Class: LIN 463 sec 001
Final draft due: [23 Apr 2003]
term paper
Expert Systems were invented as a way to decrease the reliance by corporations on human "experts" -- people who apply reasoning and experience to make judgements in a specific field, such as medicine, insurance underwriting or the operation of a power-plant. Hence, an expert system should include a database of facts and a way of reasoning about them. In many, but not all, applications it is also helpful to have a way for the system to reason with probabilities or non-Boolean truth values. Expert systems are also sometimes referred to a "knowledge-based systems".
In classical AI many different reasoning methods have been tried. A few of the common ones are "forward chaining", in which conclusions are drawn from a set of facts by using modus ponens, syllogism, and other simple tools of logic; "backward chaining", which uses trickier logic, such as modus tollens; and neural nets. Most expert systems simply use forward chaining and backward chaining, with some non-Boolean component such as Fuzzy Logic only where necessary. Expert systems tend to be more practical than AI in general.
It is quite possible to build an expert system in a conventional programming-language, such as COBOL, C or Java. However, much of the machinery inside an expert system can be abstracted away from any specific domain, and the main criterion in the success of an expert system is its ease of use and maintenance, not it's ability to make decisions in a fraction of a second. Therefore, it is possible to build a "knowledge system shell" which can then be prepared for almost any domain simply by listing rules and data in a standard form. Few expert systems are written in LISP, because most LISP implementations lack robust user-friendly input-output routines.
A good knowledge system shell includes I/O routines, a way to accurately and generally represent facts, and an easy, efficient, accurate way to give the system inference-rules. However, even the best expert system shell is limited by the problem domain to which it is applied. One researcher divided problem domains into four categories:
"Class 1. ... if the effective domain decompositions are not known and the available domain knowledge is limited to the set of allowable actions and constraints. An example of such a problem is maze traversal, where the knowledge about the structure of the maze is not available a priori.
"Class 2. These are the problems where problem solving task-reduction rules are known. However, the overall problem-solving process is not structured. Many game-playing machines (bridge, backgammon, etc) are examples.
"Class 3. Problems where the problem-solving process is well-structured, while the resulting plan exhibits flexibility... Examples of such problems are cooking instructions and trip planning.
"Class 4 problems are the problems where both the problem-solving strategy and the resulting plan have a fixed unalterable architecture. In this case, the solution is usually in the form of plan parameters that then fit into the predefined plan schema. An example of such a plan could be crop management planning." (Martínez 146-147)
Classes 4 and 3 are especially amenable to implementation as expert-systems, but some of the prototypical systems fell into Class 2.
It seems like almost every book or article written about expert systems refers to two early attempts, DENDRAL and MYCIN. The DENDRAL program dealt with the structure of chemical compounds. The MYCIN program "[helped] physicians choose appropriate antibiotics for patients with severe bacterial infections." (Text 219; see also Shadbolt 2).
Another early, commercially successful expert system was PROSPECTOR. It analyzed data that might have to do with the location of oil wells and suggested a most-profitable place to drill. Perhaps these early systems were easy to construct because the principles involved were widely known by academic experts in the relevant fields and available in published literature.
One system that was widely mentioned in subsequent papers was an application developed by Digital Equipment Corporation. Apparently DEC's computer-systems could be installed in hundreds of different configurations, and this program would evaluate a customer's needs and suggest an appropriate system. I know that the Physics department at MSU was using a VAX until a year or two ago, but new VAXen were probably not sold at all after about 1988. Today, the high-end server market has so many competing vendors that no one corporation is likely to cover all of its customers' needs from its own stock. However, the system for choosing an interactive laptop that Dr. Chai demonstrated the other week certainly bears a resemblance to the old VAX chooser.
By 1985 or so, a significant number of expert systems had been written and deployed, and the industry seems to have had the critical mass to enter a phase of wild growth. Within the next two years, dozens of books were published, many of which were collections of academic research-papers. A significant fraction of these followed a basic pattern: researchers had implemented a sample application for a specific company in a specific domain. They reported success, and encouraged other companies in the same domain to develop similar systems. They often reported the exact cost for the computer-hardware and software-packages that were involved. Apparently "workstations" costing from $5000 to $25000 each were the preferred development platform, but some reported that they had gotten their application to run on an IBM PC with acceptable performance.
These efforts in more specialized domains required the cooperation of one or a few human experts, generally experienced workers close to retirement. This caused public discussion of ethical and labor-relations issues. As the Knowledge Systems Design website says, "Loss of an employee does not automatically mean loss of their knowledge. Expertise may now be readily available and accessible day or night. Used as a training tool, an expert system provides the knowledge of the seasoned professional to the more junior employees." People expressed concern that the experts' knowledge would be drained from them and they would then be out of work; however, it was generally said that the experts were at the point of retirement anyway.
Results were published predicting a billion-dollar industry by the year 2000, based solely on the rapidly growing sales of expert-system software up to that point. More papers were written , and many conferences were held.
If the early expert-system applications were truly limited by the computer hardware of the time, advances at the end of the 1980s should have made them much more widespread and really opened up the market as had been predicted. Instead, there appears to have been a significant decline in systems that were identified as "expert" and studied as such after 1986, which continues until today. Expert systems still merit a mention in an AI textbook, and occasional software-manufacturers make the claim that their (proprietary and nonpublic) code uses "expert-system technology", but the industry as such appears to be dead or dormant. One question that arises is: why?
One explanation is that computer-professionals simply found other things to interest themselves in. The end of the 1980s saw the release of mass-market graphical-user-interface systems, such as the Apple Macintosh and the Microsoft Windows environment. In 1989 the Berlin Wall fell, and computer networks between East and West Germany also served to connect the United States with the Soviet Union. In 1991 a Finnish student released the "Linux" operating-system, and revitalized interest among programmers worldwide in basic operating-system technology. Perhaps most significantly, researchers began playing with "hypertext", a way of giving more control to the reader of a document. In about 1994 the "World Wide Web" achieved commercial power in less time than it took to go from one issue of a magazine to the next, and the "dot-com boom" was born.
An alternate explanation also hinges on the emergence of hypertext and, to a lesser extent, graphical-user-interfaces. In 1989 a group of theorists (Barlow et al) applied graph theory to what they saw as a common ground between hypertext and expert systems. They called it "expertext".
Beginning programmers often expect a user to rigidly follow certain rules, almost as if that user were a very simple computer himself or herself. In business environments, when computers were first invented, special individuals (computer operators) had the responsibility of "interpreting" the computer's output. The same people sometimes had to "fix" or "prepare" input to the computer as well; they might have to punch paper cards, or type up data-files a certain way, or patch values into a program to reflect current conditions. However, as computers became more widespread and powerful, and more people became "computer-literate", the artificial human layer became insupportable. To the extent that expert systems were only an attempt to soften the human-machine interface, they went in the wrong direction. The typical expert system asked questions in a predetermined order according to what the original expert thought was important, until a conclusion could be reached. Some companies used their expert systems to hide proprietary business processes from lower-level employees; legally, this became indefensible. Some users soon found that an expert system was too slow, and it was relegated to merely a training-function for new employees.
On the other hand, hypertext leaves the decision-making to the user of the system. The thought-pattern of a human expert may still contribute to the system, but it's not forced on the user. He may even think he freely chooses to do something at the same moment when he is surely "led down the garden path". Hypertext may seem less arbitrary when brought as evidence in a court of law, and it offers less of a guilt-factor against users who have partially memorized a system of rules and skip referring to the system when they get to a certain point.
GUIs also have advantages. They take advantage of the high bandwidth of the human visual and fine-motor systems, while traditional expert systems were more like the classic candidate for the Turing test: two entities chatting over a half-duplex text link. Instant messaging is an easy way to make killing time less boring, but people find it easier to believe and act on suggestions when they are presented with multimedia supporting data.
A few companies still sell what appear to be traditional expert-systems software and services. One is Acquired Intelligence (Acquired). They sell a framework for building expert systems and also offer consulting-services to build custom ones.
Another company is Knowledge System Design, Inc. While they claim to have been in business since 1992, their list of clients is much smaller than that of Acquired Intelligence. They mention a tool for building expert systems on their website, but it is not clear whether they have developed it or simply prefer to use it.
Research continues to be done in Expert Systems. For instance, at this university in 2001 a system was developed to organize the manufacturing-processes for composites (Martínez). An example of a something made from a composite is a Fiberglas canoe.
There are consumer-visible systems out there that seem to be influenced by expert-system principles. On would be the IBM laptop-selector, mentioned above. Microsoft and Corel include agents in their help-systems which try to integrate the search and browsing techniques that have traditionally been involved in reading online documentation. Microsoft has the "Office Assistant"; Corel has the "PerfectExpert". The Office Assistant monitors the user's actions and displays error-messages in a more-friendly manner. It even makes suggestions when the user appears to be making a mistake.
While the Office Assistant is installed by default, most people soon turn it of. Windows Help files can also be viewed as hypertext, and people seem to like interacting with it better that way.
One fear that has been expressed occasionally in science fiction is that human society will become decadent and infantile; that we will leave off the pursuit of knowledge as soon as it seems like a machine can do it better, and let the machines do all our thinking for us. While it is less commonly expressed than the theory that robots will physically and maliciously take over the world, it is implied in many stories. A multipurpose expert system with a natural-language interface and a self-extensible knowledge-base promises to be the perfect oracle. After World War III and ten thousand years of barbarism, will men find a cave where they can learn wisdom from the solid-state solar-powered computer and reuse all our fashions and technology while never developing any of their own?
In my searching for information on expert systems, I found no papers published by the NSA, British Intelligence or the KGB. This is not surprising, since these organizations conduct many of their activities in secret, but it is quite possible that expert systems are being used to organize and summarize the input from clandestine investigations around the world.
Indeed, most expert systems in corporations would be repositories of trade secrets, and many expert systems in government would be items with national security significance. In addition, their very existence could be a valuable clue for competitors or a legitimate grievance-point for customers or employees. It is quite likely that the world will soon be controlled by many hidden decision-making computers, if it is not already.
A sufficiently intelligent computer would make a perfect Platonian dictator, always making laws in the best interest of the kingdom as a whole and without even the threat of death. A computer with the capacity to model the thoughts and actions of millions of people would make an optimal Machiavellian prince. Will we see legislative bodies, or at least executive branches of governments, replaced by completely automated systems? Already, secretaries and middle-managers have been downsized out of many organizations with the help of technology.
Expert systems are proven technology, but they take significant time and knowledge to assemble. Traditional AI promises to make a self-training expert-system, but has not delivered yet. The fusion of the two, should it ever occur, could cause major changes in the world.
Conclusion
Expert systems have been around a while, and they are definitely here to stay. While they seem to consume only a tiny fraction of the resources used to develop and run computer applications, they probably cause some effect on almost every facet of modern life.
There are still significant research opportunities in expert systems, and there are opportunities to do work in private industry. Computer-programming experience is helpful in doing much of the work, but not strictly necessary, as much is done with already-built Knowledge Engineering shells.
Acquired Intelligence, Inc. Corporate website. Internet: http://www.aiinc.ca Accessed 19 Apr 2003.
Barlow, Judith, et al.. Expertext: Hypertext-Expert System Theory, Synergy and Potential Applications. In Research and Development in Expert Systems VI. Cambridge: Cambridge UP, 1989.
Knowledge Systems Design, Inc. Corporate website. Internet: http://www.knowledgesys.com/index.html Accessed 23 Apr 2003
Martínez-Bermúdez, Iliana. A Framework for Knowledge Acquisition, Representation and Problem-Solving is Knowledge-Based Planning. Thesis, Michigan State University, 2001.
Shadbolt, Nigel. Expert Systems --- a natural history. In Research and Development in Expert Systems IV. Cambridge: Cambridge UP, 1989.