turban online chapter w18

W-174

KnowledgeAcquisition,Representation,and Reasoning

Learning Objectives◆ Understand the nature of knowledge

◆ Understand the knowledge-engineering process

◆ Learn different approaches to knowledge acquisition

◆ Explain the pros and cons of different knowledge acquisition approaches

◆ Illustrate methods for knowledge verification and validation

◆ Understand inference strategies in rule-based intelligent systems

◆ Explain uncertainties and uncertainty processing in expert systems (ES)

In this book, we have described the concepts and structure of knowledge-based ES.For such systems to be useful, it is critical to have powerful knowledge in the knowl-

edge base and a good inference engine that can derive valid conclusions from theknowledge base. In this chapter, we describe the process for acquiring knowledge fromhuman experts, validating and representing the knowledge, creating and operatinginference mechanisms, and dealing with uncertainty.The chapter is divided into the fol-lowing sections:

W18.1 Opening Vignette: Development of a Real-Time Knowledge-Based System at Eli Lilly

W18.2 Concepts of Knowledge Engineering

W18.3 The Scope and Types of Knowledge

W18.4 Methods of Acquiring Knowledge from Experts

W18.5 Acquiring Knowledge from Multiple Experts

W18.6 Automated Knowledge Acquisition from Data and Documents

W18.7 Knowledge Verification and Validation

W18.8 Representation of Knowledge

W18.9 Reasoning in Intelligent Systems

CHAPTER

18

TURBMW18_0131986600.QXD 11/9/06 8:58 PM Page W-174

Opening Vignette W-175 ◆

W18.10 Explanation and Metaknowledge

W18.11 Inferencing with Uncertainty

W18.12 Expert Systems Development

W18.13 Knowledge Acquisition and the Internet

W18.1 OPENING VIGNETTE: DEVELOPMENT OF A REAL-TIME KNOWLEDGE-BASED SYSTEM AT ELI LILLY

PROBLEM

Eli Lilly (lilly.com) is a large U.S.-based, global pharmaceutical manufacturing companythat has 42,600 employees worldwide and markets it products in about 160 countries.

Production of medicines requires a special process called fermentation. A typicalfermentation process involves a series of stirred vessels in which a culture of microor-ganisms is grown and transferred to progressively larger tanks. In order to have qualityproducts, the fermentation process must be monitored carefully and controlled consis-tently. Unfortunately, some key quality parameters are difficult to control using tradi-tional statistical process control. For example, it is difficult to quantify the state of a fer-mentation seed, but without this information, predictions of the behavior of aproduction cycle may be imprecise. Furthermore, Lilly operates many plants that usethe same process but are operated by different employees. The quality of the productmay vary substantially when different operators use their experience to control theprocess. The turnover of experienced operators also causes a problem for some plants.As a result, both productivity and quality problems emerge.

SOLUTION

Eli Lilly is using ES to overcome its problems. Its motivation for adopting a new sys-tem was to make the expertise of key technical employees available to the processoperators 24 hours a day, anywhere. It was felt that an ES would allow relevant por-tions of the knowledge base to be cloned and made available to all of the company’splants. Eli Lilly installed Gensym’s G2 (gensym.com/?p=g2_platform) at its Speke siteto construct an intelligent quality-alert system that would provide consistent real-timeadvice to the operators.

DEVELOPMENT PROCESS

The Eli Lilly system focuses on the seed stage of the fermentation process. Four knowl-edge engineers were involved in the system development: three for knowledge elicitationand one for coding knowledge into G2 rules and procedures. The knowledge engineershad some domain knowledge, and they were asked to record the knowledge of theexperts but not to preempt or optimize it with personal attitudes toward the domain—and in particular not to rationalize the knowledge into a framework with which the inter-viewer was familiar.They were also asked not to intimidate the experts by virtue of theirdomain expertise. The entire development process took six months to complete, and itincluded the following major steps:

1. Knowledge elicitation. The knowledge engineers interviewed 10 experiencedexperts to acquire their knowledge. A knowledge acquisition tool, KAT, was usedto facilitate the knowledge acquisition.


◆ W-176 CHAPTER 18 Knowledge Acquisition, Representation, and Reasoning

2. Knowledge fusion. The interviews resulted in 10 different knowledge sets, repre-sented as graphs. A joint session integrated all sets.

3. Coding of the knowledge base. The resulting knowledge graph was converted intorules acceptable to G2. Initially, a total of 60 rules were produced.

4. Testing and evaluation. The system was tested using values that simulated anom-alies for the variables and output verification.

ResultsConfiguring the scalable automation was more intuitive and faster even when takingthe learning curve into consideration. The program executes exactly as drawn makingvisual proof of calculations quick and easy. Two routines can be activated during simu-lation or start-up to find which gives the best response time. Such comparisons are verydifficult with older automation.

OLE for process control (OPC) standard augments communications between flex-ible production areas. It provides data exchange standards and constructs that allowsinformation sharing among programs, much like Microsoft Office. OPC also providesmore robust, secure, and object-oriented code and greater error checking.

Sources: Adapted from A. Ranjan, J. Glassey, G. Montague, and P. Mohan, “From Process Experts toa Real-Time Knowledge-Based System,” Expert Systems, Vol. 19, No. 2, 2002, pp. 69–79; and A. Bullock, Life Sciences/Eli Lilly—Success Story, February 1988, easydeltav.com/successstories/lilly.asp (accessed April 2006).

Questions for the Opening Vignette

1. Why did Eli Lilly need to develop an intelligent system for providing advice toprocess operators?

2. Why were the coding knowledge engineers asked not to use their own knowledgeframework but to record only knowledge from the experts?

3. Why do we need knowledge engineers to develop ES? Can we ask experts todevelop the system by themselves?

4. What do you think of Eli Lilly’s approach, which developed 10 separate knowl-edge bases and then put them together through knowledge fusion? What are thepros and cons of this approach?

5. For system testing and evaluation, Eli Lilly used simulated data rather than realdata. Why is this acceptable for real-time ES? Can you think of a better approachfor system testing and evaluation?

6. What are the benefits of using knowledge acquisition tools?

WHAT WE CAN LEARN FROM THIS VIGNETTE

The opening vignette illustrates the process of acquiring knowledge and deploying it ina real-time monitoring system. It also illustrates the use of computer-aided knowledgeacquisition tools to make the development job easier and to integrate knowledge frommultiple experts.

W18.2 CONCEPTS OF KNOWLEDGE ENGINEERING

The process of acquiring knowledge from experts and building a knowledge base iscalled knowledge engineering. The activity of knowledge engineering is defined inthe pioneering work of Feigenbaum and McCorduck (1983) as the art of bringing the


Concepts of Knowledge Engineering W-177 ◆

principles and tools of artificial intelligence research to bear on difficult applicationsproblems requiring the knowledge of experts for their solutions. Knowledge engineer-ing involves the cooperation of human experts in the domain working with the knowl-edge engineer to codify and make explicit the rules (or other procedures) that a humanexpert uses to solve real problems. The field is related to software engineering.

Knowledge engineering can be viewed from two perspectives: narrow and broad.According to the narrow perspective, knowledge engineering deals with knowledgeacquisition, representation, validation, inferencing, explanation, and maintenance.Alternatively; according to the broad perspective, the term describes the entire processof developing and maintaining intelligent systems. In this book, we use the narrowdefinition.

The knowledge possessed by human experts is often unstructured and not explic-itly expressed. A major goal of knowledge engineering is to help experts articulatewhat they know and document the knowledge in a reusable form. For details, seeen.wikipedia.org/wiki/Knowledge_engineering.

THE KNOWLEDGE-ENGINEERING PROCESS

The knowledge-engineering process includes five major activities:

1. Knowledge acquisition. Knowledge acquisition involves the acquisition of knowl-edge from human experts, books, documents, sensors, or computer files. Theknowledge may be specific to the problem domain or to the problem-solving pro-cedures, it may be general knowledge (e.g., knowledge about business), or it maybe metaknowledge (knowledge about knowledge). (By metaknowledge, we meaninformation about how experts use their knowledge to solve problems and aboutproblem-solving procedures in general.) Byrd (1995) formally verified that knowl-edge acquisition is the bottleneck in ES development today; thus, much theoreticaland applied research is still being conducted in this area. An analysis of more than90 ES applications and their knowledge acquisition techniques and methods isavailable in Wagner et al. (2003).

2. Knowledge representation. Acquired knowledge is organized so that it will be readyfor use, in an activity called knowledge representation. This activity involves prepa-ration of a knowledge map and encoding of the knowledge in the knowledge base.

3. Knowledge validation. Knowledge validation (or verification) involves validatingand verifying the knowledge (e.g., by using test cases) until its quality is acceptable.Testing results are usually shown to a domain expert(s) to verify the accuracy ofthe ES.

4. Inferencing. This activity involves the design of software to enable the computer tomake inferences based on the stored knowledge and the specifics of a problem.The system can then provide advice to non-expert users.

5. Explanation and justification. This step involves the design and programming of anexplanation capability (e.g., programming the ability to answer questions such aswhy a specific piece of information is needed by the computer or how a certainconclusion was derived by the computer).

Figure W18.1 shows the process of knowledge engineering and the relationshipsamong the knowledge engineering activities. Knowledge engineers interact withhuman experts or collect documented knowledge from other sources in the knowledgeacquisition stage. The acquired knowledge is then coded into a representation schemeto create a knowledge base. The knowledge engineer can collaborate with humanexperts or use test cases to verify and validate the knowledge base. The validated



knowledge can be used in a knowledge-based system to solve new problems viamachine inference and to explain the generated recommendation. Details of theseactivities are discussed in the following sections.Application Case W18.1 illustrates thedevelopment of a tire-defect ES in Egypt.

Application Case W18.1

Development of Tire-Defect Diagnosis ES in Egypt

A manufactured tire is composed of complex formula-tions of fibers, textiles, and steel cords; in addition, trucktires are normally made up of special raw materials. Themanufacturing process includes 5 steps and more than 80different raw materials. Unless this process is well con-trolled, many factors may cause the tires to be defective.

TIREDDX (Tire Defect Diagnosis Expert System) isan integrated intelligent system developed in a leadingtruck-tire production company in Egypt. It is capable ofdiagnosing the probable cause(s) of a defect by tracing theacquired quality and production information at varioussteps of the tire-manufacturing process.

TIREDDX was developed in three major steps:

1. Acquiring and formulating the related knowledge.This step included collecting relevant pieces ofknowledge and information from experts. The knowl-edge acquisition step had two major stages: Theconstruction of a global picture of the knowledge

base and the creation of knowledge rules. The secondstage allowed more rules to be added after building aprototype from the first stage.

2. Designing and developing different knowledge bases.Knowledge is represented in the form of rules andstored in a knowledge base.

3. Testing the knowledge base for validation, using his-torical cases to evaluate the system. The performanceof the system was comparable with that of humanexperts.

The results show that the system has had the follow-ing advantages:

• It reduced the time required to reach diagnosticdecisions.

• It now formulates more consistent diagnostic decisions.

• It better utilizes the company’s information systems.

Explanation,justification

Inferencing

Knowledgebase

Knowledgevalidation

(test cases)

Sources of knowledge(experts, others)

Knowledgeacquisition

Knowledgerepresentation

Encoding

FIGURE W18.1 Process of Knowledge Engineering


The Scope and Types of Knowledge W-179 ◆

Since its development, the system has been continu-ously improved, and it is utilized by many users in thecompany.

Source: Adapted from M.G. Abou-Ali and M. Khamis,“TIREDDX: An Integrated Intelligent Defects DiagnosticSystem for Tire Production and Service,” Expert Systems withApplications, Vol. 24, 2003, pp. 247–259.

THE COMMONKADS METHODOLOGY

CommonKADS is considered the leading methodology to support structured knowl-edge engineering. It has been gradually developed and has been validated by manycompanies and universities in the context of the European ESPRIT IT program. It isnow the European de facto standard for knowledge analysis and knowledge-intensivesystem development, and many companies have adopted it in whole or partly incorpo-rated it in existing methods.

CommonKADS enables the recognition of opportunities and bottlenecks in howorganizations develop, distribute, and apply their knowledge resources, so it is a toolfor corporate knowledge management. CommonKADS also provides the methods toperform a detailed analysis of knowledge-intensive tasks and processes. Finally,CommonKADS supports the development of knowledge systems that supportselected parts of the business process. For details, see Schreiber et al. (2000) anden.wikipedia.org/wiki/CommonKADS.

Section W18.2 Review Questions

1. Define knowledge engineering.

2. List and briefly describe each of the five activities of knowledge engineering.

W18.3 THE SCOPE AND TYPES OF KNOWLEDGE

The most important factor in knowledge acquisition is the extraction of knowledgefrom sources of expertise and its transfer to the knowledge base and then to the infer-ence engine. Acquisition is actually done throughout the entire development process.

Knowledge is a collection of specialized facts, procedures, and judgment usuallyexpressed as rules. Some types of knowledge used in artificial intelligence are shown inFigure W18.2. These types of knowledge may come from one or from several sources.

SOURCES OF KNOWLEDGE

Knowledge can be collected from many sources, such as books, films, computer data-bases, pictures, maps, flow diagrams, stories, sensors, radio frequency identification(RFID), songs, or even observed behavior. Sources are of two types: documentedknowledge and undocumented knowledge. The latter resides in people’s minds.Knowledge can be identified and collected by using one or several of the humansenses. It can also be identified and collected by machines (e.g., sensors, scanners, cam-eras, pattern matchers, intelligent agents).

The multiplicity of sources and types of knowledge contributes to the complexityof knowledge acquisition. This complexity is only one reason why it is difficult toacquire knowledge. Other reasons are discussed in Section W18.4.



Knowledge Acquisition from DatabasesMany ES are constructed from knowledge extracted in whole or in part from data-bases. With the increased amount of knowledge stored in databases, the acquisition ofsuch knowledge becomes more difficult. For discussion and recent developments,see ACM Special Interest Group on Knowledge Discovery and Data Mining(SIGKDD) (acm.org/sigkdd/), the Data Mining and Knowledge Discovery Journal(springerlink.metapress.com/content/1573-756X/), and the Microsoft Web site(research.microsoft.com). Roiger and Geatz (2003) gave a very good practical descrip-tion of techniques for mining knowledge from data and their applications. Some tech-niques, such as neural networks and genetic algorithms, are described in Chapters 8and 13.

Knowledge Acquisition via the InternetWith the increased use of the Internet, it is possible to access vast amounts of knowl-edge online. The acquisition, availability, and management of knowledge via theInternet are becoming critical success issues for the construction and maintenance ofknowledge-based systems, particularly because they allow the acquisition and dissemi-nation of large quantities of knowledge in a short time across organizational and phys-ical boundaries. Knowledge acquisition methods and standards were first establishedin 1996. For details of mining knowledge from the Web, see Chakrabarti (2002).

The Internet community provides an ES with a scope far greater than is possiblewith standard computer-based systems. Benjamins et al. (1999) described ontologiesfor building acquisition tools utilizing intranet/Internet tools available within organiza-tions. Adopting such tools as HTML browsers allows users to quickly become familiarwith acquisition systems.

Behaviordescriptionand beliefs

Vocabularydefinitions

Objectsand

relationships

Heuristics anddecisions rules

Procedures forproblem solving

Typicalsituations

Hypotheses(theories)

Knowledge basic

Knowledge about knowledge(metaknowledge)

Uncertainfacts

Processes

Constraints

Facts aboutthe domain

Disjunctivefacts

General knowledge(e.g. of the world)

FIGURE W18.2 Types of Knowledge to Be Represented in a Knowledge Base

Source: Adapted from R. Fikes and T. Kehler, “The Role of Frame-Based Representation in Reasoning,”Communications of ACM, September 1985. Copyright 1985, Association for Computing Machinery, Inc.Reprinted by permission.


The Scope and Types of Knowledge W-181 ◆

Additions to standard HTML can be applied to include metadata information,allowing explicit information to be stored and retrieved. The detailed case studiesgiven by Benjamins et al. (1999) highlight the difficulties in adopting this approach andaddress differences between ontology-based and keyword-based knowledge retrieval.Also worth noting is the development of the semantic Web, which incorporates seman-tic knowledge into the Web representation (see Davies et al., 2003). See Chapter 14 fora detailed description.

LEVELS OF KNOWLEDGE

Knowledge can be represented at different levels. The two extremes are shallowknowledge (i.e., surface knowledge) and deep knowledge.

Shallow KnowledgeShallow knowledge is the representation of surface-level information that can be usedto deal with very specific situations. For example, if you don’t have gasoline in your car,the car won’t start. This knowledge can be shown as a rule:

If gasoline tank is empty, then car will not start.

The shallow version basically represents the input/output relationship of a system.As such, it can be ideally presented in terms of IF-THEN rules. Shallow representationis limited.A set of rules by itself may have little meaning for the user. It may have littleto do with the manner in which experts view the domain and solve problems. This maylimit the capability of the system to provide appropriate explanations to the user.Shallow knowledge may also be insufficient in describing complex situations.Therefore, a deeper presentation is often required.

Deep KnowledgeHuman problem solving is based on deep knowledge of a situation. Deep knowledge isthe internal and causal structure of a system and involves the interactions between thesystem’s components. Deep knowledge can be applied to different tasks and differentsituations. It is based on a completely integrated, cohesive body of human conscious-ness that includes emotions, commonsense, intuition, and so on.This type of knowledgeis difficult to computerize.The system builder must have a perfect understanding of thebasic elements and their interactions, as produced by nature. To date, such a task hasbeen found to be impossible. However, it is possible to implement a computerized rep-resentation that is deeper than shallow knowledge. To explain how this is done, let usreturn to the gasoline example. If we want to investigate at a deeper level the relation-ship between lack of gasoline and a car that won’t start, we need to know the variouscomponents of the gas system (e.g., pipes, pump, filters, starter). Such a system is shownschematically in Figure W18.3.

To represent this system and knowledge of its operation, we use special knowledgerepresentation methods such as semantic networks and frames (see Section W18.8).Theseallow the implementation of deeper-level reasoning such as abstraction and analogy,which are important expert activities. We can also represent the objects and processes ofthe domain of expertise at this level; the relationships between objects are important.

Tutoring is an important type of expertise that has been represented with a deep-level approach. The goal of tutoring is to convey to students a domain knowledge thatis best represented at the deep level: concepts, abstractions, analogies, and problem-solving strategies. Deep knowledge is much more difficult to collect and validate thanshallow knowledge. For a discussion, see Bergeron (2003). For details on Web-baseddeep knowledge, see Mussi (2006).



MAJOR CATEGORIES OF KNOWLEDGE

Knowledge can be categorized as declarative knowledge, procedural knowledge, ormetaknowledge.

Declarative KnowledgeDeclarative knowledge is a descriptive representation of knowledge. It tells us facts:what things are. It is expressed in a factual statement, such as “There is a positive asso-ciation between smoking and cancer.” Domain experts tell us about truths and associ-ations. This type of knowledge is considered shallow, or surface-level, information thatexperts can verbalize. Declarative knowledge is especially important in the initial stageof knowledge acquisition.

Procedural KnowledgeProcedural knowledge considers the manner in which things work under different setsof circumstances.The following is an example:“Compute the ratio between the price ofa share and the earnings per share. If the ratio is larger than 12, stop your investigation.Your investment is too risky. If the ratio is less than 12, check the balance sheet.” Thus,procedural knowledge includes step-by-step sequences and how-to types of instruc-tions; it may also include explanations. Procedural knowledge involves automaticresponses to stimuli. It may also tell us how to use declarative knowledge and how tomake inferences.

Declarative knowledge relates to a specific object. It includes information aboutthe meaning, roles, environment, resources, activities, associations, and outcomes of theobject. Procedural knowledge relates to the procedures used in the problem-solvingprocess (e.g., information about problem definition, data gathering, the solutionprocess, evaluation criteria).

MetaknowledgeMetaknowledge is knowledge about knowledge. In ES, metaknowledge is knowledgeabout the operation of knowledge-based systems (i.e., about their reasoningcapabilities).

Car

Starter

Brushes

Has

Gas pump

Has a

StartsHas a

Air

Gasoline

Flows via a pipe

FilterCarburetor

Stores

Connected to

Pumps Has a

gas to a

Flows to

Gas tank

FIGURE W18.3 Schematic Representation of Deep Knowledge: Automobile Gas System


Methods of Acquiring Knowledge from Experts W-183 ◆


1. List the major sources of knowledge.

2. Differentiate between documented and undocumented knowledge.

3. How is knowledge collected?

4. Define shallow knowledge and deep knowledge.

5. Define procedural knowledge and declarative knowledge.

6. Define metaknowledge.

W18.4 METHODS OF ACQUIRING KNOWLEDGE FROM EXPERTS

Knowledge acquisition is not an easy task. It includes identifying the knowledge, rep-resenting the knowledge in a proper format, structuring the knowledge, and transfer-ring the knowledge to a machine. Some difficulties in knowledge acquisition aredescribed in Technology Insights W18.2. The process of knowledge acquisition can begreatly influenced by the roles of the three major participants: the knowledge engi-neer, the expert, and the end user.

A unique approach to the interrelationships of these participants is offered bySandahl (1994). Sandahl indicated that experts should take a very active role in thecreation of a knowledge base. The knowledge engineer should act like a teacher ofknowledge structuring, a tool designer, and a catalyst at the interface between the expertand the end users. The requirements for a qualified knowledge engineer are listed inTechnology Insights W18.3.This approach could minimize such problems as inter-humanconflicts, knowledge engineering filtering, and end-user acceptance of the system. Also,knowledge maintenance problems can be reduced. Wagner and Holsapple (1997)analyzed the roles the participants play in knowledge acquisition. They suggested that itis appropriate to think of the participants as playing one or more roles, including actingas knowledge sources, agents, and targets for knowledge acquisition processes. They

TECHNOLOGY INSIGHTS W18.2

Difficulties in Knowledge Acquisition

Acquiring knowledge from experts is not an easy task.The following are some factors that add to the complex-ity of knowledge acquisition from experts and its trans-fer to a computer:

• Experts may not know how to articulate theirknowledge or may be unable to do so.

• Experts may lack time or may be unwilling tocooperate.

• Testing and refining knowledge is complicated.

• Methods for knowledge elicitation may be poorlydefined.

• System builders tend to collect knowledge fromone source, but the relevant knowledge may bescattered across several sources.

• Builders may attempt to collect documentedknowledge rather than use experts. The knowledgecollected may be incomplete.

• It is difficult to recognize specific knowledge whenit is mixed up with irrelevant data.

• Experts may change their behavior when they areobserved or interviewed.

• Problematic interpersonal communication factorsmay affect the knowledge engineer and the expert.



further developed a participant model to explain participant interactions in a metaviewof knowledge acquisition.This view allows for a more flexible consideration of the manypossible combinations that can and do occur in reality.

ROLES OF KNOWLEDGE ENGINEERS

The ability and personality of the knowledge engineer directly influence the expert.Part of successful knowledge acquisition involves developing a positive relationshipwith the expert. The knowledge engineer is responsible for creating the right impres-sion, positively communicating information about the project, understanding theexpert’s style, preparing the sessions, and so on. The skills required by knowledge engi-neers are shown in Technology Insights W18.3.

The basic model of knowledge engineering portrays teamwork in which a knowl-edge engineer mediates between the expert and the knowledge base. Figure W18.4shows the following tasks performed by knowledge engineers at different stages ofknowledge acquisition:

1. Advise the expert on the process of interactive knowledge elicitation.2. Set up and appropriately manage the interactive knowledge acquisition tools.3. Edit the unencoded and coded knowledge base in collaboration with the expert.4. Set up and appropriately manage the knowledge-encoding tools.5. Validate application of the knowledge base in collaboration with the expert.6. Train clients in effective use of the knowledge base in collaboration with the

expert by developing operational and training procedures.

The elicitation of knowledge from the expert can be seen as a process of modeling(see Technology Insights W18.4) and can be done manually or with the aid of comput-ers. Most manual elicitation techniques have been borrowed (often with modifica-


Required Skills of Knowledge Engineers

Knowledge engineers are human professionals who areable to communicate with experts and consolidateknowledge from various sources to build a valid knowl-edge base. They can use computers and special methodsto overcome difficulties in knowledge engineering. Thefollowing are some of the skills and characteristics thatare desirable in knowledge engineers:

• Computer skills (e.g., hardware, programming,software)

• Tolerance and ambivalence

• Effective communication abilities (e.g., sensitivity,tact, diplomacy)

• Broad educational background

• Advanced, socially-sophisticated verbal skills

• Fast-learning capabilities (of different domains)

• Understanding of organizations and individuals

• Wide experience in knowledge engineering

• Empathy and patience

• Persistence

• Logical thinking

• Versatility and inventiveness

• Self-confidence

Because of these requirements, knowledge engi-neers are in short supply (and they are costly as well,demanding high salaries). Some of the automationdevelopments described later in this chapter attempt toovercome the short-supply problem.



tions) from psychology or from system analysis. These elicitation methods are classi-fied in different ways and appear under different names; for a discussion, see Moody etal. (1998/1999).

KNOWLEDGE MODELING METHODS

The knowledge modeling methods can be classified into three categories: manual,semiautomatic, and automatic.

Source: Adapted from B.R. Gaines, University of Calgary, with permission.

ExpertI

ExpertII

Knowledgeacquisition

Knowledgebase

Encoding

Computerknowledge

base

ES tools

ClientI

ClientII

Train

Validate

Edit

Edit

Advise

Manageencoding

Manageacquisition

Knowledgeengineer

FIGURE W18.4 Knowledge Engineers’ Roles in Interactive Knowledge Acquisition



Manual methods are basically structured around an interview of some kind. Theknowledge engineer elicits knowledge from the expert or other sources and then codesit in the knowledge base. The three major manual methods are interviewing (i.e., struc-tured, semistructured, unstructured), tracking the reasoning process, and observing.Manual methods are slow, expensive, and sometimes inaccurate. Therefore, there is atrend toward automating the process as much as possible.

Semiautomatic methods are divided into two categories: those intended to supportthe experts by allowing them to build knowledge bases with little or no help fromknowledge engineers and those intended to help knowledge engineers by allowingthem to execute the necessary tasks in a more efficient or effective manner (sometimeswith only minimal participation by an expert).

In automatic methods, the roles of both the expert and the knowledge engineerare minimized or even eliminated. For example, the induction method, which generatesrules from a set of known cases, can be applied to build a knowledge base. The roles ofthe expert and knowledge engineers are minimal. The term automatic may be mislead-ing, but it indicates that, compared with other methods, the contributions from aknowledge engineer and an expert are relatively small.

Manual Knowledge Modeling MethodsThe manual knowledge modeling methods include: Interview (structured and unstruc-tured), process tracking, protocol analysis, observation. These methods and others arediscussed next.

Interviews The most commonly used form of knowledge acquisition is face-to-faceinterview analysis. This is an explicit technique that appears in several variations. Itinvolves a direct dialog between the expert and the knowledge engineer. Informationis collected with the aid of conventional instruments (e.g., tape recorders, question-naires) and is subsequently transcribed, analyzed, and coded. In the interview, theexpert is presented with a simulated case or, if possible, with an actual problem of the


Knowledge Modeling

Several key contributions made during the 1980s, includ-ing Allen Newell’s notion of knowledge level, WilliamClancey’s critical analyses, and the broader wave ofsecond-generation ES research, have shaped our currentperception of the knowledge acquisition problem. Centralto the current perception is the knowledge model, whichviews knowledge acquisition as the construction of amodel of problem-solving behavior—that is, a model interms of knowledge instead of representations.

The concept of knowledge-level modeling hasmatured considerably. The practical knowledge-levelmodels incorporated in today’s methodologies do notsimply reflect the knowledge content of a system; theyalso make explicit the structures within which the knowl-edge operates in solving various classes of problems.Thisenables the reuse of models across applications.

The model structures provide a framework forknowledge acquisition and a decomposition of the over-all acquisition task. Identified parts of knowledge-levelmodels—domain models or problem-solving methods—can serve in different systems or in different roles in thesame system. The main advantage remains: The knowl-edge level focuses attention on the knowledge thatmakes systems work rather than on the symbol-level,computational-design decisions that provide the opera-tional framework. The topic of knowledge modeling isdescribed by Rincon et al. (2005). Musen et al. (1999)reported on the use of a domain models to drive an inter-active knowledge editing tool, allowing expertise to beentered regardless of the representation requirements ofthe target ES. More recent work includes knowledgemodeling on the Web. For details, see Davies et al. (2003).



type that the ES will be expected to solve. The expert is asked to talk the knowledgeengineer through the solution. Sometimes this method is called the walk-throughmethod.

One variant of the interview approach begins with no information at all beinggiven to the expert.Any facts the expert requires must be asked for explicitly.This vari-ant makes the expert’s path through the domain more evident, especially in terms ofdefining the input an ES would require.

The interview process can be tedious. It places great demands on the domainexpert, who must be able not only to demonstrate expertise but also to express it. Onthe other hand, it requires little equipment, is highly flexible and portable, and canyield a considerable amount of information. There are two basic types of interviews:unstructured (informal) interviews and structured interviews.

Unstructured Interviews. Many knowledge acquisition interview sessions areconducted informally, usually as a starting point. Starting informally saves time; it helpsto move quickly to the basic structure of the domain. Usually, it is followed by a formaltechnique. Contrary to what many people believe, unstructured (informal) interviewsare not simple. In fact, they may present the knowledge engineer with some very prob-lematic after-effects.

Unstructured interviewing, according to McGraw and Harbison-Briggs (1989), sel-dom provides complete or well-organized descriptions of cognitive processes. Thereare several reasons for this: The domains are generally complex; the experts usuallyfind it very difficult to express some of the most important elements of their knowl-edge; domain experts may interpret the lack of structure as implying that they need notprepare for the interview; data acquired from an unstructured interview are oftenunrelated, exist at varying levels of complexity, and are difficult for the knowledgeengineer to review, interpret, and integrate; and few knowledge engineers have thetraining and experience to efficiently conduct an unstructured interview.

Example: Site Selection for a Hospital Extension Clinic

Here is a simple hypothetical example of acquisition of knowledge about site selection for ahospital extension clinic. The dialog between the expert (E) and the knowledge engineer (KE)might start like this:

E: I understand that you are somehow going to try to capture my knowledge aboutsite selection of clinics so that hospital administrators in our system in other cities canuse it.

KE: Yes, indeed! And thank you for spending time with me, Kathleen. We’ve noticed thatyou have a special knack for identifying the right locations. So far, you’ve pickedvery successful operations.

E: Well, thanks.KE: So, tell me, what’s the most important factor in determining where to put a new

facility?E: Really, it’s demographics. We need to locate it close to our potential customers.KE: So, is it also important that it not be too close to another agency’s operation?E: Not necessarily. It is important that it not be located too close to our main hospital or

our other facilities, but if the population density is high enough, we can locate itclose to a competitor.

KE: Tell me, then, what kind of demographics are you looking for?E: Well, in a large city, if there are at least 2,000 people per square mile over about four

square miles, generally they can support a profitable clinic.KE: What about competitors’ locations?



E: If a competitor is within two miles, the density has to exceed 3,500 people per squaremile. And if there are two competitors within two miles of each other already, there’sno point in even trying to break into the market, except for certain special services.

KE: What about the population’s income? Is that important?E: We must limit our indigent cases to no more than 2 percent of our clients, and so we

generally look for an average family income greater than $30,000 per year.KE: Is being near public transportation important?

By interviewing the expert, the knowledge engineer slowly learns what is going on. Thenhe or she builds a representation of the knowledge in the expert’s terms.

The process of knowledge acquisition involves uncovering the attributes of a problem andmaking explicit the thought process (usually expressed as rules) that the expert uses to inter-pret them.

The unstructured interview is most common and appears in several variations. In addition tothe walk-through, the knowledge expert can ask the expert to teach-through or read-through.In a teach-through, the expert acts as an instructor and the knowledge engineer as a student. Theexpert not only tells what he or she does but also explains why and instructs the knowledgeengineer in the skills and strategies needed to perform the task. In a read-through approach, theexpert is asked to instruct the knowledge engineer how to read and interpret the documents usedfor the task. More detailed descriptions of unstructured interviews can be found in Burns (2000).

Source: Adapted from R. B. Burns, Introduction to Research Methods, London: Sage, 2000.

Structured Interviews. A structured interview is a systematic, goal-orientedprocess. It forces organized communication between the knowledge engineer and theexpert. The structure reduces the interpretation problems inherent in unstructuredinterviews and allows the knowledge engineer to prevent the distortion caused by thesubjectivity of the domain expert. Structuring an interview requires attention to a num-ber of procedural issues, which are summarized as follows (see Kuhn and Zohar, 1995;and McGraw and Harbison-Briggs, 1989):

• The knowledge engineer studies available material on the domain to identifymajor demarcations of the relevant knowledge.

• The knowledge engineer reviews the planned ES capabilities. He or she identifiestargets for the questions to be asked during the knowledge acquisition session.

• The knowledge engineer formally schedules and plans the structured interviews(using a form). Planning includes attending to physical arrangements, definingknowledge acquisition session goals and agendas, and identifying or refiningmajor areas of questioning.

• The knowledge engineer may write sample questions, focusing on question type,level, and questioning techniques.

• The knowledge engineer ensures that the domain expert understands the purpose and goals of the session and encourages the expert to prepare before the interview.

• During the interview, the knowledge engineer follows guidelines for conductinginterviews.

• During the interview, the knowledge engineer uses directional control to retainthe interview’s structure.

Because every interview is different in very specific ways, it is difficult to providecomprehensive guidelines for the entire interview process. Therefore, interpersonalcommunication and analytic skills are important. However, several guidelines, check-lists, and instruments are available that are fairly generic in nature (see McGraw andHarbison-Briggs, 1989).

There are many structured interview methods. Some are based on psychology, andothers are based on disciplines, such as anthropology. Interviewing techniques,



although very popular, have many disadvantages, ranging from inaccuracy in collectinginformation to bias introduced by the interviewers.

In summary, interviews are an important technique, but they must be plannedcarefully, and the interview results must be subjected to thorough verification and val-idation methodologies. Interviews are sometimes replaced by tracking methods.Alternatively, they can be used to supplement tracking or other knowledge acquisitionmethods.

We recommend that before interviewing the main experts, the knowledge engineershould interview a less knowledgeable or minor expert, using the interviewingapproaches just described.This may help the knowledge engineer learn about the prob-lem, its significance, the experts, and the users.The interviewer will also be able to betterunderstand the basic terminology and (if he or she is a novice in the area) identifyarchived sources about the problem first. The knowledge engineer should next readabout the problem. Then the main experts can be interviewed much more effectively.

Process Tracking and Protocol Analysis Process tracking is a set of techniques thatattempt to track the reasoning process of an expert. It is a popular approach amongcognitive psychologists who are interested in discovering the expert’s train of thoughtin reaching a conclusion. The knowledge engineer can use the tracking process to findwhat information is being used and how it is being used. Tracking methods can beinformal or formal. The most common formal method is protocol analysis.

Protocol analysis, particularly the set of techniques known as verbal protocolanalysis, is a method by which the knowledge engineer acquires detailed knowledgefrom the expert. A protocol is a record or documentation of the expert’s step-by-stepinformation-processing and decision-making behavior. In this method, which is similarto interviewing but more formal and systematic, the expert is asked to perform a realtask and to verbalize his or her thought processes. The expert is asked to think aloudwhile performing the task or solving the problem under observation. Usually, a record-ing is made as the expert thinks aloud; it describes every aspect of the information-processing and decision-making behavior. The recording then becomes a record, orprotocol, of the expert’s ongoing behavior. Later, the recording is transcribed for fur-ther analysis (e.g., to deduce the decision process) and coded by the knowledge engi-neer. (For further details, see Owen et al., 2006.)

In contrast with interactive interview methods, a protocol analysis mainly involvesone-way communication. The knowledge engineer prepares the scenario and plans theprocess. During the session, the expert does most of the talking while interacting withdata to solve the problem. Concurrently, the knowledge engineer listens and recordsthe process. Later, the knowledge engineer must be able to analyze, interpret, andstructure the protocol into knowledge representation for review by the expert.

The process of protocol analysis is summarized as follows:

1. Provide the expert with a full range of information normally associated witha task.

2. Ask the expert to verbalize the task in the same manner as would be done nor-mally while verbalizing his or her decision process and record the verbalizationon tape.

3. Make statements by transcribing the verbal protocols.4. Gather the statements that seem to have high information content.5. Simplify and rewrite the collected statements and construct a table of production

rules from the collected statements.6. Produce a series of models by using the production rules.



The following are the advantages of protocol analysis:

• The expert consciously considers decision-making heuristics.• The expert consciously considers decision alternatives, attributes, and values.• The knowledge engineer can observe and analyze decision-making behavior.• The knowledge engineer can record, and later analyze with the expert, key deci-

sion points.

The following are the limitations of protocol analysis:

• The expert must be aware of why he or she makes a decision.• The expert must be able to categorize major decision alternatives.• The expert must be able to verbalize the attributes and values of a decision

alternative.• The expert must be able to reason about the selection of a given alternative.• A view of decision making is subjective. Explanations may not track with reasoning.

Observations Sometimes it is possible to observe an expert at work. In many ways,this is the most obvious and straightforward approach to knowledge acquisition.However, the difficulties involved should not be underestimated. For example, mostexperts advise several people and may work in several domains simultaneously. In thiscase, the knowledge engineer’s observations will cover all the other activities as well.Therefore, large quantities of knowledge are being collected, of which only a little isuseful. In particular, if recordings or videotapes are made, the cost of transcribing largeamounts of knowledge should be carefully considered.

Observations, which can be viewed as a special case of protocols, are of two types:motor movements and eye movements. With observations of motor movements, theexpert’s physical performance of the task (e.g., walking, reaching, talking) is docu-mented. With observations of eye movements, a record is made of where the expertfixes his or her gaze. Observations are used primarily as a way of supporting verbalprotocols. They are generally expensive and time-consuming.

Other Manual Knowledge Modeling Methods Many other manual methods can beused to elicit knowledge from experts. A representative list is given here (for a com-plete discussion, see Hart, 1992; and Scott et al., 1991):

• Case analysis. Experts are asked how they handled specific cases in the past.Usually, this method involves analyzing documentation. In addition to theexperts, other people (e.g., managers, users) may be questioned.

• Critical incident analysis. In this approach, only selected cases are investigated—usually those that are memorable, difficult, or of special interest. Both experts andnon-experts may be questioned.

• Discussions with users. Even though users are not experts, they can be quiteknowledgeable about some aspects of the problem. They can also indicate areaswhere they need help. The expert may be unaware of some of the users’ needs.

• Commentaries. With this method, the knowledge engineer asks experts to give arunning commentary on what he or she is doing. This method can be supported byvideotaping the experts in action or by asking an observer to do the commentary.

• Conceptual graphs and models. Diagrams and other graphical methods can beinstrumental in supporting other acquisition methods. A conceptual model can beused to describe how and when the expert’s knowledge will come into play as theES performs its task.



• Brainstorming. These methods can be used to solicit the opinions of multipleexperts and can help generate ideas. Electronic brainstorming can also be used,including blackboarding, which has been implemented as electronic whiteboards.The recently available software for this, such as Microsoft NetMeeting, MSNMessenger (msn.com), and the HomeMeeting program developed at theUniversity of Washington (homemeeting.com), are powerful tools for holdingremote knowledge acquisition sessions.

• Prototyping. Working with a prototype of the system is a powerful approach toinducing experts to contribute their knowledge. Experts like to criticize systems,and changes can be made to a prototype instantly.

• Multidimensional scaling. The complex technique of multidimensional scalingidentifies various dimensions of knowledge and then places the knowledge in theform of a distance matrix. With the use of least-squares fitting regression, the vari-ous dimensions are analyzed, interpreted, and integrated.

• Johnson’s hierarchical clustering. This is another scaling method, but it is much sim-pler to implement than multidimensional scaling and therefore is used more often.It combines related knowledge elements into clusters (two elements at a time).

• Performance review. Because ES development is an ongoing process, all of theabove can be applied iteratively as the system evolves.

Semiautomatic Knowledge Modeling MethodsKnowledge acquisition can be supported by computer-based tools. These tools providean environment in which knowledge engineers or experts can identify knowledgethrough an interactive process. Repertory grid analysis (RGA) is one typical method.

Repertory Grid Analysis (RGA) RGA is based on Kelly’s model of human thinkingcalled personal construct theory (see Hart, 1992; and Stewart and Mayes, 2000). Moreinformation about the theory can be found in the BizTech research archives(brint.com/PCT.htm/) and at the Personal Construct Psychology Web site (repgrid.com/pcp/). According to this theory, each person is viewed as a personal scientist whoseeks to predict and control events by forming theories, testing hypotheses, and analyz-ing results of experiments. Knowledge and perceptions about the world (or about adomain or a problem) are classified and categorized by each individual as a personal,perceptual model. Based on the model developed, each individual is able to anticipateand then act on the basis of these anticipations.

This personal model matches our view of an expert at work; it is a description ofthe development and use of the expert’s knowledge, and so it is suitable for ES, as sug-gested by Hart (1992). RGA is a method of investigating such a model.

How RGA Works. RGA uses several processes. First, the expert identifies theimportant objects in the domain of expertise. For example, computer languages (e.g.,LISP, C11, COBOL) are objects in the situation of needing to select a computer lan-guage. This identification is done in an interview.

Second, the expert identifies the important attributes considered in making deci-sions in the domain. For example, the availability of commercial packages and ease ofprogramming are important factors in selecting a computer language.

Third, for each attribute, the expert is asked to establish a bipolar scale with distin-guishable characteristics (i.e., traits) and their opposites. For example, in selecting acomputer language, the information shown in Table W18.1 can be included.

Fourth, the interviewer picks any three of the objects and asks, “What attributesand traits distinguish any two of these objects from the third?” For example, if a set



TABLE W18.1 RGA Input for Selecting a Computer Language

Attribute Trait Opposite

Availability Widely available Not availableEase of programming High LowTraining time Low HighOrientation Symbolic Numeric

TABLE W18.2 Example of a Grid

Ease of Training Attribute Orientation Programming Time Availability

Trait Symbolic (3) High (3) High (1) High (3)

Opposite Numeric (1) Low (1) Low (3) Low (1)

LISP 3 3 1 1PROLOG 3 2 2 2C++ 3 2 2 3COBOL 1 2 1 3

includes LISP, PROLOG, and COBOL, the expert may point to orientation. Then theexpert will say that LISP and PROLOG are symbolic in nature, whereas COBOL isnumeric.These answers are translated to points on a scale of 1 to 3 (or 1 to 5).This stepcontinues for several triplets of objects.The answers are recorded in a grid, as shown inTable W18.2. The numbers in the grid designate the points assigned to each attributefor each object.

When the grid is complete, the expert may change the ratings in the boxes.The gridcan be used afterward to make recommendations in situations in which the importanceof the attributes is known. For example, in a simplistic sense, it can be said that ifnumeric orientation is very important, then COBOL will be the recommended lan-guage. For an interesting application of RGA, see Hunter and Beck (2000).

Use of RGA in ES. A number of knowledge acquisition tools have been devel-oped based on RGA. These tools are aimed at helping in the conceptualization of thedomain. Three representative tools are ETS, AQUINAS, and KRITON.

An expert transfer system (ETS) is a computer program that interviews expertsand helps them build ES. ETS interviews experts to uncover vocabulary conclusions,problem-solving traits, trait structures, trait weights, and inconsistencies. It has beenused to construct prototypes rapidly (often in less than two hours for very small ES), toaid the expert in determining whether there is sufficient knowledge to solve a problem,and to create knowledge bases for a variety of different ES shells from its own internalrepresentation. An improved version of ETS, called NeoETS, has been developed toexpand the capabilities of ETS. The method is limited to classification-type problems(for details, see Boose and Gaines, 1990). ETS is now part of AQUINAS.

AQUINAS is a very complex tool (see Boose and Bradshaw, 1993, 1999) thatextends the problem-solving and knowledge representation of ETS by allowingexperts to structure knowledge in hierarchies. A set of heuristics has been defined andincorporated in Dialog Manager, a subsystem of AQUINAS, to provide guidance inthe knowledge acquisition process to domain experts and knowledge engineers.



Enquire Within (enquirewithin.co.nz) is an online interactive software tool thatcharts and clarifies thoughts and perceptions based on repertory grid interview tech-niques. Users are taken through compare and contrast processes, resulting in a graphi-cal representation indicating how someone has evaluated and described the subjectmatter for a particular session. Its uses include analyzing personal relationships oropportunities, decision-making support, and development of ES, and it can be used asa computer-based school study aid.

A few PC-based RGA tools are available (e.g., PCGRID from April Metzler atLehigh University). One of the first Web-based tools for knowledge acquisition,WebGrid, enhances collaborative knowledge acquisition (see ksi.cpsc.ucalgary.ca/KAW/KAW96/gaines/KMD.html). Also, there are software packages such asCircumgrids (W. Chambers, University of South Florida) for analyzing repertory gridsvia methods such as cluster analysis and factor analysis or principal component analysis.

Card Sorting Data Among the knowledge elicitation techniques card sorting isnotable for its simplicity of use, its focus on subjects’ terminology (rather than that ofexternal experts), and its ability to elicit semi-tacit knowledge. Card sorting involvescategorizing a set of pictures, objects, or labeled cards into distinct groups by using asingle criterion. Traditional semantic analysis methods that require direct researcherinterpretation of elicited linguistic terms are distinguished from methods that arepurely syntactic and hence can be automated. For an overview, see the special issue ofExpert Systems edited by Fincher and Tenenberg (2005).

Other Computer-Aided Tools A smart knowledge acquisition tool must be able toadd knowledge to the knowledge base incrementally and refine or even correct exist-ing knowledge. Sleeman and Mitchell (1996) described two systems, REFINER1 andTIGON, that allow a domain expert to perform knowledge acquisition directly. Thesesystems have been used, respectively, for patient management and for the diagnosis ofturbine errors. REFINER1 is a case-based system that infers a prototypical descriptionfor each class, which is labeled by the domain expert. It uses existing cases directly fromdatabases. It then allows the expert to work with proposed modifications that wouldremove particular inconsistencies.TIGON was developed to detect and diagnose faultsin a gas turbine engine. It incorporates background material from which it producesanalogous rule bases automatically. See Sleeman and Mitchell (1996) for details.

The KAVAS-2 project is a knowledge acquisition, visualization, and assessmentsystem that addresses the requirement to manage data and information flow in medi-cine (see Medical Informatics Laboratory application protocol system (APS),hiww.org/IR.html). Protocols and procedures are provided to increase data and infor-mation generation, improve the quality of patient management, and reduce costs inhealth care service. A toolbox (KAVIAR) is provided to define modeling-processgoals, identify suitable tools to address these goals, apply selected tools to the problem,measure the quality of the proposed solution, and validate the results.

Visual modeling techniques are often used to construct the initial domain model.The objective of the visual modeling approach is to give the user the ability to visualizereal-world problems and manipulate elements of them through the use of graphics (seeDemetriadis et al., 1999; Humphrey, 1999; and Lee et al., 1995). Kearney (1990) indi-cated that diagrams and drawings are useful in representing problems; they serve as aset of external memory aids and can reveal inconsistencies in an individual’s knowl-edge (see Lockwood and Chen, 1994). Machine learning methods can induce decisiontrees and rules (see Chapter 12). A tool available for download via the Internet isXpertRule from Attar Software (xpertrule.com). It provides a graphical developmentenvironment, allowing specification of decision trees based on business-process logic.



Deployment of developed applications can occur via standalone machines, networks,or Internet/intranets via the XpertRun runtime environment. Figure W18.5 shows ascreenshot of its Knowledge Builder.

TEIRESIAS (see Davis, 1993) is a classic program that was developed to assistknowledge engineers in the creation (or revision) of rules for a specific ES while work-ing with the EMYCIN shell. The program uses a natural language interface to help theknowledge engineer test and debug new knowledge, and it provides an expandedexplanation capability. For example, if system builders find that a set of knowledgerules lead to an inadequate conclusion, they can have TEIRESIAS show all the rulesused to reach that conclusion. With the rule editor, adjustments can easily be made. Toexpedite the process, TEIRESIAS translates each new rule, which is entered in naturallanguage, into LISP. Then it retranslates the rule into natural language. The programcan thus point out inconsistencies, rule conflicts, and inadequacies. (For acquisition aidsfrom databases, see the International Journal of Intelligent Systems, September 1992.)

A rather recent use of TEIRESIAS has been in assisting in the tracking of com-puter system hackers. TEIRESIAS is used to monitor systems during normal opera-tion and analyze the information flow, looking for repeating strings of information.These strings represent an individual computer system; an attempt to break into thecomputer system would modify them, allowing a computer to monitor its own opera-tion. For further information and actual applied use, see domino.research.ibm.com/comm/research.nsf/pages/r.compbio.html.

Most of the preceding tools and many others were designed as standalone toolsbased on the assumption that they would be used by a specific ES participant (e.g., anexpert) for the execution of a specific task. In reality, however, participants can playmultiple roles or even exchange roles.Therefore, there is a trend toward integrating theacquisition aids. For an overview of such integration tools, see Boose and Gaines(1990) and Diamantidis and Giakoumakis (1999). Examples of such tools are

Source: Adapted from attar.com/pages/info_kb.htm.

FIGURE W18.5 Decision Tree for Knowledge Building



PROTÉGÉ-II (see Gennari et al., 2003; and Musen et al., 1995), KSM (see Cuena andMolina, 1996), and KADS (see commonkads.uva.nl and van Harmelen, et al., 1996).

Automatic Knowledge Modeling MethodsIn addition to semiautomatic methods that use computers to support the knowledgeacquisition process, there are situations in which knowledge can be extracted automati-cally from existing data.The process of using computers to extract knowledge from datais called knowledge discovery. In early 1990s, the process was also called machine learn-ing, but knowledge discovery and data mining (KDD) is now a more popular term.

There are two major reasons for the use of automated knowledge acquisition:Good knowledge engineers are highly paid and difficult to find, and domain expertsare usually busy and sometimes uncooperative.As a result, manual and even semiauto-matic elicitation methods are slow and expensive. In addition, they have some otherdeficiencies:

• The correlation between verbal reports and mental behavior is often weak.• In certain situations, experts are unable to provide an overall account of how they

make decisions.• The quality of the system depends too much on the quality of the expert and the

knowledge engineer.• The expert does not understand the ES technology.• The knowledge engineer may not understand the business problem.• It is difficult to validate the acquired knowledge.

The difference between machine learning and knowledge discovery is vague. Somepeople tend to define machine learning as inductive learning and knowledge discoveryas including more techniques (e.g., clustering, neural networks).

XpertRule from XpertRule Software (xpertrule.com) is a software tool thatextracts information from graphical decision trees. Excel add-on products for datamining, such as XLMiner (resample.com/xlminer/index.shtml), are also handy in dis-covering knowledge from data.A glossary of terms related to knowledge discovery canbe found in Kohavi and Provost (1998). More software tools and information aboutKDD may be found at kdnuggets.com.The following are a few useful machine learningWeb sites:

• Austrian Research Institute for Artificial Intelligence, ofai.at• Machine Learning Network at UCI, ics.uci.edu/~mlearn/MLOther.html• MINEit Software Limited, kdnuggets.com/software.html• Machine Learning Network–Online Information Service, mlnet.org• Institute for Information Technology, National Research Council, Canada,

iit-iti.nrc-cnrc.gc.ca

According to Roiger and Geatz (2003), typical methods for knowledge discoveryinclude the following:

• Inductive learning. Rules are induced from existing cases with known results. Theinduced rules can then be stored in a knowledge base for consultation. Thismethod is described in Section 18.6.

• Neural computing. Neural computing is another problem-solving approach inwhich historical cases are used for deriving solutions to new problems. It mimicsthe human brain by building artificial neurons and storing knowledge in the con-nection of neurons (see Chapter 8).



• Genetic algorithms. Genetic algorithms use the principle of natural selection togradually find the best combination of knowledge from known cases. As in thenatural process of evolution, the basic operations for discovering knowledge arereproduction, crossover, and mutation (see Chapter 13).

In Chapter 13, we describe how inductive learning and other methods work.Techniques are also available for discovering knowledge from documents and theInternet.


1. List the difficulties in acquiring knowledge from experts.

2. List the required skills of knowledge engineers.

3. What are the major roles of knowledge engineers?

4. Define knowledge modeling.

5. Describe structured and unstructured interviews.

6. Define process tracking and analysis.

7. Describe observation as a method of knowledge acquisition.

8. List some other manual methods of knowledge acquisition.

9. Describe RGA.

10. List and describe other automated methods of knowledge acquisition.

W18.5 ACQUIRING KNOWLEDGE FROM MULTIPLE EXPERTS

An important element in the development of an ES is the identification of experts.Thisis a complicated task, perhaps because practitioners use so many support mechanismsfor certain tasks (e.g., questionnaires, informal and formal consultations, texts). Thesesupport mechanisms contribute to the high quality of professional output, but theymay also make it difficult to identify a “knowledge czar” whose estimates, processes, orknowledge are clearly superior to what the system and mix of staff, support tools, andconsulting skills produce in the rendering of normal client service.

The usual approach to this problem is to build ES for a very narrow domain inwhich expertise is clearly defined; then it is easy to find one expert. However, eventhough many ES have been constructed with one expert—an approach advocated as agood strategy for ES construction—there could be a need for multiple experts, espe-cially when more serious systems are being constructed or when expertise is not partic-ularly well defined.The case described in the opening vignette (see Ranjan et al., 2002)is a good example of using multiple experts.

The following are the major purposes of using multiple experts:

• To better understand the knowledge domain• To improve knowledge-base validity, consistency, completeness, accuracy, and

relevancy• To provide better productivity• To identify incorrect results more easily• To address broader domains• To be able to handle more complex problems and combine the strengths of differ-

ent reasoning approaches


Acquiring Knowledge from Multiple Experts W-197 ◆

Table W18.3 lists the benefits and limitations of multiple experts.When multiple experts are used, there are often differences of opinion and con-

flicts that must be resolved. This is especially true when knowledge bases are beingdeveloped from multiple sources where these systems address problems that involvethe use of subjective reasoning and heuristics.

Other related issues are identifying the various aspects of the problem and match-ing them to the appropriate experts, integrating knowledge from different experts,assimilating conflicting strategies, personalizing community knowledge bases, anddeveloping programming technologies to support these issues.

MULTIPLE-EXPERT SCENARIOS

There are four possible scenarios, or configurations, for using multiple experts (seeO’Leary, 1993; and Rayham and Fairhurst, 1999): individual experts, primary and sec-ondary experts, small groups, and panels.

Individual ExpertsIn this case, several experts contribute knowledge individually. Using multiple expertsin this manner relieves the knowledge engineer of the stress associated with multiple-expert teams. However, this approach requires that the knowledge engineer have ameans of resolving conflicts and handling multiple lines of reasoning. An example of aDelphi process performed with questionnaires for acquiring knowledge about hospitaloperating-room scheduling is provided by Hamilton (1996). Once acquired, the knowl-edge was deployed in an ES at multiple sites.

PRIMARY AND SECONDARY EXPERTS

A primary expert may be responsible for validating information retrieved from otherdomain experts. Knowledge engineers may initially consult the primary expert forguidance in domain familiarization, refinement of knowledge acquisition plans, andidentification of potential secondary experts. These scenarios determine in part themethod to be used for handling multiple experts.

Small GroupsSeveral experts may be consulted together and asked to provide agreed-upon informa-tion. Working with small groups of experts allows the knowledge engineer to observe

TABLE W18.3 Benefits of and Problems with Participation of Multiple Experts

Benefits

On average, fewer mistakes than with a single expert

Elimination of the need for using a world-class expert (who is difficult to get and expensive)

Wider domain than a single expert’sSynthesis of expertiseEnhanced quality due to synergy

among experts

Problems

Groupthink phenomenon

Fear on the part of some domain experts of senior experts or a supervisor (i.e., lack of confidentiality)

Compromising solutions generated by a group with conflicting opinions

Wasted time in group meetingsDifficulties in scheduling the expertsDominating experts (i.e., controlling, not

letting others speak)



alternative approaches to the solution of a problem and the key points made insolution-oriented discussions among experts.

PanelsTo meet goals for verification and validation of ongoing development efforts, a pro-gram may establish a council of experts. The members of the council typically meettogether at times scheduled by the developer for the purpose of reviewing knowledgebase development efforts, content, and plans. In many cases, the functionality of the ESis tested against the expertise of such a panel.

METHODS OF HANDLING MULTIPLE EXPERTS

The expert judgments of multiple experts may differ. Several major approaches to theissue of integrating expert opinions have been defined by Medsker and Turban (1995)and by O’Leary (1993):

• Blend several lines of reasoning through consensus methods such as Delphi, nom-inal group technique (NGT), and group support systems (GSS).

• Use an analytic approach, such as group probability (see O’Leary, 1993), or ananalytic hierarchy process (AHP; see Hurley and Lior, 2002).

• Keep the lines of reasoning distinct and select a specific line of reasoning basedon the situation.

• Automate the process, using software (see the ACACIA Web site,sop.inria.fr/acacia/Recherches/acacia-recherche-anglais.html) or a blackboardapproach.

• Decompose the knowledge acquired into specialized knowledge sources (i.e.,blackboard systems).


1. List the benefits of using multiple experts.

2. List the shortcomings of using multiple experts.

3. List some configurations of multiple experts.

4. List methods for using multiple experts.

W18.6 AUTOMATED KNOWLEDGE ACQUISITION FROM DATA AND DOCUMENTS

As described in Section W18.4, knowledge may be discovered from existing cases ordocuments, using computer software. The reasons for using automated knowledgeacquisition are summarized in Section W18.4. For rule-based ES, a classic method is touse rule-inductive methods in machine learning. Knowledge may also be discoveredfrom documents; this is often called text mining. Details of text mining are outlined inTechnology Insights W18.5. In this section, we describe rule induction in detail.

The following are the objectives of using automated knowledge acquisition:

• To increase the productivity of knowledge engineering (reduce the cost)• To reduce the skill level required from the knowledge engineer• To eliminate (or drastically reduce) the need for an expert• To eliminate (or drastically reduce) the need for a knowledge engineer• To increase the quality of the acquired knowledge


Automated Knowledge Acquisition from Data and Documents W-199 ◆


Acquisition of Documented Knowledge

Knowledge can often be acquired from other sources inaddition to, or instead of, human experts. This approachhas the major advantage of eliminating the need to usean expert. It is used in knowledge-based systems inwhich handling a large amount of or complex informa-tion rather than world-class expertise is the main con-cern. Searching through corporate policy manuals orcatalogs is an example.

At present, very few methodologies deal withknowledge acquisition from documented sources.It may be difficult to find the necessary knowledge ina database. One way to improve such searches is touse domain knowledge to guide the search (seeOwrang and Groupe, 1996). Another is the use of intel-ligent agents (see Chapter 14). Acquisition from docu-mented sources has a great potential for automation.Documented knowledge of almost any type can be eas-ily and inexpensively scanned and transferred to a com-puter’s database. The knowledge can then be analyzedmanually or with the use of artificial intelligence tech-nologies (a combination of natural language processing[NLP], intelligent agents, and ES). Thus, ES can be usedto build other ES.

Hahn et al. (1996) developed a methodology forknowledge acquisition and concept learning from texts(in German). The method relies on a quality-basedmodel of terminological reasoning, using concepts fromNLP. The goal of this method is to be able to scan twokinds of documents: test reports on information-technology products (about 100 documents with100,000 words) and reports of medical findings (about120,000 documents with 10 million words).

Another approach, proposed by Wang (2005), usesinduction of classification rules based on set theory.

The capability of constructing an ES that can scandatabases, digitized books, journals, and magazines isincreasing. Data stored in another computer system canbe retrieved electronically to create or update theknowledge base of the ES, all without the intervention ofa knowledge engineer or an expert. The field is basicallyat the stage of developing new methods that interpretmeanings in order to determine rules and other forms ofknowledge, such as frames for case-based reasoning. Anumber of new methods are being developed and imple-mented. More details about text mining and Web miningcan be found in Chapter 7 and Berry (2003a).

AUTOMATED RULE INDUCTION

Induction is the process of reasoning from the specific to the general. In ES terminol-ogy, it is the process by which a computer program generates rules from sample cases.A rule induction system is given examples of a problem (called a training set) for whichthe outcome is known.After it has been given enough examples, the rule induction sys-tem can create rules that fit the sample cases. The rules can then be used to assess newcases for which the outcome is not known. The heart of a rule induction system is analgorithm used to induce the rules from the examples (see Chapter 7).

Several other methods exist for automatic generation of knowledge; an example isdescribed in Application Case W18.6. For another application, see Chun and Kim (2004).


Induction-Based System for Selection of Retaining Walls

Retaining walls are used in building construction projectsfor excavating foundations. Design engineers oftenencounter the problem of choosing the most suitable ofvarious different alternatives, such as a slurry wall, a steel-rail pile, or an H-section pile. Making the best decision

requires professional knowledge and experience in orderto reduce the risk.

This selection task has been automated through thedevelopment of a system that combines rules and cases.The system stores cases in the case base and induces rules



TABLE W18.4 Case for Induction: A Knowledge Map (Induction Table)

Annual Age Number of Applicant Income ($) Assets ($) (years) Dependents Decision

Mr. White 50,000 100,000 30 3 YesMs. Green 70,000 None 35 1 YesMr. Smith 40,000 None 33 2 NoMs. Rich 30,000 250,000 42 0 Yes

from them. The rules are used to make decisions when anew case is encountered. The system has the followingmajor components:

• Case base. The case base serves as an organizationalmemory that stores all previous designs.

• Rule base. The rule base stores all rules induced fromthe case base.

• Rule induction mechanism. The module usesXpertRule to generate rules.

• Rule reasoning mechanism. The module derives newconclusions from existing rules.

• User interface. The interface allows the system tointeract with the user.

A preliminary performance test of this systemrecorded 71 percent accuracy on 48 test cases. The resultwas considered acceptable.

Sources: Compiled from J.B. Yang, “A Rule Induction–BasedKnowledge System for Retaining Wall Selection,” ExpertSystems with Applications, Vol. 23, 2002, pp. 273–279; andConstruction Industry Institute, Expert Systems Applicationand Tutorial, construction-institute.org/scriptcontent/more/sd75_more.cfm (accessed April 2006).

An example of a simplified rule induction can be seen in the work of a loan officerin a bank. Requests for loans include information about the applicants, such as incomelevel, assets, age, and number of dependents.These are the attributes, or characteristics,of the applicants. If we log several sample cases, each with its final decision, we find asituation that resembles the data in Table W18.4. From these cases, it is easy to infer thefollowing three rules:

• If income is $70,000 or more, approve the loan.• If income is $30,000 or more, age is at least 40, assets are above $249,000, and

there are no dependents, approve the loan.• If income is between $30,000 and $50,000 and assets are at least $100,000,

approve the loan.

Advantages of Rule InductionOne of the leading researchers in the field, Donald Michie, has pointed out that onlycertain types of knowledge can be properly acquired using manual knowledge acquisi-tion methods, such as interviews and observations.These are cases in which the domainof knowledge is certain, small, loosely-coupled, or modular. As the domain gets largeror uncertain, and as it gets more complex, experts become unable to explain how theyperform. Nonetheless, they can still supply the knowledge engineer with suitable exam-ples of problems and their solutions. Using rule induction allows ES to be used in morecomplicated and more commercially rewarding fields.

Another advantage of rule induction is that the builder does not have to be aknowledge engineer. He or she can be the expert or a system analyst. This not onlysaves time and money but also solves the difficulties in dealing with a knowledge engi-neer who may be an outsider unfamiliar with the business.


Automated Knowledge Acquisition from Data and Documents W-201 ◆

Machine induction also offers the possibility of deducing new knowledge.After rulesare generated, they are reviewed by an expert and modified if necessary.A big advantageof rule induction is that it enhances the thinking process of the reviewing expert.

Difficulties in Implementing Rule InductionDespite the advantages, there are several difficulties with the implementation of ruleinduction:

• Some induction programs may generate rules that are not easy for a human tounderstand because of the way the program classifies a problem’s attributes andproperties may not be the way a human would do it.

• Rule induction programs do not select the attributes. An expert must still beavailable to specify which attributes are significant (e.g., important factors inapproving a loan).

• The search process in rule induction is based on special algorithms that generateefficient decision trees, which reduce the number of questions that must be askedbefore a conclusion is reached. Several alternative algorithms are available. Theyvary in their processes and capabilities.

• Rule induction is only good for rule-based classification problems, especially ofthe yes/no type. (However, many problems can be rephrased or split so that theyfall into the classification category.)

• The number of attributes must be fairly small. The upper limit on the number ofattributes is approached very quickly.

• The number of examples necessary can be very large.• The set of examples must be “sanitized”; for example, cases that are exceptions to

rules must be removed. (Such exceptions can be determined by observing incon-sistent rules.)

• Rule induction is limited to situations under certainty.• A major problem with rule induction is that the builder does not know in advance

whether the number of examples is sufficient and whether the algorithm is goodenough. To be sure of this would presuppose that the builder had some idea ofthe solution, but the reason for using induction is that the builder does not knowthe solution and wants to discover it by using the rules.

Because of these limitations, the rule induction method is generally used to providea first prototype; the prototype is then translated into something more robust and craftedinto an improved system through an evolutionary design and development method.

Rule Induction Software PackagesMany software packages for induction are available. Some of them are free, and someare commercial products. A list of selected software can be found atkdnuggets.com/software/classification-rules.html. Some of them are standalone, andothers are add-on products for Excel. In fact, popular statistical packages such as SASand SPSS are adding the induction method as standard modules for data analysis.Popular ones include CART 5.0 (which incorporates classification costs into consider-ation, salford-systems.com/) and C4.5.

Most of these programs not only generate rules but also check them for possiblelogical conflict. All the programs are ES shells; that is, they can be used to generate therules and then to construct an ES that uses this knowledge. K-Vision even allows incor-poration of an entire induction table into a set of rules represented as a single node. Agood introduction to the method can be found in Roiger and Geatz (2003).

The free software PNC2 is available from newty.de/pnc2/.



INTERACTIVE INDUCTION

Instead of inducing rules from cases in a batch mode, knowledge can be induced byinteracting with experts incrementally. This is called interactive induction (see Jenget al., 1996). Interactive induction needs an expert supported by computer software.

One interesting tool that combines induction and interactive acquisition isACQUIRE from Acquired Intelligence, Inc. (aiinc.ca). ACQUIRE captures theknowledge of an expert through interactive interviews, distills it, and automaticallygenerates a rule-based knowledge base. ACQUIRE provides a variety of ways of cap-turing human knowledge, all of them focused on a qualitative representation of knowl-edge, and uses specific patterns of information to describe the situations experts canobserve or contemplate. Thus, ACQUIRE can capture both explicit and implicitknowledge.

ACQUIRE guides an expert through successive steps of knowledge structuring,yielding models of domain knowledge at varying levels of generality.The most general ofthese models is the object network, in which an object is anything that occurs, is consid-ered, or is concluded by the expert.The domain expert attributes a set of values or mean-ings to each object and organizes them into a network.The details of the individual rulesare completed by a number of methods, including action tables (i.e., induction tables)and IF-THEN-ELSE rules. Finally, the reasoning path is created, and conflicts amongreasoning paths are resolved. Contexts in which a rule is applied are also specified.

ACQUIRE interacts with experts (without knowledge engineers); it helps thembypass their cognitive defenses and biases and identify the important criteria and con-structs used in decision making. For details and demo systems, see aiinc.ca.


1. Define automated acquisition of knowledge and describe its objectives.

2. Describe the acquisition of documented knowledge.

3. Describe rule induction.

4. Describe the advantages of rule induction.

5. List difficulties in rule induction.

6. Describe interactive induction.

W18.7 KNOWLEDGE VERIFICATION AND VALIDATION

Knowledge acquired from experts needs to be evaluated for quality, including evalua-tion, validation, and verification. These terms are often used interchangeably. We usethe definitions provided by O’Keefe et al. (1987):

• Evaluation is a broad concept. Its objective is to assess an ES’s overall value. Inaddition to assessing acceptable performance levels, it analyzes whether the sys-tem would be usable, efficient, and cost-effective.

• Validation is the part of evaluation that deals with the performance of the system(e.g., as it compares to the expert’s). Simply stated, validation is building the rightsystem (i.e., substantiating that a system performs with an acceptable level ofaccuracy).

• Verification is building the system right or substantiating that the system is cor-rectly implemented to its specifications.


Knowledge Verification and Validation W-203 ◆

TABLE W18.5 Measures of Validation

Measure or Criterion Description

Accuracy How well the system reflects reality, how correct the knowledge is in the knowledge base

Adaptability Possibilities for future development, changesAdequacy (or Portion of the necessary knowledge included in the knowledge basecompleteness)Appeal How well the knowledge base matches intuition and stimulates

thought and practicabilityBreadth How well the domain is coveredDepth Degree of detailed knowledgeFace validity Credibility of knowledgeGenerality Capability of a knowledge base to be used with a broad range of

similar problemsPrecision Capability of the system to replicate particular system parameters,

consistency of advice, coverage of variables in knowledge baseRealism Accounting for relevant variables and relations, similarity to realityReliability Fraction of the ES predictions that are empirically correctRobustness Sensitivity of conclusions to model structureSensitivity Impact of changes in the knowledge base on quality of outputsTechnical and Quality of the assumed assumptions, context, constraints, and

operational validity conditions, and their impact on other measuresTuring test Ability of a human evaluator to identify whether a given conclusion is

made by an ES or by a human expertUsefulness How adequate the knowledge is (in terms of parameters and

relationships) for solving correctlyValidity The knowledge base’s capability of producing empirically correct

predictions

Sources: Compiled from B. Marcot, “Testing Your Knowledge Base,” AI Expert, August 1987; and R.M.O’Keefe, O. Balci, and E.P. Smith, “Validating Expert System Performance,” IEEE Expert, Winter 1987.

In the realm of ES, these activities are dynamic because they must be repeatedeach time the prototype is changed. In terms of the knowledge base, it is necessary toensure that we have the right knowledge base (i.e., that the knowledge is valid). It isalso essential to ensure that the knowledge base was constructed properly (i.e., verifi-cation).

In performing these quality-control tasks, we deal with several activities and con-cepts, as listed in Table W18.5. The process can be very difficult, considering the manysociotechnical issues involved (see Sharma and Conrath, 1992).

A method for validating ES, based on validation approaches from psychology, wasdeveloped by Sturman and Milkovich (1995). This approach tests the extent to whichthe system and the expert decisions agree, the inputs and processes used by an expertcompared to the machine, and the difference between expert and novice decisions.Ram and Ram (1996) described validation and verification techniques on specific ESfor innovative management.

Automated verification of knowledge is offered in the ACQUIRE productdescribed earlier. Verification is conducted by measuring the system’s performanceand is limited to classification cases with probabilities. It works as follows: When an ESis presented with a new case to classify, it assigns a confidence factor to each selection.By comparing these confidence factors with those provided by an expert, it is possible



to measure the accuracy of the ES for each case. By performing comparisons on manycases, it is possible to derive an overall measure of ES performance (see O’Keefe andO’Leary, 1993).

Rosenwald and Liu (1997) developed a validation procedure that uses the rulebase’s knowledge and structure to generate test cases that efficiently cover the entireinput space of the rule base. Thus, the entire set of cases need not be examined. A sym-bolic execution of a model of the ES is used to determine all conditions under whichthe fundamental knowledge can be used. For an extensive bibliography on validationand verification, see Juan et al. (1999). An easy approach to automating verification ofa large rule base can be found in Goldstein (2002).


1. Define evaluation, validation, and verification of knowledge.

2. Describe validation methods.

3. Describe automated verification of knowledge.

W18.8 REPRESENTATION OF KNOWLEDGE

Once validated, the knowledge acquired from experts or induced from a set of datamust be represented in a format that is both understandable by humans and exe-cutable on computers. There are many different methods for knowledge representa-tion. For an overview, see en.wikipedia.org/wiki/Knowledge_representation andHelbig (2006).The most popular method is the production rule. Frames, decision trees,objects, and logic are also useful in some cases. They are described next.

PRODUCTION RULES

Production rules are the most popular form of knowledge representation for ES andautomated decision support (ADS) systems. Knowledge is represented in the form ofcondition/action pairs: IF this condition (or premise or antecedent) occurs, THENsome action (or result or conclusion or consequence) will (or should) occur. Considerthese two examples:

• If the stop light is red AND you have stopped, THEN a right turn is okay.• If the client uses purchase requisition forms AND the purchase orders are

approved and purchasing is separate from receiving AND accounts payable ANDinventory records, THEN there is strongly suggestive evidence (90 percent proba-bility) that controls to prevent unauthorized purchases are adequate. (This exam-ple from an internal control procedure includes a probability.)

Each production rule in a knowledge base implements an autonomous chunk ofexpertise that can be developed and modified independently of other rules.When com-bined and fed to the inference engine, the set of rules behaves synergistically, yieldingbetter results than the sum of the results of the individual rules. In reality, knowledge-based rules are not independent. They quickly become highly interdependent. Forexample, a new rule that is added may conflict with an existing rule, or it may require arevision of attributes or rules.


Representation of Knowledge W-205 ◆

Rules can be viewed, in some sense, as a simulation of the cognitive behavior ofhuman experts. According to this view, rules are not just a neat formalism to representknowledge in a computer; rather, they represent a model of actual human behavior.

Rules can appear in different forms. Some examples follow:

• IF premise, THEN conclusion. For example, “If your income is high, THEN yourchance of being audited by the IRS is high.”

• Conclusion, IF premise. For example, “Your chance of being audited is high IFyour income is high.”

• Inclusion of ELSE. For example, “If your income is high OR your deductions areunusual, THEN your chance of being audited by the IRS is high, OR ELSE yourchance of being audited is low.”

• More complex rules. For example, “IF the credit rating is high AND the salary ismore than $30,000 OR assets are more than $75,000 AND pay history is not‘poor’, THEN approve a loan up to $10,000 and list the loan in Category B. Theaction part may include additional information: THEN approve the loan andrefer the applicant to an agent.”

The IF side of a rule can include dozens of IFs. The THEN side can include severalparts as well.

For further discussion of the use of production rules in active and deductive data-bases, see Helbig (2006). For an example of the use of production rules in an ES used inproduction environments, see Guth (1999).

KNOWLEDGE AND INFERENCE RULES

Two types of rules are common in artificial intelligence: knowledge and inference rules.Knowledge rules, or declarative rules, state all the facts and relationships about a prob-lem. Inference rules, or procedural rules, on the other hand, advise on how to solve aproblem, given that certain facts are known.The knowledge engineer separates the twotypes of rules: Knowledge rules go to the knowledge base, whereas inference rulesbecome part of the inference engine.

For example, assume that you are in the business of buying and selling gold. Theknowledge rules might look like this:

Rule 1: IF an international conflict begins, THEN the price of gold goes up.Rule 2: IF the inflation rate declines, THEN the price of gold goes down.Rule 3: IF the international conflict lasts more than seven days and IF it is in the

Middle East, THEN buy gold.

Inference rules contain rules about rules and thus are also called metarules.They per-tain to other rules (or even to themselves). Inference (procedural) rules may look like this:

Rule 1: IF the data needed are not in the system, THEN request them from theuser.

Rule 2: IF more than one rule applies, THEN deactivate any rules that add no newdata.

ADVANTAGES AND LIMITATIONS OF RULES

Rule representation is especially applicable when there is a need to recommend acourse of action based on observable events (see Application Case W18.7). It has sev-eral major advantages:



• Rules are easy to understand. They are communicable because they are a naturalform of knowledge.

• Inferences and explanations are easily derived.• Modifications and maintenance are relatively easy.• Uncertainty is easily combined with rules.• Each rule is often independent of all others.


Rule-Based Systems Tackle Employee Shrink

Innovative computer-based technologies are taking thebattle against employee-related shrink (i.e., theft) to ahigher level. Several new applications are aimed at identi-fying dishonest personnel, who account for an estimated38 percent of retail shrink, as well as those who may sim-ply be making errors at the point of sale.

PLATINUM Solutions offers a rule-based system thattargets employee theft at the retailer’s point of sale. “Therules are interrelated statements or business policies govern-ing what is allowed and what is not allowed,” explained CarlFijat, a management consultant with PLATINUM Solutions.

Rule-based software can help retailers find fraud pat-terns by collecting and storing information about incidentsoccurring at the point of sale and classifying them accordingto policies expressed as knowledge in the program. Forexample, it can be programmed to post “self-ringingemployee” alerts when employees ring up returns for them-selves.“Excessive credit to purchases” is posted when a staffmember processes purchases whose value exceeds a certainlimit and the account being credited is labeled “in-house.”

One of the products of PLATINUM Technologies (theparent company) is AionDS (now Cleverpath Aion from

Computer Associates), which models and encapsulatesretailers’ business policy logic into rules, Fijat explained.Each rule defines a premise and one or more resultingactions. Rules can be expressed in easy-to-understand,English-like language; this makes it easier for loss-preventionexecutives to communicate system requirements to thosedeveloping rule-based software, thus leading to more effec-tive knowledge representation and acquisition.

AionDS, and rule-based software in general, is partic-ularly valuable in isolating patterns that are not dis-cernible from reading exception reports. For example, ifan employee authorizes a legitimate cash sale, signs off,signs on under someone else’s identification number torefund the sale immediately, and then signs back on againas himself or herself, the system can use a rule to tracksuch fraudulent activity.

Sources: Modified and condensed from J.R. Ross,“New Rule-Based Systems Tackle Employee Shrink,” Stores,Vol. 78,No. 8,August 1996, pp. 71–72; and Computer Associates,www3.ca.com/solutions/Product.aspx?ID=5268 (accessedApril 2006).

The major limitations of rule representation are as follows:

• Complex knowledge requires thousands of rules, which may create difficulties inusing and maintaining the system.

• Builders like rules, so they try to force all knowledge into rules rather than lookfor more appropriate representations.

• Systems with many rules may have a search limitation in the control program.Some programs have difficulty evaluating rule-based systems and making inferences.

SEMANTIC NETWORKS

Semantic networks focus on the relationships between different concepts. They aregraphical depictions of knowledge composed of nodes and links that show hierarchicalrelationships between objects (see Russell and Norvig, 2002).



A simple semantic network is shown in Figure W18.6. It is made up of a number ofcircles, or nodes, that represent objects and descriptive information about the objects.An object can be any physical item, such as a book, a car, a desk, or even a person.Nodes can also be concepts, events, or actions. A concept might be the relationshipbetween supply and demand in economics, an event such as a picnic or an election, oran action such as building a house or writing a book.Attributes of an object can also beused as nodes. These might represent size, color, class, age, origin, or other characteris-tics. In this way, detailed information about objects can be presented.

Nodes are interconnected by links, or arcs, that show the relationships between thevarious objects and descriptive factors. Some of the most common arcs are of the IS-Aor HAS-A type. IS-A is used to show a class relationship (i.e., that an object belongs toa larger class or category of objects). HAS-A links are used to identify characteristicsor attributes of object nodes. Other arcs are used for definitional purposes.

As you can see in the example in Figure W18.6, the central figure in the domain ofknowledge is a person called Sam. One link shows that Sam is a man and that a man isa human being, or is part of a class called humans.Another arc from Sam shows that heis married to Kay. Additional arcs show that Kay is a woman, and that a woman is ahuman being. Other links show that Sam and Kay have a child, Joe, who is a boy and

Woman

Humanbeing

Boy

Is a

Is a

Is a

Is a

Is a

Is a

Man

Marriedto

VPSam

Kay

Has a child

Has a child

JoeGoes to

School

Needs

Food

Works forAcme

Ajax

Plays

Golf

Is a

a

Is a

Sport

Germany

Silver

MercedesBenz

Color

Made in

CarOwns Is a

Subsidiary of

FIGURE W18.6 Representation of Knowledge in a Semantic Network



goes to school. Some nodes and arcs show other characteristics about Sam. For exam-ple, he is a vice president of Acme, a company that is a subsidiary of Ajax. We also seethat Sam plays golf, which is a sport. Furthermore, Sam owns a Mercedes-Benz whosecolor is silver.We also see that Mercedes-Benz is a type of car that is made in Germany.

One of the most interesting and useful facts about a semantic network is that it canshow inheritance. Because a semantic network is basically a hierarchy, the variouscharacteristics of some nodes actually inherit the characteristics of others. As an exam-ple, consider the links showing that Sam is a man and that a man is in turn a humanbeing. Here, Sam inherits all the properties of human being. We can ask the question,“Does Sam need food?” Because of the inheritance links, we can say that he needsfood if human beings need food. See Mili and Pachet (1995) for details on a hierarchi-cal semantic network concept called regularity that subsumes inheritance. Mili andPachet presented novel ways in which semantic networks can support hypertext func-tionality, with a specific interest in generating argumentative documents (i.e., docu-ments that develop theses or prove assertions).

Semantic networks are used basically to provide visual representation of relation-ships and can be combined with other representation methods.

FRAMES

If we need to focus on the properties of certain objects, then using frames and objects isa good choice. A frame is a data structure that includes all the knowledge about a par-ticular object. This knowledge is organized in a special hierarchical structure that per-mits a diagnosis of knowledge independence. Frames are basically an application ofobject-oriented programming for artificial intelligence and ES. They are used exten-sively in ES. See Fensel et al. (1998) for a discussion of the frame-based KnowledgeAcquisition and Representation Language (KARL); see Jackson (1999) for a moredetailed description of frames and their uses; and see Chaudhri and Lowrance (1998)for details of the frame representation system and graphical browsing tool GKB-Editor.

Frames, as in frames of reference, provide a concise structural representation ofknowledge in a natural manner. In contrast to other representation methods, withframes, the values that describe one object are grouped together into a single unitcalled a frame. Thus, a frame encompasses complex objects, entire situations, or a man-agerial problem as a single entity. The knowledge in a frame is partitioned into slots. Aslot can describe declarative knowledge (e.g., the color of a car) or procedural knowl-edge (e.g., “activate a certain rule if a value exceeds a given level”). The following arethe major capabilities of frames:

• Ability to clearly document information about a domain model (e.g., a plant’smachines and their associated attributes)

• Related ability to constrain the allowable values that an attribute can take on• Modularity of information, permitting ease of system expansion and maintenance• Readability and consistency• Syntax for referencing domain objects in the rules• Platform for building a graphic interface with object graphics• Mechanism that allows the scope of facts considered during forward or backward

chaining to be restricted• Access to a mechanism that supports the inheritance of information down a class

hierarchy

A frame provides a means of organizing knowledge in slots that contain character-istics and attributes. In physical form, a frame is somewhat like an outline that has



Automobile Frame

Class of: TransportationName of manufacturer: AudiOrigin of manufacturer: GermanyModel: 5000 TurboType of car: SedanWeight: 3300 lb.Wheelbase: 105.8 inchesNumber of doors: 4 (defaults)Transmission: 3-speed automaticNumber of wheels: 4 (default)Engine: (Reference Engine Frame) • Type: In-line, overhead cam • Number of cylinders: 5Acceleration (procedural attachment) • 0–60: 10.4 seconds • Quarter mile: 17.1 seconds, 85 mphGas mileage: 22 mpg average (procedural attachment)Engine FrameCylinder bore: 3.19 inchesCylinder stroke: 3.4 inchesCompression ratio: 7.8 to 1Fuel system: Injection with turbochargerHorsepower: 140 hpTorque: 160 ft/LB

FIGURE W18.7 A Frame Describing an Automobile

categories and subcategories. A typical frame describing an automobile is shown inFigure W18.7. Note that the slots describe attributes such as name of manufacturer,model, origin of manufacturer, type of car, number of doors, engine, and other charac-teristics.

Content of a FrameA frame includes two basic elements: slots and facets. A slot is a set of attributes thatdescribe the object represented by the frame. For example, in the automobile frame,there are weight and engine slots. Each slot contains one or more facets. The facets(sometimes called subslots) describe some knowledge or procedural information aboutthe attribute in the slot. Facets can take many forms, including the following:

• Values. These facets describe attributes such as blue, red, and yellow for a color slot.• Default. This facet is used if the slot is empty (i.e., without any description). For

example, in the car frame, one default value is that the number of wheels on thecar is 4. It means that we can assume that the car has four wheels unless otherwiseindicated.

• Range. Range indicates what kind of information can appear in a slot (e.g., inte-ger numbers only, two decimal points, 0 to 100).

• If added. This facet contains procedural information or attachments. It specifiesan action to be taken when a value in the slot is added (or modified). Such proce-dural attachments are called daemons.



Vehicleframe

Trainframe

Boatframe

Carframe

Airplaneframe

Submarineframe

Busframe

Truckframe

Passengercar frame

Compactcar frame

Mid-sizecar frame

Sarah'scar frame

ToyotaCorolla frame

Brent'scar frame

FIGURE W18.8 Hierarchy of Frames Describing Vehicles

• If needed. This facet is used in a case in which no slot value is given. Much like theif-added situation, it triggers a procedure that goes out and gets or computes avalue.

• Other. Slots can contain frames, rules, semantic networks, or any other type ofinformation.

Certain procedures can be attached to slots and used to derive slot values. Forexample, slot-specific heuristics are procedures for deriving slot values in a particularcontext. An important aspect of such procedures is that they can be used to direct thereasoning process. In addition to filling slots, they can be triggered when a slot is filled.

In Figure W18.7, both acceleration and gas mileage are procedural attachments.They refer to a step-by-step procedure that defines how to acquire this information. Forexample, to determine acceleration, time needed to go from 0-to-60 mph, and quarter-mile elapsed time would be described. A procedural attachment to determine gasmileage would state a procedure for filling the gas tank, driving a certain numberof miles, determining the amount of gasoline used, and computing gas mileage in termsof miles per gallon.

Hierarchy and Inheritance of FramesMost artificial intelligence systems use a collection of frames linked together in a cer-tain manner to show their relationship. This is called a hierarchy of frames. For exam-ple, Figure W18.8 illustrates a hierarchy of frames that describes different kinds ofvehicles.

The hierarchical arrangement of frames permits inheritance frames. In FigureW18.8, the root of the tree is at the top, where the highest level of abstraction is repre-sented. Frames at the bottom are called leaves of the tree. The hierarchy permits inher-itance of characteristics. Each frame usually inherits the characteristics of all related



Name: Toyota Corolla

Slots

Owner ColorNo. of cylinders Range If needed

(a) Parent frame (b) Child frame

Check registration list List. per manufacturer

4 or 6 Ask owner

Sedan sport 2-4 doors Ask owner

1970-1995 Ask owner

Model Range If neededVintage (year) Range If needed

Owner

Color

No. of cylinders

Model

Vintage (year)

Brent

Blue

6

4D Sedan

1994

Facets Slots Facets

Name: Brent's car Instance of: Toyota Corolla frame

FIGURE W18.9 Parent and Child Frames

frames of higher levels. For example, a passenger car has the general properties of a car(e.g., an engine and a carburetor). These characteristics are expressed in the internalstructure of the frame. Inheritance is the mechanism for passing such knowledge, whichis provided in the value of the slots, from frame to frame.

Parent frames provide a more general description of the entities. More generaldescriptions are higher in the hierarchy. Parent frames contain the attribute definitions.When we describe actual physical objects, we instantiate the child’s frame. Theinstances (i.e., child frames) contain actual values of the attributes. An example isshown in Figure W18.9.

Note that every parent is a child of a higher-level parent. In building a frame, it ispossible to have a frame in which different slots are related to different parents. Theonly frame without a parent is the one at the top of the hierarchy.This frame is called amaster frame. It is the root frame, and it has the most general characteristics.The use offrame-based tools to build ES is demonstrated in tools such as CORVID (exsys.com).

OBJECT-ORIENTED KNOWLEDGE REPRESENTATIONS

The frame-based approach to knowledge representation has not been very popular inES design, but a follow-up approach that adopts many concepts from frames, known asthe object-oriented approach, has become a norm in system analysis and design.Clearly, hypermedia, especially over the Web, can exploit the object-orientedapproach. In this approach, different kinds of knowledge can be encapsulated inobjects, and these objects can communicate actions directly. Devedzic et al. (1996)described how knowledge can be represented as objects in ES, along with the reason-ing process. Their model, Reasoning Objects for Building Inference Engines (ROB-BIE), is based on a hierarchy in which each level is more detailed. At the highest levelare a blackboard and agents for interaction. As we move down the hierarchy, first the



system is defined, then its blocks, then its components, and finally the primitives thatform the pattern matching, conflict resolution, and so on.

Production rules and frames comprise a successful combination of knowledge rep-resentation methods. By themselves, production rules do not provide a totally effectiverepresentation facility for many ES applications. In particular, their expressive poweris inadequate for defining terms and for describing domain objects and static relation-ships among objects. In essence, frames are objects and fit the paradigm of an object-oriented approach to ES development rather well. Objects encapsulate properties andactions, just as frames store knowledge. Furthermore, frames are organized into classes,which are organized into hierarchies. Each frame inherits its properties from its parentframe, just as an object inherits properties. Rules can either guide the frames’ behav-iors or be embedded in the frame. (Some accounting applications using frames andrules adopt the latter approach.)

The object-oriented paradigm fits the hybrid ES structure well in working withframes and rules. Bogarin and Ebecken (1996) described an object-oriented approachof combining objects and rules for flexible pipe evaluation and design in the petroleumindustry.The objects select which sets of rules to apply to a given situation.The ES runsnumeric heuristics as well. CORVID and many other development tools have adoptedthe integrated approach. For applications that use CORVID as the development tool,check out exsys.com.

DECISION TABLES

Knowledge of relations can be represented in decision tables. In a decision table,knowledge is organized in a spreadsheet format, using columns and rows. The table isdivided into two parts. First, a list of attributes is developed, and for each attribute, allpossible values are listed. Then a list of conclusions is developed. Finally, the differentconfigurations of attributes are matched against the conclusion.

Knowledge for the table is collected in knowledge acquisition sessions. Once thetable is constructed, the knowledge in the table can be used as input to other knowl-edge representation methods. It is not possible to make inferences with the domaintables by themselves except when rule induction is used. Some ES shells can incorpo-rate an entire decision table into a single rule (e.g., K-Vision from GINESYSCorporation, ginesys.com). Decision tables are easy to understand and program.

DECISION TREES

Decision trees are related to decision tables and are popular in many places.These treesare similar to the decision trees used in decision theory. They are composed of nodesrepresenting goals and links representing decisions.

The major advantage of decision trees is that they can simplify the knowledge acqui-sition process. Knowledge diagramming is often more natural to experts than formal rep-resentation methods (e.g., rules, frames). For further discussion, see Vadera (2005).

Decision trees can easily be converted to rules. The conversion can be performedautomatically by a computer program. In fact, machine learning methods are capableof extracting decision trees automatically from textual sources and converting them torule bases.

In the example shown in Figure W18.10, the decision tree renders the knowledgeused to help diagnose the reasons for an automobile’s not starting.This problem can bedescribed as a number of knowledge rules, such as these:

Rule 1: IF the car does not start, THEN check whether the starter motor turns.



Yes

Yes No

No

Car starts?

Startermotorturns?

YesIs

therefuel?

YesSpark

at plugs?

YesIs fuel

reachingplugs?

Yes No

Is fuelreachingcarbu-retor?

Headlightswork?

Carburetorproblem

Blockage infuel line NoSpark

at coil?

Fault inHT circuit

Yes

Fault inLT circuit

No

Flat battery

No

Fill gas tank

Yes

Starter motorproblem

No

Problem withtiming

Sorry, only dononstarting

No

FIGURE W18.10 A Decision Tree for a Diagnosis of Car Malfunction

Rule 2: IF the starter motor turns, THEN check whether there is fuel in the tank;ELSE check that the headlights work.

Rule 3: IF the headlights do not work, THEN the battery is flat; ELSE there is astarter motor problem.

Predicate CalculusPredicate logic is useful for showing logic relationships and their reasoning. Facts andobservations in a problem domain are defined as premises, which are used by the logi-cal process to derive new facts and conclusions.

For a computer to perform reasoning using logic, some method must be used toconvert statements and the reasoning process into a form suitable for manipulation bya computer. The result is what is known as a symbolic logic system of rules and proce-dures that permit the drawing of inferences from premises, using logical techniques.



The two basic forms of computational logic are propositional logic (or proposi-tional calculus) and predicate logic (or predicate calculus). A proposition is nothingmore than a statement that is either true or false.When we know what a proposition is,it becomes a premise that can be used to derive new propositions or inferences. Rulesare used to determine the truth (T) or falsity (F) of the new proposition.

In propositional logic (or propositional calculus), we use symbols, such as letters ofthe alphabet, to represent propositions, premises, or conclusions. For example, considerthe propositions used in this simple deductive process:

Statement: A—The mail carrier comes Monday through Friday.Statement: B—Today is Sunday.Conclusion: C—The mail carrier will not come today.

Simple single propositions of this kind are not very interesting or useful. Real-world problems involve many interrelated propositions. To form more complexpremises, two or more propositions can be combined, using logical connectives such asAND, OR, NOT, IMPLIES, and EQUIVALENT. The resulting symbolic expressionlooks very much like a math formula. It can be manipulated using the rules of proposi-tional logic to infer new conclusions.

Because propositional logic deals primarily with complete statements and whetherthey are true or false, its ability to represent real-world knowledge is limited.Consequently, artificial intelligence uses predicate logic instead.

Predicate logic (or calculus) permits you to break a statement down into its com-ponent parts (i.e., an object, a characteristic of the object, or some assertion about theobject). In addition, predicate calculus lets you use variables and functions of variablesin a symbolic logic statement. Predicate logic is the basis for the artificial intelligencelanguage called PROLOG (programming in logic). (Note: Predicate refers to the verbpart of a sentence. In PROLOG, the action or relationship between objects is statedfirst.) For example, in PROLOG, we represent the fact that the mail carrier will arriveon Monday as arrive(mailcarrier, Monday), and we represent the rule “If the mailarrives on Monday, the case is acceptable,” as [accept :- arrive(mailcarrier, Monday)].(Note that the symbol :- stands for “is derived from.”) It is the same as in productionrules, except that the conclusion is stated first. In fact, predicate logic provides the the-oretical foundation for rule-based ES. Facts such as this one, along with rules expressedwithin the language, form the basis for inferencing. For further details, refer to Puls(2002).

Other Knowledge Representation MethodsSeveral other knowledge representation methods have been proposed and used. Forexample, Yan (2006) proposed a new complicated-knowledge representation methodfor the self-reconfiguration of complex systems such as complex software systems,complex manufacturing systems, and knowledgeable manufacturing systems. It isbased on a knowledge mesh and an agent mesh, along with a new knowledgemesh–based approach to complicated-knowledge representation.

The recent fashion in knowledge representation languages is to use XML as thelow-level syntax.This tends to make the output of these knowledge representation lan-guages easy for machines to parse, at the expense of human readability. First-orderpredicate calculus, discussed earlier, is commonly used as a mathematical basis forthese systems, to avoid excessive complexity. However, even simple systems basedon this simple logic can be used to represent data that is well beyond the processingcapability of current computer systems.



TABLE W18.6 Advantages and Disadvantages of Different Knowledge Representations

Scheme Advantages Disadvantages

Production Simple syntax, easy to understand, Hard to follow hierarchies,rules simple interpreter, highly modular, inefficient for large systems, not

flexible (easy to add to or modify) all knowledge can be expressed as rules, poor at representing structured descriptive knowledge

Semantic Easy to follow hierarchy, easy to trace Meaning attached to nodes might be networks associations, flexible ambiguous, exception handling is

difficult, difficult to programFrames Expressive power, easy to set up slots Difficult to program, difficult for

for new properties and relations, inference, lack of inexpensive easy to create specialized procedures, softwareeasy to include default information and easy detect missing values

Formal Facts asserted independently of use, Separation of representation and logic assurance that only valid processing, inefficient with large

consequences are asserted data sets, very slow with large (precision), completeness knowledge bases

The following are examples of artificial languages intended for knowledge repre-sentation:

• CycL• Loom• OWL• KM : the Knowledge Machine (a frame-based language used for knowledge repre-

sentation work)

For details on knowledge representation, see en.wikipedia.org/wiki/Knowledge_representation.

SELECTION OF REPRESENTATION METHODS

Representation methods all have advantages and limitations (see Table W18.6).Production rules are popular in the design of first-generation ES. The object-orientedmethod has become very popular in recent years. Predicate logic provides a theoreticalfoundation for rule-based inferences.

Knowledge representation should support the tasks of acquiring and retrievingknowledge as well as subsequent reasoning. Several factors must be taken into accountin evaluating knowledge representation for the above three tasks:

• Naturalness, uniformity, and understandability of the representation.• Degree to which knowledge is explicit (declarative) or embedded in procedural

code. For further information on implicit/explicit knowledge, see Krogh et al. (2000).• Modularity and flexibility of the knowledge base.• Efficiency of knowledge retrieval and the heuristic power of the inference proce-

dure. (Heuristic power is the reduction of the search space achieved by a heuristicmechanism.)

Sometimes, no single knowledge representation method is by itself ideally suitedfor all tasks. When several sources of knowledge are used simultaneously, the goal ofuniformity may have to be sacrificed in favor of exploiting the benefits of multipleknowledge representations, each tailored to a different subtask. The necessity of



translating among knowledge representations becomes a problem in these cases.Nevertheless, some recent ES shells use two or more knowledge representationschemes (e.g., the CORVID system, exsys.com).


1. Define production rules and knowledge/inference rules.

2. List the major advantages and limitations of rules.

3. Define semantic network.

4. Define frame and describe the advantages of frames.

5. Describe object-oriented knowledge representations.

6. Define decision tree and decision table.

7. Explain predicate calculus.

W18.9 REASONING IN INTELLIGENT SYSTEMS

When the knowledge representation in the knowledge base is completed, or is at leastat a sufficiently high level of accuracy, it is ready to be used.

The use of knowledge is the key to the input of intelligent systems. It is usuallydone with some kind of reasoning mechanism. The reasoning capability is what givesmachines their intelligence. Several methods of reasoning exist, some of which are cus-tomized (e.g., see Helvacioglu and Insel, 2005).

COMMONSENSE REASONING

Commonsense reasoning is the branch of artificial intelligence concerned with repli-cating human thinking. There are several components to commonsense reasoning,including the following:

• Developing adequately broad and deep commonsense knowledge bases• Developing reasoning methods that exhibit the features of human thinking, such

as the ability to reason with knowledge that is true by default, the ability to reasonrapidly across a broad range of domains, and the ability to tolerate uncertainty inknowledge

• Developing new kinds of cognitive architectures that support multiple reasoningmethods and representations

For details on commonsense reasoning, see en.wikipedia.org/wiki/commonsense_reasoning.

REASONING IN RULE-BASED SYSTEMS

Inferencing requires some computer program to access the stored knowledge and useit for making inferences.This program is an algorithm that controls a reasoning processand is usually called the inference engine or the control program. In rule-based sys-tems, it is also called the rule interpreter.

The inference engine directs the search through the knowledge base, a process thatmay involve the application of inference rules in what is called pattern matching. Thecontrol program decides which rule to investigate, which alternative to eliminate, andwhich attribute to match. The most popular control programs for rule-based systems,forward and backward chaining, are described in the next section.


Reasoning in Intelligent Systems W-217 ◆

TABLE W18.7 Reasoning Methods

Method Description

Deductive reasoning Move from a general principle to a specific inference. A general principle is composed of two or more premises.

Inductive reasoning Move from some established facts to draw general conclusions.Analogical reasoning Derive an answer to a question by known analogy. Analogical

reasoning is a verbalization of an internalized learning process(see Owen, 1990). It involves use of similar past experiences.

Formal reasoning Syntactically manipulate a data structure to deduce new facts,following prescribed rules of inferences (e.g., predicate calculus).

Procedural (numeric) Use mathematical models or simulation (such as model-based reasoning reasoning, qualitative reasoning, and temporal reasoning, or the

ability to reason about the time relationships between events).Metalevel reasoning Have knowledge about what is known (e.g., about the importance

and relevance of certain facts and rules).

Before we investigate the specific inferencing techniques used in ES, we examinehow people reason, which is what artificial intelligence attempts to mimic. There areseveral ways people reason and solve problems. An interesting view of the problem-solving process is one in which people draw on “sources of power.” Lenat (1982) iden-tified nine such sources:

1. Formal reasoning methods (e.g., logical deduction)2. Heuristic reasoning, or IF-THEN rules3. Focus, or commonsense applied to specific goals4. Divide and conquer, or dividing complex problems into subproblems (sometimes

called chunking)5. Parallelism—that is, neural processors (perhaps a million) operating in parallel6. Representation, or ways of organizing pieces of information7. Analogy, or the ability to associate and relate concepts8. Synergy, in which the whole is greater than the sum of its parts9. Serendipity, or fortuitous accidents

These sources of power range from the purely deductive reasoning best handledby computer systems to inductive reasoning, which is more difficult to computerize.Lenat (1982) said that the future of artificial intelligence relied on finding ways to tapsources that were only beginning to be exploited. These sources of power are trans-lated to specific reasoning or inference methods, as summarized in Table W18.7.

Inferencing with rules involves the implementation of modus ponens (simple argu-ment form), which is reflected in the search mechanism. Consider the following example:

Rule 1: IF an international conflict begins,THEN the price of gold will go up.

Let us assume that the ES knows that an international conflict has just started.Thisinformation is stored as a fact in the database (or assertion base), which means that thepremise (the IF part) of the rule is true. With modus ponens, the conclusion is thenaccepted as true. We say that Rule 1 fires. Firing a rule occurs only when all the rule’shypotheses (i.e., conditions in the IF part) are satisfied (i.e., evaluated to be true).Thenthe conclusion drawn is stored in the assertion base. In our case, the conclusion (i.e., theprice of gold will go up) is added to the assertion base and can be used to satisfy the



premises of other rules. The true (or false) values for either portion of the rules can beobtained by querying the user or by checking other rules.Testing a rule premise or con-clusion can be as simple as matching a symbolic pattern in the rule to a similar patternin the assertion base. This activity is called pattern matching.

Every rule in a knowledge base can be checked to see whether its premise or con-clusion can be satisfied by previously made assertions. This process can be done in oneof two directions, forward or backward, and it continues until no more rules can fire oruntil a goal is achieved.

There are two methods for controlling inference in rule-based ES: forward chain-ing and backward chaining (each of which has several variations). First, we provideintuitive descriptions of the two methods; then we discuss them in detail.

Example 1 What Flights Should I Take to Get to Tokyo?

Suppose you want to fly from Denver to Tokyo, and there are no direct fights between the twocities. Therefore, you try to find a chain of connecting flights, starting from Denver and end-ing in Tokyo. There are two basic ways you can search for this chain of flights:

• Start with all the flights that arrive in Tokyo and find the city where each flight originated.Then look up all the flights arriving at those cities and determine where they originated.Continue the process until you find Denver. Because you are working backward from yourgoal (Tokyo), this search process is called backward chaining (i.e., it is goal driven).

• List all the flights leaving Denver and note their destination cities. Then look up all theflights leaving those cities and find where they land; continue this process until you findTokyo. In this case, you are working forward from Denver toward your goal, so this searchprocess is called forward chaining (i.e., it is data driven).

This example demonstrates the importance of heuristics in expediting the search process.Going either backward or forward, you can use heuristics to make the search more efficient.For example, in the backward-chaining approach, you can look only at flights that go west-ward. Depending on the goals of your trip (e.g., minimize cost, minimize travel time, maxi-mize stopovers), you can develop additional rules to expedite the search further.

Example 2 Why Doesn’t the Car Start?

Suppose your car does not start. Is it because you are out of gas? Or is it because the starter isbroken? Or is it because of some other reason? Your task is to find out why the car won’t start.From what you already know (the consequence—the car won’t start), you go backward andtry to find the condition that caused it. This is a typical application of ES in the area of diag-nosis (i.e., the conclusion is known, and one or more of the causes are sought).

A good example of forward chaining is a situation in which a water system is overheating.Here, the goal is to predict the most likely reason. After reviewing the rules and checkingadditional evidence, you can finally find the answer. In forward chaining, you start with a con-dition or a symptom that is given as a fact.

As we show later, the search process in both cases involves a set of knowledge rules. Afterdetermining which rules are true and which are false, the search ends with a finding. Theword chaining signifies the linking of a set of pertinent rules.

The search process is directed by what is sometimes called a rule interpreter, which works asfollows:• In forward chaining, if the premise clauses match the situation, then the process attempts

to assert the conclusion.• In backward chaining, if the current goal is to determine the correct conclusion, then the

process attempts to determine whether the premise clauses (i.e., facts) match the situation.

Backward ChainingBackward chaining is a goal-driven approach in which you start from an expectation ofwhat is going to happen (i.e., hypothesis) and then seek evidence that supports (or con-tradicts) your expectation. Often, this entails formulating and testing intermediatehypotheses (or subhypotheses).



On a computer, the program starts with a goal to be verified as either true or false.Then it looks for a rule that has that goal in its conclusion. It then checks the premise ofthe rule in an attempt to satisfy the rule. It examines the assertion base first. If thesearch fails there, the program looks for another rule that has the same conclusion asthe first rule. An attempt is then made to satisfy the second rule. The process continuesuntil all the possibilities that apply are checked or until the rule initially checked (withthe goal) is satisfied. If the goal is proven false, then the next goal is tried. (In some infer-encing, the rest of the goals can be tried in succession, even if the goal is proven true.)

Example 3 Backward Chaining: Should I Invest in IBM Stock?

Here is an example involving an investment decision about whether to invest in IBM stock.The following variables are used:

A = Have $10,000B = Younger than 30C = Education at college levelD = Annual income of at least $40,000E = Invest in securitiesF = Invest in growth stocksG = Invest in IBM stock (the potential goal)

Each of these variables can be answered as true (yes) or false (no).The facts: We assume that an investor has $10,000 (i.e., that A is true) and that she is

25 years old (i.e., that B is true). She would like advice on investing in IBM stock (yes or nofor the goal).

The rules: Our knowledge base includes these five rules:

R1: IF a person has $10,000 to invest and she has a college degree, THEN she should invest in securities.

R2: IF a person’s annual income is at least $40,000 and she has a college degree, THEN she should invest in growth stocks.

R3: IF a person is younger than 30 and she is investing in securities, THEN she should invest in growth stocks.

R4: IF a person is younger than 30 and older than 22, THEN she has a college degree

R5: IF a person wants to invest in a growth stock, THEN the stock should be IBM.

These rules can be written as follows:

R1: IF A and C, THEN E.R2: IF D and C, THEN F.R3: IF B and E, THEN F.R4: IF B, THEN C.R5: IF F, THEN G.

Our goal is to determine whether to invest in IBM stock. With backward chaining, we startby looking for a rule that includes the goal (G) in its conclusion (THEN) part. Because R5 is theonly one that qualifies, we start with it. If several rules contain G, then the inference enginedictates a procedure for handling the situation. This is what we do:

Step 1: Try to accept or reject G. The ES goes to the assertion base to see whether G is there.At present, all we have in the assertion base is A is true. B is true. Therefore, the ESproceeds to step 2.

Step 2: R5 says that if it is true that we invest in growth stocks (F), then we should invest inIBM (G). If we can conclude that the premise of R5 is either true or false, then wehave solved the problem. However, we do not know whether F is true. What shall we



D and

F2 1

G

R5

R2

R4

R3

3

4

R15

R47

6

B C

andB B&E

A&CandA E

B C

C&D

or

do now? Note that F, which is the premise of R5, is also the conclusion of R2 andR3. Therefore, to find out whether F is true, we must check either of these two rules.

Step 3: We try R2 first (arbitrarily); if both D and C are true, then F is true. Now we have aproblem. D is not a conclusion of any rule, nor is it a fact. The computer can eithermove to another rule or try to find out whether D is true by asking the investor forwhom the consultation is given if her annual income is above $40,000.What the ES does depends on the search procedures used by the inference engine.Usually, a user is asked for additional information only if the information is not avail-able or cannot be deduced. We abandon R2 and return to the other rule, R3. Thisaction is called backtracking (i.e., knowing that we are at a dead end, we try some-thing else; the computer must be preprogrammed to handle backtracking).

Step 4: Go to R3; test B and E. We know that B is true because it is a given fact. To prove E,we go to R1, where E is the conclusion.

Step 5: Examine R1. It is necessary to determine whether A and C are true.Step 6: A is true because it is a given fact. To test C, it is necessary to test R4 (where C is the

conclusion).Step 7: R4 tells us that C is true (because B is true). Therefore, C becomes a fact (and is

added to the assertion base). Now E is true, which validates F, which validates ourgoal (i.e., the advice is to invest in IBM).A negative response to any of the preceding statements would result in a “Do notinvest in IBM stock” response. Then another investment decision (i.e., conclusion)would have to be considered, but this time, all the facts derived from the previoustries would be used.

Note that during the search, the ES moved from the THEN part to the IF part to the THENpart, and so on (see Figure W18.11). This is a typical search pattern in backward chaining. Aswe show next, forward chaining starts with the IF part, moves to the THEN part, then toanother IF part, and so on. Some systems allow a change in the direction of the search mid-stream; that is, they can go from THEN to THEN (or from IF to IF) as needed.

Forward ChainingForward chaining is a data-driven approach. We start from available information as itbecomes available or from a basic idea, and then we try to draw conclusions. The ESanalyzes the problem by looking for the facts that match the IF part of its IF-THENrules. For example, if a certain machine is not working, the computer checks the elec-tricity flow to the machine. As each rule is tested, the program works its way towardone or more conclusions.

FIGURE W18.11 Backward Chaining



D and

F4

G

R5

R2

R4

R3

1

3

R12

R41

B C

andB B&E

A&CandA E

B C

C&D

or

Example 4 Forward Chaining: Should I Invest in IBM Stock?

Let us use the same example we examined in backward chaining. Here we reproduce therules:R1: IF A and C, THEN E.R2: IF D and C, THEN F.R3: IF B and E, THEN F.R4: IF B, THEN C.R5: IF F, THEN G.

and the facts:A is true (the investor has $10,000).B is true (the investor is younger than 30).

In forward chaining (see Figure W18.12), we start with known facts and derive new facts byusing rules having known facts on the IF side.

Step 1: Because it is known that A and B are true, the ES starts deriving new facts by usingrules that have A and B on the IF side. Using R4, the ES derives a new fact C andadds it to the assertion base as true.

Step 2: R1 fires (because A and C are true) and asserts E as true in the assertion base.Step 3: Because B and E are both known to be true (they are in the assertion base), R3 fires

and establishes F as true in the assertion base.Step 4: R5 fires (because F is on its IF side), which establishes G as true. So the ES recom-

mends an investment in IBM stock. If there is more than one conclusion, more rulesmay fire, depending on the inferencing procedure.

Forward Chaining or Backward Chaining: Which to Use?We have seen that an antecedent–consequence rule system can run forward or back-ward, but which direction is better? The answer depends on the purpose of the reason-ing and the shape of the search space. For example, if the goal is to discover everythingthat can be deduced from a given set of facts, the system should run forward, asin accounting audit applications, because most facts are initially available in docu-ments and forms. In some cases, the two strategies can be mixed (i.e., the process isbidirectional).

The execution of forward or backward chaining is accomplished with the aid of arule interpreter. Its function is to examine production rules to determine which arecapable of being fired and when to fire them.The control strategy of the rule interpreter

FIGURE W18.12 Forward Chaining



(e.g., backward chaining) determines how the appropriate rules are found and when toapply them.

Inferencing with rules (as well as with logic) can be very effective, but there aresome obvious limitations to these techniques. One reason for this is summarized by thefamiliar axiom that there is an exception to every rule. For example, consider the fol-lowing argument:

Proposition 1: Birds can fly.Proposition 2: An ostrich is a bird.Conclusion: An ostrich can fly.

The conclusion is perfectly valid but is false for this limited-knowledge universe;ostriches simply do not cooperate with the facts. For this reason, as well as for a moreefficient search, we sometimes use other inferencing methods.

INFERENCE TREES

An inference tree (also called a goal tree or logical tree) provides a schematic view ofthe inference process. It is similar to a decision tree and an influence diagram(see Russell and Norvig, 2002). Note that each rule is composed of a premise and aconclusion.

In building an inference tree, the premises and conclusions are shown as nodes.The branches connect the premises and the conclusions. The operators AND and ORare used to reflect the structures of the rules. There is no deep significance to the con-struction of such trees—they just provide better insight into the structure of the rules.Inference trees map the knowledge in a convenient manner. In fact, the current gener-ation of Windows-based ES shells uses a kind of inference tree to display the knowl-edge base on the screen (e.g., G2, CORVID).

Figure W18.13 presents an inference tree for Example 3. Using the tree, we canvisualize the process of inference and movement along its branches. This is called treetraversal. To traverse an AND node, we must traverse all the nodes below it. To tra-verse an OR node, it is sufficient to traverse just one of the nodes below it. The infer-ence tree is constructed upside-down:The root is at the top (i.e., end), and the branchespoint downward.The tree starts with “leaves” at the bottom. (It can also be constructedfrom left to right, much like a decision tree.) Inference trees are basically composed ofclusters of goals. Each goal can have subgoals (i.e., children) and a supergoal (i.e., par-ent).

Single inference trees are always a mixture of AND nodes and OR nodes; they areoften called AND/OR trees. (Note that the NOT operation is allowed.) The AND nodesignifies a situation in which a goal is satisfied only when all its immediate subgoals aresatisfied. The OR node signifies a situation in which a goal is satisfied when any of itsimmediate goals are satisfied. When enough subgoals are satisfied to achieve the pri-mary goal, the tree is said to be satisfied. The inference engine contains procedures forexpressing this process as backward or forward chaining. These procedures are orga-nized as a set of instructions involving inference rules. They aim at satisfying the infer-ence tree and collectively contribute to the process of goal (i.e., problem) reduction.For further discussion, see Russell and Norvig (2002).

An inference tree has another big advantage: It provides a guide for answering thewhy and how questions in the explanation process. Users ask the how question whenthey want to know how a certain conclusion has been reached. The computer followsthe logic in the inference tree, identifies the goal (i.e., conclusion) involved in it and theAND/OR branches, and reports the immediate subgoals. Users ask the why question


Explanation and Metaknowledge W-223 ◆

when they want to know why the computer requests certain information as input (inbackward chaining). To deal with why questions, the computer identifies the goalinvolved with the computer-generated query and reports the immediate subgoals.


1. Define inferencing and inference engine.

2. Define forward chaining and backward chaining.

3. Describe inference trees.

W18.10 EXPLANATION AND METAKNOWLEDGE

A final feature of expert systems is the interactivity with the user and the capacity toprovide an explanation consisting of the sequence of inferences that were made by thesystem in arriving at a conclusion. This feature offers a means of evaluating theintegrity of the system when it is to be used by the experts themselves. Two basic typesof explanations are the why and the how. Metaknowledge is knowledge about knowl-edge. It is a structure within the system using the domain knowledge to accomplish the

GRoot

Branches

R5

F

OR

C

Dead end Fact Fact

D B

R3R2

C&D

AND

B&E

E

R1

R4

R4

Fact Fact

B C BA

AND

AND

FIGURE W18.13 An Inference Tree



system’s problem-solving strategy.This section deals with different methods used in ESfor generating explantions.

EXPLANATION

Human experts are often asked to explain their views, recommendations, or decisions.If ES are to mimic humans in performing highly specialized tasks, they too need to jus-tify and explain their actions. An explanation is an attempt by an ES to clarify its rea-soning, recommendations, or other actions (e.g., asking a question). The part of an ESthat provides explanations is called an explanation facility (or justifier). The explana-tion facility has several purposes:

• Make the system more intelligible to the user.• Uncover the shortcomings of the rules and knowledge base (i.e., debugging of the

systems by the knowledge engineer).• Explain situations that were unanticipated by the user.• Satisfy psychological and social needs by helping the user feel more assured

about the actions of the ES.• Clarify the assumptions underlying the system’s operations to both the user and

the builder.• Conduct sensitivity analyses. (Using the explanation facility as a guide, the user

can predict and test the effects of changes on the system.)

Explanation in rule-based ES is usually associated with a way of tracing the rulesthat are fired during the course of a problem-solving session. This is about the closestto a real explanation that today’s systems come, given that their knowledge is usuallyrepresented almost exclusively as rules that do not include basic principles necessaryfor a human-type explanation.

Programs such as MYCIN replay the exact rule used when asked for an explana-tion. DIGITAL ADVISOR is a slight improvement over this. Instead of feeding backthe rule verbatim, ADVISOR determines the generic principle on which the rule isbased (at that point in the consultation) and displays it. CORVID provides the rulesin the chain of reasoning leading to the conclusion that is hypothesized as the cur-rent goal.

In developing large ES, a good explanation facility is essential. Large ES alwaysinclude more facts and rules than a person can easily remember. Often, a new ruleadded during ES development interacts with other rules and data in unanticipatedways and can make the ES display strange explanations.

Explanation is an extremely important function because understanding dependson explanation, thus making implementation of proposed solutions easier. For exam-ple, an experiment that Ye and Johnson (1995) conducted with auditors indicated thatES explanation facilities can make a system’s advice more acceptable to users. Ye andJohnson also found that justification is the most effective type of ES explanation forbringing about a change in users’ attitudes toward a system. Also, explanation is essen-tial in ES used for training purposes.

To understand an event or a situation, you have to explain to yourself exactly whatis going on, the role of every participant, and their actions. You need to know the goalsof the participants, their intentions, and their potential behavior. Thus, constructingexplanations can become a very complex task, especially when it is done by machines.For this reason, many of the explanation facilities available in development tools pro-vide only two basic types of explanation: the why and the how.


Explanation and Metaknowledge W-225 ◆

Why ExplanationsA user poses a typical why question to a computer after the computer asks the user toprovide some information. For example, using our earlier investment example, wemight have the following dialog:

Computer: What is your annual income?Client: Why? (Why do you need to know?)

The computer displays R2 as a response. However, ideally, we would like to see areply like this:

Computer: In checking R2, I need to know whether your income is above $40,000.If this is true, I will conclude that because you have a college degree,you should invest in growth stocks.

The ES asks for this input when it is unable to find income information in the con-clusion part of any rule. In the example cited, the computer actually backtracked to R3and successfully arrived at a final conclusion.

How ExplanationsA user poses a typical how question when he or she would like to know how a certainconclusion or recommendation was reached. Simple systems are limited to the finalconclusion. More complex systems can handle intermediate conclusions as well. Thesystem can explain why a certain rule was fired; that is, it shows the chain of rules usedto reach the conclusion. In the investment problem, the how explanation might looklike this:

Computer: Invest in IBM stocks.Client: How? (How was the conclusion reached?)Computer: (Displays the rules R5–R3–R7–R4 in the sequence in which they were

fired.)Computer: Given that you have $10,000 to invest and you are younger than 30,

then according to R4, you have a college degree. If this is the case, thenaccording to R1, you should invest in securities. For a young investorlike you, according to R3, you should invest in growth stocks if you aregoing to invest at all. Finally, according to R5, if you need to invest ingrowth stocks, then IBM is your best bet.

The why and how explanations often show the rules as they were programmed andnot in a natural language. However, some systems have the capability to present theserules in natural language.

Other ExplanationsWick and Slagle (1989) expanded on the two basic questions (i.e., how and why) byproposing a “journalistic” explanation facility. They viewed the user–system interac-tion as an event to be reported much like a news story, which ideally includes the sixelements who, what, where, when, why, and how. When an explanation facility isdesigned along these lines, all bases are covered.

Some sophisticated ES supply other explanations. For example, some systems pro-vide a limited why not capability. Let us assume that the system in our investmentproblem selects IBM as a growth stock.The user may ask,“Why not GE?” and the sys-tem may answer, “Because the annual growth rate of GE is only 7 percent, whereasthat of IBM is 11 percent, using R78.” Then R78 might be displayed. This example



illustrates a possible connection between the ES and a regular database. The computermay need to access a database to provide explanations. In some special problems inexplaining the reasoning process, rules are (automatically) induced. Most such systemsonly display the rules and leave the real explanation up to the user.

Explanation is much more difficult in non-rule-based systems than in rule-basedones because the inference procedures are more complex. For an overview of the topicof explanation, see Cawsey (1995). For recent advances, see issues of Expert Systemswith Applications.

METAKNOWLEDGE

There are different methods for generating explanations. One easy way is to preinsertpieces of English text (i.e., scripts) into the system. For example, each question that canbe asked by the user can have an answer text associated with it. This is called a staticexplanation. Several problems are associated with static explanations. For example, allquestions and answers must be anticipated in advance, and for large systems, this is verydifficult. Also, the system essentially has no idea about what it is saying. Over time, theprogram might be modified without changing the text, thus causing inconsistency.

A better form of explanation is a dynamic explanation, which is reconstructedaccording to the execution pattern of the rules. With this method, the system recon-structs the reasons for its actions as it evaluates rules.

Most current explanation facilities fail to meet some of the objectives and require-ments listed earlier. Most ES explanation facilities consist of printing out a trace of therules being used. Explanation is not treated as a task that requires intelligence in itself.If ES are to provide satisfactory explanations, future systems must include not onlyknowledge of how to solve problems in their respective domains but also knowledge ofhow to effectively communicate to users their understanding of this problem-solvingprocess. Obviously, the balance between these two types of knowledge varies accordingto the primary function of the system. Constructing such knowledge bases involves for-malizing the heuristics used in providing good explanations.

With current ES, much of the knowledge vital to providing a good explanation(e.g., knowledge about the system’s problem-solving strategy) is not expressed explic-itly in the rules. Therefore, the purely rule-based representation may be difficult tograsp, especially when the relationships between the rules are not made explicit in theexplanation. Kidd and Cooper (1985) developed an explanation facility that can showthe inference tree and the parts of it that are relevant to specific queries, thus over-coming some of the deficiencies cited earlier.

In keeping with Kidd and Cooper’s (1985) concepts of explanation, Ye andJohnson (1995) provided a categorization of the explanation methods. Their typologyof ES explanations includes the following:

• Trace, or line of reasoning. This is a record of the inferential steps taken by an ESto reach a conclusion.

• Justification. This is an explicit description of the causal argument or rationalebehind each inferential step taken by the ES (based on empirical associationsinvolving the encoding of large chunks of knowledge).

• Strategy. This is a high-level goal structure that determines how the ES uses itsdomain knowledge to accomplish a task (or metaknowledge).

Justification requires a deeper understanding of the domain than the trace methoddoes. Demonstrating that the conclusions developed by the system are based on soundreasoning increases the confidence of ES users in the problem-solving ability of the


Inferencing with Uncertainty W-227 ◆

system and hence the acceptability of the conclusions. Because ES can achieve highperformance levels only within narrow problem areas, justification enables users tomake more informed decisions on whether to follow the advice. Strategy involvesknowledge about the problem-solving procedure. Generally, the last two explanationtypes are more difficult to incorporate in an ES because strategic knowledge tends tobe buried implicitly in the knowledge base, and an ES does not need an explicit repre-sentation of justification knowledge in order to execute (see Ye and Johnson, 1995).Future ES may provide these advanced explanation capabilities.

A system’s knowledge about how it reasons is called metaknowledge, or knowledgeabout knowledge. The inference rules presented earlier are a special case of meta-knowledge. Metaknowledge allows the system to examine the operation of the declara-tive and procedural knowledge in the knowledge base.

Explanation can be viewed as another aspect of metaknowledge. Over time, meta-knowledge will allow ES to do even more. ES will be able to create the rationalebehind individual rules by reasoning from first principles. They will tailor their expla-nations to fit the requirements of their audience. And they will be able to change theirown internal structure through rule correction, reorganization of the knowledge base,and system reconfiguration.


1. Describe explanation facility.

2. Why do we need explanations?

3. What type of explanations exit?

4. Describe metaknowledge.

W18.11 INFERENCING WITH UNCERTAINTY

Uncertainty is an important component in ES. According to Parsaye and Chignell(1988), uncertainty is treated as a three-step process in artificial intelligence, as shownin Figure W18.14. In step 1, an expert provides inexact knowledge in terms of rules withlikelihood values. These rules can be numeric (e.g., a probability value), graphical, orsymbolic (“It is most likely that”).

Representation ofuncertainty of

basic set of eventsStep 1

Step 2 Step 3Combining bodies

of uncertaininformation

Drawinginferences

Alternativeroute

FIGURE W18.14 A Three-Step Process for Dealing with Uncertainty in Artificial Intelligence



In step 2, the inexact knowledge of the basic set of events can be directly used todraw inferences in simple cases (step 3). However, in many cases, the various events areinterrelated. Therefore, it is necessary to combine the information provided in step 1into a global value for the system. Several methods can be used for such integration.The major methods are Bayesian probabilities, theory of evidence, certainty factors,and fuzzy sets.

In step 3, the purpose of the knowledge-based system is to draw inferences. Theseare derived from the inexact knowledge of steps 1 and 2, and they are usually imple-mented with the inference engine. By working with the inference engine, experts canadjust the input they give in step 1 after viewing the results in steps 2 and 3.

THE IMPORTANCE OF UNCERTAINTY

Although uncertainty is widespread in the real world, its treatment in the practicalworld of artificial intelligence is very limited. As a matter of fact, many real-worldknowledge-based systems completely avoid the issue of uncertainty. People feel that itis not necessary to represent uncertainty in dealing with uncertain knowledge. Why isthis so?

The answer given by practitioners is very simple: Even though they recognize theproblem of uncertainty, they feel that none of the methods available are accurate orconsistent enough to handle the situation. In fact, knowledge engineers who haveexperimented with proposed methods for dealing with uncertainty have found eitherno significant difference from treating the situation as being under assumed certaintyor large differences between the results when they used different methods. Does thismean that uncertainty avoidance is the best approach and therefore that this materialshould be deleted from the book? Certainly not!

Uncertainty is a serious problem.Avoiding it may not be the best strategy. Instead,we need to improve the methods for dealing with uncertainty. Theoreticians must real-ize that many of the concepts presented in this chapter are foreign to many practition-ers. Even structured methods, such as the Bayesian formula, seem extremely strangeand complex to many people.

REPRESENTING UNCERTAINTY

The three basic methods of representing uncertainty are numeric, graphical, andsymbolic.

Numeric Representation of UncertaintyThe most common method of representing uncertainty is numeric, using a scale withtwo extreme numbers. For example, 0 can be used to represent complete uncertainty,and 1 or 100 can represent complete certainty. Although such representation seemstrivial to some people (maybe because it is similar to the representation of probabili-ties), it is very difficult for others.

In addition to the difficulties of using numbers, there are problems with cognitivebias. For example, experts figure the numbers based on their own experience and areinfluenced by their own perceptions. For a discussion of these biases, see Parsaye andChignell (1988). Finally, people may inconsistently provide different numeric values atdifferent times.

Graphical Representation of UncertaintyAlthough many experts are able to describe uncertainty in terms of numbers, such as“It is 85 percent certain that,” some find this difficult. The use of horizontal bars may



help experts to express their confidence in certain events. Such a bar is shown in FigureW18.15. Experts are asked to place markers somewhere on the scale. Thus, inFigure W18.15, Expert A expresses very little confidence in the likelihood of inflation,whereas Expert B is more confident that inflation is coming.

Even though some experts prefer graphical presentation, graphs are not as accu-rate as numbers. Another problem is that most experts do not have experiencein marking graphical scales (or setting numbers on the scale). Many experts, espe-cially managers, prefer ranking (which is symbolic) over either graphical or numericmethods.

Symbolic Representation of UncertaintyThere are several ways to represent uncertainty by using symbols. Many experts use aLikert scale to express their opinions. For example, an expert may be asked to assessthe likelihood of inflation on a five-point scale: very unlikely, unlikely, neutral, likely,and very likely. Ranking is a very popular approach among experts who have non-quantitative preferences. Ranking can be either ordinal (i.e., listing of items in order ofimportance) or cardinal (i.e., ranking complemented by numeric values). Managers areespecially comfortable with ordinal ranking.

When the number of items to be ranked is large, people may have a problem withranking and may tend to be inconsistent. A method that can be used to alleviate thisproblem is a pairwise comparison combined with a consistency checker; this involvesranking the items two at a time and checking for consistencies. One methodology forsuch ranking, the analytic hierarchy process (AHP), is described in Chapter 4. Themethod of fuzzy logic, presented in Chapter 13, includes a special symbolic representa-tion combined with numbers.

Symbolic representation methods are often combined with numbers or convertedto numeric values. For example, it is customary to give a weight of 1 to 5 to the fiveoptions on a Likert-type scale.

PROBABILITIES AND RELATED APPROACHES

A common approach to deal with uncertainty is to use probabilities. This can done inseveral ways.

The Probability RatioThe degree of confidence in a premise or a conclusion can be expressed as a probabil-ity. Probability is the chance that a particular event will occur (or not occur). It is a ratiocomputed as follows: The probability of X occurring, stated as P(X), is the ratio of thenumber of times X occurs, N(X), to the total number of events that take place.

Multiple probability values occur in many systems. For example, a rule can have anantecedent with three parts, each with a probability value. The overall probability ofthe rule can be computed as the product of the individual probabilities if the parts of

Expert A Expert B

Completeconfidence

MuchSomeLittleNoconfidence

FIGURE W18.15 A Confidence Scale About the Occurrence of Inflation



the antecedent are independent of one another. In a three-part antecedent, the proba-bilities may be .9, .7, and .65, and the overall probability is figured like this:

P = (.9)(.7)(.65) = .4095

The combined probability is about 41 percent. But this is true only if the individualparts of the antecedent do not affect or interrelate with one another.

Sometimes one rule references another. In this case, the individual rule probabili-ties can propagate from one to another, so we must evaluate the total probability of asequence of rules or a path through the search tree to determine whether a specificrule fires. Or we may be able to use the combined probabilities to predict the best paththrough the search tree.

In knowledge-based systems, there are several methods for combining probabilities.For example, they can be multiplied (i.e., joint probabilities) or averaged (using a simpleor a weighted average); in other instances, only the highest or lowest values are consid-ered. In all such cases, rules and events are considered independent of each other. If thereare dependencies in the system, the Bayesian extension theorem can be used.

The Bayesian ApproachBayes’s theorem is a mechanism for combining new and existent evidence, usuallygiven as subjective probabilities. It is used to revise existing prior probabilities basedon new information.

The Bayesian approach is based on subjective probabilities (i.e., probabilities esti-mated by a manager without the benefit of a formal model); a subjective probability isprovided for each proposition. If E is the evidence (i.e., the sum of all informationavailable to the system), then each proposition, P, has associated with it a value repre-senting the probability that P holds in light of all the evidence, E, derived by usingBayesian inference. Bayes’s theorem provides a way of computing the probability of aparticular event, given some set of observations that have already been made. Themain point here is not how this value is derived but that what we know or infer about aproposition is represented by a single value for its likelihood.

This approach has two major deficiencies. The first is that the single value does nottell us much about its precision, which may be very low when the value is derived fromuncertain evidence. Saying that the probability of a proposition being true in a given sit-uation is .5 (in the range 0–1) usually refers to an average figure that is true within a givenrange. For example, .5 plus or minus .001 is completely different from .5 plus or minus .3,yet both can be reported as .5.The second deficiency is that the single value combines theevidence for and against a proposition, without indicating the individual value of each.

The subjective probability expresses the degree of belief, or how strongly a valueor a situation is believed to be true. The Bayesian approach, with or without new evi-dence, can be diagrammed as a network.

The Dempster–Shafer Theory of EvidenceThe Dempster–Shafer theory of evidence is a well-known procedure for reasoningwith uncertainty in artificial intelligence. For details, see Dempster and Chiu (2006). Itcan be considered an extension of the Bayesian approach.

The Dempster–Shafer approach distinguishes between uncertainty and ignoranceby creating belief functions. Belief functions allow us to use our knowledge to boundthe assignment of probabilities when the boundaries are unavailable (see Coletti andScozzafava, 2006).

The Dempster–Shafer approach is especially appropriate for combining expertopinions because experts differ in their opinions with a certain degree of ignorance



and, in many situations, at least some epistemic information (i.e., information con-structed from vague perceptions). The Dempster–Shafer theory can be used to handleepistemic information as well as ignorance or lack of information. Unfortunately, itassumes that the sources of information to be combined are statistically independentof each other. In reality, there are many situations in which the knowledge of expertsoverlaps (i.e., there are dependencies among sources of information). Daniel (2006)proposed transforming belief functions into probabilities.

THEORY OF CERTAINTY FACTORS

The theory of certainty factors is based on the concepts of belief and disbelief.

Certainty Factors and BeliefsStandard statistical methods are based on the assumption that an uncertainty is theprobability that an event (or fact) is true or false. Certainty theory is a framework forrepresenting and working with degrees of belief of true and false in knowledge-basedsystems. In certainty theory as in fuzzy logic, uncertainty is represented as a degree ofbelief. There are two steps in using every nonprobabilistic method of uncertainty. First,it is necessary to be able to express the degree of belief. Second, it is necessary tomanipulate (e.g., combine) degrees of belief when using knowledge-based systems.

Certainty theory relies on the use of certainty factors. Certainty factors (CF)express belief in an event (or a fact or a hypothesis) based on evidence (or on theexpert’s assessment). There are several methods of using certainty factors to handleuncertainty in knowledge-based systems. One way is to use 1.0 or 100 for absolute truth(i.e., complete confidence) and 0 for certain falsehood. Certainty factors are not prob-abilities. For example, when we say there is a 90 percent chance of rain, there is eitherrain (90 percent) or no rain (10 percent). In a nonprobabilistic approach, we can saythat a certainty factor of 90 for rain means that it is very likely to rain. It does not nec-essarily mean that we express any opinion about our argument of no rain (which is notnecessarily 10). Thus, certainty factors do not have to sum up to 100.

Certainty theory introduces the concepts of belief and disbelief (i.e., the degree ofbelief that something is not going to happen). These concepts are independent of eachother and so cannot be combined in the same way as probabilities, but they can becombined according to the following formula:

CF(P,E) = MB(P,E) 2 MD(P,E)

Where: CF = certainty factorMB = measure of beliefMD = measure of disbeliefP = probabilityE = evidence or event

Another assumption of certainty theory is that the knowledge content of rules ismuch more important than the algebra of confidences that holds the system together.Confidence measures correspond to the information evaluations that human expertsattach to their conclusions (e.g., “It is probably true” or “It is highly unlikely”).

Combining Certainty FactorsCertainty factors can be used to combine estimates by different experts in severalways. Before using any ES shell, you need to make sure that you understand how cer-tainty factors are combined.The most acceptable way of combining them in rule-based



systems is the method used in EMYCIN. In this approach, we distinguish between twocases, described next.

Combining Several Certainty Factors in One Rule Consider this rule with an ANDoperator:

IF inflation is high, CF = 50 percent (A) ANDunemployment rate is above 7 percent, CF = 70 percent (B)ANDbond prices decline, CF = 100 percent (C),THEN stock prices decline.

For this type of rule, all IFs must be true for the conclusion to be true. However, insome cases, there is uncertainty as to what is happening. Then the CF of the conclusionis the minimum CF on the IF side:

CF(A, B, C) = minimum[CF(A), CF(B), CF(C)]

Thus, in our case, the CF for stock prices to decline is 50 percent. In other words,the chain is as strong as its weakest link.

Now look at this rule with an OR operator:

IF inflation is low, CF = 70 percent ORbond prices are high, CF = 85 percent,THEN stock prices will be high.

In this case, it is sufficient that only one of the IFs is true for the conclusion to betrue. Thus, if both IFs are believed to be true (at their certainty factor), then the con-clusion will have a CF with the maximum of the two:

CF (A or B) = maximum [CF (A), CF (B)]

In our case, CF = 85 percent for stock prices to be high. Note that both cases holdfor any number of IFs.

Combining Two or More Rules Why might rules be combined? There may be severalways to reach the same goal, each with different certainty factors for a given set offacts. When we have a knowledge-based system with several interrelated rules, each ofwhich makes the same conclusion but with a different certainty factor, each rule can beviewed as a piece of evidence that supports the joint conclusion. To calculate the cer-tainty factor (or the confidence) of the conclusion, it is necessary to combine the evi-dence. For example, let us assume that there are two rules:

R1: IF the inflation rate is less than 5 percent,THEN stock market prices go up (CF = 0.7).

R2: IF the unemployment level is less than 7 percent,THEN stock market prices go up (CF = 0.6).

Now let us assume a prediction that during the next year, the inflation rate will be4 percent and the unemployment level will be 6.5 percent (i.e., we assume that thepremises of the two rules are true). The combined effect is computed as follows:

CF(R1,R2) = CF(R1) + [CF(R2)] 3 [1 2 CF(R1)]

= CF(R1) + CF(R2) 2 [CF(R1)] 3 [CF(R2)]


Development W-233 ◆

In probability, when we combine two dependent probabilities (joint probabilities)such as these, we get the follow:

CF(R1,R2) = [CF(R1)] 3 [CF(R2)]

Here we have deleted this value from the sum of the two certainty factors, assum-ing an independent relationship between the rules. Here is an example:

Given CF(R1) = 0.7 AND CF(R2) = 0.6,CF(R1,R2) = 0.7 + 0.6 2 (0.7)(0.6) = 0.88

That is, the ES tells us that there is an 88 percent chance that stock prices willincrease.

Note: If we just added the CF of R1 and the CF of R2, their combined certainty wouldbe larger than 1.We modify the amount of certainty added by the second certainty factorby multiplying it by 1 minus the first certainty factor.Thus, the greater the first CF, the lessthe certainty added by the second. But additional factors always add some certainty.

To add a third rule to be added, we can use the following formula:

CF(R1,R2,R3) = CF(R1,R2) + [CF(R3)] [1 2 CF (R1,R2)]= CF(R1,R2) + CF(R3) 2 [CF(R1,R2)][CF(R3)]

Assume that a third rule is added:

R3: IF bond price increases,THEN stock prices go up (CF = 0.85).

Now, assuming that all the rules are true in their IF part, the chance that stockprices will go up is:

CF(R1,R2,R3) = 0.88 + 0.85 2 (0.88)(0.85) = 0.982

That is, there is a 98.2 percent chance that stock prices will go up. Note that CF(R1,R2)was computed earlier as 0.88. For a situation with more rules, we can apply the sameformula incrementally. See Russell and Norvig (2002) for more information on uncer-tainty and probabilistic ES.


1. Relate uncertainty to knowledge.

2. How is uncertainty represented?

3. Define probability ratio.

4. What is the Bayesian approach?

5. Describe theory of evidence and belief functions.

6. Describe certainty factors.

W18.12 EXPERT SYSTEMS DEVELOPMENT

An ES is basically computer software—an information system; so its development fol-lows a software development process. The goal of a development process is to maxi-mize the probability of developing viable, sustainable software within cost limitations



and on schedule, while managing change (because introducing a new, computerizedapproach or even implementing software to perform an existing method of workinginvolves change). The availability of ES shells and other development tools hasenabled rapid prototyping in system development. The idea of rapid prototyping,which involves performing the analysis, design, and implementation phases concur-rently and repeatedly in short cycles, ultimately leading to deployment of the system, isfeasible. Prototyping is also known as iterative design or evolutionary development. It isthe most commonly used DSS development methodology. One of its advantages is thatthe decision maker(s) and user(s) learn more about the problem they are trying tosolve while performing the prototyping process, as occurred in the opening vignetteand in the application cases in this chapter.

Using the prototyping approach to developing ES includes the following six majorphases (as illustrated in Figure W18.16):

• Phase I: Project initialization• Phase II: System analysis and design

Phase I Project initialization

Phase II System analysis

and design

Phase III Rapid prototyping

Phase IV System development

Phase V Implementation

Phase VI Postimplementation

OperationMaintenance and upgradesPeriodic evaluation

Acceptance by usersInstallation, demonstration, deploymentOrientation, trainingSecurityDocumentationIntegration, field-testing

Completing the knowledge baseTesting, evaluating, and improving knowledge basePlanning for integration

Conceptual design and planDevelopment strategySources of knowledgeComputing resources

Building a small prototypeTesting, improving, expandingDemonstrating and analyzing feasibilityCompleting design

Problem definitionNeed assessmentEvaluation of alternative solutionsVerification of an expert systems approachFeasibility studyCost-benefit analysisConsideration of managerial issuesOrganization of the development team

FIGURE W18.16 A Schematic View of the ES Development Lifecycle



• Phase III: Rapid prototyping and the demonstration prototype• Phase IV: System development• Phase V: Implementation• Phase VI: Postimplementation

PHASE I: PROJECT INITIALIZATION

Project initialization is the first step in developing ES. Its main goal is to identify prob-lems and prepare for further actions. Major tasks that must be done in this phaseinclude the following:

• Problem definition. Identify problems that are suitable for system development.• Need assessment. Determine the functional requirements of the system.• Evaluation of alternative solutions. Examine the availability of experts, education

and training, packaged knowledge, and conventional software.• Verification of an ES approach. Further verify the requirements, justify the needs,

and assess the appropriateness of developing such a system.• Feasibility study. Evaluate whether the project is technologically feasible, whether

it is managerially feasible, and whether change can be initiated and sustained.• Cost–benefit analysis. This part of the feasibility study involves determining eco-

nomic feasibility: formally identifying and estimating the potential costs and ben-efits of the system. Many criteria, including internal rate of return (IRR) or netpresent value (NPV), can be used for evaluating ES projects.

• Managerial consideration. Examine managerial issues, including project initiation,financing, resources, legal and other constraints, selling of the project, identifica-tion of a champion, and the potential of gaining access to users and user support.

• Organization of the development team. Assemble the team and establish commit-ment before starting formal development because of the nature of the prototyp-ing process.

• Ending milestone. Approve the project in principle.

PHASE II: SYSTEM ANALYSIS AND DESIGN

After the concept of the project has been approved, a detailed system analysis must beconducted to estimate system functionality.The major tasks in this phase are explainedin the following sections.

Conceptual DesignA conceptual design of an ES is similar to an architectural sketch of a house. It pro-vides a general idea of what the system will look like and how it will solve the problem.The design shows the general capabilities of the system, the interfaces with othercomputer-based information systems, the areas of risk, the required resources, theanticipated cash flow, the final composition of the team (including identification ofthe expert(s) and an estimate of their commitment), and any other information thatwill be necessary for detailed design later.When the conceptual design is complete, it isnecessary to determine the development strategy.

Development Strategy and MethodologyThere are three artificial intelligence development strategies: developing in-house, out-sourcing, and taking a blended approach in which external consultants (possibly evenfrom software vendors) join in-house teams. The needs of the particular project andexpected development costs and benefits or the organization’s policies dictate whichapproach should be chosen:



• In-house development. This strategy is attractive for organizations that alreadyhave the needed skills and resources. It is also appropriate if the ES will containproprietary or sensitive knowledge. In-house development can be done by theend users (called end-user computing) or a professional team with experts in dif-ferent areas.

• Outsourcing. An organization may hire an outside firm to do the development.There are many different ways, including hiring a consulting firm, partnering witha university, joining an industry consortium, and buying into an artificial intelli-gence firm.

• Blended approach. The development may be done by a team consisting of inter-nal and external members. This allows expertise outside the organization to beincorporated in system development.

Sources of KnowledgeES use both human experts and documented sources for knowledge. The greater theneed for human expertise, the longer and more complicated the acquisition processbecomes. Several issues may surface in selecting experts:

• Who selects the experts?• Who is an expert (possessing what characteristics)?• How can several experts be managed, if needed?• How can the expert be motivated to cooperate?

Experts are relatively easy to identify, especially when they are recommended byothers. They solve problems quickly, find high-quality, appropriate solutions, and focuson what is important. The issue of managing several experts can be difficult, particu-larly if they disagree. Generally, it is wise to use a single expert (especially in dealingwith economics-based systems) if his or her point of view has been helpful in problemsolving in the past. Knowledge from other experts can be incorporated into the systemlater.

Expert cooperation can be problematic. Experts may ask, “What’s in it for me?”“Why should I contribute my wisdom and risk my job?” and so on. Before building anES that requires the cooperation of experts, management should be clear on the fol-lowing points:

• Should experts be compensated for their contributions (e.g., royalties, specialrewards, payment)?

• Are experts truthful in describing how they solve problems?• How can experts be assured that they will not lose their jobs or that their jobs will

not be deemphasized when the ES is fully operational?• Will other people in the organization experience change or loss of their jobs?

How will these changes be handled?

Selection of the Development EnvironmentAfter obtaining expert cooperation, the development environment must be chosen.Three alternatives are available:

• ES shells. Expert system shells are software that has most of the necessary mecha-nisms, including a built-in inference engine and diversified user-interface func-tions, but is provided by the vendor with an empty knowledge base. After theknowledge is entered into the knowledge base, the system is operational. A goodexample is CORVID. When you enter the facts and rules, the system can operate.Figure W18.17 shows the basic shell concept.



Tip of the iceberg

Shell

Inferenceengine

Explanationprogram

Consultationmanager

Knowledge baseeditor and debugger

Knowledgebase

(e.g.rules)

Knowledge basemanagement facilites

Source: B.G. Buchanan, in Texas Instruments’s first satellite program.

FIGURE W18.17 The Shell Concept for Building ES

• Programming languages. ES can be developed in higher-level computer lan-guages, particularly the so-called fifth-generation languages (5GL). Examples ofthese artificial intelligence computer programming languages are LISP (list pro-cessing) and PROLOG. PROLOG, for instance, has built-in logic inference capa-bilities, which makes the design of the inference engine much easier.

• A hybrid environment. ES may be developed in a hybrid environment that com-bines various supporting tools. The hybrid system is also called a tool kit, fromwhich the knowledge engineer can select appropriate tools for appropriate tasks.

Shells are probably the most common development environment tool.They can becategorized as general or domain specific. In selecting a shell, it is important to makesure that it can handle the specific nature of the application properly, including expla-nation and interfacing with databases and other systems.

Shells have limitations and disadvantages. Because they are inflexible, it can be dif-ficult to fit them to nonstandard problems and tasks.As a result, a builder may use sev-eral shells, as well as environments and other tools, in a single application. This maycause problems as each software package is upgraded and in training and maintenance.Shells are end-user tools (similar to DSS generators), and their use is subject to all theproblems of end-user computing (e.g., poor documentation, security, development ofinappropriate methods, use of stale data or knowledge, improper maintenance).

Despite their limitations, shells are extremely useful and are used extensively bymany organizations. In some cases, they are used primarily in training or as the initialtool in the prototyping cycle (which makes a shell that can produce code quite useful).

Domain-specific tools are designed to be used only in the development of an ESfor a specific application area. For example, there are shells for diagnostic systems, con-figuration, financial applications, help desks (see Knowledge Library at Ask.MeOnLine for computer software support, amol.com, and PlanWrite for develop-ing a business plan, brs-inc.com), and scheduling. Firepond offers SalesPerformer ande-ServicePerformer (firepond.com).

Domain-specific tools may include some rule-based tools, an inductive mecha-nism, or a knowledge verification component. Domain-specific tools enhance the useof more standard tools by providing special development support and a user interface.These features permit faster application development. Some include embedded



knowledge (e.g., Knowledge Library at Ask.Me OnLine, amol.com). There are only afew commercial domain-specific shells and tools.

PHASE III: RAPID PROTOTYPING AND THE DEMONSTRATION PROTOTYPE

The prototyping process is actually less a phase than a cycle of phases. However,because of the way that knowledge is acquired and incorporated into an ES wedescribe it as a phase. Prototyping has been crucial to the development and success ofmany systems. A prototype in ES starts as a small-scale system. It includes representa-tion of the knowledge captured in a manner that enables quick inferencing and cre-ation of the major components of an ES on a rudimentary basis. For example, in a rule-based system, the first prototype, the demonstration prototype, may include only 10–50rules and be built with a shell. A small number of rules is sufficient to produce limitedconsultations and test the proof of concept of the ES.

A prototype helps the builder decide on the structure of the knowledge basebefore spending time on acquiring and implementing more rules. Developing a proto-type has other advantages. Rapid prototyping is essential in developing large systemsbecause the cost of a poorly structured and/or discarded ES can be quite high.

The process of rapid prototyping starts with the design of a small system. Thedesigner determines what aspect (or segment) to prototype, how many rules to use inthe first iteration, and so on. The knowledge is acquired for the first iteration and rep-resented in the ES. Next, a test is conducted. The test can be done using historical orhypothetical cases, and the expert is asked to judge the results. The knowledge repre-sentation methods and the software and hardware effectiveness are also examined. Inaddition, a potential user should test the system. The results are then analyzed by theknowledge engineer, and if improvement is needed, the system is redesigned. Usuallythe system goes through several iterations with refinements, and the process continuesuntil the system is ready for a formal demonstration.After the system is demonstrated,it is tested again and improved.This process continues until the final (complete) proto-type is ready.

The prototyping phase can be short and simple, or it can take several months andbe fairly complex. Some early ES, such as the famous Campbell Soup Company’s SoupCooker diagnostic ES, took a little more than a year to develop. The lessons learnedduring rapid prototyping are automatically incorporated into the final design as theprocess iterates. After the demonstration prototype is shown, another go/no-go deci-sion is made.

PHASE IV: SYSTEM DEVELOPMENT

When the initial prototype is ready and management is satisfied, system developmentbegins. Plans must be made for how to continue. At this stage, the development strat-egy may be changed (e.g., a consultant may be hired). The detailed design is also likelyto be changed, and so are other elements of the plan. If the prototype built in a shell isuseful, redesigning in a programming language such as C++ or PROLOG for execu-tion speed is also acceptable.

Depending on the nature of the system—for example, its size, the amount and typeof required interfaces with other systems, the dynamics of the knowledge, the develop-ment strategy—one or both of the following approaches is used for system development:

• Continue with prototyping (common)• Use the traditional systems development life cycle (SDLC) (rare)



System development can be a lengthy and complex process. In this phase, theknowledge base is developed, and continuous testing, reviews, and improvements areperformed. Other activities include creating interfaces (e.g., with databases, docu-ments, multimedia objects, hypermedia, the Web), creating and testing the user’s inter-face, and so on. In the opening vignette, a user interface was created so that facts couldbe entered at once rather than screen by screen, and another interface was developedfor database interaction. In this phase, prototyping continues, but needs are handled asthey arise. The two major tasks are to develop the knowledge base and to evaluate andimprove the system.

A detailed description of the tasks in this phase is available at this book’s Web site(prenhall.com/turban).

PHASE V: IMPLEMENTATION

Finishing system development is not the end of system development. The process ofimplementing an ES can be long and complex. Major tasks in this phase include thefollowing:

• Acceptance by the user. A proper strategy must be adopted to ensure that the users accept the system in their daily operations. Demonstration of the system and user orientation and training are very important to reduce possibleresistance.

• Installation approaches and timing. When the system is ready for online opera-tions, it must be integrated into the current business process. A parallel processthat combines existing procedures and the new procedure with the system mayhelp. It is also important to choose the right timing for starting the deployment ofthe system.

• Documentation and security. The system must be fully documented to ensuremaintainability. Because the system contains sensitive knowledge belonging tothe organization, it is very important to have a good security mechanism.

• Integration and field testing. If the ES stands alone, it can be field tested.Otherwise, it must be integrated with or embedded in another information systembefore field testing can commence. Field testing is extremely important becauseconditions in the field may differ from those in cases provided by experts.

PHASE VI: POSTIMPLEMENTATION

Several activities are performed after the system is deployed to users. The most impor-tant of these activities are system operation, maintenance, upgrading and expansion,and evaluation.

System OperationAccording to Prerau (1990), if the ES is to be delivered as a service, a system opera-tions group (or several groups, if there are several sites) should be formed and trained.If the system is to be a product run by users, an operator-training group may need to beformed, and consideration should be given to providing help for user-operators whohave problems. If the system is embedded in another system, the operators of the othersystem should be trained in any new operating procedures required.

System MaintenanceBecause an ES evolves over time, it is never really finished.Thus, it is important to planmaintenance. An ES must be revised on a regular basis with regard to the applicabilityof the rules, the integrity and quality of the data feeds, the use of the interlinked data-bases, and so on (see Beerel, 1993). Because experts are constantly training themselves



on new situations or reorganizing their knowledge in accounting for unencounteredsituations, an ES must be adjusted for these cases. In addition, software and hardwarebugs must be fixed as found, and the system must be upgraded to run on new softwarereleases and hardware platforms. Just as in the traditional system development lifecy-cle, there is a maintenance cycle, and a long-term maintenance team must be formedand trained to perform these tasks.

If the ES is embedded in another system, some thought should be given to whetherone maintenance group will serve the overall system or whether the ES will be main-tained separately. For separate maintenance, procedures for coordinating the twomaintenance groups must be developed. For further discussions of maintenance andupgrading, see McCaffrey (1992) and Karimi and Briggs (1996).

System Expansion (Upgrading)ES evolve continuously and therefore expand continuously. All new knowledge mustbe added, and new features and capabilities must be included as they become avail-able. Upgrading tasks generally fall under the jurisdiction of the maintenance team.However, some expansion can also be performed by the original developer(s) or evenby a vendor.

System EvaluationES need to be evaluated periodically (e.g., every 6 or 12 months, depending on thevolatility of their environment). When evaluating an ES, questions such as the follow-ing should be answered:

• What is the actual cost of maintaining the system compared to the actual bene-fits?

• Is the maintenance provided sufficient to keep the knowledge up to date so thatsystem accuracy remains high?

• Is the system accessible to all users?• Is acceptance of the system by users increasing?

If the users have been encouraged to comment on the ES and the feedbackprocess is easy, it is much easier to maintain a system under continuous feedback. Thisalso reduces the effort involved in periodic evaluations.


1. List the six phases of prototyping ES and describe each briefly.

2. How do you select a development environment? What are the alternatives?

3. List the major topics of implementation.

4. List the activities of postimplementation.

W18.13 KNOWLEDGE ACQUISITION AND THE INTERNET

The rapid growth of Internet use has opened an opportunity for knowledge acquisitionover the Internet. The availability of the Internet can alleviate many of the traditionalreasons for the poor productivity of manual knowledge acquisition methods(addressed earlier in this chapter). In general, the Internet provides a very convenientchannel to access knowledge that would otherwise be costly to obtain, if not impossi-ble. This can be viewed from two perspectives: the Internet as a communicationmedium and the Web as an open source of knowledge.


Knowledge Acquisition and the Internet W-241 ◆

THE INTERNET AS A COMMUNICATION MEDIUM

Knowledge acquisition relies heavily on communication between experts and theknowledge engineer. Traditionally, this communication process occurred in a placewhere they got together. That is to say, they were usually in the same place atthe same time. This restriction is now removed, thanks to the Internet’s use as aknowledge acquisition tool. The knowledge engineer can communicate with theexpert even though they are far away from each other, as long as they have access tothe Internet.

The Internet or an intranet can be used to facilitate knowledge acquisition. Forexample, electronic interviewing can be conducted if the knowledge engineer andthe experts are in different locations. Experts can validate and maintain knowledgebases from a distance. Documented knowledge can be reached via the Internet.The problem is to identify the relevant knowledge, a task that can be facilitated byintelligent agents.

Tochtermann and Fathi (1994) observed that interviewing methods attempt tomap the unstructured, nonlinear knowledge of an expert into a linear structure(i.e., rules), and in the process, knowledge losses occur. The linear structure ismapped back onto a nonlinear structure when implemented, leading to more incon-sistencies. Support of hypermedia on the Web can be used to represent the expertisein a more natural way. Natural links can be created in the knowledge (see Tung et al.,1999). HypEs (see Tung et al., 1999) is a proposed architecture for the developmentof media-rich ES. It is designed to use multiple media (e.g., text, sound, image,graphics, animation, video) for knowledge acquisition, storage, and user-interfaceactivities.

Ralha (1996) addressed issues pertaining to the problem of automatically struc-turing informal knowledge available on the Internet through a distributed hyperme-dia system: the Web. Hypermedia technology via the Web provides an ideal approachto the development of knowledge-based systems by enlarging the human–machinecommunication channel. This new approach to integrating hypermedia technologywith knowledge acquisition deals with knowledge before formalizing it. A twofold,qualitative spatial reasoning approach is adopted: First, dynamic linking takes place;second, the topology of the hyperspace, with qualitative spatial relations based on aprimitive concept of connections between spatial regions (called a hypermap), isdeveloped.

THE INTERNET AS AN OPEN KNOWLEDGE SOURCE

With even more Web pages built on it, the Internet has also become a valuable sourceof knowledge. The value of this knowledge source is further empowered by many por-tals and powerful search engines.

Many Web search engines incorporate intelligent agents to identify and deliver theinformation an individual wants. Smeaton and Crimmins (1996) developed a data-fusion agent to conduct multiple Web searches and organize them for the user. Yahoo!provides recommended Web sites from a search, using automated collaborative filter-ing (ACF) technology. ACF is a technique that provides recommendations based onstatistical matches of people’s evaluations of a set of objects in some given domain(e.g., movies, books,Web pages).The agent contacts other people’s agents to determinethe likelihood that a given object will match the user’s needs. Descriptions of these andother Web-related intelligent agents are provided in Chapter 14. Ask.com provides anInternet search robot that supplies search results based on query strings entered in theform of paragraphs.



The tools described here provide intelligent search capabilities through documentsor the Web. One key to using the Web is to provide a method for eliciting knowledgefrom experts via an anyplace/anytime collaborative-computing environment. Thisis valuable when experts are unavailable or the knowledge engineer cannot find aqualified expert in a domain. WebGrid-II was one of the first Web-based knowledgeelicitation approaches. Shaw and Gaines (1996) described the features of WebGrid-II,an Internet implementation grounded in personal-construct theory. WebGrid-II allowsa group of people to collaborate through the Internet to develop knowledge in adifferent-place/different-time environment while adhering to the constructs from therepertory grid techniques to elicit personal constructs about a set of elements relevantto the domain of interest. It is available at webgrid.co.uk.

Finally, because the amount of information provided over the Web is growingexponentially, scientists are developing methods for structuring information in distrib-uted hypermedia systems. Knowledge acquisition from the Web becomes an appealingapproach to developing knowledge-based systems.Web mining is becoming very useful(see Berry, 2003b; Chakrabarti, 2002; and Chapter 7).


1. Describe the Internet as a communication media for facilitating knowledgeengineering.

2. How can knowledge sources available on the Internet facilitate knowledgeengineering?

Chapter Highlights

• Knowledge engineering involves acquisition, repre-sentation, reasoning (i.e., inference), and explanationof knowledge.

• Knowledge is available from many sources, some ofwhich are documented and others of which areundocumented (i.e., experts).

• Knowledge can be shallow, describing a narrowinput/output relationship, or deep, describing com-plex interactions and a system’s operation.

• Knowledge acquisition, especially from human experts,is difficult because of several communication andinformation-processing problems.

• The methods of knowledge acquisition can bedivided into manual, semiautomated, and automated.

• The primary manual approach to knowledge acquisi-tion is interviewing. Interviewing methods rangefrom completely unstructured to highly structured.

• The reasoning process of experts can be tracked byusing several methods. Protocol analysis is the pri-mary method used in artificial intelligence.

• RGA is the most common technique for semi-automated interviews used in artificial intelligence.Several software packages that use RGA improve theknowledge acquisition process.

• Rule induction examines historical cases and generatesthe rules used to arrive at certain recommendations.

• There are benefits as well as limitations and prob-lems in having several experts build a knowledgebase.

• The major methods of dealing with multiple expertsare consensus methods, analytic approaches, selectionof an appropriate line of reasoning, automation of theprocess, and blackboard systems.

• Validation and verification of the knowledge base arecritical success factors in ES implementation.

• Case-based reasoning, neural computing, intelligentagents, and other machine learning tools can enhancethe task of knowledge acquisition.

• To build a knowledge base, a variety of know-ledge representation schemes are used, includingproduction rules, logic, semantic networks, andframes.

• Production rules are the most popular way of repre-senting knowledge. They take the form of an IF-THEN statement, such as “IF you drink too much,THEN you should not drive.”

• There are two types of rules: declarative (i.e., describ-ing facts) and procedural (i.e., inference).


Chapter Highlights W-243 ◆

• Propositional logic is a system of using symbols torepresent and manipulate premises, prove or dis-prove propositions, and draw conclusions.

• Predicate calculus is a type of logic used to representknowledge in the form of statements that assertinformation about objects or events and apply themin reasoning.

• Semantic networks are graphical depictions of knowl-edge that show relationships (i.e., arcs) betweenobjects (i.e., nodes). Common relationships are IS-A,Has-A, OWNS, and MADE FROM.

• A major property of networks is the inheritance ofproperties through the hierarchy.

• Decision trees and tables are often used in conjunc-tion with other representation methods. They helporganize the knowledge acquired before it is coded.

• A frame is a holistic data structure based on object-oriented programming technology.

• Frames are composed of slots that may contain differ-ent types of knowledge representation (e.g., rules,scripts, formulas).

• Frames can show complex relationships, graphicalinformation, and inheritance in a concise manner.Their modular structure helps in inference andmaintenance.

• Integrating several knowledge representation meth-ods is gaining popularity due to decreasing softwarecosts and increasing capabilities.

• The seed of an inference engine is the mechanism forsearching and reasoning. The major ones are rulechaining (backward and forward) and case-basedreasoning.

• In backward chaining, the search starts from a spe-cific goal:You seek evidence that supports (or contra-dicts) the acceptance of your goal. Then, dependingon your application of the rules and whether yourgoal is contradicted or accepted, you can try anothergoal.

• In forward chaining, the search starts from the facts(i.e., evidence), and you try to arrive at one or moreconclusions.

• The chaining process can be described graphically byan inference tree.

• Inferencing with frames often involves the use of rules.

• Case-based reasoning is based on experience withsimilar situations.

• In case-based reasoning, the attributes of an existingcase are compared with critical attributes derivedfrom cases stored in the case library.

• An explanation can be static, in which case a cannedresponse is available for a specific configuration; it

can also be dynamic, in which case the explanation isreconstructed according to the execution pattern ofthe rules.

• Processing uncertainty is important because knowl-edge may be inexact, and experts may be uncertain atany given time.

• Uncertainty processing methods include probabili-ties, the Bayesian theorem, and the theory of cer-tainty factors.

• Three basic methods can be used to represent uncer-tainty: numeric (probability-like), graphical, andqualitative.

• Disbelief expresses a feeling about what is not goingto occur.

• Certainty theory combines evidence available in onerule by seeking the lowest certainty factor when sev-eral certainty factors are added and the highest cer-tainty factor when any one of several factors is usedto establish evidence.

• Building an ES is a complex process with six majorphases: system initialization, system analysis anddesign, rapid prototyping, system development,implementation, and postimplementation.

• A feasibility study and a cost–benefit analysis help tojustify the development of an ES.

• A feasibility study is essential for the success of anymedium-sized to large ES.

• The potential system users must be involved in theES development.

• The conceptual design of an ES contains a generalidea of what the system will look like and how it willsolve the problem.

• ES can be developed in-house, by outsourcing, orthrough a blend of the two. There are several varia-tions of each approach.

• Shells are integrated packages in which the majorcomponents of an ES (except for the knowledgebase) are preprogrammed.

• Although ES can be developed with several tools, thetrend is toward developing the initial prototype witha simple (and inexpensive) integrated tool (either ashell or a hybrid environment).

• An ES can be built by creating a small-scale demon-stration prototype, testing it, and improving andexpanding it. This process, which is repeated manytimes, has many advantages.

• Validation is determining whether the right systemwas built or whether the system does what it wasmeant to do and at an acceptable level of accuracy.Verification confirms that an ES has been built cor-rectly according to specifications.



Key Terms

• automatic method• backtracking• backward chaining• belief function• certainty factor (CF)• certainty theory• chunking• CommonKADS• commonsense reasoning• decision table• declarative knowledge• deep knowledge• demonstration prototype• disbelief• documented knowledge• domain-specific tool• dynamic explanation• elicitation of knowledge• expert system shell• expert transfer system (ETS)• explanation• explanation facility

• facet• fifth-generation language (5GL)• forward chaining• frame• induction• inference engine• inference rule• inference tree• inheritance• interactive induction• interview analysis• instantiate• justifier• knowledge engineering• knowledge representation• knowledge validation

(verification)• LISP (list processing)• manual method• metaknowledge• metarule• multidimensional scaling

Questions for Discussion

1. Assume that you are to collect knowledge for one ofthe following systems:

a. An advisory system on equal-opportunity hiringsituations in your organization

b. An advisory system on investment in residentialreal estate

c. An advisory system on how to prepare your fed-eral tax return (form 1040)

What sources of knowledge would you consider forthese systems? (Refer to Figure W18.2.)

2. Discuss the major advantages of rule induction. Givean example that illustrates the method and describea situation in which you think it would be mostappropriate.

3. Transfer of knowledge from a human to a machine issaid to be more difficult than transfer of knowledgefrom a human to a human. Why?

4. What are the major advantages and disadvantages ofworking with a prototype system for knowledgeacquisition?

5. Discuss some of the problems of knowledge acquisi-tion through the use of expert reports.

6. Review the functions of ACQUIRE from aiinc.ca.What are the major benefits of ACQUIRE (or simi-lar products) compared to a conventional rule induc-tion package?

7. Why is it important to have knowledge analyzed,coded, and documented in a systematic way? How isthis related to the knowledge management systemsand e-learning systems that many organizations areadopting?

8. With machine learning methods, a computer developsknowledge models automatically. Explain why (orwhy not) machine learning techniques such as ruleinduction and neural computing can enhance knowl-edge acquisition.

9. Why is frame representation considered more com-plex than production rule representation? What arethe advantages of the former over the latter?

10. Provide an example that shows how a semantic net-work can depict inheritance.

11. Class scheduling is a domain that needs certain exper-tise. The scheduler needs to consider the professor’spreferences, student course conflicts, equipment avail-able in different classrooms, equipment requirements

• personal construct theory• predicate logic (or calculus)• procedural knowledge• procedural rule• process tracking• production rule• PROLOG (programming in logic)• propositional logic (or calculus)• protocol analysis• rule interpreter• semantic network• semiautomatic method• shallow knowledge• slot• static explanation• subjective probability• tool kit• training set• undocumented knowledge• unstructured (informal) interview• walk-through


Exercises W-245 ◆

TABLE W18.8 Communication Between an Expert and a Knowledge Engineer

Type of Communication

Time Spent Face-to- Written Continuing Time Spent by Knowledge

Method Face Contact Communication for a Long Time by Expert Engineer

Interview analysisObservations of expertsQuestionnaires and expert

reportsAnalysis of documented

knowledge

of different courses, and so on. Identify a knowledgerule with a procedural rule that is useful for classscheduling.

12. Certainty factors are popular in rule-based systems.Why? What unique features does the theory of uncer-tainty provide?

13. If you had a dialog with a human expert, what ques-tions besides why and how questions would you belikely to ask? Give examples.

14. Describe the relationship between metaknowledgeand explanations.

15. What can happen if the wrong knowledge representa-tion of a problem domain or an inappropriate ESshell or tool is selected?

Exercises

Teradata University and Other Hands-onExercises

1. Fill in Table W18.8 with regard to the type of commu-nication between the expert and the knowledge engi-neer. Use the following symbols: Y for yes, N for no,H for high, M for medium, and, L for low.

2. Read the following knowledge acquisition sessioninvolving a knowledge engineer (KE) and an expert (E):

KE: Michael, you have a reputation for findingthe best real estate properties for yourclients. How do you do it?

E: Well, Marla, first I learn about the client’sobjectives.

KE: What do you mean by that?

E: Some people are interested in income, others inprice appreciation. There are some specula-tors, too.

KE: Assume that somebody is interested in priceappreciation. What would your advice be?

E: Well, I first find out how much money theinvestor can put down and to what degreehe or she can subsidize the property.

KE: Why?

E: The more cash you use as a down payment, theless subsidy you will need. Properties withhigh potential for price appreciation need tobe subsidized for about two years.

KE: What else?

E: Location is very important. As a general rule, Irecommend looking for the lowest-pricedproperty in an expensive area.

KE: What else?

E: I compute the cash flow and consider the taximpact by using built-in formulas in my calculator.

Complete the following:

a. List the heuristics cited in this interview.

b. List the algorithms mentioned.

3. The authors of this book are coming to town and wouldlike to know where to go for lunch. Unfortunately, youwill be out of town, and you don’t know specificallywhat they like to eat or the kind of atmosphere theylike. But you do know how to encode your knowledgein an induction table like the one shown in TableW18.9. Using specific restaurants as the decisions (notyes/no but the restaurants themselves), construct aninduction table that reflects your knowledge about



TABLE W18.9 Comparisons of Automated Rule Induction and Interactive Methods

Time of Skill of Method or Tool Time of Expert Knowledge Engineer Knowledge Engineer

Rule inductionAutointelligenceSmart editorsETS

local restaurants.To make this problem a good learningexperience, work with a classmate, with one of youplaying the part of the knowledge engineer and theother the expert. Then switch roles, and for a secondadvisory problem, recommend a local hotel. (Note thatbecause these are not yes/no decision problems, theadvice can conclude with more than one choice.)

4. Construct a semantic network for the following situa-tion: Mini is a robin; it lives in a nest on a pine tree inMs. Wang’s backyard. Robins are birds, they can fly,and they have wings. They are an endangered species,and they are protected by government regulations.

5. Prepare a set of frames of an organization, given thefollowing information:

• Company: 1,050 employees, $130 million annualsales, Mary Sunny is the president

• Departments: accounting, finance, marketing, pro-duction, personnel

• Production department: five lines of production

• Product: computers

• Annual budget: $50,000 + ($12,000 x number ofcomputers produced)

• Materials: $6,000 per unit produced

• Working days: 250 per year

• Number of supervisors: 1 for every 12 employees

• Range of number of employees: 400–500 per shift(two shifts per day). Overtime or part time on athird shift is possible.

6. The following is a typical instruction set found in themanuals in most auto shops (this one is based onNissan’s shop manual):

• Topic: starter system troubles

• Procedures: Try to crank the starter. If it is dead orcranks slowly, turn on the headlights. If the head-lights are bright (or dim only slightly), the troubleis in the starter itself, the solenoid, or the wiring.Tofind the trouble, short the two large solenoid ter-minals together (not to ground). If the startercranks normally, the problem is in the wiring or in

the solenoid; check their connections up to theignition switch. If the starter does not work nor-mally, check the bushings. If the bushings aregood, send the starter to a test station or replace it.If the headlights are out or are very dim, check thebattery. If the battery is okay, check the wiring forbreaks, shorts, and dirty connections. If the batteryand connecting wires are not at fault, turn theheadlights on and try to crank the starter. If thelights dim drastically, it is probably because thestarter is shorted to ground. Have the startertested or replace it. (This example is based onCarrico et al., 1989.) Translate this informationinto rules. Can you do it in only six rules?

7. You are asked “Should we buy a house or not?” andgiven the following set of rules to answer it:

R1: IF inflation is low,THEN interest rates are low,ELSE interest rates are high.

R2: IF interest rates are high,THEN housing prices are high.

R3: IF housing prices are high,THEN do not buy a house,ELSE buy a house.

a. Run backward chaining with a high inflation rate,as given.

b. Run forward chaining with a low inflation rate, asgiven.

c. Prepare an inference tree for the backward-chainingcase.

8. You are given an ES with the following rules:

R1: IF interest rates fall,THEN bond prices will increase.

R2: IF interest rates increase,THEN bond prices will decline.

R3: IF interest rates are unchanged,THEN bond prices will remain unchanged.

R4: IF the dollar rises (against other currencies),THEN interest rates will decline.


Exercises W-247 ◆

R5: IF the dollar falls,THEN interest rates will increase.

R6: IF bond prices decline,THEN buy bonds.

a. A client has just observed that the dollar exchangerate is falling. He wants to know whether to buybonds. Run forward chaining and backward chain-ing, and submit a report of your findings to him.

b. Prepare an inference tree for the backward chain-ing you did.

c. A second client has observed that interest ratesare unchanged. She asks for advice on investing inbonds. What will the ES tell her? Use forwardchaining to figure it out.

9. You are given an ES with seven rules pertaining tothe interpersonal skills of a job applicant:

R1: IF the applicant answers questions in a straight-forward manner,THEN she is easy to converse with.

R2: IF the applicant seems honest,THEN she answers in a straightforward manner.

R3: IF the applicant has items on her resume thatare found to be untrue,THEN she does not seem honest,ELSE she seems honest.

R4: IF the applicant is able to arrange an appoint-ment with the executive assistant,THEN she is able to strike up a conversationwith the executive assistant.

R5: IF the applicant strikes up a conversation withthe executive assistant and the applicant is easyto converse with,THEN she is amiable.

R6: IF the applicant is amiable,THEN she has adequate interpersonal skills.

R7: IF the applicant has adequate interpersonalskills,THEN we will offer her the job.

Solve the following three independent cases:

a. Assume that the applicant does not have any itemson her resume that are found to be untrue and thatshe is able to arrange an appointment with theexecutive assistant. Run a forward-chaining analy-sis to find out whether to offer her the job.

b. It is known that the applicant answers questions in astraightforward manner. Run a backward-chaininganalysis to find out whether to offer her the job.

c. We have just discovered that the applicant wasable to arrange an appointment with the executiveassistant. It is also known that she is honest. Doesshe have interpersonal skills?

Exercises

Team Assignments and Role-Playing1. Have everyone in the group consider the fairly easy

task of doing laundry. Individually, write down all themotions you use in sorting clothes, loading the washerand dryer, and folding the clothes. Compare notes.Are any members of the group better at the task thanothers? For simplicity, leave out details such as “go tothe Laundromat.” Code the doing-laundry facts in arule base. How many exceptions to the rules did youfind?

2. Development of a Small Rule-Based ES for WineSelectionBuilding a small knowledge base with an ES shell isfairly easy. Here is an example.Selecting an appropriate wine for a certain type of foodis not simple. Many qualitative factors influence thechoice. Delegating the decision to a waiter can be riskyin some restaurants and even riskier at home if youwere trying to impress someone. So it may be useful fora restaurant to develop an ES to advise the customer.

Using the CORVID ES shell for this project, followthese steps to develop an ES for wine selection:

1. Specify the problem (i.e., pairing wine and food at arestaurant).

2. Name the system (e.g., Sommelier).3. Write the starting text (“Sommelier will help you

select a wine . . . ”).4. Decide on an appropriate coding for an uncertainty

situation (e.g., 0/1, 0 to 10, or 2100 to 100; you shoulduse 0 to 10).

5. Decide on any other inferencing parameters, asrequired by the shell (e.g., use all the rules or onlysome of the rules; set the threshold levels).

6. Prepare any concluding note that you want the userto see at the end of the consultation (e.g., “Thank youfor using Sommelier. Bon appetit!”).

7. Work with an expert to list 12 potential choices, suchas aged Cabernet Sauvignon, Sauvignon Blanc, crispChardonnay, Gerwürztraminer, Merlot, Pinot Noir,Zinfandel, and so on.



8. Develop a working knowledge of food and wine.(This is the best part for the knowledge engineerswho must gain some familiarity with the problemdomain.)

9. Develop a knowledge map with an expert (okay,maybe we did some testing on our own . . . ) in tabu-lar form. This helps focus on the knowledge.

10. Build the initial set of rules, the knowledge base, andthe shell and debug them. In CORVID, this involvesdeveloping qualifiers, which are the questions to beposed to derive facts, and the answer set of possiblefact values for each one. (Qualifiers are explainedshortly.)

11. Demonstrate the system to the expert, refine therules, and collect additional cases and rules. Repeatsteps 9–11 until the recommendation given by theexpert(s) and the system are the same.

12. Deploy the system.

Write the rules in the standard format required bythe shell. CORVID uses the concept of qualifiers. Forexample, it is known that the wine depends on the menuselection: You might have meat, fish, poultry, or pasta, andyou may have an appetizer as well. Each fact-gatheringstatement is called a qualifier. For example, Qualifier 1 isthe meat menu selection. Literally, its text is the phrase“The meat menu selection is.” Qualifier text generallyends in a verb. The qualifier can assume the followingvalues: (1) prime rib, (2) grilled steak, and (3) filet mignon.These are called the values of the qualifier.

The first rule might be constructed as follows:

IF the meat menu selection is prime rib

THEN Pinot Noir, confidence 9/10

AND Merlot, confidence 8/10

AND aged Cabernet Sauvignon, confidence 6/10.

This means that each of the three choices is appropriate,but Pinot Noir is the best choice (9), followed by Merlot (8),and then by aged Cabernet Sauvignon (6).

To build this rule, you create a new qualifier and thenselect the number of the qualifier (1) and the appropriatevalue (1, for prime rib). This forms the IF part of the rule.Click to the THEN part of the rule and call up the choices(step 7) and select the appropriate values for the THEN partwith the appropriate confidence level (don’t forget the con-fidence level, or the system will recommend no wine, leadingto a rather dour evening). In this case, the three wines areappropriate, but Pinot Noir is the best recommendation.

When all the rules are applied to the particular meal,the system might not recommend Pinot Noir overall. Asother rules are consulted, the certainty in the choice maydecrease. For example, the appetizer and the cost of the winemay influence the recommendation. Another considerationis the universe of the system’s knowledge. If your menuselection is not on the considered list, the system will notknow what to recommend because the main course is notpart of its knowledge universe.

Construct the knowledge base and implement the pro-totype to show some sample sessions.

Source: Lisa Sandoval, a graduate student at California StateUniversity, Long Beach, developed this ES student project,called “Your Sommelier.”

Exercises

Internet Exercises1. Access the Web site for CORVID at exsys.com and try

the demo systems available there. Summarize the char-acteristics of the demo systems and any new knowledgeyou acquire from them and from the Web site.

2. Search the Internet to find a knowledge acquisitiontool that was not introduced in this chapter and sum-marize its major functional features.

3. Search the Internet to find an intelligent systemapplication used in a movie rental store. Describe thepossible rules included in the knowledge base andsummarize the requirements that are specific to thedevelopment of Web-based ES. How are Web-basedsystems different from traditional systems?


References W-249 ◆

Beerel, A.C. (1993). Expert Systems in Business: RealWorld Applications. New York: Ellis Horwood.

Benjamins,V.R., D. Fensel, S. Decker, and A.G. Perez. (1999).“Building Ontologies for the Internet:A Mid-TermReport.” International Journal of Human–ComputerStudies.

Bergeron, B. (2003). Essentials of KnowledgeManagement. New York: Wiley.

Berry, M.J.A. (2003a). Survey of Text Mining: Clustering,Classification, and Retrieval. Berlin: Springer-Verlag.

Berry, M.J.A. (2003b). Data Mining Techniques withMastering Data Mining Set. New York: Wiley.

Bogarin, J.A.G., and N.F.F. Ebecken. (1996).“Integration ofKnowledge Sources for Flexible Pipe Evaluation andDesign.” Expert Systems with Applications,Vol. 10, No. 1.

Boose, J.H., and J.M. Bradshaw. (1993). “ExpertiseTransfer and Complex Problems: Using AQUINAS asa Knowledge Acquisition Workbench for Knowledge-Based Systems.” In B.G. Buchanan and D. Wilkins(eds.). Readings in Knowledge Acquisition andLearning: Automating the Construction andImprovement of Expert Systems. San Mateo, CA:Morgan Kaufmann.

Boose, J.H., and J.M. Bradshaw. (1999, August).“Expertise Transfer and Complex Problems: UsingAQUINAS as a Knowledge Acquisition Workbenchfor Knowledge-Based Systems.” International Journalof Human–Computer Studies.

Boose, J.H., and B.R. Gaines (eds.). (1990). TheFoundations of Knowledge Acquisition. New York:Academic Press.

Byrd, T.A. (1995). “Expert Systems Implementation:Interviews with Knowledge Engineers.” IndustrialManagement and Data Systems, Vol. 95, No. 10.

Carrico, M.A., et al. (1989). Building Knowledge Systems:Developing and Managing Rule-Based Applications.New York: McGraw-Hill.

Cawsey,A. (1995).“Developing an Explanation Componentfor a Knowledge-Based System: Discussion.” ExpertSystems with Applications,Vol. 8, No. 4.

Chakrabarti, S. (2002). Mining the Web: Analysis ofHypertext and Semi-Structured Data. San Francisco:Morgan Kaufmann.

Chaudhri, V.K., and J.D. Lowrance, (1998). GenericKnowledge-Base Editor, ai.sri.com/~gkb/ (accessedApril 2006).

Chun, S.-H., and S. Kim. (2004, September). “AutomatedGeneration of New Knowledge to Support ManagerialDecision-Making: Case Study in Forecasting a StockMarket.” Expert Systems, Vol. 21, No. 4.

Coletti, G., and R. Scozzafava. (2006, February 2).“Toward a General Theory of Conditional Beliefs.”International Journal of Intelligent Systems,Vol. 21, No. 3.

Cuena, J., and M. Molina. (1996). “KSM: An Environmentfor Design of Structured Knowledge Models.” In S.G.Tzafestas (ed.). Knowledge-Based Systems: AdvancedConcepts, Techniques and Applications. Singapore:World Scientific.

Daniel, M. (2006, February 2). “On Transformations ofBelief Functions to Probabilities.” International Journalof Intelligent Systems, Vol. 21, No. 3.

Davies, J., et al. (eds.). (2003). Towards the Semantic Web:Ontology-Driven Knowledge Management. New York:Wiley.

Davis, R. (1993). “Interactive Transfer of Expertise:Acquisition of New Inference Rules.” In B.G.Buchanan and D. Wilkins (eds.). Readings inKnowledge Acquisition and Learning: Automating theConstruction and Improvement of Expert Systems. SanMateo, CA: Morgan Kaufmann.

Demetriadis, S., A. Karoulis, and A. Pombortis. (1999,May). “Graphical Jog Through: Expert-BasedMethodology for User Interface Evaluation, Applied inthe Case of an Educational Simulation Interface.”Computers and Education.

Dempster, A.P., and W.F. Chiu. (2006, February 2).“Dempster-Shafer Models for Object Recognition andClassification.” International Journal of IntelligentSystems, Vol. 21, No. 3.

Devedzic, V., J. Debenham, and D. Radovic. (1996,October). “Object-Oriented Modelling of ExpertSystem Reasoning Process.” Proceedings, IEEEConference on Systems, Man, and Cybernetics. Beijing,pp. 2716–2721.

Diamantidis, N.A., and E.A. Giakoumakis. (1999,February). “An Interactive Tool for Knowledge BaseRefinement.” Expert Systems.

Feigenbaum, E., and P. McCorduck. (1983). The FifthGeneration. Reading, MA: Addison-Wesley.

Fensel, D., J. Angele, and R. Struder. (1998, July/August).“The Knowledge Acquisition and RepresentationLanguage KARL.” IEEE Transactions on Knowledgeand Data Engineering, Vol. 10, No. 4.

Fincher, S., and J. Tenenberg. (2005, July). “Making Senseof Card Sorting Data.” Expert Systems, Vol. 22, No. 3.

Gennari, J.H., et al. (2003). “The Evolution of Protégé: AnEnvironment for Knowledge-Based SystemsDevelopment.” International Journal ofHuman–Computer Studies, Vol. 58, No. 1.

References



Goldstein, D. (2002). “Automating Large Business Ruleand Expert Systems Verification: Easy ApproachSimplifies Maintenance.” PC AI, pp. 47–48.

Guth, M. (1999). An Expert System for Curtailing ElectricPower, timesofpakistan.com/Article/An-Expert-System-for-Curtailing-Electric-Power/105 (accessedApril 2006).

Hahn, U., M. Klenner, and K. Schnattinger. (1996). “AQuality-Based Terminological Reasoning Model forText Knowledge Acquisition.” In N. Shadbolt, K.O’Hara, and G. Schreiber (eds.). Advances inKnowledge Acquisition. Berlin: Springer-Verlag.

Hamilton, D.M. (1996). “Knowledge Acquisition forMultiple Site, Related Domain Expert Systems: DelphiProcess and Applications.” Expert Systems withApplications, Vol. 11, No. 3.

Hart, A. (1992). Knowledge Acquisition for ExpertSystems. New York: McGraw-Hill.

Helbig, H. (2006). Knowledge Representation and theSemantics of Natural Language. New York: Springer.

Helvacioglu, S., and M. Insel. (2005, May). “A ReasoningMethod for a Ship Design Expert System.” ExpertSystems, Vol. 22, No 2.

Humphrey, M.C. (1999, February). “A Graphical Notationfor the Design of Information Visualisations.”International Journal of Human–Computer Studies.

Hunter, G., and J.E. Beck. (2000, March). “UsingRepertory Grids to Conduct Across CulturalInformation Systems Research.” Information SystemsResearch.

Hurley, W.J., and D.U. Lior. (2002). “Combining ExpertJudgment: On the Performance of Trimmed Mean VoteAggregation Procedures in the Presence of StrategicVoting.” European Journal of Operational Research,Vol.140, No. 1, pp. 142–147.

Jackson, P. (1999). Introduction to Expert Systems, 3rdedition. Boston, MA: Addison-Wesley.

Jeng, B., T.P. Liang, and M. Hong. (1996). “InteractiveInduction for Knowledge Acquisition.” Expert Systemswith Applications, Vol. 10, Nos. 3 and 4, pp. 393–401.

Juan, P., J. Morant, and L. Gonzalez. (1999, October).“Knowledge-Based Systems’ Validation: When to StopRunning Test Cases.” International Journal ofHuman–Computer Studies.

Karimi, J., and P.L. Briggs. (1996, September). “SoftwareMaintenance Support for Knowledge-Based Systems.”Journal of Systems and Software, Vol. 34, No. 3.

Kearney, M. (1990, July). “Making KnowledgeEngineering Productive.” AI Expert.

Kidd, A.L., and M.B. Cooper. (1985). “Man–MachineInterface Issues in the Construction and Use of anExpert System.” International Journal of Man–MachineStudies, Vol. 22.

Kohavi, R. and F. Provost. (1998).“Glossary of Terms,” spe-cial issue on applications of machine learning and theknowledge discovery process, Machine Learning, Vol. 30.

Krogh, G.V., et al. (2000). Enabling Knowledge Creation.Oxford University Press.

Kuhn, D., and A. Zohar (1995). Strategies of KnowledgeAcquisition. Chicago: University of Chicago Press.

Lee, H.-Y., H.-L. Ong, and L.-H. Quek. (1995).“Exploiting Visualization in Knowledge Discovery.” InV. M. Fayyad and R. Uthurusamy (eds.). Proceedings ofthe First International Conferences on KnowledgeDiscovery and Data Mining (KDD-95). Menlo Park,CA: AAAI Press.

Lenat, D.B. (1982). “The Ubiquity of Discovery.”Artificial Intelligence, Vol. 19, No. 2.

Lockwood, S., and Z. Chen. (1994). “Modeling Experts’Decision Making Using Knowledge Charts.”Information and Decision Technologies, Vol. 19, No. 4.

McCaffrey, M.J. (1992). “Maintenance of Expert Systems:The Upcoming Challenge.” In E. Turban and J.Liebowitz (eds.). Managing Expert Systems.Harrisburg, PA: Idea Group.

McGraw, K.L., and B.K. Harbison-Briggs. (1989).Knowledge Acquisition, Principles and Guidelines.Upper Saddle River, NJ: Prentice Hall.

Medsker, L., and E. Turban. (1995). “KnowledgeAcquisition from Multiple Experts: Problems andIssues.” Expert Systems with Applications, Vol. 9.

Mili, H., and F. Pachet. (1995). “Regularity, DocumentGeneration, and Cyc.” In R. Rada and K. Tochtermann(eds.). Expertmedia: Expert Systems and Hypermedia.Singapore: World Scientific.

Moody, J.W., J.E. Blanton, and R.P. Will. (1998/1999,Winter). “Capturing Expertise from Experts: The Needto Match Knowledge Elicitation Techniques withExpert System Types.” Journal of ComputerInformation Systems.

Musen, M.A., L.M. Fagan, D.M. Combs, and E.H.Shortliffe. (1999, August). “Use of a Domain Model toDrive an Interactive Knowledge-Editing Tool.”International Journal of Human–Computer Studies.

Musen, M.A., S.W. Tu, A.K. Das, and Y. Shahar. (1995,July 23–27). “PROTÉGÉ-II: Computer Support forDevelopment of Intelligent Systems from Libraries ofComponents.” Proceedings of MEDINFO 1995,Eighth World Congress on Medical Informatics.Vancouver, BC.

Mussi, S. (2006, February). “User Profiling on the Web-Based on Deep Knowledge and SequentialQuestioning.” Expert Systems, Vol. 23, No. 1.

O’Keefe, R.M., O. Balci, and E.P. Smith. (1987, Winter).“Validating Expert System Performance.” IEEEExpert.


References W-251 ◆

O’Keefe, R.M., and D.E. O’Leary. (1993). “Performingand Managing Expert System Validation.” In M. Grabowski and W.A. Wallace (eds.) Advances inExpert Systems for Management, Vol. 1. Greenwich,CT: JAI Press.

O’Leary, D.E. (1993, March/April). “DeterminingDifferences in Expert Judgment: Implications forKnowledge Acquisition and Validation.” DecisionSciences, Vol. 24, No. 2.

Owen, S. (1990). Analog for Automated Reasoning. NewYork: Academic Press.

Owen, S., D. Budgen, and P. Brerton. (2006, February).“Protocol Analysis: A Neglected Practice.”Communications of the ACM, Vol. 49 No. 2.

Owrang, M.M., and F.J. Groupe. (1996). “Using DomainKnowledge to Guide Database Knowledge Discovery.”Expert Systems with Applications, Vol. 10, No. 2.

Parsaye, K., and M. Chignell. (1988). Expert Systems. NewYork: Wiley.

Prerau, D.S. (1990). Developing and Managing ExpertSystems. Reading, MA: Addison-Wesley.

Puls, T.L. (2002). “Prolog Fundamentals: A LogicProgramming Tutorial.” PC AI, Vol. 16, No. 6, pp. 28–33.

Ralha, C.G. (1996). “Structuring Information in aDistributed Hypermedia System.” In N. Shadbolt, K.O’Hara, and G. Schreiber (eds.). Advances inKnowledge Acquisition. Berlin: Springer-Verlag.

Ram, S., and S. Ram. (1996, January). “Validation ofExpert Systems for Innovation Management: Issues,Methodology, and Empirical Assessment.” Journal ofProduct Innovation Management, Vol. 13, No. 1.

Ranjan, A., J. Glassey,G. Montague, and P. Mohan,“FromProcess Experts to a Real-Time Knowledge-BasedSystem,” Expert Systems,Vol. 19, No. 2, 2002,pp. 69–79.

Rayham, A.F.R., and M.C. Fairhurst. (1999, February).“Enhancing Multiple Expert Decision CombinationStrategies Through Exploration of a Priori InformationSources.” IEEE Proceedings—Vision, Image and SignalProcessing, Vol. 146, No. 1.

Rincon, M., M. Bachiller, and J. Mira. (2005, July).Knowledge Modeling for the Image UnderstandingTask as a Design Task.” Expert Sysetms withApplications, Vol. 29, No. 1.

Roiger, R.J., and M. Geatz. (2003). Data Mining: ATutorial-Based Primer. Reading, MA: Addison-Wesley.

Rosenwald, G.W., and C.-C. Liu. (1997, January).“Rule-Based System Validation Through Automatic Identifica-tion of Equivalence Classes.” IEEE Transactions onKnowledge and Data Engineering,Vol. 9, No. 1.

Russell, S.J., and P. Norvig. (2002). Artificial Intelligence:A Modern Approach, 2nd ed. Prentice Hall.

Sandahl, K. (1994). “Transferring Knowledge from ActiveExperts to End-User Environments.” KnowledgeAcquisition, Vol. 6.

Schreiber, A.T., et al. (2000). Knowledge Engineering andManagement: The Common KADS Methodology.Boston: MIT Press.

Scott, A.C., J.E. Clayton, and E.L. Gibson. (1991). APractical Guide to Knowledge Acquisition. Reading,MA: Addison-Wesley.

Sharma, R.S., and D.W. Conrath. (1992, August).“Evaluating Expert Systems: The Socio-TechnicalDimensions of Quality.” Expert Systems.

Shaw, M.L.G., and B.R. Gaines. (1996). WebGrid:Knowledge Elicitation and Modeling on the Web.KSI, University of Calgary, ksi.cpsc.ucalgary.ca/KAW/ KAW96/gaines/KMD.html (accessed April 2006).

Sleeman, D., and F. Mitchell. (1996). “Towards PainlessKnowledge Acquisition.” In N. Shadbolt, K. O’Hara,and G. Schreiber (eds.). Advances in KnowledgeAcquisition. Berlin: Springer-Verlag.

Smeaton, A.F., and F. Crimmins. (1996). Using a DataFusion Agent for Searching the WWW. Glasnevin,Dublin, Ireland: School of Computer Applications,Dublin City University (no longer available online).

Stewart, V., and J. Mayes. (2000). “Business Applicationsof Repertory Grid” Enquire Within, enquirewithin.co.nz/BUS_APP/business.htm (accessed April 2006).

Sturman, M.C., and G.T. Milkovich. (1995,January/February). “Validating Expert Systems:A Demonstration Using Personal Choice Expert,a Flexible Employee Benefit System.” DecisionSciences, Vol. 26, No. 1.

Tochtermann, K., and M. Fathi. (1994). “Making Use ofExpertext to Enhance the Process of KnowledgeAcquisition.” In J. Liebowitz (ed.). Proceedings of theSecond World Congress on Expert Systems. Lisbon.

Tung, Y., R.D. Gopal, and J.R. Marsden. (1999, October).“HypEs: An Architecture for Hypermedia-EnabledExpert Systems.” Decision Support Systems, Vol. 26,No. 4.

van Harmelen, F., M. Aben, F. Ruiz, and J. van dePlassche. (1996, February). “Evaluating a formal KBSSpecification Language.” IEEE Expert.

Wagner, W.P., and C.W. Holsapple. (1997, February). “AnAnalysis of Knowledge Acquisition Roles andParticipants.” Expert Systems, Vol. 14, No. 1.

Wagner, W.P., J.R. Otto, and Q.B. Chung. (2003). “TheImpact of Problem Domains and KnowledgeAcquisition Techniques: A Content Analysis of P/OMExpert System Case Studies.” Expert Systems withApplications, Vol. 24, pp. 79–86.



Wang, F.-H. (2005, July). “On Acquiring ClassificationKnowledge from Noisy Data Based on Rough Set.”Expert Systems with Applications, Vol. 29, No. 1.

Wick, M.R., and J.R. Slagle. (1989). “An ExplanationFacility for Today’s Expert Systems.” IEEE Expert,Vol. 4, No. 1.

Vadera, S. (2005). “Inducing Safer Oblique Trees withoutCost.” Expert Systems, Vol. 22, No. 4.

Yan, H.-S. (2006, January). “A New Complicated-Knowledge Representation Approach Based onKnowledge Meshes.” IEEE Transactions on Knowledgeand Data Engineering, Vol. 18, No. 1.

Ye, L.R., and P.E. Johnson. (1995, June). “The Impact ofExplanation Facilities on User Acceptance of ExpertSystems Advice.” MIS Quarterly, Vol. 19, No. 2.

Glossary

automatic method An automatic knowledge acquisitionmethod that involves using computer software to auto-matically discover knowledge from a set of data.

backtracking A technique used in tree searches. Theprocess involves working backward from a failedobjective or an incorrect result to examine unexploredalternatives.

backward chaining A search technique that uses IF-THEN rules and is used in production systems thatbegin with the action clause of a rule and works back-ward through a chain of rules in an attempt to find averifiable set of condition clauses.

belief function The representation of uncertainty with-out the need to specify exact probabilities.

certainty factor (CF) A percentage supplied by anexpert system that indicates the probability that theconclusion reached by the system is correct. Also, thedegree of belief an expert has that a certain conclusionwill occur if a certain premise is true.

certainty theory A framework for representing andworking with degrees of belief of true and false inknowledge-based systems.

chunking A process of dividing and conquering, ordividing complex problems into subproblems.

CommonKADS The leading methodology to supportstructured knowledge engineering. It enables therecognition of opportunities and bottlenecks in howorganizations develop, distribute, and apply theirknowledge resources, and it is a tool for corporateknowledge management. CommonKADS provides themethods to perform a detailed analysis of knowledge-intensive tasks and processes and supports the devel-opment of knowledge systems that support selectedparts of the business process.

commonsense reasoning The branch of artificial intelligence that is concerned with replicating humanthinking.

decision table A table used to represent knowledge andprepare it for analysis.

declarative knowledge A representation of facts andassertions.

deep knowledge A representation of information aboutthe internal and causal structure of a system that con-siders the interactions among the system’s components.

demonstration prototype A small-scale prototype of a(usually expert) system that demonstrates some majorcapabilities of the final system on a rudimentary basis.It is used to gain support among users and managers.

disbelief The degree of belief that something is notgoing to happen.

documented knowledge For ES, stored knowledgesources not based directly on human expertise.

domain-specific tool A software shell designed to beused only in the development of a specific area (e.g., adiagnostic system).

dynamic explanation In ES, an explanation facility thatreconstructs the reasons for its actions as it evaluatesrules.

elicitation of knowledge The act of extracting knowl-edge, generally automatically, from nonhuman sources;machine learning.

expert system shell A computer program that facili-tates relatively easy implementation of a specificexpert system. Analogous to a DSS generator.

expert transfer system (ETS) A computer programthat interviews experts and helps them build expertsystems.

explanation An attempt by an ES to clarify its reason-ing, recommendations, or other actions (e.g., asking aquestion).

explanation facility The component of an expert systemthat can explain the system’s reasoning and justify itsconclusions. Also known as a justifier.

facet An attribute or a feature that describes the contentof a slot in a frame.

fifth-generation language (5GL) An artificial intelli-gence computer programming language. The bestknown are PROLOG and LISP.

forward chaining A data-driven search in a rule-basedsystem.

frame A knowledge representation scheme that associ-ates one or more features with an object in terms ofslots and particular slot values.

induction The process of reasoning from the specific tothe general.

inference engine The part of an expert system thatactually performs the reasoning function.

inference rule See metarule.


Glossary W-253 ◆

inference tree A schematic view of the inferenceprocess that shows the order in which rules are tested.

inheritance The process by which one object takes on oris assigned the characteristics of another object higherup in a hierarchy.

interactive induction A computer-based means ofknowledge acquisition that directly supports an expertin performing knowledge acquisition by guiding theexpert through knowledge structuring.

interview analysis An explicit, face-to-face knowledgeacquisition technique that involves a direct dialogbetween the expert and the knowledge engineer.

instantiate To assign (or substitute) a specific value orname to a variable in a frame (or in a logic expression),making it a particular “instance” of that variable.

justifier See explanation facility.knowledge engineering The engineering discipline in

which knowledge is integrated into computer systemsto solve complex problems that normally require ahigh level of human expertise.

knowledge representation A formalism for represent-ing facts and rules in a computer about a subject orspecialty.

knowledge validation (verification) The process oftesting to determine whether the knowledge in an artificial intelligence system is correct and whether thesystem performs with an acceptable level of accuracy.

LISP (list processing) An artificial intelligence pro-gramming language, created by artificial intelligencepioneer John McCarthy, that is especially popular inthe United States. It is based on manipulating lists ofsymbols.

manual method A human-intensive method for knowl-edge acquisition, such as interviews and observations,used to elicit knowledge from experts.

metaknowledge In an expert system, knowledge abouthow the system operates or reasons. More generally,knowledge about knowledge.

metarule A rule that describes how other rules shouldbe used or modified to direct an ES inference engine.

multidimensional scaling A method that identifies var-ious dimensions of knowledge and then arranges themin the form of a distance matrix. It uses least-squaresfitting regression to analyze, interpret, and integratethe data.

personal construct theory An approach in which eachperson is viewed as a “personal scientist” who seeks topredict and control events by forming theories, testinghypotheses, and analyzing results of experiments.

predicate logic (or calculus) A logical system of rea-soning used in artificial intelligence programs to indi-cate relationships among data items. It is the basis ofthe computer language PROLOG.

procedural knowledge Information about courses ofaction. Procedural knowledge contrasts with declara-tive knowledge.

procedural rule A rule that advises on how to solve aproblem, given that certain facts are known.

process tracking The process of an expert system’s trac-ing the reasoning process in order to reach a conclusion.

production rule A knowledge representation method inwhich knowledge is formalized into rules that have IFparts and THEN parts (also called conditions andactions, respectively).

PROLOG (programming in logic) A high-level computer language based on the concepts of predicatecalculus.

propositional logic (or calculus) A formal logical sys-tem of reasoning in which conclusions are drawn froma series of statements according to a strict set of rules.

protocol analysis A set of instructions governing theformat and control of data in moving from onemedium to another.

rule interpreter The inference mechanism in a rule-based system.

semantic network A knowledge representation methodthat consists of a network of nodes, representing con-cepts or objects, connected by arcs describing the rela-tions between the nodes.

semiautomatic method A knowledge acquisitionmethod that uses computer-based tools to supportknowledge engineers in order to facilitate the process.

shallow knowledge A representation of only surface-level information that can be used to deal with veryspecific situations.

slot A sub-element of a frame of an object. A slot is aparticular characteristic, specification, or definitionused in forming a knowledge base.

static explanation In an ES, an association of fixedexplanation text with a rule to explain the rule’s meaning.

subjective probability A probability estimated by amanager without the benefit of a formal model.

tool kit A collection of related software items that assista system developer.

training set A set of data for inducing a knowledgemodel, such as a rule base or a neural network.

undocumented knowledge Knowledge that comesfrom sources that are not documented, such as humanexperts.

unstructured (informal) interview An informal inter-view that acquaints a knowledge engineer with anexpert’s problem-solving domain.

walk-through In knowledge engineering, a processwhereby the expert walks (or talks) the knowledgeengineer through the solution to a problem.
