AI at work: from prisons to autos to space.

The Tennessee Department of Correction had a tough problem. Faced with a prison population of 50,000 and a welter of more than 2,000 rules specifying lengths of sentences, criteria for probation, and numerous other details of incarceration, department personnel had trouble coping. They found it especially difficult to calculate accurately the duration of sentences.

Required by a 1990 federal court order to improve prison conditions and reduce overcrowding, the department decided to seek a software solution to its management woes. It turned to Andersen Consulting, a firm specializing in the application of artificial intelligence (AI) techniques to business problems, to develop the necessary system.

Now in use, the $14 million Tennessee Offender Management Information System (TOMIS) oversees the entire process, from sentencing through imprisonment to release. "[TOMIS] has reduced errors, streamlined business activities, and provided more accurate and timely information," report Tim Beck of the Tennessee Department of Correction and David Reynolds of Andersen Consulting in Nashville, Tenn.

TOMIS stands as an example of how knowledge combined with some form of reasoning -- the basis of AI techniques-can be incorporated into information systems to solve complex business problems. It was one of 16 projects highlighted this week at the Innovative Applications of Artificial Intelligence conference, held in Washington, D.C.

"We were looking for real business successes," says Philip Klahr of Inference Corp. in El Segundo, Calif. "These applications are major accomplishments for AI." Klahr headed the committee that selected these projects from a wide range of submissions.

In most of the projects, the developers took advantage of computer programming techniques unique to AI research to capture certain types of knowledge -- for example, the intricacies of how a vehicle is manufactured, together with the expertise required to estimate the cost of parts when engineers propose a new or modified design. The resulting computer program could then act as an assistant by generating, investigating, and recommending possible courses of action, and then justifying its decisions.

The Ford Motor Co. started developing its computer-aided parts estimation (CAPE) system in 1988. Initially, "senior management was skeptical that such an ambitious system would work," says Adam Cunningham, a Ford programmer in Baslidon, England. Put together in stages, the system now incorporates about 40 percent of the entire manufacturing process and is already in daily use at Ford facilities in Great Britain and Germany.

"We think CAPE is a major technical achievement," Cunningham says.

He cites the example of a car's electrical system, which by itself requires about 400 parts, most of which Ford purchases from suppliers. Estimators usually require two weeks to work out the cost of the electrical parts in a new design. Using CAPE, they can obtain a more accurate estimate in 30 minutes. "You get vastly improved control over costs," Cunningham contends.

On a considerably smaller scale, researchers at NASA's Ames Research Center in Mountain View, Calif., and their collaborators have developed a knowledge-based system that helps astronauts perform scientific experiments during space shuttle flights. Known as Principal Investigator-in-a-Box, this system collects data, analyzes the status of experiments, and recommends changes in procedure to maximize the value of the research performed within the severely limited time available to mission specialists.

Running on a single portable computer (a Macintosh PowerBook 170), PI-in-a-Box has been delivered to NASA for use on the Space Life Sciences-2 shuttle mission, scheduled for launch Sept. 10.