DECISION SUPPORT SYSTEMS

Photo by: sellingpix

Decision support systems (DSS) are a diverse group of interactive computer toolsprimarily customizable software designed to assist managerial decision making. They fall into a broader class known as management support systems (MSSs). The goal of a DSS is to make management more efficient and effective, particularly with ad hoc and discretionary decisions (versus routine or programmatic ones that require little judgment). Interactivity is key; unlike related expert systems and many artificial intelligence tools (see Figure 1), DSS generally do not attempt to make the decision themselves, but rather present information in a manner that is conducive to making an informed and efficient decision.

EVOLUTION OF DSS
DSS were introduced in the 1970s and gained mainstream attention in the 1980s. Originally run largely on mainframes, they were seen as an evolutionary step from management information systems, which at the time were relatively inflexible storehouses of corporate data. In that environment, DSS were high-end applications reserved for occasional, non-recurring strategic decisions by senior management. Since then, the rapid advances in personal computers ushered in a new breed of simple and widely used DSS. Indeed, some experts consider the built-in

Figure 1 Decision support Systems Versus Other Management Tools analytic functions in popular spreadsheet programs, such as Microsoft Excel and Lotus 1-2-3, to be mini-DSS. As a result, many DSS today are simple, informal PC software tools that users create themselves with the help of templates, macros, user-programmed modules, and other customizable features. While a simple DSS for an individual may cost a couple hundred dollars and some programming time, sophisticated ones continue to be significant business investments. At their inception they were exceptionally expensive to develop, and thus only large companies could afford them. Although relative prices have come down, they still tend to cost anywhere from $30,000 to $500,000 or more to implement company-wide. Premium systems are offered by such firms as IBM, SAS Institute, SPSS, and a host of more specialized vendors.

COMPONENTS OF A DSS
There are three basic components in a DSS:

a database a model base a user interface

Depending on the system, each of these components may be very simple or highly elaborate. The database, or in advanced systems, a database management system (DBMS) or a data warehouse, consists of structured, real-life information, such as customer account records, product sales history, employee schedules, or manufacturing process statistics. The model base, or model base management system (MBMS), contains one or more models for the kind of analysis the system will perform. For example, if the purpose of the system is to supply sales projections under different conditions, one model might be a linear regression formula derived from past sales and other factors. The user interface integrates the two into a coherent system and provides the decision maker with controls forand possibly feedback aboutmanaging the data and the models.

ANATOMY OF A DECISION
In order to discuss the support of decisions and what these tools can or should do, it is necessary to have a perspective on the nature of the decision process and thus what it means to support it. One way of looking at a decision is in terms of three areas or components. The first component is the data collected by a decision maker to be used in making the decision. The second component is the process selected by the decision maker to combine this data. Finally, there is an evaluation or learning component that compares decisions and examines them to see if there is a need to change either the data being used or the process that combines the data. These components of a decision interact with the characteristics of the decision being made. One approach to categorizing decisions is to consider the degree of structure in the decision-making activity.

THE STRUCTURE OF DECISIONS

STRUCTURED DECISIONS.
A structured decision is one in which all three components can be fairly well specified, i.e., the data, process, and evaluation are determined. Usually structured decisions are made regularly and therefore it makes sense to place a comparatively rigid framework around the decision and the people making it. An example of this type of decision may be the routine credit-granting decision made by many businesses. It is probably the case that most firms collect rather similar sets of data for credit granting decision makers to use. In addition the way in which the data is combined is likely to be consistent (for instance, household debt must be less than 25 percent of gross income). Finally, this decision can also be evaluated in a very structured way (specifically when the marginal cost of relaxing credit requirements equals the marginal revenue obtained from additional sales). For structured decisions it is possible and desirable to develop computer programs that collect and combine the data, thus giving the process a high degree of consistency. However, because these tend to be routine and predictable choices, a DSS is typically not needed for highly structured decisions. Instead, there are any number of automated tools that can make the decision based on the predefined criteria.

UNSTRUCTURED DECISIONS.

At the other end of the continuum are unstructured decisions. These decisions have the same components as structured ones; however, there is little agreement on their nature. For instance, with these types of decisions, each decision maker may use different data and processes to reach a conclusion. In addition, because of the nature of the decision there may also be few people that are even qualified to evaluate the decision. These types of decisions are generally the domain of experts in a given field. This is why firms hire consulting engineers to assist their decisionmaking activities in these areas. To support unstructured decisions requires an appreciation of individual approaches, and it may not be terribly beneficial to expend a great deal of effort to support them. Generally, unstructured decisions are not made regularly or are made in situations in which the environment is not well understood. New product decisions may fit into this category for either of these reasons. To support a decision like this requires a system that begins by focusing on the individual or team that will make the decision. These decision makers are usually entrusted with decisions that are unstructured because of their experience or expertise, and therefore it is their individual ability that is of value. One approach to support systems in this area is to construct a program that simulates the process used by a particular individual. These have been called "expert systems." It is probably not the case that an expert decision maker would be replaced by such a system, although it may offer support in terms of providing another perspective of the decision. Another approach is to monitor and document the process that was used so that the decision maker(s) can readily review what has already been examined and concluded. An even more novel approach used to support these decisions is to provide environments that are specially designed to give these decision makers an atmosphere that is conducive to their particular tastes, a task well suited for a DSS. The key to support of unstructured decisions is to understand the role that individual experience or expertise plays in the decision and to allow for individual approaches.

SEMI-STRUCTURED DECISIONS.
In the middle of the continuum are semi-structured decisions, and this is where most of what are considered to be true decision support systems are focused. Decisions of this type are characterized as having some agreement on the data, process, and/or evaluation to be used, but there is still a desire not to place too much structure on the decision and to let some human judgment be used. An initial step in analyzing which support system is required is to understand where the limitations of the decision maker may be manifested, i.e., will it be in the data acquisition portion, or in the process component, or possibly in the evaluation of outcomes. For instance, suppose an insurance executive is trying to decide whether to offer a new type of product to existing policyholders that will focus on families with two or more children that will be ready to attend college in six to nine years. The support required for this decision is essentially data oriented. The information required can be expressed in terms of the following query on the insurance company's database: "Give me a list of all of our policyholders that have a college education and have more than two children between ages 10 and 12."

PROCESSING INFORMATION

A major role of DSS is simple information processing; the program makes a large array of facts and considerations more digestible. They also automate tasks at which humans tend to be slow and inaccurate, such as sorting and mathematical calculations.

WHAT-IF ANALYSIS.
For instance, the insurance executive who wanted to offer the new product now has to decide on a price for the product. In order to make this decision, the effect of different variables (including price) on demand for the product and the subsequent profit must be evaluated. The executive's perceptions of the demand for the product can be captured in a mathematical formula that portrays the relationship between profit, price, and other variables considered important. Once the relationships have been expressed, the decision maker may now want to change the values for different variables and see what the effect on profits would be. The ability to save mathematical relationships and then obtain results for different values is a feature of many decision support systems. This is called "what-if' analysis and is a common application for DSS to automate.

EVALUATING OUTPUT.
Of course, the output from such a system is only as good as the model or data being used; if the demand model is inaccurate or outdated or based on dissimilar products, the outcome projections may be worthless. Thus, decision makers must be aware of the risk of potential inaccuracies and understand the underlying logic behind a DSS's output, as opposed to accepting its output blindly, in order to make an informed decision. The object of a good DSS is to obtain useful information for human consideration rather than to let the computer make the decision itself. Advanced DSS may contain safeguards and pointers to help users avoid misinterpreting output or creating meaningless output.

LEVEL OF INTERACTION.
Systems such as the Statistical Navigator go through a dialogue with the user to determine what the data's characteristics are and what questions are actually being asked of the data. Then the system suggests what techniques are most appropriate to use. This approach to supporting decision makers requires that the DSS possess a great deal more than database or processing capabilitiesit should actually have an understanding of the domain in which the system is being used. The Statistical Navigator has knowledge of statistical methods and the benefits, assumptions, and problems associated with each method. A future step would be a system that has an understanding of more processing options than just statistical methods. This might include linear programming or present value analysis. As DSS start to include many different processing models in the library of choices, two possibilities exist. One possibility is that the system will merely allow users to choose different methods within the same overall DSS. In this instance the user must still supply the knowledge of what is the most appropriate method and must be able to interpret the results. Another possibility is similar to the approach used in the Statistical Navigator, which would be to include a knowledge of the methods in the DSS and let it help the user select among many methods, not just statistical. Of

course each approach does have its problems. There are software packages that allow users to select among different methods, but they do not offer a great deal of guidance on their use. Thus, as is the case with certain statistical analysis packages, the conclusions may not be correct because the method was applied incorrectly. The second possibility presents a very different problem, or perhaps challenge; that is, how much knowledge to build into the DSS. A single system with general knowledge of most processing methods would be very popular with most users. However, designers would be confronted with the problem of what to include in a support system or what decision activities it should support. Should it have simple knowledge of the processing methods, such as linear programming, statistical regression, and present value? Or should it have knowledge about decision areas, such as cash budgeting, locating a new plant, or pricing policies? This second approach may keep inappropriate data from being used, but then the questions about the role of the decision maker and how structured decisions may become must be addressed. Is the decision maker merely an information provider to a DSS that performs many functions, or should the role of the DSS be simply to make whatever analysis is desired as easy as possible. As software develops more and more capabilities, designers and users of decision support systems will have to answer the question of what it actually means to support a decision.

SPECIAL KINDS OF DSS

Although all DSS are designed to tackle fairly specific types of problems, there are a number of recognized subcategories of DSS. Among them group decision support systems (GDSS) and executive information systems (EIS). At times these can be hard to distinguish from a "conventional" DSS, but both continue to enjoy solid backing in corporations and the separate terminology persists.

GROUP DECISION SUPPORT SYSTEMS.

As the name implies, GDSS are used to assist groups of decision makers who have common or overlapping responsibilities, such as executive committees, task forces, and work teams. Some of these tools are designed to be used directly when the group is convened. One example is tallying and processing group member preferences, and then presenting output for the participants to discuss. In other cases the group may never meet, but a centralized system is available to each member for common tasks they perform, such as financial monitoring and reporting.

EXECUTIVE INFORMATION SYSTEMS.

EISs are suites of data analysis tools that are meant to be applied to a company's most critical financial and performance data. In large organizations, usually this means the EIS has the ability to pull and manipulate dataincreasingly in real time instead of waiting days or weeks for the most recent dataon multiple corporate systems. EISs enjoyed a resurgence in the 1990s in part because of widespread management interest in activity-based costing, data warehousing, and enterprise resource planning systems. Software advances have also made EISs less costly and

more powerful. Many of the latest systems are run on client/server technology using a Web browser.

EXPERT SYSTEMS

Photo by: AlienCat

Expert systems are computer applications that combine computer equipment, software, and specialized information to imitate expert human reasoning and advice. As a branch of artificial intelligence, expert systems provide discipline-specific advice and explanation to their users. While artificial intelligence is a broad field covering many aspects of computer-generated thought, expert systems are more narrowly focused. Typically, expert systems function best with specific activities or problems and a discrete database of digitized facts, rules, cases, and models. Expert systems are used widely in commercial and industrial settings, including medicine, finance, manufacturing, and sales. As a software program, the expert system integrates a searching and sorting program with a knowledge database. The specific searching and sorting program for an expert system is known as the inference engine. The inference engine contains all the systematic processing rules and logic associated with the problem or task at hand. Mathematical probabilities often serve as the basis for many expert systems. The second componentthe knowledge databasestores necessary factual, procedural, and experiential information representing expert knowledge. Through a procedure known as knowledge transfer, expertise (or those skills and knowledge that sustain a much better than average performance) passes from human expert to knowledge engineer. The knowledge engineer actually creates and structures the knowledge database by

completing certain logical, physical, and psychosocial tasks. For this reason, expert systems are often referred to as knowledge-based information systems. By widely distributing human expertise through expert systems, businesses can realize benefits in consistency, accuracy, and reliability in problem-solving activities. Businesses may or may not differentiate between a decision support system (DSS) and an expert system. Some consider each one, alternately, to be a subcategory of the other. Whether or not they are one in the same, closely related, or completely independent is frequently debated in trade and professional literature. Like expert systems, the DSS relies on computer hardware, software, and information to function effectively. The debatable distinction, however, between an expert system and a DSS seems to lie in their practical applications. Decision support systems are used most often in specific decision-making activities, while expert systems operate in the area of problem-solving activities. But this distinction may be blurry in practice, and therefore investigation of an expert system often implies research on DSS as well. Four interactive roles form the activities of the expert system:

diagnosing interpreting predicting instructing

The systems accomplish each of these by applying rules and logic specified by the human expert during system creation or maintenance or determined by the system itself based on analysis of historical precedents. Instruction, in particular, emerges as a result of the expert system's justification system. Synthesizing feedback with various combinations of diagnostic, interpretative and predictive curriculum, the expert system can become a finely tuned personal tutor or a fully developed and standardized group class. Computer-aided instruction (CAI) thrives as a field of inquiry and development for businesses.

EARLY MODELS
Early expert systems appeared in the mid-1960s as an offshoot of research in artificial intelligence. Many early systems (GPPS and DENDRAL at Stanford University, XCON at Digital Equipment Corp., and CATS-1 at General Electric) pioneered the concept of a computer expert. But one, MYCIN, most clearly introduced two essential characteristics of an expert system: modularity and justification. MYCIN was developed at Stanford University as an expert system to aid in the diagnosis of bacterial meningitis. As it was developed, MYCIN emerged as a product of modular design with a facility to explain its own advice. Modular design refers to the concept and practice of developing software as stand-alone sets of programming code and informational materials. Each set connects as a module or self-contained capsule to other modules. This idea of modular design led to the further advance of expert shells. An expert shell program simplifies the development of an expert system by providing preexisting inference engines and module knowledge database components. The frontward and backward chaining effects of MYCIN (its ability to recount the steps it took to arrive at any recommendation) still influence the design of expert systems. As a result, the ability to explain or justify is a standard

facility of commercially produced expert systems and programs. Perhaps the most important discovery for MYCIN and other early expert systems was the importance of the human expert in the expert system.

BUILDING A KNOWLEDGE BASE

The basic role of an expert system is to replicate a human expert and replace him or her in a problem-solving activity. In order for this to happen, key information must be transferred from a human expert into the knowledge database and, when appropriate, the inference engine. Two different types of knowledge emerge from the human expert: facts and procedural or heuristic information. Facts encompass the definitively known data and the defined variables that comprise any given activity. Procedures capture the if-then logic the expert would use in any given activity. Through a formal knowledge acquisition process that includes identification, conceptualization, formalization, implementation, and testing, expert databases are developed. Interviews, transactional tracking, observation, case study, and self-reporting choices are common means of extracting information from a human expert. Using programmatic and physical integration of logic, data, and choice, expert systems integrate the examination and interpretation of data input with specific rules of behavior and facts to arrive at a recommended outcome.

APPLYING EXPERTISE: THE INFERENCE ENGINE

When an expert system must choose which piece of information is an appropriate answer to the specific problem at hand, uncertainty is intrinsic; thus, uncertainty is an underlying consideration in the overall conceptualization, development, and use of an expert system. One popular treatment of uncertainty uses fuzzy logic. Fuzzy logic divides the simple yes-no decision into a scale of probability. This extension of probability criteria allows the expert system to accommodate highly complex problems and activities in an attempt to more closely model human expert assistance and interaction. Probabilities of uncertainty vary from system to system based on the kind of information being stored and its intended uses. In its diagnostic role, an expert system offers to solve a problem by analyzing yes or no with the likelihood of correctly identifying a cause of a problem or disturbance. By inferring difficulties from past observations, the expert system identifies possible problems while offering possible advice and/or solutions. Diagnostic systems typically infer causes of problems. Applications include medicine, manufacturing, service, and a multitude of narrowly focused problem areas. As an aid to human problem solving, the diagnostic system or program assists by relying on past evidence and problems. By inferring descriptions from observations rather than problems, the expert system takes an interpretive rather than diagnostic role. Interpretive systems explain observations by inferring their meaning based on previous descriptions of situations. The probability of uncertainty is quantified as the likelihood of being an accurate representation. In a predictive role, the expert system forecasts future events and activities based on past information. Probabilities of uncertainty are emphasized as chances or the likelihood of being right. Finally, in an instructive role, the expert system teaches and evaluates the successful transfer of education

information back to the user. By explanation of its decision-making process, supplemental materials, and systematic testing, the instructive system accounts for uncertainty by measuring the likelihood that knowledge transfer was achieved. Regardless of the role of an expert system or how it deals with uncertainty, its anatomy is still similar. The inference engine forms the heart of the expert system. The knowledge base serves as the brain of the expert system. The inference engine chums through countless potential paths and possibilities based on some combination of rules, cases, models, or theories. Some rules, such as predicate logic, mimic human reasoning and offer various mathematical arguments to any query. A decision tree or branching steps and actions synthesize probability with rules and information to arrive at a recommendation. Probabilities mirror the human expert's own experience with an activity or problem. Other models or cases structure some systematic movement through a problem-solving exercise in different ways. Case-based reasoning uses specific incidents or models of behavior to simulate human reasoning. Other inference engines are based on semantic networks (associated nodes and links of association), scripts (preprogrammed activities and responses), objects (selfcontained variables and rule sets), and frames (more-specialized objects allowing inheritance). In all cases, the inference engine guides the processing steps and expert information together in a systematic way. The knowledge database provides the fuel for the inference engine. The knowledge database is composed of facts, records, rules, books, and countless other resources and materials. These materials are the absolute values and documented evidence associated with the database structure. If-then procedures and pertinent rules are an important part of the knowledge database. Imitating human reasoning, rules or heuristics use logic to record expert processing steps and requirements. Logic, facts, and past experience are woven together to make an expert database. As a result of knowledge transfer, significant experiences, skills, and facts fuse together in a representation of expertise. This expert database, or knowledge-based information system, is the foil for the inference engine. As such, the knowledge database must be accurately and reliably conceived, planned, and realized for optimum performance. Additionally, the knowledge database must be validated and confirmed as accurate and reliable. Expert databases containing inaccurate information or procedural steps that result in bad advice are ineffective and potentially destructive to the operation of a business. When, however, the inference engine and knowledge database synchronize correctly, businesses may realize gains in productivity and decreases in costs.

BENEFITS AND COSTS

Expert systems capture scarce expert knowledge and render it archival. This is an advantage when losing the expert would be a significant loss to the organization. Distributing the expert knowledge enhances employee productivity by offering necessary assistance to make the best decision. Improvements in reliability and quality frequently appear when expert systems distribute expert advice, opinion, and explanation on demand. Expert systems are capable of handling enormously complex tasks and activities as well as an extremely rich knowledgedatabase structure and content. As such, they are well suited to model human activities and problems. Expert systems can reduce production downtime and, as a result, increase output and quality. Additionally, expert systems facilitate the transfer of expertise to remote locations using

digital communications. In specific situations, ongoing use of an expert system may be cheaper and more consistent than the services of a human expert. Some benefits of an expert system are direct. Loma Engineering reduced its staff requirements from five engineers to a 1.5 equivalent by using an expert system to customize machine specifications. Other benefits are less direct and may include improved managerial functions. The Federal Aviation Administration uses the Smartflow Traffic Management System to better coordinate air traffic activities. The American Stock Exchange also put expert systems to use in monitoring insider trading. Hospitals use expert systems to interpret patient data through a large database of drug knowledge in order to identify harmful drug interactions and other problems. Thanks to one New England hospital system, doctors don't even have to be at the computer to get the results: if the system discovers a problem as new data is analyzed, it can automatically send a message to the doctor's pager. In manufacturing, expert systems are common and successful as well. Expert systems can track production variables, tabulate statistics, and identify processes that don't match the expected patterns, signaling potential problems. Moreover, integrated expert systems can immediately notify the appropriate person to correct a problem in the manufacturing process. The costs of expert systems vary considerably and often include post-development costs such as training and maintenance. Prices for the software development itself range from the low thousands of dollars for a very simple system to millions for a major undertaking. For large companies and complex activities, sufficiently powerful computer hardware must be available, and frequently programming must be done to integrate the new expert system with existing information systems and process controls. Additionally, depending on the application, the knowledge database must be updated frequently to maintain relevance and timeliness. Increased costs may also appear with the identification and employment of a human expert or a series of experts. Retaining an expert involves the potentially expensive task of transferring expertise to a digital format. Depending on the expert's ability to conceive and digitally represent knowledge, this process may be lengthy. Even after such efforts some expert systems fail to recover their costs because of poor design or inadequate knowledge modeling. Expert systems suffer, as well, from the systematic integration of preexisting human biases and ignorance into their original programming. Using an expert shella kind of off-the-shelf computer program for building an expert applicationis one way to reduce the costs of obtaining an expert system. The expert shell simplifies the expert system by providing preprogrammed modules and a ready-to-use inference engine structure. A number of companies provide expert shells that support business and industrial operations, including those conducted in Internet environments.