AI expert systems are often created by extracting the specialized knowledge of human experts, artistically represented by glimmering lightbulbs in this drawing. The expert information is converted into a set of probabilistic rules.

1965

EXPERT SYSTEMS

According to journalist Luke Dormehl, AI “expert systems” are “attempts to create clones of flesh-and-blood human experts . . . by extracting their specialized knowledge and turning it into a set of probabilistic rules.” In a best-case scenario, expert systems could, in principle, be used to cram the know-how of an expert gastroenterologist, financial adviser, or lawyer into a computerized device, and have the resultant AI systems give useful advice to all.

Expert systems started to be explored in the 1960s and made use of a knowledge base (containing representations of facts and rules) and inference engines (to apply the rules and perform evaluations). Rules might be of an “if-then” form, such as “If a patient with a particular demographic exhibits a particular symptom, then there is a certain probability that he or she has a particular condition.”

Applications of expert systems may broadly include diagnosis, prediction, planning, classifying, and related areas involving specialized domains of expertise, in areas ranging from medicine to evaluating insurance risks or potential locations for mineral exploration. Useful expert systems also often have inference engines that can provide explanations, so that users can understand the chain of reasoning. An example of a famous early expert system is Dendral (short for Dendritic Algorithm), a Stanford University project started in 1965 to help chemists identify unknown organic molecules based on information from mass spectra. Another famous early example is MYCIN, a Stanford University AI system developed in the 1970s to help diagnose bacterial infections and recommend antibiotics and dosages. Early expert systems were often coded in LISP or Prolog.

One of the challenges of expert systems often involves acquiring and codifying knowledge from busy experts in particular fields, or from books and papers. It can also be challenging to organize knowledge into a collection of facts and rules that experts agree on, along with various numerical weights applied (to signify likelihood or importance). Today, many people use “recommender systems,” a somewhat related area of AI that is more focused on predicting user preferences in areas ranging from movies and books to financial services and potential marriage partners.