Expert systems (also known as knowledge-based systems) are artificial intelligence programs that implement the rule-based reasoning of human experts in a relatively narrow "domain." Although they're similar to non-artificial intelligence computer programs, they have several unique characteristics and advantages. This is an area where a little jargon will have you sounding like a "knowledge engineer" (what a "kewl" occupation) in no time, so I'll throw in some technical terms here and there.

Expert systems use branching chains of if-then rules. There is a "shell" program that typically consists of an “inference engine” (rules for searching the if-then tree), a user interface, and the data being worked on. The “knowledge base” is separate from the shell, so that a business or other user can buy a shell and create its own knowledge base. The expert system may also include a special program for user interface for inputting, correcting, updating and maintaining the knowledge base rules.

An expert system can proceed along two or more if-then inference branches at a time, generate questions to the user if it needs more information, give a probability-based answer if it can't give a definite answer, and explain how it arrived at its conclusion. The individual rules are usually simple and easy to understand, so that a human user can follow the reasoning chain and troubleshoot it. The rules are not hidden inside the computer software code as in traditional computer programs. The expert system has the flexibility to use the rules in any order. To change the system, one usually only need add a rule, not re-write the entire program. Also, one can usually remove a rule without destroying the program.

Expert programs do medical diagnostics and machine diagnostics (e.g., computer, automotive), identify geologic formations for natural gas, underwrite mortgages and insurance, control military logistics and plan large events. They can work in any decision-making area that has many rules or requirements, and are especially important for double-checking, making sure all the questions get asked.

The first expert system was invented by Edward A. Feigenbaum, who was educated here in Pennsylvania at Carnegie Mellon University. In fact, he was in Herbert Simon's class when Simon announced the breakthrough made by Logic Theorist (discussed in the October 2007 column). (This Feigenbaum is not to be confused with the mathematical physicist, Mitchell J. Feigenbaum, who was, however, born in Philadelphia.)

Feigenbaum and his collaborators designed Dendral to determine the structure of complex molecules from mass spectrometer data. In their second expert system, Mycin, the team separated the shell and knowledge base, which gives expert systems much of the flexibility that has made them so successful. Mycin diagnosed, and recommended antibiotics for, blood and meningitis infections. The first successful commercial expert system, XCON, configured computers (also known as “technical editing”) at Digital Equipment Corporation

A little example will help us understand the power and limitations of expert systems. Let’s see if we can program a robot to greet visitors at our Mac User Group meetings.

The first reasoning our expert human greeters probably do is determine whether they recognize a person who is entering the meeting as a member or previous visitor. Right here we have an obstacle because face recognition programs are not available. So we will have to have the robot greeter ask whether the human has been to a meeting before. If the human’s response falls into the category of “visitor,” the robot will then ask questions about what Mac the human owns, how recently it was acquired, whether the human has used Macs before and what the human uses, or wants to use, the Mac for. We’ll ignore speech recognition obstacles too.

How would our MLMUG-Bot greet four different visitors?

Visitor A says she bought her first Mac, a Mac Mini, last week and has never been to a meeting. MLMUG-Bot reasons: IF human bought Mac in the past 3 months AND never used a Mac before, THEN direct human to the Beginners SIG AND mention the New Users Lite meeting on fourth Saturdays and the MLMUG library of useful books.

So far, so good. Notice that the robot knows that “last week” is included in the time period “the past 3 months,” and that the visitor can say this is her “first Mac” and the robot knows she’s never used a Mac before. Expert systems have this kind of flexibility.

Visitor B says he was here last month and has decided to join. MLMUG-Bot reasons: IF human wants to become a member, THEN take human gently by the arm AND lead human to the Treasurer’s table AND say “You will love being a member; there are so many benefits. We have a Dutch-treat lunch after the meeting at Casey’s Dugout Saloon, my favorite because of the baseball theme, and I hope you’ll join us.”

Visitor C says she uses her G5 for her commercial photography business. MLMUG-Bot reasons: IF human is an experienced user AND uses the Mac for media endeavors, THEN direct human to the Multimedia SIG AND say “You’ll be thrilled to know that the main meeting presentation later is about digital photography.”

Visitor D says he has used Macs for years and was a visitor once a few months ago, but he just bought a MacAir to use for international travel and has some questions about Leopard. MLMUG-Bot reasons: IF human has a MacBook or other portable Mac, THEN direct human to the Intermediate and Advanced Users SIG AND say “This SIG also covers mobile computing and the presentation today is on travel bags and accessories.” AND, IF human has questions about intermediate or advanced use, THEN direct human to Q&A session during the Intermediate and Advanced Users SIG and general Q&A after the SIG meetings, AND introduce human to Expert MLMUGer who is standing near the entrance.

Oops! Here we are back at the face recognition problem. How will MLMUG-Bot recognize the members who are standing nearby? In addition, we’d have to keep the MLMUG-Bot up-to-date with information about new Mac products, since visitors frequently come to meetings when they’ve purchased something new, the location of the Treasurer’s table that day and information about the SIG topics for each meeting. That’s besides the original programming of all the rules, which would be considerably longer and more complex than the above example. If we had two visitors who were friends and did not want to attend separate SIGs, one of whom was an experienced user, and one a beginner, we might see that classic cartoon image of smoke coming out of our MLMUG-Bot’s brain. I doubt that our human greeters have to worry about being replaced by robots any time soon.

I have to confess that, as much as I love AI, I became concerned while writing this month’s column. Could an expert system encroach on my own territory of legal expertise? I was relieved to learn that humans have had this fear with every technological advance, including writing, and that such fears have not been realized on the whole, although technological advances can in fact cause disruption for individual humans.

To defuse the fear a bit further, there are numerous disadvantages to today’s expert systems. Unlike human experts, expert systems can't go beyond their if-then rules. They can’t interact with each other or refer a human to another expert. They require a lot of memory, which was one of the reasons they weren't invented even earlier. By 1987, XCON had grown to approximately 10,000 rules. If there are too many rules, the system becomes unwieldy and can produce bizarre results. As we saw with the MLMUG-Bot example, expert systems take a lot of time to maintain and update. Also, experts often do not reason by thinking in terms of rules. “Case-based” reasoning, used in the medical and legal fields, is an area of current research.

So where are all the expert systems these days? Commercial use proliferated in the 1980s, but a number of factors contributed to changes in the field, and there has been less visible, theoretical progress in recent years.

Part of the change is a change in perception. Expert systems are now all around us, "embedded" in the computer systems we use, rather than in stand-alone LISP software language systems. They're in the online Apple Store's procedures for helping you configure your Mac, peripherals and accessories, and Macintosh Mail's Junk Mail filter. They're used in various ways in human genome sequencing, and are the topic of futuristic hopefulness for Internet search engines, which never seem to be able to do quite as much as we’d like them to.

As with many other areas of AI, greater progress in learning programs and common sense programs seems to be necessary to take expert systems to the next level. Another important avenue is “nonmonotonic” reasoning, i.e., how to deal with exceptions that change general rules. If a visitor to the MLMUG meeting is an unaccompanied dog, would our MLMUG-Bot know to pet it and direct it to the muffins before the humans eat them all?

Sources and other information for this column:

Definition of "kewl":
http://www.urbandictionary.com/define.php?term=kewl

Many links to recent expert systems research: http://www.aaai.org/AITopics/html/expert.htm

Crevier, Daniel; AI: The Tumultuous History of the Search for Artificial Intelligence, Basic Books, 1993.

Example of expert system use in genome sequencing: Matise, Perlin and Chakravarti, “Automated construction of genetic linkage maps using an expert system (MultiMap): a human genome linkage map,” Nature Genetics 6, 384-390 (1994), lhttp://www.nature.com/ng/journal/v6/n4/abs/ng0494-384.html

An expert system was held to have engaged in the unauthorized practice of law: Reynoso v. Frankfort Digital Services, Ltd., 477 F.3d 1117 (9th Cir. 2007), http://www.ca9.uscourts.gov/ca9/newopinions.nsf/780E8DE27F08D8A98825728F000085F7/$file/0417190.pdf?openelement

Kissell, Joe, “Getting to Know Apple Mail’s Spam Filter,” TidBITS, May 24, 2004, http://db.tidbits.com/article/07677.

Olsen, Stefanie, “Spying an intelligent search engine,” CNET News.com, August 18, 2006, http://www.news.com/Spying-an-intelligent-search-engine/2100-1032_3-6107048.html?tag=news.1