[ Present
Problems from Inference ]
ven as successful as expert systems have been, the strict "If..
Then" rules reaffirms how unintelligent computers are. Like an actor reading a
script, early expert systems needed complete information in order to make any kind of
decision. However, in the real world, people almost never know everything about a
particular subject, but they can still make decisions. On the contrary, people can infer--to
derive additional information from what is presented--to make up for lack of knowledge by
making assumptions from what is known.
In AI, an inference engine is a system of applying the rules on a knowledge base.
It is these rules that allow a computer to perform deductive and inductive
reasoning. The car-not-starting problem diagnosis problem in the section "Brain & Computer--Indexing" is an example of deduction
with the computer going from general knowledge("the car won't start") to
specific("the car has no gas"). Conversely, induction, otherwise known as
generalization, goes from the specific to the general. In an expert system, it
merely involves back-tracking up the decision-tree.(Kurzweil 293)
[ Top ]
Because nearly all computers today are based on the von Neumann design, they perform
calculation serially. In other words, computers process things one at a time.
For inference engines, this makes computers perform adequately when it has to work its way
through a small and simple decision-tree and knowledge base. However, as knowledge
bases become larger, the decision-trees become exponentially more complex which translates
to a more inferences for a computer to perform to reach a conclusion. This is where
the recent development of parallel-computing
can help. With numerous processors working on different branches of a decision-tree
at the same time, it would speed up expert systems exponentially.
