron su!fide, which is a copper sulfide, which may be significant evidence for a certain type of copper deposit. PROSPECTOR easily handles such taxonomies using the semantic net mechanism. However, the current state of geol­ ogy does not support the use of taxonomies indiscrim­ ~nately since, for example, many rocks can be classified in any of a number of ways. (This provides an example of an opportunity for work in artificial intelligence to contribute to the codification of another field, a point that others have also re~ marked upon).

Subset relations are exploited by PROSPECTOR in another, slightly more subtle, way to form in­ ference and to affect the flow of control. An aux­ iliatynetwork mechanism is used to record subset hierarchies implied by objects having some shared properties . For example, suppose there are two rules, one requiring "rhyolite intrusive" as an antecedant and one requiring "rhyolite" as an antecedent. If the user tells us that a rhyolite intrusive is pre­ sent (with some probability), then we can conclude that rhyolite is also present (with at least that probability) and the second rule can be invoked . Similarly, denying the existence of rhyolite also denies the existence of a rhyolite intrusive and both rules can be marked as not usable. Of course, separate rules could also have been used to repre­ sent the implicit subset hierarchies, but only at the expense of considerably more complexity.

Contexts and control

Most geologic reasoning can only occur within a semantic context that has to be established be­ fore details are considered. For example, the exis­ tence of an intrusive body must be established be­ fore considering its age and mineralogy. This leads to a number of related design issues. The domain­ specific expert must provide not only the inference rules and associated numerical values, but must fre­ quently specify orderings for establishing the ex­ istence of objects. Correspondingly, the control strategy used by the system must make use of the 'advice" provided.

PROSPECTOR addresses some of these issues through the use of mechanisms like ordered conjunc­ tions and disjunctions. Current work,however, semts to be leading to a richer design combining class­ ical production systems with PLANNER-like concepts that permit control strategies to be partially spec­ ified at design time. As this design evolves, mod­ ularity and extendability will become more difficutt to maintain,. and the problem of communicating with experts about their specialties will become more acute.

The foregoing observations touch on only a few of the issues associated with designing an expert system for a new domain of professional knowledge. It is interesting to note that most of them were un­ suspected by the system designers at the outset. These and similar problems arereing actively worked on, and new techniques are being developed for their solution. I have little doubt that, as artificial intelligence continues to address problems of prac­ tical import, new research problems will continue to unfold.

Acknowledgements

The development of PROSPECTOR is supported by the Office .of Resource Analysis of the U.S .Geolog­ ical Survey. The technical work described is the result of the active collaboration of R.O.Duda, N.J.Nilsson, R.Reboh, and G.L.Sutherland.

REMARKS ON BUILDING EXPERT SYSTEMS (J . S.Brown)

During the last two years Dick Burton, Johan DeKleer and I have been involved in the"construc­ tion of intelligent instructional systems built a­ round expert problem-solvers/reasoners in the areas of electronics and algebra. Most of our effort has been directed at developing techniques which use "expert" systems to help 1) explain correct solu­ tions (which, in addition to an articulate expert, requires a theory of explanation and a structural model of the student), 2) predict "mind bugs" in a student, 3) synthesize deep structure models of how a student is reasoning and 4) provide hints on what to do next (which depends on how he is currently at­ tacking the problem). We shall focus our remarks here on issues pertaining to building habitable ex­ pert systems.

Multiple representations of knowledge

In our first electronics tutorial system, SOPHIE!, (Brown and Burton, 1975) we sacrificed ar­ ticulateness for efficiencey and robutstness. Our intuitions led us to conceptually separate finding an answer to a query from explaining the answer. In particular, we felt that once we had determined the answer fbv oossibly non-human like inference schemes such as constraint satisfaction) we could use a logically incomplete rule based system to "ra­ tionalize" the answer in understandable terms. As further developed in our SCA paradigm (Synthesis, contingent knowledge structure [CKS], analysis), (Bobrow and Brown, 1975), SOPHIE! was .decomposable into synthesis specialists (in the form of SIMULA­ TORS), analysis specialists and an information struc­ ture (CKS) that mediated the "world views" of each . The synthesists incorporated electronic laws, mod­ els of electronic components, a particular circuit description, boundary conditions and from this know­ ledge synthesized a complete state description of the circuit. Our synthesists left only as much of an "audit trail" or "proof" as was necessary to handle a limited class of causality questions -- fault pro­ pagation. The output of the synthesis was a highly structured data base (contingent knowledge structure) whose exact form and content was known by a collec­ tion of procedural "analysts" that knew how to de­ compose a question, say about power dissipation. Into a set of primitive requests to that data base. These analysts, that also contained models of elec­ tronic laws etc., left an extensive audit trail for explanations and took the upmost advantage of the canonical forms created by the synthesists. In brief, the synthesists and analysts reflect a subtle use of multiple representations of knowledge. Each special­ ist represents its knowledge in a form ideal to car­ rying out its tasks. The tasks were determined by . finding a decomposition of the overall problem into segments which enabled leverage to be realized through the existence of special inference tech­ niques/representations.

Another form of multiple representations of knowledge in SOPHIE! is in its use of multiple le­ vels of circuit simulators (multiple synthesis specialists) . In some cases, a crude but extremely efficient functional simulator determines an approx­ imate state description; if that state looks "use­ ful",then a more accurate simulation rPfines the description. In particular, in constructing a set of possible theories of what could be wrong with the circuit,this higher level simulator helps to instan­ tiate unbound variables in the theories and/or prune existing theories. By using hierarchical simulation models, we can sometimes increase the speed by two to three orders of magnitude, thereby enabling us to use simulation in ways that would otherwise have been too costly.

Our newer electronics system, SOPHIE3 (Brown et al., 1975; DeKleer, 1976) moves closer to human­ like reasoning. It employs a mixture of circuit dependent rules capturing teleological knowledge a­ bout a g'iven circuit, with a general purpose elec­ tronics reasoner for making simple, local deductions. [The general purpose reasoner is based on Johan DeKleer's work and also draws heavily on Sussman's EL].This mixture arises out of the need to make ex­ tremely subtle deductions based on a deep understand­ ing of the underlying teleology of a circuit while maintaining generality so that the system can handle multiple circuits.

This system's task is to take a measurement (or collection of measurements) made by a student, while tiDublehooting a circuit and to drive its logical consequences, including such information as what could still be at fault and why (a module needed by our troubleshooting tutor/critic.) The problem was that some of the inferences that an "expert" makes depend on a very deep model of how the cir­ cuit was designed--its underlying teleology. Since we did not know how to build a robust teleological reasoner, we opted for expressing this knowledge as a collection of "demons" or production rules that watch for the values of certain measurements and, once these are known, assert some fact about the state of given functional blocks or components of the circuit. The conclusion reached by a "de­ mon" was not proved per se but nevertheless could be explained by simply attaching a hypothetical explanation to the "demon" which references the underlying teleological models. Eventually, we might have an adequate enough theory and inference mechanism to be able to automatically derive/prove these high level rules/demons from a circuit model. In other words, the "demons" had compiled into them our understanding of the teleology and the consequences therein.

This scheme was potentially very brittle, as have been most applications of frame-based know­ ledge systems. Each "demon" had its slots to be filled and if the information coming into it did not perfectly fit into the preordained "view of the world" as defined by its slots or portals on the circuit that the demons were watching, im­ portant deductions would be missed. What was needed was a general purpose measurement propagator that could take whatever measurements the student made and locally propagate them onto the preordained portals of the demons. Since the smashing of the student-made measurements into these slots in­ volved only local propagations, the propagator itself did not have to know the teleology of the given circuit. It only needed knowledge about how each kind of component worked -- transistors, di­ odes, resistors, etc. Since each step taken by the local propagator is justified by a general law, the sequence of steps leading to a "demon's" firing is a "proof". Thus complete explanation of a con­ clusion reached by a demon requires merging the teleological rationalization of the given "demon" with the proof generated by the local reasoner.

In this way we get both robustness (from the general purpose propagator) and powerful inferences from the "demon-based" system, each containing either explicitly or implicitly a fair amount of electronic knowledge. Of course, the success of this system depends critically on having the right level of teleological knowledge compiled into the demons.

Executable models

The relatively simple reasoner underlying our original SOPHIE system can be thought of as an example-driven inference scheme that combines an executable or runnable simulation model with a· heuristic scheme for choosing useful examples to run on the model. In particular, to determine"if shorting X would cause Y to blow", it chose boundary conditions that would stress the relevant sections of the model and then simply runs the model (note that the selection of boundary condi­ tions requires, in general, teleological knowledge of the circuit). If Y did not blow, then othe~ boundary conditions can be tried. Also, note that the kind of reasoning and knowledge used in select­ i~g examples/boundary conditions is qualitatively different from that contained in the simulation model or the analysist that interpreted its output.

Our newer SOPHIE3, also uses an "executable" model of a· circuit but its model is quite different. Very briefly, it resembles a hierarchical finite state machine that reflects a teleological and functional abstraction of the real circuit. This model functions, in part, like the mechanistic models underlying CASNET and in part like context tree mechanism in MYCIN. In particular, it helps to define the epistemology of the "demons" (as in MYCIN), as well as to organize the propagations of their conclusions.

These models are beneficial because I) they have the underlying control structure for ·glueing together segments of knowledge and hence no search is required to find relevant rules and II) unlike axiomatic based reasoning schemes they don't need rules or axioms for excluding alternative possible worlds/models; they are, in themselves, the model.

System tools

In addition to the obvious kinds of AI tools such as a Semantic Network package -- SEMNET -­ there are two kinds of tools that we have found exceptionally useful in building "friendly" and useable systems. The first pertains to an approach for building efficient and habitable natural language frontends. Although having to cope with man-machine interface problems is, at best, a distraction to solving the problem of building expert tutorial systems, it is, nevertheless, a crucial problem. Over the last several years, we have developed a technique for producing natural language front ends based on the notion of semantic grammars [Burton, 1976]. Recently, Burton has advanced the state-of-the-art of building habitable front ends by building an ATN compiler that takes a semantic grammar expressed in the ATN formalism and a specification of a desired parsing strategy and then produces a program that is the desired natural language system. What makes this task especially exciting is that by changing the code generators of the compiler he can produce a natural language processor for any virtual machine, that only needs to have a runtime environment built for dictionary lookups etc.

The second basic "tool" that we have used a great deal has to do with achieving a debugged, habitable system. This methodology involves designing a system so that any time a module of the system encounters any internal error (such as a sentence not parsing) the current context of the system along with a "transcr~ipt of the current session" is automatically packaged as a message. Then, when the system aborts or the user logs off, that message is transmitted to us via the ARPA net for our perusal. This way, no matter which user on which TENEX is running one of our programs, we have a way of tracking is performance, discovering what kinds of deficiencies are being manifested and determining if some user is experiencing exceptional difficulties. By incorporating this simple idea into our initial SOPHIE system we were able to constantly upgrade the semantic grammar on the basis of the difficulties encountered by a wide and diverse class of users, using it in a realistic context (as opposed to a demo, which often invokes quite different requests from a user).

Paradigmatic systems

Our basic methodology (which we have found to be useful in developing a theory of intelligent instructional systems) has been to explore pieces of the overall theory in "paradigmatic" systems. A paradigmatic system is an easily modified proto­ type system constructed over a carefully cho~en domain of knowledge that enables us to exper1ment with some aspect of the overall system by simplify­ ing all the other aspects. For example, if we wanted to explore ways to infer a structural model of the stud·ent using an Articulate Expert, we would choose a domain of knowledge where the Expert would be trivial to build but where using it to infer a model of the student would be problematic. The information processing techniques and experience developed from the paradigmatic systems are then fed back into our more complete system. Examples of this methodology are our WEST [Burton and Brown, 1976], ABLOCKS, and BUGGY [Brown ef al 1977] systems which have enabled us to examine/develop new modelling techniques etc. that we now apply to our more serious and complex algebra and electronics instructional- systems.

On introducing users

One of the problems we encountered when we first started demonstrating SOPHIE! was that we would describe the capabilities of this system and then expect a user to sit down and be able to use it (not an unreasonable expectation, we though, since the system could handle any "reasonable" natural language query). But what we had failed to anticipate was the sudden "terror" many users experience when confronted with a major system that they have never used. When we realized how wide-spread this phenomenon was, we added a new module on top of the system that enabled the user to passiveiy watch some hypothetical expert using the system to do some troubleshooting. This computerized hypothetical "expert" also explained what he was doing as well as why and gradually would encourage the user to take a more active role by asking him questions and thereby leading him into using the underlying system directly. Although this was added in order to ease instructors into experimenting with our system, it has since been extended and extensively used for helping students feel comfortable with the system, as well as teaching them some expert problem solving strategies.