When MIS and AI converge - includes related articles on the use of expert systems at Transamerica and on the use of a KBMS-written temperature forecasting program at Southern California Edison

Software Magazine, May, 1988 by John Desmond

WHEN MIS AND AI CONVERGE

It began as a seed of an idea in the mind of Larry Harris. It was not the first idea Harris had ever had.

He is the creator of Intellect, a tool which uniquely combines elements of artificial intelligence theory within the esoteric and often intimidating world of IBM mainframe operating systems software.

"Our company is unique in that we have people with 10 years of experience in the IBM AI world," said Harris, the founder of AICorp of Waltham, Mass. It is a 13-year-old firm with a consistent product strategy: exploit artificial intelligence theory to build software tools for commercial data processing. "Our company is positioned at the intersection of AI and the IBM world," Harris said.

KBMS, Knowledge-Based Management System, a tool used to build expert systems, would be a follow-on to Intellect, the natural language query tool which is now into the maturity phase of a typical software product life cycle. AICorp needed a new product.

The concept of KBMS took shape after university research into expert systems had matured enough to be used in commercial products, Harris said. It was not clear at the outset what synergy would exist between Intellect, which generates SQL code from English statements, and KBMS. But as the development progressed, "We found a tremendous amount of overlap, especially in the database interface," he said.

As software professionals know, development takes time. And for companies that write software to be sold as packages, fulfilling the software development life cycle--birth, life and aging--is only part of the necessary mix. To that must be added marketing, sales and support; and each presents its own challenges.

Like any good book, play, song or other creative undertaking, a truly new software product starts with an idea: "You start with a vague feeling of what you want, and you refine it through several thousand iterations," according to Harris.

"We laid out what we wanted to achieve in seven or eight areas. Then we thought about how to get there, about how far we'd have to push in each dimension to accomplish what we wanted. In each area, we had aggressive, almost breakthrough goals that we wanted to achieve. Our notion of database interface was so far removed from what people understood, we were concerned about being able to communicate it."

Because AI is so unlike many other areas of software development, in that academic researchers and not practitioners lead the way, the AICorp developers had to be careful not to go overboard on innovation. Software development of any kind is risky business, but to blaze new trails in software is much more risky.

So a reality check was required. This was done by a search of the available literature, by analyzing competitive products, and by talking to the user community. As a "periphery check," the developers also consulted with researchers in the MIT AI labs.

According to Jeff Hill, AICorp's senior vice president of corporate R&D, "We set criteria; we didn't want to pick features so new that they weren't important. We wanted features with a major success story. We weren't trying to push the state of the art from the research perspective." This process resulted in at least one feature--rule induction--eliminated from the design. "We came very close to putting it in; it was waiting right outside the door," Hill said.

They also considered neural nets, the pattern-matching expert software, but its commercial potential was judged to be too distant. Once this literature search and consulting phase, conducted from August to October of 1986, was complete, the developers locked into the product design. This encompassed decisions on what features to include in the product, and how to make them work together.

KBMS FEATURES

The eight essential features of KBMS are the following.

1. The AI engine. Offers the four paradigms of reasoning: forward and backward chaining, object-orientation and hypothetical reasoning.

2. Call-in. The ability to be invoked from another program, especially Cobol programs, which are the language of 90% of commercial business applications.

Harris: "It's a tricky concept; it's very deceptive. If you want to be callable in a useful way, you must be able to communicate down and then back up. The notion of calling a database and having nothing to say doesn't help a user solve a problem. So, to take a system that is not callable today and design it to be callable is very difficult."

(Interpretation: Inference, Intellicorp and Teknowledge, early suppliers of expert system design tools who did not design their products to work within IBM operating systems, have a tough row to hoe to make them do so.)

3. Database integration. A tighter link than that provided by a simple database interface.

Harris: "Here we wanted to break new ground. In all the examples up to now, you program call statements to your own external routine, which means the expert system shell doesn't help you at all with accessing databases. Two years ago, we first began to see systems that allow the embedding of a database language, such as SQL, in the rules themselves.