This document provides an overview of knowledge-based systems and expert systems, including their development, components, advantages, disadvantages, examples, and historical context. It discusses the prototyping process for developing expert systems, common system components like rules, interfaces, and knowledge bases. Examples are provided of early successful systems like XCON and challenges that led to failures. Neural networks are introduced as a technique to improve system learning over time.
2. System Analyst User Expert Study the problem domain Step 1. Define the problem Step 2. Specify the rule set Step 3. Need to redesign Test the prototypesystem Step 4. Need to redesign Step 5. Construct the Interface Step 6. Conduct user tests Step 7. Use the system Step 8. Maintain the system Figure 16.10 Prototyping is Incorporated in the Development of an Expert System
3. Expert System Maintenance An expert system must be maintained just as any other CBIS subsystem, and this is accomplished in Step 8. Changes are made that enable the expert system to reflect the changing nature of the problem domain and to achieve greater efficiency. Sample Expert System Expert system activity in business began in the early 1980s. Since that time, the systems have been developed for a wide variety of application areas. As shown in Figure 16.11, the financial area of the firm has seen the highest level of activity. A sample of such a financial expert system is the credit approval system developed by professors Venkat Srinivasan of North Eastern University ,Boston and Young H, Kim of University of Cincinnati, who were working with a Fortune 500 company, which we will identify as SRR. SRR’s credit policy consists of two activities: (1) setting credit limits for new customers and reviewing them once a year, and (2) handling exceptions on a daily basis. Srinivasan and Kim interviewed credits managers and observed credit analysts making the credit decisions. A senior credit manager served as the expert. The Knowledge Base The knowledge base of the expert system consists of two components: (1) rules that reflect the credit approval logic, and (2) a mathematical model that determines the credit limit. The User Interface As the interface engine proceeds through the rule set in a forward-chaining manner, the credit analyst is asked to make pair wise comparisons. For example, the interface engine might display the prompt: What is the relative importance of Customer Background over Payment Record if the objective is to improve overall credit performances? The credit analyst enters a code that reflects the comparison, and the forward chaining proceeds. When the chaining is completed, the output appears as a series of screens. Expert
4. Table 16.1 Weightings of the Information Categories The expert system then explains how it arrived at its conclusion. The second screen in Figure 16.13 shows how the Good rating on pay experience was derived. The AHP Intensity Value is a score computed by the mathematical model.
8. How the Early Expert Systems Fared Such was the case of the first expert systems that were built during the early and mid-1980s. The XCON Success Story The most thoroughly documented expert system success story is that of Digital Equipment Corporation (Digital) and its expert system called XCON. XCON was one of several expert systems developed by Digital, and it was used to validate the technical correctness of the orders. The task was not an easy one, because there were more than 30,000 hardware and software parts that could be incorporated into a particular configuration. Evidence of the difficulty is the fact that the XCON knowledge base consisted of over 10,000 rules. Of all the expert system success stories, XCON provided the best example. The savings to Digital in the form of reduced manufacturing costs were estimated to be $15 million. Other successful efforts, although not so well publicized, were Exper-Tax by Coopers and Lybrand, and Authorizer’s Assistant by American Express. The Rest of the Story In an effort to learn the eventual outcome of the early expert systems efforts, T. Grandon Gill, an MIS professor at Florida Atlantic University, conducted a survey of ninety-seven expert systems, including XCON, which were built prior to 1988. The survey respondents included managers, developers, experts, users, and support personnel. The responses revealed that fewer than one-third of the systems ever achieved widespread or universal use and that almost one-half had been abandoned. On the positive side, almost three-fourths achieved some usage during their life span, and, for more than a third of them, the firms are still making investments in improving or maintaining the systems. Reasons for Expert System Failures The survey respondents identified the following causes of failure: 1) The original task that the expert system was designed to perform had changed. 2) The cost of maintaining the expert was too great. 3) The system became incompatible with other computer-based applications in the firm. 4) The firm changed its focus or direction. 5) The developers underestimated the size of the disk. 6) The system was developed to solve a problem that was not considered to be critical to the firm’s mission. 7) The system exposed the firm to legal liability. 8) Users resisted a system developed by outsiders. 9) Users refused to assume responsibility for maintaining the system. 10) Key development personnel were lost due to attrition. None of these reasons was due to inadequate technology. Rather, responsibility can be assigned to the firms’ executives, information specialists, and users.
9. Keys to Successful Expert System Development Using feedback from the survey respondents, Professor Gill identified five areas where the development projects could be improved. 1) Coordinate expert system development with the strategic business plan and the strategic plan for information resources. 2) Clearly define the problem to be solved, and thoroughly understand the problem domain. 3) Pay particular attention to the legal (and ethical) feasibility of the proposed system. 4) Fully understand both users’ concerns about the development project and their expectations of the operational system. 5) Employ management techniques designed to keep the attrition rate for developers within acceptable limits. These are ingredients that should be incorporated in any development project. Cause for Hope Viewing this oversight in a positive light, developers of future projects should be encouraged to know that they can substantially enhance their chances of success simply by doing thing right. Another reason to expect that future expert systems efforts will be more successful than the early ones is the fact that the technology has changed in some respects. Not all new expert systems are being constructed from the same components as the early ones. A big breakthrough has been something called a neural network, or simply a neural net, which make it possible for a knowledge-based system to actually improve its performance over time. This valuable ability can provide the system with a certain measure of the judgment and intuition ingredients that make for good business decisions. Neural Networks A neural network, commonly called a neural net, is a mathematical model of the human brain that simulates the way that neurons interact to process data and learn from the experience. Neural net design is a bottom-up approach, since it looks at the physical brain for inspiration in the creation of intelligent behavior. In contrast are the top-down approaches that have been developed by proponents of the more traditional AI areas mentioned earlier.
10. Biological Comparisons The design of neural networks has been inspired by the physical design of the human brain. The component of the brain that provides an information-processing capability is the neuron, which consists of. Dendrites specialize in the input of electrochemical signals, the soma process the signals, and the axon provide output paths for the processed signals. Figure 16.14 illustrates two neurons. Dendrites form a dendritic tree, a very fine, branch-like region of thin fibers around the cell body. Dendrites are the input components of the cell. They receive the electrochemical impulses that are carried from the axons of neighboring neurons. Axons are long fibers that carry signals from the soma. The end of the axon splits into a tree-like structure, and each branch terminates in a small end bulb that almost touches the dendrites of other neurons. The end bulb called the synapse. Each neuron may be connected to a thousand or more neighbors via this network of dendrites and axons. The soma is the processor component of the neuron. It is essentially a summation device that can respond to the total of its inputs within a short time period. The aggregation of signals is compared to an output threshold, which is the level of stimulation that is necessary for the neuron to fire or send an impulse along its axon to other connected neurons. The strength of the synaptic connection between the axon of the firing cell and the dendrite of the receiving cell determines the effect of the impulse. Applying the Systems Approach That the sequence can be described much, or perhaps most, of computing activity---and that includes human computing as well. The soma in the human brain is the processor. It receives inputs by means of dendrites and produces outputs by means of axons. In terms of the computer schematic, the soma is the “central processing unit,” the dendrites are the “input devices,” and the axons are the “output devices.” Not only is the electronic computer a reflection of the systems approach the human brain is as well. Through this very simple mechanism, input signals from neighboring neurons can be assigned priorities or weights in the soma’s accumulation process. These weights most likely serve as storage or memory for the network. Even though the response time for a single neuron is approximately a thousand times slower than the digital switches in a computer, the brain is capable of solving complex problems such as vision and language. This is accomplished by linking together a tremendous number of inherently slow neurons (processors) into an immensely complex network. The number of neurons in the human brain has been estimated to be around 10, and each neuron forms approximately 104 synapses with other neurons. This is an example of parallel distributed processing (PDP), which allows each task to be broken down into a multitude of subtasks that are performed concurrently. The Evolution of Artificial Neural Systems Interest in modeling the human learning system can be traced back to the Chinese artisans as early as 200 B.C. However, most researchers consider the development of a simple neuron function by Warren McCulloch and Walter Pitts during the late 1930s as the real starting point. The output from a McCulloch-Pitts neuron has a mathematical value equal to a weighted sum of inputs.
11. Hebb’s Learning Law One of the most famous learning rules was proposed in 1949 by Donald Hebb. Hebb’s learning law states that the more frequently one neuron contributes to the firing of a second; the more efficient will be the effect of the first on the second. Thus, memory is stored in the synaptic connections of the brain, and learning occurs with changes in the strength of these connections. The First Neurocomputers In the early 1950s, Marvin Minsky developed a device called the Snark, which is considered by many to be the first neurocomputer, or computer-based analog of the human brain. Although the Snark was technically successful, it failed to perform any significant information processing function. In the mid 1950s, Frank Rosenblatt, a neurophysicist at Cornell University, developed the Perception, a hardware device used for pattern recognition. The Perceptrons, combined with a simple learning rule. The Perceptron was able to generalize and respond to unfamiliar input stimuli. The success of Rosenblatt’s work fueled speculation that artificial brains were just around the corner. However, Marvin Minsky, an AI pioneer, and his colleague Seymour Papert demonstrated that the perceptrons of Rosenblatt could not solve simple logic problems. Their demonstration put a temporary damper on neural net research. The Artificial Neural System The artificial neural system (ANS) is not an exact duplicate of the biological system of the human brains, but it does exhibit such abilities as generalization, learning, abstraction, and even intuition. An ANS is made up of a series of very simple artificial neuron structures or neurodes. These structures are often referred to as perceptrons because of the influence of Rosenblatt. However, they are a direct extension of the mathematical model developed by McCulloch and Pitts. These artificial neurons are the processing elements of the ANS architecture. The neuron sums the weighted inputs from its neighbors, compares this sum to its threshold value, and passed the result through a transfer function. The transfer function is a relationship between the output of the weighted sum and the threshold value of the cell. When the weighted sum exceeds the threshold value, the neuron “fires.”
12. The Multi-layer Perceptron these simple neurons are combined to form a multi-layer ANS, referred to as a multi-layer perceptron. Within each one, the input nodes are linked to the output nodes through one or more hidden layers, as illustrated in Figure 16.16. The multi-layer perceptron is a feedforward network, meaning that the flow of data moves in only a simple direction, from the input layer to the output layer. However, the hidden layers permit an interaction between individual input nodes. This interaction allows a flexible mapping between inputs and outputs that facilitates their training. Network Training A neural net is not programmed in the traditional sense. Rather, it is trained by example. The training consists of many repetitions of inputs that express a variety of relationships. By progressively refining the weights of the system nodes (the simulated neurons), the ANS “discovers” the relationships among the inputs. This discovery process constitutes learning.
13. W1 Y1 W2 y3 W3 y3 Wn-1 Y yn-1 Figure 16.15 A Single Artificial Neuron
14. Putting the Artificial Neural System in Perspective The ability to learn based on adaptation is the major factor that distinguishes ANS from expert system applications. Expert systems are programmed to make inferences based on data that describes the problem environment. The ANS, on the other hand, is able to adjust the nodal weights in response to the inputs and, possibly, to the desired outputs. Because of its learning ability, the ANS is insulated from the shortcomings that plague expert systems in terms of adapting to changing conditions. During the coming years, more and more expert systems will be developed that incorporate neural nets, giving the systems combined ability to provide expert consultation and to improve their own expertise over time based on learning. Putting Knowledge-Based Systems in Perspective In 1956, when John McCarthy and his group coined the term artificial intelligence, most everyone else in the newborn computer industry was struggling to solve well-structured problems like payroll and inventory. Since that time, computer and information scientists have continually pushed back the frontiers of knowledge in the AIS, MIS, DSS, and virtual office. Most of those challenges have been met, and the applications are doing a good job of keeping up with the technology. For example, when inroads are made in such technology areas as multimedia and compact disks, they are incorporated into system designs.