EXPERT SYSTEMS

by LARRY LEVITT

A doctor types a patient's symptoms into a computer and gets back a list of possible causes ...
  An oil geologist supplies the computer with site data and is told the best spot to start drilling ...
  A chemist inputs a description of a possible chemical pollutant and the computer identifies the compound ...
  These are some of the more common real-world examples of how computers use expert systems software to effectively perform research analysis that could once only be done by highly trained human technical experts.
  Expert systems are one of the three areas of artificial intelligence (AI) research. The other two categories are robotics and natural language communication.
  The idea behind expert systems is that a computer program can simulate human expertise by manipulating large stores of properly arranged knowledge.
  AI researchers divide knowledge into two distinct types. The first type is axioms-facts accepted as indisputable. The second type is rules-which computers have traditionally handled. as If ... Then statements.
  For example, a fact might "Socrates is a man." And a relevant rule might be, "If someone is a man . . . Then he is mortal."
  An expert system is primarily a collection of such snatches of "knowledge" - often over 1,000 of them in the most complex systems.
  Of course, what's needed is an algorithm that forms correct conclusions from these bits of knowledge. AI researchers call this part of the system an "inference engine," or shell.
  Shells are generally written in the language LISP (LISt Processing), mainly because of its ease in defining recursive functions and its powerful manipulation of symbols.
  However, LISP programs are extremely slow. So most expert systems are run on dedicated "LISP machines" which are large minicomputers devoted solely to interpreting LISP programs.
  Shells normally use either "forward-chaining" or "backward-chaining" techniques to generate conclusions. Forward-chaining means that the system begins with the axioms and rules, then reviews conclusions- much like one might prove a theorem in geometry. A backward-chaining system begins with a hypothesis to be proved, and then proceeds to determine what the system must know in order to prove it.
  Stand-alone shells, or "knowledge engineering tools," have attracted recent commercial interest. Users buy just the shell and then compile the knowledge base themselves.
  This opens up the market substantially. Knowledge engineers (as programmers in the field are called) can develop widely applicable shells, instead of designing complete systems which might be only useful to a few highly specialized users.
  SRI International of Palo Alto is currently selling a $20,000 expert system shell called Series, for the IBM PC XT. The system was developed in a garage by Ray Weinstock, who was subsequently hired on at SRI.
  Puff is a medical diagnosis system for respiratory ailments. Written in BASIC, the system has only about 100 rules in its knowledge base.
  The best seller among microcomputer expert systems to date is Human Edge's line of software that provides psychological advice on the best way to negotiate business and personal dealings. These programs sell for a few hundred dollars each. According to Fortune magazine, Human Edge grossed $1.8 million from sales of 10,000 programs in the first half of 1984.
  Current expert systems primarily rely on simple symbolic manipulations of rules and facts. There is no attempt to have the software examine causality-WHY a particular conclusion seems to be true. The danger here is that rules could be applied incorrectly, leading to faulty or possibly disastrous results. Simple human common sense is still needed as a fail-safe.
  Even users of today's large over-1,000-rules expert systems have a hard time seeing how a particular decision was arrived at. There have been attempts to address this problem. Some systems attempt to explain the process they are going through. Incidentally, most expert systems use some sort of natural language interface, meaning that they appear conversational.
  The discipline of artificial intelligence is still in its infancy. But even today's comparatively simple applications based on simple programming techniques are breaking new ground and achieving highly promising results.
  Larry Levitt is a student at Harvard's Kennedy School of Government. His primary interest is the field of science, technology and society
  Antic is actively seeking more information, programs and articles which might help our readers understand the new field of artificial intelligence We believe AI represents one of the most exciting computer frontiers, and we will continue to explore this new field.