Image provided by: Texas A&M University
About The Battalion. (College Station, Tex.) 1893-current | View Entire Issue (Oct. 25, 1985)
ARTIFICIAL INTELLIGENCE (Continued from cover) design circuits, but can you trust it with your health? One medical expert system, CADUCEUS, contains the medical know how of Jack Meyers of the University of Pittsburgh: 125,000 facts about internal medicine. The program, capable of recognizing 500 diseases by analyzing 3500 symptoms, is known familiarly as 'Jack in the Box'. Another expert system, DENDRAL, which iden tifies chemicals through their mass spectrographs, is a 'generator-based' rather than rule-based sys tem. It generates 'ideal' spectral patterns and then tries to match them to the real pattern it is analyz ing. Some Al experts believe a generator-based sys tem is closer to human intelligence than a rule- based system: to recognize a chair, a human compares it to an abstract concept of a chair. Nevertheless, rules are still important. Even DEN DRAL has to apply simple rules to narrow down the immense field of possibilities before it begins the time-consuming and expensive chore of generating possible spectrographs. Furthermore, one big advantage of rule-based systems is that informa tion can be added one piece at a time, as it is obtained and also subtracted one piece at a time, to see how im portant each is to the total system. ^ Other programs, like TEIRESIAS, designed by Randall W. Davis of MIT, make expert systems al most self-replicating. TEIRESIAS keeps statistics on the freguency of incidence and correlation of symp toms, providing the basis for new rules. It also helps the programmer analyze new rules. For instance, it might say, "You have said nothing about how the disease enters the body, whereas in most diseases with rules like that one, you have specified how the disease entered the body” "Artificial . techniques are becoming simply good . practice.” Davis is now pioneering yet a third method, a 'model-based' expert system, to diagnose comput er hardware problems. Given the plan of a microcir cuit and its inputs, the program builds a model of how it is supposed to behave and calculates the outputs. If the calculated outputs do not match the real outputs, the program switches off each component of the simulation one at a time to locate the problem. The Al of the program is that the rules come from electrical engineers and that it communicates in English. Yet another development has been getting machines to imitate the brain's ability to do many things at once. Researchers have started to de sign computers to process 'in paral- executmg many mstruc- tions at the same time rather than walking straight through a program one line at a time. Supercomputers, like the Grays, a family of number crunching machines used for such monumental tasks as cracking cipher and solving high-level physics problems, have proven this archi tecture is faster. Parallel processing has been carried to its logical end by the Thinking Machines Corporation (TMC), an artificial intelligence firm founded by Al guru Marvin Minsky. They have built what they call The Connection Machine, which consists of one million interconnected processors executing symbolic in structions for pattern recognition and learning. Even their test methods are illustrative of the power of these machines. "We have database jocks who feed it dictionaries and telephone books and then look things up to see how fast it is," said a TMC programmer. Most expert systems are programmed to explain their reasoning on request, making them education al tools as well. "If you ask a program how it makes its decisions, it will tell you," said one expert. An other researcher commented, "Lots of people have requested the code for DENDRAL because it's a great advanced spectromentry course." Many Al researchers share the hope that building computer simulations of intelligence may provide a way to understand human intelligence better. In order to write an expert system program, an Al spec/a/ist must nrstanaTyzenowanexperttninks^ This forces the programmer to define 'gut feelings’ very precisely or organize vague concepts into spe cific steps. Using artificial intelligence programs to simulate the workings of the human mind has other advan tages for the researcher. "Computer programs ex hibit unlimited patience. They require no feeding, and they do not bite," noted MIT's Al LAB Director Patrick H. Wilson. But some are skeptical of the benefits of Al for cognitive science. "When you made a mechanical mouse that moved, it didn't help you understand mice better," says noted MIT physics professor Phil ip Morrison. It is unclear whether Al will ever produce a duplicate of the human brain, but the pos sibility and number of years required for a project of this magnitude are hot topics of debate in Al circles. Flowever long it takes, the race is clearly on to see who will be the first to make a computer walk and chew gum at the same time. ◄ Diana ben Aaron was editor-in-chief of MIT's stu dent paper. The Tech, last year and has just re ceived a Bachelor of Science in science writing, and materials science and engineering. Richard Downs is an illustrator currently freelanc ing in Los Angeles. A recent graduate of Pasadena Art Center College of Design, he enjoys portraying an idea using abstractions and symbols. 4 * » 'i| • WBMjMPWKM » • * » • */^ flff, • * • * * u mwSi L©4v£«q/// How to Prepare for a Career in Al A s in all fields of computer science, industry needs Al specialists faster than universities can turn them out. A degree in computing is almost a prerequisite for breaking into the field. If you want to start work right away, you should major in computer science as an undergradu ate, and try to get a thesis or coop job in artificial intelligence. However, if you have the patience for a four-year graduate program in artificial intelligence, you can approach the field from an undergraduate back ground in psychology, cognitive science, linguis tics, philosophy or mathematics. You can throw out your BASIC and FORTRAN programming manuals. The languages of choice in artificial intelligence are LISP and PROLOG. These languages are designed to promote symbolic pro cessing of logical concepts--the sine qua non of artificial intelligence. Another popular option is writing your own lan guage, tailored to the specific Al application. Win's Hiring ii Al The biggest user of Al talent, currently funding al most all Al research, is the military. The Pentagon wants to build an electronic co-pilot for combat air craft to take over the pilot's complex array of cock pit tasks. They plan to use the 'Star Wars' Strategic Defense Initiative to spawn a variety of new ma chines for battlefield operations. "We want some architectures that are good for building semantic memories, memories that can hold knowledge, [and] we want architectures that can do very rapid signal processing [and] structures that can handle very, very large amounts of data in communications,'' said Defense Advanced Re search Projects Agency (DARPA) computer director Robert Kahn. Flowever, the demand for Al is everywhere, from small entrepreneurial software firms to giants like Control Data and IBM. MIT Artificial Intelligence Lab Director Patrick FI. Winston predicts tremen dous growth in expert systems for finance and banking. Financial analyst Andrew Tobias has al ready put himself on a disk. Advances in machine vision and motion will rev olutionize traditional engineering industries. And, in another instance of self-replication, expert sys tems for circuit design are taking over the tedious details of laying out transistors on a chip, allowing engineers to focus on the overall design. Says Kenneth J. Meltsner, an MIT graduate stu dent in materials science who is writing a therma- dynamics simulator to help sophomore engineering students grasp complex concepts, "Artificial intelli gence techniques are becoming simply good engi neering practice." ◄ 6 BEYOND October 1985