After years of hype and failed expectations, artificial intelligence is finally beginning to make some real money.
Once the theme that launched a thousand science fiction novels, artificial intelligence is now capturing investment dollars as well as the imagination. Entrepreneurs, venture capitalists and some of the largest companies in the world are scrambling to hoist themselves up those few branches of artificial intelligence research that are bearing fruitful results.
The branch supporting the heaviest expectations is the so-called "expert system." Expert systems represent the tangible product of "knowledge engineering," the discipline of systematically transforming human knowledge skills into machines. Unlike conventional computer programs, which use special languages to get the computer to perform calculations, expert systems contain rules; if the computer follows these rules, it can effectively simulate the expert's role.
After carefully interviewing experts and analyzing what they do and how they do it, a knowledge engineer divines the rules that an expert uses and then crafts a system that can replicate the expert's reasoning skills in a computer program. The program usually consists of chain after chain of IF/THEN rules: IF certain criteria are met THEN consider this option.
By blending facts, rules of thumb and relevant general knowledge about a specific subject, the expert system can often serve as a substitute for the human expert. If experts can be valuable commodities, then so can expert systems, reasons the market. That's where there's money to be made.
At the National Conference on Artificial Intelligence here last week, the bulletin boards were filled with want-ads from General Electric and RCA and Texas Instruments and dozens of smaller companies seeking knowledge engineers and qualified expert systems builders in subjects ranging from engineering to medical diagnosis to legal analysis. The expert systems industry, such that it is, is prepared to take off.
"The interest is enormous," says Esther Dyson, editor of RElease 1.0 and a computer industry consultant. "It's a breakthrough that's waiting to happen."
The concept of expert systems has been around for almost two decades. Dendral, a system that decides what organic molecules might look like based on spectrograph data, was devised by two Stanford scientists in the late 1960s. In the mid-1970s, another Stanford scientist designed MYCIN, an expert system to diagnose and prescribe antibiotics for infectious diseases. Both programs, and several others like them, enjoyed reasonable success when applied, and won considerable attention in academe.
But it wasn't until the success of R1 that business realized that expert systems could dramatically affect the bottom line. In cooperation with John McDermott, a computer scientist at Carnegie Mellon, Digital Equipment Corp. began creating in 1979 an expert system--R1--to help solve a multimillion-dollar problem.
DEC had discovered that, in a significant number of the computer orders it shipped, all the components to configure a customer installation properly weren't there. The expense of sending out the appropriate parts along with service engineers, the annoyed customers and the subsequent delay in getting paid was costing Digital hundreds of thousands of dollars a month.
McDermott spent weeks interviewing engineers and poring over technical manuals to determine whether the problem was susceptible to the experts systems approach. "It turned out it was the perfect size," says McDermott. "It required enough knowledge so that a program that can perform the task is interesting--but the amount of knowledge required for the task and the specificity of the configuration constraints is such that the problem does not push very hard on state-of-the-art knowledge engineering techniques."
The first version of R1 had over 250 rules, which grew to 750 as the system was rolled out into the field in 1980. Today, R1 has over 2,500 rules. Digital estimates that R1 has been used to help install computers for over 30,000 customers. Most importantly, the company estimates that it has saved between $7 million and $10 million over the last three years.
"Conventional programming techniques just weren't applicable to the problem," said McDermott. "There would have been a combinatorial explosion vast multiplicity of choices . There were too many variables to consider."
"The significance of R1 was that the real world could use AI artificial intelligence techniques to solve problems," said Stephen Polit, a DEC executive, "There's now widespread expert system activity at Digital." The company recently set up a Knowledge Engineering Advanced Development Group to apply expert systems to such areas as office automation, hardware fault diagnosis and integrated circuit design.
Through Bell Laboratories, American Telephone & Telegraph Co. has used expert systems to save it time and money--although spokesman don't know how much. ACE, for Automated Cable Expertise, is designed to both monitor and analyze problems associated with telephone line maintenenance. Located in New Jersey and linked to Southwestern Bell via a long-distance phone line, ACE uses hundreds of rules as it sorts through maintenance records in Fort Worth trying to ascertain future trouble spots.
According to a Bell Labs spokesman, ACE can do a maintenance survey in an hour that would take a human a week. ACE does its survey overnight and has a status report waiting for the appropriate Southwestern Bell managers the next morning.
"Expert systems is a normal evolution of information processing systems," says Gregg Versonder, one of the Bell Labs scientists who created ACE. "But it's an order-of-magnitude leap in capability."
Essentially, Vesonder claims, ACE doesn't just file mainteneance information into a data base, it "knows" how to make vital decisions pertaining to telephone network maintenance by drawing on that data base. It could be used in any of the Bell operating companies with only minor modifications, and conceivably could save phone companies millions of dollars a year.
The key problem in the expert system approach is that it has to be tailored to specific problems. "Each time you do one, it's different," says J. Morris Tenenbaum, who directs artificial intelligence research for Schlumberger's Fairchild subsidiary. "It's a craft industry at this point."
Tenenbaum and others believe that the industry must find "generics"--experts systems concepts and structures that can easily be applied from industry to industry. "It won't be a volume industry until you have those generic applications," Tenenbaum says.
For example, R1 configures computers. But the concept of configuration could cover industries as disparate as hospital supplies and building construction. Expert systems could be designed to track inventory levels, demand and turnover in a variety of different industries.
The challenge is coming up with an expert system that "is general enough for wide application but close enogh to the surface to be made into specific applications," says McDermott, R1's creator.
Teknowledge, a two-year old Palo Alto Expert Systems company founded by a group of Stanford computer science professors, thinks it is taking the best approach to the generics concept. The company, which grossed $2 million in fiscal 1983 and expects to make "several times that" this year, according to president Lee Hecht, offers companies CAKES--Computer Aided Knowledge Engineering Systems.
Essentially, Teknowledge will take a core group of company employes and both train them and give them the computer-based tools to design their own knowledge systems.
"It's like we sell the Xerox machine," says Hecht. "We are the Xerox machine. You take an expert, make 200 copies of him, and send him out into the field."
Teknowledge will be most successful, Hecht feels, if it lets companies create their own expert systems rather than do the designing for them. The company has already worked on CAKES with Boeing, NCR Corp. and the French oil company Elf Aquitane.
Xerox Corp. is taking a similar approach with its LOOPS language. LOOPS, which Xerox has just formally introduced, is an experts system language. By becoming a LOOPS programmer, an individual has the computer tools to design his own expert systems. Teknowledge is currently one of the LOOPS testing sites.
Several industry analysts believe that expert systems could become one of the most profitable computer software areas by the end of the decade--but, they say, it is still to early to say whether knowledge engineering techniques will improve enough to make it the cornerstone of a new generation of computer software.
The essence of the question, they feel, is knowledge representation: how does one adequately present the human cognitive process on a silicon chip? The answer to that blends psychology, computer software and computer technology. All those areas are still evolving.
However, according to Stanford professor and Teknowledge founder Doug Lenat, the next generation expert system will know how to learn from experience--and then create additional rules for itself. In essence, an expert system that can learn to become more of an expert.