Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
EXPERT SYSTEMS: AEPITCAT~ONS IN SPACE Bruce G. Buchanan ~TE~rXJOl'lON Artificial intelligent- is one of the most important trends in computing becau ~ making cc mputers behave intelligently is at least as important as manipulating data efficiently. Opportunities for using intelligent programs in NASA space station environments are numerous and abvicus. But many of those opportunities require substantial research in artificial intelligence before they can be realized. This _ paper looks at the technology of artificial intelligence, especially expert systems, to define "from the inside Out" what capabilities exist that are relevant for applications and environments In the space station, and what research needs to be promoted in order to achieve systems better able to interact symbiotically with a variety of persons for long times in space. Anderson and Chambers (1985) mention a number of characteristics of systems in a human-centered space station. These include: symbiosis with humans: human and machine capabilities may complement one another autonomous, continuing operation for a period up to 20 years, operating in an information-rich environment, consequent ~ of interactions with humans not entirely predictable, maturation of system implies flexibility to accommodate opera;~ion21 growth and minor up graces, evolution of system implies flexibility to acoom=odate new and enhanced functionality, system may include electrons c crew belters (ECHO), humans may have to learn new skills to interact productively with computers, computers may learn from humans, autonomous agents may serve a variety of roles with varying degrees of decision making power and authority. These are some of the relevant considerations in a top-5cwn design of systems for the space station. Each of these points implies a 113
114 research and development program of some intensity. This paper takes a bottc m-up view of the same considerations--i.e., starts with what exists today and asks how we can achieve these design goals. By doing so, I hope to Introduce some relevant details into the design of Systems and the pi q of research. Expert systems are now being used in many decision-ma-ding situations of direct relevance to N~SA's mission, spanning manufacturing, engineering, medic me, and science. At present, they are used more as "intelligent assistants" than as replacements for technicians or experts. That is, they help people think through difficult problems and may provide suggestions about what to do, without taking over every aspect of the Cask. Cc mar programs that reason autoncmously are also of extreme importance in space, but they, too, must be integrated into an environment that is centered around people. They are extensions of present technology along several dimensions discussed here, that involve all of the same principles of design as the Intelligent assistant prc grams. One primary consideration is why intelligent systems are necessary in space. Although there are many reasons to build an expert system, they are all based on the premise: "Expertise is a scarce resource." The corollary (by Mhrphy's Law) is: "Even when there is enough expertise, it is never close enough to those who need it in a hurry." Because this is true--almost by definition of the term ' expertise '--constructing expert systems that reason at the level of N~SA's, or their contractors', specialists may have several benefits. These we summarized in Table 1. WHAT IS AN EXPERI SYSTEMn The general nature of expert systems is familiar to everyone within NASA. A reiteration of the four major characteristic is provided below to help define the most important dimensions for research and development efforts. An expert system is a computer program with expert-level problem solving abilities, which also fits some other criteria: it is a IABIE 1 Some Perceived Benefits of Expert Systems: Responses of 86 Users of Knowledge Engineering Tools 1. Replicate expertise 2. Preserve expertise 3. Increase productivity and cost savings 4. Free human experts for more demanding problems 5. Provide expert consultations to inexperience staff SOURCE: Bauman (1984)
115 symbolic reasoning program that uses heuristics, its reasoning and knowledge base are understandable, and--most importantly--it is flexible. These characteristics are discussed below. All are important for applications in the space station, and all define research topics that will enhance current capabilities. Example One well-known expert system that has become a classic, although not actively used, is MYCIN. It was developed at Stanford by E. H. Shortliffe and others ~ the m~d-1970s. Its task is two-fold: (a) diagnose the cause ts) of infection in a patient and (b) recommend appropriate drug therapy. From a medical perspective, MYCIN's knowledge base is now dated; from the perspective of expert systems it represents much of the kind of reasoning that is captured in today's systems. MYCIN's conclusions were demonstrated to be equal in quality to those of infectious disease specialists at Stanford Medical Center. The sample typescript shown in Appendix A illustrates MYCIN's requesting information about a case and reason mg to conclusions about the best treatment. Performance Naturally we want computer programs to solve problems without error. But that is not always possible--in fact, outside of mathematics and logic we don't have flawless methods we can put into programs. Specialists in engineering, science, education, the m~litary--and every ares outside of pure logic--must solve problems with less than perfect methods. How do they do it? Mostly by building up specialized knowledge through extra years of training and experience and by reasoning carefully with that knowledge in situations they have learned to recognize. They are not infallible, though. Specialists' decisions are challenged frequently--most noti~=hly in the courts. So it is also unreasonable to expect computer programs to reason infallibly in all of these areas. Occasionally new methods are disocvered that provide much better results than the established methods of the old practitioners. But these improvements can then be put into prc grams, thus raising the overall standard of performance while still keeping the same relative standard of comparison with the best specialists. Reasoning When we say that expert systems are reasoning--and not just calculating with numters--we are saying that they belong to a class of programs using the methods of artificial intelligence thereafter AI). In the 1940s, computers were used almost exclusively for large mathematical problems. At Los Alamos, for instance, scientists had to solve complex mathematical equations in order to calculate elements in the design of
116 the atomic bomb. These applications are usually referred to as large-scale scientific computation, or "number crunching" for short. In the 1950s, IBM and other computer manufacturers, realized the enormous value m helping b~-=iness solve problems of record keeping, payroll and the like. These applications extended the concept of comput~r-as-ca1culator to ccmputer-as-data-manager. In both of these classes of applications, the method of computation is error-free. There is no question that the result is correct, providing of course that the computer has been programmed correctly. A mathematical equation is solve] correctly; an employee roster is sorted correctly--if the Beth ~ are followed precisely. And computers are better able to follow complex instructions than people are. In computer science, logic an] mathematics we call these procedures algorithms. They a ~ procedures that can be guaranteed to provide a correct answer in a finite time, if there is one, and otherwise will provide a statement that the problem is not solvable. Some algorithms are too expensive to use, however, even in camputers. A classic example is finding the shortest route a Averting salesman can tat to visit many cities once and end up at home. With more than a hard of cities, aigorit~nic me~chc~s will not finish in time to be useful. For this reason, alternative methods have been developed. Art the mid l950s and early 1960s an alternative side of Sting ~ to be romanized as i~por~cant. Instead of always using aigori~ns, a Outer may case heuristics rules of thumb that aid In finding plausible answers quirky without guaranteeing the correctness of the results. Sc~t;~s these rules of throb are introduced into large primal simulations in order ~ get the situations ~ Card out anshrers more quickly. Or approximate methods may be sup tituted for more precise ones for the same reason. The assumptions may not all be correct; thus the results of the simulation may not be correct. When heuristic (non-algorithmic) methods are combined with symbolic (non-numeric) data, we are dealing with that part of computer science known as artificial intelligence. Understandability When someone truly knows something, he or she can "give an account" of what he knows. In our terms, good performance is not enough to call a person (or pro gram) an expert he/she (it) should also be able to explain why the solution is plausible, what features of the situation were noted to be important, what knowledge and problem solving methods were used. Otherwise we label a person ~~ "consistently (but unacocNnt~h~y) lucky", or maybe "psychic". Each field has its own standards of what a reasonable explanation is. A surgeon who recommends amputation of a leg generally talks about the process of disease or extent of injury and what will happen if it is not amputated. A broker So aWised liquidation of one's stock portfolio may explain the arrive with resent to technical Carts, historical trends, or sad ec~onc~nic principles that point to a stock market
117 collapse. In their awn c~nities, both ache surgeon and the broker can usually justif~r--in hurt if necessary--the advice they give. Arxl we rears them as experts party because they have he ~a~riedge ached lets them do this. Flexibility We e; ~ experts to be flexible In their Inlet ~ . Arx] we regard persons as amateurs, not experts, when we encounter opinions that are rigid, locked-in ways of dealing with problems, or an inability to deal with new situations. In particular, there are two situations in which we want expert systems to be flexible: At advicergiving tamp we want the program (or a person) to provide good advice about situations that have never been encountered before. Novices with good memories may be able to provide the "textbook" answers for classic situations. Experts however, should, in addition, be able to reason about novel situations. 2. At the time a program is being constructed or modified (or a person is rearm ng), we want it to be flexible enough to assimilate new bodies of information. There should be a capacity for growth of knowledge, not a rigidity that freezes either the depth or breadth of the program's knowledge. SC ME APPIICAIIONS Some of the types of problems for which expert systems have been constructed are shown in Table 2. Many of these, such as small troubleshooting assistance programs, are relatively straightforward. Although the state of the art is difficult to quantify, the programs In the Tahoe represent the kinds of commercially robust systems that can be built for NASA today, provided adequate resources and an app striate problem. We don't have an adequate taxonomy of problem types. Many of these overlap, ~ being different forms of data interpretation, for example. Even this brief characterization, however, provides a reasonably good idea of what expert systems can do. In general, expert systems can reduce costs or increase quality of goods and services--in a single phrase, they can increase productivity in an organization. If you believe either that there is not enough expertise in the world, or that it is not well distributed, then you will be willing to entertain the idea that putting human expertise into an -~ily-replicated form may answer some productivity problems. Or, at least expert systems may provide a partial answer. Consider medical
118 HE 2 Some Emblems for which Exam: Systems are ~vidi~ Solutions RISK ASSESSMENT E~I~ DAIS BEEN ADA En BEG It ~1']!;~'~ SO q~ ~Nr _~ . _ Ax . ~ : A; P~L10 ~ = ~U~ AURA S"1~;= CROP ~ # St. Paul Once Co.assess risk of insuring large camme~ial clients Genera ~tors~determi~ causes of vibration retie= and reccm=end repairs Digits Equipment Corp. translate customers' orders for computer sys ~ c into Hi Fin orders U. S. Army--design loading plan of cargo and equipment into aircraft of different types U. S. Air Force--plan an aircraft's route from base to target and back to avoid detection and threats when possible Schlumberger--~nterpret down-hole data from oil well bore holes to assist in prospecting Wasting house--plan manufacturing steps in a plant so as to avoid bottlenecks and delays Stanford Medical Center--~cist in managing malti-step chemotherapy for cancer patients IBM- monitor operations of MA operas mg sat U. S. Enr~ ion Agency--determine which requests for information fall under the exceptions to the Freedom of Information Act First Financial Planning Systems (Travelers Inch- analyze an indivi~sl's financial situation and ~ types of investments Hewlett Packard~iagnose causes of problems in potolith~ra~hy steps of wafer fabricatica, Virginia Polyphonic Institute - <Ed . . . . Assist in hegira apple orchards I~T DE;IGN Del~design Specials pose, law voltage electric ~ tars TRAINING Elf Aauitaine Oil Car --~-r~tr~te , SEVERE SUB ON& I, _ , reasoning to find cause of drill bit sticking in oil well and to correct the Problem Shell Oil Corporation advise persons on which subroutines In large FORTRAN library to use for their p Emblems and how to use then Lawrence Livenmore National laboratory specify parameter settings to bring a sensitive instant (triple quadruple mass Ureter into aligrm~t NOTE: day Free examples are listed In Buchanan (1986), and Harmon (1986) .
119 diagnosis. Specialists at university medical centers generally see more of the unusual disorders than a rural practitioner and thus stand a better chance of diagnosing them correctly. Putting some of that expertise more directly at the service of the rural practitioner could al low more effective treatment, and save patients the time and trouble of travel to the medical center. Or consider troubleshooting a complex piece of equipment. Persons with the most field experience are often the ones promoted to desk jobs in the central office. When subtle combinations of causes keep a less experienced field service technician from fixing a mechanic=] failure, someone with more expertise is needed. On earth, depending on travel times and =~he criticality of the work flow in the central office, calling the experienced specialist out may be a very expensive repair procedure. The following situations are all cases where it may make good sense to build an expert system: too few specialists for the number of problems; specialists not at the sites of problems when they occur; long training time for a specialist; high turnover among technicians; combination of complex equipment and poorly trained technicians; organization's best (or only) specialist in an area is nearing retirement; too many factors for a person to think through carefully in the time available. KEY CONCEPTS The four goals that characterize expert systems can be achieved with a few key methodological ideas. In this section, the key ideas will be introduced; in successive sections they will be elaborated on so as to explain a little how they work. The main organizational principle of expert systems is to keep specialized knowledge separate frog the logical and heuristic inference methods that use it. m is is -any to say but difficult to follow, for reasons that will be described later. Keep Dc main-Specific Knowledge Separate frog General Reasoning Methods -- FEY IDEA #1 -- Another key concept, which is imported from principled design of software generally, is modularity. (The first key idea is an instance of this, but that instance has taken on more importance than all the other instances of the general concept.) Modularity at the level of knowledge about the problem area implies conceptual separation of elements in the knowledge base. For example, medical knowledge about penicillin, although not totally independent, can often be separated from knowledge of other drugs. It can be codified in major ways, or deleted, without altering the program's knowledge of other drugs. So, this is to say that the concepts used to talk about objects ~ the Clan should be chosen so as to allow talking separately about an individual object, a single
120 properly of an object, or a single relation of one type of object with another. Modularity at the level of programming constructs implies that the program's 1ntern21 representation of knowledge elements (e.g., objects, properties, relations) is similarly "clean". Keep independent pieces of knowledge independent. Keep the rest as nPa'ly-independent as possible. --KEY IDES #2 -- A third key concept is uniformity of conceptualization and representation of knowledge. The underlying intuition is that it is easier for a person or a program to build, understand, and codify a body of knowledge if it doesn't mix and merge a variety of different types of things. This is as true at the knowledge level as at the programming level. For instance, one of the most compelling aspects of Newbonts laws is that all physical bodies ore treated as quantities with mass. He didn't need one set of laws for planets and another for apples. So it is desirable to build an expert system with a "conceptually clean", well-organized, simple collection of concepts. And it is important to use a simple, well-organized collection of programming constructs as well. Otherwise there are too many different kinds of things to keep track of and reason with. There is more dispute among AI specialists about this principle. There are good reasons to violate it, as we shall see, ~ the interest of being able to say more about the objects and relations of interest than can conveniently be said in a single language. We are frequently told by bi-lingual friends, for instance, that there are some concepts that just can't be expressed fully in English. m e same is true for programming constructs, but the basic principle for constructing expert systems is to try to ma Attain uniformity as much as Possible. Strive for uniformity of language and programming constructs KEY IDES #3 A fourth principle is to design the expert system to mirror the ways experts think about problems in their dc mains. That means using the same terms and the same rules of reasoning as the experts use. One reason for this is that building and debugging a knowledge base depends necessarily on the expert, and using 1=== fa ~ liar terminology or methods will introduce confusion and error before the knowledge base is completed. Also, aft or it is completed it needs to be comprehensible and unambiguous to the practitioners using the system or else confusion and error will result. Note that we are assuming here that the expert designing the system knows how to make it understandable to users. Great care must be taken when building a system, however, to insure that this assumption is true. There are times when this principle will be, or should be, violated. For example, when efficient computer algorithms can solve part of a problem, it doesn't often make good sense to use anything else for that part, even if the experts don't think about it in that way.
121 As huh as possible, he the sane vocabulary and methods In the program as the experts and practitioners use. -- KEY IDEA #4 -- These key ideas help us achieve all of our four goals in the following ways. · HENCE In problems whose solution methods are not airea~r well formalized, which are considerable, Huh of the effort In building a knowledge base frees an expert system lies in building the conceptual frank. ~i~hpr~ies and relations of objects to describe is often not well specified at the beginning. So the Repledge base is built Duncan - lay, Here experience with one kna~riedge base guides future modifications, extensions, or reformations. · RFASON~-~en the solution methods are not well characterize, it is important to encode heuristics that experts say they use. Storing these separately and In a simple form allows them to be changed easily. Since it is nearly impossible for an expert to articulate a complete and consistent set of heuristics at one sitting, it must be easy to add, remove, or modify the heuristics that determine the reasoning. · UNDERSTAND~BILPTY--with modularity, individual elements of the knowledge base can be displayed meaningfully in isolation. Moreover, with the separation of knowledge base and inference prcce~ures it is possible to peruse the knowledge base in order to find just those elements that were used to reason about a new case. And with uniformity of data structures, it is possible to build one set of procedures that produce explanations. · FLEX~BILITY--when the elements of the knowledge base are in separate data structures, and not intents med with code for inference procedures, we can add more knowledge with considerably more ease. When the indivi*u31 items in the knowledge base are nearly separate, we have fewer interactions to worry about when we change an items. And when the representation is homogeneous, we can more easily write other programs that act as "editing assistants" or explainers that help us insure correctness of new items and help us understand what is in the knowledge bit. ~RMANCE ISSUES Expert systems constitute one class of computer programs. As such, they work the same way as every other program: they process input data to produce output data. But the nature of the processing is different from most conventional programs. The key ideas motioned earlier are the key differences In the design am] irritation of expert systems. In order to design a rezoning program, we need to provide knowI~ge to radon with ark] reasoned methods to use. Both are needed. A
122 powerful thinker needs something to think about, and a body of facts without methods for using them is sterile. Over the last few decades, research in AI has elucidated programmed y methods for making inferences and storing knowledge. We briefly characterize these topics below, although with some reservations about cversimplifying, in order to highlight research issues relevant to increasing the performance of expert systems. In addition to research on inference methods and representation of knowledge, several other issues are mentioned briefly as needing more research in order to improve the performance of expert systems. Inf~ren~- Methods Aristotle's Theory of the syllogism defined a~r~ptahie inference methods cutside of mathematics for about 2000 years. His theory has been extended in this century by Russell & Whitehe,ad, and others, in a formal theory that includes methods of reason ng with several statements and several variables in an argument. Formal logic defines several inference rules which are guaranteed to create true conclusions if the premises of the argument are true. The chain rule (mcdus ponens) is the single most important inference rule in expert systems. It allows us to chain together a string of inferences= If A then B If B then C If C then D A Many of the inferences we make In our lives are not guaranteed by the rules of logic, however, nor do we have certain knowledge about the truth of our premises. Whenever we argue that the future will be like the past, as in stock market predictions, we have to be prepared for exceptions. These inferences, labeled "plausible inferences" by George Polya, are the ones of most interest In AT. One set of p~r~ methods were In AI for making plausible inferences is to assert the facts categorically--as if they wee hood to be true with certa~n~r and then reason about exceptions that might force revisions to he conclusion. Anther set of naptha= deals explicitly with the degree- of unc~ca~n~r In the facts and In he associations. MYC]N (see Premix A) uses this style of neasonir~. Usually the degrees of uncertainty implied by words like "often" and ' ~ y" are expressed as numbers. And often these numbers are interpreted as probabilities. A third, and most powerful, set of methods is to introduce heuristic rules, or rules of plausible inference, into the reasoning. These are facts or relationships that are not guaranteed to produce correct conclusions, but will often do so. Moreover, they often produce answers more quickly than their algorithmic counterparts. In the
123 traveling salesman problem, for example, the problem is to plan a route for visiting each cider in a set exactly once and ~ at the hcme cider. Ihis is an NP~nplete problem, that is, the algorithm for solving it Woes times that is exponential with ache number of cities. One heuristic we Tray ~n~uce is to go to the nearest city Scat has rot yet been visited. This certainly is up the computation of the route, but may (and probably will) miss the route that is shortest overall. Some mules of plausible inference used, with caution, In some exit systems are Shown below: Satisficing: If it will be expensive to find the very best solution to a problem, then stop with the first solution that satisfies easier criteria of being good enough. Inheritance: (Some specified) properties of a whole are shared by all its parts. E.g., An ice cube is cold and hard. Pieces of an ice cube are cold and hard. LBut other properties, like '~eight", do not behave the same.] Single Fault: If a piece of equipment (or any organized system) is malfunctioning, and one hypothesis explains the problem, then there prod ably is only a single cause of the problem. Compelling Evidence: If you have gathered a lot of evidence in favor of hypothesis Hi, and very little evidence against it, and you have gathered little positive evidence for alternative hypotheses, then HI is a plausible hypothesis. Dcccmposability: If there are many parts to a problem that are nearly independent, assume they can be solved independently. Then adjust the composed solution to take account of known interactions. Parsimony of Design: Designs or plans with fewer elements are preferred to those with more. In pr mciple, the rules of inference (both logical and plausible) may be applied again and again to a situation description, in any order, and the resulting conclusions will be the same. This is not always possible in practice, however. There may not be enough time to reason exhaustively about all possibilities and contingencies. For that reason AI researchers talk about controlling the inferences as being a more important, and more difficult, problem than making the inflect Controlling inferences breaks down into two subtasks: (a) deciding which rules to apply now, at this stage of the problem-solving process, and (b) deciding which part of the problem to work on now. Since we believe these sNbtasks require some intelligence, all of the principles for building knowledge-based systems also apply at this level of reason m g On participle, ~t is desirable to make this control knowledge explicit and separate from the inference methods
124 Representation of ~ Fledge We have said that a key idea in building expert systens is storing knowledge separately from the inference methods. Another key idea was to avoid, as mlldh as possible, representing it in a low-level computer language. But we have not maid how to represent for the computer what an expert wants to tell it. English is too difficult for a computer to interpret unambigacusly; FORTRAN and BASIC are too low-level for an expert to deal with efficiently. Clearly we need some stylized representations that are somewhere in between. AT researchers have developed several different representation methods. There is no single one that is best in every case--they each have strengtt.5 and weaknesses One of the fun~amen:t=] trade-offs in thinking about the representation of knowledge is between simplicity and expressive power. We want a simple set of conventions for storing knowledge because that mates it Air--for a person or a program--to understand what is in the knowledge base at any moment. It is also easier to write simple statements without error. Aristotelian logic ("All As are Bs", etc.) and arithmetic are simple representations. The difficulty is they lack the expressive power to let us say everything we think is important about a problem. A hundred years ago DeMorgan noted the lack of expressive power is Aristotelian logic (and a weakness In its inference methods): if you know that all horses are animals, he said, you cannot prove that the head of a horse is the head of an animal. This sort of problem led Russell ~ Whitehe~d to develop a formalism with more expressive power. There are two major classes of r-e presentation methods, reflecting two different ways of viewing the world: action-cent~red or object-centered. Different problem areas may focus on one or the other, or different experts in the same problem area may. For example, physic) ~ taLk about disease and classes of diseases as entities with expected prcperties and also talk about clinically relevant actions that determine bat to do~e.g., sacking questions, measuring things, relating signs and ~tms to possible causes, ma~hir~ lively causes to acceptable therapies=. Neither point of view is wrong, but they focus on medical ~henc~na quite differently. And an eat system wed similarly have one focus or the other. Action-centered representations focus on conclusions that can be drawn from facts or, more generally, on relations between situations and actions. The formalism of mathematical logic is one popular choice. Another popular formalism is rules. Objec*-centered representations focus on the organization of objects _ ~ ~~ ,.r_~l ~ Am_ _~~' A__ _~ ~ ~ ~_~ ~~ L[1 ~1= wur~u, LOLL- 110~]W 1llLU l~ll~~-^r~;lll~. They still allow conclusions to be drawn when an objec* Is found to have some properties, but those inferences are triggered fr=~`'~ithin" an object father than from outside. That means that objects and their properties- and changes to any of them--5rive the inferences. But in an action-centered model, the inference rules drive the creation of new objects and prc~perties. The net effec* may be identical, as we said, but the way one thinks about the domain of discourse is distinctly different.
125 Also, m cbject-centered representations there is more machinery for saving storage space by using hierarchies. Properties of classes of object, for example, may be implicitly inherited by all of the instances without having to store it with each instance. The manager of a group is the manager of each person in the grc up, so the program only needs to store (once for each group) the name of the group manager and can use that, ply= the c~ass-instance hierarchy, to find the name of any individual's manager. There are as many different conventions for representing knowledge as there are AI researchers working on this topic. This can be confusing when reading the literature. But they are basically all variations ray r~xtures--of the two different styles just discussed. There are many expert systems built out of these two sets of ideas, but considerably more experience and analysis ~ is necessary to understand their strengths and limitations. Validation and Robustness It is impossible to prove logically that the contents of an expert system's knowledge base are correct or complete or that the inference procedures will always provide the best answers. Yet persons in a space station whose equipment and lives depend on the expertise of many systems nope to know the scope and 1;mits of each system. Or, alternatively, they need tools for determining the scope and limits of the programs they use. These range from better explanation systems to tools for checking knowledge bases. Spatial and Temporal Reasoning Many compl ~ problems in a space station ~ e autonomous computer programs that represent and reason about three-dimensional objects. Simpler representations do not allow programs to solve problems involving 3-S shapes and positions, such as problems of fitting parts or of maintaining some equipment. Building expert systems requires attention to making the systems' reasoning understandable to persons onboard the space station and changeable by them. That, in turn' requires a flexible, high-level description language as well as computation21ly efficient operations that implement the language. Similarly, reasoning about sequences of ~nter-dependent actions and about situations that may change at arbitrary times are important aspects of problem solving in space. very Tinge Knowledge Bases To date expert systems have used knowledge bases of modest size. With the complexity of operations in space, we need to design and maintain expert systems with very large knowledge bases. Although size is difficult to define, most knowledge based motion only a few thousand different facts and relations. Probably the largest today is the
126 INTERNIST system in which about 250,000 facts are encoded (Miller et al., 1982). Some of this limit results from our own inability to keep in mind the interrelationships among more facts as much from the technology of storing and retrieving them. We must improve the technology to make it "aster to build and maintain knowledge bases of much larger scale, which will be necessary In a system as large and complex ~.c the space station. Shared Knowledge Roses To~ay's systems use single knowledge bases that have been built specially for them. As more and more systems are constructed, however, it will be important to use knowledge bases in different contexts and then reuse one system's knowledge base in another system. It is wasteful--an~ should not be ner~=sary--to duplicate the contents of an old knowledge base in a new application. One should expect, for example, programs in the space station that reason about the function of life support equipment and others that reason about the mechanical structure of the same equipment, both of which must share considerable detail about the equipment itself. Distributed Databases Databases exist now on ~ ny machines. Yet it is nearly impossible to treat several of them as if they were one logical unit--frum any program. Expert systems also need this capability. Current research will allow much broader sharing of data among different databases than is currently available in commercial systems. There will be many computers in the space station. It is much sounder to think of separate specialized databases (with appropriate backup) that can accessed frog various programs than to consider separate copies of every data base on every machine. Parallel Computation Computers are fast, but never fast enough. In addition to the immense speed-ups fsum improvements in the hardware, there are potential speed-ups from software. When a problem can be divided into nearby independent subproblems, it is conceptually easy to see that multiple computers could be used to solve the subproblems in parallel, thus saving considerable time. Work In the research laboratories indicates that this is feasible. Thus it will almost certainly become a commercial reality In the near future if it is cost-effective.
127 DEVEID~ENT AND ONCE ISSUES Building an em system requires firebird out how an expert solves a problen and translating ~t es~ise into a stylized form that can be read by Inputs. mis is no different ~ principle from building a cor~v~tional program In which pr~r~n~ find out what equations or aigoriff~ns e ~ s use arm there write FO ~ AN or COBOL ~ ~ as that embody those procedures. . . . . The main difference in practice is that expert systems must ~ncorpor ate knowledge that is much more qualitative and judgmental. In fact, much of the time the expert's "know-in ~ ' is not yet written down and what he/she does is regarded as an art. Because the e ~ Is knaw~edae is often not already codified an] ~ _ ~ · . . . · · . because writing symbolic reasoning programs is itself often regarded as an art, building an expert system requires patience. It generally works best as a team effort involving one or more e ~ and one or more so-called knowledge engineers. A knowledge engineer is a programmer of knowledge-based systems who understands the conventions of the computing framework and ho assists ache en An map pirate judgment=] knowledge into ~t framework. The dialogue Bergen exam and knowledge engineer is often haled "~awle~q~ emmeerinq". _ _ One of the key ideal En knowledge eng~n~ring is to focus on case studies. It is Rich easier for any of us to Scent s ~ one how we would approach a specific situation than to say in general terms how we solve problems of a type. Of course, if we have a set method (sometimes called a "canned prccedNre") that we always use, we can say that. "Oh yes, I always use the French variation of the AleRhine Gorbachev wave theory in situations like that", you might say. But then the knowledge engineer wants to know what do you do next and--more ~terestingly--when would you maXR exceptions to your set policy. And the best way for you to think about those things is to focus on cases. As long problem solving requires more than the application of set procedures, knowledge eng Peers will need to go through many cases, and variations on them, to help codify the expert's judgementn1 expertise. Steps Involved in Knowledge Engineering It mav take months or vears to build an expert systems with the time _ . , _ ~ ~ ~ ~ ~ ~ ~ 1 1 1 capering largely on One complexity 0£ one problem and fine extent co which expertise is already codified. One reason it takes so long is that there are many steps involved. And at each step, the knowledge engineer or the expert may decide it is necessary to undo some results of previous steps. Very roughly, the steps are thought of as beginning, middle and end phases in which attention is focused on different aspects of the system, as shown below: Beginning--define the problem precisely; understand which concepts are used, what their definitions and ~nter-relationships are.
128 Mi~le--implement a substantial prototype after choosing a set of representation conventions and writing a small but substantive knowledge base. End fill out the knowledge base to fix errors and extend the scope of the system's problem solving abilities, both of which are generally discovered by testing the systems on many test cases. ~ ols to Aid in the Construction of Expert Systems Just as carpenters can construct houses faster with the right tools, knowledge engineers can build expert systems faster with software tools that boost their productivity. m ese come in several forms. The main idea, however, is to provide programmers with mechanized intelligent assistants that know about programming conventions Concluding Abbreviations and shortcuts), that can help locate and fix errors, that can display the contents and interrelationships in a program or knowledge base, and so forth. These are the kinds of extra capabilities that distinguish system-building environments four prcgramaung languages. Sump of the more powerful environments sometimes called shells" are shown below. One characteristic of a shell is its commitment to a set of representation conventions of the sort outlined previously. S Table 3. learning At present, expert systems do not learn from experience. This is a defect that many research groups are working to remedy. Idly prot ~ s of learning systems promise some automated assistance in maintaining and extending a knowledge base through the experience of routine use, but these are not yet available. TAELE 3 Some ~ cially Available Shells for Building Expert Systems S.1 KEE Knowledge-Craft AM MOPS Personal Consultant M.1 ESDE Teknowledge ~1 limp Carnegie Group Inference Corp. Xerox Texas Instruments Tekrlowledge IBM
129 It is possible, however, to learn an initial set of notes fray a case library (collected past experienced ark use it for classification problems. Trxtuction propels are being used to band simple note sets for expert systems In which there is little Chaining of the notes ark little use of uncertain inferences. These are largely marketed in Great Britain where it is bethel under stood that even simple problems may carry great economic leverage. Current research is extending the scope of induction programs to more complex rule sets. Rem ~i~ The major cost involved in building an expert system is in personnel time. Shell systems now run on most common computers, so it is not necessary to buy new equipment and, most importantly, it is not necessary to build the complete set of programming tools found in a shell. Purchasing the shell and some training in how to use it are recommended. The amount of time needed from a teem of experts an] knowledge eng beers is variable--as are their salaries. Table 4 gives some estimates for a hypo~heti~1 small system constructed within an existing shell. It is assumed here that a problem has been precisely defined before beginning, that a case library of at least a half dozen typical and hard cases has been assembled, that a commercial shell has been purchased and runs on an available computer, and that the senior knowledge engineer is very familiar with both the shell and the computer. It is also assumed that the team's primary responsibility is this activity, and that they have the blessing of their management. In this simple model, the senior knowledge engineer also fills the role of project leader, with as much as half his/her time filled with reports, briefings, budgeting, and other managerial responsibilities. The junior knowledge engineer in this model is responsible for software TABLE 4 Percent Time Per Quarter Q1 Q2 Q3 Q4 expert 75 75 100 75 sr.KE 100 100 100 100 jr.KE 100 100 100 100 NCIE: Approximate percentage of time required from an expert, a senior knowledge engineer, and a junior knowledge engineer to build a hypothetical small system over four quarters of a yew. The two main variables in determining the amount of time required are the nature of the problem and the definition of the deliverable.
130 eng~neering--that is, integration of the expert system into the run-time environment--as well as for help in building the knowledge base. And the expert, here, is (atypically)) also filling the role of ''management champion" with some time devoted to securing resources to make the project happen. One of the main factors that determines the length of time a project will take is, not surprisingly, the nature of the problem. This includes both the scope of the problem and the extent to which a ccmmercially available shell is appropriate for the problem. Another main factor is the definition of the "deliverable", that is the terms of the contractual agreement specifying whether the product delivered is a prototype or is a smoothly polished software package. There are added gains in building an expert system that offset some of the costs just mentioned. Besides the obvious gains showing up in work performed, there are very noticeable ga ems in the quality of information available. Shortening the time required to build systems and increasing our ability to maintain them are thus two of the central issues for putting expert systems in the space station. ENVIRONMENTAL ISSUES The environments in which expert systems currently operate are closely constrained. While there is wide variation in the degree of autonomy exhibited across all working systems, most systems in place are interactive, requiring intelligent input from humans. The predominant model of interaction is a consultation model in which an expert systems asks a person for the facts (and interpretations of them) and then pro via-= some advice. A consultation with MYCIN about a medical ~~= is shown in the Appendix. m ere are several reasons why the consultation model is appealing, each of which constitutes an opportunity for research. In the first place, a program that asks short-answer questions of a person can finesse the very large problem of understanding free-form English sentences and phrases. m e program knows what answers are reasonable in the current context and can have advance expectations about the ways these answers may be frame. S~rx3, tile consultation Ides provides a strong sense of context which not only helps Me program undershot a person's answers, but helps Me person undershot the sense of the questions. This is i:ppor~can~becauseTn~s~n~pretationof theprogram's questions can have SerlOUS consequence. Third, in a cor;sul~cation it is reasonable to make strong assumptions about the users of an expert system--what they know, what they don't know, what vocabulary they use, what environment they are working in' and so forth. This helps minimize problems in communication. This means also that so-called "common sense" knowledge may be supplied by users and need not all be supplied by the program.
131 Real Tim Monitoring As Herb systems beeline faster, it will be Tier to build systems ached monitor other devices or processes with rapid change. Conceptually a difficult problem is managing time~ependlent relations efficiently, which is one of the necessary ~ onents of a monitoring system. The large amounts of data received and the speed with which they are received are also critical issues. Integrating AI methods of reasoning about the data with numerical methods for digitizing and filtering is essential. Richer Ihput/Output No one likes to interact with computers by typing. Considerable work on interactive graphics has reduced the need for typing. But it will be even easier when we can communicate with programs by giving voice commands and receiving spoken English output in return. Models of Users and Situations No single style of interaction is best for all users at all times. Specialists do not need explanations of the meanings of terms, for example, while less experienced users used considerable help understanding the context of the problem. Also, the criticality of the situation may demand taking shortcuts in data acquisition or reasoning to reduce the risk immediately before taking a more systematic, detailed look at the problem. Expert systems must be sensitive to models of both the user and the situation in order to request appropriate input, reason at an appropriate level of detail, and present conclusions and suggestions in an appropriate way. CONCLUDING OBSERVATIONS Expert systems already are saving organizations millions of dollars and performing tasks routinely that ordinarily require human expertise. The number of applications of to~ay's technology is nearly boundless--consider, for example, the number of pieces of equipment in a space station that we don't readily know how to fix. The first commercial shells on the market are robust enough to be used effectively. Integrating ~nt=1ligent systems with conventional computer programs and with persons ~ the space station involve= new research in many dimensions. The single biggest advantage of AI programs, amply demonstrate in expert systems, is their flexibility. This matches precisely the single biggest design requirement on software in the space station. What we see now is just the beginning of ~ wave of intelligent software that can have as great an effect as business data prccessing software. It is impossible in any area of technology to make accurate
132 predictions. However, there are many parallels between the growth of expert systems and or computing hardware, with about a 25-30 year lag. When electronic computers became available commercially, businessmen began to ask about applications that would make a difference to them. In 1955, several of these innovators assembled at Harvard to discuss they e~ien~-C. She of the conclusions they Mew falcon their party experience are summarized belay (Sheehan, 1955~: . 1. "The initial c~v~n~chusiasm, which inevitably Zanies a project of this scope, can and does make the job harder. Too many people had the impression that this was the answer to all Problems. Plans it is. but we haven't bed sprat enough to . . develop all of them... , 2. "Some of cur original thinking has been partly confirmed in that the greatest benefits to be derived from a computer will probably consist of information impossible to obtain previously... 3. "Our experience has shown that the compute' is more adaptable to some projects than others... 4. "Programmers should be recruited within your own company...It is easier to teach men the required computer and program techniques than to acquaint them properly with the complex pracedurcs and routines of mcdern-day industry... 5. "I doubt if it is possible to overemphasize the desirability of providing for convenient corrections or deletion of errors in data. . . 6. 'the maximum justifiable amount of flexibility for extending or integrating applications must be included in the initial programming..." The parallels with expert systems are clear: we have just begun to explore the first set of applications an] have only a limited view of possibilities in the future. Over the next 5-10 years, however, we do know what extensions in capabilities will be possible because they are the ones that are active research areas in the present. Future developments will make the technology more broadly applicable and easier to use. Every professional activi~y--from medicine to manufacturing--contains many problem solving tasks that are not mathematical. The develcpment of artificial intelligence methods that can solve many of these kinds of problems will change the nature of applications of computers in cur society as much as the develcpment of transistors changed the a architecture and applicability of computers themselves.
133 Appendix: Annotates! Typescript of a Sample MYCIN Consultation Instructions for Users at the Start of an A. 1 Infectious Disease Consultation (Optional) [Here, as elsewhere. tore responses hyped by the user follow a double-asterisk prompt. MYCIN 3-Jun-78 . . . Special options (type ? for help): ·. Instructions? (y or N) YES This is a computer program named MYCIN that has been designed to advise you regarding an appropriate choice of infectious disease therapy. it is understood that you have a patient who may have an infection. Please answer the following questions, terminating each response with RETURN. To correct typing errors. use the DELETE key to delete single characters, ~ctrl>W to delete a word, and <ctrl>Q to delete the whole line. If you are not certain of your answer, you may modify the response by inserting a certainty factor (a number from 1 to 10) in parentheses after your response. Absolute certainty (10) is assumed for every unmodified answer It is likely that some of the following questions cannot be answered with certainty. You may change an answer to a previous question in two ways. If the program is waiting for a response from you (that is, has typed ), enter CHANGE followed by the number(s) of the question(s) whose answers will be altered. You may also change a previous answer at any time (even when the program is not waiting for a response from you) by typing cctr1>F (Fix), which will cause the program to interrupt its computation and ask what you want to change. (If the response to <ctrl>F is not immediate. try typing the RETURN key in addition.) Try to avoid going back because the process requires reconsidering the patient from the beginning and therefore may be slow. Note that you may also enter UNK (for UNKnown) if you do not know the answer to a question, ? if you wish to see a more precise definition of the question or some examples of recognized responses. ?? if you want to see all recognized responses, the word RULE if you would like to see the decision rule which has generated the question being asked. the word WHY if you would like to see a more detailed explanation of the question. or the letters QA if you would like to interrupt the consultation in order to ask questions regarding the decisions made so far In the consultation. If you are ever puzzled about what options are available to you during a consultation. enter the word HELP and a list of options will be listed for you. Sample Response [user input follows the 1 Does the patient have a risk factor for tuberculosis?
134 Appendix: Annotated Typescript of a Sample MYCIN Consultation One or more of the following are considered risk factors for TB: a) positive PPD (5TU). by history of close contact with a person having active TB. c) household member with a past history of active TB. d) chest X-rav showing apical scarring. e) granulomas seen on biopsy of any organ tissue. expected responses are: YES NO Ente! HELP for user options. ·- YES SUMMARY: (type ctrl-O to abort printout) UNK - Answer not known. ? - Rephrases the question and gives examples of recognized responses. ?9 - Prints a list of all recognized responses. Prints the current decision rule. Program enters question-answering mode. Go back and re-request answer to question number #. Enter comments about any aspect of the system s performance. Your comments will be forwarded to those in charge of the MYCIN program. Gives high-level explanation of the current reasoning chain that provoked this question Explains HOW the system wii achieve a goal referred to by number # in a previous explanation. Provides a more detailed explanation of a previous answer given by a WHY command Resets the explanation of the reasoning chain back to lowest level. as if you never used the WHY EXPLAIN commands. Halts the program without completing the consultation. Halts the program saving the current patient on a disk file retrievable at a later date. Prints this list. RULE - QA - CHANGE # COMM ENT WHY - EXPLAIN FORGET - STOP - SAVE HELP A.2 Infectious Disease Consultation Session [This is a typescript of a MYCIN consu/tabon on an actua/ patient treated at Stanford MYCIN was no' used to suggest treatment. Spaces have been added for readability. and comments (in italics) have been inserted Otherwise it is a verbatim transcript. Users responses follow double asterisks ----PATIENT-538-------- 1 ) Pat~ent s name: 2' l ., PT538 Age: 34 YEARS 3) Sex: MAEL = MALE 4) Race: WHY [i.e. WHY is it important to determine the race of PATIENT-538?] [Name age. sex. and race are always asked because they are INlTlALDA`-A j [MYCIN corrects misspellings [The user asks for an explanation In this Instance it is trivia! [1.0] The race of a patient is one of the 5 standard parameters that identify a patient. This may be useful later in the consultation. For example. it Is relevant In determining the organisms (other than those seen on cultures or smears) which might be causing the Infection. [back to question 4...l CAUCASIAN Please supply one answer under each heading. If your response Is not fans is the answer to the quest/on about race
135 Appendix: Annotated Typescript of a Sample MYCIN Consultation 16) Please give the date on which clinical evidence (symptoms. signs. or laboratory tests) of the meningitis first appeared. 29-Jan-77 22:15 17j Does Pt538 have an abnormal chest X-ray? NO 18) Does Pt538 have one or more of the following risk factors for TB: a) positive intermediate strength PPD b) history of close contact with a person with active TB. c) a household member with a past history of active TB d) apical scarring on chest X-ray. e) history of granulomas NO 19) Does Pt538 have any cutaneous lesions or rash on physical examination? NO 20) Does Pt538 have evidence of ocular nerve dysfunction? NO 21) is Pt538 a compromised host (e.g. alcoholic. sickle-cell-disease. immunosuppressed)? YES 22) Is Pt538 ~mmunosuppressed (e 9 corticosteroid therapy. cytotoxic drug therapy. radiation therapy. leukemia. Iymphoma)? N<:) 23) For how many days has PtS38 had abnormal necrologic signs? On biopsy of liver lymph node. or other organs? 7.2 HOURS 24) For how many days has Pt538 had necrologic symptoms? ~ DAYS 25) Has PtS38 had recent evidence of tender or enlarged salivary glands? '' NC 26) Has Pt538 been exposed to any contagious disease recently (e.g. men~ngococcal disease. mumps) ? NO iNote that the answer was given in hours. MYCIN has to be flexible enough to convert un~ts.J Please enter CSf findings In the following table CSF Percent Protein CSF Peripheral WBC PMN Glucose Glucose 27) 100 99 280 5 UNKNOWN 1Note that some information is missJng.j 28) Do you know the results of a recent CBC of PtS38? ·. YES Please enter CE3C results in the following table WBC Percent Percent In 1000s PMN BANDS 29) 1.9 UNKNOWN UNKNOWN 30) Has Pt538 had any InJury or insult to. or defect in the central- nervous-system (excluding neurosurgery)? NO 31) Has PtS38 ever undergone any type of surgery? - YES 32) Does PtS38 have a history of NEUROSURGERY? ' M0 33) Does PtS38 I've in a crowded environments NO 34! Do you suspect recent alcoholic history In Pt538? NO 35) is meningitis a hospital-acqu~red Infection? WHY [This is a more interesting question. requ/rmg presentation of the rule that caused question 35 to be asked. Note that the first five clauses of the rule have been ~nvesngated and question 35 IS motivated by the last clause. MYCIN first rephrases the question tO give the user its ,nterpretanon of the WHY|
36 Infectious Disease Consultation Session (i e. WHY is it important to determine whether the Infection was acquired while the patient was hospitalized?] [2.01 This will aid in determining the organisms (other than those seen on cultures or smears! which rn~gnt be causing the Infection It has already been established that Author: Yu [RULE5453 [back to question 35. HOW 2.3 [2 1 J the infection which requires therapy Is meningitis and [2 21 organisms were not seen on the stain of this pending CSF culture. and [2 3] the type of the infection is bacterial. and [2.41 neurosurgery is not one of the relevant items from the history of Pt538. and [2.5] the age of Pt538 is greater than 10 days Therefore. if [2.6] the infection was acquired while the patient was hospitalized then there is evidence that the organisms (other than those seen on cultures or smears) which might be causing the Infention is e cold ( 75) staphylococcus-coag-pos ( 3) pseudomonas-aerug!nosa ( 3) klebsiella-pneumoniae (.5) {MYCIN now requests the answer to question 35 again. Instead of answering. the user asks how clause 2.3 was confirmed.] [i.e. HOW was it established that the type of the infection is bacterial?] [Ramer than printing the contents of the rules. MYCIN shows the rule numbers and the strength of positive and negative evidence from each rule.| The following rules concluded about: the likelihood that the type of the infection is bacterial Rule Yes No [2.71 RULE148 (.19) [2.8] RULE500 (.51) [2.91 RULE501 (.95) [3.0] RULE502 (.97) [3.1] RULE526 ( 98) (3.2] RULE504 (.97) [3.31 RULE524 ( 98) Thus it has been established that the type of the infection is bacterial. back to question 35... HOW 2.5 [i.e. HOW was it established that the age of pt538 is greater than 10 days? In answer to question 2 you said that the age of pt538 is 34 years Thus it has been established that the age of pt538 is greater than 10 days. [back to question 35. . .] ~~ y 36) Is Pt538 s clinical history consistent with EPIGLOTTITiS? NO 37) Is Pt538s clinical history consistent with OTITIS-MEDIA? ~ NO 38) Has Pt538 ever undergone splenectomy? NO 39) Is Pt538 a burn patient? WHY [Another explanation is reglJested.
137 Appendix: Annotated Typescript of ~ Sample MYCIN Consultation [i.e. WHY is it important to determine whether PATIENT-538 has been seriously bumed?l 4.01 This will aid in determining the organisms (other than those seen on cultures or smears) which might be causing the infection. It has already been established that [4.1] the infection which requires therapy is meningitis and [4.21 organisms were not seen on the stain of this pending CSF culture. and [4.3] the type of the infection is bacterial Therefore. if then 14.41 pt538 has been seriously burned there is suggestive evidence (.5) that pseudomonas-aeruginosa is one of the organisms (other than those seen on cultures or smears) which might be causing the infection Author: Yu [RULE578] [back to question 39. . .] WHY [In this case the user asks another WHY question. backing up the subgoal tree.J [i.e. WHY is it important to determine the organisms (other than those seen on cultures or smears) which might be causing the infection?] [5.0] ... in order to determine the therapeutic regimen of PATIENT-538. If [5.1 ] information has been gathered about organisms isolated from the patient organisms noted on smears of the patient negative cultures of the patient. suspected infections without microbiological evidence. current drugs of the patient and prior drugs of the patient and 5.2] an attempt has been made to deduce the organisms which require therapy and [5.3] You have given consideration to organisms (other than those seen in cultures or smears) which might be present and found at least one organism which requires therapy. then determine the best therapy recommendation from among the drugs likely to be effective against the significant organisms tRULE092] [back to question 39. . .] ~~ y 1MYCIN now prints its intermediate conclusions about the likely causes of the infection.] I have found evidence (RULE545 RULE557) that treatment should cover for E.coli causing meningitis. I have found evidence (RULE545 RULE557 RULE578) that treatment should cover for Pseudomonas-aerug~nosa causing meningitis. I have found evidence (RULES45 RULE557) that treatment should cover for Klebsielia-pneumoniae causing meningitis. I have found evidence (RULE507) that treatment should cover for Diplococcus- pneumoniae causing meningitis. [Now MYCIN begins to ask questions related to appropriate theraPyJ 40) Does PtS38 have focal neurological signs? UNKNOWN Some questions are motivated by the organism under consideration. Thus the following lines sometimes provide useful contextua/ information to a users (Determining which drugs are desirable for use against the E.coli. . .] [Determining which drugs are desirable for use against the Pseudomonas- aerug~nosa. .
138 [Determining which drugs are desirable for use against the F<lebs~ella- pneumoniae 1 [~)eterm~n~ng which drugs are desirable for use against the Diplococcus- pneumoniae l Infectious Disease Consultation Session My therapy recommendations will be designed to treat for organisms that either are very likely or although less likely would have a significant effect on therapy selection if they were present. It Is important to cover for the following probable infection(s) and associated organisms NFECTION-1 Is MENINGITIS ITEM-1 ~ E.COLI [from clinical evidence only] ·' ITEM-2 PSEUDOMONAS-AERUGINOSA [from clinical evidence onIy ITEM-3> KLEBSIELLA-PNEUMONIAE [from clinical evidence only ·- ITEM-4:> DIPLOCOCCUS-PNEUMONIAE [from clinical evidence only 41) Does PtS38 have a clinically significant allergic reaction to any antimicrobial agent? NO [Considering AMPICILLIN-AND-GENTAMICIN for use against INFECTION-~...J 42) Do you have reason to suspect that PtS38 may have impaired renal function? NO 43) Pt538s weight in kilograms (or <number> POUNDS): 70 Finally, MYCIN prints the best therapy it has found to maximize the chances of effective coverage of the caused of the infection while minimizing the number of drugs, their toxicity, and side effects! REC-l] My preferred therapy recommendation is as follows: In order to cover for Items _1 2 3 4>: Give the following in combination: 1 ) AMPICILLIN Dose: 3.59 (28.0 ml) q4h IV [calculated on basis of 50 mg kgJ 2) GENTAMICIN Dose: 119 mg (3.0 ml 80mg!2ml ampule) q8h IV [calculated on basis of 1.7 mgJkg plus consider giving 5 mg q24h Intrathecal Comments: Monitor serum concentrations Since high concentrations of penicillins can inactivate aminoglycosides do not mix these two antibiotics in the same IV bottle. Do you wish to see the next choice therapy? YES Which. if any of tt" above drugs do you wish eliminated from consideration? AMPICILLIN [Considering GENTAMICIN for use against INFECTION-t...] [Conskbring PENICILLIN for use against INFECTION-1...1 [REC-21 Next best therapy recommendation: In order to cover for Items <1 2 3~: Give: GENTAMICIN Dose: 119 mg (3.0 ml 80mg/2ml ampule) q8h IV [calculated on basis of 1.7 m~kgl plus consider giving 5 mg q24h Intrathecal Comments: Monitor serum concentrations In order to cover Or Item <4>: Give: PENICILLIN Dose: 3.5 million units (7.0 ml) q4h IV [calculated on basis of 50000 unitsJks, tOther therapies can be examined 1
139 understood. you will be prompted by the standard questioning method. (enter ? for normal prompting, for questions that are not applicable ) Please enter the results of positive cultures using the following table. site culture date dated number collected positive 5) DONE Please enter the results of pending cultures in the following table. site culture date stain number collected examined 6) ' CSF 1234 30-Jan-77 YES Infectious Disease Consultation Session method of collection [urine,sputum only [The next question asks for severe/ items in tabular form since they are found together in the hospital record. method ofl collection [urine,sputum only| stain of morph! fMYClN follows up on the stain of smear the smear and the morphology.! [type DONE if no organism was seen on the smear 7)--DONE [No information about stain or morphology is available. MYCIN now follows up question 6, asking about other positive cuJtures. site cul # date coil. smear coil. ~nothodl examined?! 8)''DONE [There are no more positive cu/tures.1 Please enter negative cultures in the following table. site culture dated method of number collected collection [urine,sputum only] 9)--DONE 10) Do you suspect Pt538 may have an infection at a site from which you have not obtained culture specimens? NO Please enter information on current antimicrobials. drug route of dater name administration started 11). CEPHALOTHIN IV 23-Jan-77 1 2) 'DONE Please enter information on prior antimicrobials. drug route of name 1 3)--DONE dater administration started Considering organisms (other than those isolated from cultures) which might be print. . .] 14) Has Pt538 recently had symptoms of persistent headache or other abnormal necrologic symptoms (dizziness, lethargy, etc.)? YES 15) Has Pt538 recently had objective evidence of abnormal neurologic signs (nuchal rigidity, coma, seizures, etc.) documented by physician observation or examination? ~ YES The CSF cultures will be considered to be associated with MENINGITIS. [There are no negative cuitures.J dater stopped
140 Appendix: Annotated Typescript of a Sample MYCIN Consultation RECOMMENDATION-2 is equivalent to flECOMMENDATION.1. Would you like to see details of this comparison? YES Looking at the coverage of the 'most likely' items, we see that: RECOMMENDATION-1 consists of 1 first choice drug (AMPICILLIN for ITEM-1 and ITEM - ) and 1 third choice drug (GENTAMICIN for ITEM-1, ITEM-2 and ITEM-3) RECOMMENDATION-2 consists of 1 first choice drug (PENICILLIN for ITEM-4) and 1 third choice drug (GENTAMICIN for ITEM-2. ITEM-1 and ITEM-3) ORGANISMS RECOMMENDATION-1 RECOMMENDATION-2 most likely' Drug -- Choice Ranking Drug -- Choice Ranking ITEM-1 AMPICILLIN-AND-GENTAMICIN--1 st GENTAMICIN--3rd ITEM-2 GENTAMICIN--2nd GENTAMICIN--2nd ITEM-3 GENTAMICIN--3rd GENTAMICIN--3rd ITEM-4 AMPICILLIN--1st PENICILLIN--1st (The desirability of a drug is defined to be its lowest ranking for the items it covers.) Do you wish to see the next choice therapy? ' NO
141 RENGF~ Arxierson, J. L. and A. Cl~. 1985 ~nan~nter~ Space Station Design. Paper presented at Conference on Cc - uters ~ Aerospace, Lord Beadh, CA, October 21-23. Bagman, B. 1984 MOP Thesis, University of Texas, Or. Buchanan, B. G. 1986 Expert systems: working systems and the Church literature. Expert Systems 3 (1) :32-51. Also appear as S`canford REL Report 85-37 and SON CS-85-1075. Han, Paul. 1986 ~ventor)~ and analysis of existing expert system. Expert S:rst~; Strategies 2 (8): 1-16. Miller, R. A., Pople, H. E. and M\rers, J. D. 1982 ~NTERNTST-1, Art Experiments ~ter-Based Diagnostic Consultant for General Thternal Mectic~ne. N=M 307(8) :468-476. peso ahead; in 'Readings In Medical Artificial Intelligence: The First Decade', Clingy. W. J and Shortliffe, E. H., Ad., Wading, Ha, A~ison-Wesley, Pp. 190-209. Sheehan, G. M 1955 An application to payroll. Data processing Conference. Cambridge, He, September. . Feedings of the Autistic Heard University ess,
