Reuters estimates that they save over one million dollars. IBM’s estimates are in the tens of millions. The IRS estimates that they collect over one-hundred million dollars in previously uncollectible revenue. Just what are these organizations employing to save so much money?
Known as a branch of artificial intelligence, expert systems offer a way to capture and encode knowledge from experts. These computer programs emulate the way people think. Born in the labs of Stanford University over twenty years ago to help diagnose infectious blood diseases, expert systems have since moved into virtually every profession that requires human judgment.
Expert systems generally consist of three main parts: 1) the explanation generator and user interface, 2) the inference engine, and 3) the knowledge base. In the early days at Stanford, the knowledge base consisted of medical “rules” in the form of IF-THEN statements with an associated confidence factor. For example, IF the patient is experiencing symptom A AND symptom B THEN the diagnosis is X, confidence Y%. If the patient experienced symptom C, this rule would not even be considered. The selection of the rules is handled automatically by the inference engine.
The researchers found that if they replaced the medical knowledge base rules by rules from a different subject domain, the same inference engine could be used to select which rules are executed. This way, programmers don’t need to rewrite the inferencing code for different applications. The prototype inference engine was dubbed “emycin,” for “empty-mycin,” since the medicines used to treat the diseases diagnosed by the prototype system ended in “-mycin.”
Expert systems differ from conventional programming in that they do not process data sequentially. Instead, the sequence in which rules are executed is determined based on the information available and the information desired. There are two principal types of inferencing:
Backward chaining (goal-directed reasoning). This method is often used to classify information. For example, Digital Equipment Corp. recommends configurations for its mainframes based on data such as room length, width, etc. Certain pieces of equipment must be next to each other, others apart. XCON, Digital’s expert system, determines the optimum configuration. As the product lines change, only slight modifications to the code are needed.
Forward chaining (data-driven reasoning). This method derives new facts from available facts. Some financial institutions use this technique to flag suspicious foreign exchange transactions. Reports are generated instantly by linking their worldwide computers to their expert system. When the conditions of a suspicious transaction are met, managers are notified and can then investigate further. Needless to say, the rules for such systems are highly guarded.
Initially, expert systems were developed in LISP and in PROLOG. While many systems are still developed from “scratch,” today there are numerous expert system “shells” that contain the explanation generator and inferencing code. The developer only needs to enter the knowledge base rules and customize the user interface.
The rules are developed by debriefing experts in the selected field. “Knowledge engineers” extract this information and massage it into a form digestible by the expert system shell. This isn’t always easy. Experts often have the “know how,” but not the “say how.” Most shells can access database records, call graphics files, and be embedded in larger programs.
Expert system shells are available on all platforms and in all price ranges. These systems are used to operate “help desks,” to ensure manufacturing quality, to determine mass merchandise pricing, and to assist in many other areas that involve human expertise, not just processing power. What happened to the prototype in the medical field? The fine line between who gets sued if something goes wrong played a big part in its unpopularity.
