RESEARCH No'rF~
Expert Systems in Marketing: Guidelines For Development John T. Mentzer Virginia Polytechnic Institute
Nimish Gandhi Bradley University
This paper is aimed at understanding and developing marketing expert systems. A discussion of the structure of expert systems is presented. Guidelines for development of marketing expert systems and marketing areas most amenable to expert system development are also provided.
problem domain and applies their problem solving expertise to make useful inferences for the user of the system (Waterman and Hayes-Roth 1982). They are knowledge-based in nature and, hence, are also referred to as knowledge-based systems (Fikes and Kehler 1985). The knowledge of an expert system consists of facts and heuristics. The "facts" constitute a body of information that is widely shared, available, and agreed upon by experts in the field. The "heuristics" are rules of thumb that are used by experts to make judgements on the basis of their own beliefs and experience. As shown in Figure 1, the knowledge base and working memory constitute one part of the expert system. The inference engine and all the subsystems and interfaces constitute the second part of the system. The knowledge base contains facts and rules that embody the expert's knowledge. It is generally built by informal interviews between the expert and a knowledge engineer, the person who encodes the expert's knowledge. Working memory contains facts that emerge from consultation with the knowledge base. This consultation is conducted by the inference engine which contains the inference strategies and controls that an expert uses when he/she rr/anipulates the facts and rules. The inference engine performs two major tasks. First, it examines existing rules and facts, and adds new rules or facts when possible. Second, it decides the order in which inferences are made. When a user interfaces with the system, the queries of the system and his/her responses are inferred by the engine, which derives the rules and facts from the knowledge base. Marketing knowledge, in the form of experience, data, and published research, is accessed by the expert, the knowledge engineer, the user, and the expert system. The expert and knowledge engineer access the marketing data base in the process of creating the expert system. The user accesses the data base in the process of interacting with and reacting to the expert system. Once taken into the expert
Expert systems are computer programs that mimic human logic to solve problems (Newquist 1986). They are designed to make decisions in a specific task domain as well as an expert would perform the same task (Barr and Feigenbaum 1983; Hayes-Roth, Waterman, and Lenat 1983). In many ways, marketing managers are experts who draw on a variety of resources to analyze and solve marketing problems. Thus, marketing managers have in expert systems the tools that can help them understand marketing problems and generate improved solutions. In order to actualize such potential, it is necessary to understand what expert systems are and how they can be developed and applied in marketing. Toward this goal, the purpose of this article is to provide an overview and explore the possibilities for expert system development in marketing.
STRUCTURE OF EXPERT SYSTEMS An expert system can be defined as a computer system that uses the experience of one or more experts in some Journal of the Academy of Marketing Science ISSN: 0092-0703 Volume 20, Number 1, pages 71-80. Copyright 9 1992 by Academy of Marketing Science. All rights of reproduction in any form reserved. This article w a s accepted by the previous editor.
JAMS
73
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINES FOR DEVELOPMENT
MENTZER AND GANDHI
may be defined. Because interactions between frames and slots change the shape of frame-based systems, they are dynamic in nature. This enables frame-based expert systems to grow as more interactions take place. Thus, frame-based expert systems provide the potential to solve a variety of problems in a general knowledge domain (Newquist 1986). Here the underlying assumption is that after accumulating a critical mass of problem-independent knowledge about a given domain, a knowledge-based system will be capable of solving problems within that domain. Finally, logical expressions constitute yet another way of representing knowledge. Two common forms of logical expressions are propositional logic and predicate calculus. Propositions are statements that are either true or false and are linked together with connectives such as AND, OR, or NOT. Propositional logic is concerned with the truthfulness of the statements that are connected with these words. For example, if proposition A is true and proposition B is false, then the statement "A AND B" is false, where as "A OR B" is true. Predicate calculus is an extension to propositional logic. The elementary unit in predicate logic is an object. Statements about objects are called predicates and, again, are either true or false. For example, "Brand X is packaged in a blue box" is a statement that is either true or false. Thus, knowledge bases developed using logical expressions include statements that are either true or false. Correspondingly, retrieval of such knowledge also requires asking questions that are true or false. One cannot ask, "What is the brand X package color?" Rather, one must ask, "Is the brand X package color red?"
FIGURE 1 Structure of an Expert System KnowledgeBase Rules
Facts
InferenceEngine 9
Inference
_•
Knowledge ] Acquisition Subsystem
Control
Z
Explanation Subsystem
1
Expertor [_ Knowledge Engineer
l-
-E .-r-q
Adapted from Harmon and King 1985.
system, marketing knowledge becomes part of the knowledge base, for use in future problem solving.
Representing Knowledge The knowledge base essentially consists of rules and facts. There are five different strategies to encode rules, facts, and relationships that constitute knowledge (Harmon and King 1985): semantic networks, object-attribute-value triplets, rules, frames, and logical expressions. A semantic network is a collection of objects called nodes, which are conhected by links. For example, in a statement such as "John is.a male," "John" and "male" are nodes and "is a" represents a link. In object-attribute-value triplets, objects may be physical entities such as a sales manual or a conceptual entity such as a sales presentation. Attributes are general characteristics associated with objects, such as the attribute of sales manual educational value. The value identifies the degree of an attribute in a particular situation (e.g., the educationa] value of a sales manual may be "high"). Rules are used to represent relationships. Using IF-THEN clauses, rules can be used to represent knowledge. For example, "IF the product is complex, THEN it may generate more processing of the information in consumers' minds" can be used as a rule. Expert systems based on rules provide the ability to solve clearly defined, stand-alone problems (Newquist 1986). The underlying assumption is that the skills an expert uses to solve a given problem can be extracted effectively and efficiently as rules of thumb and incorporated into a system. The structure of the rules being specific (such as IF C THEN A) enables the system to solve well defined problems. A frame is a description of an object that contains slots for all the information associated with the object. Slots, like attributes, may store values. They also contain pointers to other frames and sets of rules or procedures by which values
JAMS
Drawing Inferences The inference engine uses facts and rules contained in the knowledge base and processes them to reach a decision. The system's methods of making inferences and inference control represent how a problem is diagnosed and a plausible solution offered. The most common inference strategy used in knowledge systems is the application of a logical rule called modus ponens (Harmon and King 1985). Modus ponens (latin for "a way of placing") is a type of inference rule that permits a system to draw its own inference on the basis of the information contained in the rule. That is, when A is known to be true and if a rule states, "If A, then B," it is valid to conclude that B is true. In other words, when the premises of a rule are true, the conclusions are also true. This represents a common rule for deriving new facts from existing rules and facts. Its importance arises due to the fact that expert systems must be able to make their own inferences with the information provided. It is a way for a system to conduct reasoning.
Inference Control The control portion of the inference engine addresses two primary problems: (1)
74
A knowledge system must have a way to decide where to start making inferences. Even though
WINTER, 1992
EXPERT SYSTEMSIN MARKETING: GUIDELINESFOR DEVELOPMENT
MENTZER AND GANDHI
ing the single value of an attribute to retract a conclusion is not very difficult. However, tracking down all the implications based on that fact is very complex and difficult. For this reason, most systems employ monotonic reasoning with some freedom to allow controlled types of nonmonotonic reasoning (Harmon and King 1985).
facts and rules reside in the knowledge base, they are static until activated by a process to begin reasoning. (2)
The inference engine must resolve conflicts that may occur when alternative lines of reasoning emerge. For example, it is possible for a system to reach a point where several rules could be used. The inference engine must be able to choose which rule to use next.
DEVELOPING AN EXPERT SYSTEM Expert systems are built iteratively (Kinnucan 1984), i.e., they evolve over time. The system changes as the decision making process is gradually understood and modeled (O'Leary 1986). Development of expert systems generally follows one of the following three routes: (1) custom development, where the system is built from scratch using artificial intelligence development languages; (2) semicustom development, which involves starting with an expert system shell and building a knowledge base around it; or (3) a package, which entails installing a pre-written application and making minor adjustments to fit the exact application needs. The first approach is by far the most expensive in terms of time, money, and effort. Typically, a knowledge engineer and an expert work together as a team. The knowledge engineer taps the expert's knowledge and programs it in one of the artificial intelligence languages. Current estimates indicate that an expert system development project costs from $150,000 to $200,000 for each person-year of effort (Newquist 1986). Moreover, developing a system on the scale that a strategic project requires may take 10 to 16 person-years. However, when fully functional, the system can be expected to provide handsome pay-offs. For instance, one expert system developed for inventory management purposes (IMA) was developed in less than six months and contained only 441 rules. In an early application, IMA detected and prevented a $600,000 mistake made by a manager with six years experience (Allen 1986). When problems to be solved are not very unique, building an expert system with a shell is beneficial (Newquist 1986). An expert system shell is a tool that facilitates development of an expert system. The shell contains an inference engine with proper codes for making inferences. The knowledge engineer then adds a specific knowledge base to the generic shell structure. Finally, expert systems application packages are off-the-shelf expert systems that are ready to use. End users of such packages must input new parameters in the system over time if the environment changes. Since pursuing this route entails minimal effort on the part of management, the market for application packages can be expected to flourish in the coming years. Developing an expert system is a complex task. Among other factors, the complexity varies according to the task domain, the tools used in developing the system, and the resources committed to its development. If the system is expected to be small in size (i.e., 50-350 rules), a tool can be selected first, then the knowledge-base built around it, and finally the target problem can be solved. The issue of first selecting an appropriate problem is not as critical in small systems as in large ones. If the system is expected to
Given that a knowledge base contains rules and facts, it is possible for a system to start reasoning in forward or backward mode. In a forward chaining system, the premises of a rule are examined first to check whether they are true, given the information on hand. The system then looks through the knowledge base to find the rules that are applicable. A rule may be selected and an action suggested by the rule may be inserted in the working memory. The system then proceeds to the next cycle of rules and checks what rules are applicable. The process continues until a solution is reached. This type of system is sometimes referred to as a data-driven system. In backward chaining, the inference engine starts with the possible solution and looks through the knowledge base to find the rules that justify that solution. Backward chaining systems are efficient when the possible outcomes to a situation are known and are small in number. This type of system is sometimes referred to as a goal-directed system. In addition, an inference .engine can conduct depth-first or breadth-first types of searches. In a depth-first search, the inference engine probes for all the information relevant to a particular theme. Breadth-first searches sweep across all premises in a rule before probing for detail. Most systems employ depth-first search (Harmon and King 1985). It could be said that "generalists" in a given field use a breadth-first strategy. They may start by asking questions in a general way about a problem, whereas "specialists" may tend to focus on a specific aspect of the problem and then probe for details about that aspect. The results of test markets for a new product in several cities, for example, might indicate some potential for success of the product. A depth-first knowledge system would ask all questions on the test market in a particular city at the same time. A breadth-first knowledge system would ask one question (e.g., How much were the sales?) for all test cities first and then proceed to the next question. A final distinction among control strategies used by inference engines is between monotonic and nonmonotonic reasoning. In a monotonic system, all values concluded from an attribute remain true. Facts that become true remain true. Thus, the amount of true information in the system grows steadily, i.e., monotonically. In a nonmonotonic reasoning system, facts that are true may be retracted (Harmon and King 1985). Planning is a good example of a problem type that demands nonmonotonic reasoning. In the early stages of a planning problem, it may make sense to follow a certain path. Later, as information is accumulated and if an early decision is found to be wrong, the decision and its consequences may be retracted. In expert systems, chang-
JAMS
75
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINES FOR DEVELOPMENT
MENTZER AND GANDHI
be large (i.e., 500 rules or more), the problem must be appropriate and clearly defined. Since large systems consume more resources than small ones, the loss of valuable resources can be minimized if the initial problem is selected carefully. Due to the investment of considerable resources in developing an expert system, a prototype is generally developed to determine if it is the kind of system desired. If successful, the prototype is expanded into the complete system. The system may then be evaluated before it is integrated with other existing systems in the organization.
ming is oriented toward using data and assisting the user in making a decision (Waterman 1986b). This is the essence of decision support systems (DSS), which use the interaction of data base, model, and dialog sub-systems to assist (support) the user in making a decision (Sprague and Carlson 1982). In contrast, knowledge programming deals with engineering the knowledge available in a system to come to a decision for the user. The reasoning in conventional programming is algorithmic whereas knowledge programming is heuristic. Finally, conventional programming involves manipulation of databases, which can be a repetitive process (Waterman 1986b]. On the other hand, knowledge programming involves manipulation of a knowledge base that can grow as the number of different interactions with the system grows. Mathematical models cannot fully satisfy a marketing manager's needs when the problem requires complex reasoning rather than calculating. The symbolic (non-numeric) knowledge and reasoning aspects of marketing problems often cannot be adequately quantified. For instance, determining the nature of sales promotions for products in different life cycle stages requires reasoning on the part of a product manager. Expert systems, however, operate particularly well when the thinking is pursued with reason and not calculations. With flexible knowledge handling capabilities, expert systems can integrate existing data with subjective judgment and reason numerically as well as symbolically (Winston 1984). Substantial codifiable knowledge exists in the marketing field upon which several experts can agree. Considerable knowledge has been accumulated by marketing researchers, much of which is published. Researchers have examined various marketing issues from several perspectives and their results have either been supportive of the occurrence of some phenomena or not supportive. In any case, reported results can be coded to show convergence (or nonconvergence) on some phenomena. Such an information base is essential to the development of expert systems and, thus, the marketing field lends itself very well to the development of such knowledge-based systems. Steinberg and Plank (1987) focus this view by exploring development of an expert system in the sales management area. Prominent issues such as productivity of salespeople, proper time utilization, and performance are of concern to a sales manager. An expert system can serve as an efficient tool for sales managers in controlling the activities of their sales staffs. The system's knowledge can include general marketing as well as selling information. The general marketing information can be related to the environment, products, prices, promotions, channels, and economic forecasts. Selling information can include market size, training strategies, past sales, costs of selling, and sales productivity records. Using a system encompassing information on these streams, a sales manager can focus on different aspects of a salesperson's performance. If performance is poor, the sales manager can diagnose the problem and develop guidelines for correction. Expert systems have also been developed to train salespeople, as demonstrated by the Automatic Courseware Expert (ACE) system (Lurin 1987). ACE extends the idea of providing computer-based training (CBT) to salespeople.
Transfer of Knowledge The inherent objective of an expert system is to enable a computer program to solve problems as well as an expert. The system is developed with the expert's knowledge and transferred to the system with the help of a knowledge engineer. Several approaches can be undertaken by the expert/knowledge engineer dyad. These include deriving knowledge from published sources (such as books and papers) as well as personal interviews of an expert by a knowledge engineer. Additionally, knowledge from more than one expert may be included to achieve greater robustness in the system (Luconi, Malone, and Scott Morton 1986). However, because of the cumbersome nature of such tasks, a strong commitment on the part of the expert as well as the knowledge engineer is important to the system's ultimate success.
Alternatives for transferring knowledge, such as simulation/expert interaction, need to be explored. For instance, if an expert is exposed to a simulation of decision circumstances, his actions and the rationale behind them can be captured by an interactive program. This may eliminate the frustration and inaccuracy inherent in the communication between knowledge engineer and expert.
System Assessment An expert system must be assessed before it is used. Its assessment includes validation, analysis of its strengths and weaknesses, and limitations. Validation of the system refers to evaluating the relationship between the decisions the system makes and the decisions an expert would make (O'Leary 1986). In other words, validation determines what a system does know, does not know, and knows incorrectly. Validation of a system entails having other experts examine the knowledge base of the system and testing the system against other problem solving models in the particular task domain. Naturally, it must perform in a similar or better fashion than other models and the experts themselves. Furthermore, the test problems to validate the system need to be representative of the problems the expert may encounter in practice.
EXPERT SYSTEMS IN MARKETING The distinguishing features between conventional programming and knowledge programming are related to how information is organized and used in the system. While both are essentially computer programs, conventional program-
JAMS
76
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINES FOR DEVELOPMENT
MENTZER AND GANDHI
FIGURE 2 Potential Marketing Expert System Applications Area
Description
Area
Description
Sales Territory Assignment, Quotas, and Congruency
Use territory and salesperson performance to define territorial boundaries, match salespeople with territories to maximize performance and set appropriate quotas.
Inventory Control
Use cost and profitability information to make stock keeping unit level decisions on inventory location, level, and order quantity.
Pricing Media Planning
Qualitative fine tuning of media plan by an expert system using quantitative, media model output (Schwoerer and Frappa 1986).
Use cost, competitive action, and buyer response information to develop expert system that assists in product and product line pricing decisions
Training/Education Promotions
Examine successful and unsuccessful promotions and draw guidelines for future promotions (Schwoerer and Frappa 1986).
PromotiOn Budgeting
Using an existing quantitative advertising/promotion response budgeting model develop an expert system to plan promotional budgets.
Use an expert system to provide an explanation of a given diagnosis or decision made from using a marketing practitioner training simulation model. Expert systems would provide interactive learning based on the level and progress of the student (Schwoerer and Frappa 1986).
Development of New Products
Couple results from simulated test market models with consumer and trade research about new products. Could focus on consumer packaged goods industry (Schwoerer and Frappa 1986).
Product Liability
Use information on past rulings and litigation decisions to develop expert systems that evaluate potential liability risk.
Product Life Cycle Analysis
Portfolio Analysis
Distribution Planning
Transportation Planning
JAMS
Market Research
By industry, develop expert systems that take environmental information to make recommendations on product, distribution, promotion, and price. Use environmental/consumer/ competitive information to develop product maps of potential new products. Based upon knowledge and information concerning total costs and physical distribution service, develop recommendations on product placement, ordering policies, and postponement/ speculation. Use rate and carrier service information in conjunction with traffic manager experience to develop expert system to make regular and irregular routing and scheduling decisions (Schwoerer and Frappa 1986).
77
Interviewing
On a per respondent basis, adapt the successive questions as a function of each respondent's previous answers, to avoid asking questions which do not provide pertinent information (Schwoerer and Frappa 1986).
Computer Interactive Negotiation Research
Allow development of a contingent negotiation expert system that reacts like a true negotiator to responses from the subject (Clopton and Barksdale 1987).
Questionnaire Design
Incorporate rules of good practice for survey design to help guarantee an acceptable minimum level, even for questionnaires designed by non-experts (Schwoerer and Frappa 1986).
Sampling
Determine the sample size and structure. Illustrate the cost versus quality and alpha, beta, and gamma error potential (Schwoerer and Frappa 1986).
Qualitative Research
Use an expert system to ask proper questions and aid in the analysis and synthesis of qualitative data (Schwoerer and Frappa 1986).
Project Design
Use expert system to assist in asking the right questions, plus develop schedules and proposals for evaluation by a marketing researcher and/or a client (Schwoerer and Frappa 1986).
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINESFOR DEVELOPMENT
MENTZER AND GANDHI
ACE allows development and modification of customized courses to assist a salesperson in learning about an organization, its products, and prices. Moreover, it offers the CBT flexibility of training the salespeople in any location. With ACE, CBT development and modification time is cut from 400 hours to 80 hours. Rangaswamy et al. (1987) discuss the development of ADCAD, an advertising expert system designed to assist in generating advertising copy strategy. ADCAD (ADvertising Communication Approach Designer), built by assimilation of practitioner as well as published knowledge, serves as a decision aid to agency personnel in determining communication approaches and copy strategy to position a product. Depending on the overall marketing and communication objectives, ADCAD evaluates audience as well as brand and product class characteristics to determine the best advertising strategy. For different segments of a market, ADCAD recommends, with varying degrees of confidence, what feature of the product should be emphasized in a communication. A marketing expert system related to market share has been developed by Atpar (1987). The goal of the system, called SHANEX (SHare ANalysis EXpert system prototype), is to provide possible reasons for market share changes of a product, rather than merely generating the estimates of changes in market share. SHANEX analyzes the product forms by sizes, on a national basis as well as account basis. For every product size, it evaluates the overall success of the product to determine the market share on the basis of trends, trade support, and price. Thus, the system provides its reasons for the market share analysis. The qualitative as well as quantitative information provided to product managers allows them to understand the reasons for market share changes. Marketing provides many more opportunities for expert system development where problems in specific domains of marketing cannot be solved in a limited time with common sense or algorithms alone. Figure 2 provides a list of many marketing application areas for expert systems. While the marketing areas indicated in Figure 2 do not exhaust the possibilities, they express the feasibility of expert system development in marketing. Existence and current development of a few systems amply exhibit the potential of such knowledge-based systems in marketing. However, this potential can only be achieved by a reasonable approach to the development of each marketing expert system application. To further this approach, the next section delineates the issues in marketing expert system development and the characteristics of a normative marketing expert system.
work indefinitely. While these are justifications for developing expert systems, there are several criteria that must be met in a field or domain before initiating the development of an expert system (Waterman 1986a): (1) the problem cannot be solved by common sense alone; (2) the problem takes at least one hour for a human expert to solve; and (3) there must exist sufficient codifiable knowledge in the public domain with agreement among experts on many issues. Naturally, if only common sense is needed to solve a problem, one does not need a computer program. If the problem can be solved easily or in a short period of time, usage of a computer program may not be warranted. Further, it is necessary that a substantial amount of public knowledge on a given issue exist. Moreover, the knowledge must be codifiable in a computer program. In other words, the knowledge must be structured with rules, their antecedents, and consequences. There must also exist general agreement among experts on many issues, which allows a foundation of common knowledge that could be applicable in various situations. A final criterion to develop an expert system is that the problem cannot be solved merely by algorithms. If the problem can be solved by a set of algorithms, the investment in developing an expert system cannot be justified. Further, an expert system, by definition, has an inherent capability to make inferences, a feat not achievable by algorithms. There are many potential applications for expert systems in marketing. The examples of existing expert systems given earlier merely touch this potential. From a normative perspective, efforts in developing marketing expert systems in the immediate future should focus on object-attributevalue triplets, frames, and logical expression for representing knowledge; modus ponens for drawing inferences; depth first, monotonic reasoning for inference control with either forward or backward chaining; and interviews for knowledge transfer. Existing expert system shells should be used to focus on specific marketing problems. Object-attribute-value triplets, frames, and logical expression offer the most potential for representing marketing knowledge since much of marketing knowledge involves judgment (value). Other forms of knowledge representation are too fact-based to allow for the subjective knowledge that composes so much of marketing information. When an expert system is to be applied to a variety of problems within a general area of marketing, frame-based systems are preferred. The dynamic interaction between frames and slots in these systems allows their application to a wider variety of problems within a given knowledge base. Modus ponens is recommended simply because most extant expert systems use this approach. Further, the IF-THEN logic follows the approach taken in building a sequential marketing plan. Depth first expert systems are more prevalent, and preferable, due to their compatibility with the marketing analysis process. Most marketing experts expect to be asked questions about a certain aspect of the marketing plan and answer all questions about that aspect, rather than jumping from one aspect to another and then back again. Although nonmonotonic reasoning is preferable in expert systems, the state of expert system development to date does not allow effective nonmonotonic programming. In a
MARKETING EXPERT SYSTEM DEVELOPMENT GUIDELINES There are at least five reasons for developing an expert system (Newquist 1986): (1) preserve the knowledge of an expert, (2) disseminate the knowledge of an expert, (3) store information in a form that can be updated easily, (4) aid novices in thinking the way an expert would, and (5) create a mechanism that is not subject to fatigue and can
JAMS
78
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINES FOR DEVELOPMENT
MENTZER AND GANDHI
tures they can include, where expert systems should be applied in marketing, and what features should be built into a marketing expert system. It has been the purpose of this article to provide this insight. An overview of expert systems has been provided, with several examples of existing marketing expert systems. Criteria for deciding on expert system solutions have also been provided. Alternative features of expert systems and those most appropriate for marketing were delineated. By its nature, expert system research in marketing will be application oriented. The list in Figure 2 provides the beginning of the areas in which this application will take place. Future research in the generic field of expert systems will concentrate on improvements to expert systems features, nonmonotonic reasoning for example. It should be the role of marketing expert system research to apply the existing technology to marketing problems and report the results. This case-by-case approach is typical of seminal research and can be used to develop more detailed guidelines of use and misuse of expert systems in marketing.
large, strategic marketing planning expert system, the falsification of a previous aspect of the plan will have widereaching implications for the veracity of many other aspects. Thus, nonmonotonic reasoning would be a valuable tool. However, monotonic marketing expert systems are more practical in the near term. Increasingly, expert systems employ both forward and backward chaining, depending on the problem being addressed (Rangaswamy et al. 1989). Forward chaining is most appropriate when the available resources are constrained and the objective is to achieve maximum performance from the given resources via the marketing plan. Backward chaining is practical for marketing management when specific goals are being set and the decision is to determine how many resources to commit to accomplish these goals. Knowledge transfer through the vehicle of knowledge engineer interviews of experts is time-consuming and error prone. However, to date it is the most effective method available. Future enhancements in interactive role playing simulations hold considerable potential to improve the timeliness and accuracy of expert knowledge transfer. Development of marketing expert systems should utilize existing generic shells. Development from scratch involves too high a cost in terms of money and time for most marketing applications. The relative paucity of off-the-shelf marketing expert systems negates this alternative for the short term. Generic expert system shells provide the marketing user with the advantages of time and monetary savings from not investing in the development of the fundamental expert systems logic, but allows the flexibility to tailor the expert system to specific marketing problems. Numerous shells exist on the market today to build expert systems of different sizes (Turban 1990). For small expert systems, a shell gaining acceptance among users is VP Expert. For medium sized expert systems, a notable shell is GURU. For large expert systems, IBM's ESE may serve a user well. Within the domain of marketing expert system development, the issue of system maintenance also arises. Knowledge and its applications are more dynamic in some fields than others (Bayer and Keon 1986). A system designed to diagnose a disease on the basis of a patient's symptoms can be used with confidence over an extended span of time. The disease identified by the system may be true and valid at all times because the knowledge about the symptoms of the disease is relatively static. However, a marketing system may be more susceptible to provide inappropriate advice due to the dynamic nature of the environment. What may be true for a marketing situation at one point in time may not remain true forever, Continual review of the validity of the resultant marketing expert system should be part of any development plan.
REFERENCES Allen, Mary K. 1986. The Development of An Artificial Intelligence System for Inventory Management. Chicago, IL: Council of Logistics Management. Alpar, Pavle. 1987. "Expert Systems in Marketing." Working Paper No. 86-19. Chicago, IL: University of Illinois at Chicago. Barr, A. and Edward Feigenbaum. 1983. The Handbook of Artificiallntelligence. Vol. 1-3. Los Altos, CA: William Kaufman. Clopton, Stephen W. and Hiram C. Barksdale, Jr. 1987. "Microcomputer Based Methods for Dyadic Interaction Research in Marketing." Journal of the Academy of Marketing Science 15 (Summer): 63-68. Fikes, Richard and Tom Kehler. 1985. "'The Role of Frame-Based Representation in Reasoning." Commum'cations of the ACM 28 (September): 904-920. Harmon, Paul and David King. i985. Expert Systems: Artificial Intelligence in Business. New York, NY: John Wiley and Sons. Hayes-Roth, Frederick, Donald A. Waterman, and Douglas B. Lenat. 1983. Building Expert Systems. Reading, MA: Addison-Wesley Publishing. Kinnucan, Paul. 1984. "Computers That Think Like Experts." High Technology 4 (January): 30-42. Luconi, Fred L., Thomas W. Malone, and Michael S. Scott Morton. 1986. "Expert Systems: The Next Challenge for Managers." Sloan Management Review 27 (Summer): 3-14. Lurin, Ely S. 1987. "New Expert System a Boon to Sales Training." Marketing News 21 (November 6): 18-9. Newquist, Harvey P. III. 1986. "Expert Systems: The Promise of a Smart Machine." Computerworld 20 (January): 43-60. O'Leary, Daniel E. 1986. "Validation of Business Expert Systems." Paper presented at the University of Southern California Audit Symposium. Los Angeles, CA. Rangaswamy, Arvind, Raymond R. Burke, Jerry Wind and Jehoshua Eliashberg. 1987. "Expert Systems for Marketing." Working Paper No. 87107. Cambridge, MA: Marketing Science Institute. Rangaswamy, Arvind, Jehoshua Eliashberg, Raymond R. Burke, and Jerry Wind. 1989. "Developing Marketing Expert Systems: An Application to International Negotiations." Journal of Marketing 53 (October): 24-39. Schwoerer, Juergen and Jean-Paul Frappa. 1986. "Artificial Intelligence and Expert Systems: Any Applications for Marketing and Marketing Research?" Journal of the European Society for Opinion and Marketing Research 14 (4): 10-24. Spraque, Ralph H. and Eric D. Carlson. 1982. Building Effective Decision Support Systems. Englewood Cliffs, NJ: Prentice Hall. Steinberg, Margery and Richard E. Plank. 1987. "Expert Systems: The Integrative Sales Management Tool of the Future." Journal of the Academy of Marketing Science 15 (Summer): 55-62.
CONCLUSIONS A great deal of experience already exists in the development of expert systems in other areas (Harmon and King 1985; Newquist 1986; Waterman 1986a). However, to effectively apply this experience to marketing, marketing experts must understand what expert systems are, what fea-
JAMS
79
WINTER, 1992
EXPERT SYSTEMS IN MARKETING: GUIDELINES FOR DEVELOPMENT
MENTZER AND GANDHI
Turban, Efraim. 1990, Decision Support and Expert Systems: Management Support Systems. New York, NY: Macmillan Publishing. Waterman, Donald A~ 1986a. A Guide To Expert Systems. Reading, MA: Addison-Wesley Publishing. - - . 1986b. "How Do Expert Systems Differ From Conventional Programs?" Expert Systems 3 (January): 16-18. Waterman, Donald A. and Frederick Hayes-Roth. 1982. An Investigation of Tools for Building Expert Systems. Santa Monica, CA: Rand Corporation, Winston, Patrick, H. 1984. Artificial Intelligence. Reading, MA: AddisonWesley.
technic Institute and State University. He has published in the Journal of the Academy of Marketing Science, Columbia Journal of World Business, International Journal of Physical Distribution and Materials Management, Journal of Business Logistics, Logistics and Transportation Review, Transportation Journal, Industrial Marketing Management, Research in Marketing, and other journals. He is the editor of the systems section of the Journal of Business Logistics and has served as editor of a special issue of the Journal of the Academy of Marketing Science and the International Journal of Physical Distribution and Materials Management.
ABOUT THE AUTHORS
Nimish Gandhi is Assistant Professor of Marketing at Bradley University. His research has appeared in the proceedings of the Academy of Marketing Science and the American Marketing Association.
John T. Mentzer is the Virginia Real Estate Professor of Marketing in the Department of Marketing at Virginia Poly-
JAMS
80
WINTER, 1992