Artffkfd Isadlfgesw: Using Computers to Tlslnk about Thirskfng. Part 2. Some Pmctfcsd Applicatfom of AI Number 52 December 26, 1983 This is the second of a two-part essay liited success. Techniques developed on artificial intelligence (AI). Part 1 dis- for solving problems in one domain were cussed “knowledge representations”- usually inadequate for other domains. models of cognition that AI investigators However, programs equipped with a have used in their attempts to build great deal of information about a single thinking machines. 1 One of the most domain performed as well as, and some- ambitious goals of AI research is to de- times better than, experts in that field.z velop programs which enable machines The superior performance of these to perform commonsense reasoning. It knowledge-based, or expert, systems will & many years, however, before th~ convinced many AI researchers that goal is achieved. problem solving demands huge banks of In the meantime, AI research has knowledge as well as reasoning proce- spawned numerous spin-offs that are dures.3 just beginning to enter the commercial Today, the goal of expert systems re- market. These include programs that en- search is to transfer a specialist’s knowl- able robots to “see” and “feel” and ma- edge into a program so the information chines which can follow instructions can be efficiently accessed and used by written in natural language (NL). But the the computer to solve problems. This in- most ambitious and successful AI appli- cludes “textbook leaming’’—the facts cations thus far are “expert systems. ” obtained from training and reading.4 It These computer programs are designed also includes heuristic knowledge, or to duplicate the problem-solving pro- rules of thumb developed through years cesses of experts in various fields. This of experience and judgment. Heuristics essay will cover some of these expert sys- are essentially educated guesses about tems and review a liited number of which solutions to a problem are most other applications of AI. likely to be successful. They don’t guar- During the early years of AI research, antee correct answers. But they do save investigators were intent upon dMcover- time by limiting the search for solutions ing the general principles underlying in- to those most likely to be correct. Com- telligence. In developing the knowledge puter and other scientists called “knowl- representations described in Part 1, AI edge engineers” sometimes spend years researchers tried to use these principles picking experts’ brains for these facts to create an “inference engine.” Ideally, and heuristics, and then structure them such an engine would solve any type of into computer programs. problem, from winning at chess to d~ag- Expert systems also include “infer- nosing disease. But attempts at develop- ence procedures” that determine which ing computer-based problem-solving heuristics and facts should be brought to strategies for any situation met with bear on a problem. One such inference 430 procedure is backward chaiiing, in in modeliig and assisting scientific which you suggest a possible solution to thinkiig, and Lederberg, a geneti- a problem and work backward to see if cist/molecular biologist with expertise it’s correct. MYCIPJ, an expert system in chemistry, had developed a computer that assists in medical diagnoses, reasons language for describing the structure of in this manner.z It advances a disease complex molecules.g As the project hypothesis based on a few known symp- grew, Carl Djerassi, also at Stanford tom$. Then it looks for other symptoms Univemity, contributed his expertise. 10 which support the hypothesis, request- The original DENDRAL program per- ing additional tests and information as formed three basic operations on the needed. In most expert systems, the in- chemical formulas and mass spectral ference procedure for deciding which data with which it was provided. In the facts and heuristics to use is separate fmt phase, cafled “plan,” it translated from the knowledge base of facts and general and specific prior knowledge rules. This makes it easier to add or and heuristics into a spec~lc repertoire modify facts and rules as new informa- of constraints. In the next phase, called tion becomes available. “generation,” it generated plausible The fret, and probably best-known, structures based on such constraints as AI programs that focused on a Ihnited the number of rings, double bonds, and domain were chess programs. Donald atoms of various types in each molecule. Michie, University of Edinburgh, Scot- In the final phase, caUed “test,” each land, explains that chess is ideal for mod- plausible structure was tested. The com- eliig specialist knowledge because it is a puter fiit generated sets of instrument very well-defined domains A large data that would be expected to describe amount of formal information is avail- each structure. Then it compared each able in the form of instructional works set of data to actual data about the com- and commentaries. And numerical pound. The closest fits were then ranked scales of performance are avaifable in for the user. the national and international rating In the past few years, scientists work- systems. Equally important, chess is a ing on the DENDRAL project have fo- game that calls on a wide range of cogni- cused most of their attention on the tive functions, from logical calculation planning and generation portion of the to imaginative thinking. The numerous program, and on making this portion chess-playing programs developed in the available to users. Called CONGEN, for 1950s and 1960s tested the proficiency constrained structure generation, this with which various knowledge represen- generator has been expanded to infer tations used facts and heuristics to solve plausible structures using a wide variety problems.s of instrumental data. In 1982. CONGEN By the mid- 1960s, AI researchers was made commercially avaifable began to expand beyond chess and puz- through the computer network Compu- zle playing, or what Edward A. Feigen- Serve. II baum, Stanford University, California, Another spin-off of DENDRAL is calls “toy problems, “G(p. 62) into practi- Mets-DENDRAL, a program that gen- cal problems. The fiist such project re- erates its own rules from mass spectral sulted in DENDRAL, an expert system data on chemical compounds. After re- that identifies the chemicaf structures of ceiving mass spectral data on a family of unknown compounds.7,8 DENDRAL compounds, Mets-DENDRAL gener- was launched at Stanford University in ates planning and test rules that describe 1966 by Feigenbaum and Joshua Leder- how these compounds fragment when berg, now president of Rockefeller Uni- studied with mass spectrometry. Some versity, New York. Both were interested of the roles generated by Meta- 431 DENDRAL duplicated those formulated solve problems in that domain. Figure 1 by expert chemists, while others were shows one of the 500 such rules used by entirely original. 1Z the infectious diseases expert system DENDRAL demonstrated that AI MYCIN. With MYCIN, a physician en- techniques could be used to solve real ters information about a patient into the problems within a Iiiited area of knowl- computer. The computer then searches edge.b Paradoxically, it also demon- for the rules that can be applied to this strated that it is easier to model the rea- information. If more information is soning processes of specialists than to needed, the computer will ask the physi- program the steps a chfld goes through cian to supply it. Then the rules for these in understanding language, or making additional data are applied. This process commonsense inferences.z This is continues until a diagnosis, and treat- because the facts and judgments an ex- ment, can be recommended. pert uses in making a decision are easier Since the acceptability of an expert to identify and categorize than are the system depends on the confidence with reasoning processes used for general which physicians can accept its sugges- problem solving. tions, MYCIN and several other consul- So far, the most successful expert sys- tation systems can provide explanations tems have been programs that weigh and of their reasoning processes. At the phy- balance evidence about data to deter- sician’s request, MYCIN will list the pro- mine how they should be categorized. duction rules it used in its diagnosis or Differential diagnosis, for example, is “a treatment recommendation, and cite classical medical example of such a references to the literature that support problem,”z according to Richard O. these rules. MYCIN will also explain Duda, Syntelligence, Menlo Park, Cali- why it has requested additional tests or fornia, and Edward H. Shortliife, Stan- other information.z An example of thk ford University School of Medicine. A explanation feature is shown in Figure 2, physician arrives at a diagnosis by evaluating a variety of symptoms and Ffgure 1: Sample MYCIN production rule. test results. Although this is a fairly complicated procedure, it is based on Mycfri identifiable facts and heuristics and, [f: 1) The infection which requires therefore, lends itself to computer therapy is meningitis, modeling. For this reason, and because And 2) The patient has evidence of a computers can consider many diseases serious skin or soft tissue infection, that physicians might not encounter in everyday practice, numerous expert And 3) Organisms were not seen on the systems have been designed to assist stain of the culture, doctors in diagnosing and treating disease.z Amf 4) The type of infection is bacterial, The knowledge representation used risen: most widely in these expert or “consulta- tion” systems is the “production rule” There is evidence that the organism approach, according to Wi~lam B, (other than those seen on cultures or Gevarter, National Aeronautics and smears) which might be causing the in- Space Administration, Washington, fection is staphyloccwcus-coagpos (.75) DC.