Robotics Reference
In-Depth Information
or changed, without the need to modify the program itself, so new in-
formation (i.e., new rules) may be easily accommodated. The ease with
which the rule-base can be changed is a boon to expert systems designers
and users—bring on a new expert, or have the original expert(s) add new
rules, and the rule-base can be improved rather easily and quickly. An-
other advantage of employing a knowledge-base of rules in the system is
that knowledge in this form is active, in that it can be used to infer new
information from what is already known about a problem.
Expert systems are deliberately made narrow in their domain of ex-
pertise because, like human specialists, by knowing more about less, the
system is able to perform with a higher level of understanding within the
chosen problem domain. For example, there is no medical diagnosis ex-
pert system that encompasses the entire range of medical knowledge, but
there are several systems that function in specialist areas of medicine and
which are able to perform at the level of leading human experts in these
fields. One of the earliest and best-known expert systems, MYCIN, was
developed at Stanford University in the 1970s to diagnose certain types
of blood infection, and in tests the system was able to outperform even
some members of the Stanford Medical School. Even today, after three
decades, MYCIN remains a classic example of how expert systems are
designed and how they function. But although such a level of speciali-
sation clearly has its advantages, it also means that the users of an expert
system must take care to employ it only within the appropriate domain;
otherwise the system could produce nonsense. One researcher, who used
a skin disease expert system to diagnose problems with his rusty old car,
was advised that the car had probably developed measles!
The Structure of an Expert System
Figure 49 shows the structure of a typical expert system program. The
user interacts with the system through the user interface, which will
normally ask questions via menus or using natural language. The user
interface sits between what is called the inference engine and the user.
It translates the system's answers from the internal representation cre-
ated by the program to something the user can understand, and it passes
questions from the system to the user and checks that the user's replies
are valid. For example, it would query or reject a ludicrously low num-
ber as the answer to a request for your weight. The inference engine
is employed to do the reasoning for the system, using both the expert
